xref: /linux-6.15/include/linux/blk-mq.h (revision 1ff85bfa)
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 	unsigned int		nr_expired;
61 
62 	struct hlist_node	cpuhp_dead;
63 	struct kobject		kobj;
64 
65 	unsigned long		poll_considered;
66 	unsigned long		poll_invoked;
67 	unsigned long		poll_success;
68 
69 #ifdef CONFIG_BLK_DEBUG_FS
70 	struct dentry		*debugfs_dir;
71 	struct dentry		*sched_debugfs_dir;
72 #endif
73 
74 	/* Must be the last member - see also blk_mq_hw_ctx_size(). */
75 	struct srcu_struct	srcu[0];
76 };
77 
78 struct blk_mq_queue_map {
79 	unsigned int *mq_map;
80 	unsigned int nr_queues;
81 	unsigned int queue_offset;
82 };
83 
84 enum hctx_type {
85 	HCTX_TYPE_DEFAULT,	/* all I/O not otherwise accounted for */
86 	HCTX_TYPE_READ,		/* just for READ I/O */
87 	HCTX_TYPE_POLL,		/* polled I/O of any kind */
88 
89 	HCTX_MAX_TYPES,
90 };
91 
92 struct blk_mq_tag_set {
93 	/*
94 	 * map[] holds ctx -> hctx mappings, one map exists for each type
95 	 * that the driver wishes to support. There are no restrictions
96 	 * on maps being of the same size, and it's perfectly legal to
97 	 * share maps between types.
98 	 */
99 	struct blk_mq_queue_map	map[HCTX_MAX_TYPES];
100 	unsigned int		nr_maps;	/* nr entries in map[] */
101 	const struct blk_mq_ops	*ops;
102 	unsigned int		nr_hw_queues;	/* nr hw queues across maps */
103 	unsigned int		queue_depth;	/* max hw supported */
104 	unsigned int		reserved_tags;
105 	unsigned int		cmd_size;	/* per-request extra data */
106 	int			numa_node;
107 	unsigned int		timeout;
108 	unsigned int		flags;		/* BLK_MQ_F_* */
109 	void			*driver_data;
110 
111 	struct blk_mq_tags	**tags;
112 
113 	struct mutex		tag_list_lock;
114 	struct list_head	tag_list;
115 };
116 
117 struct blk_mq_queue_data {
118 	struct request *rq;
119 	bool last;
120 };
121 
122 typedef blk_status_t (queue_rq_fn)(struct blk_mq_hw_ctx *,
123 		const struct blk_mq_queue_data *);
124 typedef void (commit_rqs_fn)(struct blk_mq_hw_ctx *);
125 typedef bool (get_budget_fn)(struct blk_mq_hw_ctx *);
126 typedef void (put_budget_fn)(struct blk_mq_hw_ctx *);
127 typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool);
128 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
129 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
130 typedef int (init_request_fn)(struct blk_mq_tag_set *set, struct request *,
131 		unsigned int, unsigned int);
132 typedef void (exit_request_fn)(struct blk_mq_tag_set *set, struct request *,
133 		unsigned int);
134 
135 typedef bool (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *,
136 		bool);
137 typedef bool (busy_tag_iter_fn)(struct request *, void *, bool);
138 typedef int (poll_fn)(struct blk_mq_hw_ctx *);
139 typedef int (map_queues_fn)(struct blk_mq_tag_set *set);
140 typedef bool (busy_fn)(struct request_queue *);
141 typedef void (complete_fn)(struct request *);
142 
143 
144 struct blk_mq_ops {
145 	/*
146 	 * Queue request
147 	 */
148 	queue_rq_fn		*queue_rq;
149 
150 	/*
151 	 * If a driver uses bd->last to judge when to submit requests to
152 	 * hardware, it must define this function. In case of errors that
153 	 * make us stop issuing further requests, this hook serves the
154 	 * purpose of kicking the hardware (which the last request otherwise
155 	 * would have done).
