Aren/NetBSD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

improve fast-forward performance when the number of flows exceeds IPFLOW_MAX.

Browse Source 
In the fast-forward case, when the number of flows exceeds IPFLOW_MAX, the
performmance degraded to about 50% compared to the case less than IPFLOW_MAX
flows. This modification suppresses the degradation to 65%. Furthermore,
the modified kernel is about the same performance as the original kernel
when the number of flows is less than IPFLOW_MAX.

The original patch is implemented by ryo@n.o. Thanks.
This commit is contained in:
knakahara 2016-08-01 10:22:53 +00:00
parent 466f21f0b9
commit 56995fecb7
2 changed files with 65 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
sys/netinet/ip_flow.c
View File

@ -1,4 +1,4 @@
/* $NetBSD: ip_flow.c,v 1.75 2016/07/27 04:23:42 knakahara Exp $ */
/* $NetBSD: ip_flow.c,v 1.76 2016/08/01 10:22:53 knakahara Exp $ */
/*-
* Copyright (c) 1998 The NetBSD Foundation, Inc.
@ -30,7 +30,7 @@
*/
#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: ip_flow.c,v 1.75 2016/07/27 04:23:42 knakahara Exp $");
__KERNEL_RCSID(0, "$NetBSD: ip_flow.c,v 1.76 2016/08/01 10:22:53 knakahara Exp $");
#include <sys/param.h>
#include <sys/systm.h>
@ -69,7 +69,7 @@ __KERNEL_RCSID(0, "$NetBSD: ip_flow.c,v 1.75 2016/07/27 04:23:42 knakahara Exp $
static struct pool ipflow_pool;
LIST_HEAD(ipflowhead, ipflow);
TAILQ_HEAD(ipflowhead, ipflow);
#define IPFLOW_TIMER (5 * PR_SLOWHZ)
#define IPFLOW_DEFAULT_HASHSIZE (1 << IPFLOW_HASHBITS)
@ -86,16 +86,17 @@ static struct ipflowhead *ipflowtable = NULL;
static struct ipflowhead ipflowlist;
static int ipflow_inuse;
#define IPFLOW_INSERT(bucket, ipf) \
#define IPFLOW_INSERT(hashidx, ipf) \
do { \
LIST_INSERT_HEAD((bucket), (ipf), ipf_hash); \
LIST_INSERT_HEAD(&ipflowlist, (ipf), ipf_list); \
(ipf)->ipf_hashidx = (hashidx); \
TAILQ_INSERT_HEAD(&ipflowtable[(hashidx)], (ipf), ipf_hash); \
TAILQ_INSERT_HEAD(&ipflowlist, (ipf), ipf_list); \
} while (/*CONSTCOND*/ 0)
#define IPFLOW_REMOVE(ipf) \
#define IPFLOW_REMOVE(hashidx, ipf) \
do { \
LIST_REMOVE((ipf), ipf_hash); \
LIST_REMOVE((ipf), ipf_list); \
TAILQ_REMOVE(&ipflowtable[(hashidx)], (ipf), ipf_hash); \
TAILQ_REMOVE(&ipflowlist, (ipf), ipf_list); \
} while (/*CONSTCOND*/ 0)
#ifndef IPFLOW_MAX
@ -135,7 +136,7 @@ ipflow_lookup(const struct ip *ip)
hash = ipflow_hash(ip);
LIST_FOREACH(ipf, &ipflowtable[hash], ipf_hash) {
TAILQ_FOREACH(ipf, &ipflowtable[hash], ipf_hash) {
if (ip->ip_dst.s_addr == ipf->ipf_dst.s_addr
&& ip->ip_src.s_addr == ipf->ipf_src.s_addr
&& ip->ip_tos == ipf->ipf_tos)
@ -172,9 +173,9 @@ ipflow_reinit(int table_size)
ipflowtable = new_table;
ip_hashsize = table_size;
LIST_INIT(&ipflowlist);
TAILQ_INIT(&ipflowlist);
for (i = 0; i < ip_hashsize; i++)
LIST_INIT(&ipflowtable[i]);
TAILQ_INIT(&ipflowtable[i]);
return 0;
}
@ -328,6 +329,18 @@ ipflow_fastforward(struct mbuf *m)
* Send the packet on its way. All we can get back is ENOBUFS
*/
ipf->ipf_uses++;
#if 0
/*
* Sorting list is too heavy for fast path(packet processing path).
* It degrades about 10% performance. So, we does not sort ipflowtable,
* and then we use FIFO cache replacement instead fo LRU.
*/
/* move to head (LRU) for ipflowlist. ipflowtable ooes not care LRU. */
TAILQ_REMOVE(&ipflowlist, ipf, ipf_list);
TAILQ_INSERT_HEAD(&ipflowlist, ipf, ipf_list);
#endif
PRT_SLOW_ARM(ipf->ipf_timer, IPFLOW_TIMER);
if (rt->rt_flags & RTF_GATEWAY)
@ -375,7 +388,7 @@ ipflow_free(struct ipflow *ipf)
* Once it's off the list, we can deal with it at normal
* network IPL.
*/
IPFLOW_REMOVE(ipf);
IPFLOW_REMOVE(ipf->ipf_hashidx, ipf);
ipflow_addstats(ipf);
rtcache_free(&ipf->ipf_ro);
@ -386,14 +399,34 @@ ipflow_free(struct ipflow *ipf)
