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 normal (non-reserved) tags
73 static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
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);
91 * If a previously busy queue goes inactive, potential waiters could now
92 * be allowed to queue. Wake them up and check.
94 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
96 struct blk_mq_tags *tags = hctx->tags;
98 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
101 atomic_dec(&tags->active_queues);
103 blk_mq_tag_wakeup_all(tags);
107 * For shared tag users, we track the number of currently active users
108 * and attempt to provide a fair share of the tag depth for each of them.
110 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
111 struct blk_mq_bitmap_tags *bt)
113 unsigned int depth, users;
115 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
117 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
121 * Don't try dividing an ant
126 users = atomic_read(&hctx->tags->active_queues);
131 * Allow at least some tags
133 depth = max((bt->depth + users - 1) / users, 4U);
134 return atomic_read(&hctx->nr_active) < depth;
137 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
139 int tag, org_last_tag, end;
140 bool wrap = last_tag != 0;
142 org_last_tag = last_tag;
146 tag = find_next_zero_bit(&bm->word, end, last_tag);
147 if (unlikely(tag >= end)) {
149 * We started with an offset, start from 0 to
161 } while (test_and_set_bit(tag, &bm->word));
167 * Straight forward bitmap tag implementation, where each bit is a tag
168 * (cleared == free, and set == busy). The small twist is using per-cpu
169 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
170 * contexts. This enables us to drastically limit the space searched,
171 * without dirtying an extra shared cacheline like we would if we stored
172 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
173 * of that, each word of tags is in a separate cacheline. This means that
174 * multiple users will tend to stick to different cachelines, at least
175 * until the map is exhausted.
177 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
178 unsigned int *tag_cache)
180 unsigned int last_tag, org_last_tag;
183 if (!hctx_may_queue(hctx, bt))
186 last_tag = org_last_tag = *tag_cache;
187 index = TAG_TO_INDEX(bt, last_tag);
189 for (i = 0; i < bt->map_nr; i++) {
190 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
192 tag += (index << bt->bits_per_word);
197 if (++index >= bt->map_nr)
205 * Only update the cache from the allocation path, if we ended
206 * up using the specific cached tag.
209 if (tag == org_last_tag) {
211 if (last_tag >= bt->depth - 1)
214 *tag_cache = last_tag;
220 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
221 struct blk_mq_hw_ctx *hctx)
223 struct bt_wait_state *bs;
229 wait_index = atomic_read(&hctx->wait_index);
230 bs = &bt->bs[wait_index];
231 bt_index_atomic_inc(&hctx->wait_index);
235 static int bt_get(struct blk_mq_alloc_data *data,
236 struct blk_mq_bitmap_tags *bt,
237 struct blk_mq_hw_ctx *hctx,
238 unsigned int *last_tag)
240 struct bt_wait_state *bs;
244 tag = __bt_get(hctx, bt, last_tag);
248 if (!(data->gfp & __GFP_WAIT))
251 bs = bt_wait_ptr(bt, hctx);
253 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
255 tag = __bt_get(hctx, bt, last_tag);
260 * We're out of tags on this hardware queue, kick any
261 * pending IO submits before going to sleep waiting for
264 blk_mq_run_hw_queue(hctx, false);
267 * Retry tag allocation after running the hardware queue,
268 * as running the queue may also have found completions.
270 tag = __bt_get(hctx, bt, last_tag);
274 blk_mq_put_ctx(data->ctx);
278 data->ctx = blk_mq_get_ctx(data->q);
279 data->hctx = data->q->mq_ops->map_queue(data->q,
281 if (data->reserved) {
282 bt = &data->hctx->tags->breserved_tags;
284 last_tag = &data->ctx->last_tag;
286 bt = &hctx->tags->bitmap_tags;
288 finish_wait(&bs->wait, &wait);
289 bs = bt_wait_ptr(bt, hctx);
292 finish_wait(&bs->wait, &wait);
296 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
300 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
301 &data->ctx->last_tag);
303 return tag + data->hctx->tags->nr_reserved_tags;
305 return BLK_MQ_TAG_FAIL;
308 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
312 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
314 return BLK_MQ_TAG_FAIL;
317 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
319 return BLK_MQ_TAG_FAIL;
324 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
327 return __blk_mq_get_tag(data);
329 return __blk_mq_get_reserved_tag(data);
332 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
336 wake_index = atomic_read(&bt->wake_index);
337 for (i = 0; i < BT_WAIT_QUEUES; i++) {
338 struct bt_wait_state *bs = &bt->bs[wake_index];
340 if (waitqueue_active(&bs->wait)) {
341 int o = atomic_read(&bt->wake_index);
343 atomic_cmpxchg(&bt->wake_index, o, wake_index);
348 wake_index = bt_index_inc(wake_index);
354 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
356 const int index = TAG_TO_INDEX(bt, tag);
357 struct bt_wait_state *bs;
360 clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
362 /* Ensure that the wait list checks occur after clear_bit(). */
365 bs = bt_wake_ptr(bt);
369 wait_cnt = atomic_dec_return(&bs->wait_cnt);
370 if (unlikely(wait_cnt < 0))
371 wait_cnt = atomic_inc_return(&bs->wait_cnt);
373 atomic_add(bt->wake_cnt, &bs->wait_cnt);
374 bt_index_atomic_inc(&bt->wake_index);
