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>
10 #include <linux/random.h>
12 #include <linux/blk-mq.h>
15 #include "blk-mq-tag.h"
17 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
22 return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
26 * If a previously inactive queue goes active, bump the active user count.
28 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
30 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
31 !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
32 atomic_inc(&hctx->tags->active_queues);
38 * Wakeup all potentially sleeping on tags
40 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
42 sbitmap_queue_wake_all(&tags->bitmap_tags);
44 sbitmap_queue_wake_all(&tags->breserved_tags);
48 * If a previously busy queue goes inactive, potential waiters could now
49 * be allowed to queue. Wake them up and check.
51 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
53 struct blk_mq_tags *tags = hctx->tags;
55 if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
58 atomic_dec(&tags->active_queues);
60 blk_mq_tag_wakeup_all(tags, false);
64 * For shared tag users, we track the number of currently active users
65 * and attempt to provide a fair share of the tag depth for each of them.
67 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
68 struct sbitmap_queue *bt)
70 unsigned int depth, users;
72 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
74 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
78 * Don't try dividing an ant
80 if (bt->sb.depth == 1)
83 users = atomic_read(&hctx->tags->active_queues);
88 * Allow at least some tags
90 depth = max((bt->sb.depth + users - 1) / users, 4U);
91 return atomic_read(&hctx->nr_active) < depth;
94 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt)
96 if (!hctx_may_queue(hctx, bt))
98 return __sbitmap_queue_get(bt);
101 static int bt_get(struct blk_mq_alloc_data *data, struct sbitmap_queue *bt,
102 struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags)
104 struct sbq_wait_state *ws;
108 tag = __bt_get(hctx, bt);
112 if (data->flags & BLK_MQ_REQ_NOWAIT)
115 ws = bt_wait_ptr(bt, hctx);
117 prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE);
119 tag = __bt_get(hctx, bt);
124 * We're out of tags on this hardware queue, kick any
125 * pending IO submits before going to sleep waiting for
126 * some to complete. Note that hctx can be NULL here for
127 * reserved tag allocation.
130 blk_mq_run_hw_queue(hctx, false);
133 * Retry tag allocation after running the hardware queue,
134 * as running the queue may also have found completions.
136 tag = __bt_get(hctx, bt);
140 blk_mq_put_ctx(data->ctx);
144 data->ctx = blk_mq_get_ctx(data->q);
145 data->hctx = data->q->mq_ops->map_queue(data->q,
147 if (data->flags & BLK_MQ_REQ_RESERVED) {
148 bt = &data->hctx->tags->breserved_tags;
151 bt = &hctx->tags->bitmap_tags;
153 finish_wait(&ws->wait, &wait);
154 ws = bt_wait_ptr(bt, hctx);
157 finish_wait(&ws->wait, &wait);
161 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
165 tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
168 return tag + data->hctx->tags->nr_reserved_tags;
170 return BLK_MQ_TAG_FAIL;
173 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
177 if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
179 return BLK_MQ_TAG_FAIL;
182 tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL,
185 return BLK_MQ_TAG_FAIL;
190 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
192 if (data->flags & BLK_MQ_REQ_RESERVED)
193 return __blk_mq_get_reserved_tag(data);
194 return __blk_mq_get_tag(data);
197 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
200 struct blk_mq_tags *tags = hctx->tags;
202 if (tag >= tags->nr_reserved_tags) {
203 const int real_tag = tag - tags->nr_reserved_tags;
205 BUG_ON(real_tag >= tags->nr_tags);
206 sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
208 BUG_ON(tag >= tags->nr_reserved_tags);
209 sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
213 struct bt_iter_data {
214 struct blk_mq_hw_ctx *hctx;
220 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
222 struct bt_iter_data *iter_data = data;
223 struct blk_mq_hw_ctx *hctx = iter_data->hctx;
224 struct blk_mq_tags *tags = hctx->tags;
225 bool reserved = iter_data->reserved;
229 bitnr += tags->nr_reserved_tags;
230 rq = tags->rqs[bitnr];
232 if (rq->q == hctx->queue)
233 iter_data->fn(hctx, rq, iter_data->data, reserved);
237 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
238 busy_iter_fn *fn, void *data, bool reserved)
240 struct bt_iter_data iter_data = {
244 .reserved = reserved,
247 sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
250 struct bt_tags_iter_data {
251 struct blk_mq_tags *tags;
252 busy_tag_iter_fn *fn;
257 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
259 struct bt_tags_iter_data *iter_data = data;
260 struct blk_mq_tags *tags = iter_data->tags;
261 bool reserved = iter_data->reserved;
