Re: perfcounters: lockdep warning

From: Ingo Molnar
Date: Mon Jun 15 2009 - 10:20:20 EST



* Simon Holm Thøgersen <odie@xxxxxxxxx> wrote:

> man, 15 06 2009 kl. 09:54 +0200, skrev Peter Zijlstra:
> > On Sat, 2009-06-13 at 21:39 +0200, Simon Holm Thøgersen wrote:
> > > Just tried kicking the tires of performance counters and perf and got
> > > the following warning that doesn't look like have been reported already.
> > >
> > > [ 7765.594591] [ INFO: possible circular locking dependency detected ]
> > > [ 7765.594602] 2.6.30debug-03217-gf3ad116 #47
> > > [ 7765.594609] -------------------------------------------------------
> > > [ 7765.594619] perf/14176 is trying to acquire lock:
> > > [ 7765.594628] (&ctx->mutex){+.+.+.}, at: [<c016f366>] perf_counter_for_each_sibling+0x50/0x7e
> > > [ 7765.594660]
> > > [ 7765.594663] but task is already holding lock:
> > > [ 7765.594672] (&counter->child_mutex){+.+...}, at: [<c0172180>] perf_ioctl+0x184/0x1de
> > > [ 7765.594696]
> > > [ 7765.594699] which lock already depends on the new lock.
> > > [ 7765.594703]
> > > [ 7765.594711]
> > > [ 7765.594714] the existing dependency chain (in reverse order) is:
> > > [ 7765.594723]
> > > [ 7765.594726] -> #1 (&counter->child_mutex){+.+...}:
> > > [ 7765.594744] [<c0146f79>] __lock_acquire+0x9a5/0xb11
> > > [ 7765.594765] [<c014719e>] lock_acquire+0xb9/0xdb
> > > [ 7765.594779] [<c0340664>] __mutex_lock_common+0x42/0x3c8
> > > [ 7765.594798] [<c0340a82>] mutex_lock_nested+0x2e/0x36
> > > [ 7765.594814] [<c016f8ad>] inherit_counter+0xdb/0x112
> > > [ 7765.594830] [<c01705bd>] perf_counter_init_task+0x15b/0x23f
> > > [ 7765.594847] [<c0124338>] copy_process+0x4fb/0xfc8
> > > [ 7765.594865] [<c0124f1c>] do_fork+0x117/0x2b4
> > > [ 7765.594881] [<c0101f4f>] sys_clone+0x29/0x30
> > > [ 7765.594897] [<c01032e8>] sysenter_do_call+0x12/0x3c
> > > [ 7765.594913] [<ffffffff>] 0xffffffff
> > > [ 7765.594967]
> > > [ 7765.594970] -> #0 (&ctx->mutex){+.+.+.}:
> > > [ 7765.594987] [<c0146e8c>] __lock_acquire+0x8b8/0xb11
> > > [ 7765.595004] [<c014719e>] lock_acquire+0xb9/0xdb
> > > [ 7765.595018] [<c0340664>] __mutex_lock_common+0x42/0x3c8
> > > [ 7765.595035] [<c0340a82>] mutex_lock_nested+0x2e/0x36
> > > [ 7765.595050] [<c016f366>] perf_counter_for_each_sibling+0x50/0x7e
> > > [ 7765.595067] [<c0172189>] perf_ioctl+0x18d/0x1de
> > > [ 7765.595083] [<c019fa6e>] vfs_ioctl+0x27/0x6e
> > > [ 7765.595100] [<c019ff4f>] do_vfs_ioctl+0x45a/0x48c
> > > [ 7765.595116] [<c019ffb2>] sys_ioctl+0x31/0x4a
> > > [ 7765.595132] [<c01032e8>] sysenter_do_call+0x12/0x3c
> > > [ 7765.595147] [<ffffffff>] 0xffffffff
> > > [ 7765.595163]
> > > [ 7765.595166] other info that might help us debug this:
> > > [ 7765.595170]
> > > [ 7765.595180] 1 lock held by perf/14176:
> > > [ 7765.595188] #0: (&counter->child_mutex){+.+...}, at: [<c0172180>] perf_ioctl+0x184/0x1de
> > > [ 7765.595215]
> > > [ 7765.595218] stack backtrace:
> > > [ 7765.595230] Pid: 14176, comm: perf Not tainted 2.6.30debug-03217-gf3ad116 #47
> > > [ 7765.595240] Call Trace:
> > > [ 7765.595254] [<c033f7f9>] ? printk+0x14/0x16
> > > [ 7765.595271] [<c0146310>] print_circular_bug_tail+0x5c/0x67
> > > [ 7765.595289] [<c0146e8c>] __lock_acquire+0x8b8/0xb11
