xref: /openbmc/linux/block/blk-mq.h (revision 48b71a9e)
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 /*
93  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
94  * @q: request queue
95  * @flags: request command flags
96  * @ctx: software queue cpu ctx
97  */
98 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
99 						     unsigned int flags,
100 						     struct blk_mq_ctx *ctx)
101 {
102 	enum hctx_type type = HCTX_TYPE_DEFAULT;
103 
104 	/*
105 	 * The caller ensure that if REQ_POLLED, poll must be enabled.
106 	 */
107 	if (flags & REQ_POLLED)
108 		type = HCTX_TYPE_POLL;
109 	else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
110 		type = HCTX_TYPE_READ;
111 
112 	return ctx->hctxs[type];
113 }
114 
115 /*
116  * sysfs helpers
117  */
118 extern void blk_mq_sysfs_init(struct request_queue *q);
119 extern void blk_mq_sysfs_deinit(struct request_queue *q);
120 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
121 extern int blk_mq_sysfs_register(struct request_queue *q);
122 extern void blk_mq_sysfs_unregister(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_release(struct request_queue *q);
128 
129 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
130 					   unsigned int cpu)
131 {
132 	return per_cpu_ptr(q->queue_ctx, cpu);
133 }
134 
135 /*
136  * This assumes per-cpu software queueing queues. They could be per-node
137  * as well, for instance. For now this is hardcoded as-is. Note that we don't
138  * care about preemption, since we know the ctx's are persistent. This does
139  * mean that we can't rely on ctx always matching the currently running CPU.
140  */
141 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
142 {
143 	return __blk_mq_get_ctx(q, raw_smp_processor_id());
144 }
145 
146 struct blk_mq_alloc_data {
147 	/* input parameter */
148 	struct request_queue *q;
149 	blk_mq_req_flags_t flags;
150 	unsigned int shallow_depth;
151 	unsigned int cmd_flags;
152 	unsigned int rq_flags;
153 
154 	/* allocate multiple requests/tags in one go */
155 	unsigned int nr_tags;
156 	struct request **cached_rq;
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_shared_tags(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->rq_flags & RQF_ELV))
171 		return data->hctx->tags;
172 	return data->hctx->sched_tags;
173 }
174 
175 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
176 {
177 	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
178 }
179 
180 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
181 {
182 	return hctx->nr_ctx && hctx->tags;
183 }
184 
185 unsigned int blk_mq_in_flight(struct request_queue *q,
186 		struct block_device *part);
187 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
188 		unsigned int inflight[2]);
189 
190 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
191 					      int budget_token)
192 {
193 	if (q->mq_ops->put_budget)
194 		q->mq_ops->put_budget(q, budget_token);
195 }
196 
197 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
198 {
199 	if (q->mq_ops->get_budget)
200 		return q->mq_ops->get_budget(q);
201 	return 0;
202 }
203 
204 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
205 {
206 	if (token < 0)
207 		return;
208 
209 	if (rq->q->mq_ops->set_rq_budget_token)
210 		rq->q->mq_ops->set_rq_budget_token(rq, token);
211 }
212 
213 static inline int blk_mq_get_rq_budget_token(struct request *rq)
214 {
215 	if (rq->q->mq_ops->get_rq_budget_token)
216 		return rq->q->mq_ops->get_rq_budget_token(rq);
217 	return -1;
218 }
219 
220 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
221 {
222 	if (blk_mq_is_shared_tags(hctx->flags))
223 		atomic_inc(&hctx->queue->nr_active_requests_shared_tags);
224 	else
225 		atomic_inc(&hctx->nr_active);
226 }
227 
228 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
229 		int val)
230 {
231 	if (blk_mq_is_shared_tags(hctx->flags))
232 		atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
233 	else
234 		atomic_sub(val, &hctx->nr_active);
235 }
236 
237 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
238 {
239 	__blk_mq_sub_active_requests(hctx, 1);
240 }
241 
242 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
243 {
