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
39 ANDROID_OEM_DATA_ARRAY(1, 2);
40 } ____cacheline_aligned_in_smp;
41
42 void blk_mq_exit_queue(struct request_queue *q);
43 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
44 void blk_mq_wake_waiters(struct request_queue *q);
45 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
46 unsigned int);
47 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
48 bool kick_requeue_list);
49 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
50 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *start);
52 void blk_mq_put_rq_ref(struct request *rq);
53
54 /*
55 * Internal helpers for allocating/freeing the request map
56 */
57 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
58 unsigned int hctx_idx);
59 void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags);
60 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
61 unsigned int hctx_idx,
62 unsigned int nr_tags,
63 unsigned int reserved_tags,
64 unsigned int flags);
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 at_head,
74 bool run_queue);
75 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
76 struct list_head *list);
77
78 /* Used by blk_insert_cloned_request() to issue request directly */
79 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
80 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
81 struct list_head *list);
82
83 /*
84 * CPU -> queue mappings
85 */
86 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
87
88 /*
89 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
90 * @q: request queue
91 * @type: the hctx type index
92 * @cpu: CPU
93 */
blk_mq_map_queue_type(struct request_queue * q,enum hctx_type type,unsigned int cpu)94 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
95 enum hctx_type type,
96 unsigned int cpu)
97 {
98 return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
99 }
100
101 /*
102 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
103 * @q: request queue
104 * @flags: request command flags
105 * @cpu: cpu ctx
106 */
blk_mq_map_queue(struct request_queue * q,unsigned int flags,struct blk_mq_ctx * ctx)107 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
108 unsigned int flags,
109 struct blk_mq_ctx *ctx)
110 {
111 enum hctx_type type = HCTX_TYPE_DEFAULT;
112
113 /*
114 * The caller ensure that if REQ_HIPRI, poll must be enabled.
115 */
116 if (flags & REQ_HIPRI)
117 type = HCTX_TYPE_POLL;
118 else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
119 type = HCTX_TYPE_READ;
120
121 return ctx->hctxs[type];
122 }
123
124 /*
125 * sysfs helpers
126 */
127 extern void blk_mq_sysfs_init(struct request_queue *q);
128 extern void blk_mq_sysfs_deinit(struct request_queue *q);
129 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
130 extern int blk_mq_sysfs_register(struct request_queue *q);
131 extern void blk_mq_sysfs_unregister(struct request_queue *q);
132 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
133
134 void blk_mq_release(struct request_queue *q);
135
__blk_mq_get_ctx(struct request_queue * q,unsigned int cpu)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 */
blk_mq_get_ctx(struct request_queue * q)148 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
149 {
150 return __blk_mq_get_ctx(q, raw_smp_processor_id());
151 }
152
153 struct blk_mq_alloc_data {
154 /* input parameter */
155 struct request_queue *q;
156 blk_mq_req_flags_t flags;
157 unsigned int shallow_depth;
158 unsigned int cmd_flags;
159
160 /* input & output parameter */
161 struct blk_mq_ctx *ctx;
162 struct blk_mq_hw_ctx *hctx;
163 };
164
blk_mq_is_sbitmap_shared(unsigned int flags)165 static inline bool blk_mq_is_sbitmap_shared(unsigned int flags)
166 {
167 return flags & BLK_MQ_F_TAG_HCTX_SHARED;
168 }
169
blk_mq_tags_from_data(struct blk_mq_alloc_data * data)170 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
171 {
172 if (data->q->elevator)
173 return data->hctx->sched_tags;
174
175 return data->hctx->tags;
176 }
177
blk_mq_hctx_stopped(struct blk_mq_hw_ctx * hctx)178 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
179 {
180 return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
181 }
182
blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx * hctx)183 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
184 {
185 return hctx->nr_ctx && hctx->tags;
186 }
187
188 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part);
189 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
190 unsigned int inflight[2]);
191
blk_mq_put_dispatch_budget(struct request_queue * q)192 static inline void blk_mq_put_dispatch_budget(struct request_queue *q)
193 {
194 if (q->mq_ops->put_budget)
