Loading...
1/*
2 * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
3 * over multiple cachelines to avoid ping-pong between multiple submitters
4 * or submitter and completer. Uses rolling wakeups to avoid falling of
5 * the scaling cliff when we run out of tags and have to start putting
6 * submitters to sleep.
7 *
8 * Uses active queue tracking to support fairer distribution of tags
9 * between multiple submitters when a shared tag map is used.
10 *
11 * Copyright (C) 2013-2014 Jens Axboe
12 */
13#include <linux/kernel.h>
14#include <linux/module.h>
15#include <linux/random.h>
16
17#include <linux/blk-mq.h>
18#include "blk.h"
19#include "blk-mq.h"
20#include "blk-mq-tag.h"
21
22static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
23{
24 int i;
25
26 for (i = 0; i < bt->map_nr; i++) {
27 struct blk_align_bitmap *bm = &bt->map[i];
28 int ret;
29
30 ret = find_first_zero_bit(&bm->word, bm->depth);
31 if (ret < bm->depth)
32 return true;
33 }
34
35 return false;
36}
37
38bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
39{
40 if (!tags)
41 return true;
42
43 return bt_has_free_tags(&tags->bitmap_tags);
44}
45
46static inline int bt_index_inc(int index)
47{
48 return (index + 1) & (BT_WAIT_QUEUES - 1);
49}
50
51static inline void bt_index_atomic_inc(atomic_t *index)
52{
53 int old = atomic_read(index);
54 int new = bt_index_inc(old);
55 atomic_cmpxchg(index, old, new);
56}
57
58/*
59 * If a previously inactive queue goes active, bump the active user count.
60 */
61bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
62{
63 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
64 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
65 atomic_inc(&hctx->tags->active_queues);
66
67 return true;
68}
69
70/*
71 * Wakeup all potentially sleeping on tags
72 */
73void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
74{
75 struct blk_mq_bitmap_tags *bt;
76 int i, wake_index;
77
78 /*
79 * Make sure all changes prior to this are visible from other CPUs.
80 */
81 smp_mb();
82 bt = &tags->bitmap_tags;
83 wake_index = atomic_read(&bt->wake_index);
84 for (i = 0; i < BT_WAIT_QUEUES; i++) {
85 struct bt_wait_state *bs = &bt->bs[wake_index];
86
87 if (waitqueue_active(&bs->wait))
88 wake_up(&bs->wait);
89
90 wake_index = bt_index_inc(wake_index);
91 }
92
93 if (include_reserve) {
94 bt = &tags->breserved_tags;
95 if (waitqueue_active(&bt->bs[0].wait))
96 wake_up(&bt->bs[0].wait);
97 }
98}
99
100/*
101 * If a previously busy queue goes inactive, potential waiters could now
102 * be allowed to queue. Wake them up and check.
103 */
104void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
105{
106 struct blk_mq_tags *tags = hctx->tags;
107
108 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
109 return;
110
111 atomic_dec(&tags->active_queues);
112
113 blk_mq_tag_wakeup_all(tags, false);
114}
115
116/*
117 * For shared tag users, we track the number of currently active users
118 * and attempt to provide a fair share of the tag depth for each of them.
119 */
120static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
121 struct blk_mq_bitmap_tags *bt)
122{
123 unsigned int depth, users;
124
125 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
126 return true;
127 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
128 return true;
129
130 /*
131 * Don't try dividing an ant
132 */
133 if (bt->depth == 1)
134 return true;
135
136 users = atomic_read(&hctx->tags->active_queues);
137 if (!users)
138 return true;
139
140 /*
141 * Allow at least some tags
142 */
143 depth = max((bt->depth + users - 1) / users, 4U);
144 return atomic_read(&hctx->nr_active) < depth;
145}
146
147static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag,
148 bool nowrap)
149{
150 int tag, org_last_tag = last_tag;
151
152 while (1) {
153 tag = find_next_zero_bit(&bm->word, bm->depth, last_tag);
154 if (unlikely(tag >= bm->depth)) {
155 /*
156 * We started with an offset, and we didn't reset the
157 * offset to 0 in a failure case, so start from 0 to
158 * exhaust the map.
159 */
160 if (org_last_tag && last_tag && !nowrap) {
161 last_tag = org_last_tag = 0;
162 continue;
163 }
164 return -1;
165 }
166
167 if (!test_and_set_bit(tag, &bm->word))
168 break;
169
170 last_tag = tag + 1;
171 if (last_tag >= bm->depth - 1)
172 last_tag = 0;
173 }
174
175 return tag;
176}
177
178#define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
179
180/*
181 * Straight forward bitmap tag implementation, where each bit is a tag
182 * (cleared == free, and set == busy). The small twist is using per-cpu
183 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
184 * contexts. This enables us to drastically limit the space searched,
185 * without dirtying an extra shared cacheline like we would if we stored
186 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
187 * of that, each word of tags is in a separate cacheline. This means that
188 * multiple users will tend to stick to different cachelines, at least
189 * until the map is exhausted.
