xfs: factor out a new xfs_log_force_inode helper

Create a new helper to force the log up to the last LSN touching an
inode.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: fix inode number overflow in ifree cluster helper

Qian Cai reports seemingly random buffer read verifier errors during
filesystem writeback. This was isolated to a recent patch that
factored out some inode cluster freeing code and happened to cast an
unsigned inode number type to a signed value. If the inode number
value overflows, we can skip marking in-core inodes associated with
the underlying buffer stale at the time the physical inodes are
freed. If such an inode happens to be dirty, xfsaild will eventually
attempt to write it back over non-inode blocks. The invalidation of
the underlying inode buffer causes writeback to read the buffer from
disk. This fails the read verifier (preventing eventual corruption)
if the buffer no longer looks like an inode cluster. Analysis by
Dave Chinner.

Fix up the helper to use the proper type for inode number values.

Fixes: 5806165a6663 ("xfs: factor inode lookup from xfs_ifree_cluster")
Reported-by: Qian Cai <cai@lca.pw>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove redundant variable assignment in xfs_symlink()

The variables 'udqp' and 'gdqp' have been initialized, so remove
redundant variable assignment in xfs_symlink().

Signed-off-by: Kaixu Xia <kaixuxia@tencent.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: ratelimit inode flush on buffered write ENOSPC

A customer reported rcu stalls and softlockup warnings on a computer
with many CPU cores and many many more IO threads trying to write to a
filesystem that is totally out of space.  Subsequent analysis pointed to
the many many IO threads calling xfs_flush_inodes -> sync_inodes_sb,
which causes a lot of wb_writeback_work to be queued.  The writeback
worker spends so much time trying to wake the many many threads waiting
for writeback completion that it trips the softlockup detector, and (in
this case) the system automatically reboots.

In addition, they complain that the lengthy xfs_flush_inodes scan traps
all of those threads in uninterruptible sleep, which hampers their
ability to kill the program or do anything else to escape the situation.

If there's thousands of threads trying to write to files on a full
filesystem, each of those threads will start separate copies of the
inode flush scan.  This is kind of pointless since we only need one
scan, so rate limit the inode flush.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: return locked status of inode buffer on xfsaild push

If the inode buffer backing a particular inode is locked,
xfs_iflush() returns -EAGAIN and xfs_inode_item_push() skips the
inode. It still returns success to xfsaild, however, which bypasses
the xfsaild backoff heuristic. Update xfs_inode_item_push() to
return locked status if the inode buffer couldn't be locked.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: trylock underlying buffer on dquot flush

A dquot flush currently blocks on the buffer lock for the underlying
dquot buffer. In turn, this causes xfsaild to block rather than
continue processing other items in the meantime. Update
xfs_qm_dqflush() to trylock the buffer, similar to how inode buffers
are handled, and return -EAGAIN if the lock fails. Fix up any
callers that don't currently handle the error properly.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove unnecessary ternary from xfs_create

Since the "no-allocation" reservations for file creations has
been removed, the resblks value should be larger than zero, so
remove unnecessary ternary conditional.

Signed-off-by: Kaixu Xia <kaixuxia@tencent.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: s/judgment/ternary/]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: don't write a corrupt unmount record to force summary counter recalc

In commit f467cad95f5e3, I added the ability to force a recalculation of
the filesystem summary counters if they seemed incorrect.  This was done
(not entirely correctly) by tweaking the log code to write an unmount
record without the UMOUNT_TRANS flag set.  At next mount, the log
recovery code will fail to find the unmount record and go into recovery,
which triggers the recalculation.

What actually gets written to the log is what ought to be an unmount
record, but without any flags set to indicate what kind of record it
actually is.  This worked to trigger the recalculation, but we shouldn't
write bogus log records when we could simply write nothing.

Fixes: f467cad95f5e3 ("xfs: force summary counter recalc at next mount")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: factor inode lookup from xfs_ifree_cluster

There's lots of indent in this code which makes it a bit hard to
follow. We are also going to completely rework the inode lookup code
as part of the inode reclaim rework, so factor out the inode lookup
code from the inode cluster freeing code.

Based on prototype code from Christoph Hellwig.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: tail updates only need to occur when LSN changes

We currently wake anything waiting on the log tail to move whenever
the log item at the tail of the log is removed. Historically this
was fine behaviour because there were very few items at any given
LSN. But with delayed logging, there may be thousands of items at
any given LSN, and we can't move the tail until they are all gone.

Hence if we are removing them in near tail-first order, we might be
waking up processes waiting on the tail LSN to change (e.g. log
space waiters) repeatedly without them being able to make progress.
This also occurs with the new sync push waiters, and can result in
thousands of spurious wakeups every second when under heavy direct
reclaim pressure.

