xfs: fix mount hang during primary superblock recovery failure

When mounting an image containing a log with sb modifications that require
log replay, the mount process hang all the time and stack as follows:

  [root@localhost ~]# cat /proc/557/stack
  [<0>] xfs_buftarg_wait+0x31/0x70
  [<0>] xfs_buftarg_drain+0x54/0x350
  [<0>] xfs_mountfs+0x66e/0xe80
  [<0>] xfs_fs_fill_super+0x7f1/0xec0
  [<0>] get_tree_bdev_flags+0x186/0x280
  [<0>] get_tree_bdev+0x18/0x30
  [<0>] xfs_fs_get_tree+0x1d/0x30
  [<0>] vfs_get_tree+0x2d/0x110
  [<0>] path_mount+0xb59/0xfc0
  [<0>] do_mount+0x92/0xc0
  [<0>] __x64_sys_mount+0xc2/0x160
  [<0>] x64_sys_call+0x2de4/0x45c0
  [<0>] do_syscall_64+0xa7/0x240
  [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e

During log recovery, while updating the in-memory superblock from the
primary SB buffer, if an error is encountered, such as superblock
corruption occurs or some other reasons, we will proceed to out_release
and release the xfs_buf. However, this is insufficient because the
xfs_buf's log item has already been initialized and the xfs_buf is held
by the buffer log item as follows, the xfs_buf will not be released,
causing the mount thread to hang.

  xlog_recover_do_primary_sb_buffer
    xlog_recover_do_reg_buffer
      xlog_recover_validate_buf_type
        xfs_buf_item_init(bp, mp)

The solution is straightforward, we simply need to allow it to be
handled by the normal buffer write process. The filesystem will be
shutdown before the submission of buffer_list in xlog_do_recovery_pass(),
ensuring the correct release of the xfs_buf as follows:

  xlog_do_recovery_pass
    error = xlog_recover_process
      xlog_recover_process_data
        xlog_recover_process_ophdr
          xlog_recovery_process_trans
            ...
              xlog_recover_buf_commit_pass2
                error = xlog_recover_do_primary_sb_buffer
                  //Encounter error and return
                if (error)
                  goto out_writebuf
                ...
              out_writebuf:
                xfs_buf_delwri_queue(bp, buffer_list) //add bp to list
                return  error
            ...
    if (!list_empty(&buffer_list))
      if (error)
        xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR); //shutdown first
      xfs_buf_delwri_submit(&buffer_list); //submit buffers in list
        __xfs_buf_submit
          if (bp->b_mount->m_log && xlog_is_shutdown(bp->b_mount->m_log))
            xfs_buf_ioend_fail(bp)  //release bp correctly

Fixes: 6a18765b54e2 ("xfs: update the file system geometry after recoverying superblock buffers")
Cc: stable@vger.kernel.org # v6.12
Signed-off-by: Long Li <leo.lilong@huawei.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Carlos Maiolino <cem@kernel.org>

diff --git a/fs/xfs/xfs_buf_item_recover.c b/fs/xfs/xfs_buf_item_recover.c
index b05d5b81f642daf12f6936d478f8c90ccb875b71..05a2f6927c121bd4adb739ec5c1d9caad3bb031a 100644 (file)
--- a/fs/xfs/xfs_buf_item_recover.c
+++ b/fs/xfs/xfs_buf_item_recover.c
@@ -1087,7 +1087,7 @@ xlog_recover_buf_commit_pass2(
                error = xlog_recover_do_primary_sb_buffer(mp, item, bp, buf_f,
                                current_lsn);
                if (error)
-                       goto out_release;
+                       goto out_writebuf;
 
                /* Update the rt superblock if we have one. */
                if (xfs_has_rtsb(mp) && mp->m_rtsb_bp) {
@@ -1104,6 +1104,15 @@ xlog_recover_buf_commit_pass2(
                xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
        }
 
+       /*
+        * Buffer held by buf log item during 'normal' buffer recovery must
+        * be committed through buffer I/O submission path to ensure proper
+        * release. When error occurs during sb buffer recovery, log shutdown
+        * will be done before submitting buffer list so that buffers can be
+        * released correctly through ioend failure path.
+        */
+out_writebuf:
+
        /*
         * Perform delayed write on the buffer.  Asynchronous writes will be
         * slower when taking into account all the buffers to be flushed.