Re: NFS: possible circular locking i_mutex <> mmap_sem

[Wu Fengguang <fengguang.wu@intel.com>]

On Fri, Jun 19, 2009 at 12:48:08AM +0800, Peter Zijlstra wrote:
> On Thu, 2009-06-18 at 11:02 +0800, Wu Fengguang wrote:
> 
> > > > [ 2638.515865] =======================================================
> > > > [ 2638.519743] [ INFO: possible circular locking dependency detected ]
> > > > [ 2638.519743] 2.6.30-rc8-mm1 #307
> > > > [ 2638.519743] -------------------------------------------------------
> > > > [ 2638.519743] firefox-bin/3399 is trying to acquire lock:
> > > > [ 2638.519743]  (&mm->mmap_sem){++++++}, at: [<ffffffff81548471>] do_page_fault+0x301/0x330
> > > > [ 2638.519743]
> > > > [ 2638.519743] but task is already holding lock:
> > > > [ 2638.519743]  (&sb->s_type->i_mutex_key#6){+.+.+.}, at: [<ffffffff810c2bd2>] generic_file_aio_write+0x52/0xd0
> > > > [ 2638.519743]
> > > > [ 2638.519743] which lock already depends on the new lock.
> > > > [ 2638.519743]
> > > > [ 2638.519743]
> > > > [ 2638.519743] the existing dependency chain (in reverse order) is:
> > > > [ 2638.519743]
> > > > [ 2638.519743] -> #1 (&sb->s_type->i_mutex_key#6){+.+.+.}:
> > > > [ 2638.519743]        [<ffffffff8107c066>] __lock_acquire+0x12b6/0x1b40
> > > > [ 2638.519743]        [<ffffffff8107c9d1>] lock_acquire+0xe1/0x120
> > > > [ 2638.519743]        [<ffffffff8154328e>] mutex_lock_nested+0x5e/0x390
> > > > [ 2638.519743]        [<ffffffff811be15c>] nfs_revalidate_mapping+0xac/0x110    ==> takes i_mutex in nfs_invalidate_mapping()
> > > > [ 2638.519743]        [<ffffffff811bba25>] nfs_file_mmap+0x55/0x80
> > > > [ 2638.519743]        [<ffffffff810e2bb7>] mmap_region+0x427/0x600
> > > > [ 2638.519743]        [<ffffffff810e305e>] do_mmap_pgoff+0x2ce/0x3f0
> > > > [ 2638.519743]        [<ffffffff81010c26>] sys_mmap+0x106/0x130                 ==> takes mmap_sem
> > > > [ 2638.519743]        [<ffffffff8100bf42>] system_call_fastpath+0x16/0x1b
> > > > [ 2638.519743]        [<ffffffffffffffff>] 0xffffffffffffffff
> > > 
> > > This is the faulty code path.  mmap_sem is supposed to nest inside i_mutex.
> > 
> > This could be a long stand bug: the nfs_revalidate_mapping() call was
> > there in the very beginning of git history.
> 
> http://programming.kicks-ass.net/kernel-patches/mmap-vs-nfs/
> 

I wonder why such an old patch was not merged back in 2007.
Nick and Trond seems to be OK with it. So is me. But now there
are so many merge conflicts because it modified so many files..

I can do some stress tests for it.

Thanks,
Fengguang