xref: /linux/block/blk-mq.h (revision 7a3a401874bea02f568aa416ac29170d8cde0dc2)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef INT_BLK_MQ_H
3 #define INT_BLK_MQ_H
4 
5 #include "blk-stat.h"
6 #include "blk-mq-tag.h"
7 
8 struct blk_mq_tag_set;
9 
10 struct blk_mq_ctxs {
11 	struct kobject kobj;
12 	struct blk_mq_ctx __percpu	*queue_ctx;
13 };
14 
15 /**
16  * struct blk_mq_ctx - State for a software queue facing the submitting CPUs
17  */
18 struct blk_mq_ctx {
19 	struct {
20 		spinlock_t		lock;
21 		struct list_head	rq_lists[HCTX_MAX_TYPES];
22 	} ____cacheline_aligned_in_smp;
23 
24 	unsigned int		cpu;
25 	unsigned short		index_hw[HCTX_MAX_TYPES];
26 	struct blk_mq_hw_ctx 	*hctxs[HCTX_MAX_TYPES];
27 
28 	struct request_queue	*queue;
29 	struct blk_mq_ctxs      *ctxs;
30 	struct kobject		kobj;
31 } ____cacheline_aligned_in_smp;
32 
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,
35 		unsigned int flags);
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 *,
40 			     unsigned int);
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);
47 
48 /*
49  * Internal helpers for allocating/freeing the request map
50  */
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);
59 /*
60  * Internal helpers for request insertion into sw queues
61  */
62 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
63 				bool at_head);
64 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
65 				  bool run_queue);
66 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
67 				struct list_head *list);
68 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
69 				    struct list_head *list);
70 
71 /*
72  * CPU -> queue mappings
73  */
74 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
75 
76 /*
77  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
78  * @q: request queue
79  * @type: the hctx type index
80  * @cpu: CPU
81  */
82 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
83 							  enum hctx_type type,
84 							  unsigned int cpu)
85 {
86 	return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]);
87 }
88 
89 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf)
90 {
91 	enum hctx_type type = HCTX_TYPE_DEFAULT;
92 
93 	/*
94 	 * The caller ensure that if REQ_POLLED, poll must be enabled.
95 	 */
96 	if (opf & REQ_POLLED)
97 		type = HCTX_TYPE_POLL;
98 	else if ((opf & REQ_OP_MASK) == REQ_OP_READ)
99 		type = HCTX_TYPE_READ;
100 	return type;
101 }
102 
103 /*
104  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
105  * @q: request queue
106  * @opf: operation type (REQ_OP_*) and flags (e.g. REQ_POLLED).
107  * @ctx: software queue cpu ctx
108  */
109 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
110 						     blk_opf_t opf,
111 						     struct blk_mq_ctx *ctx)
112 {
113 	return ctx->hctxs[blk_mq_get_hctx_type(opf)];
114 }
115 
116 /*
117  * sysfs helpers
118  */
119 extern void blk_mq_sysfs_init(struct request_queue *q);
120 extern void blk_mq_sysfs_deinit(struct request_queue *q);
121 int blk_mq_sysfs_register(struct gendisk *disk);
122 void blk_mq_sysfs_unregister(struct gendisk *disk);
123 int blk_mq_sysfs_register_hctxs(struct request_queue *q);
124 void blk_mq_sysfs_unregister_hctxs(struct request_queue *q);
125 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
126 void blk_mq_free_plug_rqs(struct blk_plug *plug);
127 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
128 
129 void blk_mq_cancel_work_sync(struct request_queue *q);
130 
131 void blk_mq_release(struct request_queue *q);
132 
133 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
134 					   unsigned int cpu)
135 {
136 	return per_cpu_ptr(q->queue_ctx, cpu);
137 }
138 
139 /*
140  * This assumes per-cpu software queueing queues. They could be per-node
141  * as well, for instance. For now this is hardcoded as-is. Note that we don't
142  * care about preemption, since we know the ctx's are persistent. This does
143  * mean that we can't rely on ctx always matching the currently running CPU.
144  */
145 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
146 {
147 	return __blk_mq_get_ctx(q, raw_smp_processor_id());
148 }
149 
150 struct blk_mq_alloc_data {
151 	/* input parameter */
152 	struct request_queue *q;
153 	blk_mq_req_flags_t flags;
154 	unsigned int shallow_depth;
155 	blk_opf_t cmd_flags;
156 	req_flags_t rq_flags;
157 
158 	/* allocate multiple requests/tags in one go */
159 	unsigned int nr_tags;
160 	struct request **cached_rq;
161 
162 	/* input & output parameter */
163 	struct blk_mq_ctx *ctx;
164 	struct blk_mq_hw_ctx *hctx;
165 };
166 
167 static inline bool blk_mq_is_shared_tags(unsigned int flags)
168 {
169 	return flags & BLK_MQ_F_TAG_HCTX_SHARED;
170 }
171 
172 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
173 {
174 	if (!(data->rq_flags & RQF_ELV))
175 		return data->hctx->tags;
176 	return data->hctx->sched_tags;
177 }
178 
179 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
180 {
181 	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
182 }
183 
184 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
185 {
186 	return hctx->nr_ctx && hctx->tags;
187 }
188 
189 unsigned int blk_mq_in_flight(struct request_queue *q,
190 		struct block_device *part);
191 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
192 		unsigned int inflight[2]);
193 
194 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
195 					      int budget_token)
196 {
197 	if (q->mq_ops->put_budget)
198 		q->mq_ops->put_budget(q, budget_token);
199 }
200 
201 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
202 {
203 	if (q->mq_ops->get_budget)
204 		return q->mq_ops->get_budget(q);
205 	return 0;
206 }
207 
208 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
209 {
210 	if (token < 0)
211 		return;
212 
213 	if (rq->q->mq_ops->set_rq_budget_token)
