]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - block/blk-mq-tag.c
Merge tag 's390-5.11-4' of git://git.kernel.org/pub/scm/linux/kernel/git/s390/linux
[mirror_ubuntu-hirsute-kernel.git] / block / blk-mq-tag.c
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-tag.h"
17
18 /*
19 * If a previously inactive queue goes active, bump the active user count.
20 * We need to do this before try to allocate driver tag, then even if fail
21 * to get tag when first time, the other shared-tag users could reserve
22 * budget for it.
23 */
24 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
25 {
26 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
27 struct request_queue *q = hctx->queue;
28 struct blk_mq_tag_set *set = q->tag_set;
29
30 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags) &&
31 !test_and_set_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
32 atomic_inc(&set->active_queues_shared_sbitmap);
33 } else {
34 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
35 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
36 atomic_inc(&hctx->tags->active_queues);
37 }
38
39 return true;
40 }
41
42 /*
43 * Wakeup all potentially sleeping on tags
44 */
45 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
46 {
47 sbitmap_queue_wake_all(tags->bitmap_tags);
48 if (include_reserve)
49 sbitmap_queue_wake_all(tags->breserved_tags);
50 }
51
52 /*
53 * If a previously busy queue goes inactive, potential waiters could now
54 * be allowed to queue. Wake them up and check.
55 */
56 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
57 {
58 struct blk_mq_tags *tags = hctx->tags;
59 struct request_queue *q = hctx->queue;
60 struct blk_mq_tag_set *set = q->tag_set;
61
62 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
63 if (!test_and_clear_bit(QUEUE_FLAG_HCTX_ACTIVE,
64 &q->queue_flags))
65 return;
66 atomic_dec(&set->active_queues_shared_sbitmap);
67 } else {
68 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
69 return;
70 atomic_dec(&tags->active_queues);
71 }
72
73 blk_mq_tag_wakeup_all(tags, false);
74 }
75
76 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
77 struct sbitmap_queue *bt)
78 {
79 if (!data->q->elevator && !(data->flags & BLK_MQ_REQ_RESERVED) &&
80 !hctx_may_queue(data->hctx, bt))
81 return BLK_MQ_NO_TAG;
82
83 if (data->shallow_depth)
84 return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
85 else
86 return __sbitmap_queue_get(bt);
87 }
88
89 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
90 {
91 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
92 struct sbitmap_queue *bt;
93 struct sbq_wait_state *ws;
94 DEFINE_SBQ_WAIT(wait);
95 unsigned int tag_offset;
96 int tag;
97
98 if (data->flags & BLK_MQ_REQ_RESERVED) {
99 if (unlikely(!tags->nr_reserved_tags)) {
100 WARN_ON_ONCE(1);
101 return BLK_MQ_NO_TAG;
102 }
103 bt = tags->breserved_tags;
104 tag_offset = 0;
105 } else {
106 bt = tags->bitmap_tags;
107 tag_offset = tags->nr_reserved_tags;
108 }
109
110 tag = __blk_mq_get_tag(data, bt);
111 if (tag != BLK_MQ_NO_TAG)
112 goto found_tag;
113
114 if (data->flags & BLK_MQ_REQ_NOWAIT)
115 return BLK_MQ_NO_TAG;
116
117 ws = bt_wait_ptr(bt, data->hctx);
118 do {
119 struct sbitmap_queue *bt_prev;
120
121 /*
122 * We're out of tags on this hardware queue, kick any
123 * pending IO submits before going to sleep waiting for
124 * some to complete.
125 */
126 blk_mq_run_hw_queue(data->hctx, false);
127
128 /*
129 * Retry tag allocation after running the hardware queue,
130 * as running the queue may also have found completions.
