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/*
2 * Tag allocation using scalable bitmaps. Uses active queue tracking to support
3 * fairer distribution of tags between multiple submitters when a shared tag map
4 * is used.
5 *
6 * Copyright (C) 2013-2014 Jens Axboe
7 */
8#include <linux/kernel.h>
9#include <linux/module.h>
10
11#include <linux/blk-mq.h>
12#include "blk.h"
13#include "blk-mq.h"
14#include "blk-mq-tag.h"
15
16bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
17{
18 if (!tags)
19 return true;
20
21 return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
22}
23
24/*
25 * If a previously inactive queue goes active, bump the active user count.
26 */
27bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
28{
29 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
30 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
31 atomic_inc(&hctx->tags->active_queues);
32
33 return true;
34}
35
36/*
37 * Wakeup all potentially sleeping on tags
38 */
39void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
40{
41 sbitmap_queue_wake_all(&tags->bitmap_tags);
42 if (include_reserve)
43 sbitmap_queue_wake_all(&tags->breserved_tags);
44}
45
46/*
47 * If a previously busy queue goes inactive, potential waiters could now
48 * be allowed to queue. Wake them up and check.
49 */
50void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
51{
52 struct blk_mq_tags *tags = hctx->tags;
53
54 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
55 return;
56
57 atomic_dec(&tags->active_queues);
58
59 blk_mq_tag_wakeup_all(tags, false);
60}
61
62/*
63 * For shared tag users, we track the number of currently active users
64 * and attempt to provide a fair share of the tag depth for each of them.
65 */
66static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
67 struct sbitmap_queue *bt)
68{
69 unsigned int depth, users;
70
71 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
72 return true;
73 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
74 return true;
75
76 /*
77 * Don't try dividing an ant
78 */
79 if (bt->sb.depth == 1)
80 return true;
81
82 users = atomic_read(&hctx->tags->active_queues);
83 if (!users)
84 return true;
85
86 /*
87 * Allow at least some tags
88 */
89 depth = max((bt->sb.depth + users - 1) / users, 4U);
90 return atomic_read(&hctx->nr_active) < depth;
91}
92
93static int __bt_get(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt)
94{
95 if (!hctx_may_queue(hctx, bt))
96 return -1;
97 return __sbitmap_queue_get(bt);
98}
99
100static int bt_get(struct blk_mq_alloc_data *data, struct sbitmap_queue *bt,
101 struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags)
102{
103 struct sbq_wait_state *ws;
104 DEFINE_WAIT(wait);
105 int tag;
106
107 tag = __bt_get(hctx, bt);
108 if (tag != -1)
109 return tag;
110
111 if (data->flags & BLK_MQ_REQ_NOWAIT)
112 return -1;
113
114 ws = bt_wait_ptr(bt, hctx);
115 do {
116 prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE);
117
118 tag = __bt_get(hctx, bt);
119 if (tag != -1)
120 break;
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. Note that hctx can be NULL here for
126 * reserved tag allocation.
127 */
128 if (hctx)
129 blk_mq_run_hw_queue(hctx, false);
130
131 /*
132 * Retry tag allocation after running the hardware queue,
133 * as running the queue may also have found completions.
