madwifi/net80211/ieee80211_node.c

2517 lines
72 KiB
C
Raw Normal View History

/*-
* Copyright (c) 2001 Atsushi Onoe
* Copyright (c) 2002-2005 Sam Leffler, Errno Consulting
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. The name of the author may not be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* Alternatively, this software may be distributed under the terms of the
* GNU General Public License ("GPL") version 2 as published by the Free
* Software Foundation.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
* THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* $Id$
*/
#ifndef EXPORT_SYMTAB
#define EXPORT_SYMTAB
#endif
/*
* IEEE 802.11 node handling support.
*/
#ifndef AUTOCONF_INCLUDED
#include <linux/config.h>
#endif
#include <linux/version.h>
#include <linux/module.h>
#include <linux/skbuff.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/random.h>
#include "if_media.h"
#include <net80211/ieee80211_var.h>
#include <net80211/if_athproto.h>
/*
* Association IDs are managed with a bit vector.
*/
#define IEEE80211_AID_SET(_vap, _b) \
((_vap)->iv_aid_bitmap[IEEE80211_AID(_b) / 32] |= \
(1 << (IEEE80211_AID(_b) % 32)))
#define IEEE80211_AID_CLR(_vap, _b) \
((_vap)->iv_aid_bitmap[IEEE80211_AID(_b) / 32] &= \
~(1 << (IEEE80211_AID(_b) % 32)))
#define IEEE80211_AID_ISSET(_vap, _b) \
((_vap)->iv_aid_bitmap[IEEE80211_AID(_b) / 32] & \
(1 << (IEEE80211_AID(_b) % 32)))
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
#define ieee80211_alloc_node(_vap, _mac) \
ieee80211_alloc_node_debug(_vap, _mac, __func__, __LINE__)
static struct ieee80211_node *ieee80211_alloc_node_debug(struct ieee80211vap *,
const u_int8_t *, const char *func, int line);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
static struct ieee80211_node *ieee80211_alloc_node(struct ieee80211vap *,
const u_int8_t *);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
static int ieee80211_sta_join1(struct ieee80211_node *);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
static struct ieee80211_node *node_alloc_debug(struct ieee80211vap *,
const char *func, int line);
static void node_cleanup_debug(struct ieee80211_node *, const char *func,
int line);
static void node_free_debug(struct ieee80211_node *, const char *func,
int line);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else /* #ifdef IEEE80211_DEBUG_REFCNT */
static struct ieee80211_node *node_alloc(struct ieee80211vap *);
static void node_cleanup(struct ieee80211_node *);
static void node_free(struct ieee80211_node *);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
static int32_t node_count(struct ieee80211com *ic);
static u_int8_t node_getrssi(const struct ieee80211_node *);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
#define node_table_leave_locked(_table, _node) \
node_table_leave_locked_debug(_table, _node, __func__, __LINE__)
static void node_table_leave_locked_debug(struct ieee80211_node_table *,
struct ieee80211_node *, const char *func, int line);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
static void node_table_leave_locked(struct ieee80211_node_table *,
struct ieee80211_node *);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
#ifdef IEEE80211_DEBUG_REFCNT
#define node_table_join_locked(_table, _node) \
node_table_join_locked_debug(_table, _node, __func__, __LINE__)
static void node_table_join_locked_debug(struct ieee80211_node_table *,
struct ieee80211_node *, const char *func, int line);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
static void node_table_join_locked(struct ieee80211_node_table *,
struct ieee80211_node *);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
static void ieee80211_node_timeout(unsigned long);
static void ieee80211_node_table_init(struct ieee80211com *,
struct ieee80211_node_table *, const char *, int);
static void ieee80211_node_table_cleanup(struct ieee80211_node_table *);
static void ieee80211_node_table_reset(struct ieee80211_node_table *,
struct ieee80211vap *);
static void ieee80211_node_wds_ageout(unsigned long);
MALLOC_DEFINE(M_80211_NODE, "80211node", "802.11 node state");
void
ieee80211_node_attach(struct ieee80211com *ic)
{
ieee80211_node_table_init(ic, &ic->ic_sta, "station",
IEEE80211_INACT_INIT);
init_timer(&ic->ic_inact);
ic->ic_inact.function = ieee80211_node_timeout;
ic->ic_inact.data = (unsigned long) ic;
mod_timer(&ic->ic_inact, jiffies + IEEE80211_INACT_WAIT * HZ);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ic->ic_node_alloc_debug = node_alloc_debug;
ic->ic_node_free_debug = node_free_debug;
ic->ic_node_cleanup_debug = node_cleanup_debug;
#else
ic->ic_node_alloc = node_alloc;
ic->ic_node_free = node_free;
ic->ic_node_cleanup = node_cleanup;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
ic->ic_node_count = node_count;
ic->ic_node_getrssi = node_getrssi;
}
void
ieee80211_node_detach(struct ieee80211com *ic)
{
del_timer(&ic->ic_inact);
ieee80211_node_table_cleanup(&ic->ic_sta);
}
void
ieee80211_node_vattach(struct ieee80211vap *vap)
{
/* default station inactivity timer setings */
vap->iv_inact_init = IEEE80211_INACT_INIT;
vap->iv_inact_auth = IEEE80211_INACT_AUTH;
vap->iv_inact_run = IEEE80211_INACT_RUN;
vap->iv_inact_probe = IEEE80211_INACT_PROBE;
}
void
ieee80211_node_latevattach(struct ieee80211vap *vap)
{
struct ieee80211com *ic = vap->iv_ic;
struct ieee80211_rsnparms *rsn;
/*
* Allocate these only if needed. Beware that we
* know adhoc mode doesn't support ATIM yet...
*/
if (vap->iv_opmode == IEEE80211_M_HOSTAP) {
if (vap->iv_max_aid == 0)
vap->iv_max_aid = IEEE80211_AID_DEF;
else if (vap->iv_max_aid > IEEE80211_AID_MAX)
vap->iv_max_aid = IEEE80211_AID_MAX;
MALLOC(vap->iv_aid_bitmap, u_int32_t *,
howmany(vap->iv_max_aid, 32) * sizeof(u_int32_t),
M_DEVBUF, M_NOWAIT | M_ZERO);
if (vap->iv_aid_bitmap == NULL) {
/* XXX no way to recover */
printk(KERN_ERR "%s: no memory for AID bitmap!\n", __func__);
vap->iv_max_aid = 0;
}
}
ieee80211_reset_bss(vap);
/*
* Setup "global settings" in the bss node so that
* each new station automatically inherits them.
*/
rsn = &vap->iv_bss->ni_rsn;
/* WEP, TKIP, and AES-CCM are always supported */
rsn->rsn_ucastcipherset |= 1 << IEEE80211_CIPHER_WEP;
rsn->rsn_ucastcipherset |= 1 << IEEE80211_CIPHER_TKIP;
rsn->rsn_ucastcipherset |= 1 << IEEE80211_CIPHER_AES_CCM;
if (ic->ic_caps & IEEE80211_C_AES)
rsn->rsn_ucastcipherset |= 1 << IEEE80211_CIPHER_AES_OCB;
if (ic->ic_caps & IEEE80211_C_CKIP)
rsn->rsn_ucastcipherset |= 1 << IEEE80211_CIPHER_CKIP;
/*
* Default unicast cipher to WEP for 802.1x use. If
* WPA is enabled the management code will set these
* values to reflect.
*/
rsn->rsn_ucastcipher = IEEE80211_CIPHER_WEP;
rsn->rsn_ucastkeylen = 104 / NBBY;
/*
* WPA says the multicast cipher is the lowest unicast
* cipher supported. But we skip WEP which would
* otherwise be used based on this criteria.
*/
rsn->rsn_mcastcipher = IEEE80211_CIPHER_TKIP;
rsn->rsn_mcastkeylen = 128 / NBBY;
/*
* We support both WPA-PSK and 802.1x; the one used
* is determined by the authentication mode and the
* setting of the PSK state.
*/
rsn->rsn_keymgmtset = WPA_ASE_8021X_UNSPEC | WPA_ASE_8021X_PSK;
rsn->rsn_keymgmt = WPA_ASE_8021X_PSK;
vap->iv_auth = ieee80211_authenticator_get(vap->iv_bss->ni_authmode);
}
void
ieee80211_node_vdetach(struct ieee80211vap *vap)
{
struct ieee80211com *ic = vap->iv_ic;
ieee80211_node_table_reset(&ic->ic_sta, vap);
if (vap->iv_bss != NULL) {
ieee80211_unref_node(&vap->iv_bss);
}
if (vap->iv_aid_bitmap != NULL) {
FREE(vap->iv_aid_bitmap, M_DEVBUF);
vap->iv_aid_bitmap = NULL;
}
}
/* Port authorize/unauthorize interfaces for use by an authenticator. */
void
ieee80211_node_authorize(struct ieee80211_node *ni)
{
ni->ni_flags |= IEEE80211_NODE_AUTH;
ni->ni_inact_reload = ni->ni_vap->iv_inact_run;
}
EXPORT_SYMBOL(ieee80211_node_authorize);
void
ieee80211_node_unauthorize(struct ieee80211_node *ni)
{
ni->ni_flags &= ~IEEE80211_NODE_AUTH;
}
EXPORT_SYMBOL(ieee80211_node_unauthorize);
/*
* Set/change the channel. The rate set is also updated
* to ensure a consistent view by drivers.
