xref: /linux/block/blk-mq.h (revision 94e48d6aafef23143f92eadd010c505c49487576)
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 	/* incremented at dispatch time */
29 	unsigned long		rq_dispatched[2];
30 	unsigned long		rq_merged;
31 
32 	/* incremented at completion time */
33 	unsigned long		____cacheline_aligned_in_smp rq_completed[2];
34 
35 	struct request_queue	*queue;
36 	struct blk_mq_ctxs      *ctxs;
37 	struct kobject		kobj;
38 } ____cacheline_aligned_in_smp;
39 
40 void blk_mq_exit_queue(struct request_queue *q);
41 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
42 void blk_mq_wake_waiters(struct request_queue *q);
43 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
44 			     unsigned int);
45 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
46 				bool kick_requeue_list);
47 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
48 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
49 					struct blk_mq_ctx *start);
50 void blk_mq_put_rq_ref(struct request *rq);
51 
52 /*
53  * Internal helpers for allocating/freeing the request map
54  */
55 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
56 		     unsigned int hctx_idx);
57 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags);
58 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
59 					unsigned int hctx_idx,
60 					unsigned int nr_tags,
61 					unsigned int reserved_tags,
62 					unsigned int flags);
63 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
64 		     unsigned int hctx_idx, unsigned int depth);
65 
66 /*
67  * Internal helpers for request insertion into sw queues
68  */
69 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
70 				bool at_head);
71 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
72 				  bool run_queue);
73 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
74 				struct list_head *list);
75 
76 /* Used by blk_insert_cloned_request() to issue request directly */
77 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
78 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
79 				    struct list_head *list);
80 
81 /*
82  * CPU -> queue mappings
83  */
84 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
85 
86 /*
87  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
88  * @q: request queue
89  * @type: the hctx type index
90  * @cpu: CPU
91  */
92 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
93 							  enum hctx_type type,
94 							  unsigned int cpu)
95 {
96 	return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
97 }
98 
99 /*
100  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
101  * @q: request queue
102  * @flags: request command flags
103  * @ctx: software queue cpu ctx
104  */
105 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
106 						     unsigned int flags,
107 						     struct blk_mq_ctx *ctx)
108 {
109 	enum hctx_type type = HCTX_TYPE_DEFAULT;
110 
111 	/*
112 	 * The caller ensure that if REQ_HIPRI, poll must be enabled.
113 	 */
114 	if (flags & REQ_HIPRI)
115 		type = HCTX_TYPE_POLL;
116 	else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
117 		type = HCTX_TYPE_READ;
118 
119 	return ctx->hctxs[type];
120 }
121 
122 /*
123  * sysfs helpers
124  */
125 extern void blk_mq_sysfs_init(struct request_queue *q);
126 extern void blk_mq_sysfs_deinit(struct request_queue *q);
127 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
128 extern int blk_mq_sysfs_register(struct request_queue *q);
129 extern void blk_mq_sysfs_unregister(struct request_queue *q);
130 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
131 
132 void blk_mq_release(struct request_queue *q);
133 
134 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
135 					   unsigned int cpu)
136 {
137 	return per_cpu_ptr(q->queue_ctx, cpu);
138 }
139 
140 /*
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.
