blk-mq: Add blk_mq_alloc_map_and_rqs()
[linux-block.git] / block / blk-mq.h
CommitLineData
b2441318 1/* SPDX-License-Identifier: GPL-2.0 */
320ae51f
JA
2#ifndef INT_BLK_MQ_H
3#define INT_BLK_MQ_H
4
cf43e6be 5#include "blk-stat.h"
244c65a3 6#include "blk-mq-tag.h"
cf43e6be 7
24d2f903
CH
8struct blk_mq_tag_set;
9
1db4909e
ML
10struct blk_mq_ctxs {
11 struct kobject kobj;
12 struct blk_mq_ctx __percpu *queue_ctx;
13};
14
fe644072
LW
15/**
16 * struct blk_mq_ctx - State for a software queue facing the submitting CPUs
17 */
320ae51f
JA
18struct blk_mq_ctx {
19 struct {
20 spinlock_t lock;
c16d6b5a
ML
21 struct list_head rq_lists[HCTX_MAX_TYPES];
22 } ____cacheline_aligned_in_smp;
320ae51f
JA
23
24 unsigned int cpu;
f31967f0 25 unsigned short index_hw[HCTX_MAX_TYPES];
8ccdf4a3 26 struct blk_mq_hw_ctx *hctxs[HCTX_MAX_TYPES];
320ae51f
JA
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;
1db4909e 36 struct blk_mq_ctxs *ctxs;
320ae51f 37 struct kobject kobj;
4bb659b1 38} ____cacheline_aligned_in_smp;
320ae51f 39
c7e2d94b 40void blk_mq_exit_queue(struct request_queue *q);
e3a2b3f9 41int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
aed3ea94 42void blk_mq_wake_waiters(struct request_queue *q);
1fd40b5e
ML
43bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
44 unsigned int);
e6c98712
BVA
45void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
46 bool kick_requeue_list);
2c3ad667 47void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
b347689f
ML
48struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
49 struct blk_mq_ctx *start);
2e315dc0 50void blk_mq_put_rq_ref(struct request *rq);
2c3ad667
JA
51
52/*
53 * Internal helpers for allocating/freeing the request map
54 */
cc71a6f4
JA
55void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
56 unsigned int hctx_idx);
1c0706a7 57void blk_mq_free_rq_map(struct blk_mq_tags *tags, unsigned int flags);
63064be1
JG
58struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
59 unsigned int hctx_idx, unsigned int depth);
2c3ad667
JA
60
61/*
62 * Internal helpers for request insertion into sw queues
63 */
64void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
65 bool at_head);
01e99aec
ML
66void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
67 bool run_queue);
bd166ef1
JA
68void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
69 struct list_head *list);
320ae51f 70
fd9c40f6
BVA
71/* Used by blk_insert_cloned_request() to issue request directly */
72blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last);
6ce3dd6e
ML
73void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
74 struct list_head *list);
396eaf21 75
320ae51f
JA
76/*
77 * CPU -> queue mappings
78 */
ed76e329 79extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
320ae51f 80
b3c661b1
JA
81/*
82 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
83 * @q: request queue
e20ba6e1 84 * @type: the hctx type index
b3c661b1
JA
85 * @cpu: CPU
86 */
87static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
e20ba6e1 88 enum hctx_type type,
b3c661b1 89 unsigned int cpu)
7d7e0f90 90{
e20ba6e1 91 return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
7d7e0f90
CH
92}
93
b3c661b1
JA
94/*
95 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
96 * @q: request queue
97 * @flags: request command flags
d220a214 98 * @ctx: software queue cpu ctx
b3c661b1
JA
99 */
100static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
101 unsigned int flags,
8ccdf4a3 102 struct blk_mq_ctx *ctx)
ff2c5660 103{
e20ba6e1
CH
104 enum hctx_type type = HCTX_TYPE_DEFAULT;
105
bb94aea1
JW
106 /*
107 * The caller ensure that if REQ_HIPRI, poll must be enabled.
