1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef BLK_MQ_H
3 #define BLK_MQ_H
4
5 #include <linux/blkdev.h>
6 #include <linux/sbitmap.h>
7 #include <linux/srcu.h>
8
9 struct blk_mq_tags;
10 struct blk_flush_queue;
11
12 /**
13 * struct blk_mq_hw_ctx - State for a hardware queue facing the hardware block device
14 */
15 struct blk_mq_hw_ctx {
16 struct {
17 spinlock_t lock;
18 struct list_head dispatch;
19 unsigned long state; /* BLK_MQ_S_* flags */
20 } ____cacheline_aligned_in_smp;
21
22 struct delayed_work run_work;
23 cpumask_var_t cpumask;
24 int next_cpu;
25 int next_cpu_batch;
26
27 unsigned long flags; /* BLK_MQ_F_* flags */
28
29 void *sched_data;
30 struct request_queue *queue;
31 struct blk_flush_queue *fq;
32
33 void *driver_data;
34
35 struct sbitmap ctx_map;
36
37 struct blk_mq_ctx *dispatch_from;
38 unsigned int dispatch_busy;
39
40 unsigned short type;
41 unsigned short nr_ctx;
42 struct blk_mq_ctx **ctxs;
43
44 spinlock_t dispatch_wait_lock;
45 wait_queue_entry_t dispatch_wait;
46 atomic_t wait_index;
47
48 struct blk_mq_tags *tags;
49 struct blk_mq_tags *sched_tags;
50
51 unsigned long queued;
52 unsigned long run;
53 #define BLK_MQ_MAX_DISPATCH_ORDER 7
54 unsigned long dispatched[BLK_MQ_MAX_DISPATCH_ORDER];
55
56 unsigned int numa_node;
57 unsigned int queue_num;
58
59 atomic_t nr_active;
60
61 struct hlist_node cpuhp_dead;
62 struct kobject kobj;
63
64 unsigned long poll_considered;
65 unsigned long poll_invoked;
66 unsigned long poll_success;
67
68 #ifdef CONFIG_BLK_DEBUG_FS
69 struct dentry *debugfs_dir;
70 struct dentry *sched_debugfs_dir;
71 #endif
72
73 struct list_head hctx_list;
74
75 /* Must be the last member - see also blk_mq_hw_ctx_size(). */
76 struct srcu_struct srcu[0];
77 };
78
79 struct blk_mq_queue_map {
80 unsigned int *mq_map;
81 unsigned int nr_queues;
82 unsigned int queue_offset;
83 };
84
85 enum hctx_type {
86 HCTX_TYPE_DEFAULT, /* all I/O not otherwise accounted for */
87 HCTX_TYPE_READ, /* just for READ I/O */
88 HCTX_TYPE_POLL, /* polled I/O of any kind */
89
90 HCTX_MAX_TYPES,
91 };
92
93 struct blk_mq_tag_set {
94 /*
95 * map[] holds ctx -> hctx mappings, one map exists for each type
96 * that the driver wishes to support. There are no restrictions
97 * on maps being of the same size, and it's perfectly legal to
98 * share maps between types.
99 */
100 struct blk_mq_queue_map map[HCTX_MAX_TYPES];
101 unsigned int nr_maps; /* nr entries in map[] */
102 const struct blk_mq_ops *ops;
103 unsigned int nr_hw_queues; /* nr hw queues across maps */
104 unsigned int queue_depth; /* max hw supported */
105 unsigned int reserved_tags;
106 unsigned int cmd_size; /* per-request extra data */
107 int numa_node;
108 unsigned int timeout;
109 unsigned int flags; /* BLK_MQ_F_* */
110 void *driver_data;
111
112 struct blk_mq_tags **tags;
113
114 struct mutex tag_list_lock;
115 struct list_head tag_list;
116 };
117
118 struct blk_mq_queue_data {
119 struct request *rq;
120 bool last;
121 };
122
123 typedef blk_status_t (queue_rq_fn)(struct blk_mq_hw_ctx *,
124 const struct blk_mq_queue_data *);
125 typedef void (commit_rqs_fn)(struct blk_mq_hw_ctx *);
126 typedef bool (get_budget_fn)(struct blk_mq_hw_ctx *);
127 typedef void (put_budget_fn)(struct blk_mq_hw_ctx *);
128 typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool);
129 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
130 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
131 typedef int (init_request_fn)(struct blk_mq_tag_set *set, struct request *,
132 unsigned int, unsigned int);
133 typedef void (exit_request_fn)(struct blk_mq_tag_set *set, struct request *,
134 unsigned int);
135
136 typedef bool (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *,
137 bool);
138 typedef bool (busy_tag_iter_fn)(struct request *, void *, bool);
139 typedef int (poll_fn)(struct blk_mq_hw_ctx *);
140 typedef int (map_queues_fn)(struct blk_mq_tag_set *set);
141 typedef bool (busy_fn)(struct request_queue *);
142 typedef void (complete_fn)(struct request *);
143 typedef void (cleanup_rq_fn)(struct request *);
144
145
146 struct blk_mq_ops {
147 /*
148 * Queue request
149 */
150 queue_rq_fn *queue_rq;
151
152 /*
153 * If a driver uses bd->last to judge when to submit requests to
154 * hardware, it must define this function. In case of errors that
155 * make us stop issuing further requests, this hook serves the
156 * purpose of kicking the hardware (which the last request otherwise
157 * would have done).
