Re: [PATCH] xfs: truncate should remove all blocks, not just to the end of the page cache

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Dec 24, 2019 at 12:21:27AM -0800, Christoph Hellwig wrote:
> On Sun, Dec 22, 2019 at 08:36:30AM -0800, Darrick J. Wong wrote:
> > From: Darrick J. Wong <darrick.wong@oracle.com>
> > 
> > xfs_itruncate_extents_flags() is supposed to unmap every block in a file
> > from EOF onwards.  Oddly, it uses s_maxbytes as the upper limit to the
> > bunmapi range, even though s_maxbytes reflects the highest offset the
> > pagecache can support, not the highest offset that XFS supports.
> > 
> > The result of this confusion is that if you create a 20T file on a
> > 64-bit machine, mount the filesystem on a 32-bit machine, and remove the
> > file, we leak everything above 16T.  Fix this by capping the bunmapi
> > request at the maximum possible block offset, not s_maxbytes.
> > 
> > Fixes: 32972383ca462 ("xfs: make largest supported offset less shouty")
> 
> Why would that fix that commit?  The commit just changed how do derive
> the value, but not the value itself.

I'm not sure what to put for a fixes tag when the code in question is
from the bitkeeper era.

> > diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
> > index 401da197f012..eaa85d5933cb 100644
> > --- a/fs/xfs/xfs_inode.c
> > +++ b/fs/xfs/xfs_inode.c
> > @@ -1544,9 +1544,12 @@ xfs_itruncate_extents_flags(
> >  	 * possible file size.  If the first block to be removed is
> >  	 * beyond the maximum file size (ie it is the same as last_block),
> >  	 * then there is nothing to do.
> > +	 *
> > +	 * We have to free all the blocks to the bmbt maximum offset, even if
> > +	 * the page cache can't scale that far.
> >  	 */
> >  	first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
> > -	last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
> > +	last_block = (1ULL << BMBT_STARTOFF_BITLEN) - 1;
> >  	if (first_unmap_block == last_block)
> >  		return 0;
> 
> That check is now never true.  I think that whole function wants some
> attenttion instead.  Kill that whole last_block calculation, switch to
> __xfs_bunmapi and pass ULLONG_MAX for the rlen input and just exit the
> loop once rlen is 0.

I'll give that a try.

--D