Loading...
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/blk-mq.h>
13#include <linux/delay.h>
14#include "blk.h"
15#include "blk-mq.h"
16#include "blk-mq-sched.h"
17#include "blk-mq-tag.h"
18
19/*
20 * Recalculate wakeup batch when tag is shared by hctx.
21 */
22static void blk_mq_update_wake_batch(struct blk_mq_tags *tags,
23 unsigned int users)
24{
25 if (!users)
26 return;
27
28 sbitmap_queue_recalculate_wake_batch(&tags->bitmap_tags,
29 users);
30 sbitmap_queue_recalculate_wake_batch(&tags->breserved_tags,
31 users);
32}
33
34/*
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
38 * budget for it.
39 */
40void __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
41{
42 unsigned int users;
43
44 if (blk_mq_is_shared_tags(hctx->flags)) {
45 struct request_queue *q = hctx->queue;
46
47 if (test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
48 return;
49 set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags);
50 } else {
51 if (test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
52 return;
53 set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state);
54 }
55
56 users = atomic_inc_return(&hctx->tags->active_queues);
57
58 blk_mq_update_wake_batch(hctx->tags, users);
59}
60
61/*
62 * Wakeup all potentially sleeping on tags
63 */
64void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
65{
66 sbitmap_queue_wake_all(&tags->bitmap_tags);
67 if (include_reserve)
68 sbitmap_queue_wake_all(&tags->breserved_tags);
69}
70
71/*
72 * If a previously busy queue goes inactive, potential waiters could now
73 * be allowed to queue. Wake them up and check.
74 */
75void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
76{
77 struct blk_mq_tags *tags = hctx->tags;
78 unsigned int users;
79
80 if (blk_mq_is_shared_tags(hctx->flags)) {
81 struct request_queue *q = hctx->queue;
82
83 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
84 &q->queue_flags))
85 return;
86 } else {
87 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
88 return;
89 }
90
91 users = atomic_dec_return(&tags->active_queues);
92
93 blk_mq_update_wake_batch(tags, users);
94
95 blk_mq_tag_wakeup_all(tags, false);
96}
97
98static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
99 struct sbitmap_queue *bt)
100{
101 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
102 !hctx_may_queue(data->hctx, bt))
103 return BLK_MQ_NO_TAG;
104
105 if (data->shallow_depth)
106 return sbitmap_queue_get_shallow(bt, data->shallow_depth);
107 else
108 return __sbitmap_queue_get(bt);
109}
110
111unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
112 unsigned int *offset)
113{
114 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
115 struct sbitmap_queue *bt = &tags->bitmap_tags;
116 unsigned long ret;
117
118 if (data->shallow_depth ||data->flags & BLK_MQ_REQ_RESERVED ||
119 data->hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
120 return 0;
121 ret = __sbitmap_queue_get_batch(bt, nr_tags, offset);
122 *offset += tags->nr_reserved_tags;
123 return ret;
124}
125
126unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
127{
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;
133 int tag;
134
135 if (data->flags & BLK_MQ_REQ_RESERVED) {
136 if (unlikely(!tags->nr_reserved_tags)) {
137 WARN_ON_ONCE(1);
138 return BLK_MQ_NO_TAG;
139 }
140 bt = &tags->breserved_tags;
141 tag_offset = 0;
142 } else {
143 bt = &tags->bitmap_tags;
144 tag_offset = tags->nr_reserved_tags;
145 }
146
147 tag = __blk_mq_get_tag(data, bt);
148 if (tag != BLK_MQ_NO_TAG)
149 goto found_tag;
150
151 if (data->flags & BLK_MQ_REQ_NOWAIT)
152 return BLK_MQ_NO_TAG;
153
154 ws = bt_wait_ptr(bt, data->hctx);
155 do {
156 struct sbitmap_queue *bt_prev;
157
158 /*
159 * We're out of tags on this hardware queue, kick any
160 * pending IO submits before going to sleep waiting for
161 * some to complete.
162 */
163 blk_mq_run_hw_queue(data->hctx, false);
164
165 /*
166 * Retry tag allocation after running the hardware queue,
167 * as running the queue may also have found completions.