To fix this, check that the tail LSN has actually changed on the
AIL before triggering wakeups. This will reduce the number of
spurious wakeups when doing bulk AIL removal and make this code much
more efficient.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: factor common AIL item deletion code

Factor the common AIL deletion code that does all the wakeups into a
helper so we only have one copy of this somewhat tricky code to
interface with all the wakeups necessary when the LSN of the log
tail changes.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: correctly acount for reclaimable slabs

The XFS inode item slab actually reclaimed by inode shrinker
callbacks from the memory reclaim subsystem. These should be marked
as reclaimable so the mm subsystem has the full picture of how much
memory it can actually reclaim from the XFS slab caches.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: Improve metadata buffer reclaim accountability

The buffer cache shrinker frees more than just the xfs_buf slab
objects - it also frees the pages attached to the buffers. Make sure
the memory reclaim code accounts for this memory being freed
correctly, similar to how the inode shrinker accounts for pages
freed from the page cache due to mapping invalidation.

We also need to make sure that the mm subsystem knows these are
reclaimable objects. We provide the memory reclaim subsystem with a
a shrinker to reclaim xfs_bufs, so we should really mark the slab
that way.

We also have a lot of xfs_bufs in a busy system, spread them around
like we do inodes.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: don't allow log IO to be throttled

Running metadata intensive workloads, I've been seeing the AIL
pushing getting stuck on pinned buffers and triggering log forces.
The log force is taking a long time to run because the log IO is
getting throttled by wbt_wait() - the block layer writeback
throttle. It's being throttled because there is a huge amount of
metadata writeback going on which is filling the request queue.

IOWs, we have a priority inversion problem here.

Mark the log IO bios with REQ_IDLE so they don't get throttled
by the block layer writeback throttle. When we are forcing the CIL,
we are likely to need to to tens of log IOs, and they are issued as
fast as they can be build and IO completed. Hence REQ_IDLE is
appropriate - it's an indication that more IO will follow shortly.

And because we also set REQ_SYNC, the writeback throttle will now
treat log IO the same way it treats direct IO writes - it will not
throttle them at all. Hence we solve the priority inversion problem
caused by the writeback throttle being unable to distinguish between
high priority log IO and background metadata writeback.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: Throttle commits on delayed background CIL push

In certain situations the background CIL push can be indefinitely
delayed. While we have workarounds from the obvious cases now, it
doesn't solve the underlying issue. This issue is that there is no
upper limit on the CIL where we will either force or wait for
a background push to start, hence allowing the CIL to grow without
bound until it consumes all log space.

To fix this, add a new wait queue to the CIL which allows background
pushes to wait for the CIL context to be switched out. This happens
when the push starts, so it will allow us to block incoming
transaction commit completion until the push has started. This will
only affect processes that are running modifications, and only when
the CIL threshold has been significantly overrun.

This has no apparent impact on performance, and doesn't even trigger
until over 45 million inodes had been created in a 16-way fsmark
test on a 2GB log. That was limiting at 64MB of log space used, so
the active CIL size is only about 3% of the total log in that case.
The concurrent removal of those files did not trigger the background
sleep at all.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: Lower CIL flush limit for large logs

The current CIL size aggregation limit is 1/8th the log size. This
means for large logs we might be aggregating at least 250MB of dirty objects
in memory before the CIL is flushed to the journal. With CIL shadow
buffers sitting around, this means the CIL is often consuming >500MB
of temporary memory that is all allocated under GFP_NOFS conditions.

Flushing the CIL can take some time to do if there is other IO
ongoing, and can introduce substantial log force latency by itself.
It also pins the memory until the objects are in the AIL and can be
written back and reclaimed by shrinkers. Hence this threshold also
tends to determine the minimum amount of memory XFS can operate in
under heavy modification without triggering the OOM killer.

Modify the CIL space limit to prevent such huge amounts of pinned
metadata from aggregating. We can have 2MB of log IO in flight at
once, so limit aggregation to 16x this size. This threshold was
chosen as it little impact on performance (on 16-way fsmark) or log
traffic but pins a lot less memory on large logs especially under
heavy memory pressure.  An aggregation limit of 8x had 5-10%
performance degradation and a 50% increase in log throughput for
the same workload, so clearly that was too small for highly
concurrent workloads on large logs.

This was found via trace analysis of AIL behaviour. e.g. insertion
from a single CIL flush:

xfs_ail_insert: old lsn 0/0 new lsn 1/3033090 type XFS_LI_INODE flags IN_AIL

$ grep xfs_ail_insert /mnt/scratch/s.t |grep "new lsn 1/3033090" |wc -l
1721823
$

So there were 1.7 million objects inserted into the AIL from this
CIL checkpoint, the first at 2323.392108, the last at 2325.667566 which
was the end of the trace (i.e. it hadn't finished). Clearly a major
problem.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove some stale comments from the log code

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: refactor unmount record writing

Separate out the unmount record writing from the rest of the
ticket and log state futzing necessary to make it work. This is
a no-op, just makes the code cleaner and places the unmount record
formatting and writing alongside the commit record formatting and
writing code.