145  */
146 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
147 {
148 	return __blk_mq_get_ctx(q, raw_smp_processor_id());
149 }
150 
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;
156 	unsigned int cmd_flags;
157 
158 	/* input & output parameter */
159 	struct blk_mq_ctx *ctx;
160 	struct blk_mq_hw_ctx *hctx;
161 };
162 
163 static inline bool blk_mq_is_sbitmap_shared(unsigned int flags)
164 {
165 	return flags & BLK_MQ_F_TAG_HCTX_SHARED;
166 }
167 
168 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
169 {
170 	if (data->q->elevator)
171 		return data->hctx->sched_tags;
172 
173 	return data->hctx->tags;
174 }
175 
176 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
177 {
178 	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
179 }
180 
181 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
182 {
183 	return hctx->nr_ctx && hctx->tags;
184 }
185 
186 unsigned int blk_mq_in_flight(struct request_queue *q,
187 		struct block_device *part);
188 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
189 		unsigned int inflight[2]);
190 
191 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
192 					      int budget_token)
193 {
194 	if (q->mq_ops->put_budget)
195 		q->mq_ops->put_budget(q, budget_token);
196 }
197 
198 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
199 {
200 	if (q->mq_ops->get_budget)
201 		return q->mq_ops->get_budget(q);
202 	return 0;
203 }
204 
205 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
206 {
207 	if (token < 0)
208 		return;
209 
210 	if (rq->q->mq_ops->set_rq_budget_token)
211 		rq->q->mq_ops->set_rq_budget_token(rq, token);
212 }
213 
214 static inline int blk_mq_get_rq_budget_token(struct request *rq)
215 {
216 	if (rq->q->mq_ops->get_rq_budget_token)
217 		return rq->q->mq_ops->get_rq_budget_token(rq);
218 	return -1;
219 }
220 
221 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
222 {
223 	if (blk_mq_is_sbitmap_shared(hctx->flags))
224 		atomic_inc(&hctx->queue->nr_active_requests_shared_sbitmap);
225 	else
226 		atomic_inc(&hctx->nr_active);
227 }
228 
229 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
230 {
231 	if (blk_mq_is_sbitmap_shared(hctx->flags))
232 		atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap);
233 	else
234 		atomic_dec(&hctx->nr_active);
235 }
236 
237 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
238 {
239 	if (blk_mq_is_sbitmap_shared(hctx->flags))
240 		return atomic_read(&hctx->queue->nr_active_requests_shared_sbitmap);
241 	return atomic_read(&hctx->nr_active);
242 }
243 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
244 					   struct request *rq)
245 {
246 	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
247 	rq->tag = BLK_MQ_NO_TAG;
248 
249 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
250 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
251 		__blk_mq_dec_active_requests(hctx);
252 	}
253 }
254 
255 static inline void blk_mq_put_driver_tag(struct request *rq)
256 {
257 	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
258 		return;
259 
260 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
261 }
262 
263 bool blk_mq_get_driver_tag(struct request *rq);
264 
265 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
266 {
267 	int cpu;
268 
269 	for_each_possible_cpu(cpu)
270 		qmap->mq_map[cpu] = 0;
271 }
272 
273 /*
274  * blk_mq_plug() - Get caller context plug
275  * @q: request queue
276  * @bio : the bio being submitted by the caller context
277  *
278  * Plugging, by design, may delay the insertion of BIOs into the elevator in
279  * order to increase BIO merging opportunities. This however can cause BIO
280  * insertion order to change from the order in which submit_bio() is being
281  * executed in the case of multiple contexts concurrently issuing BIOs to a
282  * device, even if these context are synchronized to tightly control BIO issuing
283  * order. While this is not a problem with regular block devices, this ordering
284  * change can cause write BIO failures with zoned block devices as these
285  * require sequential write patterns to zones. Prevent this from happening by
286  * ignoring the plug state of a BIO issuing context if the target request queue
287  * is for a zoned block device and the BIO to plug is a write operation.
288  *
289  * Return current->plug if the bio can be plugged and NULL otherwise
290  */
291 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
292 					   struct bio *bio)
293 {
294 	/*
295 	 * For regular block devices or read operations, use the context plug
296 	 * which may be NULL if blk_start_plug() was not executed.
297 	 */
298 	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
299 		return current->plug;
300 
301 	/* Zoned block device write operation case: do not plug the BIO */
302 	return NULL;
303 }
304 
305 /* Free all requests on the list */
306 static inline void blk_mq_free_requests(struct list_head *list)
307 {
308 	while (!list_empty(list)) {
309 		struct request *rq = list_entry_rq(list->next);
310 
311 		list_del_init(&rq->queuelist);
312 		blk_mq_free_request(rq);
313 	}
314 }
315 
316 /*
317  * For shared tag users, we track the number of currently active users
318  * and attempt to provide a fair share of the tag depth for each of them.
319  */
320 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
321 				  struct sbitmap_queue *bt)
322 {
323 	unsigned int depth, users;
324 
325 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
326 		return true;
327 
328 	/*
329 	 * Don't try dividing an ant
330 	 */
331 	if (bt->sb.depth == 1)
332 		return true;
333 
334 	if (blk_mq_is_sbitmap_shared(hctx->flags)) {
335 		struct request_queue *q = hctx->queue;
336 		struct blk_mq_tag_set *set = q->tag_set;
337 
338 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
339 			return true;
340 		users = atomic_read(&set->active_queues_shared_sbitmap);
341 	} else {
342 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
343 			return true;
344 		users = atomic_read(&hctx->tags->active_queues);
345 	}
346 
347 	if (!users)
348 		return true;
349 
350 	/*
351 	 * Allow at least some tags
352 	 */
353 	depth = max((bt->sb.depth + users - 1) / users, 4U);
354 	return __blk_mq_active_requests(hctx) < depth;
355 }
356 
357 
358 #endif
359