xref: /openbmc/linux/block/blk-mq.h (revision 36de991e)
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_cancel_work_sync(struct request_queue *q);
132 
133 void blk_mq_release(struct request_queue *q);
134 
135 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
136 					   unsigned int cpu)
137 {
138 	return per_cpu_ptr(q->queue_ctx, cpu);
139 }
140 
141 /*
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.
146  */
147 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
148 {
149 	return __blk_mq_get_ctx(q, raw_smp_processor_id());
150 }
151 
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;
159 
160 	/* allocate multiple requests/tags in one go */
161 	unsigned int nr_tags;
162 	struct request **cached_rq;
163 
164 	/* input & output parameter */
165 	struct blk_mq_ctx *ctx;
166 	struct blk_mq_hw_ctx *hctx;
167 };
168 
169 static inline bool blk_mq_is_shared_tags(unsigned int flags)
170 {
171 	return flags & BLK_MQ_F_TAG_HCTX_SHARED;
172 }
173 
174 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
175 {
176 	if (!(data->rq_flags & RQF_ELV))
177 		return data->hctx->tags;
178 	return data->hctx->sched_tags;
179 }
180 
181 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
182 {
183 	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
184 }
185 
186 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
187 {
188 	return hctx->nr_ctx && hctx->tags;
189 }
190 
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]);
195 
196 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
197 					      int budget_token)
198 {
199 	if (q->mq_ops->put_budget)
200 		q->mq_ops->put_budget(q, budget_token);
201 }
202 
203 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
204 {
205 	if (q->mq_ops->get_budget)
206 		return q->mq_ops->get_budget(q);
207 	return 0;
208 }
209 
210 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
211 {
212 	if (token < 0)
213 		return;
214 
215 	if (rq->q->mq_ops->set_rq_budget_token)
216 		rq->q->mq_ops->set_rq_budget_token(rq, token);
217 }
218 
219 static inline int blk_mq_get_rq_budget_token(struct request *rq)
220 {
221 	if (rq->q->mq_ops->get_rq_budget_token)
222 		return rq->q->mq_ops->get_rq_budget_token(rq);
223 	return -1;
224 }
225 
226 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
227 {
228 	if (blk_mq_is_shared_tags(hctx->flags))
229 		atomic_inc(&hctx->queue->nr_active_requests_shared_tags);
230 	else
231 		atomic_inc(&hctx->nr_active);
232 }
233 
234 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
235 		int val)
236 {
237 	if (blk_mq_is_shared_tags(hctx->flags))
238 		atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
239 	else
240 		atomic_sub(val, &hctx->nr_active);
241 }
242 
243 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
244 {
245 	__blk_mq_sub_active_requests(hctx, 1);
246 }
247 
248 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
249 {
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);
253 }
254 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
255 					   struct request *rq)
256 {
257 	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
258 	rq->tag = BLK_MQ_NO_TAG;
259 
260 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
261 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
262 		__blk_mq_dec_active_requests(hctx);
263 	}
264 }
265 
266 static inline void blk_mq_put_driver_tag(struct request *rq)
267 {
268 	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
269 		return;
270 
271 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
272 }
273 
274 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq);
275 
276 static inline bool blk_mq_get_driver_tag(struct request *rq)
277 {
278 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
279 
280 	if (rq->tag != BLK_MQ_NO_TAG &&
281 	    !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
282 		hctx->tags->rqs[rq->tag] = rq;
283 		return true;
284 	}
285 
286 	return __blk_mq_get_driver_tag(hctx, rq);
287 }
288 
289 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
290 {
291 	int cpu;
292 
293 	for_each_possible_cpu(cpu)
294 		qmap->mq_map[cpu] = 0;
295 }
296 
297 /*
298  * blk_mq_plug() - Get caller context plug
299  * @q: request queue
300  * @bio : the bio being submitted by the caller context
301  *
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.
312  *
313  * Return current->plug if the bio can be plugged and NULL otherwise
314  */
315 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
316 					   struct bio *bio)
317 {
318 	/*
319 	 * For regular block devices or read operations, use the context plug
320 	 * which may be NULL if blk_start_plug() was not executed.
321 	 */
322 	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
323 		return current->plug;
324 
325 	/* Zoned block device write operation case: do not plug the BIO */
326 	return NULL;
327 }
328 
329 /* Free all requests on the list */
330 static inline void blk_mq_free_requests(struct list_head *list)
331 {
332 	while (!list_empty(list)) {
333 		struct request *rq = list_entry_rq(list->next);
334 
335 		list_del_init(&rq->queuelist);
336 		blk_mq_free_request(rq);
337 	}
338 }
339 
340 /*
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.
343  */
344 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
345 				  struct sbitmap_queue *bt)
346 {
347 	unsigned int depth, users;
348 
349 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
350 		return true;
351 
352 	/*
353 	 * Don't try dividing an ant
354 	 */
355 	if (bt->sb.depth == 1)
356 		return true;
357 
358 	if (blk_mq_is_shared_tags(hctx->flags)) {
359 		struct request_queue *q = hctx->queue;
360 
361 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
362 			return true;
363 	} else {
364 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
365 			return true;
366 	}
367 
368 	users = atomic_read(&hctx->tags->active_queues);
369 
370 	if (!users)
371 		return true;
372 
373 	/*
374 	 * Allow at least some tags
375 	 */
376 	depth = max((bt->sb.depth + users - 1) / users, 4U);
377 	return __blk_mq_active_requests(hctx) < depth;
378 }
379 
380 
381 #endif
382