xref: /linux/block/blk-mq.h (revision 7f71507851fc7764b36a3221839607d3a45c2025)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef INT_BLK_MQ_H
3 #define INT_BLK_MQ_H
4 
5 #include <linux/blk-mq.h>
6 #include "blk-stat.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 enum {
34 	BLK_MQ_NO_TAG		= -1U,
35 	BLK_MQ_TAG_MIN		= 1,
36 	BLK_MQ_TAG_MAX		= BLK_MQ_NO_TAG - 1,
37 };
38 
39 #define BLK_MQ_CPU_WORK_BATCH	(8)
40 
41 typedef unsigned int __bitwise blk_insert_t;
42 #define BLK_MQ_INSERT_AT_HEAD		((__force blk_insert_t)0x01)
43 
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,
46 		unsigned int flags);
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 *,
51 			     unsigned int);
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);
56 
57 /*
58  * Internal helpers for allocating/freeing the request map
59  */
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);
68 
69 /*
70  * CPU -> queue mappings
71  */
72 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
73 
74 /*
75  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
76  * @q: request queue
77  * @type: the hctx type index
78  * @cpu: CPU
79  */
80 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
81 							  enum hctx_type type,
82 							  unsigned int cpu)
83 {
84 	return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]);
85 }
86 
87 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf)
88 {
89 	enum hctx_type type = HCTX_TYPE_DEFAULT;
90 
91 	/*
92 	 * The caller ensure that if REQ_POLLED, poll must be enabled.
93 	 */
94 	if (opf & REQ_POLLED)
95 		type = HCTX_TYPE_POLL;
96 	else if ((opf & REQ_OP_MASK) == REQ_OP_READ)
97 		type = HCTX_TYPE_READ;
98 	return type;
99 }
100 
101 /*
102  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
103  * @q: request queue
104  * @opf: operation type (REQ_OP_*) and flags (e.g. REQ_POLLED).
105  * @ctx: software queue cpu ctx
106  */
107 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
108 						     blk_opf_t opf,
109 						     struct blk_mq_ctx *ctx)
110 {
111 	return ctx->hctxs[blk_mq_get_hctx_type(opf)];
112 }
113 
114 /*
115  * sysfs helpers
116  */
117 extern void blk_mq_sysfs_init(struct request_queue *q);
118 extern void blk_mq_sysfs_deinit(struct request_queue *q);
119 int blk_mq_sysfs_register(struct gendisk *disk);
120 void blk_mq_sysfs_unregister(struct gendisk *disk);
121 int blk_mq_sysfs_register_hctxs(struct request_queue *q);
122 void blk_mq_sysfs_unregister_hctxs(struct request_queue *q);
123 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
124 void blk_mq_free_plug_rqs(struct blk_plug *plug);
125 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
126 
127 void blk_mq_cancel_work_sync(struct request_queue *q);
128 
129 void blk_mq_release(struct request_queue *q);
130 
131 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
132 					   unsigned int cpu)
133 {
134 	return per_cpu_ptr(q->queue_ctx, cpu);
135 }
136 
137 /*
138  * This assumes per-cpu software queueing queues. They could be per-node
139  * as well, for instance. For now this is hardcoded as-is. Note that we don't
140  * care about preemption, since we know the ctx's are persistent. This does
141  * mean that we can't rely on ctx always matching the currently running CPU.