190 */
191static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
192 unsigned int *tag_cache, struct blk_mq_tags *tags)
193{
194 unsigned int last_tag, org_last_tag;
195 int index, i, tag;
196
197 if (!hctx_may_queue(hctx, bt))
198 return -1;
199
200 last_tag = org_last_tag = *tag_cache;
201 index = TAG_TO_INDEX(bt, last_tag);
202
203 for (i = 0; i < bt->map_nr; i++) {
204 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag),
205 BT_ALLOC_RR(tags));
206 if (tag != -1) {
207 tag += (index << bt->bits_per_word);
208 goto done;
209 }
210
211 /*
212 * Jump to next index, and reset the last tag to be the
213 * first tag of that index
214 */
215 index++;
216 last_tag = (index << bt->bits_per_word);
217
218 if (index >= bt->map_nr) {
219 index = 0;
220 last_tag = 0;
221 }
222 }
223
224 *tag_cache = 0;
225 return -1;
226
227 /*
228 * Only update the cache from the allocation path, if we ended
229 * up using the specific cached tag.
230 */
231done:
232 if (tag == org_last_tag || unlikely(BT_ALLOC_RR(tags))) {
233 last_tag = tag + 1;
234 if (last_tag >= bt->depth - 1)
235 last_tag = 0;
236
237 *tag_cache = last_tag;
238 }
239
240 return tag;
241}
242
243static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
244 struct blk_mq_hw_ctx *hctx)
245{
246 struct bt_wait_state *bs;
247 int wait_index;
248
249 if (!hctx)
250 return &bt->bs[0];
251
252 wait_index = atomic_read(&hctx->wait_index);
253 bs = &bt->bs[wait_index];
254 bt_index_atomic_inc(&hctx->wait_index);
255 return bs;
256}
257
258static int bt_get(struct blk_mq_alloc_data *data,
259 struct blk_mq_bitmap_tags *bt,
260 struct blk_mq_hw_ctx *hctx,
261 unsigned int *last_tag, struct blk_mq_tags *tags)
262{
263 struct bt_wait_state *bs;
264 DEFINE_WAIT(wait);
265 int tag;
266
267 tag = __bt_get(hctx, bt, last_tag, tags);
268 if (tag != -1)
269 return tag;
270
271 if (data->flags & BLK_MQ_REQ_NOWAIT)
272 return -1;
273
274 bs = bt_wait_ptr(bt, hctx);
275 do {
276 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
277
278 tag = __bt_get(hctx, bt, last_tag, tags);
279 if (tag != -1)
280 break;
281
282 /*
283 * We're out of tags on this hardware queue, kick any
284 * pending IO submits before going to sleep waiting for
285 * some to complete. Note that hctx can be NULL here for
286 * reserved tag allocation.
287 */
288 if (hctx)
289 blk_mq_run_hw_queue(hctx, false);
290
291 /*
292 * Retry tag allocation after running the hardware queue,
293 * as running the queue may also have found completions.
294 */
295 tag = __bt_get(hctx, bt, last_tag, tags);
296 if (tag != -1)
297 break;
298
299 blk_mq_put_ctx(data->ctx);
300
301 io_schedule();
302
303 data->ctx = blk_mq_get_ctx(data->q);
304 data->hctx = data->q->mq_ops->map_queue(data->q,
305 data->ctx->cpu);
306 if (data->flags & BLK_MQ_REQ_RESERVED) {
307 bt = &data->hctx->tags->breserved_tags;
308 } else {
309 last_tag = &data->ctx->last_tag;
310 hctx = data->hctx;
311 bt = &hctx->tags->bitmap_tags;
312 }
313 finish_wait(&bs->wait, &wait);
314 bs = bt_wait_ptr(bt, hctx);
315 } while (1);
316
317 finish_wait(&bs->wait, &wait);
318 return tag;
319}
320
321static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
322{
323 int tag;
324
325 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
326 &data->ctx->last_tag, data->hctx->tags);
327 if (tag >= 0)
328 return tag + data->hctx->tags->nr_reserved_tags;
329
330 return BLK_MQ_TAG_FAIL;
331}
332
333static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
334{
335 int tag, zero = 0;
336
337 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
338 WARN_ON_ONCE(1);
339 return BLK_MQ_TAG_FAIL;
340 }
341
342 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero,
343 data->hctx->tags);
344 if (tag < 0)
345 return BLK_MQ_TAG_FAIL;
346
347 return tag;
348}
349
350unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
351{
352 if (data->flags & BLK_MQ_REQ_RESERVED)
353 return __blk_mq_get_reserved_tag(data);
354 return __blk_mq_get_tag(data);
355}
356
357static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
358{
359 int i, wake_index;
360
361 wake_index = atomic_read(&bt->wake_index);
362 for (i = 0; i < BT_WAIT_QUEUES; i++) {
363 struct bt_wait_state *bs = &bt->bs[wake_index];
364
365 if (waitqueue_active(&bs->wait)) {
366 int o = atomic_read(&bt->wake_index);
367 if (wake_index != o)
368 atomic_cmpxchg(&bt->wake_index, o, wake_index);
369
370 return bs;
371 }
372
373 wake_index = bt_index_inc(wake_index);
