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 void bt_index_inc(unsigned int *index)
48 *index = (*index + 1) & (BT_WAIT_QUEUES - 1);
52 * If a previously inactive queue goes active, bump the active user count.
54 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
56 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
57 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
58 atomic_inc(&hctx->tags->active_queues);
64 * Wakeup all potentially sleeping on normal (non-reserved) tags
66 static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
68 struct blk_mq_bitmap_tags *bt;
71 bt = &tags->bitmap_tags;
72 wake_index = bt->wake_index;
73 for (i = 0; i < BT_WAIT_QUEUES; i++) {
74 struct bt_wait_state *bs = &bt->bs[wake_index];
76 if (waitqueue_active(&bs->wait))
79 bt_index_inc(&wake_index);
84 * If a previously busy queue goes inactive, potential waiters could now
85 * be allowed to queue. Wake them up and check.
87 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
89 struct blk_mq_tags *tags = hctx->tags;
91 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
94 atomic_dec(&tags->active_queues);
96 blk_mq_tag_wakeup_all(tags);
100 * For shared tag users, we track the number of currently active users
101 * and attempt to provide a fair share of the tag depth for each of them.
103 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
104 struct blk_mq_bitmap_tags *bt)
106 unsigned int depth, users;
108 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
110 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
114 * Don't try dividing an ant
119 users = atomic_read(&hctx->tags->active_queues);
124 * Allow at least some tags
126 depth = max((bt->depth + users - 1) / users, 4U);
127 return atomic_read(&hctx->nr_active) < depth;
130 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
132 int tag, org_last_tag, end;
134 org_last_tag = last_tag;
138 tag = find_next_zero_bit(&bm->word, end, last_tag);
139 if (unlikely(tag >= end)) {
141 * We started with an offset, start from 0 to
144 if (org_last_tag && last_tag) {
152 } while (test_and_set_bit_lock(tag, &bm->word));
158 * Straight forward bitmap tag implementation, where each bit is a tag
159 * (cleared == free, and set == busy). The small twist is using per-cpu
160 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
161 * contexts. This enables us to drastically limit the space searched,
162 * without dirtying an extra shared cacheline like we would if we stored
163 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
164 * of that, each word of tags is in a separate cacheline. This means that
165 * multiple users will tend to stick to different cachelines, at least
166 * until the map is exhausted.
168 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
169 unsigned int *tag_cache)
171 unsigned int last_tag, org_last_tag;
174 if (!hctx_may_queue(hctx, bt))
177 last_tag = org_last_tag = *tag_cache;
178 index = TAG_TO_INDEX(bt, last_tag);
180 for (i = 0; i < bt->map_nr; i++) {
181 tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
183 tag += (index << bt->bits_per_word);
188 if (++index >= bt->map_nr)
196 * Only update the cache from the allocation path, if we ended
197 * up using the specific cached tag.
200 if (tag == org_last_tag) {
202 if (last_tag >= bt->depth - 1)
205 *tag_cache = last_tag;
211 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
212 struct blk_mq_hw_ctx *hctx)
214 struct bt_wait_state *bs;
219 bs = &bt->bs[hctx->wait_index];
220 bt_index_inc(&hctx->wait_index);
224 static int bt_get(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
225 unsigned int *last_tag, gfp_t gfp)
227 struct bt_wait_state *bs;
231 tag = __bt_get(hctx, bt, last_tag);
235 if (!(gfp & __GFP_WAIT))
238 bs = bt_wait_ptr(bt, hctx);
242 was_empty = list_empty(&wait.task_list);
243 prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
245 tag = __bt_get(hctx, bt, last_tag);
250 atomic_set(&bs->wait_cnt, bt->wake_cnt);
255 finish_wait(&bs->wait, &wait);
259 static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags,
260 struct blk_mq_hw_ctx *hctx,
261 unsigned int *last_tag, gfp_t gfp)
265 tag = bt_get(&tags->bitmap_tags, hctx, last_tag, gfp);
267 return tag + tags->nr_reserved_tags;
269 return BLK_MQ_TAG_FAIL;
272 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
277 if (unlikely(!tags->nr_reserved_tags)) {
279 return BLK_MQ_TAG_FAIL;
282 tag = bt_get(&tags->breserved_tags, NULL, &zero, gfp);
284 return BLK_MQ_TAG_FAIL;
289 unsigned int blk_mq_get_tag(struct blk_mq_hw_ctx *hctx, unsigned int *last_tag,
290 gfp_t gfp, bool reserved)
293 return __blk_mq_get_tag(hctx->tags, hctx, last_tag, gfp);
295 return __blk_mq_get_reserved_tag(hctx->tags, gfp);
298 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
302 wake_index = bt->wake_index;
303 for (i = 0; i < BT_WAIT_QUEUES; i++) {
304 struct bt_wait_state *bs = &bt->bs[wake_index];
306 if (waitqueue_active(&bs->wait)) {
307 if (wake_index != bt->wake_index)
308 bt->wake_index = wake_index;
313 bt_index_inc(&wake_index);
319 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
321 const int index = TAG_TO_INDEX(bt, tag);