134 */
135 tag = __bt_get(hctx, bt);
136 if (tag != -1)
137 break;
138
139 blk_mq_put_ctx(data->ctx);
140
141 io_schedule();
142
143 data->ctx = blk_mq_get_ctx(data->q);
144 data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
145 if (data->flags & BLK_MQ_REQ_RESERVED) {
146 bt = &data->hctx->tags->breserved_tags;
147 } else {
148 hctx = data->hctx;
149 bt = &hctx->tags->bitmap_tags;
150 }
151 finish_wait(&ws->wait, &wait);
152 ws = bt_wait_ptr(bt, hctx);
153 } while (1);
154
155 finish_wait(&ws->wait, &wait);
156 return tag;
157}
158
159static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
160{
161 int tag;
162
163 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
164 data->hctx->tags);
165 if (tag >= 0)
166 return tag + data->hctx->tags->nr_reserved_tags;
167
168 return BLK_MQ_TAG_FAIL;
169}
170
171static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
172{
173 int tag;
174
175 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
176 WARN_ON_ONCE(1);
177 return BLK_MQ_TAG_FAIL;
178 }
179
180 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL,
181 data->hctx->tags);
182 if (tag < 0)
183 return BLK_MQ_TAG_FAIL;
184
185 return tag;
186}
187
188unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
189{
190 if (data->flags & BLK_MQ_REQ_RESERVED)
191 return __blk_mq_get_reserved_tag(data);
192 return __blk_mq_get_tag(data);
193}
194
195void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
196 unsigned int tag)
197{
198 struct blk_mq_tags *tags = hctx->tags;
199
200 if (tag >= tags->nr_reserved_tags) {
201 const int real_tag = tag - tags->nr_reserved_tags;
202
203 BUG_ON(real_tag >= tags->nr_tags);
204 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
205 } else {
206 BUG_ON(tag >= tags->nr_reserved_tags);
207 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
208 }
209}
210
211struct bt_iter_data {
212 struct blk_mq_hw_ctx *hctx;
213 busy_iter_fn *fn;
214 void *data;
215 bool reserved;
216};
217
218static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
219{
220 struct bt_iter_data *iter_data = data;
221 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
222 struct blk_mq_tags *tags = hctx->tags;
223 bool reserved = iter_data->reserved;
224 struct request *rq;
225
226 if (!reserved)
227 bitnr += tags->nr_reserved_tags;
228 rq = tags->rqs[bitnr];
229
230 if (rq->q == hctx->queue)
231 iter_data->fn(hctx, rq, iter_data->data, reserved);
232 return true;
233}
234
235static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
236 busy_iter_fn *fn, void *data, bool reserved)
237{
238 struct bt_iter_data iter_data = {
239 .hctx = hctx,
240 .fn = fn,
241 .data = data,
242 .reserved = reserved,
243 };
244
245 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
246}
247
248struct bt_tags_iter_data {
249 struct blk_mq_tags *tags;
250 busy_tag_iter_fn *fn;
251 void *data;
252 bool reserved;
253};
254
255static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
256{
257 struct bt_tags_iter_data *iter_data = data;
258 struct blk_mq_tags *tags = iter_data->tags;
259 bool reserved = iter_data->reserved;
260 struct request *rq;
261
262 if (!reserved)
263 bitnr += tags->nr_reserved_tags;
264 rq = tags->rqs[bitnr];
265
266 iter_data->fn(rq, iter_data->data, reserved);
267 return true;
268}
269
270static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
271 busy_tag_iter_fn *fn, void *data, bool reserved)
272{
273 struct bt_tags_iter_data iter_data = {
274 .tags = tags,
275 .fn = fn,
276 .data = data,
277 .reserved = reserved,
278 };
279
280 if (tags->rqs)
281 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
282}
283
284static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
285 busy_tag_iter_fn *fn, void *priv)
286{
287 if (tags->nr_reserved_tags)
288 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
289 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
290}
291
292void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
293 busy_tag_iter_fn *fn, void *priv)
294{
295 int i;
296
297 for (i = 0; i < tagset->nr_hw_queues; i++) {
298 if (tagset->tags && tagset->tags[i])
299 blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
300 }
301}
302EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
303
304int blk_mq_reinit_tagset(struct blk_mq_tag_set *set)
305{
306 int i, j, ret = 0;
307
308 if (!set->ops->reinit_request)
309 goto out;
310
311 for (i = 0; i < set->nr_hw_queues; i++) {