244 	if (blk_mq_is_shared_tags(hctx->flags))
245 		return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
246 	return atomic_read(&hctx->nr_active);
247 }
248 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
249 					   struct request *rq)
250 {
251 	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
252 	rq->tag = BLK_MQ_NO_TAG;
253 
254 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
255 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
256 		__blk_mq_dec_active_requests(hctx);
257 	}
258 }
259 
260 static inline void blk_mq_put_driver_tag(struct request *rq)
261 {
262 	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
263 		return;
264 
265 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
266 }
267 
268 bool __blk_mq_get_driver_tag(struct blk_mq_hw_ctx *hctx, struct request *rq);
269 
270 static inline bool blk_mq_get_driver_tag(struct request *rq)
271 {
272 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
273 
274 	if (rq->tag != BLK_MQ_NO_TAG &&
275 	    !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) {
276 		hctx->tags->rqs[rq->tag] = rq;
277 		return true;
278 	}
279 
280 	return __blk_mq_get_driver_tag(hctx, rq);
281 }
282 
283 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
284 {
285 	int cpu;
286 
287 	for_each_possible_cpu(cpu)
288 		qmap->mq_map[cpu] = 0;
289 }
290 
291 /*
292  * blk_mq_plug() - Get caller context plug
293  * @q: request queue
294  * @bio : the bio being submitted by the caller context
295  *
296  * Plugging, by design, may delay the insertion of BIOs into the elevator in
297  * order to increase BIO merging opportunities. This however can cause BIO
298  * insertion order to change from the order in which submit_bio() is being
299  * executed in the case of multiple contexts concurrently issuing BIOs to a
300  * device, even if these context are synchronized to tightly control BIO issuing
301  * order. While this is not a problem with regular block devices, this ordering
302  * change can cause write BIO failures with zoned block devices as these
303  * require sequential write patterns to zones. Prevent this from happening by
304  * ignoring the plug state of a BIO issuing context if the target request queue
305  * is for a zoned block device and the BIO to plug is a write operation.
306  *
307  * Return current->plug if the bio can be plugged and NULL otherwise
308  */
309 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
310 					   struct bio *bio)
311 {
312 	/*
313 	 * For regular block devices or read operations, use the context plug
314 	 * which may be NULL if blk_start_plug() was not executed.
315 	 */
316 	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
317 		return current->plug;
318 
319 	/* Zoned block device write operation case: do not plug the BIO */
320 	return NULL;
321 }
322 
323 /* Free all requests on the list */
324 static inline void blk_mq_free_requests(struct list_head *list)
325 {
326 	while (!list_empty(list)) {
327 		struct request *rq = list_entry_rq(list->next);
328 
329 		list_del_init(&rq->queuelist);
330 		blk_mq_free_request(rq);
331 	}
332 }
333 
334 /*
335  * For shared tag users, we track the number of currently active users
336  * and attempt to provide a fair share of the tag depth for each of them.
337  */
338 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
339 				  struct sbitmap_queue *bt)
340 {
341 	unsigned int depth, users;
342 
343 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
344 		return true;
345 
346 	/*
347 	 * Don't try dividing an ant
348 	 */
349 	if (bt->sb.depth == 1)
350 		return true;
351 
352 	if (blk_mq_is_shared_tags(hctx->flags)) {
353 		struct request_queue *q = hctx->queue;
354 
355 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
356 			return true;
357 	} else {
358 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
359 			return true;
360 	}
361 
362 	users = atomic_read(&hctx->tags->active_queues);
363 
364 	if (!users)
365 		return true;
366 
367 	/*
368 	 * Allow at least some tags
369 	 */
370 	depth = max((bt->sb.depth + users - 1) / users, 4U);
371 	return __blk_mq_active_requests(hctx) < depth;
372 }
373 
374 
375 #endif
376