Linux 4.6-rc6
[cris-mirror.git] / block / blk-mq-tag.c
blobabdbb47405cb840c7005618cdece6b3165850ecc
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.
8 * Uses active queue tracking to support fairer distribution of tags
9 * between multiple submitters when a shared tag map is used.
11 * Copyright (C) 2013-2014 Jens Axboe
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/random.h>
17 #include <linux/blk-mq.h>
18 #include "blk.h"
19 #include "blk-mq.h"
20 #include "blk-mq-tag.h"
22 static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
24 int i;
26 for (i = 0; i < bt->map_nr; i++) {
27 struct blk_align_bitmap *bm = &bt->map[i];
28 int ret;
30 ret = find_first_zero_bit(&bm->word, bm->depth);
31 if (ret < bm->depth)
32 return true;
35 return false;
38 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
40 if (!tags)
41 return true;
43 return bt_has_free_tags(&tags->bitmap_tags);
46 static inline int bt_index_inc(int index)
48 return (index + 1) & (BT_WAIT_QUEUES - 1);
51 static inline void bt_index_atomic_inc(atomic_t *index)
53 int old = atomic_read(index);
54 int new = bt_index_inc(old);
55 atomic_cmpxchg(index, old, new);
59 * If a previously inactive queue goes active, bump the active user count.
61 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
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);
67 return true;
71 * Wakeup all potentially sleeping on tags
73 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
75 struct blk_mq_bitmap_tags *bt;
76 int i, wake_index;
79 * Make sure all changes prior to this are visible from other CPUs.
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];
87 if (waitqueue_active(&bs->wait))
88 wake_up(&bs->wait);
90 wake_index = bt_index_inc(wake_index);
93 if (include_reserve) {
94 bt = &tags->breserved_tags;
95 if (waitqueue_active(&bt->bs[0].wait))
96 wake_up(&bt->bs[0].wait);
101 * If a previously busy queue goes inactive, potential waiters could now
102 * be allowed to queue. Wake them up and check.
104 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
106 struct blk_mq_tags *tags = hctx->tags;
108 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
109 return;
111 atomic_dec(&tags->active_queues);
113 blk_mq_tag_wakeup_all(tags, false);
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.
120 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
121 struct blk_mq_bitmap_tags *bt)
123 unsigned int depth, users;
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;
131 * Don't try dividing an ant
133 if (bt->depth == 1)
134 return true;
136 users = atomic_read(&hctx->tags->active_queues);
137 if (!users)
138 return true;
141 * Allow at least some tags
143 depth = max((bt->depth + users - 1) / users, 4U);
144 return atomic_read(&hctx->nr_active) < depth;
147 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag,
148 bool nowrap)
150 int tag, org_last_tag = last_tag;
152 while (1) {
153 tag = find_next_zero_bit(&bm->word, bm->depth, last_tag);
154 if (unlikely(tag >= bm->depth)) {
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.
160 if (org_last_tag && last_tag && !nowrap) {
161 last_tag = org_last_tag = 0;
162 continue;
164 return -1;
167 if (!test_and_set_bit(tag, &bm->word))
168 break;
170 last_tag = tag + 1;
171 if (last_tag >= bm->depth - 1)
172 last_tag = 0;
175 return tag;
178 #define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
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.
191 static 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)
194 unsigned int last_tag, org_last_tag;
195 int index, i, tag;
197 if (!hctx_may_queue(hctx, bt))
198 return -1;
200 last_tag = org_last_tag = *tag_cache;
201 index = TAG_TO_INDEX(bt, last_tag);
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;
212 * Jump to next index, and reset the last tag to be the
213 * first tag of that index
215 index++;
216 last_tag = (index << bt->bits_per_word);
218 if (index >= bt->map_nr) {
219 index = 0;
220 last_tag = 0;
224 *tag_cache = 0;
225 return -1;
228 * Only update the cache from the allocation path, if we ended
229 * up using the specific cached tag.
231 done:
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;
237 *tag_cache = last_tag;
240 return tag;
243 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
244 struct blk_mq_hw_ctx *hctx)
246 struct bt_wait_state *bs;
247 int wait_index;
249 if (!hctx)
250 return &bt->bs[0];
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;
258 static 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)
263 struct bt_wait_state *bs;
264 DEFINE_WAIT(wait);
265 int tag;
267 tag = __bt_get(hctx, bt, last_tag, tags);
268 if (tag != -1)
269 return tag;
271 if (data->flags & BLK_MQ_REQ_NOWAIT)
272 return -1;
274 bs = bt_wait_ptr(bt, hctx);
275 do {
276 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
278 tag = __bt_get(hctx, bt, last_tag, tags);
279 if (tag != -1)
280 break;
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.
288 if (hctx)
289 blk_mq_run_hw_queue(hctx, false);
292 * Retry tag allocation after running the hardware queue,
293 * as running the queue may also have found completions.
