your patch "mm: Remove false WARN_ON from pagecache_isize_extended()"

your patch "mm: Remove false WARN_ON from pagecache_isize_extended()"

[Jan Beulich]

Jan,

having run into that warning too, I looked into it a little, and now
having found that patch am pretty uncertain: Both truncate_setsize()
and pagecache_isize_extended() document that they want to be
called with i_mutex held, so removing the WARN_ON() alone seems
either incomplete or wrong. What I found to work without violating
this documented requirement is the patch below.

Jan

--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -797,7 +797,7 @@ xfs_file_fallocate(
 		     FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
 		return -EOPNOTSUPP;
 
-	xfs_ilock(ip, XFS_IOLOCK_EXCL);
+	xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
 	if (mode & FALLOC_FL_PUNCH_HOLE) {
 		error = xfs_free_file_space(ip, offset, len);
 		if (error)
@@ -877,7 +877,7 @@ xfs_file_fallocate(
 	}
 
 out_unlock:
-	xfs_iunlock(ip, XFS_IOLOCK_EXCL);
+	xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
 	return error;
 }
 

Re: your patch "mm: Remove false WARN_ON from pagecache_isize_extended()"

[Dave Chinner]

On Mon, Nov 03, 2014 at 04:41:13PM +0000, Jan Beulich wrote:
> Jan,
> 
> having run into that warning too, I looked into it a little, and now
> having found that patch am pretty uncertain: Both truncate_setsize()
> and pagecache_isize_extended() document that they want to be
> called with i_mutex held, so removing the WARN_ON() alone seems
> either incomplete or wrong. What I found to work without violating
> this documented requirement is the patch below.

Or, just perhaps, the comments are wrong....

Some filesystems have stronger, more robust internal serialisation
than the VFS provides with i_mutex....

> --- a/fs/xfs/xfs_file.c
> +++ b/fs/xfs/xfs_file.c
> @@ -797,7 +797,7 @@ xfs_file_fallocate(
>  		     FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE))
>  		return -EOPNOTSUPP;
>  
> -	xfs_ilock(ip, XFS_IOLOCK_EXCL);
> +	xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
>  	if (mode & FALLOC_FL_PUNCH_HOLE) {
>  		error = xfs_free_file_space(ip, offset, len);
>  		if (error)

The i_mutex is completely redundant here. Not to mention there are
multiple callers of these xfs_*_file_space() functions used to
implement fallocate(), and we're not about to change the locking
model for them....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: your patch "mm: Remove false WARN_ON from pagecache_isize_extended()"

[Jan Beulich]

>>> On 03.11.14 at 23:18, <david@fromorbit.com> wrote:
> On Mon, Nov 03, 2014 at 04:41:13PM +0000, Jan Beulich wrote:
>> having run into that warning too, I looked into it a little, and now
>> having found that patch am pretty uncertain: Both truncate_setsize()
>> and pagecache_isize_extended() document that they want to be
>> called with i_mutex held, so removing the WARN_ON() alone seems
>> either incomplete or wrong. What I found to work without violating
>> this documented requirement is the patch below.
> 
> Or, just perhaps, the comments are wrong....

Right - that's what I was suggesting with the option of the patch
being incomplete (rather than just removing the WARN_ON() it
should also remove the respective comments then).

Jan

Re: your patch "mm: Remove false WARN_ON from pagecache_isize_extended()"

[Jan Kara]

On Tue 04-11-14 07:40:20, Jan Beulich wrote:
> >>> On 03.11.14 at 23:18, <david@fromorbit.com> wrote:
> > On Mon, Nov 03, 2014 at 04:41:13PM +0000, Jan Beulich wrote:
> >> having run into that warning too, I looked into it a little, and now
> >> having found that patch am pretty uncertain: Both truncate_setsize()
> >> and pagecache_isize_extended() document that they want to be
> >> called with i_mutex held, so removing the WARN_ON() alone seems
> >> either incomplete or wrong. What I found to work without violating
> >> this documented requirement is the patch below.
> > 
> > Or, just perhaps, the comments are wrong....
> 
> Right - that's what I was suggesting with the option of the patch
> being incomplete (rather than just removing the WARN_ON() it
> should also remove the respective comments then).
  Yeah, I'll update the comments to mention there are filesystem which use
serialization different from i_mutex. Thanks for catching this.

								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR