linux/net/core/dst.c

324 lines
7.8 KiB
C
Raw Normal View History

/*
* net/core/dst.c Protocol independent destination cache.
*
* Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
*
*/
#include <linux/bitops.h>
#include <linux/errno.h>
#include <linux/init.h>
#include <linux/kernel.h>
[PATCH] NET : convert IP route cache garbage collection from softirq processing to a workqueue When the periodic IP route cache flush is done (every 600 seconds on default configuration), some hosts suffer a lot and eventually trigger the "soft lockup" message. dst_run_gc() is doing a scan of a possibly huge list of dst_entries, eventually freeing some (less than 1%) of them, while holding the dst_lock spinlock for the whole scan. Then it rearms a timer to redo the full thing 1/10 s later... The slowdown can last one minute or so, depending on how active are the tcp sessions. This second version of the patch converts the processing from a softirq based one to a workqueue. Even if the list of entries in garbage_list is huge, host is still responsive to softirqs and can make progress. Instead of resetting gc timer to 0.1 second if one entry was freed in a gc run, we do this if more than 10% of entries were freed. Before patch : Aug 16 06:21:37 SRV1 kernel: BUG: soft lockup detected on CPU#0! Aug 16 06:21:37 SRV1 kernel: Aug 16 06:21:37 SRV1 kernel: Call Trace: Aug 16 06:21:37 SRV1 kernel: <IRQ> [<ffffffff802286f0>] wake_up_process+0x10/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80251e09>] softlockup_tick+0xe9/0x110 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd380>] dst_run_gc+0x0/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802376f3>] run_local_timers+0x13/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802379c7>] update_process_times+0x57/0x90 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80216034>] smp_local_timer_interrupt+0x34/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802165cc>] smp_apic_timer_interrupt+0x5c/0x80 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8020a816>] apic_timer_interrupt+0x66/0x70 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd3d3>] dst_run_gc+0x53/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd3c6>] dst_run_gc+0x46/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80237148>] run_timer_softirq+0x148/0x1c0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8023340c>] __do_softirq+0x6c/0xe0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8020ad6c>] call_softirq+0x1c/0x30 Aug 16 06:21:37 SRV1 kernel: <EOI> [<ffffffff8020cb34>] do_softirq+0x34/0x90 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802331cf>] local_bh_enable_ip+0x3f/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422913>] _spin_unlock_bh+0x13/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803dfde8>] rt_garbage_collect+0x1d8/0x320 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd4dd>] dst_alloc+0x1d/0xa0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803e1433>] __ip_route_output_key+0x573/0x800 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c02e2>] sock_common_recvmsg+0x32/0x50 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803e16dc>] ip_route_output_flow+0x1c/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80400160>] tcp_v4_connect+0x150/0x610 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803ebf07>] inet_bind_bucket_create+0x17/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8040cd16>] inet_stream_connect+0xa6/0x2c0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422981>] _spin_lock_bh+0x11/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c0bdf>] lock_sock_nested+0xcf/0xe0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422981>] _spin_lock_bh+0x11/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803be551>] sys_connect+0x71/0xa0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803eee3f>] tcp_setsockopt+0x1f/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c030f>] sock_common_setsockopt+0xf/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803be4bd>] sys_setsockopt+0x9d/0xc0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8028881e>] sys_ioctl+0x5e/0x80 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80209c4e>] system_call+0x7e/0x83 After patch : (RT_CACHE_DEBUG set to 2 to get following traces) dst_total: 75469 delayed: 74109 work_perf: 141 expires: 150 elapsed: 8092 us dst_total: 78725 delayed: 73366 work_perf: 743 expires: 400 elapsed: 8542 us dst_total: 86126 delayed: 71844 work_perf: 1522 expires: 775 elapsed: 8849 us dst_total: 100173 delayed: 68791 work_perf: 3053 expires: 1256 elapsed: 9748 us dst_total: 121798 delayed: 64711 work_perf: 4080 expires: 1997 elapsed: 10146 us dst_total: 154522 delayed: 58316 work_perf: 6395 expires: 25 elapsed: 11402 us dst_total: 154957 delayed: 58252 work_perf: 64 expires: 150 elapsed: 6148 us dst_total: 157377 delayed: 57843 work_perf: 409 expires: 400 elapsed: 6350 us dst_total: 163745 delayed: 56679 work_perf: 1164 expires: 775 elapsed: 7051 us dst_total: 176577 delayed: 53965 work_perf: 2714 expires: 1389 elapsed: 8120 us dst_total: 198993 delayed: 49627 work_perf: 4338 expires: 1997 elapsed: 8909 us dst_total: 226638 delayed: 46865 work_perf: 2762 expires: 2748 elapsed: 7351 us I successfully reduced the IP route cache of many hosts by a four factor thanks to this patch. Previously, I had to disable "ip route flush cache" to avoid crashes. Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-12 12:29:01 +00:00
#include <linux/workqueue.h>
#include <linux/mm.h>
#include <linux/module.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
#include <linux/slab.h>
#include <linux/netdevice.h>
#include <linux/skbuff.h>
#include <linux/string.h>
#include <linux/types.h>
