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 struct blk_mq_hw_ctx {
13 struct {
14 spinlock_t lock;
15 struct list_head dispatch;
16 unsigned long state; /* BLK_MQ_S_* flags */
17 } ____cacheline_aligned_in_smp;
18
19 struct delayed_work run_work;
20 cpumask_var_t cpumask;
21 int next_cpu;
22 int next_cpu_batch;
23
24 unsigned long flags; /* BLK_MQ_F_* flags */
25
26 void *sched_data;
27 struct request_queue *queue;
28 struct blk_flush_queue *fq;
29
30 void *driver_data;
31
32 struct sbitmap ctx_map;
33
34 struct blk_mq_ctx **ctxs;
35 unsigned int nr_ctx;
36
37 wait_queue_entry_t dispatch_wait;
38 atomic_t wait_index;
39
40 struct blk_mq_tags *tags;
41 struct blk_mq_tags *sched_tags;
42
43 unsigned long queued;
44 unsigned long run;
45 #define BLK_MQ_MAX_DISPATCH_ORDER 7
46 unsigned long dispatched[BLK_MQ_MAX_DISPATCH_ORDER];
47
48 unsigned int numa_node;
49 unsigned int queue_num;
50
51 atomic_t nr_active;
52
53 struct hlist_node cpuhp_dead;
54 struct kobject kobj;
55
56 unsigned long poll_considered;
57 unsigned long poll_invoked;
58 unsigned long poll_success;
59
60 #ifdef CONFIG_BLK_DEBUG_FS
61 struct dentry *debugfs_dir;
62 struct dentry *sched_debugfs_dir;
63 #endif
64
65 /* Must be the last member - see also blk_mq_hw_ctx_size(). */
66 struct srcu_struct queue_rq_srcu[0];
67 };
68
69 struct blk_mq_tag_set {
70 unsigned int *mq_map;
71 const struct blk_mq_ops *ops;
72 unsigned int nr_hw_queues;
73 unsigned int queue_depth; /* max hw supported */
74 unsigned int reserved_tags;
75 unsigned int cmd_size; /* per-request extra data */
76 int numa_node;
77 unsigned int timeout;
78 unsigned int flags; /* BLK_MQ_F_* */
79 void *driver_data;
80
81 struct blk_mq_tags **tags;
82
83 struct mutex tag_list_lock;
84 struct list_head tag_list;
85 };
86
87 struct blk_mq_queue_data {
88 struct request *rq;
89 bool last;
90 };
91
92 typedef blk_status_t (queue_rq_fn)(struct blk_mq_hw_ctx *,
93 const struct blk_mq_queue_data *);
94 typedef enum blk_eh_timer_return (timeout_fn)(struct request *, bool);
95 typedef int (init_hctx_fn)(struct blk_mq_hw_ctx *, void *, unsigned int);
96 typedef void (exit_hctx_fn)(struct blk_mq_hw_ctx *, unsigned int);
97 typedef int (init_request_fn)(struct blk_mq_tag_set *set, struct request *,
98 unsigned int, unsigned int);
99 typedef void (exit_request_fn)(struct blk_mq_tag_set *set, struct request *,
100 unsigned int);
101
102 typedef void (busy_iter_fn)(struct blk_mq_hw_ctx *, struct request *, void *,
103 bool);
104 typedef void (busy_tag_iter_fn)(struct request *, void *, bool);
105 typedef int (poll_fn)(struct blk_mq_hw_ctx *, unsigned int);
106 typedef int (map_queues_fn)(struct blk_mq_tag_set *set);
107
108
109 struct blk_mq_ops {
110 /*
111 * Queue request
112 */
113 queue_rq_fn *queue_rq;
114
115 /*
116 * Called on request timeout
117 */
118 timeout_fn *timeout;
119
120 /*
121 * Called to poll for completion of a specific tag.
122 */
123 poll_fn *poll;
124
125 softirq_done_fn *complete;
126
127 /*
128 * Called when the block layer side of a hardware queue has been
129 * set up, allowing the driver to allocate/init matching structures.
130 * Ditto for exit/teardown.
131 */
132 init_hctx_fn *init_hctx;
133 exit_hctx_fn *exit_hctx;
134
135 /*
136 * Called for every command allocated by the block layer to allow
137 * the driver to set up driver specific data.
138 *
139 * Tag greater than or equal to queue_depth is for setting up
140 * flush request.
141 *
142 * Ditto for exit/teardown.
143 */
144 init_request_fn *init_request;
145 exit_request_fn *exit_request;
146 /* Called from inside blk_get_request() */
147 void (*initialize_rq_fn)(struct request *rq);
148
149 map_queues_fn *map_queues;
150
151 #ifdef CONFIG_BLK_DEBUG_FS
152 /*
153 * Used by the debugfs implementation to show driver-specific
154 * information about a request.