374 }
375
376 return NULL;
377}
378
379static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
380{
381 const int index = TAG_TO_INDEX(bt, tag);
382 struct bt_wait_state *bs;
383 int wait_cnt;
384
385 clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
386
387 /* Ensure that the wait list checks occur after clear_bit(). */
388 smp_mb();
389
390 bs = bt_wake_ptr(bt);
391 if (!bs)
392 return;
393
394 wait_cnt = atomic_dec_return(&bs->wait_cnt);
395 if (unlikely(wait_cnt < 0))
396 wait_cnt = atomic_inc_return(&bs->wait_cnt);
397 if (wait_cnt == 0) {
398 atomic_add(bt->wake_cnt, &bs->wait_cnt);
399 bt_index_atomic_inc(&bt->wake_index);
400 wake_up(&bs->wait);
401 }
402}
403
404void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
405 unsigned int *last_tag)
406{
407 struct blk_mq_tags *tags = hctx->tags;
408
409 if (tag >= tags->nr_reserved_tags) {
410 const int real_tag = tag - tags->nr_reserved_tags;
411
412 BUG_ON(real_tag >= tags->nr_tags);
413 bt_clear_tag(&tags->bitmap_tags, real_tag);
414 if (likely(tags->alloc_policy == BLK_TAG_ALLOC_FIFO))
415 *last_tag = real_tag;
416 } else {
417 BUG_ON(tag >= tags->nr_reserved_tags);
418 bt_clear_tag(&tags->breserved_tags, tag);
419 }
420}
421
422static void bt_for_each(struct blk_mq_hw_ctx *hctx,
423 struct blk_mq_bitmap_tags *bt, unsigned int off,
424 busy_iter_fn *fn, void *data, bool reserved)
425{
426 struct request *rq;
427 int bit, i;
428
429 for (i = 0; i < bt->map_nr; i++) {
430 struct blk_align_bitmap *bm = &bt->map[i];
431
432 for (bit = find_first_bit(&bm->word, bm->depth);
433 bit < bm->depth;
434 bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
435 rq = hctx->tags->rqs[off + bit];
436 if (rq->q == hctx->queue)
437 fn(hctx, rq, data, reserved);
438 }
439
440 off += (1 << bt->bits_per_word);
441 }
442}
443
444static void bt_tags_for_each(struct blk_mq_tags *tags,
445 struct blk_mq_bitmap_tags *bt, unsigned int off,
446 busy_tag_iter_fn *fn, void *data, bool reserved)
447{
448 struct request *rq;
449 int bit, i;
450
451 if (!tags->rqs)
452 return;
453 for (i = 0; i < bt->map_nr; i++) {
454 struct blk_align_bitmap *bm = &bt->map[i];
455
456 for (bit = find_first_bit(&bm->word, bm->depth);
457 bit < bm->depth;
458 bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
459 rq = tags->rqs[off + bit];
460 fn(rq, data, reserved);
461 }
462
463 off += (1 << bt->bits_per_word);
464 }
465}
466
467void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
468 void *priv)
469{
470 if (tags->nr_reserved_tags)
471 bt_tags_for_each(tags, &tags->breserved_tags, 0, fn, priv, true);
472 bt_tags_for_each(tags, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
473 false);
474}
475EXPORT_SYMBOL(blk_mq_all_tag_busy_iter);
476
477void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
478 void *priv)
479{
480 struct blk_mq_hw_ctx *hctx;
481 int i;
482
483
484 queue_for_each_hw_ctx(q, hctx, i) {
485 struct blk_mq_tags *tags = hctx->tags;
486
487 /*
488 * If not software queues are currently mapped to this
489 * hardware queue, there's nothing to check
490 */
491 if (!blk_mq_hw_queue_mapped(hctx))
492 continue;
493
494 if (tags->nr_reserved_tags)
495 bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
496 bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
497 false);
498 }
499
500}
501
502static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
503{
504 unsigned int i, used;
505
506 for (i = 0, used = 0; i < bt->map_nr; i++) {
507 struct blk_align_bitmap *bm = &bt->map[i];
508
509 used += bitmap_weight(&bm->word, bm->depth);
510 }
511
512 return bt->depth - used;
513}
514
515static void bt_update_count(struct blk_mq_bitmap_tags *bt,
516 unsigned int depth)
517{
518 unsigned int tags_per_word = 1U << bt->bits_per_word;
519 unsigned int map_depth = depth;
520
521 if (depth) {
522 int i;
523
524 for (i = 0; i < bt->map_nr; i++) {
525 bt->map[i].depth = min(map_depth, tags_per_word);
526 map_depth -= bt->map[i].depth;
527 }
528 }
529
530 bt->wake_cnt = BT_WAIT_BATCH;
531 if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
532 bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
533
534 bt->depth = depth;
535}
536
537static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
538 int node, bool reserved)
539{
540 int i;
541
542 bt->bits_per_word = ilog2(BITS_PER_LONG);
543
544 /*
545 * Depth can be zero for reserved tags, that's not a failure
546 * condition.
