Skip to contentSkip to conversation

perf_counter: Fix counter inheritance

tip:perfcounters/core Ingo Molnar 2 participants 2 messages May 17, 2009 – May 19, 2009 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 51 lines · tip-bot for Peter Zijlstra
Commit-ID:  856d56b9e5de650a64a6c41c17aaed702b55d578
Gitweb:     http://git.kernel.org/tip/856d56b9e5de650a64a6c41c17aaed702b55d578
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Fri, 15 May 2009 20:45:59 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sun, 17 May 2009 07:52:24 +0200

perf_counter: Fix counter inheritance

Srivatsa Vaddagiri reported that a Java workload triggers this warning in kernel/exit.c:

   WARN_ON_ONCE(!list_empty(&tsk->perf_counter_ctx.counter_list));

Add the inherited counter propagation on self-detach, this could cause counter leaks and incomplete stats in threaded code like the below:

  #include <pthread.h>
  #include <unistd.h>

  void *thread(void *arg)
  {
          sleep(5);
          return NULL;
  }

  void main(void)
  {
          pthread_t thr;
          pthread_create(&thr, NULL, thread, NULL);
  }
Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Marcelo Tosatti <[email protected]>
---
 kernel/exit.c |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
kernel/exit.c +0 -0
index 4741376..16d74f1 100644

kernel/exit.c static void __exit_signal(struct task_struct *tsk) · line 135

Quoted 8 lines
kernel/exit.c +5 -0
@@ -128,6 +128,12 @@ static void __exit_signal(struct task_struct *tsk)
 		sig = NULL; /* Marker for below. */
 	}
 
+	/*
+	 * Flush inherited counters to the parent - before the parent
+	 * gets woken up by child-exit notifications.
+	 */
+	perf_counter_exit_task(tsk);

Causes:

[  447.882292] BUG: sleeping function called from invalid context at kernel/mutex.c:94
[  447.890094] in_atomic(): 0, irqs_disabled(): 1, pid: 23597, name: hackbench_pth
[  447.897587] Pid: 23597, comm: hackbench_pth Not tainted 2.6.30-rc6-tip #188
[  447.904678] Call Trace:
[  447.907158]  [<ffffffff814cdd0b>] ? mutex_lock+0x15/0x37
[  447.912518]  [<ffffffff8108f1e3>] ? perf_counter_exit_task+0x170/0x1e9
[  447.919134]  [<ffffffff81046182>] ? release_task+0x22c/0x402
[  447.924859]  [<ffffffff8104789d>] ? do_exit+0x655/0x6e7
[  447.930144]  [<ffffffff810479ea>] ? complete_and_exit+0x0/0x16
[  447.936054]  [<ffffffff8100baab>] ? system_call_fastpath+0x16/0x1b

when running:

perf stat ./hackbench_pth 20

release_task() is a deep-atomic context, we cannot acquire a mutex there. I'm not sure we can change that lock to a spinlock straight away.

Ingo

#2 lore raw
Quoted 78 lines · Ingo Molnar
Quoted 51 lines · tip-bot for Peter Zijlstra
Commit-ID:  856d56b9e5de650a64a6c41c17aaed702b55d578
Gitweb:     http://git.kernel.org/tip/856d56b9e5de650a64a6c41c17aaed702b55d578
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Fri, 15 May 2009 20:45:59 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sun, 17 May 2009 07:52:24 +0200

perf_counter: Fix counter inheritance

Srivatsa Vaddagiri reported that a Java workload triggers this warning in kernel/exit.c:

   WARN_ON_ONCE(!list_empty(&tsk->perf_counter_ctx.counter_list));

Add the inherited counter propagation on self-detach, this could cause counter leaks and incomplete stats in threaded code like the below:

  #include <pthread.h>
  #include <unistd.h>

  void *thread(void *arg)
  {
          sleep(5);
          return NULL;
  }

  void main(void)
  {
          pthread_t thr;
          pthread_create(&thr, NULL, thread, NULL);
  }
Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc Marcelo Tosatti <[email protected]>
---
 kernel/exit.c |    6 ++++++
 1 files changed, 6 insertions(+), 0 deletions(-)
kernel/exit.c +0 -0
index 4741376..16d74f1 100644

kernel/exit.c static void __exit_signal(struct task_struct *tsk) · line 135

Quoted 8 lines
kernel/exit.c +5 -0
@@ -128,6 +128,12 @@ static void __exit_signal(struct task_struct *tsk)
 		sig = NULL; /* Marker for below. */
 	}
 