155 */
156 void (*show_rq)(struct seq_file *m, struct request *rq);
157 #endif
158 };
159
160 enum {
161 BLK_MQ_F_SHOULD_MERGE = 1 << 0,
162 BLK_MQ_F_TAG_SHARED = 1 << 1,
163 BLK_MQ_F_SG_MERGE = 1 << 2,
164 BLK_MQ_F_BLOCKING = 1 << 5,
165 BLK_MQ_F_NO_SCHED = 1 << 6,
166 BLK_MQ_F_ALLOC_POLICY_START_BIT = 8,
167 BLK_MQ_F_ALLOC_POLICY_BITS = 1,
168
169 BLK_MQ_S_STOPPED = 0,
170 BLK_MQ_S_TAG_ACTIVE = 1,
171 BLK_MQ_S_SCHED_RESTART = 2,
172 BLK_MQ_S_TAG_WAITING = 3,
173 BLK_MQ_S_START_ON_RUN = 4,
174
175 BLK_MQ_MAX_DEPTH = 10240,
176
177 BLK_MQ_CPU_WORK_BATCH = 8,
178 };
179 #define BLK_MQ_FLAG_TO_ALLOC_POLICY(flags) \
180 ((flags >> BLK_MQ_F_ALLOC_POLICY_START_BIT) & \
181 ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1))
182 #define BLK_ALLOC_POLICY_TO_MQ_FLAG(policy) \
183 ((policy & ((1 << BLK_MQ_F_ALLOC_POLICY_BITS) - 1)) \
184 << BLK_MQ_F_ALLOC_POLICY_START_BIT)
185
186 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *);
187 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
188 struct request_queue *q);
189 int blk_mq_register_dev(struct device *, struct request_queue *);
190 void blk_mq_unregister_dev(struct device *, struct request_queue *);
191
192 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set);
193 void blk_mq_free_tag_set(struct blk_mq_tag_set *set);
194
195 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule);
196
197 void blk_mq_free_request(struct request *rq);
198 bool blk_mq_can_queue(struct blk_mq_hw_ctx *);
199
200 enum {
201 BLK_MQ_REQ_NOWAIT = (1 << 0), /* return when out of requests */
202 BLK_MQ_REQ_RESERVED = (1 << 1), /* allocate from reserved pool */
203 BLK_MQ_REQ_INTERNAL = (1 << 2), /* allocate internal/sched tag */
204 };
205
206 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
207 unsigned int flags);
208 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
209 unsigned int op, unsigned int flags, unsigned int hctx_idx);
210 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag);
211
212 enum {
213 BLK_MQ_UNIQUE_TAG_BITS = 16,
214 BLK_MQ_UNIQUE_TAG_MASK = (1 << BLK_MQ_UNIQUE_TAG_BITS) - 1,
215 };
216
217 u32 blk_mq_unique_tag(struct request *rq);
218
blk_mq_unique_tag_to_hwq(u32 unique_tag)219 static inline u16 blk_mq_unique_tag_to_hwq(u32 unique_tag)
220 {
221 return unique_tag >> BLK_MQ_UNIQUE_TAG_BITS;
222 }
223
blk_mq_unique_tag_to_tag(u32 unique_tag)224 static inline u16 blk_mq_unique_tag_to_tag(u32 unique_tag)
225 {
226 return unique_tag & BLK_MQ_UNIQUE_TAG_MASK;
227 }
228
229
230 int blk_mq_request_started(struct request *rq);
231 void blk_mq_start_request(struct request *rq);
232 void blk_mq_end_request(struct request *rq, blk_status_t error);
233 void __blk_mq_end_request(struct request *rq, blk_status_t error);
234
235 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list);
236 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
237 bool kick_requeue_list);
238 void blk_mq_kick_requeue_list(struct request_queue *q);
239 void blk_mq_delay_kick_requeue_list(struct request_queue *q, unsigned long msecs);
240 void blk_mq_complete_request(struct request *rq);
241
242 bool blk_mq_queue_stopped(struct request_queue *q);
243 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx);
244 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx);
245 void blk_mq_stop_hw_queues(struct request_queue *q);
246 void blk_mq_start_hw_queues(struct request_queue *q);
247 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
248 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async);
249 void blk_mq_quiesce_queue(struct request_queue *q);
250 void blk_mq_unquiesce_queue(struct request_queue *q);
251 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
252 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
253 void blk_mq_run_hw_queues(struct request_queue *q, bool async);
254 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs);
255 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
256 busy_tag_iter_fn *fn, void *priv);
257 void blk_mq_freeze_queue(struct request_queue *q);
258 void blk_mq_unfreeze_queue(struct request_queue *q);
259 void blk_freeze_queue_start(struct request_queue *q);
260 void blk_mq_freeze_queue_wait(struct request_queue *q);
261 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
262 unsigned long timeout);
263 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set,
264 int (reinit_request)(void *, struct request *));
265
266 int blk_mq_map_queues(struct blk_mq_tag_set *set);
267 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues);
268
269 void blk_mq_quiesce_queue_nowait(struct request_queue *q);
270
271 /*
272 * Driver command data is immediately after the request. So subtract request
273 * size to get back to the original request, add request size to get the PDU.
274 */
blk_mq_rq_from_pdu(void * pdu)275 static inline struct request *blk_mq_rq_from_pdu(void *pdu)
276 {
277 return pdu - sizeof(struct request);
278 }
blk_mq_rq_to_pdu(struct request * rq)279 static inline void *blk_mq_rq_to_pdu(struct request *rq)
280 {
281 return rq + 1;
282 }
283
284 #define queue_for_each_hw_ctx(q, hctx, i) \
285 for ((i) = 0; (i) < (q)->nr_hw_queues && \
286 ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
287
288 #define hctx_for_each_ctx(hctx, ctx, i) \
289 for ((i) = 0; (i) < (hctx)->nr_ctx && \
290 ({ ctx = (hctx)->ctxs[(i)]; 1; }); (i)++)
291
292 #endif
293