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/blk-mq.h>
13 #include <linux/delay.h>
16 #include "blk-mq-sched.h"
17 #include "blk-mq-tag.h"
20 * Recalculate wakeup batch when tag is shared by hctx.
22 static void blk_mq_update_wake_batch(struct blk_mq_tags *tags,
28 sbitmap_queue_recalculate_wake_batch(&tags->bitmap_tags,
30 sbitmap_queue_recalculate_wake_batch(&tags->breserved_tags,
35 * If a previously inactive queue goes active, bump the active user count.
36 * We need to do this before try to allocate driver tag, then even if fail
37 * to get tag when first time, the other shared-tag users could reserve
40 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
44 if (blk_mq_is_shared_tags(hctx->flags)) {
45 struct request_queue *q = hctx->queue;
47 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
49 set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags);
51 if (test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
53 set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state);
56 users = atomic_inc_return(&hctx->tags->active_queues);
58 blk_mq_update_wake_batch(hctx->tags, users);
62 * Wakeup all potentially sleeping on tags
64 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
66 sbitmap_queue_wake_all(&tags->bitmap_tags);
68 sbitmap_queue_wake_all(&tags->breserved_tags);
72 * If a previously busy queue goes inactive, potential waiters could now
73 * be allowed to queue. Wake them up and check.
75 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
77 struct blk_mq_tags *tags = hctx->tags;
80 if (blk_mq_is_shared_tags(hctx->flags)) {
81 struct request_queue *q = hctx->queue;
83 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
87 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
91 users = atomic_dec_return(&tags->active_queues);
93 blk_mq_update_wake_batch(tags, users);
95 blk_mq_tag_wakeup_all(tags, false);
98 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
99 struct sbitmap_queue *bt)
101 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
102 !hctx_may_queue(data->hctx, bt))
103 return BLK_MQ_NO_TAG;
105 if (data->shallow_depth)
106 return sbitmap_queue_get_shallow(bt, data->shallow_depth);
108 return __sbitmap_queue_get(bt);
111 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
112 unsigned int *offset)
114 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
115 struct sbitmap_queue *bt = &tags->bitmap_tags;
118 if (data->shallow_depth ||data->flags & BLK_MQ_REQ_RESERVED ||
119 data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
121 ret = __sbitmap_queue_get_batch(bt, nr_tags, offset);
122 *offset += tags->nr_reserved_tags;
126 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
128 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
129 struct sbitmap_queue *bt;
130 struct sbq_wait_state *ws;
131 DEFINE_SBQ_WAIT(wait);
132 unsigned int tag_offset;
135 if (data->flags & BLK_MQ_REQ_RESERVED) {
136 if (unlikely(!tags->nr_reserved_tags)) {
138 return BLK_MQ_NO_TAG;
140 bt = &tags->breserved_tags;
143 bt = &tags->bitmap_tags;
144 tag_offset = tags->nr_reserved_tags;
147 tag = __blk_mq_get_tag(data, bt);
148 if (tag != BLK_MQ_NO_TAG)
151 if (data->flags & BLK_MQ_REQ_NOWAIT)
152 return BLK_MQ_NO_TAG;
154 ws = bt_wait_ptr(bt, data->hctx);
156 struct sbitmap_queue *bt_prev;
159 * We're out of tags on this hardware queue, kick any
160 * pending IO submits before going to sleep waiting for
163 blk_mq_run_hw_queue(data->hctx, false);
166 * Retry tag allocation after running the hardware queue,
167 * as running the queue may also have found completions.
169 tag = __blk_mq_get_tag(data, bt);
170 if (tag != BLK_MQ_NO_TAG)
173 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
175 tag = __blk_mq_get_tag(data, bt);
176 if (tag != BLK_MQ_NO_TAG)
182 sbitmap_finish_wait(bt, ws, &wait);
184 data->ctx = blk_mq_get_ctx(data->q);
185 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
187 tags = blk_mq_tags_from_data(data);
188 if (data->flags & BLK_MQ_REQ_RESERVED)
189 bt = &tags->breserved_tags;
191 bt = &tags->bitmap_tags;
194 * If destination hw queue is changed, fake wake up on
195 * previous queue for compensating the wake up miss, so
196 * other allocations on previous queue won't be starved.