#include <net/net_namespace.h>
#include <linux/sched.h>
#include <linux/prefetch.h>
#include <net/lwtunnel.h>
#include <net/dst.h>
#include <net/dst_metadata.h>
[PATCH] NET : convert IP route cache garbage collection from softirq processing to a workqueue When the periodic IP route cache flush is done (every 600 seconds on default configuration), some hosts suffer a lot and eventually trigger the "soft lockup" message. dst_run_gc() is doing a scan of a possibly huge list of dst_entries, eventually freeing some (less than 1%) of them, while holding the dst_lock spinlock for the whole scan. Then it rearms a timer to redo the full thing 1/10 s later... The slowdown can last one minute or so, depending on how active are the tcp sessions. This second version of the patch converts the processing from a softirq based one to a workqueue. Even if the list of entries in garbage_list is huge, host is still responsive to softirqs and can make progress. Instead of resetting gc timer to 0.1 second if one entry was freed in a gc run, we do this if more than 10% of entries were freed. Before patch : Aug 16 06:21:37 SRV1 kernel: BUG: soft lockup detected on CPU#0! Aug 16 06:21:37 SRV1 kernel: Aug 16 06:21:37 SRV1 kernel: Call Trace: Aug 16 06:21:37 SRV1 kernel: <IRQ> [<ffffffff802286f0>] wake_up_process+0x10/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80251e09>] softlockup_tick+0xe9/0x110 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd380>] dst_run_gc+0x0/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802376f3>] run_local_timers+0x13/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802379c7>] update_process_times+0x57/0x90 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80216034>] smp_local_timer_interrupt+0x34/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802165cc>] smp_apic_timer_interrupt+0x5c/0x80 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8020a816>] apic_timer_interrupt+0x66/0x70 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd3d3>] dst_run_gc+0x53/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd3c6>] dst_run_gc+0x46/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80237148>] run_timer_softirq+0x148/0x1c0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8023340c>] __do_softirq+0x6c/0xe0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8020ad6c>] call_softirq+0x1c/0x30 Aug 16 06:21:37 SRV1 kernel: <EOI> [<ffffffff8020cb34>] do_softirq+0x34/0x90 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802331cf>] local_bh_enable_ip+0x3f/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422913>] _spin_unlock_bh+0x13/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803dfde8>] rt_garbage_collect+0x1d8/0x320 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd4dd>] dst_alloc+0x1d/0xa0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803e1433>] __ip_route_output_key+0x573/0x800 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c02e2>] sock_common_recvmsg+0x32/0x50 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803e16dc>] ip_route_output_flow+0x1c/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80400160>] tcp_v4_connect+0x150/0x610 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803ebf07>] inet_bind_bucket_create+0x17/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8040cd16>] inet_stream_connect+0xa6/0x2c0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422981>] _spin_lock_bh+0x11/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c0bdf>] lock_sock_nested+0xcf/0xe0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422981>] _spin_lock_bh+0x11/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803be551>] sys_connect+0x71/0xa0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803eee3f>] tcp_setsockopt+0x1f/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c030f>] sock_common_setsockopt+0xf/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803be4bd>] sys_setsockopt+0x9d/0xc0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8028881e>] sys_ioctl+0x5e/0x80 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80209c4e>] system_call+0x7e/0x83 After patch : (RT_CACHE_DEBUG set to 2 to get following traces) dst_total: 75469 delayed: 74109 work_perf: 141 expires: 150 elapsed: 8092 us dst_total: 78725 delayed: 73366 work_perf: 743 expires: 400 elapsed: 8542 us dst_total: 86126 delayed: 71844 work_perf: 1522 expires: 775 elapsed: 8849 us dst_total: 100173 delayed: 68791 work_perf: 3053 expires: 1256 elapsed: 9748 us dst_total: 121798 delayed: 64711 work_perf: 4080 expires: 1997 elapsed: 10146 us dst_total: 154522 delayed: 58316 work_perf: 6395 expires: 25 elapsed: 11402 us dst_total: 154957 delayed: 58252 work_perf: 64 expires: 150 elapsed: 6148 us dst_total: 157377 delayed: 57843 work_perf: 409 expires: 400 elapsed: 6350 us dst_total: 163745 delayed: 56679 work_perf: 1164 expires: 775 elapsed: 7051 us dst_total: 176577 delayed: 53965 work_perf: 2714 expires: 1389 elapsed: 8120 us dst_total: 198993 delayed: 49627 work_perf: 4338 expires: 1997 elapsed: 8909 us dst_total: 226638 delayed: 46865 work_perf: 2762 expires: 2748 elapsed: 7351 us I successfully reduced the IP route cache of many hosts by a four factor thanks to this patch. Previously, I had to disable "ip route flush cache" to avoid crashes. Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-12 12:29:01 +00:00
/*
* Theory of operations:
* 1) We use a list, protected by a spinlock, to add
* new entries from both BH and non-BH context.
* 2) In order to keep spinlock held for a small delay,
* we use a second list where are stored long lived
* entries, that are handled by the garbage collect thread
* fired by a workqueue.
* 3) This list is guarded by a mutex,
* so that the gc_task and dst_dev_event() can be synchronized.
