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