Re: [PATCH V3 09/12] xfs: Enable bulkstat ioctl to support 64-bit per-inode extent counters

[Dave Chinner <david@fromorbit.com>]

On Tue, Sep 28, 2021 at 03:19:29PM +0530, Chandan Babu R wrote:
> On 28 Sep 2021 at 04:36, Dave Chinner wrote:
> > On Thu, Sep 16, 2021 at 03:36:44PM +0530, Chandan Babu R wrote:
> >> @@ -492,9 +494,16 @@ struct xfs_bulk_ireq {
> >>   */
> >>  #define XFS_BULK_IREQ_METADIR	(1 << 2)
> >>  
> >> -#define XFS_BULK_IREQ_FLAGS_ALL	(XFS_BULK_IREQ_AGNO | \
> >> +#define XFS_BULK_IREQ_BULKSTAT	(1 << 3)
> >> +
> >> +#define XFS_BULK_IREQ_FLAGS_ALL	(XFS_BULK_IREQ_AGNO |	 \
> >>  				 XFS_BULK_IREQ_SPECIAL | \
> >> -				 XFS_BULK_IREQ_METADIR)
> >> +				 XFS_BULK_IREQ_METADIR | \
> >> +				 XFS_BULK_IREQ_BULKSTAT)
> >
> > What's this XFS_BULK_IREQ_METADIR thing? I haven't noticed that when
> > scanning any recent proposed patch series....
> >
> 
> XFS_BULK_IREQ_METADIR is from Darrick's tree. His "Kill XFS_BTREE_MAXLEVELS"
> patch series is based on his other patchsets. His recent "xfs: support dynamic
> btree cursor height" patch series rebases only the required patchset on top of
> v5.15-rc1 kernel eliminating the others.

OK, so how much testing has this had on just a straight v5.15-rcX
kernel?

> >> @@ -134,7 +136,26 @@ xfs_bulkstat_one_int(
> >>  
> >>  	buf->bs_xflags = xfs_ip2xflags(ip);
> >>  	buf->bs_extsize_blks = ip->i_extsize;
> >> -	buf->bs_extents = xfs_ifork_nextents(&ip->i_df);
> >> +
> >> +	nextents = xfs_ifork_nextents(&ip->i_df);
> >> +	if (!(bc->breq->flags & XFS_IBULK_NREXT64)) {
> >> +		xfs_extnum_t max_nextents = XFS_IFORK_EXTCNT_MAXS32;
> >> +
> >> +		if (unlikely(XFS_TEST_ERROR(false, mp,
> >> +				XFS_ERRTAG_REDUCE_MAX_IEXTENTS)))
> >> +			max_nextents = 10;
> >> +
> >> +		if (nextents > max_nextents) {
> >> +			xfs_iunlock(ip, XFS_ILOCK_SHARED);
> >> +			xfs_irele(ip);
> >> +			error = -EINVAL;
> >> +			goto out_advance;
> >> +		}
> >
> > So we return an EINVAL error if any extent overflows the 32 bit
> > counter? Why isn't this -EOVERFLOW?
> >
> 
> Returning -EINVAL causes xfs_bulkstat_iwalk() to skip inodes whose extent
> count is larger than that which can be fitted into a 32-bit field. Returning
> -EOVERFLOW causes the bulkstat ioctl to stop reporting remaining inodes.

Ok, that's a bad behaviour we need to fix because it will cause
things like old versions of xfs_dump to miss inodes that
have overflowing extent counts. i.e. it will cause incomplete
backups, and the failure will likely be silent.

I asked about -EOVERFLOW because that's what stat() returns when an
inode attribute value doesn't fit in the stat_buf field (e.g. 64 bit
inode number on 32 bit kernel), and if we are overflowing the
bulkstat field then we really should be telling userspace that an
overflow occurred.

/me has a sudden realisation that the xfs_dump format may not
support large extents counts and goes looking...

Yeah, xfsdump doesn't support extent counts greater than 2^32. So
that means we really do need -EOVERFLOW errors here.  i.e, if we get
an extent count overflow with a !(bc->breq->flags &
XFS_IBULK_NREXT64) bulkstat walk, xfs_dump needs bulkstat to fill
out the inode with the overflow with all the fileds that aren't
overflowed, then error out with -EOVERFLOW.

Bulkstat itself should not silently skip the inode because it would
overflow a field in the struct xfs-bstat - the decision of what to
do with the overflow is something xfsdump needs to handle, not the
kernel.  Hence we need to return -EOVERFLOW here so that userspace
can decide what to do with an inode it can't handle...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com