131 */
132 tag = __blk_mq_get_tag(data, bt);
133 if (tag != BLK_MQ_NO_TAG)
134 break;
135
136 sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
137
138 tag = __blk_mq_get_tag(data, bt);
139 if (tag != BLK_MQ_NO_TAG)
140 break;
141
142 bt_prev = bt;
143 io_schedule();
144
145 sbitmap_finish_wait(bt, ws, &wait);
146
147 data->ctx = blk_mq_get_ctx(data->q);
148 data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
149 data->ctx);
150 tags = blk_mq_tags_from_data(data);
151 if (data->flags & BLK_MQ_REQ_RESERVED)
152 bt = tags->breserved_tags;
153 else
154 bt = tags->bitmap_tags;
155
156 /*
157 * If destination hw queue is changed, fake wake up on
158 * previous queue for compensating the wake up miss, so
159 * other allocations on previous queue won't be starved.
160 */
161 if (bt != bt_prev)
162 sbitmap_queue_wake_up(bt_prev);
163
164 ws = bt_wait_ptr(bt, data->hctx);
165 } while (1);
166
167 sbitmap_finish_wait(bt, ws, &wait);
168
169 found_tag:
170 /*
171 * Give up this allocation if the hctx is inactive. The caller will
172 * retry on an active hctx.
173 */
174 if (unlikely(test_bit(BLK_MQ_S_INACTIVE, &data->hctx->state))) {
175 blk_mq_put_tag(tags, data->ctx, tag + tag_offset);
176 return BLK_MQ_NO_TAG;
177 }
178 return tag + tag_offset;
179 }
180
181 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
182 unsigned int tag)
183 {
184 if (!blk_mq_tag_is_reserved(tags, tag)) {
185 const int real_tag = tag - tags->nr_reserved_tags;
186
187 BUG_ON(real_tag >= tags->nr_tags);
188 sbitmap_queue_clear(tags->bitmap_tags, real_tag, ctx->cpu);
189 } else {
190 BUG_ON(tag >= tags->nr_reserved_tags);
191 sbitmap_queue_clear(tags->breserved_tags, tag, ctx->cpu);
192 }
193 }
194
195 struct bt_iter_data {
196 struct blk_mq_hw_ctx *hctx;
197 busy_iter_fn *fn;
198 void *data;
199 bool reserved;
200 };
201
202 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
203 {
204 struct bt_iter_data *iter_data = data;
205 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
206 struct blk_mq_tags *tags = hctx->tags;
207 bool reserved = iter_data->reserved;
208 struct request *rq;
209
210 if (!reserved)
211 bitnr += tags->nr_reserved_tags;
212 rq = tags->rqs[bitnr];
213
214 /*
215 * We can hit rq == NULL here, because the tagging functions
216 * test and set the bit before assigning ->rqs[].
217 */
218 if (rq && rq->q == hctx->queue && rq->mq_hctx == hctx)
219 return iter_data->fn(hctx, rq, iter_data->data, reserved);
220 return true;
221 }
222
223 /**
224 * bt_for_each - iterate over the requests associated with a hardware queue
225 * @hctx: Hardware queue to examine.
226 * @bt: sbitmap to examine. This is either the breserved_tags member
227 * or the bitmap_tags member of struct blk_mq_tags.
228 * @fn: Pointer to the function that will be called for each request
229 * associated with @hctx that has been assigned a driver tag.
230 * @fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
231 * where rq is a pointer to a request. Return true to continue
232 * iterating tags, false to stop.
233 * @data: Will be passed as third argument to @fn.
234 * @reserved: Indicates whether @bt is the breserved_tags member or the
235 * bitmap_tags member of struct blk_mq_tags.
236 */
237 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
238 busy_iter_fn *fn, void *data, bool reserved)
239 {
240 struct bt_iter_data iter_data = {
241 .hctx = hctx,
242 .fn = fn,
243 .data = data,
244 .reserved = reserved,
245 };
246
247 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
248 }
249
250 struct bt_tags_iter_data {
251 struct blk_mq_tags *tags;
252 busy_tag_iter_fn *fn;
253 void *data;
254 unsigned int flags;
255 };
256
257 #define BT_TAG_ITER_RESERVED (1 << 0)
258 #define BT_TAG_ITER_STARTED (1 << 1)
259 #define BT_TAG_ITER_STATIC_RQS (1 << 2)
260
261 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
262 {
263 struct bt_tags_iter_data *iter_data = data;
264 struct blk_mq_tags *tags = iter_data->tags;
265 bool reserved = iter_data->flags & BT_TAG_ITER_RESERVED;
266 struct request *rq;
267
268 if (!reserved)
269 bitnr += tags->nr_reserved_tags;
270
271 /*
272 * We can hit rq == NULL here, because the tagging functions
273 * test and set the bit before assigning ->rqs[].