168 */
169 tag = __blk_mq_get_tag(data, bt);
170 if (tag != BLK_MQ_NO_TAG)
171 break;
172
173 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
174
175 tag = __blk_mq_get_tag(data, bt);
176 if (tag != BLK_MQ_NO_TAG)
177 break;
178
179 bt_prev = bt;
180 io_schedule();
181
182 sbitmap_finish_wait(bt, ws, &wait);
183
184 data->ctx = blk_mq_get_ctx(data->q);
185 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
186 data->ctx);
187 tags = blk_mq_tags_from_data(data);
188 if (data->flags & BLK_MQ_REQ_RESERVED)
189 bt = &tags->breserved_tags;
190 else
191 bt = &tags->bitmap_tags;
192
193 /*
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.
197 */
198 if (bt != bt_prev)
199 sbitmap_queue_wake_up(bt_prev, 1);
200
201 ws = bt_wait_ptr(bt, data->hctx);
202 } while (1);
203
204 sbitmap_finish_wait(bt, ws, &wait);
205
206found_tag:
207 /*
208 * Give up this allocation if the hctx is inactive. The caller will
209 * retry on an active hctx.
210 */
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;
214 }
215 return tag + tag_offset;
216}
217
218void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
219 unsigned int tag)
220{
221 if (!blk_mq_tag_is_reserved(tags, tag)) {
222 const int real_tag = tag - tags->nr_reserved_tags;
223
224 BUG_ON(real_tag >= tags->nr_tags);
225 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
226 } else {
227 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
228 }
229}
230
231void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags)
232{
233 sbitmap_queue_clear_batch(&tags->bitmap_tags, tags->nr_reserved_tags,
234 tag_array, nr_tags);
235}
236
237struct bt_iter_data {
238 struct blk_mq_hw_ctx *hctx;
239 struct request_queue *q;
240 busy_tag_iter_fn *fn;
241 void *data;
242 bool reserved;
243};
244
245static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
246 unsigned int bitnr)
247{
248 struct request *rq;
249 unsigned long flags;
250
251 spin_lock_irqsave(&tags->lock, flags);
252 rq = tags->rqs[bitnr];
253 if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
254 rq = NULL;
255 spin_unlock_irqrestore(&tags->lock, flags);
256 return rq;
257}
258
259static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
260{
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;
266 struct request *rq;
267 bool ret = true;
268
269 if (blk_mq_is_shared_tags(set->flags))
270 tags = set->shared_tags;
271 else
272 tags = hctx->tags;
273
274 if (!iter_data->reserved)
275 bitnr += tags->nr_reserved_tags;
276 /*
277 * We can hit rq == NULL here, because the tagging functions
278 * test and set the bit before assigning ->rqs[].
279 */
280 rq = blk_mq_find_and_get_req(tags, bitnr);
281 if (!rq)
282 return true;
283
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);
287 return ret;
288}
289
290/**
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.
304 */
305static 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)
308{
309 struct bt_iter_data iter_data = {
310 .hctx = hctx,
311 .fn = fn,
312 .data = data,
313 .reserved = reserved,
314 .q = q,
315 };
316
317 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
318}
319
320struct bt_tags_iter_data {
321 struct blk_mq_tags *tags;
322 busy_tag_iter_fn *fn;
323 void *data;
324 unsigned int flags;
325};
326
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)
330
331static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
332{
333 struct bt_tags_iter_data *iter_data = data;
334 struct blk_mq_tags *tags = iter_data->tags;
335 struct request *rq;
336 bool ret = true;
337 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
338
339 if (!(iter_data->flags & BT_TAG_ITER_RESERVED))
340 bitnr += tags->nr_reserved_tags;
341
342 /*
343 * We can hit rq == NULL here, because the tagging functions
344 * test and set the bit before assigning ->rqs[].
345 */
346 if (iter_static_rqs)
347 rq = tags->static_rqs[bitnr];
348 else
349 rq = blk_mq_find_and_get_req(tags, bitnr);
350 if (!rq)
351 return true;
352
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);
358 return ret;
359}
360
361/**
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_*
372 */
373static 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)
375{
376 struct bt_tags_iter_data iter_data = {
377 .tags = tags,
378 .fn = fn,
379 .data = data,
380 .flags = flags,
381 };
382
383 if (tags->rqs)
384 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
385}
386
387static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
388 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
389{
390 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
391
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);
396}
397
398/**
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.