156 	 */
157 	commit_rqs_fn		*commit_rqs;
158 
159 	/*
160 	 * Reserve budget before queue request, once .queue_rq is
161 	 * run, it is driver's responsibility to release the
162 	 * reserved budget. Also we have to handle failure case
163 	 * of .get_budget for avoiding I/O deadlock.
164 	 */
165 	get_budget_fn		*get_budget;
166 	put_budget_fn		*put_budget;
167 
168 	/*
169 	 * Called on request timeout
170 	 */
171 	timeout_fn		*timeout;
172 
173 	/*
174 	 * Called to poll for completion of a specific tag.
175 	 */
176 	poll_fn			*poll;
177 
178 	complete_fn		*complete;
179 
180 	/*
181 	 * Called when the block layer side of a hardware queue has been
182 	 * set up, allowing the driver to allocate/init matching structures.
183 	 * Ditto for exit/teardown.
184 	 */
185 	init_hctx_fn		*init_hctx;
186 	exit_hctx_fn		*exit_hctx;
187 
188 	/*
189 	 * Called for every command allocated by the block layer to allow
190 	 * the driver to set up driver specific data.
191 	 *
192 	 * Tag greater than or equal to queue_depth is for setting up
193 	 * flush request.
194 	 *
195 	 * Ditto for exit/teardown.
196 	 */
197 	init_request_fn		*init_request;
198 	exit_request_fn		*exit_request;
199 	/* Called from inside blk_get_request() */
200 	void (*initialize_rq_fn)(struct request *rq);
201 
202 	/*
203 	 * If set, returns whether or not this queue currently is busy
204 	 */
205 	busy_fn			*busy;
206 
207 	map_queues_fn		*map_queues;
208 
209 #ifdef CONFIG_BLK_DEBUG_FS
210 	/*
211 	 * Used by the debugfs implementation to show driver-specific
212 	 * information about a request.
213 	 */
214 	void (*show_rq)(struct seq_file *m, struct request *rq);
215 #endif
216 };
217 
218 enum {
219 	BLK_MQ_F_SHOULD_MERGE	= 1 << 0,
220 	BLK_MQ_F_TAG_SHARED	= 1 << 1,
221 	BLK_MQ_F_SG_MERGE	= 1 << 2,
222 	BLK_MQ_F_BLOCKING	= 1 << 5,
223 	BLK_MQ_F_NO_SCHED	= 1 << 6,
224 	BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
225 	BLK_MQ_F_ALLOC_POLICY_BITS = 1,
226 
227 	BLK_MQ_S_STOPPED	= 0,
228 	BLK_MQ_S_TAG_ACTIVE	= 1,
229 	BLK_MQ_S_SCHED_RESTART	= 2,
230 
231 	BLK_MQ_MAX_DEPTH	= 10240,
232 
233 	BLK_MQ_CPU_WORK_BATCH	= 8,
234 };
235 #define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \
236 	((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \
237 		((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1))
238 #define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \
239 	((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \
240 		<< BLK_MQ_F_ALLOC_POLICY_START_BIT)
241 
242 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
243 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
244 						  struct request_queue *q);
245 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
246 						const struct blk_mq_ops *ops,
247 						unsigned int queue_depth,
248 						unsigned int set_flags);
249 int blk_mq_register_dev(struct device *, struct request_queue *);
250 void blk_mq_unregister_dev(struct device *, struct request_queue *);
251 
252 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
253 void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
254 
255 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
256 
257 void blk_mq_free_request(struct request *rq);
258 bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
259 
260 bool blk_mq_queue_inflight(struct request_queue *q);
261 
262 enum {
263 	/* return when out of requests */
264 	BLK_MQ_REQ_NOWAIT	= (__force blk_mq_req_flags_t)(1 << 0),
265 	/* allocate from reserved pool */
266 	BLK_MQ_REQ_RESERVED	= (__force blk_mq_req_flags_t)(1 << 1),
267 	/* allocate internal/sched tag */
268 	BLK_MQ_REQ_INTERNAL	= (__force blk_mq_req_flags_t)(1 << 2),
269 	/* set RQF_PREEMPT */
270 	BLK_MQ_REQ_PREEMPT	= (__force blk_mq_req_flags_t)(1 << 3),
271 };
272 
273 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