295 tag = __bt_get(hctx, bt, last_tag, tags);
296 if (tag != -1)
297 break;
299 blk_mq_put_ctx(data->ctx);
301 io_schedule();
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;
313 finish_wait(&bs->wait, &wait);
314 bs = bt_wait_ptr(bt, hctx);
315 } while (1);
317 finish_wait(&bs->wait, &wait);
318 return tag;
321 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
323 int tag;
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;
330 return BLK_MQ_TAG_FAIL;
333 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
335 int tag, zero = 0;
337 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
338 WARN_ON_ONCE(1);
339 return BLK_MQ_TAG_FAIL;
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;
347 return tag;
350 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
352 if (data->flags & BLK_MQ_REQ_RESERVED)
353 return __blk_mq_get_reserved_tag(data);
354 return __blk_mq_get_tag(data);
357 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
359 int i, wake_index;
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];
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);
370 return bs;
373 wake_index = bt_index_inc(wake_index);
376 return NULL;
379 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
381 const int index = TAG_TO_INDEX(bt, tag);
382 struct bt_wait_state *bs;
383 int wait_cnt;
385 clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
387 /* Ensure that the wait list checks occur after clear_bit(). */
388 smp_mb();
390 bs = bt_wake_ptr(bt);
391 if (!bs)
392 return;
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);
404 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
405 unsigned int *last_tag)
407 struct blk_mq_tags *tags = hctx->tags;
409 if (tag >= tags->nr_reserved_tags) {
410 const int real_tag = tag - tags->nr_reserved_tags;
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);
422 static 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)
426 struct request *rq;
427 int bit, i;
429 for (i = 0; i < bt->map_nr; i++) {
430 struct blk_align_bitmap *bm = &bt->map[i];
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);
440 off += (1 << bt->bits_per_word);
444 static 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)
448 struct request *rq;
449 int bit, i;
451 if (!tags->rqs)
452 return;
453 for (i = 0; i < bt->map_nr; i++) {
454 struct blk_align_bitmap *bm = &bt->map[i];
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);
463 off += (1 << bt->bits_per_word);
467 void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
468 void *priv)
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);
475 EXPORT_SYMBOL(blk_mq_all_tag_busy_iter);
477 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
478 void *priv)
480 struct blk_mq_hw_ctx *hctx;
481 int i;
484 queue_for_each_hw_ctx(q, hctx, i) {
485 struct blk_mq_tags *tags = hctx->tags;
488 * If not software queues are currently mapped to this
489 * hardware queue, there's nothing to check
491 if (!blk_mq_hw_queue_mapped(hctx))
492 continue;
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);
502 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
504 unsigned int i, used;
506 for (i = 0, used = 0; i < bt->map_nr; i++) {
507 struct blk_align_bitmap *bm = &bt->map[i];
509 used += bitmap_weight(&bm->word, bm->depth);
512 return bt->depth - used;
515 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
516 unsigned int depth)
518 unsigned int tags_per_word = 1U << bt->bits_per_word;
519 unsigned int map_depth = depth;
521 if (depth) {
522 int i;
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;
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);
534 bt->depth = depth;
537 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
538 int node, bool reserved)
540 int i;
542 bt->bits_per_word = ilog2(BITS_PER_LONG);
545 * Depth can be zero for reserved tags, that's not a failure
546 * condition.
548 if (depth) {
549 unsigned int nr, tags_per_word;
551 tags_per_word = (1 << bt->bits_per_word);
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.
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);
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;
572 bt->map_nr = nr;
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;
582 bt_update_count(bt, depth);
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);
589 return 0;
592 static void bt_free(struct blk_mq_bitmap_tags *bt)
594 kfree(bt->map);
595 kfree(bt->bs);
598 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
599 int node, int alloc_policy)
601 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
603 tags->alloc_policy = alloc_policy;
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;
610 return tags;
611 enomem:
612 bt_free(&tags->bitmap_tags);
613 kfree(tags);
614 return NULL;
617 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
618 unsigned int reserved_tags,
619 int node, int alloc_policy)
621 struct blk_mq_tags *tags;
623 if (total_tags > BLK_MQ_TAG_MAX) {
624 pr_err("blk-mq: tag depth too large\n");
625 return NULL;
628 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
629 if (!tags)
630 return NULL;
632 if (!zalloc_cpumask_var(&tags->cpumask, GFP_KERNEL)) {
633 kfree(tags);
634 return NULL;
637 tags->nr_tags = total_tags;
638 tags->nr_reserved_tags = reserved_tags;
640 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
643 void blk_mq_free_tags(struct blk_mq_tags *tags)
645 bt_free(&tags->bitmap_tags);
646 bt_free(&tags->breserved_tags);
647 free_cpumask_var(tags->cpumask);
648 kfree(tags);
651 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
653 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
655 *tag = prandom_u32() % depth;
658 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
660 tdepth -= tags->nr_reserved_tags;
661 if (tdepth > tags->nr_tags)
662 return -EINVAL;
665 * Don't need (or can't) update reserved tags here, they remain
666 * static and should never need resizing.
668 bt_update_count(&tags->bitmap_tags, tdepth);
669 blk_mq_tag_wakeup_all(tags, false);
670 return 0;
674 * blk_mq_unique_tag() - return a tag that is unique queue-wide
675 * @rq: request for which to compute a unique tag
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.
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.
685 u32 blk_mq_unique_tag(struct request *rq)
687 struct request_queue *q = rq->q;
688 struct blk_mq_hw_ctx *hctx;
689 int hwq = 0;
691 if (q->mq_ops) {
692 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
693 hwq = hctx->queue_num;
696 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
697 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
699 EXPORT_SYMBOL(blk_mq_unique_tag);
701 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
703 char *orig_page = page;
704 unsigned int free, res;
706 if (!tags)
707 return 0;
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);
714 free = bt_unused_tags(&tags->bitmap_tags);
715 res = bt_unused_tags(&tags->breserved_tags);
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));
720 return page - orig_page;