1 // SPDX-License-Identifier: GPL-2.0
3 * Tag allocation using scalable bitmaps. Uses active queue tracking to support
4 * fairer distribution of tags between multiple submitters when a shared tag map
7 * Copyright (C) 2013-2014 Jens Axboe
9 #include <linux/kernel.h>
10 #include <linux/module.h>
12 #include <linux/delay.h>
15 #include "blk-mq-sched.h"
18 * Recalculate wakeup batch when tag is shared by hctx.
20 static void blk_mq_update_wake_batch(struct blk_mq_tags *tags,
26 sbitmap_queue_recalculate_wake_batch(&tags->bitmap_tags,
28 sbitmap_queue_recalculate_wake_batch(&tags->breserved_tags,
33 * If a previously inactive queue goes active, bump the active user count.
34 * We need to do this before try to allocate driver tag, then even if fail
35 * to get tag when first time, the other shared-tag users could reserve
38 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
42 struct blk_mq_tags *tags = hctx->tags;
45 * calling test_bit() prior to test_and_set_bit() is intentional,
46 * it avoids dirtying the cacheline if the queue is already active.
48 if (blk_mq_is_shared_tags(hctx->flags)) {
49 struct request_queue *q = hctx->queue;
51 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) ||
52 test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
55 if (test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) ||
56 test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
60 spin_lock_irqsave(&tags->lock, flags);
61 users = tags->active_queues + 1;
62 WRITE_ONCE(tags->active_queues, users);
63 blk_mq_update_wake_batch(tags, users);
64 spin_unlock_irqrestore(&tags->lock, flags);
68 * Wakeup all potentially sleeping on tags
70 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
72 sbitmap_queue_wake_all(&tags->bitmap_tags);
74 sbitmap_queue_wake_all(&tags->breserved_tags);
78 * If a previously busy queue goes inactive, potential waiters could now
79 * be allowed to queue. Wake them up and check.
81 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
83 struct blk_mq_tags *tags = hctx->tags;
86 if (blk_mq_is_shared_tags(hctx->flags)) {
87 struct request_queue *q = hctx->queue;
89 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
93 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
97 spin_lock_irq(&tags->lock);
98 users = tags->active_queues - 1;
99 WRITE_ONCE(tags->active_queues, users);
100 blk_mq_update_wake_batch(tags, users);
101 spin_unlock_irq(&tags->lock);
103 blk_mq_tag_wakeup_all(tags, false);
106 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
107 struct sbitmap_queue *bt)
109 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
110 !hctx_may_queue(data->hctx, bt))
111 return BLK_MQ_NO_TAG;
113 if (data->shallow_depth)
114 return sbitmap_queue_get_shallow(bt, data->shallow_depth);
116 return __sbitmap_queue_get(bt);
119 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
120 unsigned int *offset)
122 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
123 struct sbitmap_queue *bt = &tags->bitmap_tags;
126 if (data->shallow_depth ||data->flags & BLK_MQ_REQ_RESERVED ||
127 data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
129 ret = __sbitmap_queue_get_batch(bt, nr_tags, offset);
130 *offset += tags->nr_reserved_tags;
134 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
136 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
137 struct sbitmap_queue *bt;
138 struct sbq_wait_state *ws;
139 DEFINE_SBQ_WAIT(wait);
140 unsigned int tag_offset;
143 if (data->flags & BLK_MQ_REQ_RESERVED) {
144 if (unlikely(!tags->nr_reserved_tags)) {
146 return BLK_MQ_NO_TAG;
148 bt = &tags->breserved_tags;
151 bt = &tags->bitmap_tags;
152 tag_offset = tags->nr_reserved_tags;
155 tag = __blk_mq_get_tag(data, bt);
156 if (tag != BLK_MQ_NO_TAG)
159 if (data->flags & BLK_MQ_REQ_NOWAIT)
160 return BLK_MQ_NO_TAG;
162 ws = bt_wait_ptr(bt, data->hctx);
164 struct sbitmap_queue *bt_prev;
167 * We're out of tags on this hardware queue, kick any
168 * pending IO submits before going to sleep waiting for
171 blk_mq_run_hw_queue(data->hctx, false);
174 * Retry tag allocation after running the hardware queue,
175 * as running the queue may also have found completions.