108 */
109 if (flags & REQ_HIPRI)
e20ba6e1 110 type = HCTX_TYPE_POLL;
bb94aea1 111 else if ((flags & REQ_OP_MASK) == REQ_OP_READ)
e20ba6e1 112 type = HCTX_TYPE_READ;
5aceaeb2 113
8ccdf4a3 114 return ctx->hctxs[type];
ff2c5660
JA
115}
116
67aec14c
JA
117/*
118 * sysfs helpers
119 */
737f98cf 120extern void blk_mq_sysfs_init(struct request_queue *q);
7ea5fe31 121extern void blk_mq_sysfs_deinit(struct request_queue *q);
2d0364c8 122extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
67aec14c
JA
123extern int blk_mq_sysfs_register(struct request_queue *q);
124extern void blk_mq_sysfs_unregister(struct request_queue *q);
868f2f0b 125extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
67aec14c 126
e09aae7e
ML
127void blk_mq_release(struct request_queue *q);
128
1aecfe48
ML
129static 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 */
141static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
142{
c05f4220 143 return __blk_mq_get_ctx(q, raw_smp_processor_id());
1aecfe48
ML
144}
145
cb96a42c
ML
146struct blk_mq_alloc_data {
147 /* input parameter */
148 struct request_queue *q;
9a95e4ef 149 blk_mq_req_flags_t flags;
229a9287 150 unsigned int shallow_depth;
f9afca4d 151 unsigned int cmd_flags;
cb96a42c
ML
152
153 /* input & output parameter */
154 struct blk_mq_ctx *ctx;
155 struct blk_mq_hw_ctx *hctx;
156};
157
32bc15af
JG
158static inline bool blk_mq_is_sbitmap_shared(unsigned int flags)
159{
160 return flags & BLK_MQ_F_TAG_HCTX_SHARED;
161}
162
4941115b
JA
163static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
164{
42fdc5e4 165 if (data->q->elevator)
bd166ef1
JA
166 return data->hctx->sched_tags;
167
4941115b
JA
168 return data->hctx->tags;
169}
170
5d1b25c1
BVA
171static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
172{
173 return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
174}
175
19c66e59
ML
176static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
177{
178 return hctx->nr_ctx && hctx->tags;
179}
180
8446fe92
CH
181unsigned int blk_mq_in_flight(struct request_queue *q,
182 struct block_device *part);
183void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
184 unsigned int inflight[2]);
f299b7c7 185
2a5a24aa
ML
186static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
187 int budget_token)
de148297 188{
de148297 189 if (q->mq_ops->put_budget)
2a5a24aa 190 q->mq_ops->put_budget(q, budget_token);
de148297
ML
191}
192
2a5a24aa 193static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
de148297 194{
de148297 195 if (q->mq_ops->get_budget)
65c76369 196 return q->mq_ops->get_budget(q);
2a5a24aa
ML
197 return 0;
198}
199
200static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
201{
202 if (token < 0)
203 return;
204
205 if (rq->q->mq_ops->set_rq_budget_token)
206 rq->q->mq_ops->set_rq_budget_token(rq, token);
207}
208
209static inline int blk_mq_get_rq_budget_token(struct request *rq)
210{
211 if (rq->q->mq_ops->get_rq_budget_token)
212 return rq->q->mq_ops->get_rq_budget_token(rq);
213 return -1;
de148297
ML
214}
215
bccf5e26
JG
216static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
217{
218 if (blk_mq_is_sbitmap_shared(hctx->flags))
219 atomic_inc(&hctx->queue->nr_active_requests_shared_sbitmap);
220 else
221 atomic_inc(&hctx->nr_active);
222}
223
224static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
225{
226 if (blk_mq_is_sbitmap_shared(hctx->flags))
227 atomic_dec(&hctx->queue->nr_active_requests_shared_sbitmap);
228 else
229 atomic_dec(&hctx->nr_active);
230}
231
232static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
233{
234 if (blk_mq_is_sbitmap_shared(hctx->flags))
235 return atomic_read(&hctx->queue->nr_active_requests_shared_sbitmap);
236 return atomic_read(&hctx->nr_active);
237}
4e2f62e5
JA