*/
[PATCH] NET : convert IP route cache garbage collection from softirq processing to a workqueue When the periodic IP route cache flush is done (every 600 seconds on default configuration), some hosts suffer a lot and eventually trigger the "soft lockup" message. dst_run_gc() is doing a scan of a possibly huge list of dst_entries, eventually freeing some (less than 1%) of them, while holding the dst_lock spinlock for the whole scan. Then it rearms a timer to redo the full thing 1/10 s later... The slowdown can last one minute or so, depending on how active are the tcp sessions. This second version of the patch converts the processing from a softirq based one to a workqueue. Even if the list of entries in garbage_list is huge, host is still responsive to softirqs and can make progress. Instead of resetting gc timer to 0.1 second if one entry was freed in a gc run, we do this if more than 10% of entries were freed. Before patch : Aug 16 06:21:37 SRV1 kernel: BUG: soft lockup detected on CPU#0! Aug 16 06:21:37 SRV1 kernel: Aug 16 06:21:37 SRV1 kernel: Call Trace: Aug 16 06:21:37 SRV1 kernel: <IRQ> [<ffffffff802286f0>] wake_up_process+0x10/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80251e09>] softlockup_tick+0xe9/0x110 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd380>] dst_run_gc+0x0/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802376f3>] run_local_timers+0x13/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802379c7>] update_process_times+0x57/0x90 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80216034>] smp_local_timer_interrupt+0x34/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802165cc>] smp_apic_timer_interrupt+0x5c/0x80 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8020a816>] apic_timer_interrupt+0x66/0x70 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd3d3>] dst_run_gc+0x53/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd3c6>] dst_run_gc+0x46/0x140 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80237148>] run_timer_softirq+0x148/0x1c0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8023340c>] __do_softirq+0x6c/0xe0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8020ad6c>] call_softirq+0x1c/0x30 Aug 16 06:21:37 SRV1 kernel: <EOI> [<ffffffff8020cb34>] do_softirq+0x34/0x90 Aug 16 06:21:37 SRV1 kernel: [<ffffffff802331cf>] local_bh_enable_ip+0x3f/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422913>] _spin_unlock_bh+0x13/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803dfde8>] rt_garbage_collect+0x1d8/0x320 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803cd4dd>] dst_alloc+0x1d/0xa0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803e1433>] __ip_route_output_key+0x573/0x800 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c02e2>] sock_common_recvmsg+0x32/0x50 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803e16dc>] ip_route_output_flow+0x1c/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80400160>] tcp_v4_connect+0x150/0x610 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803ebf07>] inet_bind_bucket_create+0x17/0x60 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8040cd16>] inet_stream_connect+0xa6/0x2c0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422981>] _spin_lock_bh+0x11/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c0bdf>] lock_sock_nested+0xcf/0xe0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80422981>] _spin_lock_bh+0x11/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803be551>] sys_connect+0x71/0xa0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803eee3f>] tcp_setsockopt+0x1f/0x30 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803c030f>] sock_common_setsockopt+0xf/0x20 Aug 16 06:21:37 SRV1 kernel: [<ffffffff803be4bd>] sys_setsockopt+0x9d/0xc0 Aug 16 06:21:37 SRV1 kernel: [<ffffffff8028881e>] sys_ioctl+0x5e/0x80 Aug 16 06:21:37 SRV1 kernel: [<ffffffff80209c4e>] system_call+0x7e/0x83 After patch : (RT_CACHE_DEBUG set to 2 to get following traces) dst_total: 75469 delayed: 74109 work_perf: 141 expires: 150 elapsed: 8092 us dst_total: 78725 delayed: 73366 work_perf: 743 expires: 400 elapsed: 8542 us dst_total: 86126 delayed: 71844 work_perf: 1522 expires: 775 elapsed: 8849 us dst_total: 100173 delayed: 68791 work_perf: 3053 expires: 1256 elapsed: 9748 us dst_total: 121798 delayed: 64711 work_perf: 4080 expires: 1997 elapsed: 10146 us dst_total: 154522 delayed: 58316 work_perf: 6395 expires: 25 elapsed: 11402 us dst_total: 154957 delayed: 58252 work_perf: 64 expires: 150 elapsed: 6148 us dst_total: 157377 delayed: 57843 work_perf: 409 expires: 400 elapsed: 6350 us dst_total: 163745 delayed: 56679 work_perf: 1164 expires: 775 elapsed: 7051 us dst_total: 176577 delayed: 53965 work_perf: 2714 expires: 1389 elapsed: 8120 us dst_total: 198993 delayed: 49627 work_perf: 4338 expires: 1997 elapsed: 8909 us dst_total: 226638 delayed: 46865 work_perf: 2762 expires: 2748 elapsed: 7351 us I successfully reduced the IP route cache of many hosts by a four factor thanks to this patch. Previously, I had to disable "ip route flush cache" to avoid crashes. Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-12 12:29:01 +00:00
/*
* We want to keep lock & list close together
* to dirty as few cache lines as possible in __dst_free().
* As this is not a very strong hint, we dont force an alignment on SMP.