407 *
408 * Caller has to pass the tag map from which requests are allocated.
409 */
410void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
411 void *priv)
412{
413 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
414}
415
416/**
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.
425 *
426 * We grab one request reference before calling @fn and release it after
427 * @fn returns.
428 */
429void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
430 busy_tag_iter_fn *fn, void *priv)
431{
432 unsigned int flags = tagset->flags;
433 int i, nr_tags;
434
435 nr_tags = blk_mq_is_shared_tags(flags) ? 1 : tagset->nr_hw_queues;
436
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);
441 }
442}
443EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
444
445static bool blk_mq_tagset_count_completed_rqs(struct request *rq, void *data)
446{
447 unsigned *count = data;
448
449 if (blk_mq_request_completed(rq))
450 (*count)++;
451 return true;
452}
453
454/**
455 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
456 * completions have finished.
457 * @tagset: Tag set to drain completed request
458 *
459 * Note: This function has to be run after all IO queues are shutdown
460 */
461void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
462{
463 while (true) {
464 unsigned count = 0;
465
466 blk_mq_tagset_busy_iter(tagset,
467 blk_mq_tagset_count_completed_rqs, &count);
468 if (!count)
469 break;
470 msleep(5);
471 }
472}
473EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
474
475/**
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.
484 *
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.
488 */
489void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
490 void *priv)
491{
492 /*
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
495 * racing with it.
496 */
497 if (!percpu_ref_tryget(&q->q_usage_counter))
498 return;
499
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;
504
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);
508 } else {
509 struct blk_mq_hw_ctx *hctx;
510 unsigned long i;
511
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;
516
517 /*
518 * If no software queues are currently mapped to this
519 * hardware queue, there's nothing to check
520 */
521 if (!blk_mq_hw_queue_mapped(hctx))
522 continue;
523
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);
527 }
528 }
529 blk_queue_exit(q);
530}
531
532static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
533 bool round_robin, int node)
534{
535 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
536 node);
537}
538
539int 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)
543{
544 unsigned int depth = queue_depth - reserved;
545 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
546
547 if (bt_alloc(bitmap_tags, depth, round_robin, node))
548 return -ENOMEM;
549 if (bt_alloc(breserved_tags, reserved, round_robin, node))
550 goto free_bitmap_tags;
551
552 return 0;
553
554free_bitmap_tags:
555 sbitmap_queue_free(bitmap_tags);
556 return -ENOMEM;
557}
558
559struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
560 unsigned int reserved_tags,
561 int node, int alloc_policy)
562{
563 struct blk_mq_tags *tags;
564
565 if (total_tags > BLK_MQ_TAG_MAX) {
566 pr_err("blk-mq: tag depth too large\n");
567 return NULL;
568 }
569
570 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
571 if (!tags)
572 return NULL;
573
574 tags->nr_tags = total_tags;
575 tags->nr_reserved_tags = reserved_tags;
576 spin_lock_init(&tags->lock);
577
578 if (blk_mq_init_bitmaps(&tags->bitmap_tags, &tags->breserved_tags,
579 total_tags, reserved_tags, node,
580 alloc_policy) < 0) {
581 kfree(tags);
582 return NULL;
583 }
584 return tags;
585}
586
587void blk_mq_free_tags(struct blk_mq_tags *tags)
588{
589 sbitmap_queue_free(&tags->bitmap_tags);
590 sbitmap_queue_free(&tags->breserved_tags);
591 kfree(tags);
592}
593
594int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
595 struct blk_mq_tags **tagsptr, unsigned int tdepth,
596 bool can_grow)
597{
598 struct blk_mq_tags *tags = *tagsptr;
599
600 if (tdepth <= tags->nr_reserved_tags)
601 return -EINVAL;
602
603 /*
604 * If we are allowed to grow beyond the original size, allocate
605 * a new set of tags before freeing the old one.
606 */
607 if (tdepth > tags->nr_tags) {
608 struct blk_mq_tag_set *set = hctx->queue->tag_set;
609 struct blk_mq_tags *new;
610
611 if (!can_grow)
612 return -EINVAL;
613
614 /*
615 * We need some sort of upper limit, set it high enough that
616 * no valid use cases should require more.
617 */
618 if (tdepth > MAX_SCHED_RQ)
619 return -EINVAL;
620
621 /*
622 * Only the sbitmap needs resizing since we allocated the max
623 * initially.
