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

[Mike Snitzer]

On Tue, Aug 17 2010 at 12:51pm -0400,
Tejun Heo <tj at kernel.org> wrote:

> Hello,
> 
> On 08/17/2010 04:07 PM, Mike Snitzer wrote:
> >> 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
> > recently.
> 
> Sure thing.  The block part isn't fixed yet and so the RFC tag.  Once
> the block layer part is settled, it probably should be pulled into
> dm/md and other trees and conversions should happen there.

Why base your work on a partial 2.6.36 tree?  Why not rebase to linus'
2.6.36-rc1?

Once we get the changes in a more suitable state (across the entire
tree) we can split the individual changes out to their respective
trees.  Without a comprehensive tree I fear this code isn't going to get
tested or reviewed properly.

For example: any review of DM changes, against stale DM code, is wasted
effort.

> >> * 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
> 
> Oh, I was talking about the other way around.  Passing REQ_FUA in
> bio->bi_rw down to member request_queues.  Sometimes while
> constructing clone / split bios, the bit is lost (e.g. md raid5).

Seems we need to change __blk_rq_prep_clone to propagate REQ_FUA just
like REQ_DISCARD: http://git.kernel.org/linus/3383977

Doesn't seem like we need to do the same for REQ_FLUSH (at least not for
rq-based DM's benefit) because dm.c:setup_clone already special cases
flush requests and sets REQ_FLUSH in cmd_flags.

Mike