*/
static __inline void
ieee80211_node_set_chan(struct ieee80211com *ic, struct ieee80211_node *ni)
{
struct ieee80211_channel *chan = ic->ic_bsschan;
KASSERT(chan != IEEE80211_CHAN_ANYC, ("bss channel not setup"));
ni->ni_chan = chan;
#ifdef ATH_SUPERG_XR
if (ni->ni_vap->iv_flags & IEEE80211_F_XR)
ni->ni_rates = ic->ic_sup_xr_rates;
else
#endif
ni->ni_rates = ic->ic_sup_rates[ieee80211_chan2mode(chan)];
}
static __inline void
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
copy_bss_state(struct ieee80211_node *nbss, const struct ieee80211_node *obss)
{
/* propagate useful state */
nbss->ni_authmode = obss->ni_authmode;
nbss->ni_ath_flags = obss->ni_ath_flags;
nbss->ni_txpower = obss->ni_txpower;
nbss->ni_vlan = obss->ni_vlan;
nbss->ni_rsn = obss->ni_rsn;
nbss->ni_rates = obss->ni_rates;
/* XXX statistics? */
}
void
ieee80211_create_ibss(struct ieee80211vap *vap, struct ieee80211_channel *chan)
{
struct ieee80211com *ic = vap->iv_ic;
struct ieee80211_node *ni;
IEEE80211_DPRINTF(vap, IEEE80211_MSG_SCAN,
"%s: creating ibss on channel %u\n", __func__,
ieee80211_chan2ieee(ic, chan));
/* Check to see if we already have a node for this mac
* NB: we gain a node reference here
*/
ni = ieee80211_find_node(&ic->ic_sta, vap->iv_myaddr);
if (ni == NULL) {
ni = ieee80211_alloc_node_table(vap, vap->iv_myaddr);
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: ni:%p allocated for " MAC_FMT "\n",
__func__, ni, MAC_ADDR(vap->iv_myaddr));
if (ni == NULL) {
/* XXX recovery? */
return;
}
}
IEEE80211_ADDR_COPY(ni->ni_bssid, vap->iv_myaddr);
IEEE80211_ADDR_COPY(vap->iv_bssid, vap->iv_myaddr);
ni->ni_esslen = vap->iv_des_ssid[0].len;
memcpy(ni->ni_essid, vap->iv_des_ssid[0].ssid, ni->ni_esslen);
if (vap->iv_bss != NULL)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
copy_bss_state(ni, vap->iv_bss);
ni->ni_intval = ic->ic_lintval;
#ifdef ATH_SUPERG_XR
if (vap->iv_flags & IEEE80211_F_XR) {
ni->ni_intval *= IEEE80211_XR_BEACON_FACTOR;
}
#endif
if (vap->iv_flags & IEEE80211_F_PRIVACY)
ni->ni_capinfo |= IEEE80211_CAPINFO_PRIVACY;
if (ic->ic_phytype == IEEE80211_T_FH) {
ni->ni_fhdwell = 200; /* XXX */
ni->ni_fhindex = 1;
}
if (vap->iv_opmode == IEEE80211_M_IBSS) {
vap->iv_flags |= IEEE80211_F_SIBSS;
ni->ni_capinfo |= IEEE80211_CAPINFO_IBSS; /* XXX */
if (vap->iv_flags & IEEE80211_F_DESBSSID) {
IEEE80211_ADDR_COPY(ni->ni_bssid, vap->iv_des_bssid);
IEEE80211_ADDR_COPY(vap->iv_bssid, vap->iv_des_bssid);
} else {
ni->ni_bssid[0] |= 0x02; /* local bit for IBSS */
vap->iv_bssid[0] |= 0x02;
}
} else if (vap->iv_opmode == IEEE80211_M_AHDEMO) {
if (vap->iv_flags & IEEE80211_F_DESBSSID) {
IEEE80211_ADDR_COPY(ni->ni_bssid, vap->iv_des_bssid);
IEEE80211_ADDR_COPY(vap->iv_bssid, vap->iv_des_bssid);
} else {
IEEE80211_ADDR_SET_NULL(ni->ni_bssid);
IEEE80211_ADDR_SET_NULL(vap->iv_bssid);
}
}
#ifdef ATH_SUPERG_DYNTURBO
if (vap->iv_opmode == IEEE80211_M_HOSTAP) {
ni->ni_ath_flags = vap->iv_ath_cap;
/*
* no dynamic turbo and AR on a static turbo channel.
* no dynamic turbo and AR on non-turbo channel.
* no AR on 5GHZ channel .
*/
if (IEEE80211_IS_CHAN_STURBO(chan) ||
!ieee80211_find_channel(ic, chan->ic_freq, chan->ic_flags | IEEE80211_CHAN_TURBO))
ni->ni_ath_flags &= ~(IEEE80211_ATHC_TURBOP | IEEE80211_ATHC_AR);
if (IEEE80211_IS_CHAN_5GHZ(chan))
ni->ni_ath_flags &= ~IEEE80211_ATHC_AR;
}
#endif
/*
* Fix the channel and related attributes.
*/
ic->ic_bsschan = chan;
ieee80211_node_set_chan(ic, ni);
ic->ic_curmode = ieee80211_chan2mode(chan);
/* Update country ie information */
ieee80211_build_countryie(ic);
if (IEEE80211_IS_CHAN_HALF(chan))
ni->ni_rates = ic->ic_sup_half_rates;
else if (IEEE80211_IS_CHAN_QUARTER(chan))
ni->ni_rates = ic->ic_sup_quarter_rates;
(void) ieee80211_sta_join1(PASS_NODE(ni));
}
EXPORT_SYMBOL(ieee80211_create_ibss);
/*
* Reset bss state on transition to the INIT state.
* Clear any stations from the table (they have been
* deauth'd) and reset the bss node (clears key, rate,
* etc. state).
*/
void
ieee80211_reset_bss(struct ieee80211vap *vap)
{
struct ieee80211com *ic = vap->iv_ic;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
struct ieee80211_node *ni = NULL;
struct ieee80211_node *obss = vap->iv_bss;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
/* Recreate the node table */
ieee80211_node_table_reset(&ic->ic_sta, vap);
/* XXX multi-bss wrong */
ieee80211_reset_erp(ic, ic->ic_curmode);
ni = ieee80211_alloc_node_table(vap, vap->iv_myaddr);
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: ni:%p allocated for " MAC_FMT "\n",
__func__, ni, MAC_ADDR(vap->iv_myaddr));
KASSERT(ni != NULL, ("unable to setup inital BSS node"));
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
vap->iv_bss = PASS_NODE(ni);
KASSERT((atomic_read(&vap->iv_bss->ni_refcnt) == 2),
("wrong refcount for new node."));
if (obss != NULL) {
copy_bss_state(vap->iv_bss, obss);
vap->iv_bss->ni_intval = ic->ic_lintval;
/* Caller's reference */
ieee80211_unref_node(&obss);
}
}
static int
match_ssid(const struct ieee80211_node *ni,
int nssid, const struct ieee80211_scan_ssid ssids[])
{
int i;
for (i = 0; i < nssid; i++) {
if (ni->ni_esslen == ssids[i].len &&
memcmp(ni->ni_essid, ssids[i].ssid, ni->ni_esslen) == 0)
return 1;
}
return 0;
}
/*
* Test a node for suitability/compatibility.
*/
static int
check_bss(struct ieee80211vap *vap, struct ieee80211_node *ni)
{
struct ieee80211com *ic = ni->ni_ic;
u_int8_t rate;
if (isclr(ic->ic_chan_active, ieee80211_chan2ieee(ic, ni->ni_chan)))
return 0;
if (vap->iv_opmode == IEEE80211_M_IBSS) {
if ((ni->ni_capinfo & IEEE80211_CAPINFO_IBSS) == 0)
return 0;
} else {
if ((ni->ni_capinfo & IEEE80211_CAPINFO_ESS) == 0)
return 0;
}
if (vap->iv_flags & IEEE80211_F_PRIVACY) {
if ((ni->ni_capinfo & IEEE80211_CAPINFO_PRIVACY) == 0)
return 0;
} else {
/* Reference: IEEE802.11 7.3.1.4
* This means that the data confidentiality service is required
* for all frames exchanged with this STA in IBSS and for all
* frames exchanged within the entire BSS otherwise
*/
if (ni->ni_capinfo & IEEE80211_CAPINFO_PRIVACY)
return 0;
}
rate = ieee80211_fix_rate(ni, IEEE80211_F_DONEGO | IEEE80211_F_DOFRATE);
if (rate & IEEE80211_RATE_BASIC)
return 0;
if (vap->iv_des_nssid != 0 &&
!match_ssid(ni, vap->iv_des_nssid, vap->iv_des_ssid))
return 0;
if ((vap->iv_flags & IEEE80211_F_DESBSSID) &&
!IEEE80211_ADDR_EQ(vap->iv_des_bssid, ni->ni_bssid))
return 0;
return 1;
}
#ifdef IEEE80211_DEBUG
/*
* Display node suitability/compatibility.
*/
static void
check_bss_debug(struct ieee80211vap *vap, struct ieee80211_node *ni)
{
struct ieee80211com *ic = ni->ni_ic;
u_int8_t rate;
int fail;
fail = 0;
if (isclr(ic->ic_chan_active, ieee80211_chan2ieee(ic, ni->ni_chan)))
fail |= 0x01;
if (vap->iv_opmode == IEEE80211_M_IBSS) {
if ((ni->ni_capinfo & IEEE80211_CAPINFO_IBSS) == 0)
fail |= 0x02;
} else {
if ((ni->ni_capinfo & IEEE80211_CAPINFO_ESS) == 0)
fail |= 0x02;
}
if (vap->iv_flags & IEEE80211_F_PRIVACY) {
if ((ni->ni_capinfo & IEEE80211_CAPINFO_PRIVACY) == 0)
fail |= 0x04;
} else {
/* This means that the data confidentiality service is required
* for all frames exchanged within this BSS. (IEEE802.11 7.3.1.4)
*/
if (ni->ni_capinfo & IEEE80211_CAPINFO_PRIVACY)
fail |= 0x04;
}
rate = ieee80211_fix_rate(ni, IEEE80211_F_DONEGO | IEEE80211_F_DOFRATE);
if (rate & IEEE80211_RATE_BASIC)
fail |= 0x08;
if (vap->iv_des_nssid != 0 &&
!match_ssid(ni, vap->iv_des_nssid, vap->iv_des_ssid))
fail |= 0x10;
if ((vap->iv_flags & IEEE80211_F_DESBSSID) &&
!IEEE80211_ADDR_EQ(vap->iv_des_bssid, ni->ni_bssid))
fail |= 0x20;
printk(" %c " MAC_FMT, fail ? '-' : '+', MAC_ADDR(ni->ni_macaddr));
printk(" " MAC_FMT "%c", MAC_ADDR(ni->ni_bssid), fail & 0x20 ? '!' : ' ');
printk(" %3d%c",
ieee80211_chan2ieee(ic, ni->ni_chan), fail & 0x01 ? '!' : ' ');
printk(" %+4d", ni->ni_rssi);
printk(" %2dM%c", (rate & IEEE80211_RATE_VAL) / 2,
fail & 0x08 ? '!' : ' ');
printk(" %4s%c",
(ni->ni_capinfo & IEEE80211_CAPINFO_ESS) ? "ess" :
(ni->ni_capinfo & IEEE80211_CAPINFO_IBSS) ? "ibss" :
"????",
fail & 0x02 ? '!' : ' ');
printk(" %3s%c ",
(ni->ni_capinfo & IEEE80211_CAPINFO_PRIVACY) ? "wep" : "no",
fail & 0x04 ? '!' : ' ');
ieee80211_print_essid(ni->ni_essid, ni->ni_esslen);
printk("%s\n", fail & 0x10 ? "!" : "");
}
#endif /* IEEE80211_DEBUG */
/*
* Handle 802.11 ad hoc network merge. The
* convention, set by the Wireless Ethernet Compatibility Alliance
* (WECA), is that an 802.11 station will change its BSSID to match
* the "oldest" 802.11 ad hoc network, on the same channel, that
* has the station's desired SSID. The "oldest" 802.11 network
* sends beacons with the greatest TSF timestamp.
*
* The caller is assumed to validate TSFs before attempting a merge.
*
* Return !0 if the BSSID changed, 0 otherwise.
*/
int
ieee80211_ibss_merge(struct ieee80211_node *ni)
{
struct ieee80211vap *vap = ni->ni_vap;
#ifdef IEEE80211_DEBUG
struct ieee80211com *ic = ni->ni_ic;
#endif
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: ni:%p[" MAC_FMT "] iv_bss:%p[" MAC_FMT "]\n",
__func__, ni, MAC_ADDR(ni->ni_macaddr),
vap->iv_bss, MAC_ADDR(vap->iv_bss->ni_macaddr));
if (!check_bss(vap, ni)) {
/* capabilities mismatch */
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: merge failed, capabilities mismatch\n", __func__);
#ifdef IEEE80211_DEBUG
if (ieee80211_msg_assoc(vap))
check_bss_debug(vap, ni);
#endif
vap->iv_stats.is_ibss_capmismatch++;
return 0;
}
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: new bssid " MAC_FMT ": %s preamble, %s slot time%s\n", __func__,
MAC_ADDR(ni->ni_bssid),
ic->ic_flags & IEEE80211_F_SHPREAMBLE ? "short" : "long",
ic->ic_flags & IEEE80211_F_SHSLOT ? "short" : "long",
ic->ic_flags & IEEE80211_F_USEPROT ? ", protection" : "");
return ieee80211_sta_join1(ieee80211_ref_node(ni));
}
EXPORT_SYMBOL(ieee80211_ibss_merge);
static __inline int
ssid_equal(const struct ieee80211_node *a, const struct ieee80211_node *b)
{
return (a->ni_esslen == b->ni_esslen &&
memcmp(a->ni_essid, b->ni_essid, a->ni_esslen) == 0);
}
/*
* Join the specified IBSS/BSS network. The node is assumed to
* be passed in with a reference already held for use in assigning
* to iv_bss.
