]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/commitdiff
block, bfq: reduce latency during request-pool saturation
authorPaolo Valente <paolo.valente@linaro.org>
Wed, 12 Apr 2017 16:23:15 +0000 (18:23 +0200)
committerJens Axboe <axboe@fb.com>
Wed, 19 Apr 2017 14:30:26 +0000 (08:30 -0600)
This patch introduces an heuristic that reduces latency when the
I/O-request pool is saturated. This goal is achieved by disabling
device idling, for non-weight-raised queues, when there are weight-
raised queues with pending or in-flight requests. In fact, as
explained in more detail in the comment on the function
bfq_bfqq_may_idle(), this reduces the rate at which processes
associated with non-weight-raised queues grab requests from the pool,
thereby increasing the probability that processes associated with
weight-raised queues get a request immediately (or at least soon) when
they need one. Along the same line, if there are weight-raised queues,
then this patch halves the service rate of async (write) requests for
non-weight-raised queues.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/bfq-iosched.c

index 574a5f6a2370f62d02203029845f65705affc6bb..deb1f21c535fddb180d4f896063f0798cd18776f 100644 (file)
@@ -420,6 +420,8 @@ struct bfq_data {
         * queue in service, even if it is idling).
         */
        int busy_queues;
+       /* number of weight-raised busy @bfq_queues */
+       int wr_busy_queues;
        /* number of queued requests */
        int queued;
        /* number of requests dispatched and waiting for completion */
@@ -2490,6 +2492,9 @@ static void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq,
 
        bfqd->busy_queues--;
 
+       if (bfqq->wr_coeff > 1)
+               bfqd->wr_busy_queues--;
+
        bfqg_stats_update_dequeue(bfqq_group(bfqq));
 
        bfq_deactivate_bfqq(bfqd, bfqq, true, expiration);
@@ -2506,6 +2511,9 @@ static void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 
        bfq_mark_bfqq_busy(bfqq);
        bfqd->busy_queues++;
+
+       if (bfqq->wr_coeff > 1)
+               bfqd->wr_busy_queues++;
 }
 
 #ifdef CONFIG_BFQ_GROUP_IOSCHED
@@ -3779,7 +3787,16 @@ static unsigned long bfq_serv_to_charge(struct request *rq,
        if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1)
                return blk_rq_sectors(rq);
 
-       return blk_rq_sectors(rq) * bfq_async_charge_factor;
+       /*
+        * If there are no weight-raised queues, then amplify service
+        * by just the async charge factor; otherwise amplify service
+        * by twice the async charge factor, to further reduce latency
+        * for weight-raised queues.
+        */
+       if (bfqq->bfqd->wr_busy_queues == 0)
+               return blk_rq_sectors(rq) * bfq_async_charge_factor;
+
+       return blk_rq_sectors(rq) * 2 * bfq_async_charge_factor;
 }
 
 /**
@@ -4234,6 +4251,7 @@ static void bfq_add_request(struct request *rq)
                        bfqq->wr_coeff = bfqd->bfq_wr_coeff;
                        bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
 
+                       bfqd->wr_busy_queues++;
                        bfqq->entity.prio_changed = 1;
                }
                if (prev != bfqq->next_rq)
@@ -4474,6 +4492,8 @@ end:
 /* Must be called with bfqq != NULL */
 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
 {
+       if (bfq_bfqq_busy(bfqq))
+               bfqq->bfqd->wr_busy_queues--;
        bfqq->wr_coeff = 1;
        bfqq->wr_cur_max_time = 0;
        bfqq->last_wr_start_finish = jiffies;
@@ -5497,7 +5517,8 @@ static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
 static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
 {
        struct bfq_data *bfqd = bfqq->bfqd;
-       bool idling_boosts_thr, asymmetric_scenario;
+       bool idling_boosts_thr, idling_boosts_thr_without_issues,
+               asymmetric_scenario;
 
        if (bfqd->strict_guarantees)
                return true;
@@ -5519,6 +5540,44 @@ static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
         */
        idling_boosts_thr = !bfqd->hw_tag || bfq_bfqq_IO_bound(bfqq);
 
+       /*
+        * The value of the next variable,
+        * idling_boosts_thr_without_issues, is equal to that of
+        * idling_boosts_thr, unless a special case holds. In this
+        * special case, described below, idling may cause problems to
+        * weight-raised queues.
+        *
+        * When the request pool is saturated (e.g., in the presence
+        * of write hogs), if the processes associated with
+        * non-weight-raised queues ask for requests at a lower rate,
+        * then processes associated with weight-raised queues have a
+        * higher probability to get a request from the pool
+        * immediately (or at least soon) when they need one. Thus
+        * they have a higher probability to actually get a fraction
+        * of the device throughput proportional to their high
+        * weight. This is especially true with NCQ-capable drives,
+        * which enqueue several requests in advance, and further
+        * reorder internally-queued requests.
+        *
+        * For this reason, we force to false the value of
+        * idling_boosts_thr_without_issues if there are weight-raised
+        * busy queues. In this case, and if bfqq is not weight-raised,
+        * this guarantees that the device is not idled for bfqq (if,
+        * instead, bfqq is weight-raised, then idling will be
+        * guaranteed by another variable, see below). Combined with
+        * the timestamping rules of BFQ (see [1] for details), this
+        * behavior causes bfqq, and hence any sync non-weight-raised
+        * queue, to get a lower number of requests served, and thus
+        * to ask for a lower number of requests from the request
+        * pool, before the busy weight-raised queues get served
+        * again. This often mitigates starvation problems in the
+        * presence of heavy write workloads and NCQ, thereby
+        * guaranteeing a higher application and system responsiveness
+        * in these hostile scenarios.
+        */
+       idling_boosts_thr_without_issues = idling_boosts_thr &&
+               bfqd->wr_busy_queues == 0;
+
        /*
         * There is then a case where idling must be performed not for
         * throughput concerns, but to preserve service guarantees. To
@@ -5593,7 +5652,7 @@ static bool bfq_bfqq_may_idle(struct bfq_queue *bfqq)
         *    is necessary to preserve service guarantees.
         */
        return bfq_bfqq_sync(bfqq) &&
-               (idling_boosts_thr || asymmetric_scenario);
+               (idling_boosts_thr_without_issues || asymmetric_scenario);
 }
 
 /*
@@ -6801,6 +6860,7 @@ static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
                                              * high-definition compressed
                                              * video.
                                              */
+       bfqd->wr_busy_queues = 0;
 
        /*
         * Begin by assuming, optimistically, that the device is a