[lvc-project] [PATCH v3] ext4: fix i_data_sem unlock order in ext4_ind_migrate()

[Mikhail Ukhin]

Fuzzing reports a possible deadlock in jbd2_log_wait_commit.

This issue is triggered when an EXT4_IOC_MIGRATE ioctl is set to require
synchronous updates because the file descriptor is opened with O_SYNC.
This can lead to the jbd2_journal_stop() function calling
jbd2_might_wait_for_commit(), potentially causing a deadlock if the
EXT4_IOC_MIGRATE call races with a write(2) system call.

This problem only arises when CONFIG_PROVE_LOCKING is enabled. In this
case, the jbd2_might_wait_for_commit macro locks jbd2_handle in the
jbd2_journal_stop function while i_data_sem is locked. This triggers
lockdep because the jbd2_journal_start function might also lock the same
jbd2_handle simultaneously.

Found by Linux Verification Center (linuxtesting.org) with syzkaller.

Reviewed-by: Ritesh Harjani (IBM) <ritesh.list at gmail.com>
Signed-off-by: Artem Sadovnikov <ancowi69 at gmail.com>
Signed-off-by: Mikhail Ukhin <mish.uxin2012 at yandex.ru>
---
 v2: New addresses have been added and Ritesh Harjani has been noted as a
 reviewer.
 v3: Description updated.
 fs/hugetlbfs/inode.c | 2 +-
 1 file changed, 1 insertions(+), 1 deletions(-)
 
diff --git a/fs/ext4/migrate.c b/fs/ext4/migrate.c
index b0ea646454ac..59290356aa5b 100644
--- a/fs/ext4/migrate.c
+++ b/fs/ext4/migrate.c
@@ -663,8 +663,8 @@ int ext4_ind_migrate(struct inode *inode)
 	if (unlikely(ret2 && !ret))
 		ret = ret2;
 errout:
-	ext4_journal_stop(handle);
 	up_write(&EXT4_I(inode)->i_data_sem);
+	ext4_journal_stop(handle);
 out_unlock:
 	percpu_up_write(&sbi->s_writepages_rwsem);
 	return ret;
 	
-- 
2.25.1