*/
static int
ieee80211_sta_join1(struct ieee80211_node *selbs)
{
struct ieee80211vap *vap = selbs->ni_vap;
struct ieee80211com *ic = selbs->ni_ic;
struct ieee80211_node *obss;
int canreassoc;
if (vap->iv_opmode == IEEE80211_M_IBSS) {
/*
* Delete unusable rates; we've already checked
* that the negotiated rate set is acceptable.
*/
ieee80211_fix_rate(selbs, IEEE80211_F_DODEL);
}
/*
* Committed to selbs, setup state.
*/
obss = vap->iv_bss;
/*
* Check if old+new node have the same ssid in which
* case we can reassociate when operating in sta mode.
*/
canreassoc = ((obss != NULL) &&
(vap->iv_state == IEEE80211_S_RUN) && ssid_equal(obss, selbs));
vap->iv_bss = selbs;
IEEE80211_ADDR_COPY(vap->iv_bssid, selbs->ni_bssid);
if (obss != NULL)
ieee80211_unref_node(&obss);
ic->ic_bsschan = selbs->ni_chan;
ic->ic_curchan = ic->ic_bsschan;
ic->ic_curmode = ieee80211_chan2mode(ic->ic_curchan);
ic->ic_set_channel(ic);
/*
* Set the erp state (mostly the slot time) to deal with
* the auto-select case; this should be redundant if the
* mode is locked.
*/
ieee80211_reset_erp(ic, ic->ic_curmode);
ieee80211_wme_initparams(vap);
if (vap->iv_opmode == IEEE80211_M_STA) {
/*
* Act as if we received a DEAUTH frame in case we are
* invoked from the RUN state. This will cause us to try
* to re-authenticate if we are operating as a station.
*/
if (canreassoc) {
vap->iv_nsparams.newstate = IEEE80211_S_ASSOC;
vap->iv_nsparams.arg = IEEE80211_FC0_SUBTYPE_REASSOC_REQ;
IEEE80211_SCHEDULE_TQUEUE(&vap->iv_stajoin1tq);
} else {
vap->iv_nsparams.newstate = IEEE80211_S_AUTH;
vap->iv_nsparams.arg = IEEE80211_FC0_SUBTYPE_DEAUTH;
IEEE80211_SCHEDULE_TQUEUE(&vap->iv_stajoin1tq);
}
} else {
vap->iv_nsparams.newstate = IEEE80211_S_RUN;
vap->iv_nsparams.arg = -1;
IEEE80211_SCHEDULE_TQUEUE(&vap->iv_stajoin1tq);
}
return 1;
}
void
ieee80211_sta_join1_tasklet(IEEE80211_TQUEUE_ARG data)
{
struct ieee80211vap *vap= (struct ieee80211vap *)data;
int rc;
rc = ieee80211_new_state(vap, vap->iv_nsparams.newstate, vap->iv_nsparams.arg);
vap->iv_nsparams.result = rc;
vap->iv_nsdone = 1;
}
EXPORT_SYMBOL(ieee80211_sta_join1_tasklet);
int
ieee80211_sta_join(struct ieee80211vap *vap,
const struct ieee80211_scan_entry *se)
{
struct ieee80211com *ic = vap->iv_ic;
struct ieee80211_node *ni;
ni = ieee80211_find_node(&ic->ic_sta, se->se_macaddr);
if (ni == NULL) {
ni = ieee80211_alloc_node_table(vap, se->se_macaddr);
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: ni:%p allocated for " MAC_FMT "\n",
__func__, ni, MAC_ADDR(se->se_macaddr));
if (ni == NULL) {
IEEE80211_DPRINTF(vap, IEEE80211_MSG_NODE,
"%s: Unable to allocate node for BSS: " MAC_FMT "\n", __func__,
MAC_ADDR(ni->ni_macaddr));
return 0;
}
}
/*
* Expand scan state into node's format.
* XXX may not need all this stuff
*/
ni->ni_authmode = vap->iv_bss->ni_authmode; /* inherit authmode from iv_bss */
/* inherit the WPA setup as well (structure copy!) */
ni->ni_rsn = vap->iv_bss->ni_rsn;
IEEE80211_ADDR_COPY(ni->ni_bssid, se->se_bssid);
ni->ni_esslen = se->se_ssid[1];
memcpy(ni->ni_essid, se->se_ssid + 2, ni->ni_esslen);
ni->ni_rtsf = se->se_rtsf;
ni->ni_tstamp.tsf = se->se_tstamp.tsf;
ni->ni_intval = IEEE80211_BINTVAL_SANITISE(se->se_intval);
ni->ni_capinfo = se->se_capinfo;
ni->ni_chan = se->se_chan;
ni->ni_timoff = se->se_timoff;
ni->ni_fhdwell = se->se_fhdwell;
ni->ni_fhindex = se->se_fhindex;
ni->ni_erp = se->se_erp;
ni->ni_rssi = se->se_rssi;
if (se->se_wpa_ie != NULL)
ieee80211_saveie(&ni->ni_wpa_ie, se->se_wpa_ie);
if (se->se_rsn_ie != NULL)
ieee80211_saveie(&ni->ni_rsn_ie, se->se_rsn_ie);
if (se->se_wme_ie != NULL)
ieee80211_saveie(&ni->ni_wme_ie, se->se_wme_ie);
if (se->se_ath_ie != NULL)
ieee80211_saveath(ni, se->se_ath_ie);
vap->iv_dtim_period = se->se_dtimperiod;
vap->iv_dtim_count = 0;
/* NB: must be after ni_chan is setup */
ieee80211_setup_rates(ni, se->se_rates, se->se_xrates,
IEEE80211_F_DOSORT | IEEE80211_F_DONEGO | IEEE80211_F_DODEL);
return ieee80211_sta_join1(PASS_NODE(ni));
}
EXPORT_SYMBOL(ieee80211_sta_join);
/*
* Leave the specified IBSS/BSS network. The node is assumed to
* be passed in with a held reference.
*/
void
ieee80211_sta_leave(struct ieee80211_node *ni)
{
struct ieee80211vap *vap = ni->ni_vap;
/* WDS/Repeater: Stop software beacon timer for STA */
if ((vap->iv_opmode == IEEE80211_M_STA) &&
(vap->iv_flags_ext & IEEE80211_FEXT_SWBMISS)) {
del_timer(&vap->iv_swbmiss);
}
ieee80211_notify_node_leave(ni);
}
/*
* Node table support.
*/
static void
ieee80211_node_table_init(struct ieee80211com *ic,
struct ieee80211_node_table *nt, const char *name, int inact)
{
nt->nt_ic = ic;
IEEE80211_NODE_TABLE_LOCK_INIT(nt, ic->ic_dev->name);
IEEE80211_SCAN_LOCK_INIT(nt, ic->ic_dev->name);
TAILQ_INIT(&nt->nt_node);
nt->nt_name = name;
nt->nt_scangen = 1;
nt->nt_inact_init = inact;
init_timer(&nt->nt_wds_aging_timer);
nt->nt_wds_aging_timer.function = ieee80211_node_wds_ageout;
nt->nt_wds_aging_timer.data = (unsigned long) nt;
mod_timer(&nt->nt_wds_aging_timer, jiffies + HZ * WDS_AGING_TIMER_VAL);
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
static __inline
void
#ifdef IEEE80211_DEBUG_REFCNT
node_table_join_locked_debug(struct ieee80211_node_table *nt,
struct ieee80211_node *ni, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
node_table_join_locked(struct ieee80211_node_table *nt,
struct ieee80211_node *ni)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211_node *tni = NULL;
IEEE80211_NODE_TABLE_LOCK_ASSERT(nt);
ni->ni_table = nt;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
tni = ieee80211_ref_node_debug(ni, func, line);
#else
tni = ieee80211_ref_node(ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
TAILQ_INSERT_TAIL(&nt->nt_node, tni, ni_list);
tni = NULL;
LIST_INSERT_HEAD(&nt->nt_hash[IEEE80211_NODE_HASH(ni->ni_macaddr)],
ni, ni_hash);
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
static __inline
void
#ifdef IEEE80211_DEBUG_REFCNT
node_table_leave_locked_debug(struct ieee80211_node_table *nt,
struct ieee80211_node *ni, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
node_table_leave_locked(struct ieee80211_node_table *nt,
struct ieee80211_node *ni)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211_node *hni;
IEEE80211_NODE_TABLE_LOCK_ASSERT(nt);
TAILQ_REMOVE(&nt->nt_node, ni, ni_list);
LIST_FOREACH(hni, &nt->nt_hash[IEEE80211_NODE_HASH(ni->ni_macaddr)],
ni_hash) {
LIST_REMOVE(ni, ni_hash);
}
ni->ni_table = NULL;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_unref_node_debug(&ni, func, line);
#else
ieee80211_unref_node(&ni);
#endif
}
/* This is overridden by ath_node_alloc in ath/if_ath.c, and so
* should never get called.
*/
static struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
node_alloc_debug(struct ieee80211vap *vap, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
node_alloc(struct ieee80211vap *vap)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211_node *ni;
MALLOC(ni, struct ieee80211_node *, sizeof(struct ieee80211_node),
M_80211_NODE, M_NOWAIT | M_ZERO);
printk(KERN_ERR "%s: ERROR, this function should never be called!", __func__);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
dump_stack();
return ni;
}
/*
* Reclaim any resources in a node and reset any critical
* state. Typically nodes are free'd immediately after,
* but in some cases the storage may be reused so we need
* to ensure consistent state (should probably fix that).
*
* Context: hwIRQ, softIRQ and process context
*/
static void
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
node_cleanup_debug(struct ieee80211_node *ni, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
node_cleanup(struct ieee80211_node *ni)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211vap *vap = ni->ni_vap;
/* NB: preserve ni_table */
if (ni->ni_flags & IEEE80211_NODE_PWR_MGT) {
if (vap->iv_opmode != IEEE80211_M_STA)
vap->iv_ps_sta--;
ni->ni_flags &= ~IEEE80211_NODE_PWR_MGT;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
IEEE80211_NOTE(vap, IEEE80211_MSG_POWER, ni,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
"%s (%s%d): power save mode off, %u STAs in PS mode",
__func__, func, line,
vap->iv_ps_sta);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
if (ni->ni_flags & IEEE80211_NODE_UAPSD_TRIG) {
ni->ni_flags &= ~IEEE80211_NODE_UAPSD_TRIG;
IEEE80211_LOCK_IRQ(ni->ni_ic);
ni->ni_ic->ic_uapsdmaxtriggers--;
IEEE80211_UNLOCK_IRQ(ni->ni_ic);
}
}
/*
* Drain power save queue and, if needed, clear TIM.
