xfs: always free xattri_leaf_bp when cancelling a deferred op

While running the following fstest with logged xattrs DISabled, I
noticed the following:

# FSSTRESS_AVOID="-z -f unlink=1 -f rmdir=1 -f creat=2 -f mkdir=2 -f
getfattr=3 -f listfattr=3 -f attr_remove=4 -f removefattr=4 -f
setfattr=20 -f attr_set=60" ./check generic/475

INFO: task u9:1:40 blocked for more than 61 seconds.
      Tainted: G           O      5.19.0-rc2-djwx #rc2
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
task:u9:1            state:D stack:12872 pid:   40 ppid:     2 flags:0x00004000
Workqueue: xfs-cil/dm-0 xlog_cil_push_work [xfs]
Call Trace:
 <TASK>
 __schedule+0x2db/0x1110
 schedule+0x58/0xc0
 schedule_timeout+0x115/0x160
 __down_common+0x126/0x210
 down+0x54/0x70
 xfs_buf_lock+0x2d/0xe0 [xfs 0532c1cb1d67dd81d15cb79ac6e415c8dec58f73]
 xfs_buf_item_unpin+0x227/0x3a0 [xfs 0532c1cb1d67dd81d15cb79ac6e415c8dec58f73]
 xfs_trans_committed_bulk+0x18e/0x320 [xfs 0532c1cb1d67dd81d15cb79ac6e415c8dec58f73]
 xlog_cil_committed+0x2ea/0x360 [xfs 0532c1cb1d67dd81d15cb79ac6e415c8dec58f73]
 xlog_cil_push_work+0x60f/0x690 [xfs 0532c1cb1d67dd81d15cb79ac6e415c8dec58f73]
 process_one_work+0x1df/0x3c0
 worker_thread+0x53/0x3b0
 kthread+0xea/0x110
 ret_from_fork+0x1f/0x30
 </TASK>

This appears to be the result of shortform_to_leaf creating a new leaf
buffer as part of adding an xattr to a file.  The new leaf buffer is
held and attached to the xfs_attr_intent structure, but then the
filesystem shuts down.  Instead of the usual path (which adds the attr
to the held leaf buffer which releases the hold), we instead cancel the
entire deferred operation.

Unfortunately, xfs_attr_cancel_item doesn't release any attached leaf
buffers, so we leak the locked buffer.  The CIL cannot do anything
about that, and hangs.  Fix this by teaching it to release leaf buffers,
and make XFS a little more careful about not leaving a dangling
reference.

The prologue of xfs_attri_item_recover is (in this author's opinion) a
little hard to figure out, so I'll clean that up in the next patch.

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

fs/xfs/xfs_attr_item.c

@@ -455,6 +455,8 @@ static inline void
 xfs_attr_free_item(
 	struct xfs_attr_intent		*attr)
 {
+	ASSERT(attr->xattri_leaf_bp == NULL);
+
 	if (attr->xattri_da_state)
 		xfs_da_state_free(attr->xattri_da_state);
 	xfs_attri_log_nameval_put(attr->xattri_nameval);
@@ -509,6 +511,10 @@ xfs_attr_cancel_item(
 	struct xfs_attr_intent		*attr;
 
 	attr = container_of(item, struct xfs_attr_intent, xattri_list);
+	if (attr->xattri_leaf_bp) {
+		xfs_buf_relse(attr->xattri_leaf_bp);
+		attr->xattri_leaf_bp = NULL;
+	}
 	xfs_attr_free_item(attr);
 }
 
@@ -670,9 +676,21 @@ xfs_attri_item_recover(
 	error = xfs_defer_ops_capture_and_commit(tp, capture_list);
 
 out_unlock:
-	if (attr->xattri_leaf_bp)
+	if (attr->xattri_leaf_bp) {
 		xfs_buf_relse(attr->xattri_leaf_bp);
 
+		/*
+		 * If there's more work to do to complete the attr intent, the
+		 * defer capture structure will have taken its own reference to
+		 * the attr leaf buffer and will give that to the continuation
+		 * transaction.  The attr intent struct drives the continuation
+		 * work, so release our refcount on the attr leaf buffer but
+		 * retain the pointer in the intent structure.
+		 */
+		if (ret != -EAGAIN)
+			attr->xattri_leaf_bp = NULL;
+	}
+
 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 	xfs_irele(ip);
 out: