Skip to contentSkip to conversation

Ext4: wait for log to commit when putting super

PATCH In review Josef Bacik 3 participants 5 messages Dec 7, 2009 ext4 file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Ext4: wait for log to commit when putting super

There is a problem where a transaction will be committing while we're unmounting the filesystem and you will get a panic because EXT4_SB(sb)->s_group_info has been kfree'ed in ext4_put_super. The commit code does the callback for the mballoc stuff to release free'ed blocks in the transaction and panic's trying to access s_group_info. The fix is to wait for the transaction to finish committing before we start cleaning up the mballoc stuff. This patch hasn't been tested yet, but its an obvious fix.

---
 fs/ext4/super.c |    4 ++++
 1 files changed, 4 insertions(+), 0 deletions(-)
fs/ext4/super.c +4 -0
index d4ca92a..ddd115f 100644
@@ -593,6 +593,7 @@ static void ext4_put_super(struct super_block *sb)
 {
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
 	struct ext4_super_block *es = sbi->s_es;
+	tid_t target;
 	int i, err;
 
 	flush_workqueue(sbi->dio_unwritten_wq);
@@ -603,6 +604,9 @@ static void ext4_put_super(struct super_block *sb)
 	if (sb->s_dirt)
 		ext4_commit_super(sb, 1);
 
+	if (jbd2_journal_start_commit(EXT4_SB(sb)->s_journal, &target))
+		jbd2_log_wait_commit(EXT4_SB(sb)->s_journal, target);
+
 	ext4_release_system_zone(sb);
 	ext4_mb_release(sb);
 	ext4_ext_release(sb);
Signature · 1 line
1.6.2.5
#1 lore raw
Quoted 7 lines · Josef Bacik

There is a problem where a transaction will be committing while we're unmounting the filesystem and you will get a panic because EXT4_SB(sb)->s_group_info has been kfree'ed in ext4_put_super. The commit code does the callback for the mballoc stuff to release free'ed blocks in the transaction and panic's trying to access s_group_info. The fix is to wait for the transaction to finish committing before we start cleaning up the mballoc stuff. This patch hasn't been tested yet, but its an obvious fix.

Hm, doesn't jbd2_journal_destroy already do that, but we just call it -after- we've done ext4_mb_release which leads to freeing EXT4_SB(sb)->s_group_info ?

Can we just do jbd2_journal_destroy() earlier before we tear down things it may depend on?

-Eric

Quoted 27 lines
---
 fs/ext4/super.c |    4 ++++
 1 files changed, 4 insertions(+), 0 deletions(-)
fs/ext4/super.c +4 -0
index d4ca92a..ddd115f 100644
@@ -593,6 +593,7 @@ static void ext4_put_super(struct super_block *sb)
 {
 	struct ext4_sb_info *sbi = EXT4_SB(sb);
 	struct ext4_super_block *es = sbi->s_es;
+	tid_t target;
 	int i, err;
 
 	flush_workqueue(sbi->dio_unwritten_wq);
@@ -603,6 +604,9 @@ static void ext4_put_super(struct super_block *sb)
 	if (sb->s_dirt)
 		ext4_commit_super(sb, 1);
 
+	if (jbd2_journal_start_commit(EXT4_SB(sb)->s_journal, &target))
+		jbd2_log_wait_commit(EXT4_SB(sb)->s_journal, target);
+
 	ext4_release_system_zone(sb);
 	ext4_mb_release(sb);
 	ext4_ext_release(sb);
#2 lore raw
Quoted 9 lines · Eric Sandeen
Quoted 7 lines · Josef Bacik

There is a problem where a transaction will be committing while we're unmounting the filesystem and you will get a panic because EXT4_SB(sb)->s_group_info has been kfree'ed in ext4_put_super. The commit code does the callback for the mballoc stuff to release free'ed blocks in the transaction and panic's trying to access s_group_info. The fix is to wait for the transaction to finish committing before we start cleaning up the mballoc stuff. This patch hasn't been tested yet, but its an obvious fix.

Quoted 7 lines

Hm, doesn't jbd2_journal_destroy already do that, but we just call it -after- we've done ext4_mb_release which leads to freeing EXT4_SB(sb)->s_group_info ?

Can we just do jbd2_journal_destroy() earlier before we tear down things it may depend on?

That seems reasonable. We do

        ext4_release_system_zone(sb);
        ext4_mb_release(sb);
        ext4_ext_release(sb);
        ext4_xattr_put_super(sb);

before the journal_destroy, and all thats doing is cleaning up our internal stuff. If we want to keep it close to what ext3 does we can just move

        ext4_release_system_zone(sb);
        ext4_mb_release(sb);
        ext4_ext_release(sb);

under the journal_destroy or just move the journal destroy above all of that stuff. Whichever you prefer. Thanks,

Josef

#3 lore raw
Quoted 27 lines · Josef Bacik
Quoted 8 lines · Eric Sandeen
Quoted 7 lines · Josef Bacik

There is a problem where a transaction will be committing while we're unmounting the filesystem and you will get a panic because EXT4_SB(sb)->s_group_info has been kfree'ed in ext4_put_super. The commit code does the callback for the mballoc stuff to release free'ed blocks in the transaction and panic's trying to access s_group_info. The fix is to wait for the transaction to finish committing before we start cleaning up the mballoc stuff. This patch hasn't been tested yet, but its an obvious fix.

Quoted 7 lines

Hm, doesn't jbd2_journal_destroy already do that, but we just call it -after- we've done ext4_mb_release which leads to freeing EXT4_SB(sb)->s_group_info ?

Can we just do jbd2_journal_destroy() earlier before we tear down things it may depend on?

That seems reasonable. We do

        ext4_release_system_zone(sb);
        ext4_mb_release(sb);
        ext4_ext_release(sb);
        ext4_xattr_put_super(sb);

before the journal_destroy, and all thats doing is cleaning up our internal stuff. If we want to keep it close to what ext3 does we can just move

Quoted 6 lines
        ext4_release_system_zone(sb);
        ext4_mb_release(sb);
        ext4_ext_release(sb);

under the journal_destroy or just move the journal destroy above all of that stuff. Whichever you prefer. Thanks,

I guess it doesn't matter much to me, I just preferred rearranging teardown order to adding more logic that I think is already taken care of in the jbd2_journal_destroy call ..

Thanks for finding this! :)

-Eric

Quoted 1 line

Josef

#4 lore raw
Quoted 4 lines · Eric Sandeen

I guess it doesn't matter much to me, I just preferred rearranging teardown order to adding more logic that I think is already taken care of in the jbd2_journal_destroy call ..

Agreed.

Quoted 1 line

Thanks for finding this! :)

... and seconded! :-)

- Ted