Re: [PATCH V5 12/12] xfs: Introduce error injection to allocate only minlen size extents for files

[Chandan Babu R <chandanrlinux@gmail.com>]

On Tuesday 6 October 2020 2:47:02 PM IST Chandan Babu R wrote:
> On Tuesday 6 October 2020 10:04:24 AM IST Darrick J. Wong wrote:
> > On Sat, Oct 03, 2020 at 11:26:33AM +0530, Chandan Babu R wrote:
> > > This commit adds XFS_ERRTAG_BMAP_ALLOC_MINLEN_EXTENT error tag which
> > > helps userspace test programs to get xfs_bmap_btalloc() to always
> > > allocate minlen sized extents.
> > > 
> > > This is required for test programs which need a guarantee that minlen
> > > extents allocated for a file do not get merged with their existing
> > > neighbours in the inode's BMBT. "Inode fork extent overflow check" for
> > > Directories, Xattrs and extension of realtime inodes need this since the
> > > file offset at which the extents are being allocated cannot be
> > > explicitly controlled from userspace.
> > > 
> > > One way to use this error tag is to,
> > > 1. Consume all of the free space by sequentially writing to a file.
> > > 2. Punch alternate blocks of the file. This causes CNTBT to contain
> > >    sufficient number of one block sized extent records.
> > > 3. Inject XFS_ERRTAG_BMAP_ALLOC_MINLEN_EXTENT error tag.
> > > After step 3, xfs_bmap_btalloc() will issue space allocation
> > > requests for minlen sized extents only.
> > 
> > Is step #2 required?  What happens if I only turn the knob?
> 
> If there are no minlen sized free space extents in the CNTBT, we would return
> -ENOSPC to the userspace process. The reason behind forcing allocation of
> minlen sized CNTBT records is to make sure that these newly allocated extents
> do not get merged with their neighbouring extents in the inode's BMBT. On the
> other hand, if we did allow slicing off minlen sized chunks of a larger free
> space extent record in the CNTBT, the newly allocated extent records could be
> contiguous (w.r.t both disk offset and file offset) with its neighbours in the
> BMBT and hence merged, therby reducing inode fork extent count. This will
> prevent us from writing deterministic "Inode extent count overflow" tests for
> Directories, xattrs and realtime inodes.
> 
> > > ENOSPC error code is returned to userspace when there aren't any "one
> > > block sized" extents left in any of the AGs.
> > > 
> > > Signed-off-by: Chandan Babu R <chandanrlinux@gmail.com>
> > > ---
> > >  fs/xfs/libxfs/xfs_alloc.c    | 46 ++++++++++++++++++++++++++++++++++++
> > >  fs/xfs/libxfs/xfs_alloc.h    |  1 +
> > >  fs/xfs/libxfs/xfs_bmap.c     | 26 ++++++++++++++------
> > >  fs/xfs/libxfs/xfs_errortag.h |  4 +++-
> > >  fs/xfs/xfs_error.c           |  3 +++
> > >  5 files changed, 72 insertions(+), 8 deletions(-)
> > > 
> > > diff --git a/fs/xfs/libxfs/xfs_alloc.c b/fs/xfs/libxfs/xfs_alloc.c
> > > index 852b536551b5..d8d8ab1478db 100644
> > > --- a/fs/xfs/libxfs/xfs_alloc.c
> > > +++ b/fs/xfs/libxfs/xfs_alloc.c
> > > @@ -2473,6 +2473,45 @@ xfs_defer_agfl_block(
> > >  	xfs_defer_add(tp, XFS_DEFER_OPS_TYPE_AGFL_FREE, &new->xefi_list);
> > >  }
> > >  
> > > +STATIC int
> > > +minlen_freespace_available(
> > 
> > This ought to have an 'xfs_' prefix.
> 
> Ok. I will fix this up.
> > 
> > Also, what does this function do?  Does it decide if there's even enough
> > space to go ahead with a minlen allocation?
> 
> I will come up with a better name for this function. This function checks if
> there is a freespace extent record whose length is exactly equal to
> args->minlen.
> 
> > 
> > > +	struct xfs_alloc_arg	*args,
> > > +	struct xfs_buf		*agbp,
> > > +	int			*stat)
> > > +{
> > > +	xfs_btree_cur_t		*cnt_cur;
> > 
> > struct xfs_btree_cur	*cnt_cur;
> 
> Sorry, I will fix that up.
> 
> > 
> > > +	xfs_agblock_t		fbno;
> > > +	xfs_extlen_t		flen;
> > > +	int			btree_error = XFS_BTREE_NOERROR;
> > > +	int			error = 0;
> > > +
> > > +	cnt_cur = xfs_allocbt_init_cursor(args->mp, args->tp, agbp,