379 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
380 unsigned int *last_tag)
382 struct blk_mq_tags *tags = hctx->tags;
384 if (tag >= tags->nr_reserved_tags) {
385 const int real_tag = tag - tags->nr_reserved_tags;
387 BUG_ON(real_tag >= tags->nr_tags);
388 bt_clear_tag(&tags->bitmap_tags, real_tag);
389 *last_tag = real_tag;
391 BUG_ON(tag >= tags->nr_reserved_tags);
392 bt_clear_tag(&tags->breserved_tags, tag);
396 static void bt_for_each(struct blk_mq_hw_ctx *hctx,
397 struct blk_mq_bitmap_tags *bt, unsigned int off,
398 busy_iter_fn *fn, void *data, bool reserved)
403 for (i = 0; i < bt->map_nr; i++) {
404 struct blk_align_bitmap *bm = &bt->map[i];
406 for (bit = find_first_bit(&bm->word, bm->depth);
408 bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
409 rq = blk_mq_tag_to_rq(hctx->tags, off + bit);
410 if (rq->q == hctx->queue)
411 fn(hctx, rq, data, reserved);
414 off += (1 << bt->bits_per_word);
418 void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn,
421 struct blk_mq_tags *tags = hctx->tags;
423 if (tags->nr_reserved_tags)
424 bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
425 bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
428 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
430 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
432 unsigned int i, used;
434 for (i = 0, used = 0; i < bt->map_nr; i++) {
435 struct blk_align_bitmap *bm = &bt->map[i];
437 used += bitmap_weight(&bm->word, bm->depth);
440 return bt->depth - used;
443 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
446 unsigned int tags_per_word = 1U << bt->bits_per_word;
447 unsigned int map_depth = depth;
452 for (i = 0; i < bt->map_nr; i++) {
453 bt->map[i].depth = min(map_depth, tags_per_word);
454 map_depth -= bt->map[i].depth;
458 bt->wake_cnt = BT_WAIT_BATCH;
459 if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
460 bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
465 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
466 int node, bool reserved)
470 bt->bits_per_word = ilog2(BITS_PER_LONG);
473 * Depth can be zero for reserved tags, that's not a failure
477 unsigned int nr, tags_per_word;
479 tags_per_word = (1 << bt->bits_per_word);
482 * If the tag space is small, shrink the number of tags
483 * per word so we spread over a few cachelines, at least.
484 * If less than 4 tags, just forget about it, it's not
485 * going to work optimally anyway.
488 while (tags_per_word * 4 > depth) {
490 tags_per_word = (1 << bt->bits_per_word);
494 nr = ALIGN(depth, tags_per_word) / tags_per_word;
495 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
503 bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
509 bt_update_count(bt, depth);
511 for (i = 0; i < BT_WAIT_QUEUES; i++) {
512 init_waitqueue_head(&bt->bs[i].wait);
513 atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
519 static void bt_free(struct blk_mq_bitmap_tags *bt)
525 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
528 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
530 if (bt_alloc(&tags->bitmap_tags, depth, node, false))
532 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
537 bt_free(&tags->bitmap_tags);
542 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
543 unsigned int reserved_tags, int node)
545 struct blk_mq_tags *tags;
547 if (total_tags > BLK_MQ_TAG_MAX) {
548 pr_err("blk-mq: tag depth too large\n");
552 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
556 tags->nr_tags = total_tags;
557 tags->nr_reserved_tags = reserved_tags;
559 return blk_mq_init_bitmap_tags(tags, node);
562 void blk_mq_free_tags(struct blk_mq_tags *tags)
564 bt_free(&tags->bitmap_tags);
565 bt_free(&tags->breserved_tags);
569 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
571 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
573 *tag = prandom_u32() % depth;
576 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
578 tdepth -= tags->nr_reserved_tags;
579 if (tdepth > tags->nr_tags)
583 * Don't need (or can't) update reserved tags here, they remain
584 * static and should never need resizing.
586 bt_update_count(&tags->bitmap_tags, tdepth);
587 blk_mq_tag_wakeup_all(tags);
592 * blk_mq_unique_tag() - return a tag that is unique queue-wide
593 * @rq: request for which to compute a unique tag
595 * The tag field in struct request is unique per hardware queue but not over
596 * all hardware queues. Hence this function that returns a tag with the
597 * hardware context index in the upper bits and the per hardware queue tag in
600 * Note: When called for a request that is queued on a non-multiqueue request
601 * queue, the hardware context index is set to zero.
603 u32 blk_mq_unique_tag(struct request *rq)
605 struct request_queue *q = rq->q;
606 struct blk_mq_hw_ctx *hctx;
610 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
611 hwq = hctx->queue_num;
614 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
615 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
617 EXPORT_SYMBOL(blk_mq_unique_tag);
619 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
621 char *orig_page = page;
622 unsigned int free, res;
627 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
628 "bits_per_word=%u\n",
629 tags->nr_tags, tags->nr_reserved_tags,
630 tags->bitmap_tags.bits_per_word);
632 free = bt_unused_tags(&tags->bitmap_tags);
633 res = bt_unused_tags(&tags->breserved_tags);
635 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
636 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
638 return page - orig_page;