265 bitnr += tags->nr_reserved_tags;
266 rq = tags->rqs[bitnr];
268 iter_data->fn(rq, iter_data->data, reserved);
272 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
273 busy_tag_iter_fn *fn, void *data, bool reserved)
275 struct bt_tags_iter_data iter_data = {
279 .reserved = reserved,
283 sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
286 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
287 busy_tag_iter_fn *fn, void *priv)
289 if (tags->nr_reserved_tags)
290 bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
291 bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
294 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
295 busy_tag_iter_fn *fn, void *priv)
299 for (i = 0; i < tagset->nr_hw_queues; i++) {
300 if (tagset->tags && tagset->tags[i])
301 blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
304 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
306 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set)
310 if (!set->ops->reinit_request)
313 for (i = 0; i < set->nr_hw_queues; i++) {
314 struct blk_mq_tags *tags = set->tags[i];
316 for (j = 0; j < tags->nr_tags; j++) {
320 ret = set->ops->reinit_request(set->driver_data,
330 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset);
332 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
335 struct blk_mq_hw_ctx *hctx;
339 queue_for_each_hw_ctx(q, hctx, i) {
340 struct blk_mq_tags *tags = hctx->tags;
343 * If not software queues are currently mapped to this
344 * hardware queue, there's nothing to check
346 if (!blk_mq_hw_queue_mapped(hctx))
349 if (tags->nr_reserved_tags)
350 bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
351 bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
356 static unsigned int bt_unused_tags(const struct sbitmap_queue *bt)
358 return bt->sb.depth - sbitmap_weight(&bt->sb);
361 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
362 bool round_robin, int node)
364 return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
368 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
369 int node, int alloc_policy)
371 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
372 bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
374 if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
376 if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
378 goto free_bitmap_tags;
382 sbitmap_queue_free(&tags->bitmap_tags);
388 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
389 unsigned int reserved_tags,
390 int node, int alloc_policy)
392 struct blk_mq_tags *tags;
394 if (total_tags > BLK_MQ_TAG_MAX) {
395 pr_err("blk-mq: tag depth too large\n");
399 tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
403 if (!zalloc_cpumask_var(&tags->cpumask, GFP_KERNEL)) {
408 tags->nr_tags = total_tags;
409 tags->nr_reserved_tags = reserved_tags;
411 return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
414 void blk_mq_free_tags(struct blk_mq_tags *tags)
416 sbitmap_queue_free(&tags->bitmap_tags);
417 sbitmap_queue_free(&tags->breserved_tags);
418 free_cpumask_var(tags->cpumask);
422 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
424 unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
426 *tag = prandom_u32() % depth;
429 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
431 tdepth -= tags->nr_reserved_tags;
432 if (tdepth > tags->nr_tags)
436 * Don't need (or can't) update reserved tags here, they remain
437 * static and should never need resizing.
439 sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
441 blk_mq_tag_wakeup_all(tags, false);
446 * blk_mq_unique_tag() - return a tag that is unique queue-wide
447 * @rq: request for which to compute a unique tag
449 * The tag field in struct request is unique per hardware queue but not over
450 * all hardware queues. Hence this function that returns a tag with the
451 * hardware context index in the upper bits and the per hardware queue tag in
454 * Note: When called for a request that is queued on a non-multiqueue request
455 * queue, the hardware context index is set to zero.
457 u32 blk_mq_unique_tag(struct request *rq)
459 struct request_queue *q = rq->q;
460 struct blk_mq_hw_ctx *hctx;
464 hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
465 hwq = hctx->queue_num;
468 return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
469 (rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
471 EXPORT_SYMBOL(blk_mq_unique_tag);
473 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
475 char *orig_page = page;
476 unsigned int free, res;
481 page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
482 "bits_per_word=%u\n",
483 tags->nr_tags, tags->nr_reserved_tags,
484 1U << tags->bitmap_tags.sb.shift);
486 free = bt_unused_tags(&tags->bitmap_tags);
487 res = bt_unused_tags(&tags->breserved_tags);
489 page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
490 page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
492 return page - orig_page;