> > > +			args->agno, XFS_BTNUM_CNT);
> > > +	error = xfs_alloc_lookup_ge(cnt_cur, 0, args->minlen, stat);
> > > +	if (error) {
> > > +		btree_error = XFS_BTREE_ERROR;
> > > +		goto out;
> > > +	}
> > > +
> > > +	ASSERT(*stat == 1);
> > 
> > Is it ok to keep going with stat==0?  Or should we just ... I don't
> > know?  Bail out with -EFSCORRUPTED?
> 
> I think returning with -EFSCORRUPTED is a better option since before
> executing the code here, we would have already executed
> xfs_alloc_space_available() to make sure that atleast minlen free space is
> available in the AG whose CNTBT is being traversed. Thanks for the
> suggestion.
> 
> > 
> > > +
> > > +	error = xfs_alloc_get_rec(cnt_cur, &fbno, &flen, stat);
> > > +	if (error) {
> > > +		btree_error = XFS_BTREE_ERROR;
> > > +		goto out;
> > > +	}
> > > +
> > > +	if (flen == args->minlen)
> > > +		*stat = 1;
> > > +	else
> > > +		*stat = 0;
> > > +
> > > +out:
> > > +	xfs_btree_del_cursor(cnt_cur, btree_error);
> > 
> > Note that due to a sloppy quirk of error handling, you can pass @error
> > to this function, no need for a separate btree_error.
> 
> Ok. Thanks for pointing that out. I will fix this.
> 
> > 
> > > +
> > > +	return error;
> > > +}
> > > +
> > >  /*
> > >   * Decide whether to use this allocation group for this allocation.
> > >   * If so, fix up the btree freelist's size.
> > > @@ -2490,6 +2529,7 @@ xfs_alloc_fix_freelist(
> > >  	struct xfs_alloc_arg	targs;	/* local allocation arguments */
> > >  	xfs_agblock_t		bno;	/* freelist block */
> > >  	xfs_extlen_t		need;	/* total blocks needed in freelist */
> > > +	int			i;
> > >  	int			error = 0;
> > >  
> > >  	/* deferred ops (AGFL block frees) require permanent transactions */
> > > @@ -2544,6 +2584,12 @@ xfs_alloc_fix_freelist(
> > >  	if (!xfs_alloc_space_available(args, need, flags))
> > >  		goto out_agbp_relse;
> > >  
> > > +	if (args->alloc_minlen_only) {
> > > +		error = minlen_freespace_available(args, agbp, &i);
> > > +		if (error || !i)
> > > +			goto out_agbp_relse;
> > > +	}
> > > +
> > >  	/*
> > >  	 * Make the freelist shorter if it's too long.
> > >  	 *
> > > diff --git a/fs/xfs/libxfs/xfs_alloc.h b/fs/xfs/libxfs/xfs_alloc.h
> > > index 6c22b12176b8..1d04089b7fb4 100644
> > > --- a/fs/xfs/libxfs/xfs_alloc.h
> > > +++ b/fs/xfs/libxfs/xfs_alloc.h
> > > @@ -75,6 +75,7 @@ typedef struct xfs_alloc_arg {
> > >  	char		wasfromfl;	/* set if allocation is from freelist */
> > >  	struct xfs_owner_info	oinfo;	/* owner of blocks being allocated */
> > >  	enum xfs_ag_resv_type	resv;	/* block reservation to use */
> > > +	bool		alloc_minlen_only;
> > >  } xfs_alloc_arg_t;
> > >  
> > >  /*
> > > diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
> > > index 5156cbd476f2..fab4097e7492 100644
> > > --- a/fs/xfs/libxfs/xfs_bmap.c
> > > +++ b/fs/xfs/libxfs/xfs_bmap.c
> > > @@ -3510,12 +3510,19 @@ xfs_bmap_btalloc(
> > >  		ASSERT(ap->length);
> > >  	}
> > >  
> > > +	memset(&args, 0, sizeof(args));
> > > +
> > > +	args.alloc_minlen_only = XFS_TEST_ERROR(false, mp,
> > > +					XFS_ERRTAG_BMAP_ALLOC_MINLEN_EXTENT);
> > 
> > Can we just set maxlen = minlen here?
> 
> I had noticed that xfs_bmap_btalloc() is structured as described below,
> 1. Compute the appropriate filesystem-wide block number (and hence the AG)
>    to start searching for free space extents.
> 2. Compute xfs_alloc_arg->{type, total, minlen, maxlen}.
> 3. Compute xfs_alloc_arg->alignment and adjust xfs_alloc_arg->{type, maxlen}
>    as required.
> 4. Invoke xfs_alloc_vextent().
> 
> To keep up with the existing code flow, I had set
> xfs_alloc_args->{minlen, maxlen, total} to xfs_bmalloca->minlen at function
> location corresponding to step 2.
> 
> > 
> > Also, should this debug knob also be applied to rt file allocations?
> 
> I had missed xfs_bmap_alloc_userdata() => xfs_bmap_rtalloc() sequence. I will
> add the error tag to rt file allocations as well. Thanks for pointing that
> out.