We can also get rid of the ticket flag clearing before the
xlog_write() call because it no longer cares about the state of
XLOG_TIC_INITED.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: merge xlog_commit_record with xlog_write_done

xlog_write_done() is just a thin wrapper around xlog_commit_record(), so
they can be merged together easily.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: split xlog_ticket_done

Remove xlog_ticket_done and just call the renamed low-level helpers for
ungranting or regranting log space directly.  To make that a little
the reference put on the ticket and all tracing is moved into the actual
helpers.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: kill XLOG_TIC_INITED

It is not longer used or checked by anything, so remove the last
traces from the log ticket code.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: refactor and split xfs_log_done()

xfs_log_done() does two separate things. Firstly, it triggers commit
records to be written for permanent transactions, and secondly it
releases or regrants transaction reservation space.

Since delayed logging was introduced, transactions no longer write
directly to the log, hence they never have the XLOG_TIC_INITED flag
cleared on them. Hence transactions never write commit records to
the log and only need to modify reservation space.

Split up xfs_log_done into two parts, and only call the parts of the
operation needed for the context xfs_log_done() is currently being
called from.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: re-order initial space accounting checks in xlog_write

Commit and unmount records records do not need start records to be
written, so rearrange the logic in xlog_write() to remove the need
to check for XLOG_TIC_INITED to determine if we should account for
the space used by a start record.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: don't try to write a start record into every iclog

The xlog_write() function iterates over iclogs until it completes
writing all the log vectors passed in. The ticket tracks whether
a start record has been written or not, so only the first iclog gets
a start record. We only ever pass single use tickets to
xlog_write() so we only ever need to write a start record once per
xlog_write() call.

Hence we don't need to store whether we should write a start record
in the ticket as the callers provide all the information we need to
determine if a start record should be written. For the moment, we
have to ensure that we clear the XLOG_TIC_INITED appropriately so
the code in xfs_log_done() still works correctly for committing
transactions.

(darrick: Note the slight behavior change that we always deduct the
size of the op header from the ticket, even for unmount records)

Signed-off-by: Dave Chinner <dchinner@redhat.com>
[hch: pass an explicit need_start_rec argument]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: validate the realtime geometry in xfs_validate_sb_common

Validate the geometry of the realtime geometry when we mount the
filesystem, so that we don't abruptly shut down the filesystem later on.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: prohibit fs freezing when using empty transactions

I noticed that fsfreeze can take a very long time to freeze an XFS if
there happens to be a GETFSMAP caller running in the background.  I also
happened to notice the following in dmesg:

------------[ cut here ]------------
WARNING: CPU: 2 PID: 43492 at fs/xfs/xfs_super.c:853 xfs_quiesce_attr+0x83/0x90 [xfs]
Modules linked in: xfs libcrc32c ip6t_REJECT nf_reject_ipv6 ipt_REJECT nf_reject_ipv4 ip_set_hash_ip ip_set_hash_net xt_tcpudp xt_set ip_set_hash_mac ip_set nfnetlink ip6table_filter ip6_tables bfq iptable_filter sch_fq_codel ip_tables x_tables nfsv4 af_packet [last unloaded: xfs]
CPU: 2 PID: 43492 Comm: xfs_io Not tainted 5.6.0-rc4-djw #rc4
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.10.2-1ubuntu1 04/01/2014
RIP: 0010:xfs_quiesce_attr+0x83/0x90 [xfs]
Code: 7c 07 00 00 85 c0 75 22 48 89 df 5b e9 96 c1 00 00 48 c7 c6 b0 2d 38 a0 48 89 df e8 57 64 ff ff 8b 83 7c 07 00 00 85 c0 74 de <0f> 0b 48 89 df 5b e9 72 c1 00 00 66 90 0f 1f 44 00 00 41 55 41 54
RSP: 0018:ffffc900030f3e28 EFLAGS: 00010202
RAX: 0000000000000001 RBX: ffff88802ac54000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: ffffffff81e4a6f0 RDI: 00000000ffffffff
RBP: ffff88807859f070 R08: 0000000000000001 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000010 R12: 0000000000000000
R13: ffff88807859f388 R14: ffff88807859f4b8 R15: ffff88807859f5e8
FS:  00007fad1c6c0fc0(0000) GS:ffff88807e000000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f0c7d237000 CR3: 0000000077f01003 CR4: 00000000001606a0
Call Trace:
 xfs_fs_freeze+0x25/0x40 [xfs]
 freeze_super+0xc8/0x180
 do_vfs_ioctl+0x70b/0x750
 ? __fget_files+0x135/0x210
 ksys_ioctl+0x3a/0xb0
 __x64_sys_ioctl+0x16/0x20
 do_syscall_64+0x50/0x1a0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe

These two things appear to be related.  The assertion trips when another
thread initiates a fsmap request (which uses an empty transaction) after
the freezer waited for m_active_trans to hit zero but before the the
freezer executes the WARN_ON just prior to calling xfs_log_quiesce.

The lengthy delays in freezing happen because the freezer calls
xfs_wait_buftarg to clean out the buffer lru list.  Meanwhile, the
GETFSMAP caller is continuing to grab and release buffers, which means
that it can take a very long time for the buffer lru list to empty out.