624 */
625 if (blk_mq_is_shared_tags(set->flags))
626 return 0;
627
628 new = blk_mq_alloc_map_and_rqs(set, hctx->queue_num, tdepth);
629 if (!new)
630 return -ENOMEM;
631
632 blk_mq_free_map_and_rqs(set, *tagsptr, hctx->queue_num);
633 *tagsptr = new;
634 } else {
635 /*
636 * Don't need (or can't) update reserved tags here, they
637 * remain static and should never need resizing.
638 */
639 sbitmap_queue_resize(&tags->bitmap_tags,
640 tdepth - tags->nr_reserved_tags);
641 }
642
643 return 0;
644}
645
646void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, unsigned int size)
647{
648 struct blk_mq_tags *tags = set->shared_tags;
649
650 sbitmap_queue_resize(&tags->bitmap_tags, size - set->reserved_tags);
651}
652
653void blk_mq_tag_update_sched_shared_tags(struct request_queue *q)
654{
655 sbitmap_queue_resize(&q->sched_shared_tags->bitmap_tags,
656 q->nr_requests - q->tag_set->reserved_tags);
657}
658
659/**
660 * blk_mq_unique_tag() - return a tag that is unique queue-wide
661 * @rq: request for which to compute a unique tag
662 *
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
666 * the lower bits.
667 *
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.
670 */
671u32 blk_mq_unique_tag(struct request *rq)
672{
673 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
674 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
675}
676EXPORT_SYMBOL(blk_mq_unique_tag);
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/blk-mq.h>
13#include <linux/delay.h>
14#include "blk.h"
15#include "blk-mq.h"
16#include "blk-mq-sched.h"
17#include "blk-mq-tag.h"
18
19/*
20 * If a previously inactive queue goes active, bump the active user count.
21 * We need to do this before try to allocate driver tag, then even if fail
22 * to get tag when first time, the other shared-tag users could reserve
23 * budget for it.
24 */
25bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
26{
27 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
28 struct request_queue *q = hctx->queue;
29 struct blk_mq_tag_set *set = q->tag_set;
30
31 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) &&
32 !test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
33 atomic_inc(&set->active_queues_shared_sbitmap);
34 } else {
35 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
36 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
37 atomic_inc(&hctx->tags->active_queues);
38 }
39
40 return true;
41}
42
43/*
44 * Wakeup all potentially sleeping on tags
45 */
46void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
47{
48 sbitmap_queue_wake_all(tags->bitmap_tags);
49 if (include_reserve)
50 sbitmap_queue_wake_all(tags->breserved_tags);
51}
52
53/*
54 * If a previously busy queue goes inactive, potential waiters could now
55 * be allowed to queue. Wake them up and check.
56 */
57void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
58{
59 struct blk_mq_tags *tags = hctx->tags;
60 struct request_queue *q = hctx->queue;
61 struct blk_mq_tag_set *set = q->tag_set;
62
63 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
64 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
65 &q->queue_flags))
66 return;
67 atomic_dec(&set->active_queues_shared_sbitmap);
68 } else {
69 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
70 return;
71 atomic_dec(&tags->active_queues);
72 }
73
74 blk_mq_tag_wakeup_all(tags, false);
75}
76
77static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
78 struct sbitmap_queue *bt)
79{
80 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
81 !hctx_may_queue(data->hctx, bt))
82 return BLK_MQ_NO_TAG;
83
84 if (data->shallow_depth)
85 return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
86 else
87 return __sbitmap_queue_get(bt);
88}
89
90unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
91{
92 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
93 struct sbitmap_queue *bt;
94 struct sbq_wait_state *ws;
95 DEFINE_SBQ_WAIT(wait);
96 unsigned int tag_offset;
97 int tag;
98
99 if (data->flags & BLK_MQ_REQ_RESERVED) {
100 if (unlikely(!tags->nr_reserved_tags)) {
101 WARN_ON_ONCE(1);
102 return BLK_MQ_NO_TAG;
103 }
104 bt = tags->breserved_tags;
105 tag_offset = 0;
106 } else {
107 bt = tags->bitmap_tags;
108 tag_offset = tags->nr_reserved_tags;
109 }
110
111 tag = __blk_mq_get_tag(data, bt);
112 if (tag != BLK_MQ_NO_TAG)
113 goto found_tag;
114
115 if (data->flags & BLK_MQ_REQ_NOWAIT)
116 return BLK_MQ_NO_TAG;
117
118 ws = bt_wait_ptr(bt, data->hctx);
119 do {
120 struct sbitmap_queue *bt_prev;
121
122 /*
123 * We're out of tags on this hardware queue, kick any
124 * pending IO submits before going to sleep waiting for
125 * some to complete.
