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
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>
20 #include "blk-mq-tag.h"
22 static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
26 for (i = 0; i < bt->map_nr; i++) {
27 struct blk_align_bitmap *bm = &bt->map[i];
30 ret = find_first_zero_bit(&bm->word, bm->depth);
38 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
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);
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;
78 bt = &tags->bitmap_tags;
79 wake_index = atomic_read(&bt->wake_index);
80 for (i = 0; i < BT_WAIT_QUEUES; i++) {
81 struct bt_wait_state *bs = &bt->bs[wake_index];
83 if (waitqueue_active(&bs->wait))
86 wake_index = bt_index_inc(wake_index);
89 if (include_reserve) {
90 bt = &tags->breserved_tags;
91 if (waitqueue_active(&bt->bs[0].wait))
92 wake_up(&bt->bs[0].wait);
97 * If a previously busy queue goes inactive, potential waiters could now
98 * be allowed to queue. Wake them up and check.
100 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
102 struct blk_mq_tags *tags = hctx->tags;
104 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
107 atomic_dec(&tags->active_queues);
109 blk_mq_tag_wakeup_all(tags, false);
113 * For shared tag users, we track the number of currently active users
114 * and attempt to provide a fair share of the tag depth for each of them.
116 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
117 struct blk_mq_bitmap_tags *bt)
119 unsigned int depth, users;
121 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
123 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
127 * Don't try dividing an ant
132 users = atomic_read(&hctx->tags->active_queues);
137 * Allow at least some tags
139 depth = max((bt->depth + users - 1) / users, 4U);
140 return atomic_read(&hctx->nr_active) < depth;
143 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
145 int tag, org_last_tag, end;
146 bool wrap = last_tag != 0;
148 org_last_tag = last_tag;
152 tag = find_next_zero_bit(&bm->word, end, last_tag);
153 if (unlikely(tag >= end)) {
155 * We started with an offset, start from 0 to
167 } while (test_and_set_bit(tag, &bm->word));
173 * Straight forward bitmap tag implementation, where each bit is a tag
174 * (cleared == free, and set == busy). The small twist is using per-cpu
175 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
176 * contexts. This enables us to drastically limit the space searched,
177 * without dirtying an extra shared cacheline like we would if we stored
178 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
179 * of that, each word of tags is in a separate cacheline. This means that
180 * multiple users will tend to stick to different cachelines, at least
181 * until the map is exhausted.
183 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
184 unsigned int *tag_cache)
186 unsigned int last_tag, org_last_tag;
189 if (!hctx_may_queue(hctx, bt))
192 last_tag = org_last_tag = *tag_cache;
193 index = TAG_TO_INDEX(bt, last_tag);
195 for (i = 0; i < bt->map_nr; i++) {
196 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
198 tag += (index << bt->bits_per_word);
203 if (++index >= bt->map_nr)
211 * Only update the cache from the allocation path, if we ended
212 * up using the specific cached tag.
215 if (tag == org_last_tag) {
217 if (last_tag >= bt->depth - 1)
220 *tag_cache = last_tag;
226 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
227 struct blk_mq_hw_ctx *hctx)
229 struct bt_wait_state *bs;
235 wait_index = atomic_read(&hctx->wait_index);
236 bs = &bt->bs[wait_index];
237 bt_index_atomic_inc(&hctx->wait_index);
241 static int bt_get(struct blk_mq_alloc_data *data,
242 struct blk_mq_bitmap_tags *bt,
243 struct blk_mq_hw_ctx *hctx,
244 unsigned int *last_tag)
246 struct bt_wait_state *bs;
250 tag = __bt_get(hctx, bt, last_tag);
254 if (!(data->gfp & __GFP_WAIT))
257 bs = bt_wait_ptr(bt, hctx);
259 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
261 tag = __bt_get(hctx, bt, last_tag);
266 * We're out of tags on this hardware queue, kick any
267 * pending IO submits before going to sleep waiting for
270 blk_mq_run_hw_queue(hctx, false);
273 * Retry tag allocation after running the hardware queue,
274 * as running the queue may also have found completions.