177 tag = __blk_mq_get_tag(data, bt);
178 if (tag != BLK_MQ_NO_TAG)
181 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
183 tag = __blk_mq_get_tag(data, bt);
184 if (tag != BLK_MQ_NO_TAG)
190 sbitmap_finish_wait(bt, ws, &wait);
192 data->ctx = blk_mq_get_ctx(data->q);
193 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
195 tags = blk_mq_tags_from_data(data);
196 if (data->flags & BLK_MQ_REQ_RESERVED)
197 bt = &tags->breserved_tags;
199 bt = &tags->bitmap_tags;
202 * If destination hw queue is changed, fake wake up on
203 * previous queue for compensating the wake up miss, so
204 * other allocations on previous queue won't be starved.
207 sbitmap_queue_wake_up(bt_prev, 1);
209 ws = bt_wait_ptr(bt, data->hctx);
212 sbitmap_finish_wait(bt, ws, &wait);
216 * Give up this allocation if the hctx is inactive. The caller will
217 * retry on an active hctx.
219 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
220 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
221 return BLK_MQ_NO_TAG;
223 return tag + tag_offset;
226 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
229 if (!blk_mq_tag_is_reserved(tags, tag)) {
230 const int real_tag = tag - tags->nr_reserved_tags;
232 BUG_ON(real_tag >= tags->nr_tags);
233 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
235 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
239 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags)
241 sbitmap_queue_clear_batch(&tags->bitmap_tags, tags->nr_reserved_tags,
245 struct bt_iter_data {
246 struct blk_mq_hw_ctx *hctx;
247 struct request_queue *q;
248 busy_tag_iter_fn *fn;
253 static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
259 spin_lock_irqsave(&tags->lock, flags);
260 rq = tags->rqs[bitnr];
261 if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
263 spin_unlock_irqrestore(&tags->lock, flags);
267 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
269 struct bt_iter_data *iter_data = data;
270 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
271 struct request_queue *q = iter_data->q;
272 struct blk_mq_tag_set *set = q->tag_set;
273 struct blk_mq_tags *tags;
277 if (blk_mq_is_shared_tags(set->flags))
278 tags = set->shared_tags;
282 if (!iter_data->reserved)
283 bitnr += tags->nr_reserved_tags;
285 * We can hit rq == NULL here, because the tagging functions
286 * test and set the bit before assigning ->rqs[].
288 rq = blk_mq_find_and_get_req(tags, bitnr);
292 if (rq->q == q && (!hctx || rq->mq_hctx == hctx))
293 ret = iter_data->fn(rq, iter_data->data);
294 blk_mq_put_rq_ref(rq);
299 * bt_for_each - iterate over the requests associated with a hardware queue
300 * @hctx: Hardware queue to examine.
301 * @q: Request queue to examine.
302 * @bt: sbitmap to examine. This is either the breserved_tags member
303 * or the bitmap_tags member of struct blk_mq_tags.
304 * @fn: Pointer to the function that will be called for each request
305 * associated with @hctx that has been assigned a driver tag.
306 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
307 * where rq is a pointer to a request. Return true to continue
308 * iterating tags, false to stop.
309 * @data: Will be passed as third argument to @fn.
310 * @reserved: Indicates whether @bt is the breserved_tags member or the
311 * bitmap_tags member of struct blk_mq_tags.
313 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct request_queue *q,
314 struct sbitmap_queue *bt, busy_tag_iter_fn *fn,
315 void *data, bool reserved)
317 struct bt_iter_data iter_data = {
321 .reserved = reserved,
325 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
328 struct bt_tags_iter_data {
329 struct blk_mq_tags *tags;
330 busy_tag_iter_fn *fn;
335 #define BT_TAG_ITER_RESERVED (1 << 0)
336 #define BT_TAG_ITER_STARTED (1 << 1)
337 #define BT_TAG_ITER_STATIC_RQS (1 << 2)
339 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
341 struct bt_tags_iter_data *iter_data = data;
342 struct blk_mq_tags *tags = iter_data->tags;
345 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
347 if (!(iter_data->flags & BT_TAG_ITER_RESERVED))
348 bitnr += tags->nr_reserved_tags;
351 * We can hit rq == NULL here, because the tagging functions
352 * test and set the bit before assigning ->rqs[].
