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-tag.h"
19 * If a previously inactive queue goes active, bump the active user count.
20 * We need to do this before try to allocate driver tag, then even if fail
21 * to get tag when first time, the other shared-tag users could reserve
24 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
26 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
27 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
28 atomic_inc(&hctx->tags->active_queues);
34 * Wakeup all potentially sleeping on tags
36 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
38 sbitmap_queue_wake_all(&tags->bitmap_tags);
40 sbitmap_queue_wake_all(&tags->breserved_tags);
44 * If a previously busy queue goes inactive, potential waiters could now
45 * be allowed to queue. Wake them up and check.
47 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
49 struct blk_mq_tags *tags = hctx->tags;
51 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
54 atomic_dec(&tags->active_queues);
56 blk_mq_tag_wakeup_all(tags, false);
60 * For shared tag users, we track the number of currently active users
61 * and attempt to provide a fair share of the tag depth for each of them.
63 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
64 struct sbitmap_queue *bt)
66 unsigned int depth, users;
68 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
70 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
74 * Don't try dividing an ant
76 if (bt->sb.depth == 1)
79 users = atomic_read(&hctx->tags->active_queues);
84 * Allow at least some tags
86 depth = max((bt->sb.depth + users - 1) / users, 4U);
87 return atomic_read(&hctx->nr_active) < depth;
90 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
91 struct sbitmap_queue *bt)
93 if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
94 !hctx_may_queue(data->hctx, bt))
96 if (data->shallow_depth)
97 return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
99 return __sbitmap_queue_get(bt);
102 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
104 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
105 struct sbitmap_queue *bt;
106 struct sbq_wait_state *ws;
107 DEFINE_SBQ_WAIT(wait);
108 unsigned int tag_offset;
111 if (data->flags & BLK_MQ_REQ_RESERVED) {
112 if (unlikely(!tags->nr_reserved_tags)) {
114 return BLK_MQ_NO_TAG;
116 bt = &tags->breserved_tags;
119 bt = &tags->bitmap_tags;
120 tag_offset = tags->nr_reserved_tags;
123 tag = __blk_mq_get_tag(data, bt);
124 if (tag != BLK_MQ_NO_TAG)
127 if (data->flags & BLK_MQ_REQ_NOWAIT)
128 return BLK_MQ_NO_TAG;
130 ws = bt_wait_ptr(bt, data->hctx);
132 struct sbitmap_queue *bt_prev;
135 * We're out of tags on this hardware queue, kick any
136 * pending IO submits before going to sleep waiting for
139 blk_mq_run_hw_queue(data->hctx, false);
142 * Retry tag allocation after running the hardware queue,
143 * as running the queue may also have found completions.
145 tag = __blk_mq_get_tag(data, bt);
146 if (tag != BLK_MQ_NO_TAG)
149 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
151 tag = __blk_mq_get_tag(data, bt);
152 if (tag != BLK_MQ_NO_TAG)
158 sbitmap_finish_wait(bt, ws, &wait);
160 data->ctx = blk_mq_get_ctx(data->q);
161 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
163 tags = blk_mq_tags_from_data(data);
164 if (data->flags & BLK_MQ_REQ_RESERVED)
165 bt = &tags->breserved_tags;
167 bt = &tags->bitmap_tags;
170 * If destination hw queue is changed, fake wake up on
171 * previous queue for compensating the wake up miss, so
172 * other allocations on previous queue won't be starved.
175 sbitmap_queue_wake_up(bt_prev);
177 ws = bt_wait_ptr(bt, data->hctx);
180 sbitmap_finish_wait(bt, ws, &wait);
184 * Give up this allocation if the hctx is inactive. The caller will
185 * retry on an active hctx.
187 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
188 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
189 return BLK_MQ_NO_TAG;
191 return tag + tag_offset;
194 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
197 if (!blk_mq_tag_is_reserved(tags, tag)) {
198 const int real_tag = tag - tags->nr_reserved_tags;
200 BUG_ON(real_tag >= tags->nr_tags);
201 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
203 BUG_ON(tag >= tags->nr_reserved_tags);
204 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
208 struct bt_iter_data {
209 struct blk_mq_hw_ctx *hctx;
215 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
217 struct bt_iter_data *iter_data = data;
218 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
219 struct blk_mq_tags *tags = hctx->tags;
220 bool reserved = iter_data->reserved;
224 bitnr += tags->nr_reserved_tags;
225 rq = tags->rqs[bitnr];
228 * We can hit rq == NULL here, because the tagging functions
229 * test and set the bit before assigning ->rqs[].
