Re: [PATCH] Fix speculative allocation beyond eof

[Lachlan McIlroy <lachlan@sgi.com>]

Eric Sandeen wrote:
> Lachlan McIlroy wrote:
>> Speculative allocation beyond eof doesn't work properly.  It was broken some
>> time ago after a code cleanup that moved what is now xfs_iomap_eof_align_last_fsb()
>> and xfs_iomap_eof_want_preallocate() out of xfs_iomap_write_delay() into
>> separate functions.  The code used to use the current file size in various checks
>> but got changed to be max(file_size, i_new_size).  Since i_new_size is the result
>> of 'offset + count' then in xfs_iomap_eof_want_preallocate() the check for
>> '(offset + count) <= isize' will always be true.
>>
>> ie if 'offset + count' is > ip->i_size then isize will be i_new_size and equal to
>> 'offset + count'.
>>
>> This change fixes all the places that used to use the current file size.
> 
> Lachlan, so what's the failure mode?  (is it that we simply don't do
> that speculative preallocation anymore due to the problem?)
Yes, we are not doing speculative allocation anymore.

Actually it's not that clear cut.  The 'offset + count' used in these functions
is rounded up to a page boundary by the calling code so if the user is not page
aligning their I/O then 'offset + count' will extend further into the file than
what the user specified.  And i_new_size is the result of the actual user's
'offset + count'.  The result is users who do not page align their I/O will get
some speculative allocation but users who do page align their I/O (in order to
get better performance) will be punished.

> 
>>From your description of the change above, it seems like you're talking
> about:
> 
> dd9f438e32900d67def49fa1b8961b3e19b6fefc
> 
> [XFS] Implement the di_extsize allocator hint for non-realtime files as
> well.  Also provides a mechanism for inheriting this property from the
> parent directory for new files.
> 
> SGI-PV: 945264
> SGI-Modid: xfs-linux-melb:xfs-kern:24367a
> 
> Signed-off-by: Nathan Scott <nathans@sgi.com>
> 
> but that isn't really a cleanup; and I don't think it changed the isize
> behavior.
That change did a lot of things.  I was referring the large chunks of code that
were moved into new functions.

Before the change we had this
-       if (!(ioflag & BMAPI_SYNC) && ((offset + count) > ip->i_d.di_size)) {

and after it is now
+       if ((ioflag & BMAPI_SYNC) || (offset + count) <= isize)

where isize is max(ip->i_d.di_size, io->io_new_size).  Note that the logic has
been inverted but it's clear that we're using the wrong size.  That's where the
choice to do speculative allocation got busted.  The code that does the stripe
width/unit rounding has been broken much longer - in version 1.3 of xfs_iomap.c
we changed XFS_SIZE(mp, io) to isize.  The code has morphed a lot since then but
we kept the use of isize.

> 
> Was it possibly [XFS] Fix to prevent the notorious 'NULL files' problem
> after a crash. that caused the regression?
No, that just changes uses of the on-disk file size to be the in-memory file size.