We fix both of these races by calling sb_start_write to obtain freeze
protection while using empty transactions for GETFSMAP and for metadata
scrubbing.  The other two users occur during mount, during which time we
cannot fs freeze.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: shutdown on failure to add page to log bio

If the bio_add_page() call fails, we proceed to write out a
partially constructed log buffer. This corrupts the physical log
such that log recovery is not possible. Worse, persistent
occurrences of this error eventually lead to a BUG_ON() failure in
bio_split() as iclogs wrap the end of the physical log, which
triggers log recovery on subsequent mount.

Rather than warn about writing out a corrupted log buffer, shutdown
the fs as is done for any log I/O related error. This preserves the
consistency of the physical log such that log recovery succeeds on a
subsequent mount. Note that this was observed on a 64k page debug
kernel without upstream commit 59bb47985c1d ("mm, sl[aou]b:
guarantee natural alignment for kmalloc(power-of-two)"), which
demonstrated frequent iclog bio overflows due to unaligned (slab
allocated) iclog data buffers.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: directory bestfree check should release buffers

When we're checking bestfree information in directory blocks, always
drop the block buffer at the end of the function.  We should always
release resources when we're done using them.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: drop all altpath buffers at the end of the sibling check

The dirattr btree checking code uses the altpath substructure of the
dirattr state structure to check the sibling pointers of dir/attr tree
blocks.  At the end of sibling checks, xfs_da3_path_shift could have
changed multiple levels of buffer pointers in the altpath structure.
Although we release the leaf level buffer, this isn't enough -- we also
need to release the node buffers that are unique to the altpath.

Not releasing all of the altpath buffers leaves them locked to the
transaction.  This is suboptimal because we should release resources
when we don't need them anymore.  Fix the function to loop all levels of
the altpath, and fix the return logic so that we always run the loop.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: preserve default grace interval during quotacheck

When quotacheck runs, it zeroes all the timer fields in every dquot.
Unfortunately, it also does this to the root dquot, which erases any
preconfigured grace intervals and warning limits that the administrator
may have set.  Worse yet, the incore copies of those variables remain
set.  This cache coherence problem manifests itself as the grace
interval mysteriously being reset back to the defaults at the /next/
mount.

Fix it by not resetting the root disk dquot's timer and warning fields.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: remove xlog_state_want_sync

Open code the xlog_state_want_sync logic in its two callers given that
this function is a trivial wrapper around xlog_state_switch_iclogs.

Move the lockdep assert into xlog_state_switch_iclogs to not lose this
debugging aid, and improve the comment that documents
xlog_state_switch_iclogs as well.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: move the ioerror check out of xlog_state_clean_iclog

Use the shutdown flag in the log to bypass xlog_state_clean_iclog
entirely in case of a shut down log.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: refactor xlog_state_clean_iclog

Factor out a few self-contained helpers from xlog_state_clean_iclog, and
update the documentation so it primarily documents why things happens
instead of how.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove the aborted parameter to xlog_state_done_syncing

We can just check for a shut down log all the way down in
xlog_cil_committed instead of passing the parameter.  This means a
slight behavior change in that we now also abort log items if the
shutdown came in halfway into the I/O completion processing, which
actually is the right thing to do.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: simplify log shutdown checking in xfs_log_release_iclog

There is no need to check for the ioerror state before the lock, as
the shutdown case is not a fast path.  Also remove the call to force
shutdown the file system, as it must have been shut down already
for an iclog to be in the ioerror state.  Also clean up the flow of
the function a bit.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: simplify the xfs_log_release_iclog calling convention

The only caller of xfs_log_release_iclog doesn't care about the return
value, so remove it.  Also don't bother passing the mount pointer,
given that we can trivially derive it from the iclog.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: factor out a xlog_wait_on_iclog helper

Factor out the shared code to wait for a log force into a new helper.
This helper uses the XLOG_FORCED_SHUTDOWN check previous only used
by the unmount code over the equivalent iclog ioerror state used by
the other two functions.

There is a slight behavior change in that the force of the unmount
record is now accounted in the log force statistics.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: merge xlog_cil_push into xlog_cil_push_work

xlog_cil_push is only called by xlog_cil_push_work, so merge the two
functions.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove the di_version field from struct icdinode

We know the version is 3 if on a v5 file system.   For earlier file
systems formats we always upgrade the remaining v1 inodes to v2 and
thus only use v2 inodes.  Use the xfs_sb_version_has_large_dinode
helper to check if we deal with small or large dinodes, and thus
remove the need for the di_version field in struct icdinode.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: simplify a check in xfs_ioctl_setattr_check_cowextsize

Only v5 file systems can have the reflink feature, and those will
always use the large dinode format.  Remove the extra check for the
inode version.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: simplify di_flags2 inheritance in xfs_ialloc

di_flags2 is initialized to zero for v4 and earlier file systems.  This
means di_flags2 can only be non-zero for a v5 file systems, in which
case both the parent and child inodes can store the field.  Remove the
extra di_version check, and also remove the rather pointless local
di_flags2 variable while at it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: only check the superblock version for dinode size calculation

The size of the dinode structure is only dependent on the file system
version, so instead of checking the individual inode version just use
the newly added xfs_sb_version_has_large_dinode helper, and simplify
various calling conventions.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: add a new xfs_sb_version_has_v3inode helper

Add a new wrapper to check if a file system supports the v3 inode format
with a larger dinode core.  Previously we used xfs_sb_version_hascrc for
that, which is technically correct but a little confusing to read.

