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)
43 * calling test_bit() prior to test_and_set_bit() is intentional,
44 * it avoids dirtying the cacheline if the queue is already active.
46 if (blk_mq_is_shared_tags(hctx->flags)) {
47 struct request_queue *q = hctx->queue;
49 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) ||
50 test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
53 if (test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) ||
54 test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
58 users = atomic_inc_return(&hctx->tags->active_queues);
60 blk_mq_update_wake_batch(hctx->tags, users);
64 * Wakeup all potentially sleeping on tags
66 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
68 sbitmap_queue_wake_all(&tags->bitmap_tags);
70 sbitmap_queue_wake_all(&tags->breserved_tags);
74 * If a previously busy queue goes inactive, potential waiters could now
75 * be allowed to queue. Wake them up and check.
77 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
79 struct blk_mq_tags *tags = hctx->tags;
82 if (blk_mq_is_shared_tags(hctx->flags)) {
83 struct request_queue *q = hctx->queue;
85 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
89 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
93 users = atomic_dec_return(&tags->active_queues);
95 blk_mq_update_wake_batch(tags, users);
97 blk_mq_tag_wakeup_all(tags, false);
100 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
101 struct sbitmap_queue *bt)
103 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
104 !hctx_may_queue(data->hctx, bt))
105 return BLK_MQ_NO_TAG;
107 if (data->shallow_depth)
108 return sbitmap_queue_get_shallow(bt, data->shallow_depth);
110 return __sbitmap_queue_get(bt);
113 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
114 unsigned int *offset)
116 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
117 struct sbitmap_queue *bt = &tags->bitmap_tags;
120 if (data->shallow_depth ||data->flags & BLK_MQ_REQ_RESERVED ||
121 data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
123 ret = __sbitmap_queue_get_batch(bt, nr_tags, offset);
124 *offset += tags->nr_reserved_tags;
128 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
130 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
131 struct sbitmap_queue *bt;
132 struct sbq_wait_state *ws;
133 DEFINE_SBQ_WAIT(wait);
134 unsigned int tag_offset;
137 if (data->flags & BLK_MQ_REQ_RESERVED) {
138 if (unlikely(!tags->nr_reserved_tags)) {
140 return BLK_MQ_NO_TAG;
142 bt = &tags->breserved_tags;
145 bt = &tags->bitmap_tags;
146 tag_offset = tags->nr_reserved_tags;
149 tag = __blk_mq_get_tag(data, bt);
150 if (tag != BLK_MQ_NO_TAG)
153 if (data->flags & BLK_MQ_REQ_NOWAIT)
154 return BLK_MQ_NO_TAG;
156 ws = bt_wait_ptr(bt, data->hctx);
158 struct sbitmap_queue *bt_prev;
161 * We're out of tags on this hardware queue, kick any
162 * pending IO submits before going to sleep waiting for
165 blk_mq_run_hw_queue(data->hctx, false);
168 * Retry tag allocation after running the hardware queue,
169 * as running the queue may also have found completions.
171 tag = __blk_mq_get_tag(data, bt);
172 if (tag != BLK_MQ_NO_TAG)
175 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
177 tag = __blk_mq_get_tag(data, bt);
178 if (tag != BLK_MQ_NO_TAG)
184 sbitmap_finish_wait(bt, ws, &wait);
186 data->ctx = blk_mq_get_ctx(data->q);
187 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
189 tags = blk_mq_tags_from_data(data);
190 if (data->flags & BLK_MQ_REQ_RESERVED)
191 bt = &tags->breserved_tags;
193 bt = &tags->bitmap_tags;
196 * If destination hw queue is changed, fake wake up on
197 * previous queue for compensating the wake up miss, so
198 * other allocations on previous queue won't be starved.
201 sbitmap_queue_wake_up(bt_prev, 1);
203 ws = bt_wait_ptr(bt, data->hctx);
206 sbitmap_finish_wait(bt, ws, &wait);
210 * Give up this allocation if the hctx is inactive. The caller will
211 * retry on an active hctx.
213 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
214 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
215 return BLK_MQ_NO_TAG;
217 return tag + tag_offset;
220 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
223 if (!blk_mq_tag_is_reserved(tags, tag)) {
224 const int real_tag = tag - tags->nr_reserved_tags;
226 BUG_ON(real_tag >= tags->nr_tags);
227 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
229 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
233 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags)
235 sbitmap_queue_clear_batch(&tags->bitmap_tags, tags->nr_reserved_tags,
239 struct bt_iter_data {
240 struct blk_mq_hw_ctx *hctx;
241 struct request_queue *q;
242 busy_tag_iter_fn *fn;
247 static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
253 spin_lock_irqsave(&tags->lock, flags);
254 rq = tags->rqs[bitnr];
255 if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
257 spin_unlock_irqrestore(&tags->lock, flags);
261 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
263 struct bt_iter_data *iter_data = data;
264 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
265 struct request_queue *q = iter_data->q;
266 struct blk_mq_tag_set *set = q->tag_set;
267 struct blk_mq_tags *tags;
271 if (blk_mq_is_shared_tags(set->flags))
272 tags = set->shared_tags;
276 if (!iter_data->reserved)
277 bitnr += tags->nr_reserved_tags;
279 * We can hit rq == NULL here, because the tagging functions
280 * test and set the bit before assigning ->rqs[].