547 */
548 if (depth) {
549 unsigned int nr, tags_per_word;
550
551 tags_per_word = (1 << bt->bits_per_word);
552
553 /*
554 * If the tag space is small, shrink the number of tags
555 * per word so we spread over a few cachelines, at least.
556 * If less than 4 tags, just forget about it, it's not
557 * going to work optimally anyway.
558 */
559 if (depth >= 4) {
560 while (tags_per_word * 4 > depth) {
561 bt->bits_per_word--;
562 tags_per_word = (1 << bt->bits_per_word);
563 }
564 }
565
566 nr = ALIGN(depth, tags_per_word) / tags_per_word;
567 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
568 GFP_KERNEL, node);
569 if (!bt->map)
570 return -ENOMEM;
571
572 bt->map_nr = nr;
573 }
574
575 bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
576 if (!bt->bs) {
577 kfree(bt->map);
578 bt->map = NULL;
579 return -ENOMEM;
580 }
581
582 bt_update_count(bt, depth);
583
584 for (i = 0; i < BT_WAIT_QUEUES; i++) {
585 init_waitqueue_head(&bt->bs[i].wait);
586 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
587 }
588
589 return 0;
590}
591
592static void bt_free(struct blk_mq_bitmap_tags *bt)
593{
594 kfree(bt->map);
595 kfree(bt->bs);
596}
597
598static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
599 int node, int alloc_policy)
600{
601 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
602
603 tags->alloc_policy = alloc_policy;
604
605 if (bt_alloc(&tags->bitmap_tags, depth, node, false))
606 goto enomem;
607 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
608 goto enomem;
609
610 return tags;
611enomem:
612 bt_free(&tags->bitmap_tags);
613 kfree(tags);
614 return NULL;
615}
616
617struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
618 unsigned int reserved_tags,
619 int node, int alloc_policy)
620{
621 struct blk_mq_tags *tags;
622
623 if (total_tags > BLK_MQ_TAG_MAX) {
624 pr_err("blk-mq: tag depth too large\n");
625 return NULL;
626 }
627
628 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
629 if (!tags)
630 return NULL;
631
632 if (!zalloc_cpumask_var(&tags->cpumask, GFP_KERNEL)) {
633 kfree(tags);
634 return NULL;
635 }
636
637 tags->nr_tags = total_tags;
638 tags->nr_reserved_tags = reserved_tags;
639
640 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
641}
642
643void blk_mq_free_tags(struct blk_mq_tags *tags)
644{
645 bt_free(&tags->bitmap_tags);
646 bt_free(&tags->breserved_tags);
647 free_cpumask_var(tags->cpumask);
648 kfree(tags);
649}
650
651void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
652{
653 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
654
655 *tag = prandom_u32() % depth;
656}
657
658int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
659{
660 tdepth -= tags->nr_reserved_tags;
661 if (tdepth > tags->nr_tags)
662 return -EINVAL;
663
664 /*
665 * Don't need (or can't) update reserved tags here, they remain
666 * static and should never need resizing.
667 */
668 bt_update_count(&tags->bitmap_tags, tdepth);
669 blk_mq_tag_wakeup_all(tags, false);
670 return 0;
671}
672
673/**
674 * blk_mq_unique_tag() - return a tag that is unique queue-wide
675 * @rq: request for which to compute a unique tag
676 *
677 * The tag field in struct request is unique per hardware queue but not over
678 * all hardware queues. Hence this function that returns a tag with the
679 * hardware context index in the upper bits and the per hardware queue tag in
680 * the lower bits.
681 *
682 * Note: When called for a request that is queued on a non-multiqueue request
683 * queue, the hardware context index is set to zero.
684 */
685u32 blk_mq_unique_tag(struct request *rq)
686{
687 struct request_queue *q = rq->q;
688 struct blk_mq_hw_ctx *hctx;
689 int hwq = 0;
690
691 if (q->mq_ops) {
692 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
693 hwq = hctx->queue_num;
694 }
695
696 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
697 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
698}
699EXPORT_SYMBOL(blk_mq_unique_tag);
700
701ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
702{
703 char *orig_page = page;
704 unsigned int free, res;
705
706 if (!tags)
707 return 0;
708
709 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
710 "bits_per_word=%u\n",
711 tags->nr_tags, tags->nr_reserved_tags,
712 tags->bitmap_tags.bits_per_word);
713
714 free = bt_unused_tags(&tags->bitmap_tags);
715 res = bt_unused_tags(&tags->breserved_tags);
716
717 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
718 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
719
720 return page - orig_page;
721}
1// SPDX-License-Identifier: GPL-2.0
2/*
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
5 * is used.
6 *
7 * Copyright (C) 2013-2014 Jens Axboe
8 */
9#include <linux/kernel.h>
10#include <linux/module.h>
11
12#include <linux/delay.h>
13#include "blk.h"
14#include "blk-mq.h"
15#include "blk-mq-sched.h"
16
17/*
18 * Recalculate wakeup batch when tag is shared by hctx.
