xref: /openbmc/linux/block/blk-mq.h (revision a20eefae)
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 request_queue *, struct list_head *, bool);
44 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
45 				bool kick_requeue_list);
46 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
47 bool blk_mq_get_driver_tag(struct request *rq);
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);
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 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
62 		     unsigned int hctx_idx, unsigned int depth);
63 
64 /*
65  * Internal helpers for request insertion into sw queues
66  */
67 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
68 				bool at_head);
69 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue);
70 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
71 				struct list_head *list);
72 
73 /* Used by blk_insert_cloned_request() to issue request directly */
74 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
75 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
76 				    struct list_head *list);
77 
78 /*
79  * CPU -> queue mappings
80  */
81 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
82 
83 /*
84  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
85  * @q: request queue
86  * @type: the hctx type index
87  * @cpu: CPU
88  */
89 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
90 							  enum hctx_type type,
91 							  unsigned int cpu)
92 {
93 	return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
94 }
95 
96 /*
97  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
98  * @q: request queue
99  * @flags: request command flags
100  * @cpu: cpu ctx
101  */
102 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
103 						     unsigned int flags,
104 						     struct blk_mq_ctx *ctx)
105 {
106 	enum hctx_type type = HCTX_TYPE_DEFAULT;
107 
108 	/*
109 	 * The caller ensure that if REQ_HIPRI, poll must be enabled.
110 	 */
111 	if (flags & REQ_HIPRI)
112 		type = HCTX_TYPE_POLL;
113 	else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
114 		type = HCTX_TYPE_READ;
115 
116 	return ctx->hctxs[type];
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 
129 void blk_mq_release(struct request_queue *q);
130 
131 /**
132  * blk_mq_rq_state() - read the current MQ_RQ_* state of a request
133  * @rq: target request.
134  */
135 static inline enum mq_rq_state blk_mq_rq_state(struct request *rq)
136 {
137 	return READ_ONCE(rq->state);
138 }
139 
140 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
141 					   unsigned int cpu)
142 {
143 	return per_cpu_ptr(q->queue_ctx, cpu);
144 }
145 
146 /*
147  * This assumes per-cpu software queueing queues. They could be per-node
148  * as well, for instance. For now this is hardcoded as-is. Note that we don't
149  * care about preemption, since we know the ctx's are persistent. This does
150  * mean that we can't rely on ctx always matching the currently running CPU.
151  */
152 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
153 {
154 	return __blk_mq_get_ctx(q, raw_smp_processor_id());
155 }
156 
157 struct blk_mq_alloc_data {
158 	/* input parameter */
159 	struct request_queue *q;
160 	blk_mq_req_flags_t flags;
161 	unsigned int shallow_depth;
162 	unsigned int cmd_flags;
163 
164 	/* input & output parameter */
165 	struct blk_mq_ctx *ctx;
166 	struct blk_mq_hw_ctx *hctx;
167 };
168 
169 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
170 {
171 	if (data->flags & BLK_MQ_REQ_INTERNAL)
172 		return data->hctx->sched_tags;
173 
174 	return data->hctx->tags;
175 }
176 
177 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
178 {
179 	return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
180 }
181 
182 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
183 {
184 	return hctx->nr_ctx && hctx->tags;
185 }
186 
187 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part);
188 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
189 			 unsigned int inflight[2]);
190 
191 static inline void blk_mq_put_dispatch_budget(struct blk_mq_hw_ctx *hctx)
192 {
193 	struct request_queue *q = hctx->queue;
194 
195 	if (q->mq_ops->put_budget)
196 		q->mq_ops->put_budget(hctx);
197 }
198 
199 static inline bool blk_mq_get_dispatch_budget(struct blk_mq_hw_ctx *hctx)
200 {
201 	struct request_queue *q = hctx->queue;
202 
203 	if (q->mq_ops->get_budget)
204 		return q->mq_ops->get_budget(hctx);
205 	return true;
206 }
207 
208 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
209 					   struct request *rq)
210 {
211 	blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
212 	rq->tag = -1;
213 
214 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
215 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
216 		atomic_dec(&hctx->nr_active);
217 	}
218 }
219 
220 static inline void blk_mq_put_driver_tag(struct request *rq)
221 {
222 	if (rq->tag == -1 || rq->internal_tag == -1)
223 		return;
224 
225 	__blk_mq_put_driver_tag(rq->mq_hctx, rq);
226 }
227 
228 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
229 {
230 	int cpu;
231 
232 	for_each_possible_cpu(cpu)
233 		qmap->mq_map[cpu] = 0;
234 }
235 
236 /*
237  * blk_mq_plug() - Get caller context plug
238  * @q: request queue
239  * @bio : the bio being submitted by the caller context
240  *
241  * Plugging, by design, may delay the insertion of BIOs into the elevator in
242  * order to increase BIO merging opportunities. This however can cause BIO
243  * insertion order to change from the order in which submit_bio() is being
244  * executed in the case of multiple contexts concurrently issuing BIOs to a
245  * device, even if these context are synchronized to tightly control BIO issuing
246  * order. While this is not a problem with regular block devices, this ordering
247  * change can cause write BIO failures with zoned block devices as these
248  * require sequential write patterns to zones. Prevent this from happening by
249  * ignoring the plug state of a BIO issuing context if the target request queue
250  * is for a zoned block device and the BIO to plug is a write operation.
251  *
252  * Return current->plug if the bio can be plugged and NULL otherwise
253  */
254 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
255 					   struct bio *bio)
256 {
257 	/*
258 	 * For regular block devices or read operations, use the context plug
259 	 * which may be NULL if blk_start_plug() was not executed.
260 	 */
261 	if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
262 		return current->plug;
263 
264 	/* Zoned block device write operation case: do not plug the BIO */
265 	return NULL;
266 }
267 
268 #endif
269