*/
int dst_discard_out(struct net *net, struct sock *sk, struct sk_buff *skb)
{
kfree_skb(skb);
return 0;
}
EXPORT_SYMBOL(dst_discard_out);
const struct dst_metrics dst_default_metrics = {
/* This initializer is needed to force linker to place this variable
* into const section. Otherwise it might end into bss section.
* We really want to avoid false sharing on this variable, and catch
* any writes on it.
*/
.refcnt = ATOMIC_INIT(1),
};
void dst_init(struct dst_entry *dst, struct dst_ops *ops,
struct net_device *dev, int initial_ref, int initial_obsolete,
unsigned short flags)
{
dst->child = NULL;
dst->dev = dev;
if (dev)
dev_hold(dev);
dst->ops = ops;
dst_init_metrics(dst, dst_default_metrics.metrics, true);
dst->expires = 0UL;
dst->path = dst;
ipv6: fix race condition regarding dst->expires and dst->from. Eric Dumazet wrote: | Some strange crashes happen in rt6_check_expired(), with access | to random addresses. | | At first glance, it looks like the RTF_EXPIRES and | stuff added in commit 1716a96101c49186b | (ipv6: fix problem with expired dst cache) | are racy : same dst could be manipulated at the same time | on different cpus. | | At some point, our stack believes rt->dst.from contains a dst pointer, | while its really a jiffie value (as rt->dst.expires shares the same area | of memory) | | rt6_update_expires() should be fixed, or am I missing something ? | | CC Neil because of https://bugzilla.redhat.com/show_bug.cgi?id=892060 Because we do not have any locks for dst_entry, we cannot change essential structure in the entry; e.g., we cannot change reference to other entity. To fix this issue, split 'from' and 'expires' field in dst_entry out of union. Once it is 'from' is assigned in the constructor, keep the reference until the very last stage of the life time of the object. Of course, it is unsafe to change 'from', so make rt6_set_from simple just for fresh entries. Reported-by: Eric Dumazet <eric.dumazet@gmail.com> Reported-by: Neil Horman <nhorman@tuxdriver.com> CC: Gao Feng <gaofeng@cn.fujitsu.com> Signed-off-by: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Reviewed-by: Eric Dumazet <edumazet@google.com> Reported-by: Steinar H. Gunderson <sesse@google.com> Reviewed-by: Neil Horman <nhorman@tuxdriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2013-02-20 00:29:08 +00:00
dst->from = NULL;
#ifdef CONFIG_XFRM
dst->xfrm = NULL;
#endif
dst->input = dst_discard;
dst->output = dst_discard_out;
dst->error = 0;
dst->obsolete = initial_obsolete;
dst->header_len = 0;
dst->trailer_len = 0;
#ifdef CONFIG_IP_ROUTE_CLASSID
dst->tclassid = 0;
#endif
dst->lwtstate = NULL;
atomic_set(&dst->__refcnt, initial_ref);
dst->__use = 0;
dst->lastuse = jiffies;
dst->flags = flags;
dst->next = NULL;
if (!(flags & DST_NOCOUNT))
dst_entries_add(ops, 1);
}
EXPORT_SYMBOL(dst_init);
void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
int initial_ref, int initial_obsolete, unsigned short flags)
{
struct dst_entry *dst;
if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
if (ops->gc(ops))
return NULL;
}
dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
if (!dst)
return NULL;
dst_init(dst, ops, dev, initial_ref, initial_obsolete, flags);
return dst;
}
EXPORT_SYMBOL(dst_alloc);
struct dst_entry *dst_destroy(struct dst_entry * dst)
{
struct dst_entry *child;
smp_rmb();
child = dst->child;
if (!(dst->flags & DST_NOCOUNT))
dst_entries_add(dst->ops, -1);
if (dst->ops->destroy)
dst->ops->destroy(dst);
if (dst->dev)
dev_put(dst->dev);