Actually the debug knob is not required for rt file allocations because they
take the same path as direct i/o writes and hence a userspace test program
could control the file offsets at which writes take place in order to prevent
neighbouring extents from getting merged into a single one. An example test
program is given below,

# realtime file
add_nosplit_5_iext_count_overflow_check()
{
        umount $dev

        mkfs.xfs -f -K -d size=${fssize} -r rtdev=${rtdev} -m reflink=0,rmapbt=0 $dev || \
                { print "Unable to mkfs.xfs $dev"; exit 1 }

        mount -o rtdev=${rtdev} $dev $mntpnt || { print "Unable to mount $dev"; exit 1 }

        testfile=${mntpnt}/testfile

        nr_blks=$((15 * 2))

        xfs_io -x -c 'inject reduce_max_iextents' $mntpnt

        for i in $(seq 0 2 $(($nr_blks - 1))); do
                xfs_io -Rf -c "pwrite $(($i * $bsize)) $bsize" -c fsync $testfile > /dev/null 2>&1
                [[ $? != 0 ]] && { echo "Failed to write at block $i"; break; }
        done

        ls -i $testfile
        # Make sure that this is a realtime file
        xfs_io -c 'lsattr' $testfile
        xfs_io -f -c "fiemap" $testfile | grep -i -v hole
}

In the above script, we write at non-contiguous file offsets and hence this is
sufficient to guarantee that the resulting file extents do not get merged with
their neighbours.