274 */
275 if (iter_data->flags & BT_TAG_ITER_STATIC_RQS)
276 rq = tags->static_rqs[bitnr];
277 else
278 rq = tags->rqs[bitnr];
279 if (!rq)
280 return true;
281 if ((iter_data->flags & BT_TAG_ITER_STARTED) &&
282 !blk_mq_request_started(rq))
283 return true;
284 return iter_data->fn(rq, iter_data->data, reserved);
285 }
286
287 /**
288 * bt_tags_for_each - iterate over the requests in a tag map
289 * @tags: Tag map to iterate over.
290 * @bt: sbitmap to examine. This is either the breserved_tags member
291 * or the bitmap_tags member of struct blk_mq_tags.
292 * @fn: Pointer to the function that will be called for each started
293 * request. @fn will be called as follows: @fn(rq, @data,
294 * @reserved) where rq is a pointer to a request. Return true
295 * to continue iterating tags, false to stop.
296 * @data: Will be passed as second argument to @fn.
297 * @flags: BT_TAG_ITER_*
298 */
299 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
300 busy_tag_iter_fn *fn, void *data, unsigned int flags)
301 {
302 struct bt_tags_iter_data iter_data = {
303 .tags = tags,
304 .fn = fn,
305 .data = data,
306 .flags = flags,
307 };
308
309 if (tags->rqs)
310 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
311 }
312
313 static void __blk_mq_all_tag_iter(struct blk_mq_tags *tags,
314 busy_tag_iter_fn *fn, void *priv, unsigned int flags)
315 {
316 WARN_ON_ONCE(flags & BT_TAG_ITER_RESERVED);
317
318 if (tags->nr_reserved_tags)
319 bt_tags_for_each(tags, tags->breserved_tags, fn, priv,
320 flags | BT_TAG_ITER_RESERVED);
321 bt_tags_for_each(tags, tags->bitmap_tags, fn, priv, flags);
322 }
323
324 /**
325 * blk_mq_all_tag_iter - iterate over all requests in a tag map
326 * @tags: Tag map to iterate over.
327 * @fn: Pointer to the function that will be called for each
328 * request. @fn will be called as follows: @fn(rq, @priv,
329 * reserved) where rq is a pointer to a request. 'reserved'
330 * indicates whether or not @rq is a reserved request. Return
331 * true to continue iterating tags, false to stop.
332 * @priv: Will be passed as second argument to @fn.
333 *
334 * Caller has to pass the tag map from which requests are allocated.
335 */
336 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
337 void *priv)
338 {
339 __blk_mq_all_tag_iter(tags, fn, priv, BT_TAG_ITER_STATIC_RQS);
340 }
341
342 /**
343 * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
344 * @tagset: Tag set to iterate over.
345 * @fn: Pointer to the function that will be called for each started
346 * request. @fn will be called as follows: @fn(rq, @priv,
347 * reserved) where rq is a pointer to a request. 'reserved'
348 * indicates whether or not @rq is a reserved request. Return
349 * true to continue iterating tags, false to stop.
350 * @priv: Will be passed as second argument to @fn.