355 rq = tags->static_rqs[bitnr];
357 rq = blk_mq_find_and_get_req(tags, bitnr);
361 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
362 blk_mq_request_started(rq))
363 ret = iter_data->fn(rq, iter_data->data);
364 if (!iter_static_rqs)
365 blk_mq_put_rq_ref(rq);
370 * bt_tags_for_each - iterate over the requests in a tag map
371 * @tags: Tag map to iterate over.
372 * @bt: sbitmap to examine. This is either the breserved_tags member
373 * or the bitmap_tags member of struct blk_mq_tags.
374 * @fn: Pointer to the function that will be called for each started
375 * request. @fn will be called as follows: @fn(rq, @data,
376 * @reserved) where rq is a pointer to a request. Return true
377 * to continue iterating tags, false to stop.
378 * @data: Will be passed as second argument to @fn.
379 * @flags: BT_TAG_ITER_*
381 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
382 busy_tag_iter_fn *fn, void *data, unsigned int flags)
384 struct bt_tags_iter_data iter_data = {
392 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
395 static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
396 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
398 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
400 if (tags->nr_reserved_tags)
401 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv,
402 flags | BT_TAG_ITER_RESERVED);
403 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, flags);
407 * blk_mq_all_tag_iter - iterate over all requests in a tag map
408 * @tags: Tag map to iterate over.
409 * @fn: Pointer to the function that will be called for each
410 * request. @fn will be called as follows: @fn(rq, @priv,
411 * reserved) where rq is a pointer to a request. 'reserved'
412 * indicates whether or not @rq is a reserved request. Return
413 * true to continue iterating tags, false to stop.
414 * @priv: Will be passed as second argument to @fn.
416 * Caller has to pass the tag map from which requests are allocated.
418 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
421 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
425 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
426 * @tagset: Tag set to iterate over.
427 * @fn: Pointer to the function that will be called for each started
428 * request. @fn will be called as follows: @fn(rq, @priv,
429 * reserved) where rq is a pointer to a request. 'reserved'
430 * indicates whether or not @rq is a reserved request. Return
431 * true to continue iterating tags, false to stop.
432 * @priv: Will be passed as second argument to @fn.
434 * We grab one request reference before calling @fn and release it after
437 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
438 busy_tag_iter_fn *fn, void *priv)
440 unsigned int flags = tagset->flags;
443 nr_tags = blk_mq_is_shared_tags(flags) ? 1 : tagset->nr_hw_queues;
445 for (i = 0; i < nr_tags; i++) {
446 if (tagset->tags && tagset->tags[i])
447 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
448 BT_TAG_ITER_STARTED);
451 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
453 static bool blk_mq_tagset_count_completed_rqs(struct request *rq, void *data)
455 unsigned *count = data;
457 if (blk_mq_request_completed(rq))
463 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
464 * completions have finished.
465 * @tagset: Tag set to drain completed request
467 * Note: This function has to be run after all IO queues are shutdown
469 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
474 blk_mq_tagset_busy_iter(tagset,
475 blk_mq_tagset_count_completed_rqs, &count);
481 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
484 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
485 * @q: Request queue to examine.
486 * @fn: Pointer to the function that will be called for each request
487 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
488 * reserved) where rq is a pointer to a request and hctx points
489 * to the hardware queue associated with the request. 'reserved'
490 * indicates whether or not @rq is a reserved request.
491 * @priv: Will be passed as third argument to @fn.
493 * Note: if @q->tag_set is shared with other request queues then @fn will be
494 * called for all requests on all queues that share that tag set and not only
495 * for requests associated with @q.