231 if (rq && rq->q == hctx->queue)
232 return iter_data->fn(hctx, rq, iter_data->data, reserved);
237 * bt_for_each - iterate over the requests associated with a hardware queue
238 * @hctx: Hardware queue to examine.
239 * @bt: sbitmap to examine. This is either the breserved_tags member
240 * or the bitmap_tags member of struct blk_mq_tags.
241 * @fn: Pointer to the function that will be called for each request
242 * associated with @hctx that has been assigned a driver tag.
243 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
244 * where rq is a pointer to a request. Return true to continue
245 * iterating tags, false to stop.
246 * @data: Will be passed as third argument to @fn.
247 * @reserved: Indicates whether @bt is the breserved_tags member or the
248 * bitmap_tags member of struct blk_mq_tags.
250 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
251 busy_iter_fn *fn, void *data, bool reserved)
253 struct bt_iter_data iter_data = {
257 .reserved = reserved,
260 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
263 struct bt_tags_iter_data {
264 struct blk_mq_tags *tags;
265 busy_tag_iter_fn *fn;
270 #define BT_TAG_ITER_RESERVED (1 << 0)
271 #define BT_TAG_ITER_STARTED (1 << 1)
273 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
275 struct bt_tags_iter_data *iter_data = data;
276 struct blk_mq_tags *tags = iter_data->tags;
277 bool reserved = iter_data->flags & BT_TAG_ITER_RESERVED;
281 bitnr += tags->nr_reserved_tags;
284 * We can hit rq == NULL here, because the tagging functions
285 * test and set the bit before assining ->rqs[].
287 rq = tags->rqs[bitnr];
290 if ((iter_data->flags & BT_TAG_ITER_STARTED) &&
291 !blk_mq_request_started(rq))
293 return iter_data->fn(rq, iter_data->data, reserved);
297 * bt_tags_for_each - iterate over the requests in a tag map
298 * @tags: Tag map to iterate over.
299 * @bt: sbitmap to examine. This is either the breserved_tags member
300 * or the bitmap_tags member of struct blk_mq_tags.
301 * @fn: Pointer to the function that will be called for each started
302 * request. @fn will be called as follows: @fn(rq, @data,
303 * @reserved) where rq is a pointer to a request. Return true
304 * to continue iterating tags, false to stop.
305 * @data: Will be passed as second argument to @fn.
306 * @flags: BT_TAG_ITER_*
308 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
309 busy_tag_iter_fn *fn, void *data, unsigned int flags)
311 struct bt_tags_iter_data iter_data = {
319 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
322 static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
323 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
325 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
327 if (tags->nr_reserved_tags)
328 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv,
329 flags | BT_TAG_ITER_RESERVED);
330 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, flags);
334 * blk_mq_all_tag_iter - iterate over all requests in a tag map
335 * @tags: Tag map to iterate over.
336 * @fn: Pointer to the function that will be called for each
337 * request. @fn will be called as follows: @fn(rq, @priv,
338 * reserved) where rq is a pointer to a request. 'reserved'
339 * indicates whether or not @rq is a reserved request. Return
340 * true to continue iterating tags, false to stop.
341 * @priv: Will be passed as second argument to @fn.
343 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
346 return __blk_mq_all_tag_iter(tags, fn, priv, 0);
350 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
351 * @tagset: Tag set to iterate over.
352 * @fn: Pointer to the function that will be called for each started
353 * request. @fn will be called as follows: @fn(rq, @priv,
354 * reserved) where rq is a pointer to a request. 'reserved'
355 * indicates whether or not @rq is a reserved request. Return
356 * true to continue iterating tags, false to stop.
357 * @priv: Will be passed as second argument to @fn.