route: fix a use-after-free This patch fixes the following crash: general protection fault: 0000 [#1] SMP DEBUG_PAGEALLOC CPU: 1 PID: 0 Comm: swapper/1 Not tainted 4.2.0-rc7+ #166 Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011 task: ffff88010656d280 ti: ffff880106570000 task.ti: ffff880106570000 RIP: 0010:[<ffffffff8182f91b>] [<ffffffff8182f91b>] dst_destroy+0xa6/0xef RSP: 0018:ffff880107603e38 EFLAGS: 00010202 RAX: 0000000000000001 RBX: ffff8800d225a000 RCX: ffffffff82250fd0 RDX: 0000000000000001 RSI: ffffffff82250fd0 RDI: 6b6b6b6b6b6b6b6b RBP: ffff880107603e58 R08: 0000000000000001 R09: 0000000000000001 R10: 000000000000b530 R11: ffff880107609000 R12: 0000000000000000 R13: ffffffff82343c40 R14: 0000000000000000 R15: ffffffff8182fb4f FS: 0000000000000000(0000) GS:ffff880107600000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: 00007fcabd9d3000 CR3: 00000000d7279000 CR4: 00000000000006e0 Stack: ffffffff82250fd0 ffff8801077d6f00 ffffffff82253c40 ffff8800d225a000 ffff880107603e68 ffffffff8182fb5d ffff880107603f08 ffffffff810d795e ffffffff810d7648 ffff880106574000 ffff88010656d280 ffff88010656d280 Call Trace: <IRQ> [<ffffffff8182fb5d>] dst_destroy_rcu+0xe/0x1d [<ffffffff810d795e>] rcu_process_callbacks+0x618/0x7eb [<ffffffff810d7648>] ? rcu_process_callbacks+0x302/0x7eb [<ffffffff8182fb4f>] ? dst_gc_task+0x1eb/0x1eb [<ffffffff8107e11b>] __do_softirq+0x178/0x39f [<ffffffff8107e52e>] irq_exit+0x41/0x95 [<ffffffff81a4f215>] smp_apic_timer_interrupt+0x34/0x40 [<ffffffff81a4d5cd>] apic_timer_interrupt+0x6d/0x80 <EOI> [<ffffffff8100b968>] ? default_idle+0x21/0x32 [<ffffffff8100b966>] ? default_idle+0x1f/0x32 [<ffffffff8100bf19>] arch_cpu_idle+0xf/0x11 [<ffffffff810b0bc7>] default_idle_call+0x1f/0x21 [<ffffffff810b0dce>] cpu_startup_entry+0x1ad/0x273 [<ffffffff8102fe67>] start_secondary+0x135/0x156 dst is freed right before lwtstate_put(), this is not correct... Fixes: 61adedf3e3f1 ("route: move lwtunnel state to dst_entry") Acked-by: Jiri Benc <jbenc@redhat.com> Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com> Signed-off-by: Cong Wang <cwang@twopensource.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-08-25 17:38:53 +00:00
lwtstate_put(dst->lwtstate);
if (dst->flags & DST_METADATA)
metadata_dst_free((struct metadata_dst *)dst);
else
kmem_cache_free(dst->ops->kmem_cachep, dst);
dst = child;
xfrm: take refcnt of dst when creating struct xfrm_dst bundle During the creation of xfrm_dst bundle, always take ref count when allocating the dst. This way, xfrm_bundle_create() will form a linked list of dst with dst->child pointing to a ref counted dst child. And the returned dst pointer is also ref counted. This makes the link from the flow cache to this dst now ref counted properly. As the dst is always ref counted properly, we can safely mark DST_NOGC flag so dst_release() will release dst based on refcnt only. And dst gc is no longer needed and all dst_free() and its related function calls should be replaced with dst_release() or dst_release_immediate(). The special handling logic for dst->child in dst_destroy() can be replaced with a simple dst_release_immediate() call on the child to release the whole list linked by dst->child pointer. Previously used DST_NOHASH flag is not needed anymore as well. The reason that DST_NOHASH is used in the existing code is mainly to prevent the dst inserted in the fib tree to be wrongly destroyed during the deletion of the xfrm_dst bundle. So in the existing code, DST_NOHASH flag is marked in all the dst children except the one which is in the fib tree. However, with this patch series to remove dst gc logic and release dst only based on ref count, it is safe to release all the children from a xfrm_dst bundle as long as the dst children are all ref counted properly which is already the case in the existing code. So, this patch removes the use of DST_NOHASH flag. Signed-off-by: Wei Wang <weiwan@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-17 17:42:38 +00:00
if (dst)
dst_release_immediate(dst);
return NULL;
}
EXPORT_SYMBOL(dst_destroy);
static void dst_destroy_rcu(struct rcu_head *head)
{
struct dst_entry *dst = container_of(head, struct dst_entry, rcu_head);
dst = dst_destroy(dst);
}
/* Operations to mark dst as DEAD and clean up the net device referenced
* by dst:
* 1. put the dst under loopback interface and discard all tx/rx packets
* on this route.
* 2. release the net_device
* This function should be called when removing routes from the fib tree
* in preparation for a NETDEV_DOWN/NETDEV_UNREGISTER event and also to
* make the next dst_ops->check() fail.
