Re: nfs: LOCKDEP warning with 2.6.31-rc6

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2009-08-20 at 12:18 +0200, Takashi Iwai wrote:
> Hi,
> 
> my machine got a lockdep warning regarding NFS client on 2.6.31-rc6.
> Here is the log:
> 
> =======================================================
> [ INFO: possible circular locking dependency detected ]
> 2.6.31-rc6-test #12
> -------------------------------------------------------
> soffice.bin/31490 is trying to acquire lock:
>  (&sb->s_type->i_mutex_key#13){+.+.+.}, at: [<f85ea8bb>] nfs_revalidate_mapping+0x72/0xbe [nfs]
> 
> but task is already holding lock:
>  (&mm->mmap_sem){++++++}, at: [<c0106c5b>] sys_mmap2+0x6d/0xb4
> 
> which lock already depends on the new lock.
> 
> 
> the existing dependency chain (in reverse order) is:
> 
> -> #1 (&mm->mmap_sem){++++++}:
>        [<c0161b06>] __lock_acquire+0x100e/0x130d
>        [<c0161ebc>] lock_acquire+0xb7/0xeb
>        [<c01a2a15>] might_fault+0x69/0x9a
>        [<c026d399>] copy_to_user+0x3c/0x127
>        [<c01d4427>] filldir64+0xc3/0x108
>        [<f85e6270>] nfs_do_filldir+0x383/0x4ad [nfs]
>        [<f85e6b4e>] nfs_readdir+0x7b4/0x830 [nfs]
>        [<c01d469e>] vfs_readdir+0x76/0xb3
>        [<c01d474e>] sys_getdents64+0x73/0xc3
>        [<c0102df3>] sysenter_do_call+0x12/0x38
>        [<ffffffff>] 0xffffffff
> 
> -> #0 (&sb->s_type->i_mutex_key#13){+.+.+.}:
>        [<c0161856>] __lock_acquire+0xd5e/0x130d
>        [<c0161ebc>] lock_acquire+0xb7/0xeb
>        [<c03af059>] mutex_lock_nested+0x43/0x272
>        [<f85ea8bb>] nfs_revalidate_mapping+0x72/0xbe [nfs]
>        [<f85e8128>] nfs_file_mmap+0x5e/0x77 [nfs]
>        [<c01ab5f7>] mmap_region+0x263/0x40f
>        [<c01aba07>] do_mmap_pgoff+0x264/0x2c7
>        [<c0106c6f>] sys_mmap2+0x81/0xb4
>        [<c0102df3>] sysenter_do_call+0x12/0x38
>        [<ffffffff>] 0xffffffff
> 
> other info that might help us debug this:
> 
> 1 lock held by soffice.bin/31490:
>  #0:  (&mm->mmap_sem){++++++}, at: [<c0106c5b>] sys_mmap2+0x6d/0xb4
> 
> stack backtrace:
> Pid: 31490, comm: soffice.bin Not tainted 2.6.31-rc6-test #12
> Call Trace:
>  [<c03ad742>] ? printk+0x1d/0x33
>  [<c01606d1>] print_circular_bug_tail+0xaf/0xcb
>  [<c0161856>] __lock_acquire+0xd5e/0x130d
>  [<f85ea8bb>] ? nfs_revalidate_mapping+0x72/0xbe [nfs]
>  [<c0161ebc>] lock_acquire+0xb7/0xeb
>  [<f85ea8bb>] ? nfs_revalidate_mapping+0x72/0xbe [nfs]
>  [<f85ea8bb>] ? nfs_revalidate_mapping+0x72/0xbe [nfs]
>  [<c03af059>] mutex_lock_nested+0x43/0x272
>  [<f85ea8bb>] ? nfs_revalidate_mapping+0x72/0xbe [nfs]
>  [<f8609339>] ? nfs_have_delegation+0x68/0x82 [nfs]
>  [<f85ea8bb>] nfs_revalidate_mapping+0x72/0xbe [nfs]
>  [<f85e8128>] nfs_file_mmap+0x5e/0x77 [nfs]
>  [<c01ab5f7>] mmap_region+0x263/0x40f
>  [<c01aba07>] do_mmap_pgoff+0x264/0x2c7
>  [<c0106c6f>] sys_mmap2+0x81/0xb4
>  [<c0102df3>] sysenter_do_call+0x12/0x38
> 
> =======================================================
> 
> Indeed this looks like a circular lock dependency since copy_*_user()
> invokes mm->mmap_sem mutex lock internally.  (I hit a similar bug in
> the ALSA core ago, thus I know it :)
> 
> The inode lock in vfs_readdir is killable, so it's not critical,
> though...

Right, I did a number of patches to split up ->mmap() a while ago (ok, a
long while ago).

  http://www.mail-archive.com/linux-arch@vger.kernel.org/msg04493.html

Unless someone sees another way out, we should reconsider these, as this
warning seems to pop up more frequently.