[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [dm-devel] [PATCH 5/5] dm: implement REQ_FLUSH/FUA support

On Tue, Aug 17 2010 at  5:33am -0400,
Tejun Heo <tj kernel org> wrote:

> Hello,
> On 08/16/2010 09:02 PM, Mike Snitzer wrote:
> > On Mon, Aug 16 2010 at 12:52pm -0400,
> > Tejun Heo <tj kernel org> wrote:
> > 
> >> From: Tejun Heo <tj kernle org>
> >>
> >> This patch converts dm to support REQ_FLUSH/FUA instead of now
> >> deprecated REQ_HARDBARRIER.
> > 
> > What tree does this patch apply to?  I know it doesn't apply to
> > v2.6.36-rc1, e.g.: http://git.kernel.org/linus/708e929513502fb0
> (from the head message)
> These patches are on top of
>   block#for-2.6.36-post (c047ab2dddeeafbd6f7c00e45a13a5c4da53ea0b)
> + block-replace-barrier-with-sequenced-flush patchset[1]
> + block-fix-incorrect-bio-request-flag-conversion-in-md patch[2]
> and available in the following git tree.
>   git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git flush-fua
> [1] http://thread.gmane.org/gmane.linux.kernel/1022363
> [2] http://thread.gmane.org/gmane.linux.kernel/1023435
> Probably fetching the git tree is the easist way to review?

OK, I missed this info because I just looked at the DM patch.
> >> For bio-based dm,
> >> * -EOPNOTSUPP retry logic dropped.
> >
> > That logic wasn't just about retries (at least not in the latest
> > kernel).  With commit 708e929513502fb0 the -EOPNOTSUPP checking also
> > serves to optimize the barrier+discard case (when discards aren't
> > supported).
> With the patch applied, there's no second flush.  Those requests would
> now be REQ_FLUSH + REQ_DISCARD.  The first can't be avoided anyway and
> there won't be the second flush to begin with, so I don't think this
> worsens anything.

Makes sense, but your patches still need to be refreshed against the
latest (2.6.36-rc1) upstream code.  Numerous changes went in to DM

> >> * Nothing much changes.  It just needs to handle FLUSH requests as
> >>   before.  It would be beneficial to advertise FUA capability so that
> >>   it can propagate FUA flags down to member request_queues instead of
> >>   sequencing it as WRITE + FLUSH at the top queue.
> > 
> > Can you expand on that TODO a bit?  What is the mechanism to propagate
> > FUA down to a DM device's members?  I'm only aware of propagating member
> > devices' features up to the top-level DM device's request-queue (not the
> > opposite).
> > 
> > Are you saying that establishing the FUA capability on the top-level DM
> > device's request_queue is sufficient?  If so then why not make the
> > change?
> Yeah, I think it would be enough to always advertise FLUSH|FUA if the
> member devices support FLUSH (regardless of FUA support).  The reason
> why I didn't do it was, umm, laziness, I suppose.

I don't buy it.. you're far from lazy! ;)

> >> Lightly tested linear, stripe, raid1, snap and crypt targets.  Please
> >> proceed with caution as I'm not familiar with the code base.
> > 
> > This is concerning...
> Yeap, I want you to be concerned. :-) This was the first time I looked
> at the dm code and there are many different disjoint code paths and I
> couldn't fully follow or test all of them, so it definitely needs a
> careful review from someone who understands the whole thing.

You'll need Mikulas (bio-based) and NEC (request-based, Kiyoshi and
Jun'ichi) to give it serious review.

NOTE: NEC has already given some preliminary feedback to hch in the
"[PATCH, RFC 2/2] dm: support REQ_FLUSH directly" thread:

> > if we're to offer more comprehensive review I think we need more
> > detail on what guided your changes rather than details of what the
> > resulting changes are.
> I'll try to explain it.  If you have any further questions, please let
> me know.

Thanks for the additional details.

> * For bio based dm:
>   * Unlike REQ_HARDBARRIER, REQ_FLUSH/FUA doesn't have any ordering
>     requirements.  Remove assumptions of ordering and/or draining.
>     A related question: Is dm_wait_for_completion() used in
>     process_flush() safe against starvation under continuous influx of
>     other commands?

OK, so you folded dm_flush() directly into process_flush() -- the code
that was dm_flush() only needs to be called once now.

As for your specific dm_wait_for_completion() concern -- I'll defer to
Mikulas.  But I'll add: we haven't had any reported starvation issues
with DM's existing barrier support.  DM uses a mempool for its clones,
so it should naturally throttle (without starvation) when memory gets

>   * As REQ_FLUSH/FUA doesn't require any ordering of requests before
>     or after it, on array devices, the latter part - REQ_FUA - can be
>     handled like other writes.  ie. REQ_FLUSH needs to be broadcasted
>     to all devices but once that is complete the data/REQ_FUA bio can
>     be sent to only the affected devices.  This needs some care as
>     there are bio cloning/splitting code paths where REQ_FUA bit isn't
>     preserved.
>   * Guarantee that REQ_FLUSH w/ data never reaches targets (this in
>     part is to put it in alignment with request based dm).

bio-based DM already split the barrier out from the data (in
process_barrier).  You've renamed process_barrier to process_flush and
added the REQ_FLUSH logic like I'd expect.

> * For request based dm:
>   * The sequencing is done by the block layer for the top level
>     request_queue, so the only things request based dm needs to make
>     sure is 1. handling empty REQ_FLUSH correctly (block layer will
>     only send down empty REQ_FLUSHes) and 2. propagate REQ_FUA bit to
>     member devices.

OK, so seems 1 is done, 2 is still TODO.  Looking at your tree it seems
2 would be as simple as using the following in
dm_init_request_based_queue (on the most current upstream dm.c):

blk_queue_flush(q, REQ_FLUSH | REQ_FUA);

(your current patch only sets REQ_FLUSH in alloc_dev).

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]