David Chinner wrote:
you are understanding barriers to be the same as syncronous writes. (and
therefor the data is on persistant media before the call returns)
No, I'm describing the high level behaviour that is expected by
a filesystem. The reasons for this are below....
You say no, but then you go on to contradict yourself below.
Ok, that's my understanding of how *device based barriers* can work,
but there's more to it than that. As far as the filesystem is
concerned the barrier write needs to *behave* exactly like a sync
write because of the guarantees the filesystem has to provide
userspace. Specifically - sync, sync writes and fsync.
There, you just ascribed the synchronous property to barrier requests.
This is false. Barriers are about ordering, synchronous writes are
another thing entirely. The filesystem is supposed to use barriers to
maintain ordering for journal data. If you are trying to handle a
synchronous write request, that's another flag.
This is the big problem, right? If we use barriers for commit
writes, the filesystem can return to userspace after a sync write or
fsync() and an *ordered barrier device implementation* may not have
written the blocks to persistent media. If we then pull the plug on
the box, we've just lost data that sync or fsync said was
successfully on disk. That's BAD.
That's why for synchronous writes, you set the flag to mark the request
as synchronous, which has nothing at all to do with barriers. You are
trying to use barriers to solve two different problems. Use one flag to
indicate ordering, and another to indicate synchronisity.
Right now a barrier write on the last block of the fsync/sync write
is sufficient to prevent that because of the FUA on the barrier
block write. A purely ordered barrier implementation does not
provide this guarantee.
This is a side effect of the implementation of the barrier, not part of
the semantics of barriers, so you shouldn't rely on this behavior. You
don't have to use FUA to handle the barrier request, and if you don't,
then the request can be completed while the data is still in the write
cache. You just have to make sure to flush it before any subsequent
requests.
IOWs, there are two parts to the problem:
1 - guaranteeing I/O ordering
2 - guaranteeing blocks are on persistent storage.
Right now, a single barrier I/O is used to provide both of these
guarantees. In most cases, all we really need to provide is 1); the
need for 2) is a much rarer condition but still needs to be
provided.
Yep... two problems... two flags.
Yes, if we define a barrier to only guarantee 1), then yes this
would be a big win (esp. for XFS). But that requires all filesystems
to handle sync writes differently, and sync_blockdev() needs to
call blkdev_issue_flush() as well....
So, what do we do here? Do we define a barrier I/O to only provide
ordering, or do we define it to also provide persistent storage
writeback? Whatever we decide, it needs to be documented....
We do the former or we end up in the same boat as O_DIRECT; where you
have one flag that means several things, and no way to specify you only
need some of those and not the others.
-
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