19 */
20static void blk_mq_update_wake_batch(struct blk_mq_tags *tags,
21 unsigned int users)
22{
23 if (!users)
24 return;
25
26 sbitmap_queue_recalculate_wake_batch(&tags->bitmap_tags,
27 users);
28 sbitmap_queue_recalculate_wake_batch(&tags->breserved_tags,
29 users);
30}
31
32/*
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
36 * budget for it.
37 */
38void __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
39{
40 unsigned int users;
41 struct blk_mq_tags *tags = hctx->tags;
42
43 /*
44 * calling test_bit() prior to test_and_set_bit() is intentional,
45 * it avoids dirtying the cacheline if the queue is already active.
46 */
47 if (blk_mq_is_shared_tags(hctx->flags)) {
48 struct request_queue *q = hctx->queue;
49
50 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) ||
51 test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
52 return;
53 } else {
54 if (test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) ||
55 test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
56 return;
57 }
58
59 spin_lock_irq(&tags->lock);
60 users = tags->active_queues + 1;
61 WRITE_ONCE(tags->active_queues, users);
62 blk_mq_update_wake_batch(tags, users);
63 spin_unlock_irq(&tags->lock);
64}
65
66/*
67 * Wakeup all potentially sleeping on tags
68 */
69void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
70{
71 sbitmap_queue_wake_all(&tags->bitmap_tags);
72 if (include_reserve)
73 sbitmap_queue_wake_all(&tags->breserved_tags);
74}
75
76/*
77 * If a previously busy queue goes inactive, potential waiters could now
78 * be allowed to queue. Wake them up and check.
79 */
80void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
81{
82 struct blk_mq_tags *tags = hctx->tags;
83 unsigned int users;
84
85 if (blk_mq_is_shared_tags(hctx->flags)) {
86 struct request_queue *q = hctx->queue;
87
88 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
89 &q->queue_flags))
90 return;
91 } else {
92 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
93 return;
94 }
95
96 spin_lock_irq(&tags->lock);
97 users = tags->active_queues - 1;
98 WRITE_ONCE(tags->active_queues, users);
99 blk_mq_update_wake_batch(tags, users);
100 spin_unlock_irq(&tags->lock);
101
102 blk_mq_tag_wakeup_all(tags, false);
103}
104
105static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
106 struct sbitmap_queue *bt)
107{
108 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
109 !hctx_may_queue(data->hctx, bt))
110 return BLK_MQ_NO_TAG;
111
112 if (data->shallow_depth)
113 return sbitmap_queue_get_shallow(bt, data->shallow_depth);
114 else
115 return __sbitmap_queue_get(bt);
116}
117
118unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
119 unsigned int *offset)
120{
121 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
122 struct sbitmap_queue *bt = &tags->bitmap_tags;
123 unsigned long ret;
124
125 if (data->shallow_depth ||data->flags & BLK_MQ_REQ_RESERVED ||
126 data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
127 return 0;
128 ret = __sbitmap_queue_get_batch(bt, nr_tags, offset);
129 *offset += tags->nr_reserved_tags;
130 return ret;
131}
132
133unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
134{
135 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
136 struct sbitmap_queue *bt;
137 struct sbq_wait_state *ws;
138 DEFINE_SBQ_WAIT(wait);
139 unsigned int tag_offset;
140 int tag;
141
142 if (data->flags & BLK_MQ_REQ_RESERVED) {
143 if (unlikely(!tags->nr_reserved_tags)) {
144 WARN_ON_ONCE(1);
145 return BLK_MQ_NO_TAG;
146 }
147 bt = &tags->breserved_tags;
148 tag_offset = 0;
149 } else {
150 bt = &tags->bitmap_tags;
151 tag_offset = tags->nr_reserved_tags;
152 }
153
154 tag = __blk_mq_get_tag(data, bt);
155 if (tag != BLK_MQ_NO_TAG)
156 goto found_tag;
157
158 if (data->flags & BLK_MQ_REQ_NOWAIT)
159 return BLK_MQ_NO_TAG;
160
161 ws = bt_wait_ptr(bt, data->hctx);
162 do {
163 struct sbitmap_queue *bt_prev;
164
165 /*
166 * We're out of tags on this hardware queue, kick any
167 * pending IO submits before going to sleep waiting for
168 * some to complete.
169 */
170 blk_mq_run_hw_queue(data->hctx, false);
171
172 /*
173 * Retry tag allocation after running the hardware queue,
174 * as running the queue may also have found completions.
175 */
176 tag = __blk_mq_get_tag(data, bt);
177 if (tag != BLK_MQ_NO_TAG)
178 break;
179
180 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
181
182 tag = __blk_mq_get_tag(data, bt);
183 if (tag != BLK_MQ_NO_TAG)
184 break;
185
186 bt_prev = bt;
187 io_schedule();
188
189 sbitmap_finish_wait(bt, ws, &wait);
190
191 data->ctx = blk_mq_get_ctx(data->q);
192 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
193 data->ctx);
194 tags = blk_mq_tags_from_data(data);
195 if (data->flags & BLK_MQ_REQ_RESERVED)
196 bt = &tags->breserved_tags;
197 else
198 bt = &tags->bitmap_tags;
199
200 /*
201 * If destination hw queue is changed, fake wake up on
202 * previous queue for compensating the wake up miss, so
203 * other allocations on previous queue won't be starved.