> 
> > 
> > >  
> > >  	nullfb = ap->tp->t_firstblock == NULLFSBLOCK;
> > >  	fb_agno = nullfb ? NULLAGNUMBER : XFS_FSB_TO_AGNO(mp,
> > >  							ap->tp->t_firstblock);
> > >  	if (nullfb) {
> > > -		if ((ap->datatype & XFS_ALLOC_USERDATA) &&
> > > +		if (args.alloc_minlen_only) {
> > > +			ag = 0;
> > 
> > Hm, so setting this magic knob also makes everyone fight for space in AG 0?
> 
> For the normal use case, each AGF tracks the longest extent via
> xfs_agf->agf_longest. When the transaction is allocating its first
> extent, xfs_bmap_btalloc_nullfb() loops over each AG until it finds an AG
> whose longest extent can be used for allocating xfs_alloc_arg->maxlen free
> space extent. 
> 
> However, there is no such existing facility for tracking "minimum length"
> extent in an AG. This could be done by adding a new member to the in-memory
> data structure and intializing the new member by assigning the "length" value
> of the leftmost record from CNTBT during xfs_alloc_read_agf(). However I
> refrained from doing this since we will never need this on production
> machines.
> 
> Also, since xfs_alloc_arg->type is being to XFS_ALLOCTYPE_FIRST_AG later in
> the code, AG 0 is just the first AG being scanned for "exact minlen"
> extents. We end up looping across remaining AGs if previously searched AGs do
> not contain "exact minlen" extents.
> 
> > 
> > > +			ap->blkno = XFS_AGB_TO_FSB(mp, ag, 0);
> > > +		} else if ((ap->datatype & XFS_ALLOC_USERDATA) &&
> > >  		    xfs_inode_is_filestream(ap->ip)) {
> > >  			ag = xfs_filestream_lookup_ag(ap->ip);
> > >  			ag = (ag != NULLAGNUMBER) ? ag : 0;
> > > @@ -3523,10 +3530,12 @@ xfs_bmap_btalloc(
> > >  		} else {
> > >  			ap->blkno = XFS_INO_TO_FSB(mp, ap->ip->i_ino);
> > >  		}
> > > -	} else
> > > +	} else {
> > >  		ap->blkno = ap->tp->t_firstblock;
> > > +	}
> > >  
> > > -	xfs_bmap_adjacent(ap);
> > > +	if (!args.alloc_minlen_only)
> > > +		xfs_bmap_adjacent(ap);
> > >  
> > >  	/*
> > >  	 * If allowed, use ap->blkno; otherwise must use firstblock since
> > > @@ -3540,7 +3549,6 @@ xfs_bmap_btalloc(
> > >  	 * Normal allocation, done through xfs_alloc_vextent.
> > >  	 */
> > >  	tryagain = isaligned = 0;
> > > -	memset(&args, 0, sizeof(args));
> > >  	args.tp = ap->tp;
> > >  	args.mp = mp;
> > >  	args.fsbno = ap->blkno;
> > > @@ -3549,7 +3557,10 @@ xfs_bmap_btalloc(
> > >  	/* Trim the allocation back to the maximum an AG can fit. */
> > >  	args.maxlen = min(ap->length, mp->m_ag_max_usable);
> > >  	blen = 0;
> > > -	if (nullfb) {
> > > +	if (args.alloc_minlen_only) {
> > > +		args.type = XFS_ALLOCTYPE_START_AG;
> > > +		args.total = args.minlen = args.maxlen = ap->minlen;
> > > +	} else if (nullfb) {
> > >  		/*
> > >  		 * Search for an allocation group with a single extent large
> > >  		 * enough for the request.  If one isn't found, then adjust
> > > @@ -3595,7 +3606,8 @@ xfs_bmap_btalloc(
> > >  	 * is only set if the allocation length is >= the stripe unit and the