126 */
127 blk_mq_run_hw_queue(data->hctx, false);
128
129 /*
130 * Retry tag allocation after running the hardware queue,
131 * as running the queue may also have found completions.
132 */
133 tag = __blk_mq_get_tag(data, bt);
134 if (tag != BLK_MQ_NO_TAG)
135 break;
136
137 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
138
139 tag = __blk_mq_get_tag(data, bt);
140 if (tag != BLK_MQ_NO_TAG)
141 break;
142
143 bt_prev = bt;
144 io_schedule();
145
146 sbitmap_finish_wait(bt, ws, &wait);
147
148 data->ctx = blk_mq_get_ctx(data->q);
149 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
150 data->ctx);
151 tags = blk_mq_tags_from_data(data);
152 if (data->flags & BLK_MQ_REQ_RESERVED)
153 bt = tags->breserved_tags;
154 else
155 bt = tags->bitmap_tags;
156
157 /*
158 * If destination hw queue is changed, fake wake up on
159 * previous queue for compensating the wake up miss, so
160 * other allocations on previous queue won't be starved.
161 */
162 if (bt != bt_prev)
163 sbitmap_queue_wake_up(bt_prev);
164
165 ws = bt_wait_ptr(bt, data->hctx);
166 } while (1);
167
168 sbitmap_finish_wait(bt, ws, &wait);
169
170found_tag:
171 /*
172 * Give up this allocation if the hctx is inactive. The caller will
173 * retry on an active hctx.
174 */
175 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
176 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
177 return BLK_MQ_NO_TAG;
178 }
179 return tag + tag_offset;
180}
181
182void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
183 unsigned int tag)
184{
185 if (!blk_mq_tag_is_reserved(tags, tag)) {
186 const int real_tag = tag - tags->nr_reserved_tags;
187
188 BUG_ON(real_tag >= tags->nr_tags);
189 sbitmap_queue_clear(tags->bitmap_tags, real_tag, ctx->cpu);
190 } else {
191 BUG_ON(tag >= tags->nr_reserved_tags);
192 sbitmap_queue_clear(tags->breserved_tags, tag, ctx->cpu);
193 }
194}
195
196struct bt_iter_data {
197 struct blk_mq_hw_ctx *hctx;
198 busy_iter_fn *fn;
199 void *data;
200 bool reserved;
201};
202
203static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
204 unsigned int bitnr)
205{
206 struct request *rq;
207 unsigned long flags;
208
209 spin_lock_irqsave(&tags->lock, flags);
210 rq = tags->rqs[bitnr];
211 if (!rq || rq->tag != bitnr || !refcount_inc_not_zero(&rq->ref))
212 rq = NULL;
213 spin_unlock_irqrestore(&tags->lock, flags);
214 return rq;
215}
216
217static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
218{
219 struct bt_iter_data *iter_data = data;
220 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
221 struct blk_mq_tags *tags = hctx->tags;
222 bool reserved = iter_data->reserved;
223 struct request *rq;
224 bool ret = true;
225
226 if (!reserved)
227 bitnr += tags->nr_reserved_tags;
228 /*
229 * We can hit rq == NULL here, because the tagging functions
230 * test and set the bit before assigning ->rqs[].
231 */
232 rq = blk_mq_find_and_get_req(tags, bitnr);
233 if (!rq)
234 return true;
235
236 if (rq->q == hctx->queue && rq->mq_hctx == hctx)
237 ret = iter_data->fn(hctx, rq, iter_data->data, reserved);
238 blk_mq_put_rq_ref(rq);
239 return ret;
240}
241
242/**
243 * bt_for_each - iterate over the requests associated with a hardware queue
244 * @hctx: Hardware queue to examine.