204 */
205 if (bt != bt_prev)
206 sbitmap_queue_wake_up(bt_prev, 1);
207
208 ws = bt_wait_ptr(bt, data->hctx);
209 } while (1);
210
211 sbitmap_finish_wait(bt, ws, &wait);
212
213found_tag:
214 /*
215 * Give up this allocation if the hctx is inactive. The caller will
216 * retry on an active hctx.
217 */
218 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
219 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
220 return BLK_MQ_NO_TAG;
221 }
222 return tag + tag_offset;
223}
224
225void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
226 unsigned int tag)
227{
228 if (!blk_mq_tag_is_reserved(tags, tag)) {
229 const int real_tag = tag - tags->nr_reserved_tags;
230
231 BUG_ON(real_tag >= tags->nr_tags);
232 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
233 } else {
234 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
235 }
236}
237
238void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags)
239{
240 sbitmap_queue_clear_batch(&tags->bitmap_tags, tags->nr_reserved_tags,
241 tag_array, nr_tags);
242}
243
244struct bt_iter_data {
245 struct blk_mq_hw_ctx *hctx;
246 struct request_queue *q;
247 busy_tag_iter_fn *fn;
248 void *data;
249 bool reserved;
250};
251
252static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
253 unsigned int bitnr)
254{
255 struct request *rq;
256 unsigned long flags;
257
258 spin_lock_irqsave(&tags->lock, flags);
259 rq = tags->rqs[bitnr];
260 if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
261 rq = NULL;
262 spin_unlock_irqrestore(&tags->lock, flags);
263 return rq;
264}
265
266static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
267{
268 struct bt_iter_data *iter_data = data;
269 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
270 struct request_queue *q = iter_data->q;
271 struct blk_mq_tag_set *set = q->tag_set;
272 struct blk_mq_tags *tags;
273 struct request *rq;
274 bool ret = true;
275
276 if (blk_mq_is_shared_tags(set->flags))
277 tags = set->shared_tags;
278 else
279 tags = hctx->tags;
280
281 if (!iter_data->reserved)
282 bitnr += tags->nr_reserved_tags;
283 /*
284 * We can hit rq == NULL here, because the tagging functions
285 * test and set the bit before assigning ->rqs[].
286 */
287 rq = blk_mq_find_and_get_req(tags, bitnr);
288 if (!rq)
289 return true;
290
291 if (rq->q == q && (!hctx || rq->mq_hctx == hctx))
292 ret = iter_data->fn(rq, iter_data->data);
293 blk_mq_put_rq_ref(rq);
294 return ret;
295}
296
297/**
298 * bt_for_each - iterate over the requests associated with a hardware queue
299 * @hctx: Hardware queue to examine.
300 * @q: Request queue to examine.
301 * @bt: sbitmap to examine. This is either the breserved_tags member
302 * or the bitmap_tags member of struct blk_mq_tags.
303 * @fn: Pointer to the function that will be called for each request
304 * associated with @hctx that has been assigned a driver tag.
305 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
306 * where rq is a pointer to a request. Return true to continue
307 * iterating tags, false to stop.
308 * @data: Will be passed as third argument to @fn.
309 * @reserved: Indicates whether @bt is the breserved_tags member or the
310 * bitmap_tags member of struct blk_mq_tags.
311 */
312static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct request_queue *q,
313 struct sbitmap_queue *bt, busy_tag_iter_fn *fn,
314 void *data, bool reserved)
315{
316 struct bt_iter_data iter_data = {
317 .hctx = hctx,
318 .fn = fn,
319 .data = data,
320 .reserved = reserved,
321 .q = q,
322 };
323
324 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
325}
326
327struct bt_tags_iter_data {
328 struct blk_mq_tags *tags;
329 busy_tag_iter_fn *fn;
330 void *data;
331 unsigned int flags;
332};
333
334#define BT_TAG_ITER_RESERVED (1 << 0)
335#define BT_TAG_ITER_STARTED (1 << 1)
336#define BT_TAG_ITER_STATIC_RQS (1 << 2)
337
338static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
339{
340 struct bt_tags_iter_data *iter_data = data;
341 struct blk_mq_tags *tags = iter_data->tags;
342 struct request *rq;
343 bool ret = true;
344 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
345
346 if (!(iter_data->flags & BT_TAG_ITER_RESERVED))
347 bitnr += tags->nr_reserved_tags;
348
349 /*
350 * We can hit rq == NULL here, because the tagging functions
351 * test and set the bit before assigning ->rqs[].