*/
if (ieee80211_node_saveq_drain(ni) != 0 && vap->iv_set_tim != NULL)
vap->iv_set_tim(ni, 0);
ni->ni_associd = 0;
/*
* Preserve SSID, WPA, and WME IEs so the bss node is
* reusable during a re-auth/re-assoc state transition.
* If we remove these data they will not be recreated
* because they come from a probe-response or beacon frame
* which cannot be expected prior to the association-response.
* This should not be an issue when operating in other modes
* as stations leaving always go through a full state transition
* which will rebuild this state.
*
* XXX does this leave us open to inheriting old state?
*/
if (ni->ni_rxfrag != NULL) {
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_dev_kfree_skb_debug(&ni->ni_rxfrag, func, line);
#else
ieee80211_dev_kfree_skb(&ni->ni_rxfrag);
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
}
ieee80211_crypto_delkey(vap, &ni->ni_ucastkey, ni);
ni->ni_rxkeyoff = 0;
}
#ifdef IEEE80211_DEBUG
static void
node_print_message(
u_int32_t flags,
int show_counter,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
int refcnt_adjust,
const struct ieee80211_node *ni,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
const char *func1, int line1,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
const char *func2, int line2,
const char *message,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
...)
{
va_list args;
int adjusted_refcount = atomic_read(&ni->ni_refcnt) + refcnt_adjust;
char node_count[10] = { '\0' };
char expanded_message[192] = { '\0' };
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
if (0 == (ni->ni_ic->ic_debug & flags))
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
return;
if (adjusted_refcount == 0)
show_counter = 1;
if (show_counter) {
snprintf(node_count,
sizeof(node_count),
"[#NODES=%05d] ",
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
atomic_read(&ni->ni_ic->ic_node_counter));
}
va_start(args, message);
vsnprintf(expanded_message, sizeof(expanded_message), message, args);
#ifdef IEEE80211_DEBUG_REFCNT
printk(KERN_DEBUG "%s/%s: %s%s:%d -> %s:%d %s [node %p<" MAC_FMT ">%s%s%s%s, refs=%02d]\n",
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
printk(KERN_DEBUG "%s/%s: %s%s:%d %s [node %p<" MAC_FMT ">%s%s%s%s, refs=%02d]\n",
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
ni->ni_ic->ic_dev->name,
ni->ni_vap->iv_dev->name,
node_count,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
func1, line1,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
func2, line2,
expanded_message,
ni, MAC_ADDR(ni->ni_macaddr),
ni->ni_table != NULL ? " in " : "",
ni->ni_table != NULL ? ni->ni_table->nt_name : "",
ni->ni_table != NULL ? " table" : "",
ni->ni_table != NULL ? "" : " (not in any tables)",
adjusted_refcount);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
va_end(args);
}
EXPORT_SYMBOL(node_print_message);
#endif
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
static void
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
node_free_debug(struct ieee80211_node *ni, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
node_free(struct ieee80211_node *ni)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
KASSERT(atomic_read(&ni->ni_refcnt) == 0, ("node being free whilst still referenced"));
if (ni->ni_challenge != NULL)
FREE(ni->ni_challenge, M_DEVBUF);
if (ni->ni_wpa_ie != NULL)
FREE(ni->ni_wpa_ie, M_DEVBUF);
if (ni->ni_rsn_ie != NULL)
FREE(ni->ni_rsn_ie, M_DEVBUF);
if (ni->ni_wme_ie != NULL)
FREE(ni->ni_wme_ie, M_DEVBUF);
if (ni->ni_ath_ie != NULL)
FREE(ni->ni_ath_ie, M_DEVBUF);
if (ni->ni_suppchans != NULL)
FREE(ni->ni_suppchans, M_DEVBUF);
if (ni->ni_suppchans_new != NULL)
FREE(ni->ni_suppchans_new, M_DEVBUF);
if (ni->ni_needed_chans != NULL)
FREE(ni->ni_needed_chans, M_DEVBUF);
IEEE80211_NODE_SAVEQ_DESTROY(ni);
FREE(ni, M_80211_NODE);
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
static int32_t
node_count(struct ieee80211com *ic)
{
return atomic_read(&ic->ic_node_counter);
}
static u_int8_t
node_getrssi(const struct ieee80211_node *ni)
{
return ni->ni_rssi;
}
/*
* Create an entry in the specified node table. The node
* is setup with the mac address, an initial reference count,
* and some basic parameters obtained from global state.
* This interface is not intended for general use, it is
* used by the routines below to create entries with a
* specific purpose.
* Dont assume a BSS?
* Allocates a new struct ieee80211_node that has a reference
* count of one, and adds it to the node table.
*/
struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_alloc_node_table_debug(struct ieee80211vap *vap,
const u_int8_t *macaddr,
const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_alloc_node_table(struct ieee80211vap *vap,
const u_int8_t *macaddr)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211com *ic = vap->iv_ic;
struct ieee80211_node_table *nt = &ic->ic_sta;
struct ieee80211_node *ni;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_alloc_node_debug(vap, macaddr, func, line);
#else
ni = ieee80211_alloc_node(vap, macaddr);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
if (ni != NULL) {
ni->ni_inact = ni->ni_inact_reload = nt->nt_inact_init;
WME_UAPSD_NODE_TRIGSEQINIT(ni);
IEEE80211_NODE_SAVEQ_INIT(ni, "unknown");
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
node_table_join_locked_debug(nt, ni, func, line);
#else
node_table_join_locked(nt, ni);
#endif
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
}
else {
printk(KERN_ERR "Failed to allocate node for " MAC_FMT ".\n", MAC_ADDR(macaddr));
}
return ni;
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_alloc_node_table_debug);
#else
EXPORT_SYMBOL(ieee80211_alloc_node_table);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
/* Allocate a node structure and initialise specialised structures
* This function does not add the node to the node table, thus this
* node will not be found using ieee80211_find_*node.
* This is useful when sending one off errors or request denials.
*/
static struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_alloc_node_debug(struct ieee80211vap *vap, const u_int8_t *macaddr, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_alloc_node(struct ieee80211vap *vap, const u_int8_t *macaddr)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211com *ic = vap->iv_ic;
struct ieee80211_node *ni;
/* This always allocates zeroed memoery */
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ic->ic_node_alloc_debug(vap, func, line);
#else
ni = ic->ic_node_alloc(vap);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
if (ni != NULL) {
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
atomic_set(&ni->ni_refcnt, 1);
IEEE80211_ADDR_COPY(ni->ni_macaddr, macaddr);
ni->ni_chan = IEEE80211_CHAN_ANYC;
ni->ni_authmode = IEEE80211_AUTH_OPEN;
ni->ni_txpower = ic->ic_txpowlimit;
ieee80211_crypto_resetkey(vap, &ni->ni_ucastkey,
IEEE80211_KEYIX_NONE);
ni->ni_ath_defkeyindex = IEEE80211_INVAL_DEFKEY;
ni->ni_vap = vap;
ni->ni_ic = ic;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
atomic_inc(&ni->ni_ic->ic_node_counter);
#ifdef IEEE80211_DEBUG
node_print_message(IEEE80211_MSG_NODE|IEEE80211_MSG_NODE_REF,
1 /* show counter */,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
0 /* adjust refcount */,
ni,
#ifdef IEEE80211_DEBUG_REFCNT
func, line,
#endif
__func__, __LINE__,
"alloc" /* message */);
#endif
} else {
/* XXX msg */
vap->iv_stats.is_rx_nodealloc++;
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
return ni;
}
/* Add wds address to the node table */
int
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_add_wds_addr_debug(struct ieee80211_node_table *nt,
struct ieee80211_node *ni, const u_int8_t *macaddr, u_int8_t wds_static,
const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_add_wds_addr(struct ieee80211_node_table *nt,
struct ieee80211_node *ni, const u_int8_t *macaddr, u_int8_t wds_static)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
int hash;
struct ieee80211_wds_addr *wds;
MALLOC(wds, struct ieee80211_wds_addr *, sizeof(struct ieee80211_wds_addr),
M_80211_WDS, M_NOWAIT | M_ZERO);
if (wds == NULL) {
/* XXX msg */
return 1;
}
if (wds_static)
wds->wds_agingcount = WDS_AGING_STATIC;
else
wds->wds_agingcount = WDS_AGING_COUNT;
hash = IEEE80211_NODE_HASH(macaddr);
IEEE80211_ADDR_COPY(wds->wds_macaddr, macaddr);
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
wds->wds_ni = ieee80211_ref_node_debug(ni, func, line);
#else
wds->wds_ni = ieee80211_ref_node(ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
LIST_INSERT_HEAD(&nt->nt_wds_hash[hash], wds, wds_hash);
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
return 0;
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_add_wds_addr_debug);
#else
EXPORT_SYMBOL(ieee80211_add_wds_addr);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
/* remove wds address from the wds hash table */
void
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_remove_wds_addr_debug(struct ieee80211_node_table *nt, const u_int8_t *macaddr,
const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_remove_wds_addr(struct ieee80211_node_table *nt, const u_int8_t *macaddr)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
int hash;
struct ieee80211_wds_addr *wds, *twds;
hash = IEEE80211_NODE_HASH(macaddr);
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
LIST_FOREACH_SAFE(wds, &nt->nt_wds_hash[hash], wds_hash, twds) {
if (IEEE80211_ADDR_EQ(wds->wds_macaddr, macaddr)) {
LIST_REMOVE(wds, wds_hash);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_unref_node_debug(&wds->wds_ni, func, line);
#else
ieee80211_unref_node(&wds->wds_ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
FREE(wds, M_80211_WDS);
break;
}
}
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_remove_wds_addr_debug);
#else
EXPORT_SYMBOL(ieee80211_remove_wds_addr);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
/* Remove node references from wds table */
void
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_del_wds_node_debug(struct ieee80211_node_table *nt, struct ieee80211_node *ni,
const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_del_wds_node(struct ieee80211_node_table *nt, struct ieee80211_node *ni)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
int hash;
struct ieee80211_wds_addr *wds, *twds;
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
for (hash = 0; hash < IEEE80211_NODE_HASHSIZE; hash++) {
LIST_FOREACH_SAFE(wds, &nt->nt_wds_hash[hash], wds_hash, twds) {
if (wds->wds_ni == ni) {
LIST_REMOVE(wds, wds_hash);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_unref_node_debug(&wds->wds_ni, func, line);
#else
ieee80211_unref_node(&wds->wds_ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
FREE(wds, M_80211_WDS);
}
}
}
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_del_wds_node_debug);
#else
EXPORT_SYMBOL(ieee80211_del_wds_node);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
static void
ieee80211_node_wds_ageout(unsigned long data)
{
struct ieee80211_node_table *nt = (struct ieee80211_node_table *)data;
int hash;
struct ieee80211_wds_addr *wds, *twds;
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
for (hash = 0; hash < IEEE80211_NODE_HASHSIZE; hash++) {
LIST_FOREACH_SAFE(wds, &nt->nt_wds_hash[hash], wds_hash, twds) {
if (wds->wds_agingcount != WDS_AGING_STATIC) {
if (!wds->wds_agingcount) {
LIST_REMOVE(wds, wds_hash);
ieee80211_unref_node(&wds->wds_ni);
FREE(wds, M_80211_WDS);
} else
wds->wds_agingcount--;
}
}
}
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
mod_timer(&nt->nt_wds_aging_timer, jiffies + HZ * WDS_AGING_TIMER_VAL);
}
/* Add the specified station to the station table.