312 struct blk_mq_tags *tags = set->tags[i];
313
314 for (j = 0; j < tags->nr_tags; j++) {
315 if (!tags->rqs[j])
316 continue;
317
318 ret = set->ops->reinit_request(set->driver_data,
319 tags->rqs[j]);
320 if (ret)
321 goto out;
322 }
323 }
324
325out:
326 return ret;
327}
328EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset);
329
330void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
331 void *priv)
332{
333 struct blk_mq_hw_ctx *hctx;
334 int i;
335
336
337 queue_for_each_hw_ctx(q, hctx, i) {
338 struct blk_mq_tags *tags = hctx->tags;
339
340 /*
341 * If not software queues are currently mapped to this
342 * hardware queue, there's nothing to check
343 */
344 if (!blk_mq_hw_queue_mapped(hctx))
345 continue;
346
347 if (tags->nr_reserved_tags)
348 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
349 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
350 }
351
352}
353
354static unsigned int bt_unused_tags(const struct sbitmap_queue *bt)
355{
356 return bt->sb.depth - sbitmap_weight(&bt->sb);
357}
358
359static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
360 bool round_robin, int node)
361{
362 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
363 node);
364}
365
366static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
367 int node, int alloc_policy)
368{
369 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
370 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
371
372 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
373 goto free_tags;
374 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
375 node))
376 goto free_bitmap_tags;
377
378 return tags;
379free_bitmap_tags:
380 sbitmap_queue_free(&tags->bitmap_tags);
381free_tags:
382 kfree(tags);
383 return NULL;
384}
385
386struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
387 unsigned int reserved_tags,
388 int node, int alloc_policy)
389{
390 struct blk_mq_tags *tags;
391
392 if (total_tags > BLK_MQ_TAG_MAX) {
393 pr_err("blk-mq: tag depth too large\n");
394 return NULL;
395 }
396
397 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
398 if (!tags)
399 return NULL;
400
401 tags->nr_tags = total_tags;
402 tags->nr_reserved_tags = reserved_tags;
403
404 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
405}
406
407void blk_mq_free_tags(struct blk_mq_tags *tags)
408{
409 sbitmap_queue_free(&tags->bitmap_tags);
410 sbitmap_queue_free(&tags->breserved_tags);
411 kfree(tags);
412}
413
414int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
415{
416 tdepth -= tags->nr_reserved_tags;
417 if (tdepth > tags->nr_tags)
418 return -EINVAL;
419
420 /*
421 * Don't need (or can't) update reserved tags here, they remain
422 * static and should never need resizing.
423 */
424 sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
425
426 blk_mq_tag_wakeup_all(tags, false);
427 return 0;
428}
429
430/**
431 * blk_mq_unique_tag() - return a tag that is unique queue-wide
432 * @rq: request for which to compute a unique tag
433 *
434 * The tag field in struct request is unique per hardware queue but not over
435 * all hardware queues. Hence this function that returns a tag with the
436 * hardware context index in the upper bits and the per hardware queue tag in
437 * the lower bits.
438 *
439 * Note: When called for a request that is queued on a non-multiqueue request
440 * queue, the hardware context index is set to zero.
441 */
442u32 blk_mq_unique_tag(struct request *rq)
443{
444 struct request_queue *q = rq->q;
445 struct blk_mq_hw_ctx *hctx;
446 int hwq = 0;
447
448 if (q->mq_ops) {
449 hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
450 hwq = hctx->queue_num;
451 }
452
453 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
454 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
455}
456EXPORT_SYMBOL(blk_mq_unique_tag);
457
458ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
459{
460 char *orig_page = page;
461 unsigned int free, res;
462
463 if (!tags)
464 return 0;
465
466 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
467 "bits_per_word=%u\n",
468 tags->nr_tags, tags->nr_reserved_tags,
469 1U << tags->bitmap_tags.sb.shift);
470
471 free = bt_unused_tags(&tags->bitmap_tags);
472 res = bt_unused_tags(&tags->breserved_tags);
473
474 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
475 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
476
477 return page - orig_page;
478}