mirrors/madwifi
1
0
Fork 0
mirror of https://github.com/proski/madwifi synced 2024-11-22 06:21:47 +03:00
Code Issues Packages Projects Releases Wiki Activity
5890429464
madwifi/net80211/ieee80211_skb.h

204 lines
7.8 KiB
C
Raw Normal View History

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
/*-
* Copyright (c) 2007 Michael Taylor, Apprion
* 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.
*
* 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: ieee80211_skb.h 2647 2007-08-09 08:43:58Z mtaylor $
*/
#ifndef _NET80211_IEEE80211_SKB_H_
#define _NET80211_IEEE80211_SKB_H_
/*******************************************************************************
* Globals
******************************************************************************/
#ifdef IEEE80211_DEBUG_REFCNT
/* Count of currently tracked skbs */
extern atomic_t skb_total_counter;
/* Count of currently tracked skbs' references */
extern atomic_t skb_refs_counter;
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
/*******************************************************************************
* Public API
******************************************************************************/
Rename and move, into the right headers, the SKB and BF macros git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3596 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-05-04 21:48:59 +04:00
/* SKB_XX(...) macros will blow up if _skb is NULL (detect problems early) */
#define SKB_CB(_skb) ((struct ieee80211_cb *)(_skb)->cb)
#define SKB_NI(_skb) (SKB_CB(_skb)->ni)
#define M_FLAG_SET(_skb, _flag) \
(SKB_CB(_skb)->flags |= (_flag))
#define M_FLAG_CLR(_skb, _flag) \
(SKB_CB(_skb)->flags &= ~(_flag))
#define M_FLAG_GET(_skb, _flag) \
(SKB_CB(_skb)->flags & (_flag))
#define M_FLAG_KEEP_ONLY(_skb, _flag) \
(SKB_CB(_skb)->flags &= (_flag))
#define M_PWR_SAV_SET(skb) M_FLAG_SET((skb), M_PWR_SAV)
#define M_PWR_SAV_CLR(skb) M_FLAG_CLR((skb), M_PWR_SAV)
#define M_PWR_SAV_GET(skb) M_FLAG_GET((skb), M_PWR_SAV)
/* SKBs on the power save queue are tagged with an age and
* timed out. We reuse the hardware checksum field in the
* mbuf packet header to store this data.
Make pointer declarations and casting consistent. Some other formatting changes are made in passing git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3710 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-07 00:32:29 +04:00
* XXX: use private CB area. */
Rename and move, into the right headers, the SKB and BF macros git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3596 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-05-04 21:48:59 +04:00
#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,20)
#define skb_age csum_offset
#else
#define skb_age csum
#endif
#define M_AGE_SET(_skb,_v) ((_skb)->skb_age = (_v))
#define M_AGE_GET(_skb) ((_skb)->skb_age)
#define M_AGE_SUB(_skb,_adj) ((_skb)->skb_age -= (_adj))
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 will release one reference from SKB.
* If SKB refcount is going to zero:
* - Free the node reference and set it to null.
* - Break the linked list, clearing next skb's prev pointer if possible. */
Make pointer declarations and casting consistent. Some other formatting changes are made in passing git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3710 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-07 00:32:29 +04:00
void ieee80211_dev_kfree_skb(struct sk_buff **pskb);
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
Add ieee80211_skb_copy_noderef, which hopefully does what it says on the tin (i.e., copies a node reference, not a pointer, from one skb to another). git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3163 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-01-16 23:07:34 +03:00
static inline void ieee80211_skb_copy_noderef(struct sk_buff *src,
struct sk_buff *dst)
{
SKB handling code cleanup: * Convert last SKB_CB()->ni to SKB_NI * Lots of comments * Refactor various pieces of code for consistency, readability, and removal of goto git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3666 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-05-22 16:58:24 +04:00
if (SKB_NI(src) != NULL)
SKB_NI(dst) = ieee80211_ref_node(SKB_NI(src));
Add ieee80211_skb_copy_noderef, which hopefully does what it says on the tin (i.e., copies a node reference, not a pointer, from one skb to another). git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3163 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-01-16 23:07:34 +03:00
}
Make pointer declarations and casting consistent. Some other formatting changes are made in passing git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3710 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-07 00:32:29 +04:00
/* ieee80211_dev_kfree_skb_list will invoke ieee80211_dev_kfree_skb on each node in
* a list of skbs, starting with the first. */
void ieee80211_dev_kfree_skb_list(struct sk_buff **pskb);
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
Make pointer declarations and casting consistent. Some other formatting changes are made in passing git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3710 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-07 00:32:29 +04:00
struct sk_buff *ieee80211_dev_alloc_skb(int size);
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
