How to fix broken freezing?

Hello,

I was looking at what causes filesystem to have dirty data after it is
frozen. After some thought I realized freezing code is inherently racy and all filesystems (ext3, ext4, xfs) can have dirty data on frozen filesystem.

The race is basically following:

	Task 1					Task 2

freeze_super() __generic_file_aio_write()
... vfs_check_frozen(sb, SB_FREEZE_WRITE)
sb->s_frozen = SB_FREEZE_WRITE;
sync_filesystem(sb);
do the write
/* Here we create dirty data
* which is left on frozen fs */
sb->s_frozen = SB_FREEZE_TRANS;
...
->freeze_fs()

The problem is that you can never make checking for frozen filesystem race-free with the current s_frozen scheme - the filesystem can always be frozen the instant after you check for it and you end up creating dirty data on frozen filesystem.

The question is what to do with this problem. I outline the possibilities that come to my mind below:
1) Ignore the problem - depending on the exact fs details this could lead to fs snapshot being corrupted, also flusher thread can hang on the frozen filesystem (e.g. because of sync(1)) creating all sorts of secondary issues. So I don't think this is really an option.
2) Have a rwlock in the superblock that is held for writing while filesystem freezing is in progress and held for reading by the filesystem while a transaction is running except for transactions that are required to do writeback. This is kind of ugly but at least for ext3/4 relatively easy to implement.

3) Have the same rwlock but already VFS will take the lock in kernel entry points which modify a filesystem. Lot of these places is already guarded by mnt_want_write/mnt_drop_write pair so we could hook into it but there are entry points which use file descriptor and thus are not guarded by mnt_want_write/mnt_drop_write so we would have to modify these places. Note that this in particular also means ioctl calls and such so it won't be trivial to catch all the places. This approach looks the cleanest to me but it's quite some work and it's a bit fragile - requires all people adding an entry point modifying filesystem to think of fs freezing.

What do people think about this? Any idea other idea how to solve the problem?

Honza

Quoted 24 lines · Jan Kara

Hello,

I was looking at what causes filesystem to have dirty data after it is
frozen. After some thought I realized freezing code is inherently racy and all filesystems (ext3, ext4, xfs) can have dirty data on frozen filesystem.

The race is basically following:

	Task 1					Task 2

freeze_super() __generic_file_aio_write()
... vfs_check_frozen(sb, SB_FREEZE_WRITE)
sb->s_frozen = SB_FREEZE_WRITE;
sync_filesystem(sb);
do the write
/* Here we create dirty data
* which is left on frozen fs */
sb->s_frozen = SB_FREEZE_TRANS;
...
->freeze_fs()

The problem is that you can never make checking for frozen filesystem race-free with the current s_frozen scheme - the filesystem can always be frozen the instant after you check for it and you end up creating dirty data on frozen filesystem.

Quoted 11 lines

The question is what to do with this problem. I outline the possibilities that come to my mind below:
1) Ignore the problem - depending on the exact fs details this could lead to fs snapshot being corrupted, also flusher thread can hang on the frozen filesystem (e.g. because of sync(1)) creating all sorts of secondary issues. So I don't think this is really an option.
2) Have a rwlock in the superblock that is held for writing while filesystem freezing is in progress and held for reading by the filesystem while a transaction is running except for transactions that are required to do writeback. This is kind of ugly but at least for ext3/4 relatively easy to implement.

This is as far as I had gotten while independently thinking about it ;)

But talking with dchinner, he had concerns about the scalability of any rwlock, and I think we (ok, mostly Dave) came up with another idea.

What if we had 2 counters in the superblock, one for the equivalent of SB_FREEZE_WRITE and one for SB_FREEZE_TRANS. These would use similar infrastructure to mnt_want_write et al.

Everywhere we currently vfs_check_frozen() we'd have a better-named function which increments the counter, then checks the freeze level. If we are being frozen, we drop the counter & wait. If not frozen, we continue; like this pseudo-code:

void super_freeze_wait(sb, level) {
	while (1) {
		level_ref++;
		if (!frozen(sb, level))
			return;	/* not freezing */
		level_ref--;
		wait_unfrozen(sb, level);
	}
}

There would also be new code to drop the counter when the dirtying is complete.

The freezing functions then just have to wait until the counters hit zero before they can consider themselves done, and freezing is complete. That way if someone sneaks in while the freeze level is being set, they have already notified their intent, and freeze can wait for it anyway before returning.

Quoted 61 lines · Eric Sandeen

Quoted 24 lines · Jan Kara

Hello,

I was looking at what causes filesystem to have dirty data after it is
frozen. After some thought I realized freezing code is inherently racy and all filesystems (ext3, ext4, xfs) can have dirty data on frozen filesystem.

The race is basically following:

	Task 1					Task 2

freeze_super() __generic_file_aio_write()
... vfs_check_frozen(sb, SB_FREEZE_WRITE)
sb->s_frozen = SB_FREEZE_WRITE;
sync_filesystem(sb);
do the write
/* Here we create dirty data
* which is left on frozen fs */
sb->s_frozen = SB_FREEZE_TRANS;
...
->freeze_fs()

The problem is that you can never make checking for frozen filesystem race-free with the current s_frozen scheme - the filesystem can always be frozen the instant after you check for it and you end up creating dirty data on frozen filesystem.

Quoted 11 lines

The question is what to do with this problem. I outline the possibilities that come to my mind below:
1) Ignore the problem - depending on the exact fs details this could lead to fs snapshot being corrupted, also flusher thread can hang on the frozen filesystem (e.g. because of sync(1)) creating all sorts of secondary issues. So I don't think this is really an option.
2) Have a rwlock in the superblock that is held for writing while filesystem freezing is in progress and held for reading by the filesystem while a transaction is running except for transactions that are required to do writeback. This is kind of ugly but at least for ext3/4 relatively easy to implement.

This is as far as I had gotten while independently thinking about it ;)

But talking with dchinner, he had concerns about the scalability of any rwlock, and I think we (ok, mostly Dave) came up with another idea.

What if we had 2 counters in the superblock, one for the equivalent of SB_FREEZE_WRITE and one for SB_FREEZE_TRANS. These would use similar infrastructure to mnt_want_write et al.

Everywhere we currently vfs_check_frozen() we'd have a better-named function which increments the counter, then checks the freeze level. If we are being frozen, we drop the counter & wait. If not frozen, we continue; like this pseudo-code:

void super_freeze_wait(sb, level) {
	while (1) {
		level_ref++;
		if (!frozen(sb, level))
			return;	/* not freezing */
		level_ref--;
		wait_unfrozen(sb, level);
	}
}

Quoted 6 lines

There would also be new code to drop the counter when the dirtying is complete.

The freezing functions then just have to wait until the counters hit zero before they can consider themselves done, and freezing is complete. That way if someone sneaks in while the freeze level is being set, they have already notified their intent, and freeze can wait for it anyway before returning.

Just to clarify, freeze_super would do:

	sb->s_frozen = SB_FREEZE_WRITE;
	smp_wmb();

	while (sb->s_active_write_cnt > 0)
		wait;

	/* no new or existing dirtying writers now, safe to sync */
	sync_filesystem(sb);

	sb->s_frozen = SB_FREEZE_TRANS;
	smp_wmb();

	while (sb->s_active_trans_cnt > 0)
		wait;

	/* no new or existing transactions in progress now, so freeze */
	sb->s_op->freeze_fs(sb);

The counter implemetations will need to scale (e.g. per-cpu counters) and we could probably use a generic waitqueue, but I think this can all be implemented at the superblock level and we only need to call the inc/dec helper functions in the correct places to make it all work.