*/
void dst_dev_put(struct dst_entry *dst)
{
struct net_device *dev = dst->dev;
dst->obsolete = DST_OBSOLETE_DEAD;
if (dst->ops->ifdown)
dst->ops->ifdown(dst, dev, true);
dst->input = dst_discard;
dst->output = dst_discard_out;
dst->dev = dev_net(dst->dev)->loopback_dev;
dev_hold(dst->dev);
dev_put(dev);
}
EXPORT_SYMBOL(dst_dev_put);
void dst_release(struct dst_entry *dst)
{
if (dst) {
int newrefcnt;
net: speedup dst_release() During tbench/oprofile sessions, I found that dst_release() was in third position. CPU: Core 2, speed 2999.68 MHz (estimated) Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count 100000 samples % symbol name 483726 9.0185 __copy_user_zeroing_intel 191466 3.5697 __copy_user_intel 185475 3.4580 dst_release 175114 3.2648 ip_queue_xmit 153447 2.8608 tcp_sendmsg 108775 2.0280 tcp_recvmsg 102659 1.9140 sysenter_past_esp 101450 1.8914 tcp_current_mss 95067 1.7724 __copy_from_user_ll 86531 1.6133 tcp_transmit_skb Of course, all CPUS fight on the dst_entry associated with 127.0.0.1 Instead of first checking the refcount value, then decrement it, we use atomic_dec_return() to help CPU to make the right memory transaction (ie getting the cache line in exclusive mode) dst_release() is now at the fifth position, and tbench a litle bit faster ;) CPU: Core 2, speed 3000.1 MHz (estimated) Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count 100000 samples % symbol name 647107 8.8072 __copy_user_zeroing_intel 258840 3.5229 ip_queue_xmit 258302 3.5155 __copy_user_intel 209629 2.8531 tcp_sendmsg 165632 2.2543 dst_release 149232 2.0311 tcp_current_mss 147821 2.0119 tcp_recvmsg 137893 1.8767 sysenter_past_esp 127473 1.7349 __copy_from_user_ll 121308 1.6510 ip_finish_output 118510 1.6129 tcp_transmit_skb 109295 1.4875 tcp_v4_rcv Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-11-14 08:53:54 +00:00
newrefcnt = atomic_dec_return(&dst->__refcnt);
if (unlikely(newrefcnt < 0))
net_warn_ratelimited("%s: dst:%p refcnt:%d\n",
__func__, dst, newrefcnt);
if (!newrefcnt)
call_rcu(&dst->rcu_head, dst_destroy_rcu);
}
}
EXPORT_SYMBOL(dst_release);
net: introduce DST_NOGC in dst_release() to destroy dst based on refcnt The current mechanism of freeing dst is a bit complicated. dst has its ref count and when user grabs the reference to the dst, the ref count is properly taken in most cases except in IPv4/IPv6/decnet/xfrm routing code due to some historic reasons. If the reference to dst is always taken properly, we should be able to simplify the logic in dst_release() to destroy dst when dst->__refcnt drops from 1 to 0. And this should be the only condition to determine if we can call dst_destroy(). And as dst is always ref counted, there is no need for a dst garbage list to hold the dst entries that already get removed by the routing code but are still held by other users. And the task to periodically check the list to free dst if ref count become 0 is also not needed anymore. This patch introduces a temporary flag DST_NOGC(no garbage collector). If it is set in the dst, dst_release() will call dst_destroy() when dst->__refcnt drops to 0. dst_hold_safe() will also check for this flag and do atomic_inc_not_zero() similar as DST_NOCACHE to avoid double free issue. This temporary flag is mainly used so that we can make the transition component by component without breaking other parts. This flag will be removed after all components are properly transitioned. This patch also introduces a new function dst_release_immediate() which destroys dst without waiting on the rcu when refcnt drops to 0. It will be used in later patches. Follow-up patches will correct all the places to properly take ref count on dst and mark DST_NOGC. dst_release() or dst_release_immediate() will be used to release the dst instead of dst_free() and its related functions. And final clean-up patch will remove the DST_NOGC flag. Signed-off-by: Wei Wang <weiwan@google.com> Acked-by: Martin KaFai Lau <kafai@fb.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-17 17:42:27 +00:00
void dst_release_immediate(struct dst_entry *dst)
{
if (dst) {
int newrefcnt;
newrefcnt = atomic_dec_return(&dst->__refcnt);
if (unlikely(newrefcnt < 0))
net_warn_ratelimited("%s: dst:%p refcnt:%d\n",
__func__, dst, newrefcnt);
if (!newrefcnt)
dst_destroy(dst);
}
}
EXPORT_SYMBOL(dst_release_immediate);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
{
struct dst_metrics *p = kmalloc(sizeof(*p), GFP_ATOMIC);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
if (p) {
struct dst_metrics *old_p = (struct dst_metrics *)__DST_METRICS_PTR(old);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
unsigned long prev, new;
atomic_set(&p->refcnt, 1);
memcpy(p->metrics, old_p->metrics, sizeof(p->metrics));
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
new = (unsigned long) p;
prev = cmpxchg(&dst->_metrics, old, new);
if (prev != old) {
kfree(p);
p = (struct dst_metrics *)__DST_METRICS_PTR(prev);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
if (prev & DST_METRICS_READ_ONLY)
p = NULL;
} else if (prev & DST_METRICS_REFCOUNTED) {
if (atomic_dec_and_test(&old_p->refcnt))
kfree(old_p);
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
}
}
BUILD_BUG_ON(offsetof(struct dst_metrics, metrics) != 0);
return (u32 *)p;
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
}
EXPORT_SYMBOL(dst_cow_metrics_generic);
/* Caller asserts that dst_metrics_read_only(dst) is false. */
void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
{
unsigned long prev, new;
new = ((unsigned long) &dst_default_metrics) | DST_METRICS_READ_ONLY;
