1 /* SPDX-License-Identifier: GPL-2.0 */
5 #include <linux/blk-mq.h>
12 struct blk_mq_ctx __percpu *queue_ctx;
16 * struct blk_mq_ctx - State for a software queue facing the submitting CPUs
21 struct list_head rq_lists[HCTX_MAX_TYPES];
22 } ____cacheline_aligned_in_smp;
25 unsigned short index_hw[HCTX_MAX_TYPES];
26 struct blk_mq_hw_ctx *hctxs[HCTX_MAX_TYPES];
28 struct request_queue *queue;
29 struct blk_mq_ctxs *ctxs;
31 } ____cacheline_aligned_in_smp;
36 BLK_MQ_TAG_MAX = BLK_MQ_NO_TAG - 1,
39 typedef unsigned int __bitwise blk_insert_t;
40 #define BLK_MQ_INSERT_AT_HEAD ((__force blk_insert_t)0x01)
42 void blk_mq_submit_bio(struct bio *bio);
43 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
45 void blk_mq_exit_queue(struct request_queue *q);
46 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
47 void blk_mq_wake_waiters(struct request_queue *q);
48 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
50 void blk_mq_add_to_requeue_list(struct request *rq, blk_insert_t insert_flags);
51 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
52 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
53 struct blk_mq_ctx *start);
54 void blk_mq_put_rq_ref(struct request *rq);
57 * Internal helpers for allocating/freeing the request map
59 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
60 unsigned int hctx_idx);
61 void blk_mq_free_rq_map(struct blk_mq_tags *tags);
62 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
63 unsigned int hctx_idx, unsigned int depth);
64 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
65 struct blk_mq_tags *tags,
66 unsigned int hctx_idx);
68 * Internal helpers for request insertion into sw queues
70 void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags);
73 * CPU -> queue mappings
75 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
78 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
80 * @type: the hctx type index
83 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
87 return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]);
90 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf)
92 enum hctx_type type = HCTX_TYPE_DEFAULT;
95 * The caller ensure that if REQ_POLLED, poll must be enabled.
98 type = HCTX_TYPE_POLL;
99 else if ((opf & REQ_OP_MASK) == REQ_OP_READ)
100 type = HCTX_TYPE_READ;
105 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
107 * @opf: operation type (REQ_OP_*) and flags (e.g. REQ_POLLED).
108 * @ctx: software queue cpu ctx
110 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
112 struct blk_mq_ctx *ctx)
114 return ctx->hctxs[blk_mq_get_hctx_type(opf)];
120 extern void blk_mq_sysfs_init(struct request_queue *q);
121 extern void blk_mq_sysfs_deinit(struct request_queue *q);
122 int blk_mq_sysfs_register(struct gendisk *disk);
123 void blk_mq_sysfs_unregister(struct gendisk *disk);
124 int blk_mq_sysfs_register_hctxs(struct request_queue *q);
125 void blk_mq_sysfs_unregister_hctxs(struct request_queue *q);
126 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
127 void blk_mq_free_plug_rqs(struct blk_plug *plug);
128 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
130 void blk_mq_cancel_work_sync(struct request_queue *q);
132 void blk_mq_release(struct request_queue *q);
134 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
137 return per_cpu_ptr(q->queue_ctx, cpu);
141 * This assumes per-cpu software queueing queues. They could be per-node
142 * as well, for instance. For now this is hardcoded as-is. Note that we don't
143 * care about preemption, since we know the ctx's are persistent. This does
144 * mean that we can't rely on ctx always matching the currently running CPU.
146 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
148 return __blk_mq_get_ctx(q, raw_smp_processor_id());
151 struct blk_mq_alloc_data {
152 /* input parameter */
153 struct request_queue *q;
154 blk_mq_req_flags_t flags;
155 unsigned int shallow_depth;
157 req_flags_t rq_flags;
159 /* allocate multiple requests/tags in one go */
160 unsigned int nr_tags;
161 struct request **cached_rq;
163 /* input & output parameter */
164 struct blk_mq_ctx *ctx;
165 struct blk_mq_hw_ctx *hctx;
168 struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags,
169 unsigned int reserved_tags, int node, int alloc_policy);
170 void blk_mq_free_tags(struct blk_mq_tags *tags);
171 int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
172 struct sbitmap_queue *breserved_tags, unsigned int queue_depth,
173 unsigned int reserved, int node, int alloc_policy);
175 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data);
176 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
177 unsigned int *offset);
178 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
180 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags);
181 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
182 struct blk_mq_tags **tags, unsigned int depth, bool can_grow);
183 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set,
185 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q);
187 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool);
188 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
190 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
193 static inline struct sbq_wait_state *bt_wait_ptr(struct sbitmap_queue *bt,
194 struct blk_mq_hw_ctx *hctx)
198 return sbq_wait_ptr(bt, &hctx->wait_index);
201 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *);
202 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *);
204 static inline void blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
206 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
207 __blk_mq_tag_busy(hctx);
210 static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
212 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
213 __blk_mq_tag_idle(hctx);
216 static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags,
