large directory block size operations are assert failing because
xfs_bunmapi() is not completely removing fragmented directory blocks
like so:

XFS: Assertion failed: done, file: fs/xfs/libxfs/xfs_dir2.c, line: 677
....
Call Trace:
 xfs_dir2_shrink_inode+0x1a8/0x210
 xfs_dir2_block_to_sf+0x2ae/0x410
 xfs_dir2_block_removename+0x21a/0x280
 xfs_dir_removename+0x195/0x1d0
 xfs_rename+0xb79/0xc50
 ? avc_has_perm+0x8d/0x1a0
 ? avc_has_perm_noaudit+0x9a/0x120
 xfs_vn_rename+0xdb/0x150
 vfs_rename+0x719/0xb50
 ? __lookup_hash+0x6a/0xa0
 do_renameat2+0x413/0x5e0
 __x64_sys_rename+0x45/0x50
 do_syscall_64+0x3a/0x70
 entry_SYSCALL_64_after_hwframe+0x44/0xae

We are aborting the bunmapi() pass because of this specific chunk of
code:

                /*
                 * Make sure we don't touch multiple AGF headers out of order
                 * in a single transaction, as that could cause AB-BA deadlocks.
                 */
                if (!wasdel && !isrt) {
                        agno = XFS_FSB_TO_AGNO(mp, del.br_startblock);
                        if (prev_agno != NULLAGNUMBER && prev_agno > agno)
                                break;
                        prev_agno = agno;
                }

This is designed to prevent deadlocks in AGF locking when freeing
multiple extents by ensuring that we only ever lock in increasing
AG number order. Unfortunately, this also violates the "bunmapi will
always succeed" semantic that some high level callers depend on,
such as xfs_dir2_shrink_inode(), xfs_da_shrink_inode() and
xfs_inactive_symlink_rmt().

This AG lock ordering was introduced back in 2017 to fix deadlocks
triggered by generic/299 as reported here:

https://lore.kernel.org/linux-xfs/800468eb-3ded-9166-20a4-047de8018582@gmail.com/

This codebase is old enough that it was before we were defering all
AG based extent freeing from within xfs_bunmapi(). THat is, we never
actually lock AGs in xfs_bunmapi() any more - every non-rt based
extent free is added to the defer ops list, as is all BMBT block
freeing. And RT extents are not RT based, so there's no lock
ordering issues associated with them.

Hence this AGF lock ordering code is both broken and dead. Let's
just remove it so that the large directory block code works reliably
again.

Tested against xfs/538 and generic/299 which is the original test
that exposed the deadlocks that this code fixed.

Fixes: 5b094d6dac04 ("xfs: fix multi-AG deadlock in xfs_bunmapi")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>

xfs/538 is assert failing with this trace when testing with
directory block sizes of 64kB:

XFS: Assertion failed: !xfs_need_iread_extents(ifp), file: fs/xfs/libxfs/xfs_bmap.c, line: 608
....
Call Trace:
 xfs_bmap_btree_to_extents+0x2a9/0x470
 ? kmem_cache_alloc+0xe7/0x220
 __xfs_bunmapi+0x4ca/0xdf0
 xfs_bunmapi+0x1a/0x30
 xfs_dir2_shrink_inode+0x71/0x210
 xfs_dir2_block_to_sf+0x2ae/0x410
 xfs_dir2_block_removename+0x21a/0x280
 xfs_dir_removename+0x195/0x1d0
 xfs_remove+0x244/0x460
 xfs_vn_unlink+0x53/0xa0
 ? selinux_inode_unlink+0x13/0x20
 vfs_unlink+0x117/0x220
 do_unlinkat+0x1a2/0x2d0
 __x64_sys_unlink+0x42/0x60
 do_syscall_64+0x3a/0x70
 entry_SYSCALL_64_after_hwframe+0x44/0xae

This is a check to ensure that the extents have been read into
memory before we are doing a ifork btree manipulation. This assert
is bogus in the above case.

We have a fragmented directory block that has more extents in it
than can fit in extent format, so the inode data fork is in btree
format. xfs_dir2_shrink_inode() asks to remove all remaining 16
filesystem blocks from the inode so it can convert to short form,
and __xfs_bunmapi() removes all the extents. We now have a data fork
in btree format but have zero extents in the fork. This incorrectly
trips the xfs_need_iread_extents() assert because it assumes that an
empty extent btree means the extent tree has not been read into
memory yet. This is clearly not the case with xfs_bunmapi(), as it
has an explicit call to xfs_iread_extents() in it to pull the
extents into memory before it starts unmapping.

Also, the assert directly after this bogus one is:

	ASSERT(ifp->if_format == XFS_DINODE_FMT_BTREE);

Which covers the context in which it is legal to call
xfs_bmap_btree_to_extents just fine. Hence we should just remove the
bogus assert as it is clearly wrong and causes a regression.

