2 * Tag allocation using scalable bitmaps. Uses active queue tracking to support
3 * fairer distribution of tags between multiple submitters when a shared tag map
6 * Copyright (C) 2013-2014 Jens Axboe
8 #include <linux/kernel.h>
9 #include <linux/module.h>
11 #include <linux/blk-mq.h>
14 #include "blk-mq-tag.h"
16 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
21 return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
25 * If a previously inactive queue goes active, bump the active user count.
27 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
29 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
30 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
31 atomic_inc(&hctx->tags->active_queues);
37 * Wakeup all potentially sleeping on tags
39 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
41 sbitmap_queue_wake_all(&tags->bitmap_tags);
43 sbitmap_queue_wake_all(&tags->breserved_tags);
47 * If a previously busy queue goes inactive, potential waiters could now
48 * be allowed to queue. Wake them up and check.
50 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
52 struct blk_mq_tags *tags = hctx->tags;
54 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
57 atomic_dec(&tags->active_queues);
59 blk_mq_tag_wakeup_all(tags, false);
63 * For shared tag users, we track the number of currently active users
64 * and attempt to provide a fair share of the tag depth for each of them.
66 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
67 struct sbitmap_queue *bt)
69 unsigned int depth, users;
71 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
73 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
77 * Don't try dividing an ant
79 if (bt->sb.depth == 1)
82 users = atomic_read(&hctx->tags->active_queues);
87 * Allow at least some tags
89 depth = max((bt->sb.depth + users - 1) / users, 4U);
90 return atomic_read(&hctx->nr_active) < depth;
93 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt)
95 if (!hctx_may_queue(hctx, bt))
97 return __sbitmap_queue_get(bt);
100 static int bt_get(struct blk_mq_alloc_data *data, struct sbitmap_queue *bt,
101 struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags)
103 struct sbq_wait_state *ws;
107 tag = __bt_get(hctx, bt);
111 if (data->flags & BLK_MQ_REQ_NOWAIT)
114 ws = bt_wait_ptr(bt, hctx);
116 prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE);
118 tag = __bt_get(hctx, bt);
123 * We're out of tags on this hardware queue, kick any
124 * pending IO submits before going to sleep waiting for
125 * some to complete. Note that hctx can be NULL here for
126 * reserved tag allocation.
129 blk_mq_run_hw_queue(hctx, false);
132 * Retry tag allocation after running the hardware queue,
133 * as running the queue may also have found completions.
135 tag = __bt_get(hctx, bt);
139 blk_mq_put_ctx(data->ctx);
143 data->ctx = blk_mq_get_ctx(data->q);
144 data->hctx = data->q->mq_ops->map_queue(data->q,
146 if (data->flags & BLK_MQ_REQ_RESERVED) {
147 bt = &data->hctx->tags->breserved_tags;
150 bt = &hctx->tags->bitmap_tags;
152 finish_wait(&ws->wait, &wait);
153 ws = bt_wait_ptr(bt, hctx);
156 finish_wait(&ws->wait, &wait);
160 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
164 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
167 return tag + data->hctx->tags->nr_reserved_tags;
169 return BLK_MQ_TAG_FAIL;
172 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
176 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
178 return BLK_MQ_TAG_FAIL;
181 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL,
184 return BLK_MQ_TAG_FAIL;
189 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
191 if (data->flags & BLK_MQ_REQ_RESERVED)
192 return __blk_mq_get_reserved_tag(data);
193 return __blk_mq_get_tag(data);
196 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
199 struct blk_mq_tags *tags = hctx->tags;
201 if (tag >= tags->nr_reserved_tags) {
202 const int real_tag = tag - tags->nr_reserved_tags;
204 BUG_ON(real_tag >= tags->nr_tags);
205 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
207 BUG_ON(tag >= tags->nr_reserved_tags);
208 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
212 struct bt_iter_data {
213 struct blk_mq_hw_ctx *hctx;
219 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
221 struct bt_iter_data *iter_data = data;
222 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
223 struct blk_mq_tags *tags = hctx->tags;
224 bool reserved = iter_data->reserved;
228 bitnr += tags->nr_reserved_tags;
229 rq = tags->rqs[bitnr];
231 if (rq->q == hctx->queue)
232 iter_data->fn(hctx, rq, iter_data->data, reserved);
236 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
237 busy_iter_fn *fn, void *data, bool reserved)
239 struct bt_iter_data iter_data = {
243 .reserved = reserved,
246 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
249 struct bt_tags_iter_data {
250 struct blk_mq_tags *tags;
251 busy_tag_iter_fn *fn;
