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 #define BLK_MQ_CPU_WORK_BATCH (8)
41 typedef unsigned int __bitwise blk_insert_t;
42 #define BLK_MQ_INSERT_AT_HEAD ((__force blk_insert_t)0x01)
44 void blk_mq_submit_bio(struct bio *bio);
45 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
47 void blk_mq_exit_queue(struct request_queue *q);
48 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
49 void blk_mq_wake_waiters(struct request_queue *q);
50 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
52 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
53 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
54 struct blk_mq_ctx *start);
55 void blk_mq_put_rq_ref(struct request *rq);
58 * Internal helpers for allocating/freeing the request map
60 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
61 unsigned int hctx_idx);
62 void blk_mq_free_rq_map(struct blk_mq_tags *tags);
63 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
64 unsigned int hctx_idx, unsigned int depth);
65 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
66 struct blk_mq_tags *tags,
67 unsigned int hctx_idx);
70 * CPU -> queue mappings
72 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
75 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
77 * @type: the hctx type index
80 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
84 return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]);
87 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf)
89 enum hctx_type type = HCTX_TYPE_DEFAULT;
92 * The caller ensure that if REQ_POLLED, poll must be enabled.
95 type = HCTX_TYPE_POLL;
96 else if ((opf & REQ_OP_MASK) == REQ_OP_READ)
97 type = HCTX_TYPE_READ;
102 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
103 * @opf: operation type (REQ_OP_*) and flags (e.g. REQ_POLLED).
104 * @ctx: software queue cpu ctx
106 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(blk_opf_t opf,
107 struct blk_mq_ctx *ctx)
109 return ctx->hctxs[blk_mq_get_hctx_type(opf)];
115 extern void blk_mq_sysfs_init(struct request_queue *q);
116 extern void blk_mq_sysfs_deinit(struct request_queue *q);
117 int blk_mq_sysfs_register(struct gendisk *disk);
118 void blk_mq_sysfs_unregister(struct gendisk *disk);
119 int blk_mq_sysfs_register_hctxs(struct request_queue *q);
120 void blk_mq_sysfs_unregister_hctxs(struct request_queue *q);
121 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
122 void blk_mq_free_plug_rqs(struct blk_plug *plug);
123 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
125 void blk_mq_cancel_work_sync(struct request_queue *q);
127 void blk_mq_release(struct request_queue *q);
129 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
132 return per_cpu_ptr(q->queue_ctx, cpu);
136 * This assumes per-cpu software queueing queues. They could be per-node
137 * as well, for instance. For now this is hardcoded as-is. Note that we don't
138 * care about preemption, since we know the ctx's are persistent. This does
139 * mean that we can't rely on ctx always matching the currently running CPU.
141 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
143 return __blk_mq_get_ctx(q, raw_smp_processor_id());
146 struct blk_mq_alloc_data {
147 /* input parameter */
148 struct request_queue *q;
149 blk_mq_req_flags_t flags;
150 unsigned int shallow_depth;
152 req_flags_t rq_flags;
154 /* allocate multiple requests/tags in one go */
155 unsigned int nr_tags;
156 struct rq_list *cached_rqs;
158 /* input & output parameter */
159 struct blk_mq_ctx *ctx;
160 struct blk_mq_hw_ctx *hctx;
163 struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags,
164 unsigned int reserved_tags, unsigned int flags, int node);
165 void blk_mq_free_tags(struct blk_mq_tags *tags);
167 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data);
168 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
169 unsigned int *offset);
170 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
172 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags);
173 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
174 struct blk_mq_tags **tags, unsigned int depth, bool can_grow);
175 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set,
177 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q);
179 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool);
180 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
182 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
185 static inline struct sbq_wait_state *bt_wait_ptr(struct sbitmap_queue *bt,
186 struct blk_mq_hw_ctx *hctx)
190 return sbq_wait_ptr(bt, &hctx->wait_index);
193 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *);
194 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *);
196 static inline void blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
198 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
199 __blk_mq_tag_busy(hctx);
202 static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
204 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
205 __blk_mq_tag_idle(hctx);
208 static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags,
211 return tag < tags->nr_reserved_tags;
214 static inline bool blk_mq_is_shared_tags(unsigned int flags)
216 return flags & BLK_MQ_F_TAG_HCTX_SHARED;
219 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
221 if (data->rq_flags & RQF_SCHED_TAGS)
222 return data->hctx->sched_tags;
223 return data->hctx->tags;
226 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
228 /* Fast path: hardware queue is not stopped most of the time. */
229 if (likely(!test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
233 * This barrier is used to order adding of dispatch list before and
234 * the test of BLK_MQ_S_STOPPED below. Pairs with the memory barrier
235 * in blk_mq_start_stopped_hw_queue() so that dispatch code could
236 * either see BLK_MQ_S_STOPPED is cleared or dispatch list is not
237 * empty to avoid missing dispatching requests.