199 sbitmap_queue_wake_up(bt_prev, 1);
201 ws = bt_wait_ptr(bt, data->hctx);
204 sbitmap_finish_wait(bt, ws, &wait);
208 * Give up this allocation if the hctx is inactive. The caller will
209 * retry on an active hctx.
211 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
212 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
213 return BLK_MQ_NO_TAG;
215 return tag + tag_offset;
218 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
221 if (!blk_mq_tag_is_reserved(tags, tag)) {
222 const int real_tag = tag - tags->nr_reserved_tags;
224 BUG_ON(real_tag >= tags->nr_tags);
225 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
227 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
231 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags)
233 sbitmap_queue_clear_batch(&tags->bitmap_tags, tags->nr_reserved_tags,
237 struct bt_iter_data {
238 struct blk_mq_hw_ctx *hctx;
239 struct request_queue *q;
240 busy_tag_iter_fn *fn;
245 static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
251 spin_lock_irqsave(&tags->lock, flags);
252 rq = tags->rqs[bitnr];
253 if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
255 spin_unlock_irqrestore(&tags->lock, flags);
259 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
261 struct bt_iter_data *iter_data = data;
262 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
263 struct request_queue *q = iter_data->q;
264 struct blk_mq_tag_set *set = q->tag_set;
265 struct blk_mq_tags *tags;
269 if (blk_mq_is_shared_tags(set->flags))
270 tags = set->shared_tags;
274 if (!iter_data->reserved)
275 bitnr += tags->nr_reserved_tags;
277 * We can hit rq == NULL here, because the tagging functions
278 * test and set the bit before assigning ->rqs[].
280 rq = blk_mq_find_and_get_req(tags, bitnr);
284 if (rq->q == q && (!hctx || rq->mq_hctx == hctx))
285 ret = iter_data->fn(rq, iter_data->data);
286 blk_mq_put_rq_ref(rq);
291 * bt_for_each - iterate over the requests associated with a hardware queue
292 * @hctx: Hardware queue to examine.
293 * @q: Request queue to examine.
294 * @bt: sbitmap to examine. This is either the breserved_tags member
295 * or the bitmap_tags member of struct blk_mq_tags.
296 * @fn: Pointer to the function that will be called for each request
297 * associated with @hctx that has been assigned a driver tag.
298 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
299 * where rq is a pointer to a request. Return true to continue
300 * iterating tags, false to stop.
301 * @data: Will be passed as third argument to @fn.
302 * @reserved: Indicates whether @bt is the breserved_tags member or the
303 * bitmap_tags member of struct blk_mq_tags.
305 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct request_queue *q,
306 struct sbitmap_queue *bt, busy_tag_iter_fn *fn,
307 void *data, bool reserved)
309 struct bt_iter_data iter_data = {
313 .reserved = reserved,
317 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
320 struct bt_tags_iter_data {
321 struct blk_mq_tags *tags;
322 busy_tag_iter_fn *fn;
327 #define BT_TAG_ITER_RESERVED (1 << 0)
328 #define BT_TAG_ITER_STARTED (1 << 1)
329 #define BT_TAG_ITER_STATIC_RQS (1 << 2)
331 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
333 struct bt_tags_iter_data *iter_data = data;
334 struct blk_mq_tags *tags = iter_data->tags;
337 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
339 if (!(iter_data->flags & BT_TAG_ITER_RESERVED))
340 bitnr += tags->nr_reserved_tags;
343 * We can hit rq == NULL here, because the tagging functions
344 * test and set the bit before assigning ->rqs[].