142  */
143 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
144 {
145 	return __blk_mq_get_ctx(q, raw_smp_processor_id());
146 }
147 
148 struct blk_mq_alloc_data {
149 	/* input parameter */
150 	struct request_queue *q;
151 	blk_mq_req_flags_t flags;
152 	unsigned int shallow_depth;
153 	blk_opf_t cmd_flags;
154 	req_flags_t rq_flags;
155 
156 	/* allocate multiple requests/tags in one go */
157 	unsigned int nr_tags;
158 	struct rq_list *cached_rqs;
159 
160 	/* input & output parameter */
161 	struct blk_mq_ctx *ctx;
162 	struct blk_mq_hw_ctx *hctx;
163 };
164 
165 struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags,
166 		unsigned int reserved_tags, int node, int alloc_policy);
167 void blk_mq_free_tags(struct blk_mq_tags *tags);
168 int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
169 		struct sbitmap_queue *breserved_tags, unsigned int queue_depth,
170 		unsigned int reserved, int node, int alloc_policy);
171 
172 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data);
173 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
174 		unsigned int *offset);
175 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
176 		unsigned int tag);
177 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags);
178 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
179 		struct blk_mq_tags **tags, unsigned int depth, bool can_grow);
180 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set,
181 		unsigned int size);
182 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q);
183 
184 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool);
185 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
186 		void *priv);
187 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
188 		void *priv);
189 
190 static inline struct sbq_wait_state *bt_wait_ptr(struct sbitmap_queue *bt,
191 						 struct blk_mq_hw_ctx *hctx)
192 {
193 	if (!hctx)
194 		return &bt->ws[0];
195 	return sbq_wait_ptr(bt, &hctx->wait_index);
196 }
197 
198 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *);
199 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *);
200 
201 static inline void blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
202 {
203 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
204 		__blk_mq_tag_busy(hctx);
205 }
206 
207 static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
208 {
209 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
210 		__blk_mq_tag_idle(hctx);
211 }
212 
213 static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags,
214 					  unsigned int tag)
215 {
216 	return tag < tags->nr_reserved_tags;
217 }
218 
219 static inline bool blk_mq_is_shared_tags(unsigned int flags)
220 {
221 	return flags & BLK_MQ_F_TAG_HCTX_SHARED;
222 }
223 
224 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
225 {
226 	if (data->rq_flags & RQF_SCHED_TAGS)
227 		return data->hctx->sched_tags;
228 	return data->hctx->tags;
229 }
230 
231 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
232 {
233 	/* Fast path: hardware queue is not stopped most of the time. */
234 	if (likely(!test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
235 		return false;
236 
237 	/*
238 	 * This barrier is used to order adding of dispatch list before and
239 	 * the test of BLK_MQ_S_STOPPED below. Pairs with the memory barrier
240 	 * in blk_mq_start_stopped_hw_queue() so that dispatch code could
241 	 * either see BLK_MQ_S_STOPPED is cleared or dispatch list is not
242 	 * empty to avoid missing dispatching requests.
243 	 */
244 	smp_mb();
245 
246 	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
247 }
248 
249 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
250 {
251 	return hctx->nr_ctx && hctx->tags;
252 }
253 
254 unsigned int blk_mq_in_flight(struct request_queue *q,
255 		struct block_device *part);
256 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
257 		unsigned int inflight[2]);
258 
259 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
260 					      int budget_token)
261 {
262 	if (q->mq_ops->put_budget)
263 		q->mq_ops->put_budget(q, budget_token);
264 }
265 
266 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
267 {
268 	if (q->mq_ops->get_budget)
269 		return q->mq_ops->get_budget(q);
270 	return 0;
271 }
272 
273 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
274 {
275 	if (token < 0)
276 		return;
277 
278 	if (rq->q->mq_ops->set_rq_budget_token)
279 		rq->q->mq_ops->set_rq_budget_token(rq, token);
280 }
281 
282 static inline int blk_mq_get_rq_budget_token(struct request *rq)
283 {
284 	if (rq->q->mq_ops->get_rq_budget_token)
285 		return rq->q->mq_ops->get_rq_budget_token(rq);
286 	return -1;
287 }
288 
289 static inline void __blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx,