241 return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
244 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
246 return hctx->nr_ctx && hctx->tags;
249 void blk_mq_in_driver_rw(struct block_device *part, unsigned int inflight[2]);
251 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
254 if (q->mq_ops->put_budget)
255 q->mq_ops->put_budget(q, budget_token);
258 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
260 if (q->mq_ops->get_budget)
261 return q->mq_ops->get_budget(q);
265 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
270 if (rq->q->mq_ops->set_rq_budget_token)
271 rq->q->mq_ops->set_rq_budget_token(rq, token);
274 static inline int blk_mq_get_rq_budget_token(struct request *rq)
276 if (rq->q->mq_ops->get_rq_budget_token)
277 return rq->q->mq_ops->get_rq_budget_token(rq);
281 static inline void __blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx,
284 if (blk_mq_is_shared_tags(hctx->flags))
285 atomic_add(val, &hctx->queue->nr_active_requests_shared_tags);
287 atomic_add(val, &hctx->nr_active);
290 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
292 __blk_mq_add_active_requests(hctx, 1);
295 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
298 if (blk_mq_is_shared_tags(hctx->flags))
299 atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
301 atomic_sub(val, &hctx->nr_active);
304 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
306 __blk_mq_sub_active_requests(hctx, 1);
309 static inline void blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx,
312 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
313 __blk_mq_add_active_requests(hctx, val);
316 static inline void blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
318 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
319 __blk_mq_inc_active_requests(hctx);
322 static inline void blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
325 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
326 __blk_mq_sub_active_requests(hctx, val);
329 static inline void blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
331 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
332 __blk_mq_dec_active_requests(hctx);
335 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
337 if (blk_mq_is_shared_tags(hctx->flags))
338 return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
339 return atomic_read(&hctx->nr_active);
341 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
344 blk_mq_dec_active_requests(hctx);
345 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
346 rq->tag = BLK_MQ_NO_TAG;
349 static inline void blk_mq_put_driver_tag(struct request *rq)
351 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
354 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
357 bool __blk_mq_alloc_driver_tag(struct request *rq);
359 static inline bool blk_mq_get_driver_tag(struct request *rq)
361 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
367 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
371 for_each_possible_cpu(cpu)
372 qmap->mq_map[cpu] = 0;
375 /* Free all requests on the list */
376 static inline void blk_mq_free_requests(struct list_head *list)
378 while (!list_empty(list)) {
379 struct request *rq = list_entry_rq(list->next);
381 list_del_init(&rq->queuelist);
382 blk_mq_free_request(rq);
387 * For shared tag users, we track the number of currently active users
388 * and attempt to provide a fair share of the tag depth for each of them.
390 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
391 struct sbitmap_queue *bt)
393 unsigned int depth, users;
395 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
399 * Don't try dividing an ant
401 if (bt->sb.depth == 1)
404 if (blk_mq_is_shared_tags(hctx->flags)) {
405 struct request_queue *q = hctx->queue;
407 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
410 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
414 users = READ_ONCE(hctx->tags->active_queues);
419 * Allow at least some tags
421 depth = max((bt->sb.depth + users - 1) / users, 4U);
422 return __blk_mq_active_requests(hctx) < depth;
425 /* run the code block in @dispatch_ops with rcu/srcu read lock held */
426 #define __blk_mq_run_dispatch_ops(q, check_sleep, dispatch_ops) \
428 if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) { \
429 struct blk_mq_tag_set *__tag_set = (q)->tag_set; \
432 might_sleep_if(check_sleep); \
433 srcu_idx = srcu_read_lock(__tag_set->srcu); \
435 srcu_read_unlock(__tag_set->srcu, srcu_idx); \
443 #define blk_mq_run_dispatch_ops(q, dispatch_ops) \
444 __blk_mq_run_dispatch_ops(q, true, dispatch_ops) \
446 static inline bool blk_mq_can_poll(struct request_queue *q)
448 return (q->limits.features & BLK_FEAT_POLL) &&
449 q->tag_set->map[HCTX_TYPE_POLL].nr_queues;