347 rq = tags->static_rqs[bitnr];
349 rq = blk_mq_find_and_get_req(tags, bitnr);
353 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
354 blk_mq_request_started(rq))
355 ret = iter_data->fn(rq, iter_data->data);
356 if (!iter_static_rqs)
357 blk_mq_put_rq_ref(rq);
362 * bt_tags_for_each - iterate over the requests in a tag map
363 * @tags: Tag map to iterate over.
364 * @bt: sbitmap to examine. This is either the breserved_tags member
365 * or the bitmap_tags member of struct blk_mq_tags.
366 * @fn: Pointer to the function that will be called for each started
367 * request. @fn will be called as follows: @fn(rq, @data,
368 * @reserved) where rq is a pointer to a request. Return true
369 * to continue iterating tags, false to stop.
370 * @data: Will be passed as second argument to @fn.
371 * @flags: BT_TAG_ITER_*
373 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
374 busy_tag_iter_fn *fn, void *data, unsigned int flags)
376 struct bt_tags_iter_data iter_data = {
384 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
387 static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
388 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
390 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
392 if (tags->nr_reserved_tags)
393 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv,
394 flags | BT_TAG_ITER_RESERVED);
395 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, flags);
399 * blk_mq_all_tag_iter - iterate over all requests in a tag map
400 * @tags: Tag map to iterate over.
401 * @fn: Pointer to the function that will be called for each
402 * request. @fn will be called as follows: @fn(rq, @priv,
403 * reserved) where rq is a pointer to a request. 'reserved'
404 * indicates whether or not @rq is a reserved request. Return
405 * true to continue iterating tags, false to stop.
406 * @priv: Will be passed as second argument to @fn.
408 * Caller has to pass the tag map from which requests are allocated.
410 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
413 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
417 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
418 * @tagset: Tag set to iterate over.
419 * @fn: Pointer to the function that will be called for each started
420 * request. @fn will be called as follows: @fn(rq, @priv,
421 * reserved) where rq is a pointer to a request. 'reserved'
422 * indicates whether or not @rq is a reserved request. Return
423 * true to continue iterating tags, false to stop.
424 * @priv: Will be passed as second argument to @fn.
426 * We grab one request reference before calling @fn and release it after
429 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
430 busy_tag_iter_fn *fn, void *priv)
432 unsigned int flags = tagset->flags;
435 nr_tags = blk_mq_is_shared_tags(flags) ? 1 : tagset->nr_hw_queues;
437 for (i = 0; i < nr_tags; i++) {
438 if (tagset->tags && tagset->tags[i])
439 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
440 BT_TAG_ITER_STARTED);
443 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
445 static bool blk_mq_tagset_count_completed_rqs(struct request *rq, void *data)
447 unsigned *count = data;
449 if (blk_mq_request_completed(rq))
455 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
456 * completions have finished.
457 * @tagset: Tag set to drain completed request
459 * Note: This function has to be run after all IO queues are shutdown
461 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
466 blk_mq_tagset_busy_iter(tagset,
467 blk_mq_tagset_count_completed_rqs, &count);
473 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
476 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
477 * @q: Request queue to examine.
478 * @fn: Pointer to the function that will be called for each request
479 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
480 * reserved) where rq is a pointer to a request and hctx points
481 * to the hardware queue associated with the request. 'reserved'
482 * indicates whether or not @rq is a reserved request.
483 * @priv: Will be passed as third argument to @fn.
485 * Note: if @q->tag_set is shared with other request queues then @fn will be
486 * called for all requests on all queues that share that tag set and not only
487 * for requests associated with @q.
