On Tue, May 20 2008, Jamie Lokier wrote:

Jens Axboe wrote:

quoted

On Tue, May 20 2008, Jamie Lokier wrote:

quoted

Does WRITE_BARRIER always cause a flush?  It does not have to
according to Documentation/block/barrier.txt.  There are caveats about
tagged queuing "not yet implemented" in the text, but can we rely on
that?  The documentation is older than the current implementation;
those caveats might no longer apply.

It does, if you use ordered tags then that assumes write through
caching (or ordered tag + drain + flush after completion).

Oh.  That's really unclear from the opening paragraph of barrier.txt,
which _defines_ what I/O barriers are for, and does not mention flushing:

   I/O barrier requests are used to guarantee ordering around the barrier
   requests.  Unless you're crazy enough to use disk drives for
   implementing synchronization constructs (wow, sounds interesting...),
   the ordering is meaningful only for write requests for things like
   journal checkpoints.  All requests queued before a barrier request
   must be finished (made it to the physical medium) before the barrier
   request is started, and all requests queued after the barrier request
   must be started only after the barrier request is finished (again,
   made it to the physical medium).

So I assumed the reason flush is talked about later was only because
most devices don't offer an alternative.

It may not mention flushing explicitly, but it does not have to since
the flushing is one way to implement what the above describes. Note how
it says that request must have made it to physical medium before
allowing others to continue? That means you have to either write through
or flush caches, otherwise you cannot make that guarentee.

Later in barrier.txt, in the section about flushing, it says:

   the reason you use I/O barriers is mainly to protect filesystem
   integrity when power failure or some other events abruptly stop the
   drive from operating and possibly make the drive lose data in its
   cache.  So, I/O barriers need to guarantee that requests actually
   get written to non-volatile medium in order

Woa!  Nothing about flushing being important, just "to guarantee
... in order".

Thus flushing looks like an implementation detail - all we could do at
the time.  It does not seem to be the _point_ of WRITE_BARRIER
(according to the text), which is to ensure journalling integrity by
ordering writes.

Yeah, that is precisely what it is and why it does not mention flushing
explicitly!

Really, the main reason I was confused was that I imagine some
SCSI-like devices letting you do partially ordered writes to
write-back cache - with their cache preserving ordering constraints
the same way as some CPU or database caches.  (Perhaps I've been
thinking about CPUs too long!)

Right, that is what ordered tags give you. But if you have write back
caching enabled, then you get a completion event before the data is
actually on disk. Perhaps that is OK for some cases, perhaps not. The
Linux barrier implementation has always guarenteed that the data is
actually on platter before considering a barrier write done, as
described in the text you quote.

Anyway, moving on....  Let's admit I am wrong about that :-)

And get back to my idea.  Ignoring actual disks for a moment ( ;-) ),
there are some I/O scheduling optimisations possible in the kernel
itself by distinguishing between barriers (for journalling) and
flushes (for fsync).

Basically, barriers can be moved around between ordered writes,
including postponing indefinitely (i.e. a "needs barrier" flag).
Unordered writes (in-place data?) can be reordered somewhat around
barriers and other writes.  Nobody should wait for a barrier to complete.

On the other hand, flushes must be issued soon, and fdatasync/fsync
wait for the result.  Reordering possibilities are different: all
writes can be moved before a flush (unless it's a barrier too), and
in-place data writes cannot be moved after one.

Both barriers and flushes can be merged if there are no intervening
writes except unordered writes.  Double flushing, e.g. calling fsync
twice, or calling blkdev_issue_flush just to be sure somewhere,
shouldn't have any overhead.

The request elevator seems a good place to apply those heuristics.
I've written earlier about how to remove some barriers from ext3/4
journalling.  This stuff seems to suggest even more I/O scheduling
optimisations with tree-like journalling (as in BTRFS?).

There are some good ideas in there, I'll punt that to the fs people.
Ignoring actual disk drives makes things easier :-). While SCSI has
ordered IO with write back caching, SATA does not. You basically have to
drain and flush there, or use one of the other variants for getting data
in disk - see blkdev.h:

        /*
         * Hardbarrier is supported with one of the following methods.
         *
         * NONE         : hardbarrier unsupported
         * DRAIN        : ordering by draining is enough
         * DRAIN_FLUSH  : ordering by draining w/ pre and post flushes
         * DRAIN_FUA    : ordering by draining w/ pre flush and FUA
         * write
         * TAG          : ordering by tag is enough
         * TAG_FLUSH    : ordering by tag w/ pre and post flushes
         * TAG_FUA      : ordering by tag w/ pre flush and FUA write

-- 
Jens Axboe