* Allocates a new struct ieee80211_node that has a reference count of one
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
* If tmp is 0, it is added to the node table and the reference is used.
* If tmp is 1, then the caller gets to use the reference. */
struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_dup_bss_debug(struct ieee80211vap *vap, const u_int8_t *macaddr,
unsigned char tmp, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_dup_bss(struct ieee80211vap *vap, const u_int8_t *macaddr,
unsigned char tmp)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211_node *ni;
/* FIXME: Hack */
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
if (tmp) {
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_alloc_node_debug(vap, macaddr, func, line);
#else
ni = ieee80211_alloc_node(vap, macaddr);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
else {
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_alloc_node_table_debug(vap, macaddr, func, line);
#else
ni = ieee80211_alloc_node_table(vap, macaddr);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: ni:%p allocated for " MAC_FMT "\n",
__func__, ni, MAC_ADDR(macaddr));
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
}
if (ni != NULL) {
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
copy_bss_state(ni, vap->iv_bss);
IEEE80211_ADDR_COPY(ni->ni_bssid, vap->iv_bssid);
/* Do this only for nodes that already have a BSS. Otherwise
* ic_bsschan is not set and we get a KASSERT failure.
* Required by ieee80211_fix_rate */
ieee80211_node_set_chan(vap->iv_ic, ni);
}
return ni;
}
static struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_find_wds_node_locked_debug(struct ieee80211_node_table *nt,
const u_int8_t *macaddr, const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_find_wds_node_locked(struct ieee80211_node_table *nt,
const u_int8_t *macaddr)
#endif
{
struct ieee80211_wds_addr *wds;
int hash;
IEEE80211_NODE_TABLE_LOCK_ASSERT(nt);
hash = IEEE80211_NODE_HASH(macaddr);
LIST_FOREACH(wds, &nt->nt_wds_hash[hash], wds_hash) {
if (IEEE80211_ADDR_EQ(wds->wds_macaddr, macaddr)) {
if (wds->wds_agingcount != WDS_AGING_STATIC)
wds->wds_agingcount = WDS_AGING_COUNT; /* reset the aging count */
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
return ieee80211_ref_node_debug(wds->wds_ni, func, line);
#else
return ieee80211_ref_node(wds->wds_ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
}
return NULL;
}
/* NB: A node reference is acquired here; the caller MUST release it. */
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
#define ieee80211_find_node_locked(nt, mac) \
ieee80211_find_node_locked_debug(nt, mac, __func__, __LINE__)
#endif
static struct ieee80211_node *
#ifdef IEEE80211_DEBUG_REFCNT
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
ieee80211_find_node_locked_debug(struct ieee80211_node_table *nt,
const u_int8_t *macaddr, const char *func, int line)
#else
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
ieee80211_find_node_locked(struct ieee80211_node_table *nt,
const u_int8_t *macaddr)
#endif
{
struct ieee80211_node *ni;
int hash;
struct ieee80211_wds_addr *wds;
IEEE80211_NODE_TABLE_LOCK_ASSERT(nt);
hash = IEEE80211_NODE_HASH(macaddr);
LIST_FOREACH(ni, &nt->nt_hash[hash], ni_hash) {
if (IEEE80211_ADDR_EQ(ni->ni_macaddr, macaddr)) {
#ifdef IEEE80211_DEBUG_REFCNT
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
ieee80211_ref_node_debug(ni, func, line);
#else
ieee80211_ref_node(ni);
#endif
return ni;
}
}
/* Now, we look for the desired mac address in the 4 address
nodes. */
LIST_FOREACH(wds, &nt->nt_wds_hash[hash], wds_hash) {
if (IEEE80211_ADDR_EQ(wds->wds_macaddr, macaddr)) {
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
return ieee80211_ref_node_debug(wds->wds_ni, func, line);
#else
return ieee80211_ref_node(wds->wds_ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
}
return NULL;
}
struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_find_wds_node_debug(struct ieee80211_node_table *nt, const u_int8_t *macaddr,
const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_find_wds_node(struct ieee80211_node_table *nt, const u_int8_t *macaddr)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211_node *ni;
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_find_wds_node_locked_debug(nt, macaddr, func, line);
#else
ni = ieee80211_find_wds_node_locked(nt, macaddr);
#endif
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
return ni;
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_find_wds_node_debug);
#else
EXPORT_SYMBOL(ieee80211_find_wds_node);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
struct ieee80211_node *
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_find_node_debug(struct ieee80211_node_table *nt,
const u_int8_t *macaddr, const char *func, int line)
#else
ieee80211_find_node(struct ieee80211_node_table *nt, const u_int8_t *macaddr)
#endif
{
struct ieee80211_node *ni;
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_find_node_locked_debug(nt, macaddr, func, line);
#else
ni = ieee80211_find_node_locked(nt, macaddr);
#endif
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
return ni;
}
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_find_node_debug);
#else
EXPORT_SYMBOL(ieee80211_find_node);
#endif
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
/*
* Fake up a node; this handles node discovery in adhoc mode.
* Note that for the driver's benefit we we treat this like
* an association so the driver has an opportunity to setup
* its private state.
*
* Caller must ieee80211_ref_node()
*/
struct ieee80211_node *
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_fakeup_adhoc_node_debug(struct ieee80211vap *vap,
const u_int8_t macaddr[IEEE80211_ADDR_LEN],
const char *func, int line)
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ieee80211_fakeup_adhoc_node(struct ieee80211vap *vap,
const u_int8_t macaddr[IEEE80211_ADDR_LEN])
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
{
struct ieee80211_node *ni;
ni = ieee80211_dup_bss(vap, macaddr, 0);
if (ni != NULL) {
/* XXX no rate negotiation; just dup */
ni->ni_rates = vap->iv_bss->ni_rates;
if (vap->iv_ic->ic_newassoc != NULL)
vap->iv_ic->ic_newassoc(ni, 1);
/* XXX not right for 802.1x/WPA */
ieee80211_node_authorize(ni);
}
return ni;
}
/*
* Do node discovery in adhoc mode on receipt of a beacon
* or probe response frame. Note that for the driver's
* benefit we treat this like an association so the
* driver has an opportunity to setup its private state.
*/
struct ieee80211_node *
ieee80211_add_neighbor(struct ieee80211vap *vap, const struct ieee80211_frame *wh,
const struct ieee80211_scanparams *sp)
{
struct ieee80211com *ic = vap->iv_ic;
struct ieee80211_node *ni;
ni = ieee80211_dup_bss(vap, wh->i_addr2, 0);
if (ni != NULL) {
ni->ni_esslen = sp->ssid[1];
memcpy(ni->ni_essid, sp->ssid + 2, sp->ssid[1]);
IEEE80211_ADDR_COPY(ni->ni_bssid, wh->i_addr3);
memcpy(ni->ni_tstamp.data, sp->tstamp, sizeof(ni->ni_tstamp));
ni->ni_intval = IEEE80211_BINTVAL_SANITISE(sp->bintval);
ni->ni_capinfo = sp->capinfo;
ni->ni_chan = ic->ic_curchan;
ni->ni_fhdwell = sp->fhdwell;
ni->ni_fhindex = sp->fhindex;
ni->ni_erp = sp->erp;
ni->ni_timoff = sp->timoff;
if (sp->wme != NULL)
ieee80211_saveie(&ni->ni_wme_ie, sp->wme);
if (sp->wpa != NULL)
ieee80211_saveie(&ni->ni_wpa_ie, sp->wpa);
if (sp->rsn != NULL)
ieee80211_saveie(&ni->ni_rsn_ie, sp->rsn);
if (sp->ath != NULL)
ieee80211_saveath(ni, sp->ath);
/* NB: must be after ni_chan is setup */
ieee80211_setup_rates(ni, sp->rates, sp->xrates, IEEE80211_F_DOSORT);
if (ic->ic_newassoc != NULL)
ic->ic_newassoc(ni, 1);
/* XXX not right for 802.1x/WPA */
ieee80211_node_authorize(ni);
if (vap->iv_opmode == IEEE80211_M_AHDEMO) {
/*
* Blindly propagate capabilities based on the
* local configuration. In particular this permits
* us to use QoS to disable ACKs and to use short
* preamble on 2.4G channels.
*/
if (vap->iv_flags & IEEE80211_F_WME)
ni->ni_flags |= IEEE80211_NODE_QOS;
if (vap->iv_flags & IEEE80211_F_SHPREAMBLE)
ni->ni_capinfo |= IEEE80211_CAPINFO_SHORT_PREAMBLE;
}
}
return ni;
}
/*
* Return the node for the sender of a frame; if the sender is unknown return
* NULL. The caller is expected to deal with this. (The frame is sent to all
* VAPs in this case).
*
* NB: A node reference is acquired here; the caller MUST release it.
*/
struct ieee80211_node *
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_find_rxnode_debug(struct ieee80211com *ic,
const struct ieee80211_frame_min *wh, const char *func, int line)
#else
ieee80211_find_rxnode(struct ieee80211com *ic,
const struct ieee80211_frame_min *wh)
#endif
{
#define IS_CTL(wh) \
((wh->i_fc[0] & IEEE80211_FC0_TYPE_MASK) == IEEE80211_FC0_TYPE_CTL)
#define IS_PSPOLL(wh) \
((wh->i_fc[0] & IEEE80211_FC0_SUBTYPE_MASK) == IEEE80211_FC0_SUBTYPE_PS_POLL)
#define IS_RTS(wh) \
((wh->i_fc[0] & IEEE80211_FC0_SUBTYPE_MASK) == IEEE80211_FC0_SUBTYPE_RTS)
struct ieee80211_node_table *nt;
struct ieee80211_node *ni;
/* XXX check ic_bss first in station mode */
/* XXX 4-address frames? */
nt = &ic->ic_sta;
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
/* NB: Control frames typically have one address, except
* for RTS and PSPOLL */
if (IS_CTL(wh) && !IS_PSPOLL(wh) && !IS_RTS(wh))
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_find_node_locked_debug(nt, wh->i_addr1, func, line);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ni = ieee80211_find_node_locked(nt, wh->i_addr1);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
else
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_find_node_locked_debug(nt, wh->i_addr2, func, line);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#else
ni = ieee80211_find_node_locked(nt, wh->i_addr2);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
return ni;
#undef IS_PSPOLL
#undef IS_CTL
#undef IS_RTS
}
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_find_rxnode_debug);
#else
EXPORT_SYMBOL(ieee80211_find_rxnode);
#endif
/*
* Return the appropriate node for sending a data frame. This handles node
* discovery in adhoc networks.
*
* NB: A node reference is acquired here; the caller MUST release it.
*/
struct ieee80211_node *
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_find_txnode_debug(struct ieee80211vap *vap, const u_int8_t *mac,
const char *func, int line)
#else
ieee80211_find_txnode(struct ieee80211vap *vap, const u_int8_t *mac)
#endif
{
struct ieee80211_node_table *nt;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
struct ieee80211_node *ni = NULL;
/*
* The destination address should be in the node table
* unless we are operating in station mode or this is a
* multicast/broadcast frame.
