xref: /openbmc/linux/block/blk-mq.h (revision d9d6b822)
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 
51 /*
52  * Internal helpers for allocating/freeing the request map
53  */
54 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
55 		     unsigned int hctx_idx);
56 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags);
57 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
58 					unsigned int hctx_idx,
59 					unsigned int nr_tags,
60 					unsigned int reserved_tags,
61 					unsigned int flags);
62 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
63 		     unsigned int hctx_idx, unsigned int depth);
64 
65 /*
66  * Internal helpers for request insertion into sw queues
67  */
68 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
69 				bool at_head);
70 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
71 				  bool run_queue);
72 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
73 				struct list_head *list);
74 
75 /* Used by blk_insert_cloned_request() to issue request directly */
76 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
77 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
78 				    struct list_head *list);
79 
80 /*
81  * CPU -> queue mappings
82  */
83 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
84 
85 /*
86  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
87  * @q: request queue
88  * @type: the hctx type index
89  * @cpu: CPU
90  */
91 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
92 							  enum hctx_type type,
93 							  unsigned int cpu)
94 {
95 	return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
96 }
97 
98 /*
99  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
100  * @q: request queue
101  * @flags: request command flags
102  * @ctx: software queue cpu ctx
103  */
104 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
105 						     unsigned int flags,
106 						     struct blk_mq_ctx *ctx)
107 {
108 	enum hctx_type type = HCTX_TYPE_DEFAULT;
109 
110 	/*
111 	 * The caller ensure that if REQ_HIPRI, poll must be enabled.
112 	 */
113 	if (flags & REQ_HIPRI)
114 		type = HCTX_TYPE_POLL;
115 	else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
116 		type = HCTX_TYPE_READ;
117 
118 	return ctx->hctxs[type];
119 }
120 
121 /*
122  * sysfs helpers
123  */
124 extern void blk_mq_sysfs_init(struct request_queue *q);
125 extern void blk_mq_sysfs_deinit(struct request_queue *q);
126 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
127 extern int blk_mq_sysfs_register(struct request_queue *q);
128 extern void blk_mq_sysfs_unregister(struct request_queue *q);
129 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
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 
157 	/* input & output parameter */
158 	struct blk_mq_ctx *ctx;
159 	struct blk_mq_hw_ctx *hctx;
160 };
161 
162 static inline bool blk_mq_is_sbitmap_shared(unsigned int flags)
163 {
164 	return flags & BLK_MQ_F_TAG_HCTX_SHARED;
165 }
166 
167 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
168 {
169 	if (data->q->elevator)
170 		return data->hctx->sched_tags;
171 
172 	return data->hctx->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_sbitmap_shared(hctx->flags))
223 		atomic_inc(&hctx->queue->nr_active_requests_shared_sbitmap);
224 	else
225 		atomic_inc(&hctx->nr_active);
226 }
227 
228 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
229 {
230 	if (blk_mq_is_sbitmap_shared(hctx->flags))
231 		atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap);
232 	else
233 		atomic_dec(&hctx->nr_active);
234 }
235 
236 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
237 {
238 	if (blk_mq_is_sbitmap_shared(hctx->flags))
239 		return atomic_read(&hctx->queue->nr_active_requests_shared_sbitmap);
240 	return atomic_read(&hctx->nr_active);
241 }
242 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
243 					   struct request *rq)
244 {
245 	blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
246 	rq->tag = BLK_MQ_NO_TAG;
247 
248 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
249 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
250 		__blk_mq_dec_active_requests(hctx);
251 	}
252 }
253 
254 static inline void blk_mq_put_driver_tag(struct request *rq)
255 {
256 	if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
257 		return;
258 
259 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
260 }
261 
262 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
263 {
264 	int cpu;
265 
266 	for_each_possible_cpu(cpu)
267 		qmap->mq_map[cpu] = 0;
268 }
269 
270 /*
271  * blk_mq_plug() - Get caller context plug
272  * @q: request queue
273  * @bio : the bio being submitted by the caller context
274  *
275  * Plugging, by design, may delay the insertion of BIOs into the elevator in
276  * order to increase BIO merging opportunities. This however can cause BIO
277  * insertion order to change from the order in which submit_bio() is being
278  * executed in the case of multiple contexts concurrently issuing BIOs to a
279  * device, even if these context are synchronized to tightly control BIO issuing
280  * order. While this is not a problem with regular block devices, this ordering
281  * change can cause write BIO failures with zoned block devices as these
282  * require sequential write patterns to zones. Prevent this from happening by
283  * ignoring the plug state of a BIO issuing context if the target request queue
284  * is for a zoned block device and the BIO to plug is a write operation.
285  *
286  * Return current->plug if the bio can be plugged and NULL otherwise
287  */
288 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
289 					   struct bio *bio)
290 {
291 	/*
292 	 * For regular block devices or read operations, use the context plug
293 	 * which may be NULL if blk_start_plug() was not executed.
294 	 */
295 	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
296 		return current->plug;
297 
298 	/* Zoned block device write operation case: do not plug the BIO */
299 	return NULL;
300 }
301 
302 /*
303  * For shared tag users, we track the number of currently active users
304  * and attempt to provide a fair share of the tag depth for each of them.
305  */
306 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
307 				  struct sbitmap_queue *bt)
308 {
309 	unsigned int depth, users;
310 
311 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
312 		return true;
313 
314 	/*
315 	 * Don't try dividing an ant
316 	 */
317 	if (bt->sb.depth == 1)
318 		return true;
319 
320 	if (blk_mq_is_sbitmap_shared(hctx->flags)) {
321 		struct request_queue *q = hctx->queue;
322 		struct blk_mq_tag_set *set = q->tag_set;
323 
324 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
325 			return true;
326 		users = atomic_read(&set->active_queues_shared_sbitmap);
327 	} else {
328 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
329 			return true;
330 		users = atomic_read(&hctx->tags->active_queues);
331 	}
332 
333 	if (!users)
334 		return true;
335 
336 	/*
337 	 * Allow at least some tags
338 	 */
339 	depth = max((bt->sb.depth + users - 1) / users, 4U);
340 	return __blk_mq_active_requests(hctx) < depth;
341 }
342 
343 
344 #endif
345