> 
> Thanks,
> -Eric
> 
>> --- a/fs/xfs/xfs_iomap.c	2008-09-23 12:52:12.000000000 +1000
>> +++ b/fs/xfs/xfs_iomap.c	2008-09-23 12:51:29.000000000 +1000
>> @@ -290,7 +290,6 @@ STATIC int
>>  xfs_iomap_eof_align_last_fsb(
>>  	xfs_mount_t	*mp,
>>  	xfs_inode_t	*ip,
>> -	xfs_fsize_t	isize,
>>  	xfs_extlen_t	extsize,
>>  	xfs_fileoff_t	*last_fsb)
>>  {
>> @@ -306,14 +305,14 @@ xfs_iomap_eof_align_last_fsb(
>>  	 * stripe width and we are allocating past the allocation eof.
>>  	 */
>>  	else if (mp->m_swidth && (mp->m_flags & XFS_MOUNT_SWALLOC) &&
>> -	        (isize >= XFS_FSB_TO_B(mp, mp->m_swidth)))
>> +	        (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_swidth)))
>>  		new_last_fsb = roundup_64(*last_fsb, mp->m_swidth);
>>  	/*
>>  	 * Roundup the allocation request to a stripe unit (m_dalign) boundary
>>  	 * if the file size is >= stripe unit size, and we are allocating past
>>  	 * the allocation eof.
>>  	 */
>> -	else if (mp->m_dalign && (isize >= XFS_FSB_TO_B(mp, mp->m_dalign)))
>> +	else if (mp->m_dalign && (ip->i_size >= XFS_FSB_TO_B(mp, mp->m_dalign)))
>>  		new_last_fsb = roundup_64(*last_fsb, mp->m_dalign);
>>  
>>  	/*
>> @@ -403,7 +402,6 @@ xfs_iomap_write_direct(
>>  	xfs_filblks_t	count_fsb, resaligned;
>>  	xfs_fsblock_t	firstfsb;
>>  	xfs_extlen_t	extsz, temp;
>> -	xfs_fsize_t	isize;
>>  	int		nimaps;
>>  	int		bmapi_flag;
>>  	int		quota_flag;
>> @@ -426,15 +424,10 @@ xfs_iomap_write_direct(
>>  	rt = XFS_IS_REALTIME_INODE(ip);
>>  	extsz = xfs_get_extsz_hint(ip);
>>  
>> -	isize = ip->i_size;
>> -	if (ip->i_new_size > isize)
>> -		isize = ip->i_new_size;
>> -
>>  	offset_fsb = XFS_B_TO_FSBT(mp, offset);
>>  	last_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)(offset + count)));
>> -	if ((offset + count) > isize) {
>> -		error = xfs_iomap_eof_align_last_fsb(mp, ip, isize, extsz,
>> -							&last_fsb);
>> +	if ((offset + count) > ip->i_size) {
>> +		error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
>>  		if (error)
>>  			goto error_out;
>>  	} else {
>> @@ -559,7 +552,6 @@ STATIC int
>>  xfs_iomap_eof_want_preallocate(
>>  	xfs_mount_t	*mp,
>>  	xfs_inode_t	*ip,
>> -	xfs_fsize_t	isize,
>>  	xfs_off_t	offset,
>>  	size_t		count,
>>  	int		ioflag,
>> @@ -573,7 +565,7 @@ xfs_iomap_eof_want_preallocate(
>>  	int		n, error, imaps;
>>  
>>  	*prealloc = 0;
>> -	if ((ioflag & BMAPI_SYNC) || (offset + count) <= isize)
>> +	if ((ioflag & BMAPI_SYNC) || (offset + count) <= ip->i_size)
>>  		return 0;
>>  
>>  	/*
>> @@ -617,7 +609,6 @@ xfs_iomap_write_delay(
>>  	xfs_fileoff_t	ioalign;
>>  	xfs_fsblock_t	firstblock;
>>  	xfs_extlen_t	extsz;
>> -	xfs_fsize_t	isize;
>>  	int		nimaps;
>>  	xfs_bmbt_irec_t imap[XFS_WRITE_IMAPS];
>>  	int		prealloc, fsynced = 0;
>> @@ -637,11 +628,7 @@ xfs_iomap_write_delay(
>>  	offset_fsb = XFS_B_TO_FSBT(mp, offset);
>>  
>>  retry:
>> -	isize = ip->i_size;
>> -	if (ip->i_new_size > isize)
>> -		isize = ip->i_new_size;
>> -
>> -	error = xfs_iomap_eof_want_preallocate(mp, ip, isize, offset, count,
>> +	error = xfs_iomap_eof_want_preallocate(mp, ip, offset, count,
>>  				ioflag, imap, XFS_WRITE_IMAPS, &prealloc);
>>  	if (error)
>>  		return error;
>> @@ -655,8 +642,7 @@ retry:
>>  	}
>>  
>>  	if (prealloc || extsz) {
>> -		error = xfs_iomap_eof_align_last_fsb(mp, ip, isize, extsz,
>> -							&last_fsb);
>> +		error = xfs_iomap_eof_align_last_fsb(mp, ip, extsz, &last_fsb);
>>  		if (error)
>>  			return error;
>>  	}
>>
>>
> 
>