352 */
353 if (iter_static_rqs)
354 rq = tags->static_rqs[bitnr];
355 else
356 rq = blk_mq_find_and_get_req(tags, bitnr);
357 if (!rq)
358 return true;
359
360 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
361 blk_mq_request_started(rq))
362 ret = iter_data->fn(rq, iter_data->data);
363 if (!iter_static_rqs)
364 blk_mq_put_rq_ref(rq);
365 return ret;
366}
367
368/**
369 * bt_tags_for_each - iterate over the requests in a tag map
370 * @tags: Tag map to iterate over.
371 * @bt: sbitmap to examine. This is either the breserved_tags member
372 * or the bitmap_tags member of struct blk_mq_tags.
373 * @fn: Pointer to the function that will be called for each started
374 * request. @fn will be called as follows: @fn(rq, @data,
375 * @reserved) where rq is a pointer to a request. Return true
376 * to continue iterating tags, false to stop.
377 * @data: Will be passed as second argument to @fn.
378 * @flags: BT_TAG_ITER_*
379 */
380static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
381 busy_tag_iter_fn *fn, void *data, unsigned int flags)
382{
383 struct bt_tags_iter_data iter_data = {
384 .tags = tags,
385 .fn = fn,
386 .data = data,
387 .flags = flags,
388 };
389
390 if (tags->rqs)
391 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
392}
393
394static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
395 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
396{
397 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
398
399 if (tags->nr_reserved_tags)
400 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv,
401 flags | BT_TAG_ITER_RESERVED);
402 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, flags);
403}
404
405/**
406 * blk_mq_all_tag_iter - iterate over all requests in a tag map
407 * @tags: Tag map to iterate over.
408 * @fn: Pointer to the function that will be called for each
409 * request. @fn will be called as follows: @fn(rq, @priv,
410 * reserved) where rq is a pointer to a request. 'reserved'
411 * indicates whether or not @rq is a reserved request. Return
412 * true to continue iterating tags, false to stop.
413 * @priv: Will be passed as second argument to @fn.
414 *
415 * Caller has to pass the tag map from which requests are allocated.
416 */
417void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
418 void *priv)
419{
420 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
421}
422
423/**
424 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
425 * @tagset: Tag set to iterate over.
426 * @fn: Pointer to the function that will be called for each started
427 * request. @fn will be called as follows: @fn(rq, @priv,
428 * reserved) where rq is a pointer to a request. 'reserved'
429 * indicates whether or not @rq is a reserved request. Return
430 * true to continue iterating tags, false to stop.
431 * @priv: Will be passed as second argument to @fn.
432 *
433 * We grab one request reference before calling @fn and release it after
434 * @fn returns.
435 */
436void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
437 busy_tag_iter_fn *fn, void *priv)
438{
439 unsigned int flags = tagset->flags;
440 int i, nr_tags;
441
442 nr_tags = blk_mq_is_shared_tags(flags) ? 1 : tagset->nr_hw_queues;
443
444 for (i = 0; i < nr_tags; i++) {
445 if (tagset->tags && tagset->tags[i])
446 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
447 BT_TAG_ITER_STARTED);
448 }
449}
450EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
451
452static bool blk_mq_tagset_count_completed_rqs(struct request *rq, void *data)
453{
454 unsigned *count = data;
455
456 if (blk_mq_request_completed(rq))
457 (*count)++;
458 return true;
459}
460
461/**
462 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
463 * completions have finished.
464 * @tagset: Tag set to drain completed request
465 *
466 * Note: This function has to be run after all IO queues are shutdown
467 */
468void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
469{
470 while (true) {
471 unsigned count = 0;
472
473 blk_mq_tagset_busy_iter(tagset,
474 blk_mq_tagset_count_completed_rqs, &count);
475 if (!count)
476 break;
477 msleep(5);
478 }
479}
480EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
481
482/**
483 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
484 * @q: Request queue to examine.
485 * @fn: Pointer to the function that will be called for each request
486 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
487 * reserved) where rq is a pointer to a request and hctx points
488 * to the hardware queue associated with the request. 'reserved'
489 * indicates whether or not @rq is a reserved request.
490 * @priv: Will be passed as third argument to @fn.
491 *
492 * Note: if @q->tag_set is shared with other request queues then @fn will be
493 * called for all requests on all queues that share that tag set and not only
494 * for requests associated with @q.
495 */
496void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
497 void *priv)
498{
499 /*
500 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and hctx_table
501 * while the queue is frozen. So we can use q_usage_counter to avoid
502 * racing with it.
