xfs: xfs_is_shutdown vs xlog_is_shutdown cage fight

I've been chasing a recent resurgence in generic/388 recovery
failure and/or corruption events. The events have largely been
uninitialised inode chunks being tripped over in log recovery
such as:

 XFS (pmem1): User initiated shutdown received.
 pmem1: writeback error on inode 12621949, offset 1019904, sector 12968096
 XFS (pmem1): Log I/O Error (0x6) detected at xfs_fs_goingdown+0xa3/0xf0 (fs/xfs/xfs_fsops.c:500).  Shutting down filesystem.
 XFS (pmem1): Please unmount the filesystem and rectify the problem(s)
 XFS (pmem1): Unmounting Filesystem
 XFS (pmem1): Mounting V5 Filesystem
 XFS (pmem1): Starting recovery (logdev: internal)
 XFS (pmem1): bad inode magic/vsn daddr 8723584 #0 (magic=1818)
 XFS (pmem1): Metadata corruption detected at xfs_inode_buf_verify+0x180/0x190, xfs_inode block 0x851c80 xfs_inode_buf_verify
 XFS (pmem1): Unmount and run xfs_repair
 XFS (pmem1): First 128 bytes of corrupted metadata buffer:
 00000000: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000010: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000020: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000030: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000040: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000050: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000060: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 00000070: 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18 18  ................
 XFS (pmem1): metadata I/O error in "xlog_recover_items_pass2+0x52/0xc0" at daddr 0x851c80 len 32 error 117
 XFS (pmem1): log mount/recovery failed: error -117
 XFS (pmem1): log mount failed

There have been isolated random other issues, too - xfs_repair fails
because it finds some corruption in symlink blocks, rmap
inconsistencies, etc - but they are nowhere near as common as the
uninitialised inode chunk failure.

The problem has clearly happened at runtime before recovery has run;
I can see the ICREATE log item in the log shortly before the
actively recovered range of the log. This means the ICREATE was
definitely created and written to the log, but for some reason the
tail of the log has been moved past the ordered buffer log item that
tracks INODE_ALLOC buffers and, supposedly, prevents the tail of the
log moving past the ICREATE log item before the inode chunk buffer
is written to disk.

Tracing the fsstress processes that are running when the filesystem
shut down immediately pin-pointed the problem:

user shutdown marks xfs_mount as shutdown

         godown-213341 [008]  6398.022871: console:              [ 6397.915392] XFS (pmem1): User initiated shutdown received.
.....

aild tries to push ordered inode cluster buffer

  xfsaild/pmem1-213314 [001]  6398.022974: xfs_buf_trylock:      dev 259:1 daddr 0x851c80 bbcount 0x20 hold 16 pincount 0 lock 0 flags DONE|INODES|PAGES caller xfs_inode_item_push+0x8e
  xfsaild/pmem1-213314 [001]  6398.022976: xfs_ilock_nowait:     dev 259:1 ino 0x851c80 flags ILOCK_SHARED caller xfs_iflush_cluster+0xae

xfs_iflush_cluster() checks xfs_is_shutdown(), returns true,
calls xfs_iflush_abort() to kill writeback of the inode.
Inode is removed from AIL, drops cluster buffer reference.

  xfsaild/pmem1-213314 [001]  6398.022977: xfs_ail_delete:       dev 259:1 lip 0xffff88880247ed80 old lsn 7/20344 new lsn 7/21000 type XFS_LI_INODE flags IN_AIL
  xfsaild/pmem1-213314 [001]  6398.022978: xfs_buf_rele:         dev 259:1 daddr 0x851c80 bbcount 0x20 hold 17 pincount 0 lock 0 flags DONE|INODES|PAGES caller xfs_iflush_abort+0xd7

.....

All inodes on cluster buffer are aborted, then the cluster buffer
itself is aborted and removed from the AIL *without writeback*:

xfsaild/pmem1-213314 [001]  6398.023011: xfs_buf_error_relse:  dev 259:1 daddr 0x851c80 bbcount 0x20 hold 2 pincount 0 lock 0 flags ASYNC|DONE|STALE|INODES|PAGES caller xfs_buf_ioend_fail+0x33
   xfsaild/pmem1-213314 [001]  6398.023012: xfs_ail_delete:       dev 259:1 lip 0xffff8888053efde8 old lsn 7/20344 new lsn 7/20344 type XFS_LI_BUF flags IN_AIL

The inode buffer was at 7/20344 when it was removed from the AIL.

   xfsaild/pmem1-213314 [001]  6398.023012: xfs_buf_item_relse:   dev 259:1 daddr 0x851c80 bbcount 0x20 hold 2 pincount 0 lock 0 flags ASYNC|DONE|STALE|INODES|PAGES caller xfs_buf_item_done+0x31
   xfsaild/pmem1-213314 [001]  6398.023012: xfs_buf_rele:         dev 259:1 daddr 0x851c80 bbcount 0x20 hold 2 pincount 0 lock 0 flags ASYNC|DONE|STALE|INODES|PAGES caller xfs_buf_item_relse+0x39

.....