Also move xfs_dinode_good_version next to xfs_sb_version_has_v3inode
so that we have one place that documents the superblock version to
inode version relationship.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: fix unmount hang and memory leak on shutdown during quotaoff

AIL removal of the quotaoff start intent and free of both quotaoff
intents is currently limited to the ->iop_committed() handler of the
end intent. This executes when the end intent is committed to the
on-disk log and marks the completion of the operation. The problem
with this is it assumes the success of the operation. If a shutdown
or other error occurs during the quotaoff, it's possible for the
quotaoff task to exit without removing the start intent from the
AIL. This results in an unmount hang as the AIL cannot be emptied.
Further, no other codepath frees the intents and so this is also a
memory leak vector.

First, update the high level quotaoff error path to directly remove
and free the quotaoff start intent if it still exists in the AIL at
the time of the error. Next, update both of the start and end
quotaoff intents with an ->iop_release() callback to properly handle
transaction abort.

This means that If the quotaoff start transaction aborts, it frees
the start intent in the transaction commit path. If the filesystem
shuts down before the end transaction allocates, the quotaoff
sequence removes and frees the start intent. If the end transaction
aborts, it removes the start intent and frees both. This ensures
that a shutdown does not result in a hung unmount and that memory is
not leaked regardless of when a quotaoff error occurs.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: factor out quotaoff intent AIL removal and memory free

AIL removal of the quotaoff start intent and free of both intents is
hardcoded to the ->iop_committed() handler of the end intent. Factor
out the start intent handling code so it can be used in a future
patch to properly handle quotaoff errors. Use xfs_trans_ail_remove()
instead of the _delete() variant to acquire the AIL lock and also
handle cases where an intent might not reside in the AIL at the
time of a failure.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: add support for rmap btree staging cursors

Add support for btree staging cursors for the rmap btrees.  This is
needed both for online repair and also to convert xfs_repair to use
btree bulk loading.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: add support for refcount btree staging cursors

Add support for btree staging cursors for the refcount btrees.  This
is needed both for online repair and also to convert xfs_repair to use
btree bulk loading.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: add support for inode btree staging cursors

Add support for btree staging cursors for the inode btrees.  This
is needed both for online repair and also to convert xfs_repair to use
btree bulk loading.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: add support for free space btree staging cursors

Add support for btree staging cursors for the free space btrees.  This
is needed both for online repair and also to convert xfs_repair to use
btree bulk loading.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: support bulk loading of staged btrees

Add a new btree function that enables us to bulk load a btree cursor.
This will be used by the upcoming online repair patches to generate new
btrees.  This avoids the programmatic inefficiency of calling
xfs_btree_insert in a loop (which generates a lot of log traffic) in
favor of stamping out new btree blocks with ordered buffers, and then
committing both the new root and scheduling the removal of the old btree
blocks in a single transaction commit.

The design of this new generic code is based off the btree rebuilding
code in xfs_repair's phase 5 code, with the explicit goal of enabling us
to share that code between scrub and repair.  It has the additional
feature of being able to control btree block loading factors.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: introduce fake roots for inode-rooted btrees

Create an in-core fake root for inode-rooted btree types so that callers
can generate a whole new btree using the upcoming btree bulk load
function without making the new tree accessible from the rest of the
filesystem.  It is up to the individual btree type to provide a function
to create a staged cursor (presumably with the appropriate callouts to
update the fakeroot) and then commit the staged root back into the
filesystem.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: introduce fake roots for ag-rooted btrees

Create an in-core fake root for AG-rooted btree types so that callers
can generate a whole new btree using the upcoming btree bulk load
function without making the new tree accessible from the rest of the
filesystem.  It is up to the individual btree type to provide a function
to create a staged cursor (presumably with the appropriate callouts to
update the fakeroot) and then commit the staged root back into the
filesystem.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: replace open-coded bitmap weight logic

Add a xbitmap_hweight helper function so that we can get rid of the
open-coded loop.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: rename xfs_bitmap to xbitmap

Shorten the name of xfs_bitmap to xbitmap since the scrub bitmap has
nothing to do with the libxfs bitmap.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: xrep_reap_extents should not destroy the bitmap

Remove the xfs_bitmap_destroy call from the end of xrep_reap_extents
because this sort of violates our rule that the function initializing a
structure should destroy it.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: fix incorrect test in xfs_alloc_ag_vextent_lastblock

When I lifted the code in xfs_alloc_ag_vextent_lastblock out of a loop,
I forgot to convert all the accesses to len to be pointer dereferences.

Coverity-id: 1457918
Fixes: 5113f8ec3753ed ("xfs: clean up weird while loop in xfs_alloc_ag_vextent_near")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: xfs_dabuf_map should return ENOMEM when map allocation fails

If the xfs_buf_map array allocation in xfs_dabuf_map fails for whatever
reason, we bail out with error code zero.  This will confuse callers, so
make sure that we return ENOMEM.  Allocation failure should never happen
with the small size of the array, but code defensively anyway.