351 */
352 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
353 busy_tag_iter_fn *fn, void *priv)
354 {
355 int i;
356
357 for (i = 0; i < tagset->nr_hw_queues; i++) {
358 if (tagset->tags && tagset->tags[i])
359 __blk_mq_all_tag_iter(tagset->tags[i], fn, priv,
360 BT_TAG_ITER_STARTED);
361 }
362 }
363 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
364
365 static bool blk_mq_tagset_count_completed_rqs(struct request *rq,
366 void *data, bool reserved)
367 {
368 unsigned *count = data;
369
370 if (blk_mq_request_completed(rq))
371 (*count)++;
372 return true;
373 }
374
375 /**
376 * blk_mq_tagset_wait_completed_request - wait until all completed req's
377 * complete funtion is run
378 * @tagset: Tag set to drain completed request
379 *
380 * Note: This function has to be run after all IO queues are shutdown
381 */
382 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
383 {
384 while (true) {
385 unsigned count = 0;
386
387 blk_mq_tagset_busy_iter(tagset,
388 blk_mq_tagset_count_completed_rqs, &count);
389 if (!count)
390 break;
391 msleep(5);
392 }
393 }
394 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
395
396 /**
397 * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
398 * @q: Request queue to examine.
399 * @fn: Pointer to the function that will be called for each request
400 * on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
401 * reserved) where rq is a pointer to a request and hctx points
402 * to the hardware queue associated with the request. 'reserved'
403 * indicates whether or not @rq is a reserved request.
404 * @priv: Will be passed as third argument to @fn.
405 *
406 * Note: if @q->tag_set is shared with other request queues then @fn will be
407 * called for all requests on all queues that share that tag set and not only
408 * for requests associated with @q.
409 */
410 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
411 void *priv)
412 {
413 struct blk_mq_hw_ctx *hctx;
414 int i;
415
416 /*
417 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
418 * while the queue is frozen. So we can use q_usage_counter to avoid
419 * racing with it.
420 */
421 if (!percpu_ref_tryget(&q->q_usage_counter))
422 return;
423
424 queue_for_each_hw_ctx(q, hctx, i) {
425 struct blk_mq_tags *tags = hctx->tags;
426
427 /*
428 * If no software queues are currently mapped to this
429 * hardware queue, there's nothing to check
430 */
431 if (!blk_mq_hw_queue_mapped(hctx))
432 continue;
433
434 if (tags->nr_reserved_tags)
435 bt_for_each(hctx, tags->breserved_tags, fn, priv, true);
436 bt_for_each(hctx, tags->bitmap_tags, fn, priv, false);
437 }
438 blk_queue_exit(q);
439 }
440
441 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
442 bool round_robin, int node)
443 {
444 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
445 node);
446 }
447
448 static int blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
449 int node, int alloc_policy)
450 {
451 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
452 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
453
454 if (bt_alloc(&tags->__bitmap_tags, depth, round_robin, node))
455 return -ENOMEM;
456 if (bt_alloc(&tags->__breserved_tags, tags->nr_reserved_tags,
457 round_robin, node))
458 goto free_bitmap_tags;
459
460 tags->bitmap_tags = &tags->__bitmap_tags;
461 tags->breserved_tags = &tags->__breserved_tags;
462
463 return 0;
464 free_bitmap_tags:
465 sbitmap_queue_free(&tags->__bitmap_tags);
466 return -ENOMEM;
467 }
468
469 int blk_mq_init_shared_sbitmap(struct blk_mq_tag_set *set, unsigned int flags)
470 {
471 unsigned int depth = set->queue_depth - set->reserved_tags;
472 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags);
473 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
474 int i, node = set->numa_node;
475
476 if (bt_alloc(&set->__bitmap_tags, depth, round_robin, node))
477 return -ENOMEM;
478 if (bt_alloc(&set->__breserved_tags, set->reserved_tags,