245 * @bt: sbitmap to examine. This is either the breserved_tags member
246 * or the bitmap_tags member of struct blk_mq_tags.
247 * @fn: Pointer to the function that will be called for each request
248 * associated with @hctx that has been assigned a driver tag.
249 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
250 * where rq is a pointer to a request. Return true to continue
251 * iterating tags, false to stop.
252 * @data: Will be passed as third argument to @fn.
253 * @reserved: Indicates whether @bt is the breserved_tags member or the
254 * bitmap_tags member of struct blk_mq_tags.
255 */
256static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
257 busy_iter_fn *fn, void *data, bool reserved)
258{
259 struct bt_iter_data iter_data = {
260 .hctx = hctx,
261 .fn = fn,
262 .data = data,
263 .reserved = reserved,
264 };
265
266 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
267}
268
269struct bt_tags_iter_data {
270 struct blk_mq_tags *tags;
271 busy_tag_iter_fn *fn;
272 void *data;
273 unsigned int flags;
274};
275
276#define BT_TAG_ITER_RESERVED (1 << 0)
277#define BT_TAG_ITER_STARTED (1 << 1)
278#define BT_TAG_ITER_STATIC_RQS (1 << 2)
279
280static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
281{
282 struct bt_tags_iter_data *iter_data = data;
283 struct blk_mq_tags *tags = iter_data->tags;
284 bool reserved = iter_data->flags & BT_TAG_ITER_RESERVED;
285 struct request *rq;
286 bool ret = true;
287 bool iter_static_rqs = !!(iter_data->flags & BT_TAG_ITER_STATIC_RQS);
288
289 if (!reserved)
290 bitnr += tags->nr_reserved_tags;
291
292 /*
293 * We can hit rq == NULL here, because the tagging functions
294 * test and set the bit before assigning ->rqs[].
295 */
296 if (iter_static_rqs)
297 rq = tags->static_rqs[bitnr];
298 else
299 rq = blk_mq_find_and_get_req(tags, bitnr);
300 if (!rq)
301 return true;
302
303 if (!(iter_data->flags & BT_TAG_ITER_STARTED) ||
304 blk_mq_request_started(rq))
305 ret = iter_data->fn(rq, iter_data->data, reserved);
306 if (!iter_static_rqs)
307 blk_mq_put_rq_ref(rq);
308 return ret;
309}
310
311/**
312 * bt_tags_for_each - iterate over the requests in a tag map
313 * @tags: Tag map to iterate over.
314 * @bt: sbitmap to examine. This is either the breserved_tags member
315 * or the bitmap_tags member of struct blk_mq_tags.
316 * @fn: Pointer to the function that will be called for each started
317 * request. @fn will be called as follows: @fn(rq, @data,
318 * @reserved) where rq is a pointer to a request. Return true
319 * to continue iterating tags, false to stop.
320 * @data: Will be passed as second argument to @fn.
321 * @flags: BT_TAG_ITER_*
322 */
323static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
324 busy_tag_iter_fn *fn, void *data, unsigned int flags)
325{
326 struct bt_tags_iter_data iter_data = {
327 .tags = tags,
328 .fn = fn,
329 .data = data,
330 .flags = flags,
331 };
332
333 if (tags->rqs)
334 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
335}
336
337static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
338 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
339{
340 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
341
342 if (tags->nr_reserved_tags)
343 bt_tags_for_each(tags, tags->breserved_tags, fn, priv,
344 flags | BT_TAG_ITER_RESERVED);
345 bt_tags_for_each(tags, tags->bitmap_tags, fn, priv, flags);
346}
347
348/**
349 * blk_mq_all_tag_iter - iterate over all requests in a tag map
350 * @tags: Tag map to iterate over.
351 * @fn: Pointer to the function that will be called for each
352 * request. @fn will be called as follows: @fn(rq, @priv,
353 * reserved) where rq is a pointer to a request. 'reserved'
354 * indicates whether or not @rq is a reserved request. Return
355 * true to continue iterating tags, false to stop.
356 * @priv: Will be passed as second argument to @fn.
357 *
358 * Caller has to pass the tag map from which requests are allocated.
359 */
360void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
361 void *priv)
362{
363 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
364}
365
366/**
367 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
368 * @tagset: Tag set to iterate over.
