Re: [RFC PATCH 1/2] blk-mq: don't call callbacks for requests that bypassed the scheduler

From: Niklas Cassel
Date: Tue Sep 07 2021 - 08:59:37 EST


On Mon, Aug 30, 2021 at 06:11:12PM +0800, Ming Lei wrote:
> On Mon, Aug 30, 2021 at 09:48:06AM +0000, Niklas Cassel wrote:
> > On Fri, Aug 27, 2021 at 09:28:07PM +0800, Ming Lei wrote:
> > > On Fri, Aug 27, 2021 at 12:41:31PM +0000, Niklas Cassel wrote:
> > > > From: Niklas Cassel <niklas.cassel@xxxxxxx>
> > > >
> > > > Currently, __blk_mq_alloc_request() calls ops.prepare_request and sets
> > > > RQF_ELVPRIV.
> > > >
> > > > Therefore, (if the request is not a flush) the RQF_ELVPRIV flag will be
> > > > set for the request in blk_mq_submit_bio(), regardless if the request
> > > > was submitted to a scheduler, or bypassed the scheduler.
> > > >
> > > > Later, blk_mq_free_request() checks if the RQF_ELVPRIV flag is set,
> > > > if it is, the ops.finish_request callback will be called.
> > > >
> > > > The problem with this is that the finish_request scheduler callback
> > > > will be called for requests that bypassed the scheduler.
> > > >
> > > > Fix this by calling the scheduler ops.prepare_request callback, and
> > > > set the RQF_ELVPRIV flag only immediately before calling the insert
> > > > callback.
> > >
> > > One request could be inserted more than one times, such as requeue,
> > > however __blk_mq_alloc_request() is just run once, so is it fine to
> > > call ->prepare_request more than one time for same request?
> >
> > Calling ->prepare_request multiple times is fine.
> > All the different I/O schedulers (BFQ, mq-deadline, kyber)
> > simply use .prepare_request to clear/set elv->priv to a fixed value.
> >
> > >
> > > Or I am wondering why not call ->prepare_request when the following
> > > check is true?
> > >
> > > if (e && e->type->ops.prepare_request && !op_is_flush(data->cmd_flags) &&
> > > !blk_op_is_passthrough(data->cmd_flags))
> > > e->type->ops.prepare_request()
> >
> >
> > That might work, and might be a nicer solution indeed.
> >
> > If a request got plugged, it will be inserted to the scheduler through
> > blk_flush_plug_list() -> blk_mq_flush_plug_list() -> blk_mq_sched_insert_requests()
> > which will insert them unconditionally.
> > In this case. we know that !op_is_flush() (because if it was, blk_mq_submit_bio()
> > would have inserted directly.)
> >
> >
> > If we didn't plug, we do blk_mq_sched_insert_request(), which will add it if
> > blk_mq_sched_bypass_insert() returns false:
> >
> > blk_mq_sched_bypass_insert() is defined as:
> >
> > if ((rq->rq_flags & RQF_FLUSH_SEQ) || blk_rq_is_passthrough(rq))
> > return true;
> > Also in this case. we know that !op_is_flush() (blk_mq_submit_bio() would have
> > inserted directly.)
> >
> >
> > So, we could easily add && !blk_op_is_passthrough(data->cmd_flags) to the
> > ->prepare_request condition in blk_mq_rq_ctx_init() like you suggested,
> > but since the bypass condition also seems to look at RQF_FLUSH_SEQ, wouldn't
> > we need to add RQF_FLUSH_SEQ to the condition in blk_mq_rq_ctx_init() as well?
> >
> > This flag is set after blk_mq_rq_ctx_init(). Are we sure that RQF_FLUSH_SEQ
> > flag will only be set for a request which op_is_flush() returned true?
> >
> > (If so, then only adding && !blk_op_is_passthrough(data->cmd_flags) should
> > be fine.)
>
> BTW, what I meant is the following change, is it fine?
>
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 0a33d16a7298..f98f8cc05644 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -327,20 +327,6 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
>
> data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
> refcount_set(&rq->ref, 1);
> -
> - if (!op_is_flush(data->cmd_flags)) {
> - struct elevator_queue *e = data->q->elevator;
> -
> - rq->elv.icq = NULL;
> - if (e && e->type->ops.prepare_request) {
> - if (e->type->icq_cache)
> - blk_mq_sched_assign_ioc(rq);
> -
> - e->type->ops.prepare_request(rq);
> - rq->rq_flags |= RQF_ELVPRIV;
> - }
> - }
> -
> data->hctx->queued++;
> return rq;
> }
> @@ -359,17 +345,25 @@ static struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data)
> if (data->cmd_flags & REQ_NOWAIT)
> data->flags |= BLK_MQ_REQ_NOWAIT;
>
> - if (e) {
> + if (e && !op_is_flush(data->cmd_flags) &&
> + !blk_op_is_passthrough(data->cmd_flags)) {
> /*
> * Flush/passthrough requests are special and go directly to the
> * dispatch list. Don't include reserved tags in the
> * limiting, as it isn't useful.
> */
> - if (!op_is_flush(data->cmd_flags) &&
> - !blk_op_is_passthrough(data->cmd_flags) &&
> - e->type->ops.limit_depth &&
> - !(data->flags & BLK_MQ_REQ_RESERVED))
> + if (e->type->ops.limit_depth &&
> + !(data->flags & BLK_MQ_REQ_RESERVED))
> e->type->ops.limit_depth(data->cmd_flags, data);
> +
> + rq->elv.icq = NULL;
> + if (e->type->ops.prepare_request) {
> + if (e->type->icq_cache)
> + blk_mq_sched_assign_ioc(rq);
> +
> + e->type->ops.prepare_request(rq);
> + rq->rq_flags |= RQF_ELVPRIV;
> + }
> }
>
> retry:
>

Hello Ming,


Sorry for the delayed reply.

Your patch does not compile, because rq is not defined in this function.

Another problem seems to be that in __blk_mq_alloc_request(), at the end
of the function, calls blk_mq_rq_ctx_init(), which will unconditionally
set rq->rq_flags = 0;


The simple patch:

--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -328,7 +328,8 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
data->ctx->rq_dispatched[op_is_sync(data->cmd_flags)]++;
refcount_set(&rq->ref, 1);

- if (!op_is_flush(data->cmd_flags)) {
+ if (!op_is_flush(data->cmd_flags) &&
+ !blk_op_is_passthrough(data->cmd_flags)) {
struct elevator_queue *e = data->q->elevator;

rq->elv.icq = NULL;



Does appear to solve the problem.

My only worry was RQF_FLUSH_SEQ flag, but as far as I can tell, it is
only ever set for a request that which op_is_flush() returned true.


Kind regards,
Niklas