The returns the test behaviour to the pre-existing assert failure in
xfs_dir2_shrink_inode() that indicates xfs_bunmapi() has failed to
remove all the extents in the range it was asked to unmap.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>

The RTINHERIT bit can be set on a directory so that newly created
regular files will have the REALTIME bit set to store their data on the
realtime volume.  If an extent size hint (and EXTSZINHERIT) are set on
the directory, the hint will also be copied into the new file.

As pointed out in previous patches, for realtime files we require the
extent size hint be an integer multiple of the realtime extent, but we
don't perform the same validation on a directory with both RTINHERIT and
EXTSZINHERIT set, even though the only use-case of that combination is
to propagate extent size hints into new realtime files.  This leads to
inode corruption errors when the bad values are propagated.

Because there may be existing filesystems with such a configuration, we
cannot simply amend the inode verifier to trip on these directories and
call it a day because that will cause previously "working" filesystems
to start throwing errors abruptly.  Note that it's valid to have
directories with rtinherit set even if there is no realtime volume, in
which case the problem does not manifest because rtinherit is ignored if
there's no realtime device; and it's possible that someone set the flag,
crashed, repaired the filesystem (which clears the hint on the realtime
file) and continued.

Therefore, mitigate this issue in several ways: First, if we try to
write out an inode with both rtinherit/extszinherit set and an unaligned
extent size hint, turn off the hint to correct the error.  Second, if
someone tries to misconfigure a directory via the fssetxattr ioctl, fail
the ioctl.  Third, reverify both extent size hint values when we
propagate heritable inode attributes from parent to child, to prevent
misconfigurations from spreading.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

While chasing a bug involving invalid extent size hints being propagated
into newly created realtime files, I noticed that the xfs_ioctl_setattr
checks for the extent size hints weren't the same as the ones now
encoded in libxfs and used for validation in repair and mkfs.

Because the checks in libxfs are more stringent than the ones in the
ioctl, it's possible for a live system to set inode flags that
immediately result in corruption warnings.  Specifically, it's possible
to set an extent size hint on an rtinherit directory without checking if
the hint is aligned to the realtime extent size, which makes no sense
since that combination is used only to seed new realtime files.

Replace the open-coded and inadequate checks with the libxfs verifier
versions and update the code comments a bit.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

The new online shrink code exposed a gap in the per-AG reservation
code, which is that we only return ENOSPC to callers if the entire fs
doesn't have enough free blocks.  Except for debugging mode, the
reservation init code doesn't ever check that there's enough free space
in that AG to cover the reservation.

Not having enough space is not considered an immediate fatal error that
requires filesystem offlining because (a) it's shouldn't be possible to
wind up in that state through normal file operations and (b) even if
one did, freeing data blocks would recover the situation.

However, online shrink now needs to know if shrinking would not leave
enough space so that it can abort the shrink operation.  Hence we need
to promote this assertion into an actual error return.

Observed by running xfs/168 with a 1k block size, though in theory this
could happen with any configuration.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>

These ioctl definitions in xfs_fs.h are part of the userspace ABI and
were mistakenly removed during the 5.13 merge window.

Fixes: 9fefd5db08ce ("xfs: convert to fileattr")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

sc->ip is the inode that's being scrubbed, which means that it's not set
for scrub types that don't involve inodes.  If one of those scrubbers
(e.g. inode btrees) returns EDEADLOCK, we'll trip over the null pointer.
Fix that by reporting either the file being examined or the file that
was used to call scrub.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>

If a realtime allocation fails because we can't find a sufficiently
large free extent satisfying locality rules, relax the locality rules
and try again.  This reduces the occurrence of short writes to realtime
files when the write size is large and the free space is fragmented.

This was originally discovered by running generic/186 with the realtime
reflink patchset and a 128k cow extent size hint, but the short write
symptoms can manifest with a 128k extent size hint and no reflink, so
apply the fix now.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

xfs_bmap_rtalloc doesn't handle realtime extent files with extent size
hints larger than the rt volume's extent size properly, because
xfs_bmap_extsize_align can adjust the offset/length parameters to try to
fit the extent size hint.

Under these conditions, minlen has to be large enough so that any
allocation returned by xfs_rtallocate_extent will be large enough to
cover at least one of the blocks that the caller asked for.  If the
allocation is too short, bmapi_write will return no mapping for the
requested range, which causes ENOSPC errors in other parts of the
filesystem.

Therefore, adjust minlen upwards to fix this.  This can be found by
running generic/263 (g/127 or g/522) with a realtime extent size hint
that's larger than the rt volume extent size.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

Pull more iomap updates from Darrick Wong:
 "Remove the now unused 'io_private' field from struct iomap_ioend, for
  a modest savings in memory allocation"

* tag 'iomap-5.13-merge-3' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
  iomap: remove unused private field from ioend