hfs: rework MDB locking scheme

hfs_mdb_commit() used to hold the primary MDB buffer_head (mdb_bh)
locked for the whole commit, including writing the alternate MDB and
the volume bitmap. A corrupted image can set drVBMSt so the bitmap
block aliases mdb_bh. The bitmap writeback path then calls
lock_buffer() on that same buffer_head while it is already locked by
hfs_mdb_commit(). Finally, we have a deadlock during flushing the MDB
to the file system volume. However, even for valid images, the locking
scheme of holding the mdb_bh locked across this much unrelated I/O is
fragile anyway.

This patch adds a dedicated sbi->mdb_lock and take it around every
hfs_mdb_commit() caller (hfs_sync_fs(), flush_mdb(), and the initial
hfs_mdb_get() in hfs_fill_super()).

Additionally, this patch makes sbi->mdb/sbi->alt_mdb independent
in-memory copies (allocated with kmemdup()) instead of pointers into
mdb_bh's/alt_mdb_bh's page cache data. Every MDB field read or write in
hfs_mdb_commit(), hfs_mdb_close() and hfs_mdb_get() now operates on
these copies. Also, hfs_mdb_publish() and hfs_alt_mdb_publish() are
the only places left that touch the buffer_heads. The hfs_mdb_put()
frees the two copies.

Reported-by: Yue Sun <samsun1006219@gmail.com>
Link: https://lore.kernel.org/all/CAEkJfYMB47v1yOWHB8q2dc8kf=uj-rLO=+yMyudwPguJ8Kd3jA@mail.gmail.com/
Signed-off-by: Yue Sun <samsun1006219@gmail.com>
cc: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
cc: Yangtao Li <frank.li@vivo.com>
cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Viacheslav Dubeyko <slava@dubeyko.com>
Link: https://lore.kernel.org/r/20260720184414.195213-2-slava@dubeyko.com
Signed-off-by: Viacheslav Dubeyko <slava@dubeyko.com>
This commit is contained in:
Viacheslav Dubeyko 2026-07-20 11:44:15 -07:00
commit f26da7e038
3 changed files with 87 additions and 23 deletions