*/
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
if (vap->iv_opmode == IEEE80211_M_STA || IEEE80211_IS_MULTICAST(mac)) {
#ifdef IEEE80211_DEBUG_REFCNT
return ieee80211_ref_node_debug(vap->iv_bss, func, line);
#else
return ieee80211_ref_node(vap->iv_bss);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
/* XXX: Can't hold lock across dup_bss due to recursive locking. */
nt = &vap->iv_ic->ic_sta;
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_find_node_locked_debug(nt, mac, func, line);
#else
ni = ieee80211_find_node_locked(nt, mac);
#endif
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
if (ni == NULL) {
if (vap->iv_opmode == IEEE80211_M_IBSS ||
vap->iv_opmode == IEEE80211_M_AHDEMO) {
/*
* In adhoc mode cons up a node for the destination.
* Note that we need an additional reference for the
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
* caller to be consistent with ieee80211_find_node.
*/
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ni = ieee80211_fakeup_adhoc_node_debug(vap, mac,
func, line);
#else
ni = ieee80211_fakeup_adhoc_node(vap, mac);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
if (ni != NULL) {
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_ref_node_debug(ni, func, line);
#else
ieee80211_ref_node(ni);
#endif
}
} else {
IEEE80211_NOTE_MAC(vap, IEEE80211_MSG_OUTPUT, mac,
"no node, discard frame (%s)", __func__);
vap->iv_stats.is_tx_nonode++;
}
}
return ni;
}
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_find_txnode_debug);
#else
EXPORT_SYMBOL(ieee80211_find_txnode);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
/* Context: hwIRQ, softIRQ and process context
*/
void
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_free_node_debug(struct ieee80211_node *ni, const char *func, int line)
#else /* #ifdef IEEE80211_DEBUG_REFCNT */
ieee80211_free_node(struct ieee80211_node *ni)
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
{
struct ieee80211vap *vap = ni->ni_vap;
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
atomic_dec(&ni->ni_ic->ic_node_counter);
#ifdef IEEE80211_DEBUG
node_print_message(IEEE80211_MSG_NODE|IEEE80211_MSG_NODE_REF,
1 /* show counter */,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
0 /* adjust refcount */,
ni,
#ifdef IEEE80211_DEBUG_REFCNT
func, line,
#endif
__func__, __LINE__,
"free" /* message */);
#endif
if (vap->iv_aid_bitmap != NULL)
IEEE80211_AID_CLR(vap, ni->ni_associd);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
vap->iv_ic->ic_node_free_debug(ni, func, line);
#else
vap->iv_ic->ic_node_free(ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_free_node_debug);
#else
EXPORT_SYMBOL(ieee80211_free_node);
#endif
static void _reset_node(void *arg, struct ieee80211_node *ni)
{
if (ni->ni_associd != 0) {
struct ieee80211vap *vap = ni->ni_vap;
if (vap->iv_auth->ia_node_leave != NULL)
vap->iv_auth->ia_node_leave(ni);
if (vap->iv_aid_bitmap != NULL)
IEEE80211_AID_CLR(vap, ni->ni_associd);
}
ieee80211_node_leave(ni);
}
static void
ieee80211_node_table_reset(struct ieee80211_node_table *nt,
struct ieee80211vap *vap)
{
ieee80211_iterate_dev_nodes(vap->iv_dev, nt, _reset_node, NULL);
}
static void
ieee80211_node_table_cleanup(struct ieee80211_node_table *nt)
{
struct ieee80211com *ic = nt->nt_ic;
struct ieee80211_node *ni, *next;
TAILQ_FOREACH_SAFE(ni, &nt->nt_node, ni_list, next) {
if (ni->ni_associd != 0) {
struct ieee80211vap *vap = ni->ni_vap;
if (vap->iv_auth->ia_node_leave != NULL)
vap->iv_auth->ia_node_leave(ni);
if (vap->iv_aid_bitmap != NULL)
IEEE80211_AID_CLR(vap, ni->ni_associd);
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ic->ic_node_cleanup_debug(ni, __func__, __LINE__);
#else
ic->ic_node_cleanup(ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
del_timer(&nt->nt_wds_aging_timer);
IEEE80211_SCAN_LOCK_DESTROY(nt);
IEEE80211_NODE_TABLE_LOCK_DESTROY(nt);
}
/*
* Timeout inactive stations and do related housekeeping.
* Note that we cannot hold the node lock while sending a
* frame as this would lead to a LOR. Instead we use a
* generation number to mark nodes that we've scanned and
* drop the lock and restart a scan if we have to time out
* a node. Since we are single-threaded by virtue of
* controlling the inactivity timer we can be sure this will
* process each node only once.
*
* Context: softIRQ (tasklet)
*/
static void
ieee80211_timeout_stations(struct ieee80211_node_table *nt)
{
struct ieee80211com *ic = nt->nt_ic;
struct ieee80211_node *ni;
u_int gen;
int isadhoc;
isadhoc = (ic->ic_opmode == IEEE80211_M_IBSS ||
ic->ic_opmode == IEEE80211_M_AHDEMO);
IEEE80211_SCAN_LOCK_IRQ(nt);
gen = ++nt->nt_scangen;
restart:
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
TAILQ_FOREACH(ni, &nt->nt_node, ni_list) {
if (ni->ni_scangen == gen) /* previously handled */
continue;
ni->ni_scangen = gen;
/*
* Free fragment if not needed anymore
* (last fragment older than 1s).
* XXX doesn't belong here
*/
if (ni->ni_rxfrag != NULL &&
time_after(jiffies, ni->ni_rxfragstamp + HZ)) {
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
ieee80211_dev_kfree_skb(&ni->ni_rxfrag);
}
/*
* Special case ourself; we may be idle for extended periods
* of time and regardless reclaiming our state is wrong.
* Special case a WDS link: it may be dead or idle, but it is
* never ok to reclaim it, as this will block transmissions
* and nobody will recreate the node when the WDS peer is
* available again. */
if ((ni == ni->ni_vap->iv_bss) ||
(ni->ni_vap->iv_opmode == IEEE80211_M_WDS &&
!memcmp(ni->ni_macaddr, ni->ni_vap->wds_mac, ETH_ALEN)))
{
/* NB: don't permit it to go negative */
if (ni->ni_inact > 0)
ni->ni_inact--;
continue;
}
ni->ni_inact--;
if (ni->ni_associd != 0 || isadhoc) {
struct ieee80211vap *vap = ni->ni_vap;
/*
* Age frames on the power save queue.
*/
if (ieee80211_node_saveq_age(ni) != 0 &&
IEEE80211_NODE_SAVEQ_QLEN(ni) == 0 &&
vap->iv_set_tim != NULL)
vap->iv_set_tim(ni, 0);
/*
* Probe the station before time it out. We
* send a null data frame which may not be
* universally supported by drivers (need it
* for ps-poll support so it should be...).
*/
if (0 < ni->ni_inact &&
ni->ni_inact <= vap->iv_inact_probe) {
IEEE80211_NOTE(vap,
IEEE80211_MSG_INACT | IEEE80211_MSG_NODE,
ni, "%s",
"probe station due to inactivity");
/*
* Grab a reference before unlocking the table
* so the node cannot be reclaimed before we
* send the frame. ieee80211_send_nulldata
* understands we've done this and reclaims the
* ref for us as needed.
*/
ieee80211_ref_node(ni);
IEEE80211_NODE_TABLE_UNLOCK_IRQ_EARLY(nt);
ieee80211_send_nulldata(ni);
/* XXX stat? */
goto restart;
}
}
if (ni->ni_inact <= 0) {
IEEE80211_NOTE(ni->ni_vap,
IEEE80211_MSG_INACT | IEEE80211_MSG_NODE, ni,
"station timed out due to inactivity (refcnt %u)",
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
atomic_read(&ni->ni_refcnt));
/*
* Send a deauthenticate frame and drop the station.
* We grab a reference before unlocking the table so
* the node cannot be reclaimed before we complete our
* work.
*
* Separately we must drop the node lock before sending
* in case the driver takes a lock, as this may result
* in a LOR between the node lock and the driver lock.
*/
ni->ni_vap->iv_stats.is_node_timeout++;
ieee80211_ref_node(ni);
IEEE80211_NODE_TABLE_UNLOCK_IRQ_EARLY(nt);
if (ni->ni_associd != 0) {
IEEE80211_SEND_MGMT(ni,
IEEE80211_FC0_SUBTYPE_DEAUTH,
IEEE80211_REASON_AUTH_EXPIRE);
}
ieee80211_node_leave(ni);
ieee80211_unref_node(&ni);
goto restart;
}
}
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
IEEE80211_SCAN_UNLOCK_IRQ(nt);
}
/*
* Per-ieee80211com inactivity timer callback.
*/
static void
ieee80211_node_timeout(unsigned long arg)
{
struct ieee80211com *ic = (struct ieee80211com *)arg;
ieee80211_scan_timeout(ic);
ieee80211_timeout_stations(&ic->ic_sta);
mod_timer(&ic->ic_inact, jiffies + IEEE80211_INACT_WAIT * HZ);
}
void
ieee80211_iterate_nodes(struct ieee80211_node_table *nt, ieee80211_iter_func *f, void *arg)
{
ieee80211_iterate_dev_nodes(NULL, nt, f, arg);
}
EXPORT_SYMBOL(ieee80211_iterate_nodes);
void
ieee80211_iterate_dev_nodes(struct net_device *dev, struct ieee80211_node_table *nt, ieee80211_iter_func *f, void *arg)
{
struct ieee80211_node *ni;
u_int gen;
IEEE80211_SCAN_LOCK_IRQ(nt);
gen = ++nt->nt_scangen;
restart:
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
TAILQ_FOREACH(ni, &nt->nt_node, ni_list) {
if (dev != NULL && ni->ni_vap->iv_dev != dev)
continue; /* skip node not for this vap */
if (ni->ni_scangen != gen) {
ni->ni_scangen = gen;
(void) ieee80211_ref_node(ni);
IEEE80211_NODE_TABLE_UNLOCK_IRQ_EARLY(nt);
(*f)(arg, ni);
ieee80211_unref_node(&ni);
goto restart;
}
}
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
IEEE80211_SCAN_UNLOCK_IRQ(nt);
}
EXPORT_SYMBOL(ieee80211_iterate_dev_nodes);
void
ieee80211_dump_node(struct ieee80211_node_table *nt, struct ieee80211_node *ni)
{
int i;
printk("0x%p: mac " MAC_FMT " (refcnt %d)\n", ni,
MAC_ADDR(ni->ni_macaddr), atomic_read(&ni->ni_refcnt));
printk("\tscangen %u authmode %u flags 0x%x\n",
ni->ni_scangen, ni->ni_authmode, ni->ni_flags);
printk("\tassocid 0x%x txpower %u vlan %u\n",
ni->ni_associd, ni->ni_txpower, ni->ni_vlan);
printk ("rxfragstamp %lu\n", ni->ni_rxfragstamp);
for (i = 0; i < 17; i++) {
printk("\t%d: txseq %u rxseq %u fragno %u\n", i,
ni->ni_txseqs[i],
ni->ni_rxseqs[i] >> IEEE80211_SEQ_SEQ_SHIFT,
ni->ni_rxseqs[i] & IEEE80211_SEQ_FRAG_MASK);
}
printk("\trtsf %10llu rssi %u intval %u capinfo 0x%x\n",
ni->ni_rtsf, ni->ni_rssi, ni->ni_intval, ni->ni_capinfo);
printk("\tbssid " MAC_FMT " essid \"%.*s\" channel %u:0x%x\n",
MAC_ADDR(ni->ni_bssid),
ni->ni_esslen, ni->ni_essid,
ni->ni_chan != IEEE80211_CHAN_ANYC ?
ni->ni_chan->ic_freq : IEEE80211_CHAN_ANY,
ni->ni_chan != IEEE80211_CHAN_ANYC ? ni->ni_chan->ic_flags : 0);
printk("\tinact %u txrate %u\n",
ni->ni_inact, ni->ni_txrate);
}
void
ieee80211_dump_nodes(struct ieee80211_node_table *nt)
{
ieee80211_iterate_nodes(nt,
(ieee80211_iter_func *)ieee80211_dump_node, nt);
}
/*
* Handle a station joining an 11g network.