158 */
159 commit_rqs_fn *commit_rqs;
160
161 /*
162 * Reserve budget before queue request, once .queue_rq is
163 * run, it is driver's responsibility to release the
164 * reserved budget. Also we have to handle failure case
165 * of .get_budget for avoiding I/O deadlock.
166 */
167 get_budget_fn *get_budget;
168 put_budget_fn *put_budget;
169
170 /*
171 * Called on request timeout
172 */
173 timeout_fn *timeout;
174
175 /*
176 * Called to poll for completion of a specific tag.
177 */
178 poll_fn *poll;
179
180 complete_fn *complete;
181
182 /*
183 * Called when the block layer side of a hardware queue has been
184 * set up, allowing the driver to allocate/init matching structures.
185 * Ditto for exit/teardown.
186 */
187 init_hctx_fn *init_hctx;
188 exit_hctx_fn *exit_hctx;
189
190 /*
191 * Called for every command allocated by the block layer to allow
192 * the driver to set up driver specific data.
193 *
194 * Tag greater than or equal to queue_depth is for setting up
195 * flush request.
196 *
197 * Ditto for exit/teardown.
198 */
199 init_request_fn *init_request;
200 exit_request_fn *exit_request;
201 /* Called from inside blk_get_request() */
202 void (*initialize_rq_fn)(struct request *rq);
203
204 /*
205 * Called before freeing one request which isn't completed yet,
206 * and usually for freeing the driver private data
207 */
208 cleanup_rq_fn *cleanup_rq;
209
210 /*
211 * If set, returns whether or not this queue currently is busy
212 */
213 busy_fn *busy;
214
215 map_queues_fn *map_queues;
216
217 #ifdef CONFIG_BLK_DEBUG_FS
218 /*
219 * Used by the debugfs implementation to show driver-specific
220 * information about a request.
221 */
222 void (*show_rq)(struct seq_file *m, struct request *rq);
223 #endif
224 };
225
226 enum {
227 BLK_MQ_F_SHOULD_MERGE = 1 << 0,
228 BLK_MQ_F_TAG_SHARED = 1 << 1,
229 BLK_MQ_F_BLOCKING = 1 << 5,
230 BLK_MQ_F_NO_SCHED = 1 << 6,
231 BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
232 BLK_MQ_F_ALLOC_POLICY_BITS = 1,
233
234 BLK_MQ_S_STOPPED = 0,
235 BLK_MQ_S_TAG_ACTIVE = 1,
236 BLK_MQ_S_SCHED_RESTART = 2,
237
238 BLK_MQ_MAX_DEPTH = 10240,
239
240 BLK_MQ_CPU_WORK_BATCH = 8,
241 };
242 #define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \
243 ((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \
244 ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1))
245 #define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \
246 ((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \
247 << BLK_MQ_F_ALLOC_POLICY_START_BIT)
248
249 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
250 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
251 struct request_queue *q,
252 bool elevator_init);
253 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
254 const struct blk_mq_ops *ops,
255 unsigned int queue_depth,
256 unsigned int set_flags);
257 void blk_mq_unregister_dev(struct device *, struct request_queue *);
258
259 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
260 void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
261
262 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
263
264 void blk_mq_free_request(struct request *rq);
265 bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
266
267 bool blk_mq_queue_inflight(struct request_queue *q);
268
269 enum {
270 /* return when out of requests */
271 BLK_MQ_REQ_NOWAIT = (__force blk_mq_req_flags_t)(1 << 0),
272 /* allocate from reserved pool */
273 BLK_MQ_REQ_RESERVED = (__force blk_mq_req_flags_t)(1 << 1),
274 /* allocate internal/sched tag */
275 BLK_MQ_REQ_INTERNAL = (__force blk_mq_req_flags_t)(1 << 2),
276 /* set RQF_PREEMPT */
277 BLK_MQ_REQ_PREEMPT = (__force blk_mq_req_flags_t)(1 << 3),
278 };
279
280 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
281 blk_mq_req_flags_t flags);
282 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
283 unsigned int op, blk_mq_req_flags_t flags,
284 unsigned int hctx_idx);
285 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
