Re: 3.1.0-rc3 -- INFO: possible circular locking dependency detected

[Hugh Dickins <hughd@google.com>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 3767 bytes --]

On Tue, 23 Aug 2011, Josh Boyer wrote:
> On Tue, Aug 23, 2011 at 9:04 AM, Dave Chinner <david@fromorbit.com> wrote:
> > On Tue, Aug 23, 2011 at 07:59:20AM -0400, Josh Boyer wrote:
> >> On Tue, Aug 23, 2011 at 7:49 AM, Dave Chinner <david@fromorbit.com> wrote:
> >> >> >  Possible unsafe locking scenario:
> >> >> >
> >> >> >       CPU0                    CPU1
> >> >> >       ----                    ----
> >> >> >  lock(&mm->mmap_sem);
> >> >> >                               lock(&sb->s_type->i_mutex_key);
> >> >> >                               lock(&mm->mmap_sem);
> >> >> >  lock(&sb->s_type->i_mutex_key);
> >> >> >
> >> >> >  *** DEADLOCK ***
> >> >>
> >> >> This one was reported yesterday: https://lkml.org/lkml/2011/8/21/163
> >> >> and we're hoping Ted (or someone else from the ext4 camp) can comment
> >> >> on why ext4_evict_inode is holding i_mutex.
> >> >
> >> > Actually, the problem has nothing to do with ext4. the problem is
> >> > that remove_vma() is holding the mmap_sem while calling fput(). The
> >> > correct locking order is i_mutex->mmap_sem, as documented in
> >> > mm/filemap.c:
> >> >
> >> >  *  ->i_mutex                   (generic_file_buffered_write)
> >> >  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
> >> >
> >> >
> >> > The way remove_vma() calls fput() also triggers lockdep reports in
> >> > XFS and it will do so with any filesystem that takes an inode
> >> > specific lock in it's evict() processing. IOWs, remove_vma() needs
> >> > fixing, not ext4....
> >>
> >> Er... ok.  So the remove_vma code hasn't changed since 2008.  We're
> >> only seeing this issue now because the debugging code has improved,
> >> or?
> >
> > The problem has been there since at least 2008.  Here's an early
> > XFS report from 2.6.24:
> >
> > http://oss.sgi.com/archives/xfs/2008-02/msg00931.html
> >
> > Here's an XFS report
> > to match the ext4 one in this thread from 2009:
> >
> > http://oss.sgi.com/archives/xfs/2009-03/msg00149.html
> >
> > You won't find reports much older than this - it only started to be
> > reported when lockdep support in XFS matured and it started to be
> > widely used....
> >
> >> At any rate, the proposed solution is to make remove_vma drop mmap_sem
> >> before calling fput, or make it not call fput, or?
> >
> > Ask the VM folk - this is the only response I can remember from them
> > is this:
> >
> > http://oss.sgi.com/archives/xfs/2009-03/msg00224.html
> >
> > Maybe now that ext4 is hitting the problem something will be done
> > about it...
> 
> OK.  I've CC'd Andrew and Hugh, so maybe we can get a discussion going.

My first reaction would be that this is quite simply a filesystem bug.
Apparently a long-standing bug in the XFS case, but one in which ext4
has just now (3.1-rc) joined it.

The mm/fs locking hierarchy has been that way forever: mm does not assume
that the fs will not take any inode-specific lock in its fput(), but yes,
it does expect fput() not to take the i_mutex.

In this new ext4 case, it appears to be just an issue on final eviction
of the inode, which I think makes actual deadlock (when writing to file
needs to fault in a page from mm) impossible - we wouldn't be evicting
it if there were still references.  Just needs some lockdep notation?

Dropping mmap_sem while doing fput() in munmap() doesn't sound appealing
to me: although we have converted a number of paths to drop mmap_sem in
strategic places, getting back to the (possibly changed) vma sequence
afterwards is tiresome (when the munmap covers multiple vmas), and
instinct says that munmap() might be a more difficult case to get
right than most.  Leave the fput()s to a workqueue instead?

Hugh