*/
static void
ieee80211_node_join_11g(struct ieee80211_node *ni)
{
struct ieee80211com *ic = ni->ni_ic;
#ifdef IEEE80211_DEBUG
struct ieee80211vap *vap = ni->ni_vap;
#endif
IEEE80211_LOCK_ASSERT(ic);
KASSERT(IEEE80211_IS_CHAN_ANYG(ic->ic_bsschan),
("not in 11g, bss %u:0x%x, curmode %u", ic->ic_bsschan->ic_freq,
ic->ic_bsschan->ic_flags, ic->ic_curmode));
/*
* Station isn't capable of short slot time. Bump
* the count of long slot time stations and disable
* use of short slot time. Note that the actual switch
* over to long slot time use may not occur until the
* next beacon transmission (per sec. 7.3.1.4 of 11g).
*/
if ((ni->ni_capinfo & IEEE80211_CAPINFO_SHORT_SLOTTIME) == 0) {
ic->ic_longslotsta++;
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC, ni,
"station needs long slot time, count %d",
ic->ic_longslotsta);
/* XXX VAPs w/ conflicting needs won't work */
if (!IEEE80211_IS_CHAN_108G(ic->ic_bsschan)) {
/*
* Don't force slot time when switched to turbo
* mode as non-ERP stations won't be present; this
* need only be done when on the normal G channel.
*/
ieee80211_set_shortslottime(ic, 0);
}
}
/*
* If the new station is not an ERP station
* then bump the counter and enable protection
* if configured.
*/
if (!ieee80211_iserp_rateset(ic, &ni->ni_rates)) {
ic->ic_nonerpsta++;
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC, ni,
"station is !ERP, %d non-ERP stations associated",
ic->ic_nonerpsta);
/*
* If protection is configured, enable it.
*/
if (ic->ic_protmode != IEEE80211_PROT_NONE) {
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: enable use of protection\n", __func__);
ic->ic_flags |= IEEE80211_F_USEPROT;
}
/*
* If station does not support short preamble
* then we must enable use of Barker preamble.
*/
if ((ni->ni_capinfo & IEEE80211_CAPINFO_SHORT_PREAMBLE) == 0) {
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC, ni,
"%s", "station needs long preamble");
ic->ic_flags |= IEEE80211_F_USEBARKER;
ic->ic_flags &= ~IEEE80211_F_SHPREAMBLE;
}
/* Update ERP element if this is first non ERP station */
if (ic->ic_nonerpsta == 1)
ic->ic_flags_ext |= IEEE80211_FEXT_ERPUPDATE;
} else
ni->ni_flags |= IEEE80211_NODE_ERP;
}
static void
count_suppchans(struct ieee80211com *ic, struct ieee80211_node *ni, int inc)
{
int i, tmp1, tmp2 = 0;
if (ni->ni_suppchans == NULL)
return;
CHANNEL_FOREACH(i, ic, tmp1, tmp2)
if (isset(ni->ni_suppchans, i))
ic->ic_chan_nodes[i] += inc;
ic->ic_cn_total += inc;
}
static void
remove_worse_nodes(void *arg, struct ieee80211_node *ni)
{
struct ieee80211_node *better = (struct ieee80211_node *)arg;
int i;
if (ni->ni_suppchans == NULL)
return;
if (ni == better)
return;
for (i = 0; i < better->ni_n_needed_chans; i++)
if (isclr(ni->ni_suppchans, better->ni_needed_chans[i])) {
/* this is the one of the nodes to be killed, do it now */
IEEE80211_NOTE_MAC(ni->ni_vap, IEEE80211_MSG_ASSOC|IEEE80211_MSG_DOTH, better->ni_macaddr,
"forcing [" MAC_FMT "] (aid %d) to leave", MAC_ADDR(ni->ni_macaddr),
IEEE80211_NODE_AID(ni));
IEEE80211_SEND_MGMT(ni,
IEEE80211_FC0_SUBTYPE_DISASSOC,
IEEE80211_REASON_SUPPCHAN_UNACCEPTABLE);
ni->ni_vap->iv_stats.is_node_fdisassoc++;
ieee80211_node_leave(ni);
return;
}
}
void
ieee80211_node_join(struct ieee80211_node *ni, int resp)
{
struct ieee80211com *ic = ni->ni_ic;
struct ieee80211vap *vap = ni->ni_vap;
int newassoc;
if (ni->ni_associd == 0) {
u_int16_t aid;
KASSERT(vap->iv_aid_bitmap != NULL, ("no aid bitmap"));
/*
* It would be good to search the bitmap
* more efficiently, but this will do for now.
*/
for (aid = 1; aid < vap->iv_max_aid; aid++)
if (!IEEE80211_AID_ISSET(vap, aid))
break;
if (aid >= vap->iv_max_aid) {
IEEE80211_SEND_MGMT(ni, resp,
IEEE80211_REASON_ASSOC_TOOMANY);
ieee80211_node_leave(ni);
return;
}
ni->ni_associd = aid | 0xc000;
IEEE80211_LOCK_IRQ(ic);
IEEE80211_AID_SET(vap, ni->ni_associd);
vap->iv_sta_assoc++;
ic->ic_sta_assoc++;
#ifdef ATH_SUPERG_XR
if (ni->ni_vap->iv_flags & IEEE80211_F_XR)
ic->ic_xr_sta_assoc++;
#endif
if (IEEE80211_ATH_CAP(vap, ni, IEEE80211_ATHC_TURBOP))
ic->ic_dt_sta_assoc++;
if (IEEE80211_IS_CHAN_ANYG(ic->ic_bsschan))
ieee80211_node_join_11g(ni);
KASSERT(ni->ni_suppchans == NULL, ("not a reassociation, but suppchans bitmap not NULL\n"));
/* Use node's new suppchans as the current */
ni->ni_suppchans = ni->ni_suppchans_new;
ni->ni_suppchans_new = NULL;
/* Add node's suppchans to ic->ic_chan_nodes */
count_suppchans(ic, ni, 1);
IEEE80211_UNLOCK_IRQ(ic);
newassoc = 1;
} else {
IEEE80211_LOCK_IRQ(ic);
/* Remove node's previous suppchans from ic->ic_chan_nodes */
count_suppchans(ic, ni, -1);
if (ni->ni_suppchans != NULL) {
FREE(ni->ni_suppchans, M_DEVBUF);
ni->ni_suppchans = NULL;
}
/* Use node's new suppchans as the current */
ni->ni_suppchans = ni->ni_suppchans_new;
ni->ni_suppchans_new = NULL;
/* Add node's new suppchans to ic->ic_chan_nodes */
count_suppchans(ic, ni, 1);
IEEE80211_UNLOCK_IRQ(ic);
newassoc = 0;
}
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC | IEEE80211_MSG_DEBUG, ni,
"station %sassociated at aid %d: %s preamble, %s slot time"
"%s%s%s%s%s%s%s",
newassoc ? "" : "re",
IEEE80211_NODE_AID(ni),
(ic->ic_flags & IEEE80211_F_SHPREAMBLE) &&
(ni->ni_capinfo & IEEE80211_CAPINFO_SHORT_PREAMBLE) ? "short" : "long",
ic->ic_flags & IEEE80211_F_SHSLOT ? "short" : "long",
ic->ic_flags & IEEE80211_F_USEPROT ? ", protection" : "",
ni->ni_flags & IEEE80211_NODE_QOS ? ", QoS" : "",
IEEE80211_ATH_CAP(vap, ni, IEEE80211_NODE_TURBOP) ?
", turbo" : "",
IEEE80211_ATH_CAP(vap, ni, IEEE80211_NODE_COMP) ?
", compression" : "",
IEEE80211_ATH_CAP(vap, ni, IEEE80211_NODE_FF) ?
", fast-frames" : "",
IEEE80211_ATH_CAP(vap, ni, IEEE80211_NODE_XR) ? ", XR" : "",
IEEE80211_ATH_CAP(vap, ni, IEEE80211_NODE_AR) ? ", AR" : ""
);
/* give driver a chance to setup state like ni_txrate */
if (ic->ic_newassoc != NULL)
ic->ic_newassoc(ni, newassoc);
ni->ni_inact_reload = vap->iv_inact_auth;
ni->ni_inact = ni->ni_inact_reload;
IEEE80211_SEND_MGMT(ni, resp, IEEE80211_STATUS_SUCCESS);
if (ni->ni_needed_chans != NULL) {
/* remove nodes which don't support one of ni->ni_needed_chans */
ieee80211_iterate_nodes(&ic->ic_sta, &remove_worse_nodes, (void *)ni);
FREE(ni->ni_needed_chans, M_DEVBUF);
ni->ni_needed_chans = NULL;
}
/* tell the authenticator about new station */
if (vap->iv_auth->ia_node_join != NULL)
vap->iv_auth->ia_node_join(ni);
ieee80211_notify_node_join(ni, newassoc);
}
/*
* Handle a station leaving an 11g network.
*/
static void
ieee80211_node_leave_11g(struct ieee80211_node *ni)
{
struct ieee80211com *ic = ni->ni_ic;
struct ieee80211vap *vap = ni->ni_vap;
IEEE80211_LOCK_ASSERT(ic);
KASSERT(IEEE80211_IS_CHAN_ANYG(ic->ic_bsschan),
("not in 11g, bss %u:0x%x, curmode %u",
ic->ic_bsschan->ic_freq, ic->ic_bsschan->ic_flags,
ic->ic_curmode));
/*
* If a long slot station do the slot time bookkeeping.
*/
if ((ni->ni_capinfo & IEEE80211_CAPINFO_SHORT_SLOTTIME) == 0) {
/* this can be 0 on mode changes from B -> G */
if (ic->ic_longslotsta > 0)
ic->ic_longslotsta--;
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC, ni,
"long slot time station leaves, count now %d",
ic->ic_longslotsta);
if (ic->ic_longslotsta == 0) {
/*
* Re-enable use of short slot time if supported
* and not operating in IBSS mode (per spec).
*/
if ((ic->ic_caps & IEEE80211_C_SHSLOT) &&
vap->iv_opmode != IEEE80211_M_IBSS) {
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: re-enable use of short slot time\n",
__func__);
ieee80211_set_shortslottime(ic, 1);
}
}
}
/*
* If a non-ERP station do the protection-related bookkeeping.