> > >  	 * allocation offset is at the end of file.
> > >  	 */
> > > -	if (!(ap->tp->t_flags & XFS_TRANS_LOWMODE) && ap->aeof) {
> > > +	if (!(ap->tp->t_flags & XFS_TRANS_LOWMODE) && ap->aeof &&
> > > +		!args.alloc_minlen_only) {
> > >  		if (!ap->offset) {
> > 
> > Yikes, the conditional lines up with the body!
> 
> Sorry, I will fix this.
> 
> > 
> > --D
> > 
> > >  			args.alignment = stripe_align;
> > >  			atype = args.type;
> > > @@ -3681,7 +3693,7 @@ xfs_bmap_btalloc(
> > >  		if ((error = xfs_alloc_vextent(&args)))
> > >  			return error;
> > >  	}
> > > -	if (args.fsbno == NULLFSBLOCK && nullfb) {
> > > +	if (args.fsbno == NULLFSBLOCK && nullfb && !args.alloc_minlen_only) {
> > >  		args.fsbno = 0;
> > >  		args.type = XFS_ALLOCTYPE_FIRST_AG;
> > >  		args.total = ap->minlen;
> > > diff --git a/fs/xfs/libxfs/xfs_errortag.h b/fs/xfs/libxfs/xfs_errortag.h
> > > index 1c56fcceeea6..6ca9084b6934 100644
> > > --- a/fs/xfs/libxfs/xfs_errortag.h
> > > +++ b/fs/xfs/libxfs/xfs_errortag.h
> > > @@ -57,7 +57,8 @@
> > >  #define XFS_ERRTAG_IUNLINK_FALLBACK			34
> > >  #define XFS_ERRTAG_BUF_IOERROR				35
> > >  #define XFS_ERRTAG_REDUCE_MAX_IEXTENTS			36
> > > -#define XFS_ERRTAG_MAX					37
> > > +#define XFS_ERRTAG_BMAP_ALLOC_MINLEN_EXTENT		37
> > > +#define XFS_ERRTAG_MAX					38
> > >  
> > >  /*
> > >   * Random factors for above tags, 1 means always, 2 means 1/2 time, etc.
> > > @@ -99,5 +100,6 @@
> > >  #define XFS_RANDOM_IUNLINK_FALLBACK			(XFS_RANDOM_DEFAULT/10)
> > >  #define XFS_RANDOM_BUF_IOERROR				XFS_RANDOM_DEFAULT
> > >  #define XFS_RANDOM_REDUCE_MAX_IEXTENTS			1
> > > +#define XFS_RANDOM_BMAP_ALLOC_MINLEN_EXTENT		1
> > >  
> > >  #endif /* __XFS_ERRORTAG_H_ */
> > > diff --git a/fs/xfs/xfs_error.c b/fs/xfs/xfs_error.c
> > > index 3780b118cc47..028560bb596a 100644
> > > --- a/fs/xfs/xfs_error.c
> > > +++ b/fs/xfs/xfs_error.c
> > > @@ -55,6 +55,7 @@ static unsigned int xfs_errortag_random_default[] = {
> > >  	XFS_RANDOM_IUNLINK_FALLBACK,
> > >  	XFS_RANDOM_BUF_IOERROR,
> > >  	XFS_RANDOM_REDUCE_MAX_IEXTENTS,
> > > +	XFS_RANDOM_BMAP_ALLOC_MINLEN_EXTENT,
> > >  };
> > >  
> > >  struct xfs_errortag_attr {
> > > @@ -166,6 +167,7 @@ XFS_ERRORTAG_ATTR_RW(bad_summary,	XFS_ERRTAG_FORCE_SUMMARY_RECALC);
> > >  XFS_ERRORTAG_ATTR_RW(iunlink_fallback,	XFS_ERRTAG_IUNLINK_FALLBACK);
> > >  XFS_ERRORTAG_ATTR_RW(buf_ioerror,	XFS_ERRTAG_BUF_IOERROR);
> > >  XFS_ERRORTAG_ATTR_RW(reduce_max_iextents,	XFS_ERRTAG_REDUCE_MAX_IEXTENTS);
> > > +XFS_ERRORTAG_ATTR_RW(bmap_alloc_minlen_extent, XFS_ERRTAG_BMAP_ALLOC_MINLEN_EXTENT);
> > >  
> > >  static struct attribute *xfs_errortag_attrs[] = {
> > >  	XFS_ERRORTAG_ATTR_LIST(noerror),
> > > @@ -205,6 +207,7 @@ static struct attribute *xfs_errortag_attrs[] = {
> > >  	XFS_ERRORTAG_ATTR_LIST(iunlink_fallback),
> > >  	XFS_ERRORTAG_ATTR_LIST(buf_ioerror),
> > >  	XFS_ERRORTAG_ATTR_LIST(reduce_max_iextents),
> > > +	XFS_ERRORTAG_ATTR_LIST(bmap_alloc_minlen_extent),
> > >  	NULL,
> > >  };
> > >  
> > 
> 
> 
> 

-- 
chandan