Fixes: 45feef8f50b94d ("xfs: refactor xfs_dabuf_map")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>

xfs: cleanup xfs_log_unmount_write

Move the code for verifying the iclog state on a clean unmount into a
helper, and instead of checking the iclog state just rely on the shutdown
check as they are equivalent.  Also remove the ifdef DEBUG as the
compiler is smart enough to eliminate the dead code for non-debug builds.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove dead code from xfs_log_unmount_write

When the log is shut down all iclogs are in the XLOG_STATE_IOERROR state,
which means that xlog_state_want_sync and xlog_state_release_iclog are
no-ops.  Remove the whole section of code.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove the unused return value from xfs_log_unmount_write

Remove the ignored return value from xfs_log_unmount_write, and also
remove a rather pointless assert on the return value from xfs_log_force.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove the unused XLOG_UNMOUNT_REC_TYPE define

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: mark XLOG_FORCED_SHUTDOWN as unlikely

A shutdown log is a slow failure path.  Add an unlikely annotation to
it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: make the btree ag cursor private union anonymous

This is much less widely used than the bc_private union was, so this
is done as a single patch. The named union xfs_btree_cur_private
goes away and is embedded into the struct xfs_btree_cur_ag as an
anonymous union, and the code is modified via this script:

$ sed -i 's/priv\.\([abt|refc]\)/\1/g' fs/xfs/*[ch] fs/xfs/*/*[ch]

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: make the btree cursor union members named structure

we need to name the btree cursor private structures to be able
to pull them out of the deeply nested structure definition they are
in now.

Based on code extracted from a patchset by Darrick Wong.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: make btree cursor private union anonymous

Rename the union and it's internal structures to the new name and
remove the temporary defines that facilitated the change.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: rename btree cursor private btree member flags

BPRV is not longer appropriate because bc_private is going away.
Script:

$ sed -i 's/BTCUR_BPRV/BTCUR_BMBT/g' fs/xfs/*[ch] fs/xfs/*/*[ch]

With manual cleanup to the definitions in fs/xfs/libxfs/xfs_btree.h

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: change "BC_BT" to "BTCUR_BMBT", fix subject line typo]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: convert btree cursor inode-private member names

bc_private.b -> bc_ino conversion via script:

$ sed -i 's/bc_private\.b/bc_ino/g' fs/xfs/*[ch] fs/xfs/*/*[ch]

And then revert the change to the bc_ino #define in
fs/xfs/libxfs/xfs_btree.h manually.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: tweak the subject line slightly]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: convert btree cursor ag-private member name

bc_private.a -> bc_ag conversion via script:

`sed -i 's/bc_private\.a/bc_ag/g' fs/xfs/*[ch] fs/xfs/*/*[ch]`

And then revert the change to the bc_ag #define in
fs/xfs/libxfs/xfs_btree.h manually.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: introduce new private btree cursor names

Just the defines of the new names - the conversion will be in
scripted commits after this.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
[darrick: change "bc_bt" to "bc_ino"]
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: fix regression in "cleanup xfs_dir2_block_getdents"

Commit 263dde869bd09 ("xfs: cleanup xfs_dir2_block_getdents") introduced
a getdents regression, when it converted the pointer arithmetics to
offset calculations: offset is updated in the loop already for the next
iteration, but the updated offset value is used incorrectly in two
places, where we should have used the not-yet-updated value.

This caused for example "git clean -ffdx" failures to cleanup certain
directory structures when running in a container.

Fix the regression by making sure we use proper offset in the loop body.
Thanks to Christoph Hellwig for suggestion how to best fix the code.

Cc: Christoph Hellwig <hch@lst.de>
Fixes: 263dde869bd09 ("xfs: cleanup xfs_dir2_block_getdents")
Signed-off-by: Tommi Rantala <tommi.t.rantala@nokia.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: Use scnprintf() for avoiding potential buffer overflow

Since snprintf() returns the would-be-output size instead of the
actual output size, the succeeding calls may go beyond the given
buffer limit.  Fix it by replacing with scnprintf().

Signed-off-by: Takashi Iwai <tiwai@suse.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: mark extended attr corrupt when lookup-by-hash fails

In xchk_xattr_listent, we attempt to validate the extended attribute
hash structures by performing a attr lookup by (hashed) name.  If the
lookup returns ENODATA, that means that the hash information is corrupt.
The _process_error functions don't catch this, so we have to add that
explicitly.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: mark dir corrupt when lookup-by-hash fails

In xchk_dir_actor, we attempt to validate the directory hash structures
by performing a directory entry lookup by (hashed) name.  If the lookup
returns ENOENT, that means that the hash information is corrupt.  The
_process_error functions don't catch this, so we have to add that
explicitly.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: check owner of dir3 blocks

Check the owner field of dir3 block headers.  If it's corrupt, release
the buffer and return EFSCORRUPTED.  All callers handle this properly.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: check owner of dir3 data blocks

Check the owner field of dir3 data block headers.  If it's corrupt,
release the buffer and return EFSCORRUPTED.  All callers handle this
properly.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: check owner of dir3 free blocks

Check the owner field of dir3 free block headers and reject the metadata
if there's something wrong with it.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: don't ever return a stale pointer from __xfs_dir3_free_read

If we decide that a directory free block is corrupt, we must take care
not to leak a buffer pointer to the caller.  After xfs_trans_brelse
returns, the buffer can be freed or reused, which means that we have to
set *bpp back to NULL.

