nilfs2: fix hang in nilfs_lookup_dirty_data_buffers()

Syzbot reported a hang issue in migrate_pages_batch() called by mbind()
and nilfs_lookup_dirty_data_buffers() called in the log writer of nilfs2.

While migrate_pages_batch() locks a folio and waits for the writeback to
complete, the log writer thread that should bring the writeback to
completion picks up the folio being written back in
nilfs_lookup_dirty_data_buffers() that it calls for subsequent log
creation and was trying to lock the folio.  Thus causing a deadlock.

In the first place, it is unexpected that folios/pages in the middle of
writeback will be updated and become dirty.  Nilfs2 adds a checksum to
verify the validity of the log being written and uses it for recovery at
mount, so data changes during writeback are suppressed.  Since this is
broken, an unclean shutdown could potentially cause recovery to fail.

Investigation revealed that the root cause is that the wait for writeback
completion in nilfs_page_mkwrite() is conditional, and if the backing
device does not require stable writes, data may be modified without
waiting.

Fix these issues by making nilfs_page_mkwrite() wait for writeback to
finish regardless of the stable write requirement of the backing device.

Link: https://lkml.kernel.org/r/20240131145657.4209-1-konishi.ryusuke@gmail.com
Fixes: 1d1d1a767206 ("mm: only enforce stable page writes if the backing device requires it")
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@gmail.com>
Reported-by: syzbot+ee2ae68da3b22d04cd8d@syzkaller.appspotmail.com
Closes: https://lkml.kernel.org/r/00000000000047d819061004ad6c@google.com
Tested-by: Ryusuke Konishi <konishi.ryusuke@gmail.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

diff --git a/fs/nilfs2/file.c b/fs/nilfs2/file.c
index bec33b89a075858ebf289a95fa4c83dbf6e86103..0e3fc5ba33c73d7f22deefc1cb68ee8395a1efa4 100644 (file)
--- a/fs/nilfs2/file.c
+++ b/fs/nilfs2/file.c
@@ -107,7 +107,13 @@ static vm_fault_t nilfs_page_mkwrite(struct vm_fault *vmf)
        nilfs_transaction_commit(inode->i_sb);
 
  mapped:
-       folio_wait_stable(folio);
+       /*
+        * Since checksumming including data blocks is performed to determine
+        * the validity of the log to be written and used for recovery, it is
+        * necessary to wait for writeback to finish here, regardless of the
+        * stable write requirement of the backing device.
+        */
+       folio_wait_writeback(folio);
  out:
        sb_end_pagefault(inode->i_sb);
        return vmf_fs_error(ret);