286
287 enum {
288 BLK_MQ_UNIQUE_TAG_BITS = 16,
289 BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1,
290 };
291
292 u32 blk_mq_unique_tag(struct request *rq);
293
blk_mq_unique_tag_to_hwq(u32 unique_tag)294 static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag)
295 {
296 return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS;
297 }
298
blk_mq_unique_tag_to_tag(u32 unique_tag)299 static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
300 {
301 return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
302 }
303
304
305 int blk_mq_request_started(struct request *rq);
306 int blk_mq_request_completed(struct request *rq);
307 void blk_mq_start_request(struct request *rq);
308 void blk_mq_end_request(struct request *rq, blk_status_t error);
309 void __blk_mq_end_request(struct request *rq, blk_status_t error);
310
311 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list);
312 void blk_mq_kick_requeue_list(struct request_queue *q);
313 void blk_mq_delay_kick_requeue_list(struct request_queue *q, unsigned long msecs);
314 bool blk_mq_complete_request(struct request *rq);
315 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
316 struct bio *bio, unsigned int nr_segs);
317 bool blk_mq_queue_stopped(struct request_queue *q);
318 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
319 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
320 void blk_mq_stop_hw_queues(struct request_queue *q);
321 void blk_mq_start_hw_queues(struct request_queue *q);
322 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
323 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
324 void blk_mq_quiesce_queue(struct request_queue *q);
325 void blk_mq_unquiesce_queue(struct request_queue *q);
326 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
327 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
328 void blk_mq_run_hw_queues(struct request_queue *q, bool async);
329 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
330 busy_tag_iter_fn *fn, void *priv);
331 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset);
332 void blk_mq_freeze_queue(struct request_queue *q);
333 void blk_mq_unfreeze_queue(struct request_queue *q);
334 void blk_freeze_queue_start(struct request_queue *q);
335 void blk_mq_freeze_queue_wait(struct request_queue *q);
336 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
337 unsigned long timeout);
338
339 int blk_mq_map_queues(struct blk_mq_queue_map *qmap);
340 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues);
341
342 void blk_mq_quiesce_queue_nowait(struct request_queue *q);
343
344 unsigned int blk_mq_rq_cpu(struct request *rq);
345
346 /*
347 * Driver command data is immediately after the request. So subtract request
348 * size to get back to the original request, add request size to get the PDU.
349 */
blk_mq_rq_from_pdu(void * pdu)350 static inline struct request *blk_mq_rq_from_pdu(void *pdu)
351 {
352 return pdu - sizeof(struct request);
353 }
blk_mq_rq_to_pdu(struct request * rq)354 static inline void *blk_mq_rq_to_pdu(struct request *rq)
355 {
356 return rq + 1;
357 }
358
359 #define queue_for_each_hw_ctx(q, hctx, i) \
360 for ((i) = 0; (i) < (q)->nr_hw_queues && \
361 ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
362
363 #define hctx_for_each_ctx(hctx, ctx, i) \
364 for ((i) = 0; (i) < (hctx)->nr_ctx && \
365 ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
366
request_to_qc_t(struct blk_mq_hw_ctx * hctx,struct request * rq)367 static inline blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx,
368 struct request *rq)
369 {
370 if (rq->tag != -1)
371 return rq->tag | (hctx->queue_num << BLK_QC_T_SHIFT);
372
373 return rq->internal_tag | (hctx->queue_num << BLK_QC_T_SHIFT) |
374 BLK_QC_T_INTERNAL;
375 }
376
blk_mq_cleanup_rq(struct request * rq)377 static inline void blk_mq_cleanup_rq(struct request *rq)
378 {
379 if (rq->q->mq_ops->cleanup_rq)
380 rq->q->mq_ops->cleanup_rq(rq);
381 }
382
383 #endif
384