282 rq = blk_mq_find_and_get_req(tags, bitnr);
286 if (rq->q == q && (!hctx || rq->mq_hctx == hctx))
287 ret = iter_data->fn(rq, iter_data->data);
288 blk_mq_put_rq_ref(rq);
293 * bt_for_each - iterate over the requests associated with a hardware queue
294 * @hctx: Hardware queue to examine.
295 * @q: Request queue to examine.
296 * @bt: sbitmap to examine. This is either the breserved_tags member
297 * or the bitmap_tags member of struct blk_mq_tags.
298 * @fn: Pointer to the function that will be called for each request
299 * associated with @hctx that has been assigned a driver tag.
300 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
301 * where rq is a pointer to a request. Return true to continue
302 * iterating tags, false to stop.
303 * @data: Will be passed as third argument to @fn.
304 * @reserved: Indicates whether @bt is the breserved_tags member or the
305 * bitmap_tags member of struct blk_mq_tags.
307 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct request_queue *q,
308 struct sbitmap_queue *bt, busy_tag_iter_fn *fn,
309 void *data, bool reserved)
311 struct bt_iter_data iter_data = {
315 .reserved = reserved,
319 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
322 struct bt_tags_iter_data {
323 struct blk_mq_tags *tags;
324 busy_tag_iter_fn *fn;
329 #define BT_TAG_ITER_RESERVED (1 << 0)
330 #define BT_TAG_ITER_STARTED (1 << 1)
331 #define BT_TAG_ITER_STATIC_RQS (1 << 2)
333 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
335 struct bt_tags_iter_data *iter_data = data;
336 struct blk_mq_tags *tags = iter_data->tags;
339 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
341 if (!(iter_data->flags & BT_TAG_ITER_RESERVED))
342 bitnr += tags->nr_reserved_tags;
345 * We can hit rq == NULL here, because the tagging functions
346 * test and set the bit before assigning ->rqs[].
349 rq = tags->static_rqs[bitnr];
351 rq = blk_mq_find_and_get_req(tags, bitnr);
355 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
356 blk_mq_request_started(rq))
357 ret = iter_data->fn(rq, iter_data->data);
358 if (!iter_static_rqs)
359 blk_mq_put_rq_ref(rq);
364 * bt_tags_for_each - iterate over the requests in a tag map
365 * @tags: Tag map to iterate over.
366 * @bt: sbitmap to examine. This is either the breserved_tags member
367 * or the bitmap_tags member of struct blk_mq_tags.
368 * @fn: Pointer to the function that will be called for each started
369 * request. @fn will be called as follows: @fn(rq, @data,
370 * @reserved) where rq is a pointer to a request. Return true
371 * to continue iterating tags, false to stop.
372 * @data: Will be passed as second argument to @fn.
373 * @flags: BT_TAG_ITER_*
375 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
376 busy_tag_iter_fn *fn, void *data, unsigned int flags)
378 struct bt_tags_iter_data iter_data = {
386 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
389 static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
390 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
392 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
394 if (tags->nr_reserved_tags)
395 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv,
396 flags | BT_TAG_ITER_RESERVED);
397 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, flags);
401 * blk_mq_all_tag_iter - iterate over all requests in a tag map
402 * @tags: Tag map to iterate over.
403 * @fn: Pointer to the function that will be called for each
404 * request. @fn will be called as follows: @fn(rq, @priv,
405 * reserved) where rq is a pointer to a request. 'reserved'
406 * indicates whether or not @rq is a reserved request. Return
407 * true to continue iterating tags, false to stop.
408 * @priv: Will be passed as second argument to @fn.
410 * Caller has to pass the tag map from which requests are allocated.
412 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
415 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
419 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
420 * @tagset: Tag set to iterate over.
421 * @fn: Pointer to the function that will be called for each started
422 * request. @fn will be called as follows: @fn(rq, @priv,
423 * reserved) where rq is a pointer to a request. 'reserved'
424 * indicates whether or not @rq is a reserved request. Return
425 * true to continue iterating tags, false to stop.
426 * @priv: Will be passed as second argument to @fn.
428 * We grab one request reference before calling @fn and release it after
431 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
432 busy_tag_iter_fn *fn, void *priv)
434 unsigned int flags = tagset->flags;
437 nr_tags = blk_mq_is_shared_tags(flags) ? 1 : tagset->nr_hw_queues;
439 for (i = 0; i < nr_tags; i++) {
440 if (tagset->tags && tagset->tags[i])
441 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
442 BT_TAG_ITER_STARTED);
445 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
447 static bool blk_mq_tagset_count_completed_rqs(struct request *rq, void *data)
449 unsigned *count = data;
451 if (blk_mq_request_completed(rq))
457 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
458 * completions have finished.
459 * @tagset: Tag set to drain completed request
461 * Note: This function has to be run after all IO queues are shutdown
463 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
468 blk_mq_tagset_busy_iter(tagset,
469 blk_mq_tagset_count_completed_rqs, &count);
475 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
478 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
479 * @q: Request queue to examine.