void ieee80211_skb_track(struct sk_buff *skb);
void ieee80211_skb_untrack(struct sk_buff *skb);
#ifdef IEEE80211_DEBUG_REFCNT
int ieee80211_skb_counter(void);
int ieee80211_skb_references(void);
#else
#define ieee80211_skb_counter() (0)
#define ieee80211_skb_references() (0)
#endif
/*******************************************************************************
* skbuff leak/refcount debugging Replacement Functions
******************************************************************************/
#ifdef IEEE80211_DEBUG_REFCNT
Use vlan_hwaccel_rx instead of vlan_hwaccel_receive_skb, as this is the non-polling (non-NAPI) version. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3709 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-06 23:42:24 +04:00
int vlan_hwaccel_rx_debug(struct sk_buff *skb, struct vlan_group *grp,
Make pointer declarations and casting consistent. Some other formatting changes are made in passing git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3710 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-07 00:32:29 +04:00
unsigned short vlan_tag, const char *func, int line);
int netif_rx_debug(struct sk_buff *skb, const char *func, int line);
struct sk_buff *alloc_skb_debug(unsigned int length, gfp_t gfp_mask,
const char *func, int line);
struct sk_buff *dev_alloc_skb_debug(unsigned int length,
const char *func, int line);
struct sk_buff *skb_clone_debug(struct sk_buff *skb, gfp_t pri,
const char *func, int line);
struct sk_buff *skb_copy_debug(struct sk_buff *skb, gfp_t pri,
const char *func, int line);
struct sk_buff *skb_get_debug(struct sk_buff *skb,
const char *func, int line);
struct sk_buff *skb_realloc_headroom_debug(struct sk_buff *skb, unsigned int headroom,
const char *func, int line);
struct sk_buff *pskb_copy_debug(struct sk_buff *skb, gfp_t pri,
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
int dev_queue_xmit_debug(struct sk_buff *skb,
Make pointer declarations and casting consistent. Some other formatting changes are made in passing git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3710 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-07 00:32:29 +04:00
const char *func, int line);
struct sk_buff *skb_share_check_debug(struct sk_buff *skb, gfp_t pri,
const char *func, int line);
void kfree_skb_fast_debug(struct sk_buff *skb,
const char *func, int line);
struct sk_buff *skb_unshare_debug(struct sk_buff *skb, gfp_t pri,
const char *func, int line);
struct sk_buff *skb_copy_expand_debug(const struct sk_buff *skb, int newheadroom,
int newtailroom, gfp_t gfp_mask,
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
#undef alloc_skb
#undef dev_alloc_skb
#undef dev_kfree_skb
#undef dev_kfree_skb_any
#undef dev_kfree_skb_irq
#undef dev_queue_xmit
#undef kfree_skb
#undef kfree_skb_fast
#undef netif_rx
#undef pskb_copy
#undef skb_clone
#undef skb_copy
#undef skb_copy_expand
#undef skb_get
#undef skb_realloc_headroom
#undef skb_share_check
#undef skb_unshare
Use vlan_hwaccel_rx instead of vlan_hwaccel_receive_skb, as this is the non-polling (non-NAPI) version. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3709 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-06 23:42:24 +04:00
#undef vlan_hwaccel_rx
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
#define skb_unshare(_skb, _pri) \
skb_unshare_debug(_skb, _pri, __func__, __LINE__)
#define skb_copy_expand(_skb, _newheadroom, _newtailroom, _gfp_mask) \
skb_copy_expand_debug(_skb, _newheadroom, _newtailroom, _gfp_mask, __func__, __LINE__)
Use vlan_hwaccel_rx instead of vlan_hwaccel_receive_skb, as this is the non-polling (non-NAPI) version. git-svn-id: http://madwifi-project.org/svn/madwifi/trunk@3709 0192ed92-7a03-0410-a25b-9323aeb14dbd
2008-06-06 23:42:24 +04:00
#define vlan_hwaccel_rx(_skb, _grp, _tag) \
vlan_hwaccel_rx_debug(_skb, _grp, _tag, __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
#define netif_rx(_skb) \
netif_rx_debug(_skb, __func__, __LINE__)
#define alloc_skb(_length, _gfp_mask) \
alloc_skb_debug(_length, _gfp_mask, __func__, __LINE__)
#define dev_alloc_skb(_length) \
dev_alloc_skb_debug(_length, __func__, __LINE__)
#define dev_kfree_skb_irq(_skb) \
unref_skb(_skb, UNREF_USE_DEV_KFREE_SKB_IRQ, __func__, __LINE__)
#define dev_kfree_skb_any(_skb) \
unref_skb(_skb, UNREF_USE_DEV_KFREE_SKB_ANY, __func__, __LINE__)
#define dev_kfree_skb(_skb) \
unref_skb(_skb, UNREF_USE_DEV_KFREE_SKB, __func__, __LINE__)
#define kfree_skb(_skb) \
unref_skb(_skb, UNREF_USE_KFREE_SKB, __func__, __LINE__)
#define skb_clone(_skb, _pri) \
skb_clone_debug(_skb, _pri, __func__, __LINE__)
#define skb_share_check(_skb, _pri) \
skb_share_check_debug(_skb, _pri, __func__, __LINE__)
#define kfree_skb_fast(_skb) \
kfree_skb_fast_debug(_skb, __func__, __LINE__)
#define skb_realloc_headroom(_skb, _headroom) \
skb_realloc_headroom_debug(_skb, _headroom, __func__, __LINE__)
#define pskb_copy(_skb, _pri) \
pskb_copy_debug(_skb, _pri, __func__, __LINE__)
#define skb_get(_skb) \
skb_get_debug(_skb, __func__, __LINE__)
#define skb_copy(_skb, _pri) \
skb_copy_debug(_skb, _pri, __func__, __LINE__)
#define dev_queue_xmit(_skb) \
dev_queue_xmit_debug(_skb, __func__, __LINE__)
#endif /* #ifdef IEEE80211_DEBUG_REFCNT */
#endif /* _NET80211_IEEE80211_SKB_H_ */
Reference in New Issue Copy Permalink