497 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
501 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and hctx_table
502 * while the queue is frozen. So we can use q_usage_counter to avoid
505 if (!percpu_ref_tryget(&q->q_usage_counter))
508 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
509 struct blk_mq_tags *tags = q->tag_set->shared_tags;
510 struct sbitmap_queue *bresv = &tags->breserved_tags;
511 struct sbitmap_queue *btags = &tags->bitmap_tags;
513 if (tags->nr_reserved_tags)
514 bt_for_each(NULL, q, bresv, fn, priv, true);
515 bt_for_each(NULL, q, btags, fn, priv, false);
517 struct blk_mq_hw_ctx *hctx;
520 queue_for_each_hw_ctx(q, hctx, i) {
521 struct blk_mq_tags *tags = hctx->tags;
522 struct sbitmap_queue *bresv = &tags->breserved_tags;
523 struct sbitmap_queue *btags = &tags->bitmap_tags;
526 * If no software queues are currently mapped to this
527 * hardware queue, there's nothing to check
529 if (!blk_mq_hw_queue_mapped(hctx))
532 if (tags->nr_reserved_tags)
533 bt_for_each(hctx, q, bresv, fn, priv, true);
534 bt_for_each(hctx, q, btags, fn, priv, false);
540 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
541 bool round_robin, int node)
543 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
547 int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
548 struct sbitmap_queue *breserved_tags,
549 unsigned int queue_depth, unsigned int reserved,
550 int node, int alloc_policy)
552 unsigned int depth = queue_depth - reserved;
553 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
555 if (bt_alloc(bitmap_tags, depth, round_robin, node))
557 if (bt_alloc(breserved_tags, reserved, round_robin, node))
558 goto free_bitmap_tags;
563 sbitmap_queue_free(bitmap_tags);
567 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
568 unsigned int reserved_tags,
569 int node, int alloc_policy)
571 struct blk_mq_tags *tags;
573 if (total_tags > BLK_MQ_TAG_MAX) {
574 pr_err("blk-mq: tag depth too large\n");
578 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
582 tags->nr_tags = total_tags;
583 tags->nr_reserved_tags = reserved_tags;
584 spin_lock_init(&tags->lock);
586 if (blk_mq_init_bitmaps(&tags->bitmap_tags, &tags->breserved_tags,
587 total_tags, reserved_tags, node,
595 void blk_mq_free_tags(struct blk_mq_tags *tags)
597 sbitmap_queue_free(&tags->bitmap_tags);
598 sbitmap_queue_free(&tags->breserved_tags);
602 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
603 struct blk_mq_tags **tagsptr, unsigned int tdepth,
606 struct blk_mq_tags *tags = *tagsptr;
608 if (tdepth <= tags->nr_reserved_tags)
612 * If we are allowed to grow beyond the original size, allocate
613 * a new set of tags before freeing the old one.
615 if (tdepth > tags->nr_tags) {
616 struct blk_mq_tag_set *set = hctx->queue->tag_set;
617 struct blk_mq_tags *new;
623 * We need some sort of upper limit, set it high enough that
624 * no valid use cases should require more.
626 if (tdepth > MAX_SCHED_RQ)
630 * Only the sbitmap needs resizing since we allocated the max
633 if (blk_mq_is_shared_tags(set->flags))
636 new = blk_mq_alloc_map_and_rqs(set, hctx->queue_num, tdepth);
640 blk_mq_free_map_and_rqs(set, *tagsptr, hctx->queue_num);
644 * Don't need (or can't) update reserved tags here, they
645 * remain static and should never need resizing.
647 sbitmap_queue_resize(&tags->bitmap_tags,
648 tdepth - tags->nr_reserved_tags);
654 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, unsigned int size)
656 struct blk_mq_tags *tags = set->shared_tags;
658 sbitmap_queue_resize(&tags->bitmap_tags, size - set->reserved_tags);
661 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q)
663 sbitmap_queue_resize(&q->sched_shared_tags->bitmap_tags,
664 q->nr_requests - q->tag_set->reserved_tags);
668 * blk_mq_unique_tag() - return a tag that is unique queue-wide
669 * @rq: request for which to compute a unique tag
671 * The tag field in struct request is unique per hardware queue but not over
672 * all hardware queues. Hence this function that returns a tag with the
673 * hardware context index in the upper bits and the per hardware queue tag in
676 * Note: When called for a request that is queued on a non-multiqueue request
677 * queue, the hardware context index is set to zero.
679 u32 blk_mq_unique_tag(struct request *rq)
681 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
682 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
684 EXPORT_SYMBOL(blk_mq_unique_tag);