290 						int val)
291 {
292 	if (blk_mq_is_shared_tags(hctx->flags))
293 		atomic_add(val, &hctx->queue->nr_active_requests_shared_tags);
294 	else
295 		atomic_add(val, &hctx->nr_active);
296 }
297 
298 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
299 {
300 	__blk_mq_add_active_requests(hctx, 1);
301 }
302 
303 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
304 		int val)
305 {
306 	if (blk_mq_is_shared_tags(hctx->flags))
307 		atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
308 	else
309 		atomic_sub(val, &hctx->nr_active);
310 }
311 
312 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
313 {
314 	__blk_mq_sub_active_requests(hctx, 1);
315 }
316 
317 static inline void blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx,
318 					      int val)
319 {
320 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
321 		__blk_mq_add_active_requests(hctx, val);
322 }
323 
324 static inline void blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
325 {
326 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
327 		__blk_mq_inc_active_requests(hctx);
328 }
329 
330 static inline void blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
331 					      int val)
332 {
333 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
334 		__blk_mq_sub_active_requests(hctx, val);
335 }
336 
337 static inline void blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
338 {
339 	if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
340 		__blk_mq_dec_active_requests(hctx);
341 }
342 
343 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
344 {
345 	if (blk_mq_is_shared_tags(hctx->flags))
346 		return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
347 	return atomic_read(&hctx->nr_active);
348 }
349 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
350 					   struct request *rq)
351 {
352 	blk_mq_dec_active_requests(hctx);
353 	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
354 	rq->tag = BLK_MQ_NO_TAG;
355 }
356 
357 static inline void blk_mq_put_driver_tag(struct request *rq)
358 {
359 	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
360 		return;
361 
362 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
363 }
364 
365 bool __blk_mq_alloc_driver_tag(struct request *rq);
366 
367 static inline bool blk_mq_get_driver_tag(struct request *rq)
368 {
369 	if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
370 		return false;
371 
372 	return true;
373 }
374 
375 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
376 {
377 	int cpu;
378 
379 	for_each_possible_cpu(cpu)
380 		qmap->mq_map[cpu] = 0;
381 }
382 
383 /* Free all requests on the list */
384 static inline void blk_mq_free_requests(struct list_head *list)
385 {
386 	while (!list_empty(list)) {
387 		struct request *rq = list_entry_rq(list->next);
388 
389 		list_del_init(&rq->queuelist);
390 		blk_mq_free_request(rq);
391 	}
392 }
393 
394 /*
395  * For shared tag users, we track the number of currently active users
396  * and attempt to provide a fair share of the tag depth for each of them.
397  */
398 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
399 				  struct sbitmap_queue *bt)
400 {
401 	unsigned int depth, users;
402 
403 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
404 		return true;
405 
406 	/*
407 	 * Don't try dividing an ant
408 	 */
409 	if (bt->sb.depth == 1)
410 		return true;
411 
412 	if (blk_mq_is_shared_tags(hctx->flags)) {
413 		struct request_queue *q = hctx->queue;
414 
415 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
416 			return true;
417 	} else {
418 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
419 			return true;
420 	}
421 
422 	users = READ_ONCE(hctx->tags->active_queues);
423 	if (!users)
424 		return true;
425 
426 	/*
427 	 * Allow at least some tags
428 	 */
429 	depth = max((bt->sb.depth + users - 1) / users, 4U);
430 	return __blk_mq_active_requests(hctx) < depth;
431 }
432 
433 /* run the code block in @dispatch_ops with rcu/srcu read lock held */
434 #define __blk_mq_run_dispatch_ops(q, check_sleep, dispatch_ops)	\
435 do {								\
436 	if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) {		\
437 		struct blk_mq_tag_set *__tag_set = (q)->tag_set; \
438 		int srcu_idx;					\
439 								\
440 		might_sleep_if(check_sleep);			\
441 		srcu_idx = srcu_read_lock(__tag_set->srcu);	\
442 		(dispatch_ops);					\
443 		srcu_read_unlock(__tag_set->srcu, srcu_idx);	\
444 	} else {						\
445 		rcu_read_lock();				\
446 		(dispatch_ops);					\
447 		rcu_read_unlock();				\
448 	}							\
449 } while (0)
450 
451 #define blk_mq_run_dispatch_ops(q, dispatch_ops)		\
452 	__blk_mq_run_dispatch_ops(q, true, dispatch_ops)	\
453 
454 #endif
455