256 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
258 struct bt_tags_iter_data *iter_data = data;
259 struct blk_mq_tags *tags = iter_data->tags;
260 bool reserved = iter_data->reserved;
264 bitnr += tags->nr_reserved_tags;
265 rq = tags->rqs[bitnr];
267 iter_data->fn(rq, iter_data->data, reserved);
271 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
272 busy_tag_iter_fn *fn, void *data, bool reserved)
274 struct bt_tags_iter_data iter_data = {
278 .reserved = reserved,
282 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
285 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
286 busy_tag_iter_fn *fn, void *priv)
288 if (tags->nr_reserved_tags)
289 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
290 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
293 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
294 busy_tag_iter_fn *fn, void *priv)
298 for (i = 0; i < tagset->nr_hw_queues; i++) {
299 if (tagset->tags && tagset->tags[i])
300 blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
303 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
305 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set)
309 if (!set->ops->reinit_request)
312 for (i = 0; i < set->nr_hw_queues; i++) {
313 struct blk_mq_tags *tags = set->tags[i];
315 for (j = 0; j < tags->nr_tags; j++) {
319 ret = set->ops->reinit_request(set->driver_data,
329 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset);
331 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
334 struct blk_mq_hw_ctx *hctx;
338 queue_for_each_hw_ctx(q, hctx, i) {
339 struct blk_mq_tags *tags = hctx->tags;
342 * If not software queues are currently mapped to this
343 * hardware queue, there's nothing to check
345 if (!blk_mq_hw_queue_mapped(hctx))
348 if (tags->nr_reserved_tags)
349 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
350 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
355 static unsigned int bt_unused_tags(const struct sbitmap_queue *bt)
357 return bt->sb.depth - sbitmap_weight(&bt->sb);
360 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
361 bool round_robin, int node)
363 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
367 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
368 int node, int alloc_policy)
370 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
371 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
373 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
375 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
377 goto free_bitmap_tags;
381 sbitmap_queue_free(&tags->bitmap_tags);
387 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
388 unsigned int reserved_tags,
389 int node, int alloc_policy)
391 struct blk_mq_tags *tags;
393 if (total_tags > BLK_MQ_TAG_MAX) {
394 pr_err("blk-mq: tag depth too large\n");
398 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
402 if (!zalloc_cpumask_var(&tags->cpumask, GFP_KERNEL)) {
407 tags->nr_tags = total_tags;
408 tags->nr_reserved_tags = reserved_tags;
410 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
413 void blk_mq_free_tags(struct blk_mq_tags *tags)
415 sbitmap_queue_free(&tags->bitmap_tags);
416 sbitmap_queue_free(&tags->breserved_tags);
417 free_cpumask_var(tags->cpumask);
421 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
423 tdepth -= tags->nr_reserved_tags;
424 if (tdepth > tags->nr_tags)
428 * Don't need (or can't) update reserved tags here, they remain
429 * static and should never need resizing.
431 sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
433 blk_mq_tag_wakeup_all(tags, false);
438 * blk_mq_unique_tag() - return a tag that is unique queue-wide
439 * @rq: request for which to compute a unique tag
441 * The tag field in struct request is unique per hardware queue but not over
442 * all hardware queues. Hence this function that returns a tag with the
443 * hardware context index in the upper bits and the per hardware queue tag in
446 * Note: When called for a request that is queued on a non-multiqueue request
447 * queue, the hardware context index is set to zero.
449 u32 blk_mq_unique_tag(struct request *rq)
451 struct request_queue *q = rq->q;
452 struct blk_mq_hw_ctx *hctx;
456 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
457 hwq = hctx->queue_num;
460 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
461 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
463 EXPORT_SYMBOL(blk_mq_unique_tag);
465 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
467 char *orig_page = page;
468 unsigned int free, res;
473 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
474 "bits_per_word=%u\n",
475 tags->nr_tags, tags->nr_reserved_tags,
476 1U << tags->bitmap_tags.sb.shift);
478 free = bt_unused_tags(&tags->bitmap_tags);
479 res = bt_unused_tags(&tags->breserved_tags);
481 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
482 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
484 return page - orig_page;