> > > [ 7765.595306] [<c014719e>] lock_acquire+0xb9/0xdb
> > > [ 7765.595322] [<c016f366>] ? perf_counter_for_each_sibling+0x50/0x7e
> > > [ 7765.595338] [<c0340664>] __mutex_lock_common+0x42/0x3c8
> > > [ 7765.595354] [<c016f366>] ? perf_counter_for_each_sibling+0x50/0x7e
> > > [ 7765.595371] [<c034099b>] ? __mutex_lock_common+0x379/0x3c8
> > > [ 7765.595387] [<c0340a82>] mutex_lock_nested+0x2e/0x36
> > > [ 7765.595402] [<c016f366>] ? perf_counter_for_each_sibling+0x50/0x7e
> > > [ 7765.595419] [<c016f366>] perf_counter_for_each_sibling+0x50/0x7e
> > > [ 7765.595434] [<c0170f78>] ? perf_counter_enable+0x0/0xad
> > > [ 7765.595449] [<c0171ffc>] ? perf_ioctl+0x0/0x1de
> > > [ 7765.595464] [<c0170f78>] ? perf_counter_enable+0x0/0xad
> > > [ 7765.595479] [<c0172189>] perf_ioctl+0x18d/0x1de
> > > [ 7765.595494] [<c0171ffc>] ? perf_ioctl+0x0/0x1de
> > > [ 7765.595509] [<c019fa6e>] vfs_ioctl+0x27/0x6e
> > > [ 7765.595525] [<c019ff4f>] do_vfs_ioctl+0x45a/0x48c
> > > [ 7765.595540] [<c0107003>] ? native_sched_clock+0x45/0x5e
> > > [ 7765.595556] [<c0144a37>] ? put_lock_stats+0x1e/0x29
> > > [ 7765.595572] [<c0144af8>] ? lock_release_holdtime+0xb6/0xbb
> > > [ 7765.595589] [<c0105e7b>] ? sys_mmap2+0x67/0x7f
> > > [ 7765.595604] [<c0103321>] ? sysenter_exit+0xf/0x1a
> > > [ 7765.595620] [<c019ffb2>] sys_ioctl+0x31/0x4a
> > > [ 7765.595635] [<c01032e8>] sysenter_do_call+0x12/0x3c
> >
> > Right, good catch ;-)
>
> Full credit goes to lockdep, I merely tried perf record <cmd> :-)
>
> > Does this fix it?
> >
> > Signed-off-by: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
> > ---
> > diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
> > index e914daf..35fa30b 100644
> > --- a/kernel/perf_counter.c
> > +++ b/kernel/perf_counter.c
> > @@ -1658,14 +1658,18 @@ static void perf_counter_for_each_child(struct perf_counter *counter,
> > static void perf_counter_for_each(struct perf_counter *counter,
> > void (*func)(struct perf_counter *))
> > {
> > - struct perf_counter *child;
> > + struct perf_counter_context *ctx = counter->ctx;
> > + struct perf_counter *sibling;
> >
> > - WARN_ON_ONCE(counter->ctx->parent_ctx);
> > - mutex_lock(&counter->child_mutex);
> > - perf_counter_for_each_sibling(counter, func);
> > - list_for_each_entry(child, &counter->child_list, child_list)
> > - perf_counter_for_each_sibling(child, func);
> > - mutex_unlock(&counter->child_mutex);
> > + WARN_ON_ONCE(ctx->parent_ctx);
> > + mutex_lock(&ctx->mutex);
> > + counter = counter->group_leader;
> > +
> > + perf_counter_for_each_child(counter, func)
>
> missing ; goes here ^
>
> > + func(counter);
> > + list_for_each_entry(sibling, &counter->sibling_list, list_entry)
> > + perf_counter_for_each_child(counter, func)
>
> and here ^
>
> > + mutex_unlock(&ctx->mutex);
> > }
> >
> > static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
>
> It also introduces the following warning:
>
> kernel/perf_counter.c:1625: warning: 'perf_counter_for_each_sibling'
> defined but not used
>
> but otherwise it is good and you can add
>
> Reported-by: Simon Holm Thøgersen <odie@xxxxxxxxx>
> Tested-by: Simon Holm Thøgersen <odie@xxxxxxxxx>

Thanks guys, applied.

Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/