static struct ipflow *
ipflow_reap(bool just_one)
{
struct ipflow *ipf;
KASSERT(mutex_owned(&ipflow_lock));
while (just_one || ipflow_inuse > ip_maxflows) {
struct ipflow *ipf, *maybe_ipf = NULL;
/*
* This case must remove one ipflow. Furthermore, this case is used in
* fast path(packet processing path). So, simply remove TAILQ_LAST one.
*/
if (just_one) {
ipf = TAILQ_LAST(&ipflowlist, ipflowhead);
KASSERT(ipf != NULL);
ipf = LIST_FIRST(&ipflowlist);
while (ipf != NULL) {
IPFLOW_REMOVE(ipf->ipf_hashidx, ipf);
ipflow_addstats(ipf);
rtcache_free(&ipf->ipf_ro);
return ipf;
}
/*
* This case is used in slow path(sysctl).
* At first, remove invalid rtcache ipflow, and then remove TAILQ_LAST
* ipflow if it is ensured least recently used by comparing last_uses.
*/
while (ipflow_inuse > ip_maxflows) {
struct ipflow *maybe_ipf = TAILQ_LAST(&ipflowlist, ipflowhead);
TAILQ_FOREACH(ipf, &ipflowlist, ipf_list) {
/*
* If this no longer points to a valid route
* reclaim it.
@ -405,26 +438,21 @@ ipflow_reap(bool just_one)
* used or has had the least uses in the
* last 1.5 intervals.
*/
if (maybe_ipf == NULL ||
ipf->ipf_timer < maybe_ipf->ipf_timer ||
(ipf->ipf_timer == maybe_ipf->ipf_timer &&
ipf->ipf_last_uses + ipf->ipf_uses <
maybe_ipf->ipf_last_uses +
maybe_ipf->ipf_uses))
if (ipf->ipf_timer < maybe_ipf->ipf_timer
|| ((ipf->ipf_timer == maybe_ipf->ipf_timer)
&& (ipf->ipf_last_uses + ipf->ipf_uses
< maybe_ipf->ipf_last_uses + maybe_ipf->ipf_uses)))
maybe_ipf = ipf;
ipf = LIST_NEXT(ipf, ipf_list);
}
ipf = maybe_ipf;
done:
/*
* Remove the entry from the flow table.
*/
IPFLOW_REMOVE(ipf);
IPFLOW_REMOVE(ipf->ipf_hashidx, ipf);
ipflow_addstats(ipf);
rtcache_free(&ipf->ipf_ro);
if (just_one)
return ipf;
pool_put(&ipflow_pool, ipf);
ipflow_inuse--;
}
@ -446,8 +474,8 @@ ipflow_slowtimo_work(struct work *wk, void *arg)
mutex_enter(softnet_lock);
mutex_enter(&ipflow_lock);
KERNEL_LOCK(1, NULL);
for (ipf = LIST_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
next_ipf = LIST_NEXT(ipf, ipf_list);
for (ipf = TAILQ_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
next_ipf = TAILQ_NEXT(ipf, ipf_list);
if (PRT_SLOW_ISEXPIRED(ipf->ipf_timer) ||
(rt = rtcache_validate(&ipf->ipf_ro)) == NULL) {
ipflow_free(ipf);
@ -514,7 +542,7 @@ ipflow_create(const struct route *ro, struct mbuf *m)
}
memset(ipf, 0, sizeof(*ipf));
} else {
IPFLOW_REMOVE(ipf);
IPFLOW_REMOVE(ipf->ipf_hashidx, ipf);
ipflow_addstats(ipf);
rtcache_free(&ipf->ipf_ro);
@ -535,7 +563,7 @@ ipflow_create(const struct route *ro, struct mbuf *m)
* Insert into the approriate bucket of the flow table.
*/
hash = ipflow_hash(ip);
IPFLOW_INSERT(&ipflowtable[hash], ipf);
IPFLOW_INSERT(hash, ipf);
out:
KERNEL_UNLOCK_ONE(NULL);
@ -552,8 +580,8 @@ ipflow_invalidate_all(int new_size)
mutex_enter(&ipflow_lock);
for (ipf = LIST_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
next_ipf = LIST_NEXT(ipf, ipf_list);
for (ipf = TAILQ_FIRST(&ipflowlist); ipf != NULL; ipf = next_ipf) {
next_ipf = TAILQ_NEXT(ipf, ipf_list);
ipflow_free(ipf);
}

7
sys/netinet/ip_var.h
View File

@ -1,4 +1,4 @@
/* $NetBSD: ip_var.h,v 1.114 2016/06/21 03:28:27 ozaki-r Exp $ */
/* $NetBSD: ip_var.h,v 1.115 2016/08/01 10:22:53 knakahara Exp $ */
/*
* Copyright (c) 1982, 1986, 1993
@ -52,8 +52,9 @@ struct ipovly {
* IP Flow structure
*/
struct ipflow {
LIST_ENTRY(ipflow) ipf_list; /* next in active list */
LIST_ENTRY(ipflow) ipf_hash; /* next ipflow in bucket */
TAILQ_ENTRY(ipflow) ipf_list; /* next in active list */
TAILQ_ENTRY(ipflow) ipf_hash; /* next ipflow in bucket */
size_t ipf_hashidx; /* own hash index of ipflowtable[] */
struct in_addr ipf_dst; /* destination address */
struct in_addr ipf_src; /* source address */
uint8_t ipf_tos; /* type-of-service */