274 		blk_mq_req_flags_t flags);
275 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
276 		unsigned int op, blk_mq_req_flags_t flags,
277 		unsigned int hctx_idx);
278 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
279 
280 enum {
281 	BLK_MQ_UNIQUE_TAG_BITS = 16,
282 	BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1,
283 };
284 
285 u32 blk_mq_unique_tag(struct request *rq);
286 
287 static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag)
288 {
289 	return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS;
290 }
291 
292 static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
293 {
294 	return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
295 }
296 
297 
298 int blk_mq_request_started(struct request *rq);
299 void blk_mq_start_request(struct request *rq);
300 void blk_mq_end_request(struct request *rq, blk_status_t error);
301 void __blk_mq_end_request(struct request *rq, blk_status_t error);
302 
303 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list);
304 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
305 				bool kick_requeue_list);
306 void blk_mq_kick_requeue_list(struct request_queue *q);
307 void blk_mq_delay_kick_requeue_list(struct request_queue *q, unsigned long msecs);
308 bool blk_mq_complete_request(struct request *rq);
309 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
310 			   struct bio *bio);
311 bool blk_mq_queue_stopped(struct request_queue *q);
312 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
313 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
314 void blk_mq_stop_hw_queues(struct request_queue *q);
315 void blk_mq_start_hw_queues(struct request_queue *q);
316 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
317 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
318 void blk_mq_quiesce_queue(struct request_queue *q);
319 void blk_mq_unquiesce_queue(struct request_queue *q);
320 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
321 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
322 void blk_mq_run_hw_queues(struct request_queue *q, bool async);
323 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
324 		busy_tag_iter_fn *fn, void *priv);
325 void blk_mq_freeze_queue(struct request_queue *q);
326 void blk_mq_unfreeze_queue(struct request_queue *q);
327 void blk_freeze_queue_start(struct request_queue *q);
328 void blk_mq_freeze_queue_wait(struct request_queue *q);
329 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
330 				     unsigned long timeout);
331 
332 int blk_mq_map_queues(struct blk_mq_queue_map *qmap);
333 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues);
334 
335 void blk_mq_quiesce_queue_nowait(struct request_queue *q);
336 
337 unsigned int blk_mq_rq_cpu(struct request *rq);
338 
339 /*
340  * Driver command data is immediately after the request. So subtract request
341  * size to get back to the original request, add request size to get the PDU.
342  */
343 static inline struct request *blk_mq_rq_from_pdu(void *pdu)
344 {
345 	return pdu - sizeof(struct request);
346 }
347 static inline void *blk_mq_rq_to_pdu(struct request *rq)
348 {
349 	return rq + 1;
350 }
351 
352 #define queue_for_each_hw_ctx(q, hctx, i)				\
353 	for ((i) = 0; (i) < (q)->nr_hw_queues &&			\
354 	     ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
355 
356 #define hctx_for_each_ctx(hctx, ctx, i)					\
357 	for ((i) = 0; (i) < (hctx)->nr_ctx &&				\
358 	     ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
359 
360 static inline blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx,
361 		struct request *rq)
362 {
363 	if (rq->tag != -1)
364 		return rq->tag | (hctx->queue_num << BLK_QC_T_SHIFT);
365 
366 	return rq->internal_tag | (hctx->queue_num << BLK_QC_T_SHIFT) |
367 			BLK_QC_T_INTERNAL;
368 }
369 
370 #endif
371