xref: /openbmc/linux/block/blk-mq.h (revision 67559900)
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 
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);
73 
74 /*
75  * CPU -> queue mappings
76  */
77 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
78 
79 /*
80  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
81  * @q: request queue
82  * @type: the hctx type index
83  * @cpu: CPU
84  */
85 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
86 							  enum hctx_type type,
87 							  unsigned int cpu)
88 {
89 	return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
90 }
91 
92 static inline enum hctx_type blk_mq_get_hctx_type(unsigned int flags)
93 {
94 	enum hctx_type type = HCTX_TYPE_DEFAULT;
95 
96 	/*
97 	 * The caller ensure that if REQ_POLLED, poll must be enabled.
98 	 */
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;
103 	return type;
104 }
105 
106 /*
107  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
108  * @q: request queue
109  * @flags: request command flags
110  * @ctx: software queue cpu ctx
111  */
112 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
113 						     unsigned int flags,
114 						     struct blk_mq_ctx *ctx)
115 {
116 	return ctx->hctxs[blk_mq_get_hctx_type(flags)];
117 }
118 
119 /*
120  * sysfs helpers
121  */
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);
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 	unsigned int 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  * @q: request queue
298  * @bio : the bio being submitted by the caller context
299  *
300  * Plugging, by design, may delay the insertion of BIOs into the elevator in
301  * order to increase BIO merging opportunities. This however can cause BIO
302  * insertion order to change from the order in which submit_bio() is being
303  * executed in the case of multiple contexts concurrently issuing BIOs to a
304  * device, even if these context are synchronized to tightly control BIO issuing
305  * order. While this is not a problem with regular block devices, this ordering
306  * change can cause write BIO failures with zoned block devices as these
307  * require sequential write patterns to zones. Prevent this from happening by
308  * ignoring the plug state of a BIO issuing context if the target request queue
309  * is for a zoned block device and the BIO to plug is a write operation.
310  *
311  * Return current->plug if the bio can be plugged and NULL otherwise
312  */
313 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
314 					   struct bio *bio)
315 {
316 	/*
317 	 * For regular block devices or read operations, use the context plug
318 	 * which may be NULL if blk_start_plug() was not executed.
319 	 */
320 	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
321 		return current->plug;
322 
323 	/* Zoned block device write operation case: do not plug the BIO */
324 	return NULL;
325 }
326 
327 /* Free all requests on the list */
328 static inline void blk_mq_free_requests(struct list_head *list)
329 {
330 	while (!list_empty(list)) {
331 		struct request *rq = list_entry_rq(list->next);
332 
333 		list_del_init(&rq->queuelist);
334 		blk_mq_free_request(rq);
335 	}
336 }
337 
338 /*
339  * For shared tag users, we track the number of currently active users
340  * and attempt to provide a fair share of the tag depth for each of them.
341  */
342 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
343 				  struct sbitmap_queue *bt)
344 {
345 	unsigned int depth, users;
346 
347 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
348 		return true;
349 
350 	/*
351 	 * Don't try dividing an ant
352 	 */
353 	if (bt->sb.depth == 1)
354 		return true;
355 
356 	if (blk_mq_is_shared_tags(hctx->flags)) {
357 		struct request_queue *q = hctx->queue;
358 
359 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
360 			return true;
361 	} else {
362 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
363 			return true;
364 	}
365 
366 	users = atomic_read(&hctx->tags->active_queues);
367 
368 	if (!users)
369 		return true;
370 
371 	/*
372 	 * Allow at least some tags
373 	 */
374 	depth = max((bt->sb.depth + users - 1) / users, 4U);
375 	return __blk_mq_active_requests(hctx) < depth;
376 }
377 
378 
379 #endif
380