Callers are supposed to notice the nonzero return value and not use the
buffer pointer, but we should code more defensively, even if all current
callers handle this situation correctly.

Fixes: de14c5f541e7 ("xfs: verify free block header fields")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: fix buffer corruption reporting when xfs_dir3_free_header_check fails

xfs_verifier_error is supposed to be called on a corrupt metadata buffer
from within a buffer verifier function, whereas xfs_buf_mark_corrupt
is the function to be called when a piece of code has read a buffer and
catches something that a read verifier cannot.  The first function sets
b_error anticipating that the low level buffer handling code will see
the nonzero b_error and clear XBF_DONE on the buffer, whereas the second
function does not.

Since xfs_dir3_free_header_check examines fields in the dir free block
header that require more context than can be provided to read verifiers,
we must call xfs_buf_mark_corrupt when it finds a problem.

Switching the calls has a secondary effect that we no longer corrupt the
buffer state by setting b_error and leaving XBF_DONE set.  When /that/
happens, we'll trip over various state assertions (most commonly the
b_error check in xfs_buf_reverify) on a subsequent attempt to read the
buffer.

Fixes: bc1a09b8e334bf5f ("xfs: refactor verifier callers to print address of failing check")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: xfs_buf_corruption_error should take __this_address

Add a xfs_failaddr_t parameter to this function so that callers can
potentially pass in (and therefore report) the exact point in the code
where we decided that a metadata buffer was corrupt.  This enables us to
wire it up to checking functions that have to run outside of verifiers.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add a function to deal with corrupt buffers post-verifiers

Add a helper function to get rid of buffers that we have decided are
corrupt after the verifiers have run.  This function is intended to
handle metadata checks that can't happen in the verifiers, such as
inter-block relationship checking.  Note that we now mark the buffer
stale so that it will not end up on any LRU and will be purged on
release.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: fix xfs_rmap_has_other_keys usage of ECANCELED

In e7ee96dfb8c26, we converted all ITER_ABORT users to use ECANCELED
instead, but we forgot to teach xfs_rmap_has_other_keys not to return
that magic value to callers.  Fix it now by using ECANCELED both to
abort the iteration and to signal that we found another reverse mapping.
This enables us to drop the separate boolean flag.

Fixes: e7ee96dfb8c26 ("xfs: remove all *_ITER_ABORT values")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: fix use-after-free when aborting corrupt attr inactivation

Log the corrupt buffer before we release the buffer.

Fixes: a5155b870d687 ("xfs: always log corruption errors")
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: remove XFS_BUF_TO_SBP

Just dereference bp->b_addr directly and make the code a little
simpler and more clear.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove XFS_BUF_TO_AGF

Just dereference bp->b_addr directly and make the code a little
simpler and more clear.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove XFS_BUF_TO_AGI

Just dereference bp->b_addr directly and make the code a little
simpler and more clear.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove the xfs_agfl_t typedef

There is just a single user left, so remove it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove the agfl_bno member from struct xfs_agfl

struct xfs_agfl is a header in front of the AGFL entries that exists
for CRC enabled file systems.  For not CRC enabled file systems the AGFL
is simply a list of agbno.  Make the CRC case similar to that by just
using the list behind the new header.  This indirectly solves a problem
with modern gcc versions that warn about taking addresses of packed
structures (and we have to pack the AGFL given that gcc rounds up
structure sizes).  Also replace the helper macro to get from a buffer
with an inline function in xfs_alloc.h to make the code easier to
read.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: clear PF_MEMALLOC before exiting xfsaild thread

Leaving PF_MEMALLOC set when exiting a kthread causes it to remain set
during do_exit().  That can confuse things.  In particular, if BSD
process accounting is enabled, then do_exit() writes data to an
accounting file.  If that file has FS_SYNC_FL set, then this write
occurs synchronously and can misbehave if PF_MEMALLOC is set.

For example, if the accounting file is located on an XFS filesystem,
then a WARN_ON_ONCE() in iomap_do_writepage() is triggered and the data
doesn't get written when it should.  Or if the accounting file is
located on an ext4 filesystem without a journal, then a WARN_ON_ONCE()
in ext4_write_inode() is triggered and the inode doesn't get written.

Fix this in xfsaild() by using the helper functions to save and restore
PF_MEMALLOC.