322 struct bt_wait_state *bs;
325 * The unlock memory barrier need to order access to req in free
326 * path and clearing tag bit
328 clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
330 bs = bt_wake_ptr(bt);
331 if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
332 atomic_set(&bs->wait_cnt, bt->wake_cnt);
333 bt_index_inc(&bt->wake_index);
338 static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
340 BUG_ON(tag >= tags->nr_tags);
342 bt_clear_tag(&tags->bitmap_tags, tag);
345 static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
348 BUG_ON(tag >= tags->nr_reserved_tags);
350 bt_clear_tag(&tags->breserved_tags, tag);
353 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
354 unsigned int *last_tag)
356 struct blk_mq_tags *tags = hctx->tags;
358 if (tag >= tags->nr_reserved_tags) {
359 const int real_tag = tag - tags->nr_reserved_tags;
361 __blk_mq_put_tag(tags, real_tag);
362 *last_tag = real_tag;
364 __blk_mq_put_reserved_tag(tags, tag);
367 static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
368 unsigned long *free_map, unsigned int off)
372 for (i = 0; i < bt->map_nr; i++) {
373 struct blk_align_bitmap *bm = &bt->map[i];
377 bit = find_next_zero_bit(&bm->word, bm->depth, bit);
378 if (bit >= bm->depth)
381 __set_bit(bit + off, free_map);
385 off += (1 << bt->bits_per_word);
389 void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
390 void (*fn)(void *, unsigned long *), void *data)
392 unsigned long *tag_map;
395 map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
396 tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
400 bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
401 if (tags->nr_reserved_tags)
402 bt_for_each_free(&tags->breserved_tags, tag_map, 0);
407 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
409 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
411 unsigned int i, used;
413 for (i = 0, used = 0; i < bt->map_nr; i++) {
414 struct blk_align_bitmap *bm = &bt->map[i];
416 used += bitmap_weight(&bm->word, bm->depth);
419 return bt->depth - used;
422 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
425 unsigned int tags_per_word = 1U << bt->bits_per_word;
426 unsigned int map_depth = depth;
431 for (i = 0; i < bt->map_nr; i++) {
432 bt->map[i].depth = min(map_depth, tags_per_word);
433 map_depth -= bt->map[i].depth;
437 bt->wake_cnt = BT_WAIT_BATCH;
438 if (bt->wake_cnt > depth / 4)
439 bt->wake_cnt = max(1U, depth / 4);
444 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
445 int node, bool reserved)
449 bt->bits_per_word = ilog2(BITS_PER_LONG);
452 * Depth can be zero for reserved tags, that's not a failure
456 unsigned int nr, tags_per_word;
458 tags_per_word = (1 << bt->bits_per_word);
461 * If the tag space is small, shrink the number of tags
462 * per word so we spread over a few cachelines, at least.
463 * If less than 4 tags, just forget about it, it's not
464 * going to work optimally anyway.
467 while (tags_per_word * 4 > depth) {
469 tags_per_word = (1 << bt->bits_per_word);
473 nr = ALIGN(depth, tags_per_word) / tags_per_word;
474 bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
482 bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
488 for (i = 0; i < BT_WAIT_QUEUES; i++)
489 init_waitqueue_head(&bt->bs[i].wait);
491 bt_update_count(bt, depth);
495 static void bt_free(struct blk_mq_bitmap_tags *bt)
501 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
504 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
506 if (bt_alloc(&tags->bitmap_tags, depth, node, false))
508 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
513 bt_free(&tags->bitmap_tags);
518 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
519 unsigned int reserved_tags, int node)
521 struct blk_mq_tags *tags;
523 if (total_tags > BLK_MQ_TAG_MAX) {
524 pr_err("blk-mq: tag depth too large\n");
528 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
532 tags->nr_tags = total_tags;
533 tags->nr_reserved_tags = reserved_tags;
535 return blk_mq_init_bitmap_tags(tags, node);
538 void blk_mq_free_tags(struct blk_mq_tags *tags)
540 bt_free(&tags->bitmap_tags);
541 bt_free(&tags->breserved_tags);
545 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
547 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
549 *tag = prandom_u32() % depth;
552 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
554 tdepth -= tags->nr_reserved_tags;
555 if (tdepth > tags->nr_tags)
559 * Don't need (or can't) update reserved tags here, they remain
560 * static and should never need resizing.
562 bt_update_count(&tags->bitmap_tags, tdepth);
563 blk_mq_tag_wakeup_all(tags);
567 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
569 char *orig_page = page;
570 unsigned int free, res;
575 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
576 "bits_per_word=%u\n",
577 tags->nr_tags, tags->nr_reserved_tags,
578 tags->bitmap_tags.bits_per_word);
580 free = bt_unused_tags(&tags->bitmap_tags);
581 res = bt_unused_tags(&tags->breserved_tags);
583 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
584 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
586 return page - orig_page;