On Wed, Jan 17, 2018 at 10:37:23PM -0500, Mike Snitzer wrote:
> On Wed, Jan 17 2018 at 10:25pm -0500,
> Ming Lei <ming.lei@xxxxxxxxxx> wrote:
>
> > Hi Mike,
> >
> > On Wed, Jan 17, 2018 at 11:25:57AM -0500, Mike Snitzer wrote:
> > > From: Ming Lei <ming.lei@xxxxxxxxxx>
> > >
> > > blk_insert_cloned_request() is called in the fast path of a dm-rq driver
> > > (e.g. blk-mq request-based DM mpath). blk_insert_cloned_request() uses
> > > blk_mq_request_bypass_insert() to directly append the request to the
> > > blk-mq hctx->dispatch_list of the underlying queue.
> > >
> > > 1) This way isn't efficient enough because the hctx spinlock is always
> > > used.
> > >
> > > 2) With blk_insert_cloned_request(), we completely bypass underlying
> > > queue's elevator and depend on the upper-level dm-rq driver's elevator
> > > to schedule IO. But dm-rq currently can't get the underlying queue's
> > > dispatch feedback at all. Without knowing whether a request was issued
> > > or not (e.g. due to underlying queue being busy) the dm-rq elevator will
> > > not be able to provide effective IO merging (as a side-effect of dm-rq
> > > currently blindly destaging a request from its elevator only to requeue
> > > it after a delay, which kills any opportunity for merging). This
> > > obviously causes very bad sequential IO performance.
> > >
> > > Fix this by updating blk_insert_cloned_request() to use
> > > blk_mq_request_direct_issue(). blk_mq_request_direct_issue() allows a
> > > request to be issued directly to the underlying queue and returns the
> > > dispatch feedback (blk_status_t). If blk_mq_request_direct_issue()
> > > returns BLK_SYS_RESOURCE the dm-rq driver will now use DM_MAPIO_REQUEUE
> > > to _not_ destage the request. Whereby preserving the opportunity to
> > > merge IO.
> > >
> > > With this, request-based DM's blk-mq sequential IO performance is vastly
> > > improved (as much as 3X in mpath/virtio-scsi testing).
> > >
> > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
> > > [blk-mq.c changes heavily influenced by Ming Lei's initial solution, but
> > > they were refactored to make them less fragile and easier to read/review]
> > > Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx>
> ...
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index c117c2baf2c9..f5f0d8456713 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -1731,15 +1731,19 @@ static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
> > >
> > > static void __blk_mq_fallback_to_insert(struct blk_mq_hw_ctx *hctx,
> > > struct request *rq,
> > > - bool run_queue)
> > > + bool run_queue, bool bypass_insert)
> > > {
> > > - blk_mq_sched_insert_request(rq, false, run_queue, false,
> > > - hctx->flags & BLK_MQ_F_BLOCKING);
> > > + if (!bypass_insert)
> > > + blk_mq_sched_insert_request(rq, false, run_queue, false,
> > > + hctx->flags & BLK_MQ_F_BLOCKING);
> > > + else
> > > + blk_mq_request_bypass_insert(rq, run_queue);
> > > }
> >
> > If 'bypass_insert' is true, we don't need to insert the request into
> > hctx->dispatch_list for dm-rq, then it causes the issue(use after free)
> > reported by Bart and Laurence.
> >
> > Also this way is the exact opposite of the idea of the improvement,
> > we do not want to dispatch request if underlying queue is busy.
>
> Yeap, please see the patch I just posted to fix it.
Looks your patch is a bit complicated, then __blk_mq_fallback_to_insert()
can be removed.
>
> But your v4 does fallback to using blk_mq_request_bypass_insert() as
> well, just in a much narrower case -- specifically:
> if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))
Yeah, I just found it, and you can add 'bypass_insert = false' under
condition.
--
Ming
