xref: /openbmc/linux/block/blk-mq.h (revision ddc141e5)
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_ctx {
11 	struct {
12 		spinlock_t		lock;
13 		struct list_head	rq_list;
14 	}  ____cacheline_aligned_in_smp;
15 
16 	unsigned int		cpu;
17 	unsigned int		index_hw;
18 
19 	/* incremented at dispatch time */
20 	unsigned long		rq_dispatched[2];
21 	unsigned long		rq_merged;
22 
23 	/* incremented at completion time */
24 	unsigned long		____cacheline_aligned_in_smp rq_completed[2];
25 
26 	struct request_queue	*queue;
27 	struct kobject		kobj;
28 } ____cacheline_aligned_in_smp;
29 
30 /*
31  * Bits for request->gstate.  The lower two bits carry MQ_RQ_* state value
32  * and the upper bits the generation number.
33  */
34 enum mq_rq_state {
35 	MQ_RQ_IDLE		= 0,
36 	MQ_RQ_IN_FLIGHT		= 1,
37 	MQ_RQ_COMPLETE		= 2,
38 
39 	MQ_RQ_STATE_BITS	= 2,
40 	MQ_RQ_STATE_MASK	= (1 << MQ_RQ_STATE_BITS) - 1,
41 	MQ_RQ_GEN_INC		= 1 << MQ_RQ_STATE_BITS,
42 };
43 
44 void blk_mq_freeze_queue(struct request_queue *q);
45 void blk_mq_free_queue(struct request_queue *q);
46 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
47 void blk_mq_wake_waiters(struct request_queue *q);
48 bool blk_mq_dispatch_rq_list(struct request_queue *, struct list_head *, bool);
49 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
50 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
51 				bool wait);
52 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
53 					struct blk_mq_ctx *start);
54 
55 /*
56  * Internal helpers for allocating/freeing the request map
57  */
58 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
59 		     unsigned int hctx_idx);
60 void blk_mq_free_rq_map(struct blk_mq_tags *tags);
61 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
62 					unsigned int hctx_idx,
63 					unsigned int nr_tags,
64 					unsigned int reserved_tags);
65 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
66 		     unsigned int hctx_idx, unsigned int depth);
67 
68 /*
69  * Internal helpers for request insertion into sw queues
70  */
71 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
72 				bool at_head);
73 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue);
74 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
75 				struct list_head *list);
76 
77 /* Used by blk_insert_cloned_request() to issue request directly */
78 blk_status_t blk_mq_request_issue_directly(struct request *rq);
79 
80 /*
81  * CPU -> queue mappings
82  */
83 extern int blk_mq_hw_queue_to_node(unsigned int *map, unsigned int);
84 
85 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
86 		int cpu)
87 {
88 	return q->queue_hw_ctx[q->mq_map[cpu]];
89 }
90 
91 /*
92  * sysfs helpers
93  */
94 extern void blk_mq_sysfs_init(struct request_queue *q);
95 extern void blk_mq_sysfs_deinit(struct request_queue *q);
96 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
97 extern int blk_mq_sysfs_register(struct request_queue *q);
98 extern void blk_mq_sysfs_unregister(struct request_queue *q);
99 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
100 
101 void blk_mq_release(struct request_queue *q);
102 
103 /**
104  * blk_mq_rq_state() - read the current MQ_RQ_* state of a request
105  * @rq: target request.
106  */
107 static inline int blk_mq_rq_state(struct request *rq)
108 {
109 	return READ_ONCE(rq->gstate) & MQ_RQ_STATE_MASK;
110 }
111 
112 /**
113  * blk_mq_rq_update_state() - set the current MQ_RQ_* state of a request
114  * @rq: target request.
115  * @state: new state to set.
116  *
117  * Set @rq's state to @state.  The caller is responsible for ensuring that
118  * there are no other updaters.  A request can transition into IN_FLIGHT
119  * only from IDLE and doing so increments the generation number.
120  */
121 static inline void blk_mq_rq_update_state(struct request *rq,
122 					  enum mq_rq_state state)
123 {
124 	u64 old_val = READ_ONCE(rq->gstate);
125 	u64 new_val = (old_val & ~MQ_RQ_STATE_MASK) | state;
126 
127 	if (state == MQ_RQ_IN_FLIGHT) {
128 		WARN_ON_ONCE((old_val & MQ_RQ_STATE_MASK) != MQ_RQ_IDLE);
129 		new_val += MQ_RQ_GEN_INC;
130 	}
131 
132 	/* avoid exposing interim values */
133 	WRITE_ONCE(rq->gstate, new_val);
134 }
135 
136 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
137 					   unsigned int cpu)
138 {
139 	return per_cpu_ptr(q->queue_ctx, cpu);
140 }
141 
142 /*
143  * This assumes per-cpu software queueing queues. They could be per-node
144  * as well, for instance. For now this is hardcoded as-is. Note that we don't
145  * care about preemption, since we know the ctx's are persistent. This does
146  * mean that we can't rely on ctx always matching the currently running CPU.
147  */
148 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
149 {
150 	return __blk_mq_get_ctx(q, get_cpu());
151 }
152 
153 static inline void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
154 {
155 	put_cpu();
156 }
157 
158 struct blk_mq_alloc_data {
159 	/* input parameter */
160 	struct request_queue *q;
161 	blk_mq_req_flags_t flags;
162 	unsigned int shallow_depth;
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 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
188 			unsigned int inflight[2]);
189 
190 static inline void blk_mq_put_dispatch_budget(struct blk_mq_hw_ctx *hctx)
191 {
192 	struct request_queue *q = hctx->queue;
193 
194 	if (q->mq_ops->put_budget)
195 		q->mq_ops->put_budget(hctx);
196 }
197 
198 static inline bool blk_mq_get_dispatch_budget(struct blk_mq_hw_ctx *hctx)
199 {
200 	struct request_queue *q = hctx->queue;
201 
202 	if (q->mq_ops->get_budget)
203 		return q->mq_ops->get_budget(hctx);
204 	return true;
205 }
206 
207 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
208 					   struct request *rq)
209 {
210 	blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
211 	rq->tag = -1;
212 
213 	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
214 		rq->rq_flags &= ~RQF_MQ_INFLIGHT;
215 		atomic_dec(&hctx->nr_active);
216 	}
217 }
218 
219 static inline void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
220 				       struct request *rq)
221 {
222 	if (rq->tag == -1 || rq->internal_tag == -1)
223 		return;
224 
225 	__blk_mq_put_driver_tag(hctx, rq);
226 }
227 
228 static inline void blk_mq_put_driver_tag(struct request *rq)
229 {
230 	struct blk_mq_hw_ctx *hctx;
231 
232 	if (rq->tag == -1 || rq->internal_tag == -1)
233 		return;
234 
235 	hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
236 	__blk_mq_put_driver_tag(hctx, rq);
237 }
238 
239 #endif
240