block: schedule queue restart after BLK_STS_ZONE_RESOURCE

When dispatching a zone append write request to a SCSI zoned block device,
if the target zone of the request is already locked, the device driver will
return BLK_STS_ZONE_RESOURCE and the request will be pushed back to the
hctx dipatch queue. The queue will be marked as RESTART in
dd_finish_request() and restarted in __blk_mq_free_request(). However, this
restart applies to the hctx of the completed request. If the requeued
request is on a different hctx, dispatch will no be retried until another
request is submitted or the next periodic queue run triggers, leading to up
to 30 seconds latency for the requeued request.

Fix this problem by scheduling a queue restart similarly to the
BLK_STS_RESOURCE case or when we cannot get the budget.

Also, consolidate the checks into the "need_resource" variable to simplify
the condition.

Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Niklas Cassel <Niklas.Cassel@wdc.com>
Link: https://lore.kernel.org/r/20211026165127.4151055-1-naohiro.aota@wdc.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/block/blk-mq.c b/block/blk-mq.c
index bc02637..652a31f 100644 (file)
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -1325,6 +1325,7 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
        int errors, queued;
        blk_status_t ret = BLK_STS_OK;
        LIST_HEAD(zone_list);
+       bool needs_resource = false;
 
        if (list_empty(list))
                return false;
@@ -1370,6 +1371,8 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
                        queued++;
                        break;
                case BLK_STS_RESOURCE:
+                       needs_resource = true;
+                       fallthrough;
                case BLK_STS_DEV_RESOURCE:
                        blk_mq_handle_dev_resource(rq, list);
                        goto out;
@@ -1380,6 +1383,7 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
                         * accept.
                         */
                        blk_mq_handle_zone_resource(rq, &zone_list);
+                       needs_resource = true;
                        break;
                default:
                        errors++;
@@ -1406,7 +1410,6 @@ out:
                /* For non-shared tags, the RESTART check will suffice */
                bool no_tag = prep == PREP_DISPATCH_NO_TAG &&
                        (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
-               bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
 
                if (nr_budgets)
                        blk_mq_release_budgets(q, list);
@@ -1447,14 +1450,16 @@ out:
                 * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
                 * bit is set, run queue after a delay to avoid IO stalls
                 * that could otherwise occur if the queue is idle.  We'll do
-                * similar if we couldn't get budget and SCHED_RESTART is set.
+                * similar if we couldn't get budget or couldn't lock a zone
+                * and SCHED_RESTART is set.
                 */
                needs_restart = blk_mq_sched_needs_restart(hctx);
+               if (prep == PREP_DISPATCH_NO_BUDGET)
+                       needs_resource = true;
                if (!needs_restart ||
                    (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
                        blk_mq_run_hw_queue(hctx, true);
-               else if (needs_restart && (ret == BLK_STS_RESOURCE ||
-                                          no_budget_avail))
+               else if (needs_restart && needs_resource)
                        blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
 
                blk_mq_update_dispatch_busy(hctx, true);