219 return tag < tags->nr_reserved_tags;
222 static inline bool blk_mq_is_shared_tags(unsigned int flags)
224 return flags & BLK_MQ_F_TAG_HCTX_SHARED;
227 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
229 if (!(data->rq_flags & RQF_ELV))
230 return data->hctx->tags;
231 return data->hctx->sched_tags;
234 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
236 return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
239 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
241 return hctx->nr_ctx && hctx->tags;
244 unsigned int blk_mq_in_flight(struct request_queue *q,
245 struct block_device *part);
246 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
247 unsigned int inflight[2]);
249 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
252 if (q->mq_ops->put_budget)
253 q->mq_ops->put_budget(q, budget_token);
256 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
258 if (q->mq_ops->get_budget)
259 return q->mq_ops->get_budget(q);
263 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
268 if (rq->q->mq_ops->set_rq_budget_token)
269 rq->q->mq_ops->set_rq_budget_token(rq, token);
272 static inline int blk_mq_get_rq_budget_token(struct request *rq)
274 if (rq->q->mq_ops->get_rq_budget_token)
275 return rq->q->mq_ops->get_rq_budget_token(rq);
279 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
281 if (blk_mq_is_shared_tags(hctx->flags))
282 atomic_inc(&hctx->queue->nr_active_requests_shared_tags);
284 atomic_inc(&hctx->nr_active);
287 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
290 if (blk_mq_is_shared_tags(hctx->flags))
291 atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
293 atomic_sub(val, &hctx->nr_active);
296 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
298 __blk_mq_sub_active_requests(hctx, 1);
301 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
303 if (blk_mq_is_shared_tags(hctx->flags))
304 return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
305 return atomic_read(&hctx->nr_active);
307 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
310 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
311 rq->tag = BLK_MQ_NO_TAG;
313 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
314 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
315 __blk_mq_dec_active_requests(hctx);
319 static inline void blk_mq_put_driver_tag(struct request *rq)
321 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
324 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
327 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq);
329 static inline bool blk_mq_get_driver_tag(struct request *rq)
331 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
333 if (rq->tag != BLK_MQ_NO_TAG &&
334 !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
335 hctx->tags->rqs[rq->tag] = rq;
339 return __blk_mq_get_driver_tag(hctx, rq);
342 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
346 for_each_possible_cpu(cpu)
347 qmap->mq_map[cpu] = 0;
351 * blk_mq_plug() - Get caller context plug
352 * @bio : the bio being submitted by the caller context
354 * Plugging, by design, may delay the insertion of BIOs into the elevator in
355 * order to increase BIO merging opportunities. This however can cause BIO
356 * insertion order to change from the order in which submit_bio() is being
357 * executed in the case of multiple contexts concurrently issuing BIOs to a
358 * device, even if these context are synchronized to tightly control BIO issuing
359 * order. While this is not a problem with regular block devices, this ordering
360 * change can cause write BIO failures with zoned block devices as these
361 * require sequential write patterns to zones. Prevent this from happening by
362 * ignoring the plug state of a BIO issuing context if it is for a zoned block
363 * device and the BIO to plug is a write operation.
365 * Return current->plug if the bio can be plugged and NULL otherwise
367 static inline struct blk_plug *blk_mq_plug( struct bio *bio)
369 /* Zoned block device write operation case: do not plug the BIO */
370 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
371 bdev_op_is_zoned_write(bio->bi_bdev, bio_op(bio)))
375 * For regular block devices or read operations, use the context plug
376 * which may be NULL if blk_start_plug() was not executed.
378 return current->plug;
381 /* Free all requests on the list */
382 static inline void blk_mq_free_requests(struct list_head *list)
384 while (!list_empty(list)) {
385 struct request *rq = list_entry_rq(list->next);
387 list_del_init(&rq->queuelist);
388 blk_mq_free_request(rq);
393 * For shared tag users, we track the number of currently active users
394 * and attempt to provide a fair share of the tag depth for each of them.
396 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
397 struct sbitmap_queue *bt)
399 unsigned int depth, users;
401 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
405 * Don't try dividing an ant
407 if (bt->sb.depth == 1)
410 if (blk_mq_is_shared_tags(hctx->flags)) {
411 struct request_queue *q = hctx->queue;
413 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
416 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
420 users = atomic_read(&hctx->tags->active_queues);
426 * Allow at least some tags
428 depth = max((bt->sb.depth + users - 1) / users, 4U);
429 return __blk_mq_active_requests(hctx) < depth;
432 /* run the code block in @dispatch_ops with rcu/srcu read lock held */
433 #define __blk_mq_run_dispatch_ops(q, check_sleep, dispatch_ops) \
435 if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) { \
436 struct blk_mq_tag_set *__tag_set = (q)->tag_set; \
439 might_sleep_if(check_sleep); \
440 srcu_idx = srcu_read_lock(__tag_set->srcu); \
442 srcu_read_unlock(__tag_set->srcu, srcu_idx); \
450 #define blk_mq_run_dispatch_ops(q, dispatch_ops) \
451 __blk_mq_run_dispatch_ops(q, true, dispatch_ops) \