Jan Kara [off-list ref] writes:

  Hello,

On Fri 27-01-12 16:15:49, Jeff Moyer wrote:

quoted

As it stands, generic_file_aio_write will call into generic_write_sync
when -EIOCBQUEUED is returned from __generic_file_aio_write.  EIOCBQUEUED
indicates that an I/O was submitted but NOT completed.  Thus, we will
flush the disk cache, potentially before the write(s) even make it to
the disk!

  Yeah. It seems to be a problem introduced by Tejun's rewrite of barrier
code, right? Before that we'd drain the IO queue when cache flush is issued
and thus effectively wait for IO completion...

Right, though hch seems to think even then the problem existed.

quoted

 Up until now, this has been the best we could do, as file
systems didn't bother to flush the disk cache after an O_SYNC AIO+DIO
write.  After applying the prior two patches to xfs and ext4, at least
the major two file systems do the right thing.  So, let's go ahead and
fix this backwards logic.

  But doesn't this break filesystems which you didn't fix explicitely even
more than they were? You are right they might have sent cache flush too
early but they'd at least propely force all metadata modifications (e.g.
from allocation) to disk. But after this patch O_SYNC will have simply no
effect for these filesystems.

Yep.  Note that we're calling into generic_write_sync with a negative
value.  I followed that call chain all the way down and convinced myself
that it was "mostly harmless," but it sure as heck ain't right.  I'll
audit other file systems to see whether it's a problem.  btrfs, at
least, isn't affected by this.

Also I was thinking whether we couldn't implement the fix in VFS. Basically
it would be the same like the fix for ext4. Like having a per-sb workqueue
and queue work calling generic_write_sync() from end_io handler when the
file is O_SYNC? That would solve the issue for all filesystems...

Well, that would require buy-in from the other file system developers.
What do the XFS folks think?

Cheers,
Jeff