Re: dio_get_page() lockdep complaints

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 19 Apr 2007 10:01:57 +0200 Jens Axboe <jens.axboe@oracle.com> wrote:

> On Thu, Apr 19 2007, Andrew Morton wrote:
> > On Thu, 19 Apr 2007 09:38:30 +0200 Jens Axboe <jens.axboe@oracle.com> wrote:
> > 
> > > Hi,
> > > 
> > > Doing some testing on CFQ, I ran into this 100% reproducible report:
> > > 
> > > =======================================================
> > > [ INFO: possible circular locking dependency detected ]
> > > 2.6.21-rc7 #5
> > > -------------------------------------------------------
> > > fio/9741 is trying to acquire lock:
> > >  (&mm->mmap_sem){----}, at: [<b018cb34>] dio_get_page+0x54/0x161
> > > 
> > > but task is already holding lock:
> > >  (&inode->i_mutex){--..}, at: [<b038c6e5>] mutex_lock+0x1c/0x1f
> > > 
> > > which lock already depends on the new lock.
> > > 
> > 
> > This is the correct ranking: i_mutex outside mmap_sem.
> > 
> > > 
> > > the existing dependency chain (in reverse order) is:
> > > 
> > > -> #1 (&inode->i_mutex){--..}:
> > >        [<b013e3fb>] __lock_acquire+0xdee/0xf9c
> > >        [<b013e600>] lock_acquire+0x57/0x70
> > >        [<b038c4a5>] __mutex_lock_slowpath+0x73/0x297
> > >        [<b038c6e5>] mutex_lock+0x1c/0x1f
> > >        [<b01b17e9>] reiserfs_file_release+0x54/0x447
> > >        [<b016afe7>] __fput+0x53/0x101
> > >        [<b016b0ee>] fput+0x19/0x1c
> > >        [<b015bcd5>] remove_vma+0x3b/0x4d
> > >        [<b015c659>] do_munmap+0x17f/0x1cf
> > >        [<b015c6db>] sys_munmap+0x32/0x42
> > >        [<b0103f04>] sysenter_past_esp+0x5d/0x99
> > >        [<ffffffff>] 0xffffffff
> > > 
> > > -> #0 (&mm->mmap_sem){----}:
> > >        [<b013e259>] __lock_acquire+0xc4c/0xf9c
> > >        [<b013e600>] lock_acquire+0x57/0x70
> > >        [<b0137b92>] down_read+0x3a/0x4c
> > >        [<b018cb34>] dio_get_page+0x54/0x161
> > >        [<b018d7a9>] __blockdev_direct_IO+0x514/0xe2a
> > >        [<b01cf449>] ext3_direct_IO+0x98/0x1e5
> > >        [<b014e8df>] generic_file_direct_IO+0x63/0x133
> > >        [<b01500e9>] generic_file_aio_read+0x16b/0x222
> > >        [<b017f8b6>] aio_rw_vect_retry+0x5a/0x116
> > >        [<b0180147>] aio_run_iocb+0x69/0x129
> > >        [<b0180a78>] io_submit_one+0x194/0x2eb
> > >        [<b0181331>] sys_io_submit+0x92/0xe7
> > >        [<b0103f90>] syscall_call+0x7/0xb
> > >        [<ffffffff>] 0xffffffff
> > 
> > But here reiserfs is taking i_mutex in its file_operations.release(),
> > which can be called under mmap_sem.
> > 
> > Vladimir's recent de14569f94513279e3d44d9571a421e9da1759ae.
> > "resierfs: avoid tail packing if an inode was ever mmapped" comes real
> > close to this code, but afaict it did not cause this bug.
> > 
> > I can't think of anything which we've done in the 2.6.21 cycle which
> > would have caused this to start happening.  Odd.
> 
> The bug may be holder, let me know if you want me to check 2.6.20 or
> earlier.

Would be great if you could test 2.6.20.  I have a feeling that I missed
something, but what?  We didn't change the refcounting of lifetime of
vma.vm_file...


> > > The test run was fio, the job file used is:
> > > 
> > > # fio job file snip below
> > > [global]
> > > bs=4k
> > > buffered=0
> > > ioengine=libaio
> > > iodepth=4
> > > thread
> > > 
> > > [readers]
> > > numjobs=8
> > > size=128m
> > > rw=read
> > > # fio job file snip above
> > > 
> > > Filesystem was ext3, default mkfs and mount options. Kernel was
> > > 2.6.21-rc7 as of this morning, with some CFQ patches applied.
> > > 
> > 
> > It's interesting that lockdep learned the (wrong) ranking from a reiserfs
> > operation then later detected it being violated by ext3.
> 
> It's a scratch test box, which for some reason has reiserfs as the
> rootfs. So reiser gets to run first :-)

direct-io reads against reiserfs also will take i_mutex outside mmap_sem. 
As will pagefaults inside generic_file_write() (which is where this ranking
is primarily defined).

So an all-reiserfs system should be getting the same reports.  Obviously,
that isn't happening.

It's a bit odd that reiserfs is playing with file contents within
file_operations.release(): there could be other files open against that
inode.  One would expect this sort of thing to be happening in an
inode_operation.  But it's been like that for a long time.

Is it possible that fio was changed?  That it was changed to close() the fd
before doing the munmapping whereas it used to hold the file open?