[PATCH] xfs_file_last_byte() needs to acquire ilock

[Lachlan McIlroy]

----- "Felix Blyakher" <felixb at sgi.com> wrote:

> On Apr 27, 2009, at 11:11 PM, Lachlan McIlroy wrote:
> 
> > ----- "Felix Blyakher" <felixb at sgi.com> wrote:
> >
> >> On Apr 23, 2009, at 10:46 PM, Lachlan McIlroy wrote:
> >>
> >>>
> >>> ----- "Eric Sandeen" <sandeen at sandeen.net> wrote:
> >>>
> >>>> Lachlan McIlroy wrote:
> >>>>> We had some systems crash with this stack:
> >>>>>
> >>>>> [<a00000010000cb20>] ia64_leave_kernel+0x0/0x280
> >>>>> [<a00000021291ca00>] xfs_bmbt_get_startoff+0x0/0x20 [xfs]
> >>>>> [<a0000002129080b0>] xfs_bmap_last_offset+0x210/0x280 [xfs]
> >>>>> [<a00000021295b010>] xfs_file_last_byte+0x70/0x1a0 [xfs]
> >>>>> [<a00000021295b200>] xfs_itruncate_start+0xc0/0x1a0 [xfs]
> >>>>> [<a0000002129935f0>] xfs_inactive_free_eofblocks+0x290/0x460
> >> [xfs]
> >>>>> [<a000000212998fb0>] xfs_release+0x1b0/0x240 [xfs]
> >>>>> [<a0000002129ad930>] xfs_file_release+0x70/0xa0 [xfs]
> >>>>> [<a000000100162ea0>] __fput+0x1a0/0x420
> >>>>> [<a000000100163160>] fput+0x40/0x60
> >>>>>
> >>>>> The problem here is that xfs_file_last_byte() does not acquire
> >> the
> >>>>> inode lock and can therefore race with another thread that is
> >>>> modifying
> >>>>> the extext list.  While xfs_bmap_last_offset() is trying to
> >> lookup
> >>>>> what was the last extent some extents were merged and the
> extent
> >>>> list
> >>>>> shrunk so the index we lookup is now beyond the end of the
> extent
> >>>> list
> >>>>> and potentially in a freed buffer.
> >>>>>
> >>>>> diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
> >>>>> index e7ae08d..cf62d9d 100644
> >>>>> --- a/fs/xfs/xfs_inode.c
> >>>>> +++ b/fs/xfs/xfs_inode.c
> >>>>> @@ -1258,8 +1258,10 @@ xfs_file_last_byte(
> >>>>
> >>>>       /*
> >>>>        * Only check for blocks beyond the EOF if the extents
> have
> >>>>        * been read in.  This eliminates the need for the inode
> >>>> lock,
> >>>>        * and it also saves us from looking when it really isn't
> >>>>> 	 * necessary.
> >>>>> 	 */
> >>>>
> >>>> I suppose that comment should be modified too, and maybe the
> >> commit
> >>>> log
> >>>> should say why, exactly, it was wrong? :)
> >>> Ha, I didn't even read the comment!  It's still kind of correct
> in
> >>> that we wont have to get the inode lock if the extents have not
> >> been
> >>>
> >>> read in.
> >>
> >> I'd still think the comments could be made less confusing
> >> if we're adding the inode lock here.
> > The more I read the comment the more it makes sense and it seems to
> > make more sense now with the change because it is clear how we can
> > avoid the inode lock if the extents are not read in.
> 
> OK, now after your explanation and reading the comments the Nth time,
> I think, I know what you mean.
> 
> I think, the original comment intention was the following:
> 
>          if (ip->i_df.if_flags & XFS_IFEXTENTS) {
> 		// extents have been read in. This (the fact that the extents
> 		// have been read in) eliminates the need for the inode lock, as
> 		// we are not going to read them in through xfs_iread_extents().
>                  error = xfs_bmap_last_offset(NULL, ip, &last_block,
>                          XFS_DATA_FORK);
>                  if (error) {
>                          last_block = 0;
>                  }
>          } else {
>                  last_block = 0;
>          }
> 
> while in the patched version it'll become:
> 
>          if (ip->i_df.if_flags & XFS_IFEXTENTS) {
> 		// extents have been read in ...
> 		xfs_ilock(ip, XFS_ILOCK_SHARED);
>                  error = xfs_bmap_last_offset(NULL, ip, &last_block,
>                          XFS_DATA_FORK);
> 		xfs_iunlock(ip, XFS_ILOCK_SHARED);
>                  if (error) {
>                          last_block = 0;
>                  }
>          } else {
> 		// this (the fact that the extents have _NOT_ been read in)
> 		// eliminates the need for the inode lock.
> 		// Doh, obvious.
>                  last_block = 0;
>          }
> 
> Is that how you see the comment now?
Yes.  And I think that was the intention of the comment all along.

> 
> Was the assumption in the original comment about not needing the
> ilock
> simply incorrect?
I don't think so.  I think the more likely scenario is that someone
erroneously removed the locking or it was never there to begin with.
The locking is not there in version 1.1 in git so you'll have to look
at ptools or the IRIX source to get the full history.

> 
> > How would you prefer the comment reads?
> 
> I'd just leave the first sentence from the original comment.
> 
>           * Only check for blocks beyond the EOF if the extents have
>           * been read in.
> 
> The mentioning about the ilock is too confusing now, imho.
> 
> Felix
> 
> >
> >
> >>
> >> Felix
> >>
> >>>
> >>>
> >>>>
> >>>> -Eric
> >>>>
> >>>>> 	if (ip->i_df.if_flags & XFS_IFEXTENTS) {
> >>>>> +		xfs_ilock(ip, XFS_ILOCK_SHARED);
> >>>>> 		error = xfs_bmap_last_offset(NULL, ip, &last_block,
> >>>>> 			XFS_DATA_FORK);
> >>>>> +		xfs_iunlock(ip, XFS_ILOCK_SHARED);
> >>>>> 		if (error) {
> >>>>> 			last_block = 0;
> >>>>> 		}
> >>>>>
> >>>>> _______________________________________________
> >>>>> xfs mailing list
> >>>>> xfs at oss.sgi.com
> >>>>> http://oss.sgi.com/mailman/listinfo/xfs
> >>>>>
> >>>>
> >>>> _______________________________________________
> >>>> xfs mailing list
> >>>> xfs at oss.sgi.com
> >>>> http://oss.sgi.com/mailman/listinfo/xfs
> >>>
> >>> _______________________________________________
> >>> xfs mailing list
> >>> xfs at oss.sgi.com
> >>> http://oss.sgi.com/mailman/listinfo/xfs
> >>
> >> _______________________________________________
> >> xfs mailing list
> >> xfs at oss.sgi.com
> >> http://oss.sgi.com/mailman/listinfo/xfs
> 
> _______________________________________________
> xfs mailing list
> xfs at oss.sgi.com
> http://oss.sgi.com/mailman/listinfo/xfs