net: Implement read-only protection and COW'ing of metrics. Routing metrics are now copy-on-write. Initially a route entry points it's metrics at a read-only location. If a routing table entry exists, it will point there. Else it will point at the all zero metric place-holder called 'dst_default_metrics'. The writeability state of the metrics is stored in the low bits of the metrics pointer, we have two bits left to spare if we want to store more states. For the initial implementation, COW is implemented simply via kmalloc. However future enhancements will change this to place the writable metrics somewhere else, in order to increase sharing. Very likely this "somewhere else" will be the inetpeer cache. Note also that this means that metrics updates may transiently fail if we cannot COW the metrics successfully. But even by itself, this patch should decrease memory usage and increase cache locality especially for routing workloads. In those cases the read-only metric copies stay in place and never get written to. TCP workloads where metrics get updated, and those rare cases where PMTU triggers occur, will take a very slight performance hit. But that hit will be alleviated when the long-term writable metrics move to a more sharable location. Since the metrics storage went from a u32 array of RTAX_MAX entries to what is essentially a pointer, some retooling of the dst_entry layout was necessary. Most importantly, we need to preserve the alignment of the reference count so that it doesn't share cache lines with the read-mostly state, as per Eric Dumazet's alignment assertion checks. The only non-trivial bit here is the move of the 'flags' member into the writeable cacheline. This is OK since we are always accessing the flags around the same moment when we made a modification to the reference count. Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-27 04:51:05 +00:00
prev = cmpxchg(&dst->_metrics, old, new);
if (prev == old)
kfree(__DST_METRICS_PTR(old));
}
EXPORT_SYMBOL(__dst_destroy_metrics_generic);
static struct dst_ops md_dst_ops = {
.family = AF_UNSPEC,
};
static int dst_md_discard_out(struct net *net, struct sock *sk, struct sk_buff *skb)
{
WARN_ONCE(1, "Attempting to call output on metadata dst\n");
kfree_skb(skb);
return 0;
}
static int dst_md_discard(struct sk_buff *skb)
{
WARN_ONCE(1, "Attempting to call input on metadata dst\n");
kfree_skb(skb);
return 0;
}
net: store port/representator id in metadata_dst Switches and modern SR-IOV enabled NICs may multiplex traffic from Port representators and control messages over single set of hardware queues. Control messages and muxed traffic may need ordered delivery. Those requirements make it hard to comfortably use TC infrastructure today unless we have a way of attaching metadata to skbs at the upper device. Because single set of queues is used for many netdevs stopping TC/sched queues of all of them reliably is impossible and lower device has to retreat to returning NETDEV_TX_BUSY and usually has to take extra locks on the fastpath. This patch attempts to enable port/representative devs to attach metadata to skbs which carry port id. This way representatives can be queueless and all queuing can be performed at the lower netdev in the usual way. Traffic arriving on the port/representative interfaces will be have metadata attached and will subsequently be queued to the lower device for transmission. The lower device should recognize the metadata and translate it to HW specific format which is most likely either a special header inserted before the network headers or descriptor/metadata fields. Metadata is associated with the lower device by storing the netdev pointer along with port id so that if TC decides to redirect or mirror the new netdev will not try to interpret it. This is mostly for SR-IOV devices since switches don't have lower netdevs today. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-23 20:11:58 +00:00
static void __metadata_dst_init(struct metadata_dst *md_dst,
enum metadata_type type, u8 optslen)
{
struct dst_entry *dst;
dst = &md_dst->dst;
dst_init(dst, &md_dst_ops, NULL, 1, DST_OBSOLETE_NONE,
DST_METADATA | DST_NOCOUNT);
dst->input = dst_md_discard;
dst->output = dst_md_discard_out;
memset(dst + 1, 0, sizeof(*md_dst) + optslen - sizeof(*dst));
net: store port/representator id in metadata_dst Switches and modern SR-IOV enabled NICs may multiplex traffic from Port representators and control messages over single set of hardware queues. Control messages and muxed traffic may need ordered delivery. Those requirements make it hard to comfortably use TC infrastructure today unless we have a way of attaching metadata to skbs at the upper device. Because single set of queues is used for many netdevs stopping TC/sched queues of all of them reliably is impossible and lower device has to retreat to returning NETDEV_TX_BUSY and usually has to take extra locks on the fastpath. This patch attempts to enable port/representative devs to attach metadata to skbs which carry port id. This way representatives can be queueless and all queuing can be performed at the lower netdev in the usual way. Traffic arriving on the port/representative interfaces will be have metadata attached and will subsequently be queued to the lower device for transmission. The lower device should recognize the metadata and translate it to HW specific format which is most likely either a special header inserted before the network headers or descriptor/metadata fields. Metadata is associated with the lower device by storing the netdev pointer along with port id so that if TC decides to redirect or mirror the new netdev will not try to interpret it. This is mostly for SR-IOV devices since switches don't have lower netdevs today. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-23 20:11:58 +00:00
md_dst->type = type;
}
net: store port/representator id in metadata_dst Switches and modern SR-IOV enabled NICs may multiplex traffic from Port representators and control messages over single set of hardware queues. Control messages and muxed traffic may need ordered delivery. Those requirements make it hard to comfortably use TC infrastructure today unless we have a way of attaching metadata to skbs at the upper device. Because single set of queues is used for many netdevs stopping TC/sched queues of all of them reliably is impossible and lower device has to retreat to returning NETDEV_TX_BUSY and usually has to take extra locks on the fastpath. This patch attempts to enable port/representative devs to attach metadata to skbs which carry port id. This way representatives can be queueless and all queuing can be performed at the lower netdev in the usual way. Traffic arriving on the port/representative interfaces will be have metadata attached and will subsequently be queued to the lower device for transmission. The lower device should recognize the metadata and translate it to HW specific format which is most likely either a special header inserted before the network headers or descriptor/metadata fields. Metadata is associated with the lower device by storing the netdev pointer along with port id so that if TC decides to redirect or mirror the new netdev will not try to interpret it. This is mostly for SR-IOV devices since switches don't have lower netdevs today. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-23 20:11:58 +00:00