+	/*
+	 * Flush inherited counters to the parent - before the parent
+	 * gets woken up by child-exit notifications.
+	 */
+	perf_counter_exit_task(tsk);

Causes:

[  447.882292] BUG: sleeping function called from invalid context at kernel/mutex.c:94
[  447.890094] in_atomic(): 0, irqs_disabled(): 1, pid: 23597, name: hackbench_pth
[  447.897587] Pid: 23597, comm: hackbench_pth Not tainted 2.6.30-rc6-tip #188
[  447.904678] Call Trace:
[  447.907158]  [<ffffffff814cdd0b>] ? mutex_lock+0x15/0x37
[  447.912518]  [<ffffffff8108f1e3>] ? perf_counter_exit_task+0x170/0x1e9
[  447.919134]  [<ffffffff81046182>] ? release_task+0x22c/0x402
[  447.924859]  [<ffffffff8104789d>] ? do_exit+0x655/0x6e7
[  447.930144]  [<ffffffff810479ea>] ? complete_and_exit+0x0/0x16
[  447.936054]  [<ffffffff8100baab>] ? system_call_fastpath+0x16/0x1b

when running:

Quoted 7 lines

perf stat ./hackbench_pth 20

release_task() is a deep-atomic context, we cannot acquire a mutex there. I'm not sure we can change that lock to a spinlock straight away.

Ingo

Reverting this patch (0203026b58b4299ba7281c0b4b417207c1f05d0e) fixes the following oops for me:

general protection fault: 0000 [#2] PREEMPT SMP DEBUG_PAGEALLOC

last sysfs file:
/sys/devices/system/cpu/cpu7/cache/index2/shared_cpu_map CPU 2
Pid: 4648, comm: bash Tainted: G D
2.6.30-rc6-tip-01582-g1acb813-dirty #6 PowerEdge 1900

RIP: 0010:[<ffffffff8029c492>]  [<ffffffff8029c492>]
list_del_counter+0x46/0xad
RSP: 0018:ffff88022e96fd58  EFLAGS: 00010296
RAX: dead000000200200 RBX: 0000000000000000 RCX: 0000000000000000
RDX: ffff880229065448 RSI: ffff880229065390 RDI: ffff88021acd3bf0
RBP: ffff88022e96fd88 R08: dead000000200200 R09: ffff88022e96fcf8
R10: 0000000000000292 R11: 0000000000000282 R12: ffff88021acd3bf0
R13: ffff880229065390 R14: ffff880229064480 R15: ffff880229065438
FS:  00007f6456a656f0(0000) GS:ffff8800280b8000(0000)

knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 0000003697414948 CR3: 0000000228d76000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Process bash (pid: 4648, threadinfo ffff88022e96e000, task ffff88021c5d8640)

Stack:
 0000000000000287 0000000000000000 ffff88021acd3bf0 ffff880229064480
 ffff880229064480 ffff880229065438 ffff88022e96fdd8 ffffffff8029ddea
 000000000000e930 ffff880229065438 ffff880229065390 0000000000000000
Call Trace:
 [<ffffffff8029ddea>] perf_counter_exit_task+0xbc/0x221
 [<ffffffff80243f2f>] wait_consider_task+0x30d/0x971
 [<ffffffff802446c8>] ? do_wait+0x135/0x403
 [<ffffffff805f30a7>] ? _read_lock+0x1b/0x4b
 [<ffffffff80244728>] do_wait+0x195/0x403
 [<ffffffff80239dda>] ? default_wake_function+0x0/0x14
 [<ffffffff80244a20>] sys_wait4+0x8a/0xa5
 [<ffffffff8020b21b>] system_call_fastpath+0x16/0x1b
Code: 00 49 b8 00 02 20 00 00 00 ad de 48 8b 17 48 8b 47 08 49 89 f5 48

89 42 08 48 89 10 48 8b 47 18 48 8b 57 10 48 89 3f 48 89 7f 08 <48> 89 10 48 89 42 08 48 8b 47 38 4c 89 47 18 48 39 f8 74 03 ff RIP [<ffffffff8029c492>] list_del_counter+0x46/0xad
RSP <ffff88022e96fd58>

---[ end trace 71d51ad4e24097f6 ]---