*/
if ((ni->ni_flags & IEEE80211_NODE_ERP) == 0) {
/* this can be 0 on mode changes from B -> G */
if (ic->ic_nonerpsta > 0)
ic->ic_nonerpsta--;
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC, ni,
"non-ERP station leaves, count now %d", ic->ic_nonerpsta);
if (ic->ic_nonerpsta == 0) {
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: disable use of protection\n", __func__);
ic->ic_flags &= ~IEEE80211_F_USEPROT;
/* XXX verify mode? */
if (ic->ic_caps & IEEE80211_C_SHPREAMBLE) {
IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
"%s: re-enable use of short preamble\n",
__func__);
ic->ic_flags |= IEEE80211_F_SHPREAMBLE;
ic->ic_flags &= ~IEEE80211_F_USEBARKER;
}
ic->ic_flags_ext |= IEEE80211_FEXT_ERPUPDATE;
}
}
}
/*
* Handle bookkeeping for a station/neighbor leaving
* the bss when operating in ap or adhoc modes.
*/
void
ieee80211_node_leave(struct ieee80211_node *ni)
{
struct ieee80211com *ic = ni->ni_ic;
struct ieee80211vap *vap = ni->ni_vap;
struct ieee80211_node_table *nt = ni->ni_table;
if (IEEE80211_NODE_AID(ni) != 0)
IEEE80211_NOTE(vap, IEEE80211_MSG_ASSOC | IEEE80211_MSG_DEBUG,
ni, "station with aid %d leaves (refcnt %u)",
IEEE80211_NODE_AID(ni), atomic_read(&ni->ni_refcnt));
/* From this point onwards we can no longer find the node,
* so no more references are generated
*/
ieee80211_remove_wds_addr(nt, ni->ni_macaddr);
ieee80211_del_wds_node(nt, ni);
IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
node_table_leave_locked(nt, ni);
IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
/*
* If node wasn't previously associated all
* we need to do is reclaim the reference.
* This also goes for nodes that are auth'ed but
* not associated.
*/
/* XXX ibss mode bypasses 11g and notification */
if (ni->ni_associd == 0)
goto done;
/*
* Tell the authenticator the station is leaving.
* Note that we must do this before yanking the
* association id as the authenticator uses the
* associd to locate its state block.
*/
if (vap->iv_auth->ia_node_leave != NULL)
vap->iv_auth->ia_node_leave(ni);
ieee80211_notify_sta_stats(ni);
IEEE80211_LOCK_IRQ(ic);
if (vap->iv_aid_bitmap != NULL)
IEEE80211_AID_CLR(vap, ni->ni_associd);
ni->ni_associd = 0;
vap->iv_sta_assoc--;
ic->ic_sta_assoc--;
#ifdef ATH_SUPERG_XR
if (ni->ni_vap->iv_flags & IEEE80211_F_XR)
ic->ic_xr_sta_assoc--;
#endif
if (IEEE80211_ATH_CAP(vap, ni, IEEE80211_ATHC_TURBOP))
ic->ic_dt_sta_assoc--;
if (IEEE80211_IS_CHAN_ANYG(ic->ic_bsschan))
ieee80211_node_leave_11g(ni);
/* Remove node's suppchans from ic->ic_chan_nodes */
if (ni->ni_suppchans != NULL)
count_suppchans(ic, ni, -1);
IEEE80211_UNLOCK_IRQ(ic);
/*
* Cleanup station state. In particular clear various
* state that might otherwise be reused if the node
* is reused before the reference count goes to zero
* (and memory is reclaimed).
*/
ieee80211_sta_leave(ni);
done:
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
/* Run a cleanup */
#ifdef IEEE80211_DEBUG_REFCNT
ic->ic_node_cleanup_debug(ni, __func__, __LINE__);
#else
ic->ic_node_cleanup(ni);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#endif
}
EXPORT_SYMBOL(ieee80211_node_leave);
u_int8_t
ieee80211_getrssi(struct ieee80211com *ic)
{
#define NZ(x) ((x) == 0 ? 1 : (x))
struct ieee80211_node_table *nt = &ic->ic_sta;
struct ieee80211vap *vap;
u_int32_t rssi_samples, rssi_total;
struct ieee80211_node *ni;
rssi_total = 0;
rssi_samples = 0;
switch (ic->ic_opmode) {
case IEEE80211_M_IBSS: /* average of all ibss neighbors */
/* XXX locking */
TAILQ_FOREACH(ni, &nt->nt_node, ni_list)
if (ni->ni_capinfo & IEEE80211_CAPINFO_IBSS) {
rssi_samples++;
rssi_total += ic->ic_node_getrssi(ni);
}
break;
case IEEE80211_M_AHDEMO: /* average of all neighbors */
/* XXX locking */
TAILQ_FOREACH(ni, &nt->nt_node, ni_list) {
if (memcmp(ni->ni_vap->iv_myaddr, ni->ni_macaddr,
IEEE80211_ADDR_LEN)!=0) {
rssi_samples++;
rssi_total += ic->ic_node_getrssi(ni);
}
}
break;
case IEEE80211_M_HOSTAP: /* average of all associated stations */
/* XXX locking */
TAILQ_FOREACH(ni, &nt->nt_node, ni_list)
if (IEEE80211_AID(ni->ni_associd) != 0) {
rssi_samples++;
rssi_total += ic->ic_node_getrssi(ni);
}
break;
case IEEE80211_M_MONITOR: /* XXX */
case IEEE80211_M_STA: /* use stats from associated ap */
default:
TAILQ_FOREACH(vap, &ic->ic_vaps, iv_next)
if (vap->iv_bss != NULL) {
rssi_samples++;
rssi_total += ic->ic_node_getrssi(vap->iv_bss);
}
break;
}
return rssi_total / NZ(rssi_samples);
#undef NZ
}
EXPORT_SYMBOL(ieee80211_getrssi);
void
ieee80211_node_reset(struct ieee80211_node *ni, struct ieee80211vap *vap)
{
/* XXX: Untested use of iv_bssid. */
IEEE80211_ADDR_COPY(ni->ni_bssid, vap->iv_bssid);
ni->ni_prev_vap = ni->ni_vap;
ni->ni_vap = vap;
ni->ni_ic = vap->iv_ic;
}
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
EXPORT_SYMBOL(ieee80211_node_reset);
struct ieee80211_node *
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_ref_node_debug(struct ieee80211_node *ni, const char *func, int line)
#else
ieee80211_ref_node(struct ieee80211_node *ni)
#endif
{
if (ni == NULL) {
printk(KERN_ERR "%s: NULL node.\n", __func__);
dump_stack();
return ni;
}
if (atomic_read(&ni->ni_refcnt) < 1) {
#ifdef IEEE80211_DEBUG
node_print_message(IEEE80211_MSG_ANY,
0 /* show counter */,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
0 /* adjust refcount */,
ni,
#ifdef IEEE80211_DEBUG_REFCNT
func, line,
#endif
__func__, __LINE__,
"attempt to access node with invalid refcount of %d."
" No changes made." /* message */,
atomic_read(&ni->ni_refcnt));
#endif
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
dump_stack();
return ni;
}
atomic_inc(&ni->ni_refcnt);
#ifdef IEEE80211_DEBUG
node_print_message(IEEE80211_MSG_NODE_REF,
0 /* show counter */,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
0 /* adjust refcount */,
ni,
#ifdef IEEE80211_DEBUG_REFCNT
func, line,
#endif
__func__, __LINE__,
"ref" /* message */);
#endif
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
return ni;
}
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_ref_node_debug);
#else
EXPORT_SYMBOL(ieee80211_ref_node);
#endif
void
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_unref_node_debug(struct ieee80211_node **pni, const char *func, int line)
#else
ieee80211_unref_node(struct ieee80211_node **pni)
#endif
{
struct ieee80211_node *ni = NULL;
if (pni == NULL) {
printk(KERN_ERR "%s: NULL ieee80211_node **\n", __func__);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
dump_stack();
return;
}
ni = *pni;
if (ni == NULL) {
printk(KERN_ERR "%s: NULL ieee80211_node *\n", __func__);
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
dump_stack();
return;
}
if (atomic_read(&ni->ni_refcnt) < 1) {
#ifdef IEEE80211_DEBUG
node_print_message(IEEE80211_MSG_ANY,
0 /* show counter */,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
0 /* adjust refcount */,
ni,
#ifdef IEEE80211_DEBUG_REFCNT
func, line,
#endif
__func__, __LINE__,
"attempt to access node with invalid refcount of %d."
" No changes made." /* message */,
atomic_read(&ni->ni_refcnt));
#endif
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
dump_stack();
return;
}
#ifdef IEEE80211_DEBUG
node_print_message(IEEE80211_MSG_NODE_REF,
0 /* show counter */,
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
-1 /* adjust refcount */,
ni,
#ifdef IEEE80211_DEBUG_REFCNT
func, line,
#endif
__func__, __LINE__,
"unref" /* message */);
#endif
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
if (atomic_dec_and_test(&ni->ni_refcnt)) {
This patch augments the current reference counting code with: * Counters for total outstanding instances for each resource type (skb, ath_node and ath_buf) * One pair of acquisition/release functions per resource type in unlocked and one in locked * Adds some more _debug versions of functions in the call chain that acquire/release resources so that the original func/line in the driver as well as the func/line that affected the resource use can be shown in the trace. Intermediate stack frames aren't necessary to trace the leaks. * Changes naming convention for "lock-required" functions to suffix _locked for the versions that expect locking, to be consistent with some other places in the code. * Consolidate debug messages to the helper functions that actually affect the reference count or acquire/release a resource * Additional sanity checks and leak detection (esp for detecting node ref leaks through skb) * skb references are nulled out by the new sbk unref/free function. I've tested these changes extensively and found lots of cases where we didn't get enough node references when cloning skbuff, and where the kernel drops packets due to performance issues and leaks our node references. With these changes and the tracing enabled I have verified that: * TX BUF: tx buffers always go down to zero when the tx queue is done, and you can watch tx queue usage ratio go up and down again as the driver is working. There are no leaks here at the moment, although there *are* some in the madwifi-dfs branch during CAC at the moment. * skbuff leaks in all the common flows are fixed. We were leaking node references in a lot of places where kernel was dropping skb's due to congestion and we were failing to increment node references when cloning skbuffs. These are now detected, as are skbuffs that are reaped by the kernel while still holding a node reference. * the ath_node count works correctly and on an idle system we get about 5 references per station table node, with 2 node instances per VAP. One for the bss and one for the node in the station table, I believe. The ath_node count goes up and down but always lands back at the stable number based on the vaps you have configured and the number of actual stations in the station table. The point here is that it's pretty constant what you will see over time, despite excessive node creation/release in our code during input (esp input_all). Thank god for the slab allocator. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@2902 0192ed92-7a03-0410-a25b-9323aeb14dbd
2007-11-21 23:14:11 +03:00
#ifdef IEEE80211_DEBUG_REFCNT
ieee80211_free_node_debug(ni, func, line);
#else /* #ifdef IEEE80211_DEBUG_REFCNT */
ieee80211_free_node(ni);
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
}
*pni = NULL;
}
#ifdef IEEE80211_DEBUG_REFCNT
EXPORT_SYMBOL(ieee80211_unref_node_debug);
#else
EXPORT_SYMBOL(ieee80211_unref_node);
#endif
int32_t
ieee80211_get_node_count(struct ieee80211com *ic)
{
return node_count(ic);
}
EXPORT_SYMBOL(ieee80211_get_node_count);