195 q->mq_ops->put_budget(q);
196 }
197
blk_mq_get_dispatch_budget(struct request_queue * q)198 static inline bool blk_mq_get_dispatch_budget(struct request_queue *q)
199 {
200 if (q->mq_ops->get_budget)
201 return q->mq_ops->get_budget(q);
202 return true;
203 }
204
__blk_mq_inc_active_requests(struct blk_mq_hw_ctx * hctx)205 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
206 {
207 if (blk_mq_is_sbitmap_shared(hctx->flags))
208 atomic_inc(&hctx->queue->nr_active_requests_shared_sbitmap);
209 else
210 atomic_inc(&hctx->nr_active);
211 }
212
__blk_mq_dec_active_requests(struct blk_mq_hw_ctx * hctx)213 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
214 {
215 if (blk_mq_is_sbitmap_shared(hctx->flags))
216 atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap);
217 else
218 atomic_dec(&hctx->nr_active);
219 }
220
__blk_mq_active_requests(struct blk_mq_hw_ctx * hctx)221 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
222 {
223 if (blk_mq_is_sbitmap_shared(hctx->flags))
224 return atomic_read(&hctx->queue->nr_active_requests_shared_sbitmap);
225 return atomic_read(&hctx->nr_active);
226 }
__blk_mq_put_driver_tag(struct blk_mq_hw_ctx * hctx,struct request * rq)227 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
228 struct request *rq)
229 {
230 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
231 rq->tag = BLK_MQ_NO_TAG;
232
233 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
234 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
235 __blk_mq_dec_active_requests(hctx);
236 }
237 }
238
blk_mq_put_driver_tag(struct request * rq)239 static inline void blk_mq_put_driver_tag(struct request *rq)
240 {
241 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
242 return;
243
244 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
245 }
246
blk_mq_clear_mq_map(struct blk_mq_queue_map * qmap)247 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
248 {
249 int cpu;
250
251 for_each_possible_cpu(cpu)
252 qmap->mq_map[cpu] = 0;
253 }
254
255 /*
256 * blk_mq_plug() - Get caller context plug
257 * @q: request queue
258 * @bio : the bio being submitted by the caller context
259 *
260 * Plugging, by design, may delay the insertion of BIOs into the elevator in
261 * order to increase BIO merging opportunities. This however can cause BIO
262 * insertion order to change from the order in which submit_bio() is being
263 * executed in the case of multiple contexts concurrently issuing BIOs to a
264 * device, even if these context are synchronized to tightly control BIO issuing
265 * order. While this is not a problem with regular block devices, this ordering
266 * change can cause write BIO failures with zoned block devices as these
267 * require sequential write patterns to zones. Prevent this from happening by
268 * ignoring the plug state of a BIO issuing context if the target request queue
269 * is for a zoned block device and the BIO to plug is a write operation.
270 *
271 * Return current->plug if the bio can be plugged and NULL otherwise
272 */
blk_mq_plug(struct request_queue * q,struct bio * bio)273 static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
274 struct bio *bio)
275 {
276 /*
277 * For regular block devices or read operations, use the context plug
278 * which may be NULL if blk_start_plug() was not executed.
279 */
280 if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
281 return current->plug;
282
283 /* Zoned block device write operation case: do not plug the BIO */
284 return NULL;
285 }
286
287 /*
288 * For shared tag users, we track the number of currently active users
289 * and attempt to provide a fair share of the tag depth for each of them.
290 */
hctx_may_queue(struct blk_mq_hw_ctx * hctx,struct sbitmap_queue * bt)291 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
292 struct sbitmap_queue *bt)
293 {
294 unsigned int depth, users;
295
296 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
297 return true;
298
299 /*
300 * Don't try dividing an ant
301 */
302 if (bt->sb.depth == 1)
303 return true;
304
305 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
306 struct request_queue *q = hctx->queue;
307 struct blk_mq_tag_set *set = q->tag_set;
308
309 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
310 return true;
311 users = atomic_read(&set->active_queues_shared_sbitmap);
312 } else {
313 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
314 return true;
315 users = atomic_read(&hctx->tags->active_queues);
316 }
317
318 if (!users)
319 return true;
320
321 /*
322 * Allow at least some tags
323 */
324 depth = max((bt->sb.depth + users - 1) / users, 4U);
325 return __blk_mq_active_requests(hctx) < depth;
326 }
327
328
329 #endif
330