214 		rq->q->mq_ops->set_rq_budget_token(rq, token);
215 }
216 
217 static inline int blk_mq_get_rq_budget_token(struct request *rq)
218 {
219 	if (rq->q->mq_ops->get_rq_budget_token)
220 		return rq->q->mq_ops->get_rq_budget_token(rq);
221 	return -1;
222 }
223 
224 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
225 {
226 	if (blk_mq_is_shared_tags(hctx->flags))
227 		atomic_inc(&hctx->queue->nr_active_requests_shared_tags);
228 	else
229 		atomic_inc(&hctx->nr_active);
230 }
231 
232 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
233 		int val)
234 {
235 	if (blk_mq_is_shared_tags(hctx->flags))
236 		atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
237 	else
238 		atomic_sub(val, &hctx->nr_active);
239 }
240 
241 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
242 {
243 	__blk_mq_sub_active_requests(hctx, 1);
244 }
245 
246 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
247 {
248 	if (blk_mq_is_shared_tags(hctx->flags))
249 		return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
250 	return atomic_read(&hctx->nr_active);
251 }
252 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
253 					   struct request *rq)
254 {
255 	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
256 	rq->tag = BLK_MQ_NO_TAG;
257 
258 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
259 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
260 		__blk_mq_dec_active_requests(hctx);
261 	}
262 }
263 
264 static inline void blk_mq_put_driver_tag(struct request *rq)
265 {
266 	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
267 		return;
268 
269 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
270 }
271 
272 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq);
273 
274 static inline bool blk_mq_get_driver_tag(struct request *rq)
275 {
276 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
277 
278 	if (rq->tag != BLK_MQ_NO_TAG &&
279 	    !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
280 		hctx->tags->rqs[rq->tag] = rq;
281 		return true;
282 	}
283 
284 	return __blk_mq_get_driver_tag(hctx, rq);
285 }
286 
287 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
288 {
289 	int cpu;
290 
291 	for_each_possible_cpu(cpu)
292 		qmap->mq_map[cpu] = 0;
293 }
294 
295 /*
296  * blk_mq_plug() - Get caller context plug
297  * @bio : the bio being submitted by the caller context
298  *
299  * Plugging, by design, may delay the insertion of BIOs into the elevator in
300  * order to increase BIO merging opportunities. This however can cause BIO
301  * insertion order to change from the order in which submit_bio() is being
302  * executed in the case of multiple contexts concurrently issuing BIOs to a
303  * device, even if these context are synchronized to tightly control BIO issuing
304  * order. While this is not a problem with regular block devices, this ordering
305  * change can cause write BIO failures with zoned block devices as these
306  * require sequential write patterns to zones. Prevent this from happening by
307  * ignoring the plug state of a BIO issuing context if it is for a zoned block
308  * device and the BIO to plug is a write operation.
309  *
310  * Return current->plug if the bio can be plugged and NULL otherwise
311  */
312 static inline struct blk_plug *blk_mq_plug( struct bio *bio)
313 {
314 	/* Zoned block device write operation case: do not plug the BIO */
315 	if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
316 	    bdev_op_is_zoned_write(bio->bi_bdev, bio_op(bio)))
317 		return NULL;
318 
319 	/*
320 	 * For regular block devices or read operations, use the context plug
321 	 * which may be NULL if blk_start_plug() was not executed.
322 	 */
323 	return current->plug;
324 }
325 
326 /* Free all requests on the list */
327 static inline void blk_mq_free_requests(struct list_head *list)
328 {
329 	while (!list_empty(list)) {
330 		struct request *rq = list_entry_rq(list->next);
331 
332 		list_del_init(&rq->queuelist);
333 		blk_mq_free_request(rq);
334 	}
335 }
336 
337 /*
338  * For shared tag users, we track the number of currently active users
339  * and attempt to provide a fair share of the tag depth for each of them.
340  */
341 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
342 				  struct sbitmap_queue *bt)
343 {
344 	unsigned int depth, users;
345 
346 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
347 		return true;
348 
349 	/*
350 	 * Don't try dividing an ant
351 	 */
352 	if (bt->sb.depth == 1)
353 		return true;
354 
355 	if (blk_mq_is_shared_tags(hctx->flags)) {
356 		struct request_queue *q = hctx->queue;
357 
358 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
359 			return true;
360 	} else {
361 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
362 			return true;
363 	}
364 
365 	users = atomic_read(&hctx->tags->active_queues);
366 
367 	if (!users)
368 		return true;
369 
370 	/*
371 	 * Allow at least some tags
372 	 */
373 	depth = max((bt->sb.depth + users - 1) / users, 4U);
374 	return __blk_mq_active_requests(hctx) < depth;
375 }
376 
377 /* run the code block in @dispatch_ops with rcu/srcu read lock held */
378 #define __blk_mq_run_dispatch_ops(q, check_sleep, dispatch_ops)	\
379 do {								\
380 	if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) {		\
381 		struct blk_mq_tag_set *__tag_set = (q)->tag_set; \
382 		int srcu_idx;					\
383 								\
384 		might_sleep_if(check_sleep);			\
385 		srcu_idx = srcu_read_lock(__tag_set->srcu);	\
386 		(dispatch_ops);					\
387 		srcu_read_unlock(__tag_set->srcu, srcu_idx);	\
388 	} else {						\
389 		rcu_read_lock();				\
390 		(dispatch_ops);					\
391 		rcu_read_unlock();				\
392 	}							\
393 } while (0)
394 
395 #define blk_mq_run_dispatch_ops(q, dispatch_ops)		\
396 	__blk_mq_run_dispatch_ops(q, true, dispatch_ops)	\
397 
398 #endif
399