Userspace is still running, doing stuff. an fsstress process runs
syncfs() or sync() and we end up in sync_fs_one_sb() which issues
a log force. This pushes on the CIL:

        fsstress-213322 [001]  6398.024430: xfs_fs_sync_fs:       dev 259:1 m_features 0x20000000019ff6e9 opstate (clean|shutdown|inodegc|blockgc) s_flags 0x70810000 caller sync_fs_one_sb+0x26
        fsstress-213322 [001]  6398.024430: xfs_log_force:        dev 259:1 lsn 0x0 caller xfs_fs_sync_fs+0x82
        fsstress-213322 [001]  6398.024430: xfs_log_force:        dev 259:1 lsn 0x5f caller xfs_log_force+0x7c
           <...>-194402 [001]  6398.024467: kmem_alloc:           size 176 flags 0x14 caller xlog_cil_push_work+0x9f

And the CIL fills up iclogs with pending changes. This picks up
the current tail from the AIL:

           <...>-194402 [001]  6398.024497: xlog_iclog_get_space: dev 259:1 state XLOG_STATE_ACTIVE refcnt 1 offset 0 lsn 0x0 flags  caller xlog_write+0x149
           <...>-194402 [001]  6398.024498: xlog_iclog_switch:    dev 259:1 state XLOG_STATE_ACTIVE refcnt 1 offset 0 lsn 0x700005408 flags  caller xlog_state_get_iclog_space+0x37e
           <...>-194402 [001]  6398.024521: xlog_iclog_release:   dev 259:1 state XLOG_STATE_WANT_SYNC refcnt 1 offset 32256 lsn 0x700005408 flags  caller xlog_write+0x5f9
           <...>-194402 [001]  6398.024522: xfs_log_assign_tail_lsn: dev 259:1 new tail lsn 7/21000, old lsn 7/20344, last sync 7/21448

And it moves the tail of the log to 7/21000 from 7/20344. This
*moves the tail of the log beyond the ICREATE transaction* that was
at 7/20344 and pinned by the inode cluster buffer that was cancelled
above.

....

         godown-213341 [008]  6398.027005: xfs_force_shutdown:   dev 259:1 tag logerror flags log_io|force_umount file fs/xfs/xfs_fsops.c line_num 500
          godown-213341 [008]  6398.027022: console:              [ 6397.915406] pmem1: writeback error on inode 12621949, offset 1019904, sector 12968096
          godown-213341 [008]  6398.030551: console:              [ 6397.919546] XFS (pmem1): Log I/O Error (0x6) detected at xfs_fs_goingdown+0xa3/0xf0 (fs/

And finally the log itself is now shutdown, stopping all further
writes to the log. But this is too late to prevent the corruption
that moving the tail of the log forwards after we start cancelling
writeback causes.

The fundamental problem here is that we are using the wrong shutdown
checks for log items. We've long conflated mount shutdown with log
shutdown state, and I started separating that recently with the
atomic shutdown state changes in commit b36d4651e165 ("xfs: make
forced shutdown processing atomic"). The changes in that commit
series are directly responsible for being able to diagnose this
issue because it clearly separated mount shutdown from log shutdown.

Essentially, once we start cancelling writeback of log items and
removing them from the AIL because the filesystem is shut down, we
*cannot* update the journal because we may have cancelled the items
that pin the tail of the log. That moves the tail of the log
forwards without having written the metadata back, hence we have
corrupt in memory state and writing to the journal propagates that
to the on-disk state.

What commit b36d4651e165 makes clear is that log item state needs to
change relative to log shutdown, not mount shutdown. IOWs, anything
that aborts metadata writeback needs to check log shutdown state
because log items directly affect log consistency. Having them check
mount shutdown state introduces the above race condition where we
cancel metadata writeback before the log shuts down.

To fix this, this patch works through all log items and converts
shutdown checks to use xlog_is_shutdown() rather than
xfs_is_shutdown(), so that we don't start aborting metadata
writeback before we shut off journal writes.

AFAICT, this race condition is a zero day IO error handling bug in
XFS that dates back to the introduction of XLOG_IO_ERROR,
XLOG_STATE_IOERROR and XFS_FORCED_SHUTDOWN back in January 1997.

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>

1 files changed, 9 insertions, 1 deletions

diff --git a/fs/xfs/xfs_icache.c b/fs/xfs/xfs_icache.c
index 9644f938990c..4148cdf7ce4a 100644
--- a/fs/xfs/xfs_icache.c
+++ b/fs/xfs/xfs_icache.c
@@ -23,6 +23,7 @@
 #include "xfs_reflink.h"
 #include "xfs_ialloc.h"
 #include "xfs_ag.h"
+#include "xfs_log_priv.h"
 
 #include <linux/iversion.h>
 
@@ -873,7 +874,14 @@ xfs_reclaim_inode(
 	if (xfs_iflags_test_and_set(ip, XFS_IFLUSHING))
 		goto out_iunlock;
 
-	if (xfs_is_shutdown(ip->i_mount)) {
+	/*
+	 * Check for log shutdown because aborting the inode can move the log
+	 * tail and corrupt in memory state. This is fine if the log is shut
+	 * down, but if the log is still active and only the mount is shut down
+	 * then the in-memory log tail movement caused by the abort can be
+	 * incorrectly propagated to disk.
+	 */
+	if (xlog_is_shutdown(ip->i_mount->m_log)) {
 		xfs_iunpin_wait(ip);
 		xfs_iflush_abort(ip);
 		goto reclaim;