This can be reproduced as follows in the kvm-xfstests test appliance
modified to add the 'acct' Debian package, and with kvm-xfstests's
recommended kconfig modified to add CONFIG_BSD_PROCESS_ACCT=y:

        mkfs.xfs -f /dev/vdb
        mount /vdb
        touch /vdb/file
        chattr +S /vdb/file
        accton /vdb/file
        mkfs.xfs -f /dev/vdc
        mount /vdc
        umount /vdc

It causes:
WARNING: CPU: 1 PID: 336 at fs/iomap/buffered-io.c:1534
CPU: 1 PID: 336 Comm: xfsaild/vdc Not tainted 5.6.0-rc5 #3
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20191223_100556-anatol 04/01/2014
RIP: 0010:iomap_do_writepage+0x16b/0x1f0 fs/iomap/buffered-io.c:1534
[...]
Call Trace:
 write_cache_pages+0x189/0x4d0 mm/page-writeback.c:2238
 iomap_writepages+0x1c/0x33 fs/iomap/buffered-io.c:1642
 xfs_vm_writepages+0x65/0x90 fs/xfs/xfs_aops.c:578
 do_writepages+0x41/0xe0 mm/page-writeback.c:2344
 __filemap_fdatawrite_range+0xd2/0x120 mm/filemap.c:421
 file_write_and_wait_range+0x71/0xc0 mm/filemap.c:760
 xfs_file_fsync+0x7a/0x2b0 fs/xfs/xfs_file.c:114
 generic_write_sync include/linux/fs.h:2867 [inline]
 xfs_file_buffered_aio_write+0x379/0x3b0 fs/xfs/xfs_file.c:691
 call_write_iter include/linux/fs.h:1901 [inline]
 new_sync_write+0x130/0x1d0 fs/read_write.c:483
 __kernel_write+0x54/0xe0 fs/read_write.c:515
 do_acct_process+0x122/0x170 kernel/acct.c:522
 slow_acct_process kernel/acct.c:581 [inline]
 acct_process+0x1d4/0x27c kernel/acct.c:607
 do_exit+0x83d/0xbc0 kernel/exit.c:791
 kthread+0xf1/0x140 kernel/kthread.c:257
 ret_from_fork+0x27/0x50 arch/x86/entry/entry_64.S:352

This bug was originally reported by syzbot at
https://lore.kernel.org/r/0000000000000e7156059f751d7b@google.com.

Reported-by: syzbot+1f9dc49e8de2582d90c2@syzkaller.appspotmail.com
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: switch xfs_attrmulti_attr_get to lazy attr buffer allocation

Let the low-level attr code only allocate the needed buffer size
for xfs_attrmulti_attr_get instead of allocating the upper bound
at the top of the call chain.

Suggested-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: only allocate the buffer size actually needed in __xfs_set_acl

No need to allocate the max size if we can just allocate the easily
known actual ACL size.

Suggested-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: clean up bufsize alignment in xfs_ioc_attr_list

Use the round_down macro, and use the size of the uint32 type we
use in the callback that fills the buffer to make the code a little
more clear - the size of it is always the same as int for platforms
that Linux runs on.

Suggested-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: embedded the attrlist cursor into struct xfs_attr_list_context

The attrlist cursor only exists as part of an attr list context, so
embedd the structure instead of pointing to it.  Also give it a proper
xfs_ prefix and remove the obsolete typedef.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: remove XFS_DA_OP_INCOMPLETE

Now that we use the on-disk flags field also for the interface to the
lower level attr routines we can use the XFS_ATTR_INCOMPLETE definition
from the on-disk format directly instead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: clean up the attr flag confusion

The ATTR_* flags have a long IRIX history, where they a userspace
interface, the on-disk format and an internal interface.  We've split
out the on-disk interface to the XFS_ATTR_* values, but despite (or
because?) of that the flag have still been a mess.  Switch the
internal interface to pass the on-disk XFS_ATTR_* flags for the
namespace and the Linux XATTR_* flags for the actual flags instead.
The ATTR_* values that are actually used are move to xfs_fs.h with a
new XFS_IOC_* prefix to not conflict with the userspace version that
has the same name and must have the same value.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: clean up the ATTR_REPLACE checks

Remove superflous braces, elses after return statements and use a goto
label to merge common error handling.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: improve xfs_forget_acl

Move the function to xfs_acl.c and provide a proper stub for the
!CONFIG_XFS_POSIX_ACL case.  Lift the flags check to the caller as it
nicely fits in there.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: lift cursor copy in/out into xfs_ioc_attr_list

Lift the common code to copy the cursor from and to user space into
xfs_ioc_attr_list.  Note that this means we copy in twice now as
the cursor is in the middle of the conaining structure, but we never
touch the memory for the original copy.  Doing so keeps the cursor
handling isolated in the common helper.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: lift buffer allocation into xfs_ioc_attr_list

Lift the buffer allocation from the two callers into xfs_ioc_attr_list.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: lift common checks into xfs_ioc_attr_list

Lift the flags and bufsize checks from both callers into the common code
in xfs_ioc_attr_list.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: rename xfs_attr_list_int to xfs_attr_list

The version taking the context structure is the main interface to list
attributes, so drop the _int postfix.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Rajendra <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>