struct metadata_dst *metadata_dst_alloc(u8 optslen, enum metadata_type type,
gfp_t flags)
{
struct metadata_dst *md_dst;
md_dst = kmalloc(sizeof(*md_dst) + optslen, flags);
if (!md_dst)
return NULL;
net: store port/representator id in metadata_dst Switches and modern SR-IOV enabled NICs may multiplex traffic from Port representators and control messages over single set of hardware queues. Control messages and muxed traffic may need ordered delivery. Those requirements make it hard to comfortably use TC infrastructure today unless we have a way of attaching metadata to skbs at the upper device. Because single set of queues is used for many netdevs stopping TC/sched queues of all of them reliably is impossible and lower device has to retreat to returning NETDEV_TX_BUSY and usually has to take extra locks on the fastpath. This patch attempts to enable port/representative devs to attach metadata to skbs which carry port id. This way representatives can be queueless and all queuing can be performed at the lower netdev in the usual way. Traffic arriving on the port/representative interfaces will be have metadata attached and will subsequently be queued to the lower device for transmission. The lower device should recognize the metadata and translate it to HW specific format which is most likely either a special header inserted before the network headers or descriptor/metadata fields. Metadata is associated with the lower device by storing the netdev pointer along with port id so that if TC decides to redirect or mirror the new netdev will not try to interpret it. This is mostly for SR-IOV devices since switches don't have lower netdevs today. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-23 20:11:58 +00:00
__metadata_dst_init(md_dst, type, optslen);
return md_dst;
}
EXPORT_SYMBOL_GPL(metadata_dst_alloc);
void metadata_dst_free(struct metadata_dst *md_dst)
{
#ifdef CONFIG_DST_CACHE
dst_cache_destroy(&md_dst->u.tun_info.dst_cache);
#endif
kfree(md_dst);
}
net: store port/representator id in metadata_dst Switches and modern SR-IOV enabled NICs may multiplex traffic from Port representators and control messages over single set of hardware queues. Control messages and muxed traffic may need ordered delivery. Those requirements make it hard to comfortably use TC infrastructure today unless we have a way of attaching metadata to skbs at the upper device. Because single set of queues is used for many netdevs stopping TC/sched queues of all of them reliably is impossible and lower device has to retreat to returning NETDEV_TX_BUSY and usually has to take extra locks on the fastpath. This patch attempts to enable port/representative devs to attach metadata to skbs which carry port id. This way representatives can be queueless and all queuing can be performed at the lower netdev in the usual way. Traffic arriving on the port/representative interfaces will be have metadata attached and will subsequently be queued to the lower device for transmission. The lower device should recognize the metadata and translate it to HW specific format which is most likely either a special header inserted before the network headers or descriptor/metadata fields. Metadata is associated with the lower device by storing the netdev pointer along with port id so that if TC decides to redirect or mirror the new netdev will not try to interpret it. This is mostly for SR-IOV devices since switches don't have lower netdevs today. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-23 20:11:58 +00:00
struct metadata_dst __percpu *
metadata_dst_alloc_percpu(u8 optslen, enum metadata_type type, gfp_t flags)
{
int cpu;
struct metadata_dst __percpu *md_dst;
md_dst = __alloc_percpu_gfp(sizeof(struct metadata_dst) + optslen,
__alignof__(struct metadata_dst), flags);
if (!md_dst)
return NULL;
for_each_possible_cpu(cpu)
net: store port/representator id in metadata_dst Switches and modern SR-IOV enabled NICs may multiplex traffic from Port representators and control messages over single set of hardware queues. Control messages and muxed traffic may need ordered delivery. Those requirements make it hard to comfortably use TC infrastructure today unless we have a way of attaching metadata to skbs at the upper device. Because single set of queues is used for many netdevs stopping TC/sched queues of all of them reliably is impossible and lower device has to retreat to returning NETDEV_TX_BUSY and usually has to take extra locks on the fastpath. This patch attempts to enable port/representative devs to attach metadata to skbs which carry port id. This way representatives can be queueless and all queuing can be performed at the lower netdev in the usual way. Traffic arriving on the port/representative interfaces will be have metadata attached and will subsequently be queued to the lower device for transmission. The lower device should recognize the metadata and translate it to HW specific format which is most likely either a special header inserted before the network headers or descriptor/metadata fields. Metadata is associated with the lower device by storing the netdev pointer along with port id so that if TC decides to redirect or mirror the new netdev will not try to interpret it. This is mostly for SR-IOV devices since switches don't have lower netdevs today. Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com> Signed-off-by: Simon Horman <horms@verge.net.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-06-23 20:11:58 +00:00
__metadata_dst_init(per_cpu_ptr(md_dst, cpu), type, optslen);
return md_dst;
}
EXPORT_SYMBOL_GPL(metadata_dst_alloc_percpu);