479 round_robin, node))
480 goto free_bitmap_tags;
481
482 for (i = 0; i < set->nr_hw_queues; i++) {
483 struct blk_mq_tags *tags = set->tags[i];
484
485 tags->bitmap_tags = &set->__bitmap_tags;
486 tags->breserved_tags = &set->__breserved_tags;
487 }
488
489 return 0;
490 free_bitmap_tags:
491 sbitmap_queue_free(&set->__bitmap_tags);
492 return -ENOMEM;
493 }
494
495 void blk_mq_exit_shared_sbitmap(struct blk_mq_tag_set *set)
496 {
497 sbitmap_queue_free(&set->__bitmap_tags);
498 sbitmap_queue_free(&set->__breserved_tags);
499 }
500
501 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
502 unsigned int reserved_tags,
503 int node, unsigned int flags)
504 {
505 int alloc_policy = BLK_MQ_FLAG_TO_ALLOC_POLICY(flags);
506 struct blk_mq_tags *tags;
507
508 if (total_tags > BLK_MQ_TAG_MAX) {
509 pr_err("blk-mq: tag depth too large\n");
510 return NULL;
511 }
512
513 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
514 if (!tags)
515 return NULL;
516
517 tags->nr_tags = total_tags;
518 tags->nr_reserved_tags = reserved_tags;
519
520 if (flags & BLK_MQ_F_TAG_HCTX_SHARED)
521 return tags;
522
523 if (blk_mq_init_bitmap_tags(tags, node, alloc_policy) < 0) {
524 kfree(tags);
525 return NULL;
526 }
527 return tags;
528 }
529
530 void blk_mq_free_tags(struct blk_mq_tags *tags, unsigned int flags)
531 {
532 if (!(flags & BLK_MQ_F_TAG_HCTX_SHARED)) {
533 sbitmap_queue_free(tags->bitmap_tags);
534 sbitmap_queue_free(tags->breserved_tags);
535 }
536 kfree(tags);
537 }
538
539 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
540 struct blk_mq_tags **tagsptr, unsigned int tdepth,
541 bool can_grow)
542 {
543 struct blk_mq_tags *tags = *tagsptr;
544
545 if (tdepth <= tags->nr_reserved_tags)
546 return -EINVAL;
547
548 /*
549 * If we are allowed to grow beyond the original size, allocate
550 * a new set of tags before freeing the old one.
551 */
552 if (tdepth > tags->nr_tags) {
553 struct blk_mq_tag_set *set = hctx->queue->tag_set;
554 /* Only sched tags can grow, so clear HCTX_SHARED flag */
555 unsigned int flags = set->flags & ~BLK_MQ_F_TAG_HCTX_SHARED;
556 struct blk_mq_tags *new;
557 bool ret;
558
559 if (!can_grow)
560 return -EINVAL;
561
562 /*
563 * We need some sort of upper limit, set it high enough that
564 * no valid use cases should require more.
565 */
566 if (tdepth > 16 * BLKDEV_MAX_RQ)
567 return -EINVAL;
568
569 new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
570 tags->nr_reserved_tags, flags);
571 if (!new)
572 return -ENOMEM;
573 ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
574 if (ret) {
575 blk_mq_free_rq_map(new, flags);
576 return -ENOMEM;
577 }
578
579 blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
580 blk_mq_free_rq_map(*tagsptr, flags);
581 *tagsptr = new;
582 } else {
583 /*
584 * Don't need (or can't) update reserved tags here, they
585 * remain static and should never need resizing.
586 */
587 sbitmap_queue_resize(tags->bitmap_tags,
588 tdepth - tags->nr_reserved_tags);
589 }
590
591 return 0;
592 }
593
594 void blk_mq_tag_resize_shared_sbitmap(struct blk_mq_tag_set *set, unsigned int size)
595 {
596 sbitmap_queue_resize(&set->__bitmap_tags, size - set->reserved_tags);
597 }
598
599 /**
600 * blk_mq_unique_tag() - return a tag that is unique queue-wide
601 * @rq: request for which to compute a unique tag
602 *
603 * The tag field in struct request is unique per hardware queue but not over
604 * all hardware queues. Hence this function that returns a tag with the
605 * hardware context index in the upper bits and the per hardware queue tag in
606 * the lower bits.
607 *
608 * Note: When called for a request that is queued on a non-multiqueue request
609 * queue, the hardware context index is set to zero.
610 */
611 u32 blk_mq_unique_tag(struct request *rq)
612 {
613 return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
614 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
615 }
616 EXPORT_SYMBOL(blk_mq_unique_tag);