369 * @fn: Pointer to the function that will be called for each started
370 * request. @fn will be called as follows: @fn(rq, @priv,
371 * reserved) where rq is a pointer to a request. 'reserved'
372 * indicates whether or not @rq is a reserved request. Return
373 * true to continue iterating tags, false to stop.
374 * @priv: Will be passed as second argument to @fn.
375 *
376 * We grab one request reference before calling @fn and release it after
377 * @fn returns.
378 */
379void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
380 busy_tag_iter_fn *fn, void *priv)
381{
382 int i;
383
384 for (i = 0; i < tagset->nr_hw_queues; i++) {
385 if (tagset->tags && tagset->tags[i])
386 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
387 BT_TAG_ITER_STARTED);
388 }
389}
390EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
391
392static bool blk_mq_tagset_count_completed_rqs(struct request *rq,
393 void *data, bool reserved)
394{
395 unsigned *count = data;
396
397 if (blk_mq_request_completed(rq))
398 (*count)++;
399 return true;
400}
401
402/**
403 * blk_mq_tagset_wait_completed_request - Wait until all scheduled request
404 * completions have finished.
405 * @tagset: Tag set to drain completed request
406 *
407 * Note: This function has to be run after all IO queues are shutdown
408 */
409void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
410{
411 while (true) {
412 unsigned count = 0;
413
414 blk_mq_tagset_busy_iter(tagset,
415 blk_mq_tagset_count_completed_rqs, &count);
416 if (!count)
417 break;
418 msleep(5);
419 }
420}
421EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
422
423/**
424 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
425 * @q: Request queue to examine.
426 * @fn: Pointer to the function that will be called for each request
427 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
428 * reserved) where rq is a pointer to a request and hctx points
429 * to the hardware queue associated with the request. 'reserved'
430 * indicates whether or not @rq is a reserved request.
431 * @priv: Will be passed as third argument to @fn.
432 *
433 * Note: if @q->tag_set is shared with other request queues then @fn will be
434 * called for all requests on all queues that share that tag set and not only
435 * for requests associated with @q.
436 */
437void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
438 void *priv)
439{
440 struct blk_mq_hw_ctx *hctx;
441 int i;
442
443 /*
444 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
445 * while the queue is frozen. So we can use q_usage_counter to avoid
446 * racing with it.
447 */
448 if (!percpu_ref_tryget(&q->q_usage_counter))
449 return;
450
451 queue_for_each_hw_ctx(q, hctx, i) {
452 struct blk_mq_tags *tags = hctx->tags;
453
454 /*
455 * If no software queues are currently mapped to this
456 * hardware queue, there's nothing to check
457 */
458 if (!blk_mq_hw_queue_mapped(hctx))
459 continue;
460
461 if (tags->nr_reserved_tags)
462 bt_for_each(hctx, tags->breserved_tags, fn, priv, true);
463 bt_for_each(hctx, tags->bitmap_tags, fn, priv, false);
464 }
465 blk_queue_exit(q);
466}
467
468static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
469 bool round_robin, int node)
470{
471 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
472 node);
473}
474
475int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
476 struct sbitmap_queue *breserved_tags,
477 unsigned int queue_depth, unsigned int reserved,
478 int node, int alloc_policy)
479{
480 unsigned int depth = queue_depth - reserved;
481 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
482
483 if (bt_alloc(bitmap_tags, depth, round_robin, node))
484 return -ENOMEM;
485 if (bt_alloc(breserved_tags, reserved, round_robin, node))
486 goto free_bitmap_tags;
487
488 return 0;
489
490free_bitmap_tags:
491 sbitmap_queue_free(bitmap_tags);
492 return -ENOMEM;
493}
494
495static int blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
496 int node, int alloc_policy)
497{
498 int ret;
499
500 ret = blk_mq_init_bitmaps(&tags->__bitmap_tags,
501 &tags->__breserved_tags,
502 tags->nr_tags, tags->nr_reserved_tags,
503 node, alloc_policy);
504 if (ret)