238static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
239 struct request *rq)
240{
241 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
242 rq->tag = BLK_MQ_NO_TAG;
243
244 if (rq->rq_flags & RQF_MQ_INFLIGHT) {
245 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
bccf5e26 246 __blk_mq_dec_active_requests(hctx);
4e2f62e5
JA
247 }
248}
249
250static inline void blk_mq_put_driver_tag(struct request *rq)
251{
252 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
253 return;
254
255 __blk_mq_put_driver_tag(rq->mq_hctx, rq);
256}
257
61347154
JK
258bool blk_mq_get_driver_tag(struct request *rq);
259
ed76e329 260static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
0da73d00
MI
261{
262 int cpu;
263
264 for_each_possible_cpu(cpu)
ed76e329 265 qmap->mq_map[cpu] = 0;
0da73d00
MI
266}
267
b49773e7
DLM
268/*
269 * blk_mq_plug() - Get caller context plug
270 * @q: request queue
271 * @bio : the bio being submitted by the caller context
272 *
273 * Plugging, by design, may delay the insertion of BIOs into the elevator in
274 * order to increase BIO merging opportunities. This however can cause BIO
275 * insertion order to change from the order in which submit_bio() is being
276 * executed in the case of multiple contexts concurrently issuing BIOs to a
277 * device, even if these context are synchronized to tightly control BIO issuing
278 * order. While this is not a problem with regular block devices, this ordering
279 * change can cause write BIO failures with zoned block devices as these
280 * require sequential write patterns to zones. Prevent this from happening by
281 * ignoring the plug state of a BIO issuing context if the target request queue
282 * is for a zoned block device and the BIO to plug is a write operation.
283 *
284 * Return current->plug if the bio can be plugged and NULL otherwise
285 */
286static inline struct blk_plug *blk_mq_plug(struct request_queue *q,
287 struct bio *bio)
288{
289 /*
290 * For regular block devices or read operations, use the context plug
291 * which may be NULL if blk_start_plug() was not executed.
292 */
293 if (!blk_queue_is_zoned(q) || !op_is_write(bio_op(bio)))
294 return current->plug;
295
296 /* Zoned block device write operation case: do not plug the BIO */
297 return NULL;
298}
299
fd2ef39c
JK
300/* Free all requests on the list */
301static inline void blk_mq_free_requests(struct list_head *list)
302{
303 while (!list_empty(list)) {
304 struct request *rq = list_entry_rq(list->next);
305
306 list_del_init(&rq->queuelist);
307 blk_mq_free_request(rq);
308 }
309}
310
a0235d23
JG
311/*
312 * For shared tag users, we track the number of currently active users
313 * and attempt to provide a fair share of the tag depth for each of them.
314 */
315static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
316 struct sbitmap_queue *bt)
317{
318 unsigned int depth, users;
319
320 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
321 return true;
a0235d23
JG
322
323 /*
324 * Don't try dividing an ant
325 */
326 if (bt->sb.depth == 1)
327 return true;
328
f1b49fdc
JG
329 if (blk_mq_is_sbitmap_shared(hctx->flags)) {
330 struct request_queue *q = hctx->queue;
331 struct blk_mq_tag_set *set = q->tag_set;
332
2569063c 333 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
f1b49fdc
JG
334 return true;
335 users = atomic_read(&set->active_queues_shared_sbitmap);
336 } else {
337 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
338 return true;
339 users = atomic_read(&hctx->tags->active_queues);
340 }
341
a0235d23
JG
342 if (!users)
343 return true;
344
345 /*
346 * Allow at least some tags
347 */
348 depth = max((bt->sb.depth + users - 1) / users, 4U);
bccf5e26 349 return __blk_mq_active_requests(hctx) < depth;
a0235d23
JG
350}
351
352
320ae51f 353#endif