1 /* SPDX-License-Identifier: GPL-2.0 */
6 #include "blk-mq-tag.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;
33 void blk_mq_submit_bio(struct bio *bio);
34 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
36 void blk_mq_exit_queue(struct request_queue *q);
37 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
38 void blk_mq_wake_waiters(struct request_queue *q);
39 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
41 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
42 bool kick_requeue_list);
43 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
44 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
45 struct blk_mq_ctx *start);
46 void blk_mq_put_rq_ref(struct request *rq);
49 * Internal helpers for allocating/freeing the request map
51 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
52 unsigned int hctx_idx);
53 void blk_mq_free_rq_map(struct blk_mq_tags *tags);
54 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
55 unsigned int hctx_idx, unsigned int depth);
56 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
57 struct blk_mq_tags *tags,
58 unsigned int hctx_idx);
60 * Internal helpers for request insertion into sw queues
62 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
64 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
66 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
67 struct list_head *list);
69 /* Used by blk_insert_cloned_request() to issue request directly */
70 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
71 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
72 struct list_head *list);
75 * CPU -> queue mappings
77 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
80 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
82 * @type: the hctx type index
85 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
89 return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
92 static inline enum hctx_type blk_mq_get_hctx_type(unsigned int flags)
94 enum hctx_type type = HCTX_TYPE_DEFAULT;
97 * The caller ensure that if REQ_POLLED, poll must be enabled.
99 if (flags & REQ_POLLED)
100 type = HCTX_TYPE_POLL;
101 else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
102 type = HCTX_TYPE_READ;
107 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
109 * @flags: request command flags
110 * @ctx: software queue cpu ctx
112 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
114 struct blk_mq_ctx *ctx)
116 return ctx->hctxs[blk_mq_get_hctx_type(flags)];
122 extern void blk_mq_sysfs_init(struct request_queue *q);
123 extern void blk_mq_sysfs_deinit(struct request_queue *q);
124 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
125 extern int blk_mq_sysfs_register(struct request_queue *q);
126 extern void blk_mq_sysfs_unregister(struct request_queue *q);
127 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
128 void blk_mq_free_plug_rqs(struct blk_plug *plug);
129 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
131 void blk_mq_cancel_work_sync(struct request_queue *q);
133 void blk_mq_release(struct request_queue *q);
135 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
138 return per_cpu_ptr(q->queue_ctx, cpu);
142 * This assumes per-cpu software queueing queues. They could be per-node
143 * as well, for instance. For now this is hardcoded as-is. Note that we don't
144 * care about preemption, since we know the ctx's are persistent. This does
145 * mean that we can't rely on ctx always matching the currently running CPU.
147 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
149 return __blk_mq_get_ctx(q, raw_smp_processor_id());
152 struct blk_mq_alloc_data {
153 /* input parameter */
154 struct request_queue *q;
155 blk_mq_req_flags_t flags;
156 unsigned int shallow_depth;
157 unsigned int cmd_flags;
158 req_flags_t rq_flags;
160 /* allocate multiple requests/tags in one go */
161 unsigned int nr_tags;
162 struct request **cached_rq;
164 /* input & output parameter */
165 struct blk_mq_ctx *ctx;
166 struct blk_mq_hw_ctx *hctx;
169 static inline bool blk_mq_is_shared_tags(unsigned int flags)
171 return flags & BLK_MQ_F_TAG_HCTX_SHARED;
174 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
176 if (!(data->rq_flags & RQF_ELV))
177 return data->hctx->tags;
178 return data->hctx->sched_tags;
181 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
183 return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
186 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
188 return hctx->nr_ctx && hctx->tags;
191 unsigned int blk_mq_in_flight(struct request_queue *q,
192 struct block_device *part);
193 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
194 unsigned int inflight[2]);
196 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
199 if (q->mq_ops->put_budget)
200 q->mq_ops->put_budget(q, budget_token);
203 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
205 if (q->mq_ops->get_budget)
206 return q->mq_ops->get_budget(q);
210 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
215 if (rq->q->mq_ops->set_rq_budget_token)
216 rq->q->mq_ops->set_rq_budget_token(rq, token);
219 static inline int blk_mq_get_rq_budget_token(struct request *rq)
221 if (rq->q->mq_ops->get_rq_budget_token)
222 return rq->q->mq_ops->get_rq_budget_token(rq);
226 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
228 if (blk_mq_is_shared_tags(hctx->flags))
229 atomic_inc(&hctx->queue->nr_active_requests_shared_tags);
231 atomic_inc(&hctx->nr_active);
234 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
237 if (blk_mq_is_shared_tags(hctx->flags))
238 atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
240 atomic_sub(val, &hctx->nr_active);
243 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
245 __blk_mq_sub_active_requests(hctx, 1);
248 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
250 if (blk_mq_is_shared_tags(hctx->flags))
251 return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
252 return atomic_read(&hctx->nr_active);
254 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
257 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
258 rq->tag = BLK_MQ_NO_TAG;
260 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
261 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
262 __blk_mq_dec_active_requests(hctx);
266 static inline void blk_mq_put_driver_tag(struct request *rq)
268 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
271 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
274 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq);
276 static inline bool blk_mq_get_driver_tag(struct request *rq)
278 struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
280 if (rq->tag != BLK_MQ_NO_TAG &&
281 !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
282 hctx->tags->rqs[rq->tag] = rq;
286 return __blk_mq_get_driver_tag(hctx, rq);
289 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
293 for_each_possible_cpu(cpu)
294 qmap->mq_map[cpu] = 0;
298 * blk_mq_plug() - Get caller context plug
300 * @bio : the bio being submitted by the caller context
302 * Plugging, by design, may delay the insertion of BIOs into the elevator in
303 * order to increase BIO merging opportunities. This however can cause BIO
304 * insertion order to change from the order in which submit_bio() is being
305 * executed in the case of multiple contexts concurrently issuing BIOs to a
306 * device, even if these context are synchronized to tightly control BIO issuing
307 * order. While this is not a problem with regular block devices, this ordering
308 * change can cause write BIO failures with zoned block devices as these
309 * require sequential write patterns to zones. Prevent this from happening by
310 * ignoring the plug state of a BIO issuing context if the target request queue
311 * is for a zoned block device and the BIO to plug is a write operation.
313 * Return current->plug if the bio can be plugged and NULL otherwise
315 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
319 * For regular block devices or read operations, use the context plug
320 * which may be NULL if blk_start_plug() was not executed.
322 if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
323 return current->plug;
325 /* Zoned block device write operation case: do not plug the BIO */
329 /* Free all requests on the list */
330 static inline void blk_mq_free_requests(struct list_head *list)
332 while (!list_empty(list)) {
333 struct request *rq = list_entry_rq(list->next);
335 list_del_init(&rq->queuelist);
336 blk_mq_free_request(rq);
341 * For shared tag users, we track the number of currently active users
342 * and attempt to provide a fair share of the tag depth for each of them.
344 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
345 struct sbitmap_queue *bt)
347 unsigned int depth, users;
349 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
353 * Don't try dividing an ant
355 if (bt->sb.depth == 1)
358 if (blk_mq_is_shared_tags(hctx->flags)) {
359 struct request_queue *q = hctx->queue;
361 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
364 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
368 users = atomic_read(&hctx->tags->active_queues);
374 * Allow at least some tags
376 depth = max((bt->sb.depth + users - 1) / users, 4U);
377 return __blk_mq_active_requests(hctx) < depth;