503 */
504 if (!percpu_ref_tryget(&q->q_usage_counter))
505 return;
506
507 if (blk_mq_is_shared_tags(q->tag_set->flags)) {
508 struct blk_mq_tags *tags = q->tag_set->shared_tags;
509 struct sbitmap_queue *bresv = &tags->breserved_tags;
510 struct sbitmap_queue *btags = &tags->bitmap_tags;
511
512 if (tags->nr_reserved_tags)
513 bt_for_each(NULL, q, bresv, fn, priv, true);
514 bt_for_each(NULL, q, btags, fn, priv, false);
515 } else {
516 struct blk_mq_hw_ctx *hctx;
517 unsigned long i;
518
519 queue_for_each_hw_ctx(q, hctx, i) {
520 struct blk_mq_tags *tags = hctx->tags;
521 struct sbitmap_queue *bresv = &tags->breserved_tags;
522 struct sbitmap_queue *btags = &tags->bitmap_tags;
523
524 /*
525 * If no software queues are currently mapped to this
526 * hardware queue, there's nothing to check
527 */
528 if (!blk_mq_hw_queue_mapped(hctx))
529 continue;
530
531 if (tags->nr_reserved_tags)
532 bt_for_each(hctx, q, bresv, fn, priv, true);
533 bt_for_each(hctx, q, btags, fn, priv, false);
534 }
535 }
536 blk_queue_exit(q);
537}
538
539static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
540 bool round_robin, int node)
541{
542 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
543 node);
544}
545
546int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
547 struct sbitmap_queue *breserved_tags,
548 unsigned int queue_depth, unsigned int reserved,
549 int node, int alloc_policy)
550{
551 unsigned int depth = queue_depth - reserved;
552 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
553
554 if (bt_alloc(bitmap_tags, depth, round_robin, node))
555 return -ENOMEM;
556 if (bt_alloc(breserved_tags, reserved, round_robin, node))
557 goto free_bitmap_tags;
558
559 return 0;
560
561free_bitmap_tags:
562 sbitmap_queue_free(bitmap_tags);
563 return -ENOMEM;
564}
565
566struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
567 unsigned int reserved_tags,
568 int node, int alloc_policy)
569{
570 struct blk_mq_tags *tags;
571
572 if (total_tags > BLK_MQ_TAG_MAX) {
573 pr_err("blk-mq: tag depth too large\n");
574 return NULL;
575 }
576
577 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
578 if (!tags)
579 return NULL;
580
581 tags->nr_tags = total_tags;
582 tags->nr_reserved_tags = reserved_tags;
583 spin_lock_init(&tags->lock);
584
585 if (blk_mq_init_bitmaps(&tags->bitmap_tags, &tags->breserved_tags,
586 total_tags, reserved_tags, node,
587 alloc_policy) < 0) {
588 kfree(tags);
589 return NULL;
590 }
591 return tags;
592}
593
594void blk_mq_free_tags(struct blk_mq_tags *tags)
595{
596 sbitmap_queue_free(&tags->bitmap_tags);
597 sbitmap_queue_free(&tags->breserved_tags);
598 kfree(tags);
599}
600
601int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
602 struct blk_mq_tags **tagsptr, unsigned int tdepth,
603 bool can_grow)
604{
605 struct blk_mq_tags *tags = *tagsptr;
606
607 if (tdepth <= tags->nr_reserved_tags)
608 return -EINVAL;
609
610 /*
611 * If we are allowed to grow beyond the original size, allocate
612 * a new set of tags before freeing the old one.
613 */
614 if (tdepth > tags->nr_tags) {
615 struct blk_mq_tag_set *set = hctx->queue->tag_set;
616 struct blk_mq_tags *new;
617
618 if (!can_grow)
619 return -EINVAL;
620
621 /*
622 * We need some sort of upper limit, set it high enough that
623 * no valid use cases should require more.
624 */
625 if (tdepth > MAX_SCHED_RQ)
626 return -EINVAL;
627
628 /*
629 * Only the sbitmap needs resizing since we allocated the max
630 * initially.
631 */
632 if (blk_mq_is_shared_tags(set->flags))
633 return 0;
634
635 new = blk_mq_alloc_map_and_rqs(set, hctx->queue_num, tdepth);
636 if (!new)
637 return -ENOMEM;
638
639 blk_mq_free_map_and_rqs(set, *tagsptr, hctx->queue_num);
640 *tagsptr = new;
641 } else {
642 /*
643 * Don't need (or can't) update reserved tags here, they
644 * remain static and should never need resizing.
645 */
646 sbitmap_queue_resize(&tags->bitmap_tags,
647 tdepth - tags->nr_reserved_tags);
648 }
649
650 return 0;
651}
652
653void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, unsigned int size)
654{
655 struct blk_mq_tags *tags = set->shared_tags;
656
657 sbitmap_queue_resize(&tags->bitmap_tags, size - set->reserved_tags);
658}
659
660void blk_mq_tag_update_sched_shared_tags(struct request_queue *q)
661{
662 sbitmap_queue_resize(&q->sched_shared_tags->bitmap_tags,
663 q->nr_requests - q->tag_set->reserved_tags);
664}
665
666/**
667 * blk_mq_unique_tag() - return a tag that is unique queue-wide
668 * @rq: request for which to compute a unique tag
669 *
670 * The tag field in struct request is unique per hardware queue but not over
671 * all hardware queues. Hence this function that returns a tag with the
672 * hardware context index in the upper bits and the per hardware queue tag in
673 * the lower bits.
674 *
675 * Note: When called for a request that is queued on a non-multiqueue request
676 * queue, the hardware context index is set to zero.
677 */
678u32 blk_mq_unique_tag(struct request *rq)
679{
680 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
681 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
682}
683EXPORT_SYMBOL(blk_mq_unique_tag);