Re: [RFC/PATCH 4/4] block: Add URGENT request notification supportto CFQ scheduler

From: Tanya Brokhman
Date: Fri Jul 12 2013 - 11:29:23 EST


Hello Jeff

Thank you for your comments. Please see inline.

On 7/11/2013 9:41 PM, Jeff Moyer wrote:
Tanya Brokhman <tlinder@xxxxxxxxxxxxxx> writes:

When the scheduler reports to the block layer that there is an urgent
request pending, the device driver may decide to stop the transmission
of the current request in order to handle the urgent one. This is done
in order to reduce the latency of an urgent request. For example:
long WRITE may be stopped to handle an urgent READ.

In general, I don't like the approach taken. I would much rather see
a low-level cancellation method, and layer your urgent request handling
on top of that. That could actually be used by the aio subsystem as
well (with a lot of work, of course). That aside, I've provided some
comments below.


We shared a support for the low-level cancellation method some time ago. Please look for a patch from Konstantin Dorfman from June 30 subjected:
[RFC/PATCH v2] mmc: Add support to handle Urgent data transfer
This patch set was released as a usage example for the mmc patch. They should go together actually in order to achieve the benefit described bellow.



@@ -2111,12 +2114,13 @@ static void blk_account_io_done(struct request *req)
cpu = part_stat_lock();
part = req->part;

- part_stat_inc(cpu, part, ios[rw]);
- part_stat_add(cpu, part, ticks[rw], duration);
- part_round_stats(cpu, part);
- part_dec_in_flight(part, rw);
-
- hd_struct_put(part);
+ if (req->part != NULL) {
+ part_stat_inc(cpu, part, ios[rw]);
+ part_stat_add(cpu, part, ticks[rw], duration);
+ part_round_stats(cpu, part);
+ part_dec_in_flight(part, rw);
+ hd_struct_put(part);
+ }

A comment about why we now expect req->part might be null would be nice.


This is not related to my patch. Have no idea how it got in. I'll upload a new version soon. Sorry about that.


@@ -2783,6 +2786,14 @@ static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
(RQ_CFQG(rq))->dispatched++;
elv_dispatch_sort(q, rq);

+ if (rq->cmd_flags & REQ_URGENT) {
+ if (!cfqd->nr_urgent_pending)
+ WARN_ON(1);
+ else
+ cfqd->nr_urgent_pending--;
+ cfqd->nr_urgent_in_flight++;
+ }
+

This is a rather ugly construct, and gets repeated later. I'd be
inclined to just BUG.

will fix.

+/*
+ * Called when a request (rq) is reinserted (to cfqq). Check if there's
+ * something we should do about it
+ */
+static void
+cfq_rq_requeued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
+ struct request *rq)
+{
+ struct cfq_io_cq *cic = RQ_CIC(rq);
+
+ cfqd->rq_queued++;
+ if (rq->cmd_flags & REQ_PRIO)
+ cfqq->prio_pending++;
+
+ cfqq->dispatched--;
+ (RQ_CFQG(rq))->dispatched--;
+
+ cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
+
+ cfq_update_io_thinktime(cfqd, cfqq, cic);
+ cfq_update_io_seektime(cfqd, cfqq, rq);
+ cfq_update_idle_window(cfqd, cfqq, cic);
+
+ cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
+
+ if (cfqq == cfqd->active_queue) {
+ if (cfq_cfqq_wait_request(cfqq)) {
+ if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
+ cfqd->busy_queues > 1) {
+ cfq_del_timer(cfqd, cfqq);
+ cfq_clear_cfqq_wait_request(cfqq);
+ } else {
+ cfqg_stats_update_idle_time(cfqq->cfqg);
+ cfq_mark_cfqq_must_dispatch(cfqq);
+ }
+ }
+ } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
+ cfq_preempt_queue(cfqd, cfqq);
+ }
+}

Huge cut-n-paste of cfq_rq_enqueued. Please factor the code out.

ok.


+
static void cfq_insert_request(struct request_queue *q, struct request *rq)
{
struct cfq_data *cfqd = q->elevator->elevator_data;
@@ -3923,6 +3996,43 @@ static void cfq_insert_request(struct request_queue *q, struct request *rq)
cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
rq->cmd_flags);
cfq_rq_enqueued(cfqd, cfqq, rq);
+
+ if (rq->cmd_flags & REQ_URGENT) {
+ WARN_ON(1);
+ blk_dump_rq_flags(rq, "");
+ rq->cmd_flags &= ~REQ_URGENT;
+ }
+
+ /*
+ * Request is considered URGENT if:
+ * 1. The queue being served is of a lower IO priority then the new
+ * request
+ * OR:
+ * 2. The workload being performed is ASYNC
+ * Only READ requests may be considered as URGENT
+ */
+ if ((cfqd->active_queue &&
+ cfqq->ioprio_class < cfqd->active_queue->ioprio_class) ||
+ (cfqd->serving_wl_type == ASYNC_WORKLOAD &&
+ rq_data_dir(rq) == READ)) {
+ rq->cmd_flags |= REQ_URGENT;
+ cfqd->nr_urgent_pending++;
+ }

If requests are queued from a higher priority queue, then that queue
will preempt the existing queue. Why do we also need to interrupt read
requests from the lower priority queue? You seemed to indicate that
long-running writes were the primary concern.

You're right, our main concern are long-running writes. Preempting lower priority read request won't give us much benefit because read requests are usually short.
In the current implementation of urgent request in the mmc driver, short requests or reads won't be preempted. The block layer notifies the device driver of an urgent pending but it's the device driver decision whether to stop an ongoing request or not.
CFQ doesn't distinguish between read and sync write. And we do want to preempt sync write requests.


--
QUALCOMM ISRAEL, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/