480 * @fn: Pointer to the function that will be called for each request
481 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
482 * reserved) where rq is a pointer to a request and hctx points
483 * to the hardware queue associated with the request. 'reserved'
484 * indicates whether or not @rq is a reserved request.
485 * @priv: Will be passed as third argument to @fn.
487 * Note: if @q->tag_set is shared with other request queues then @fn will be
488 * called for all requests on all queues that share that tag set and not only
489 * for requests associated with @q.
491 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
495 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and hctx_table
496 * while the queue is frozen. So we can use q_usage_counter to avoid
499 if (!percpu_ref_tryget(&q->q_usage_counter))
502 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
503 struct blk_mq_tags *tags = q->tag_set->shared_tags;
504 struct sbitmap_queue *bresv = &tags->breserved_tags;
505 struct sbitmap_queue *btags = &tags->bitmap_tags;
507 if (tags->nr_reserved_tags)
508 bt_for_each(NULL, q, bresv, fn, priv, true);
509 bt_for_each(NULL, q, btags, fn, priv, false);
511 struct blk_mq_hw_ctx *hctx;
514 queue_for_each_hw_ctx(q, hctx, i) {
515 struct blk_mq_tags *tags = hctx->tags;
516 struct sbitmap_queue *bresv = &tags->breserved_tags;
517 struct sbitmap_queue *btags = &tags->bitmap_tags;
520 * If no software queues are currently mapped to this
521 * hardware queue, there's nothing to check
523 if (!blk_mq_hw_queue_mapped(hctx))
526 if (tags->nr_reserved_tags)
527 bt_for_each(hctx, q, bresv, fn, priv, true);
528 bt_for_each(hctx, q, btags, fn, priv, false);
534 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
535 bool round_robin, int node)
537 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
541 int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
542 struct sbitmap_queue *breserved_tags,
543 unsigned int queue_depth, unsigned int reserved,
544 int node, int alloc_policy)
546 unsigned int depth = queue_depth - reserved;
547 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
549 if (bt_alloc(bitmap_tags, depth, round_robin, node))
551 if (bt_alloc(breserved_tags, reserved, round_robin, node))
552 goto free_bitmap_tags;
557 sbitmap_queue_free(bitmap_tags);
561 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
562 unsigned int reserved_tags,
563 int node, int alloc_policy)
565 struct blk_mq_tags *tags;
567 if (total_tags > BLK_MQ_TAG_MAX) {
568 pr_err("blk-mq: tag depth too large\n");
572 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
576 tags->nr_tags = total_tags;
577 tags->nr_reserved_tags = reserved_tags;
578 spin_lock_init(&tags->lock);
580 if (blk_mq_init_bitmaps(&tags->bitmap_tags, &tags->breserved_tags,
581 total_tags, reserved_tags, node,
589 void blk_mq_free_tags(struct blk_mq_tags *tags)
591 sbitmap_queue_free(&tags->bitmap_tags);
592 sbitmap_queue_free(&tags->breserved_tags);
596 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
597 struct blk_mq_tags **tagsptr, unsigned int tdepth,
600 struct blk_mq_tags *tags = *tagsptr;
602 if (tdepth <= tags->nr_reserved_tags)
606 * If we are allowed to grow beyond the original size, allocate
607 * a new set of tags before freeing the old one.
609 if (tdepth > tags->nr_tags) {
610 struct blk_mq_tag_set *set = hctx->queue->tag_set;
611 struct blk_mq_tags *new;
617 * We need some sort of upper limit, set it high enough that
618 * no valid use cases should require more.
620 if (tdepth > MAX_SCHED_RQ)
624 * Only the sbitmap needs resizing since we allocated the max
627 if (blk_mq_is_shared_tags(set->flags))
630 new = blk_mq_alloc_map_and_rqs(set, hctx->queue_num, tdepth);
634 blk_mq_free_map_and_rqs(set, *tagsptr, hctx->queue_num);
638 * Don't need (or can't) update reserved tags here, they
639 * remain static and should never need resizing.
641 sbitmap_queue_resize(&tags->bitmap_tags,
642 tdepth - tags->nr_reserved_tags);
648 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, unsigned int size)
650 struct blk_mq_tags *tags = set->shared_tags;
652 sbitmap_queue_resize(&tags->bitmap_tags, size - set->reserved_tags);
655 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q)
657 sbitmap_queue_resize(&q->sched_shared_tags->bitmap_tags,
658 q->nr_requests - q->tag_set->reserved_tags);
662 * blk_mq_unique_tag() - return a tag that is unique queue-wide
663 * @rq: request for which to compute a unique tag
665 * The tag field in struct request is unique per hardware queue but not over
666 * all hardware queues. Hence this function that returns a tag with the
667 * hardware context index in the upper bits and the per hardware queue tag in
670 * Note: When called for a request that is queued on a non-multiqueue request
671 * queue, the hardware context index is set to zero.
673 u32 blk_mq_unique_tag(struct request *rq)
675 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
676 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
678 EXPORT_SYMBOL(blk_mq_unique_tag);