276 tag = __bt_get(hctx, bt, last_tag);
280 blk_mq_put_ctx(data->ctx);
284 data->ctx = blk_mq_get_ctx(data->q);
285 data->hctx = data->q->mq_ops->map_queue(data->q,
287 if (data->reserved) {
288 bt = &data->hctx->tags->breserved_tags;
290 last_tag = &data->ctx->last_tag;
292 bt = &hctx->tags->bitmap_tags;
294 finish_wait(&bs->wait, &wait);
295 bs = bt_wait_ptr(bt, hctx);
298 finish_wait(&bs->wait, &wait);
302 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
306 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
307 &data->ctx->last_tag);
309 return tag + data->hctx->tags->nr_reserved_tags;
311 return BLK_MQ_TAG_FAIL;
314 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
318 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
320 return BLK_MQ_TAG_FAIL;
323 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
325 return BLK_MQ_TAG_FAIL;
330 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
333 return __blk_mq_get_tag(data);
335 return __blk_mq_get_reserved_tag(data);
338 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
342 wake_index = atomic_read(&bt->wake_index);
343 for (i = 0; i < BT_WAIT_QUEUES; i++) {
344 struct bt_wait_state *bs = &bt->bs[wake_index];
346 if (waitqueue_active(&bs->wait)) {
347 int o = atomic_read(&bt->wake_index);
349 atomic_cmpxchg(&bt->wake_index, o, wake_index);
354 wake_index = bt_index_inc(wake_index);
360 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
362 const int index = TAG_TO_INDEX(bt, tag);
363 struct bt_wait_state *bs;
366 clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
368 /* Ensure that the wait list checks occur after clear_bit(). */
371 bs = bt_wake_ptr(bt);
375 wait_cnt = atomic_dec_return(&bs->wait_cnt);
376 if (unlikely(wait_cnt < 0))
377 wait_cnt = atomic_inc_return(&bs->wait_cnt);
379 atomic_add(bt->wake_cnt, &bs->wait_cnt);
380 bt_index_atomic_inc(&bt->wake_index);
385 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
386 unsigned int *last_tag)
388 struct blk_mq_tags *tags = hctx->tags;
390 if (tag >= tags->nr_reserved_tags) {
391 const int real_tag = tag - tags->nr_reserved_tags;
393 BUG_ON(real_tag >= tags->nr_tags);
394 bt_clear_tag(&tags->bitmap_tags, real_tag);
395 *last_tag = real_tag;
397 BUG_ON(tag >= tags->nr_reserved_tags);
398 bt_clear_tag(&tags->breserved_tags, tag);
402 static void bt_for_each(struct blk_mq_hw_ctx *hctx,
403 struct blk_mq_bitmap_tags *bt, unsigned int off,
404 busy_iter_fn *fn, void *data, bool reserved)
409 for (i = 0; i < bt->map_nr; i++) {
410 struct blk_align_bitmap *bm = &bt->map[i];
412 for (bit = find_first_bit(&bm->word, bm->depth);
414 bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
415 rq = blk_mq_tag_to_rq(hctx->tags, off + bit);
416 if (rq->q == hctx->queue)
417 fn(hctx, rq, data, reserved);
420 off += (1 << bt->bits_per_word);
424 void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn,
427 struct blk_mq_tags *tags = hctx->tags;
429 if (tags->nr_reserved_tags)
430 bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
431 bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
434 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
436 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
438 unsigned int i, used;
440 for (i = 0, used = 0; i < bt->map_nr; i++) {
441 struct blk_align_bitmap *bm = &bt->map[i];
443 used += bitmap_weight(&bm->word, bm->depth);
446 return bt->depth - used;
449 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
452 unsigned int tags_per_word = 1U << bt->bits_per_word;
453 unsigned int map_depth = depth;
458 for (i = 0; i < bt->map_nr; i++) {
459 bt->map[i].depth = min(map_depth, tags_per_word);
460 map_depth -= bt->map[i].depth;
464 bt->wake_cnt = BT_WAIT_BATCH;
465 if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
466 bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
471 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
472 int node, bool reserved)
476 bt->bits_per_word = ilog2(BITS_PER_LONG);
479 * Depth can be zero for reserved tags, that's not a failure
483 unsigned int nr, tags_per_word;
485 tags_per_word = (1 << bt->bits_per_word);
488 * If the tag space is small, shrink the number of tags
489 * per word so we spread over a few cachelines, at least.
490 * If less than 4 tags, just forget about it, it's not
491 * going to work optimally anyway.
494 while (tags_per_word * 4 > depth) {
496 tags_per_word = (1 << bt->bits_per_word);
500 nr = ALIGN(depth, tags_per_word) / tags_per_word;
501 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
509 bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
515 bt_update_count(bt, depth);
517 for (i = 0; i < BT_WAIT_QUEUES; i++) {
518 init_waitqueue_head(&bt->bs[i].wait);
519 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
525 static void bt_free(struct blk_mq_bitmap_tags *bt)
531 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
534 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
536 if (bt_alloc(&tags->bitmap_tags, depth, node, false))
538 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
543 bt_free(&tags->bitmap_tags);
548 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
549 unsigned int reserved_tags, int node)
551 struct blk_mq_tags *tags;
553 if (total_tags > BLK_MQ_TAG_MAX) {
554 pr_err("blk-mq: tag depth too large\n");
558 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
562 tags->nr_tags = total_tags;
563 tags->nr_reserved_tags = reserved_tags;
565 return blk_mq_init_bitmap_tags(tags, node);
568 void blk_mq_free_tags(struct blk_mq_tags *tags)
570 bt_free(&tags->bitmap_tags);
571 bt_free(&tags->breserved_tags);
575 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
577 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
579 *tag = prandom_u32() % depth;
582 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
584 tdepth -= tags->nr_reserved_tags;
585 if (tdepth > tags->nr_tags)
589 * Don't need (or can't) update reserved tags here, they remain
590 * static and should never need resizing.
592 bt_update_count(&tags->bitmap_tags, tdepth);
593 blk_mq_tag_wakeup_all(tags, false);
598 * blk_mq_unique_tag() - return a tag that is unique queue-wide
599 * @rq: request for which to compute a unique tag
601 * The tag field in struct request is unique per hardware queue but not over
602 * all hardware queues. Hence this function that returns a tag with the
603 * hardware context index in the upper bits and the per hardware queue tag in
606 * Note: When called for a request that is queued on a non-multiqueue request
607 * queue, the hardware context index is set to zero.
609 u32 blk_mq_unique_tag(struct request *rq)
611 struct request_queue *q = rq->q;
612 struct blk_mq_hw_ctx *hctx;
616 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
617 hwq = hctx->queue_num;
620 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
621 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
623 EXPORT_SYMBOL(blk_mq_unique_tag);
625 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
627 char *orig_page = page;
628 unsigned int free, res;
633 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
634 "bits_per_word=%u\n",
635 tags->nr_tags, tags->nr_reserved_tags,
636 tags->bitmap_tags.bits_per_word);
638 free = bt_unused_tags(&tags->bitmap_tags);
639 res = bt_unused_tags(&tags->breserved_tags);
641 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
642 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
644 return page - orig_page;