Re: [BUG] perf_event: circular lock dependency

[Peter Zijlstra]

On Thu, 2010-01-28 at 10:19 +0100, Stephane Eranian wrote:

> On Intel Core, one of my test programs generate this kind of
> warning when it unmaps the sampling buffer after it has closed
> the events fds.

> [ 1729.441066] the existing dependency chain (in reverse order) is:
> [ 1729.441092]
> [ 1729.441093] -> #1 (&mm->mmap_sem){++++++}:
> [ 1729.441123]        [<ffffffff81077f97>] validate_chain+0xc17/0x1360
> [ 1729.441151]        [<ffffffff81078a53>] __lock_acquire+0x373/0xb30
> [ 1729.441170]        [<ffffffff810792ac>] lock_acquire+0x9c/0x100
> [ 1729.441189]        [<ffffffff810e74a4>] might_fault+0x84/0xb0
> [ 1729.441207]        [<ffffffff810c3605>] perf_read+0x135/0x2d0
> [ 1729.441225]        [<ffffffff8110c604>] vfs_read+0xc4/0x180
> [ 1729.441245]        [<ffffffff8110ca10>] sys_read+0x50/0x90
> [ 1729.441263]        [<ffffffff81002ceb>] system_call_fastpath+0x16/0x1b
> [ 1729.441284]
> [ 1729.441284] -> #0 (&ctx->mutex){+.+...}:
> [ 1729.441313]        [<ffffffff810786cd>] validate_chain+0x134d/0x1360
> [ 1729.441332]        [<ffffffff81078a53>] __lock_acquire+0x373/0xb30
> [ 1729.441351]        [<ffffffff810792ac>] lock_acquire+0x9c/0x100
> [ 1729.441369]        [<ffffffff81442e59>] mutex_lock_nested+0x69/0x340
> [ 1729.441389]        [<ffffffff810c2ebd>] perf_event_release_kernel+0x2d/0xe0
> [ 1729.441409]        [<ffffffff810c2f8b>] perf_release+0x1b/0x20
> [ 1729.441426]        [<ffffffff8110d051>] __fput+0x101/0x230
> [ 1729.441444]        [<ffffffff8110d457>] fput+0x17/0x20
> [ 1729.441462]        [<ffffffff810e98d1>] remove_vma+0x51/0x90
> [ 1729.441480]        [<ffffffff810ea708>] do_munmap+0x2e8/0x340
> [ 1729.441498]        [<ffffffff810ebac0>] sys_munmap+0x50/0x80
> [ 1729.441516]        [<ffffffff81002ceb>] system_call_fastpath+0x16/0x1b
> [ 1729.441535]

Crap, the thing is right.. you've been using group reads, which require
holding the ctx->mutex to ensure the group doesn't change while you're
reading it, leading to this inversion thing...

Not sure where to break this loop though, the hacky way is pushing all
of perf_event_release_kernel() into a work, but that's yucky.. Let me
ponder this a bit more.

--
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/