505 return ret;
506
507 tags->bitmap_tags = &tags->__bitmap_tags;
508 tags->breserved_tags = &tags->__breserved_tags;
509
510 return 0;
511}
512
513int blk_mq_init_shared_sbitmap(struct blk_mq_tag_set *set)
514{
515 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags);
516 int i, ret;
517
518 ret = blk_mq_init_bitmaps(&set->__bitmap_tags, &set->__breserved_tags,
519 set->queue_depth, set->reserved_tags,
520 set->numa_node, alloc_policy);
521 if (ret)
522 return ret;
523
524 for (i = 0; i < set->nr_hw_queues; i++) {
525 struct blk_mq_tags *tags = set->tags[i];
526
527 tags->bitmap_tags = &set->__bitmap_tags;
528 tags->breserved_tags = &set->__breserved_tags;
529 }
530
531 return 0;
532}
533
534void blk_mq_exit_shared_sbitmap(struct blk_mq_tag_set *set)
535{
536 sbitmap_queue_free(&set->__bitmap_tags);
537 sbitmap_queue_free(&set->__breserved_tags);
538}
539
540struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
541 unsigned int reserved_tags,
542 int node, unsigned int flags)
543{
544 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(flags);
545 struct blk_mq_tags *tags;
546
547 if (total_tags > BLK_MQ_TAG_MAX) {
548 pr_err("blk-mq: tag depth too large\n");
549 return NULL;
550 }
551
552 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
553 if (!tags)
554 return NULL;
555
556 tags->nr_tags = total_tags;
557 tags->nr_reserved_tags = reserved_tags;
558 spin_lock_init(&tags->lock);
559
560 if (blk_mq_is_sbitmap_shared(flags))
561 return tags;
562
563 if (blk_mq_init_bitmap_tags(tags, node, alloc_policy) < 0) {
564 kfree(tags);
565 return NULL;
566 }
567 return tags;
568}
569
570void blk_mq_free_tags(struct blk_mq_tags *tags, unsigned int flags)
571{
572 if (!blk_mq_is_sbitmap_shared(flags)) {
573 sbitmap_queue_free(tags->bitmap_tags);
574 sbitmap_queue_free(tags->breserved_tags);
575 }
576 kfree(tags);
577}
578
579int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
580 struct blk_mq_tags **tagsptr, unsigned int tdepth,
581 bool can_grow)
582{
583 struct blk_mq_tags *tags = *tagsptr;
584
585 if (tdepth <= tags->nr_reserved_tags)
586 return -EINVAL;
587
588 /*
589 * If we are allowed to grow beyond the original size, allocate
590 * a new set of tags before freeing the old one.
591 */
592 if (tdepth > tags->nr_tags) {
593 struct blk_mq_tag_set *set = hctx->queue->tag_set;
594 struct blk_mq_tags *new;
595 bool ret;
596
597 if (!can_grow)
598 return -EINVAL;
599
600 /*
601 * We need some sort of upper limit, set it high enough that
602 * no valid use cases should require more.
603 */
604 if (tdepth > MAX_SCHED_RQ)
605 return -EINVAL;
606
607 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
608 tags->nr_reserved_tags, set->flags);
609 if (!new)
610 return -ENOMEM;
611 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
612 if (ret) {
613 blk_mq_free_rq_map(new, set->flags);
614 return -ENOMEM;
615 }
616
617 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
618 blk_mq_free_rq_map(*tagsptr, set->flags);
619 *tagsptr = new;
620 } else {
621 /*
622 * Don't need (or can't) update reserved tags here, they
623 * remain static and should never need resizing.
624 */
625 sbitmap_queue_resize(tags->bitmap_tags,
626 tdepth - tags->nr_reserved_tags);
627 }
628
629 return 0;
630}
631
632void blk_mq_tag_resize_shared_sbitmap(struct blk_mq_tag_set *set, unsigned int size)
633{
634 sbitmap_queue_resize(&set->__bitmap_tags, size - set->reserved_tags);
635}
636
637/**
638 * blk_mq_unique_tag() - return a tag that is unique queue-wide
639 * @rq: request for which to compute a unique tag
640 *
641 * The tag field in struct request is unique per hardware queue but not over
642 * all hardware queues. Hence this function that returns a tag with the
643 * hardware context index in the upper bits and the per hardware queue tag in
644 * the lower bits.
645 *
646 * Note: When called for a request that is queued on a non-multiqueue request
647 * queue, the hardware context index is set to zero.
648 */
649u32 blk_mq_unique_tag(struct request *rq)
650{
651 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
652 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
653}
654EXPORT_SYMBOL(blk_mq_unique_tag);