489 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
493 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and hctx_table
494 * while the queue is frozen. So we can use q_usage_counter to avoid
497 if (!percpu_ref_tryget(&q->q_usage_counter))
500 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
501 struct blk_mq_tags *tags = q->tag_set->shared_tags;
502 struct sbitmap_queue *bresv = &tags->breserved_tags;
503 struct sbitmap_queue *btags = &tags->bitmap_tags;
505 if (tags->nr_reserved_tags)
506 bt_for_each(NULL, q, bresv, fn, priv, true);
507 bt_for_each(NULL, q, btags, fn, priv, false);
509 struct blk_mq_hw_ctx *hctx;
512 queue_for_each_hw_ctx(q, hctx, i) {
513 struct blk_mq_tags *tags = hctx->tags;
514 struct sbitmap_queue *bresv = &tags->breserved_tags;
515 struct sbitmap_queue *btags = &tags->bitmap_tags;
518 * If no software queues are currently mapped to this
519 * hardware queue, there's nothing to check
521 if (!blk_mq_hw_queue_mapped(hctx))
524 if (tags->nr_reserved_tags)
525 bt_for_each(hctx, q, bresv, fn, priv, true);
526 bt_for_each(hctx, q, btags, fn, priv, false);
532 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
533 bool round_robin, int node)
535 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
539 int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
540 struct sbitmap_queue *breserved_tags,
541 unsigned int queue_depth, unsigned int reserved,
542 int node, int alloc_policy)
544 unsigned int depth = queue_depth - reserved;
545 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
547 if (bt_alloc(bitmap_tags, depth, round_robin, node))
549 if (bt_alloc(breserved_tags, reserved, round_robin, node))
550 goto free_bitmap_tags;
555 sbitmap_queue_free(bitmap_tags);
559 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
560 unsigned int reserved_tags,
561 int node, int alloc_policy)
563 struct blk_mq_tags *tags;
565 if (total_tags > BLK_MQ_TAG_MAX) {
566 pr_err("blk-mq: tag depth too large\n");
570 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
574 tags->nr_tags = total_tags;
575 tags->nr_reserved_tags = reserved_tags;
576 spin_lock_init(&tags->lock);
578 if (blk_mq_init_bitmaps(&tags->bitmap_tags, &tags->breserved_tags,
579 total_tags, reserved_tags, node,
587 void blk_mq_free_tags(struct blk_mq_tags *tags)
589 sbitmap_queue_free(&tags->bitmap_tags);
590 sbitmap_queue_free(&tags->breserved_tags);
594 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
595 struct blk_mq_tags **tagsptr, unsigned int tdepth,
598 struct blk_mq_tags *tags = *tagsptr;
600 if (tdepth <= tags->nr_reserved_tags)
604 * If we are allowed to grow beyond the original size, allocate
605 * a new set of tags before freeing the old one.
607 if (tdepth > tags->nr_tags) {
608 struct blk_mq_tag_set *set = hctx->queue->tag_set;
609 struct blk_mq_tags *new;
615 * We need some sort of upper limit, set it high enough that
616 * no valid use cases should require more.
618 if (tdepth > MAX_SCHED_RQ)
622 * Only the sbitmap needs resizing since we allocated the max
625 if (blk_mq_is_shared_tags(set->flags))
628 new = blk_mq_alloc_map_and_rqs(set, hctx->queue_num, tdepth);
632 blk_mq_free_map_and_rqs(set, *tagsptr, hctx->queue_num);
636 * Don't need (or can't) update reserved tags here, they
637 * remain static and should never need resizing.
639 sbitmap_queue_resize(&tags->bitmap_tags,
640 tdepth - tags->nr_reserved_tags);
646 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, unsigned int size)
648 struct blk_mq_tags *tags = set->shared_tags;
650 sbitmap_queue_resize(&tags->bitmap_tags, size - set->reserved_tags);
653 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q)
655 sbitmap_queue_resize(&q->sched_shared_tags->bitmap_tags,
656 q->nr_requests - q->tag_set->reserved_tags);
660 * blk_mq_unique_tag() - return a tag that is unique queue-wide
661 * @rq: request for which to compute a unique tag
663 * The tag field in struct request is unique per hardware queue but not over
664 * all hardware queues. Hence this function that returns a tag with the
665 * hardware context index in the upper bits and the per hardware queue tag in
668 * Note: When called for a request that is queued on a non-multiqueue request
669 * queue, the hardware context index is set to zero.
671 u32 blk_mq_unique_tag(struct request *rq)
673 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
674 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
676 EXPORT_SYMBOL(blk_mq_unique_tag);