359 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
360 busy_tag_iter_fn *fn, void *priv)
364 for (i = 0; i < tagset->nr_hw_queues; i++) {
365 if (tagset->tags && tagset->tags[i])
366 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
367 BT_TAG_ITER_STARTED);
370 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
372 static bool blk_mq_tagset_count_completed_rqs(struct request *rq,
373 void *data, bool reserved)
375 unsigned *count = data;
377 if (blk_mq_request_completed(rq))
383 * blk_mq_tagset_wait_completed_request - wait until all completed req's
384 * complete funtion is run
385 * @tagset: Tag set to drain completed request
387 * Note: This function has to be run after all IO queues are shutdown
389 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
394 blk_mq_tagset_busy_iter(tagset,
395 blk_mq_tagset_count_completed_rqs, &count);
401 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
404 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
405 * @q: Request queue to examine.
406 * @fn: Pointer to the function that will be called for each request
407 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
408 * reserved) where rq is a pointer to a request and hctx points
409 * to the hardware queue associated with the request. 'reserved'
410 * indicates whether or not @rq is a reserved request.
411 * @priv: Will be passed as third argument to @fn.
413 * Note: if @q->tag_set is shared with other request queues then @fn will be
414 * called for all requests on all queues that share that tag set and not only
415 * for requests associated with @q.
417 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
420 struct blk_mq_hw_ctx *hctx;
424 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
425 * while the queue is frozen. So we can use q_usage_counter to avoid
426 * racing with it. __blk_mq_update_nr_hw_queues() uses
427 * synchronize_rcu() to ensure this function left the critical section
430 if (!percpu_ref_tryget(&q->q_usage_counter))
433 queue_for_each_hw_ctx(q, hctx, i) {
434 struct blk_mq_tags *tags = hctx->tags;
437 * If no software queues are currently mapped to this
438 * hardware queue, there's nothing to check
440 if (!blk_mq_hw_queue_mapped(hctx))
443 if (tags->nr_reserved_tags)
444 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
445 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
450 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
451 bool round_robin, int node)
453 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
457 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
458 int node, int alloc_policy)
460 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
461 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
463 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
465 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
467 goto free_bitmap_tags;
471 sbitmap_queue_free(&tags->bitmap_tags);
477 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
478 unsigned int reserved_tags,
479 int node, int alloc_policy)
481 struct blk_mq_tags *tags;
483 if (total_tags > BLK_MQ_TAG_MAX) {
484 pr_err("blk-mq: tag depth too large\n");
488 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
492 tags->nr_tags = total_tags;
493 tags->nr_reserved_tags = reserved_tags;
495 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
498 void blk_mq_free_tags(struct blk_mq_tags *tags)
500 sbitmap_queue_free(&tags->bitmap_tags);
501 sbitmap_queue_free(&tags->breserved_tags);
505 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
506 struct blk_mq_tags **tagsptr, unsigned int tdepth,
509 struct blk_mq_tags *tags = *tagsptr;
511 if (tdepth <= tags->nr_reserved_tags)
515 * If we are allowed to grow beyond the original size, allocate
516 * a new set of tags before freeing the old one.
518 if (tdepth > tags->nr_tags) {
519 struct blk_mq_tag_set *set = hctx->queue->tag_set;
520 struct blk_mq_tags *new;
527 * We need some sort of upper limit, set it high enough that
528 * no valid use cases should require more.
530 if (tdepth > 16 * BLKDEV_MAX_RQ)
533 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
534 tags->nr_reserved_tags);
537 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
539 blk_mq_free_rq_map(new);
543 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
544 blk_mq_free_rq_map(*tagsptr);
548 * Don't need (or can't) update reserved tags here, they
549 * remain static and should never need resizing.
551 sbitmap_queue_resize(&tags->bitmap_tags,
552 tdepth - tags->nr_reserved_tags);
559 * blk_mq_unique_tag() - return a tag that is unique queue-wide
560 * @rq: request for which to compute a unique tag
562 * The tag field in struct request is unique per hardware queue but not over
563 * all hardware queues. Hence this function that returns a tag with the
564 * hardware context index in the upper bits and the per hardware queue tag in
567 * Note: When called for a request that is queued on a non-multiqueue request
568 * queue, the hardware context index is set to zero.
570 u32 blk_mq_unique_tag(struct request *rq)
572 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
573 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
575 EXPORT_SYMBOL(blk_mq_unique_tag);