1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef INT_BLK_MQ_H 3 #define INT_BLK_MQ_H 4 5 #include <linux/blk-mq.h> 6 #include "blk-stat.h" 7 8 struct blk_mq_tag_set; 9 10 struct blk_mq_ctxs { 11 struct kobject kobj; 12 struct blk_mq_ctx __percpu *queue_ctx; 13 }; 14 15 /** 16 * struct blk_mq_ctx - State for a software queue facing the submitting CPUs 17 */ 18 struct blk_mq_ctx { 19 struct { 20 spinlock_t lock; 21 struct list_head rq_lists[HCTX_MAX_TYPES]; 22 } ____cacheline_aligned_in_smp; 23 24 unsigned int cpu; 25 unsigned short index_hw[HCTX_MAX_TYPES]; 26 struct blk_mq_hw_ctx *hctxs[HCTX_MAX_TYPES]; 27 28 struct request_queue *queue; 29 struct blk_mq_ctxs *ctxs; 30 struct kobject kobj; 31 } ____cacheline_aligned_in_smp; 32 33 enum { 34 BLK_MQ_NO_TAG = -1U, 35 BLK_MQ_TAG_MIN = 1, 36 BLK_MQ_TAG_MAX = BLK_MQ_NO_TAG - 1, 37 }; 38 39 #define BLK_MQ_CPU_WORK_BATCH (8) 40 41 typedef unsigned int __bitwise blk_insert_t; 42 #define BLK_MQ_INSERT_AT_HEAD ((__force blk_insert_t)0x01) 43 44 void blk_mq_submit_bio(struct bio *bio); 45 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob, 46 unsigned int flags); 47 void blk_mq_exit_queue(struct request_queue *q); 48 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr); 49 void blk_mq_wake_waiters(struct request_queue *q); 50 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *, 51 unsigned int); 52 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list); 53 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx, 54 struct blk_mq_ctx *start); 55 void blk_mq_put_rq_ref(struct request *rq); 56 57 /* 58 * Internal helpers for allocating/freeing the request map 59 */ 60 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags, 61 unsigned int hctx_idx); 62 void blk_mq_free_rq_map(struct blk_mq_tags *tags); 63 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set, 64 unsigned int hctx_idx, unsigned int depth); 65 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set, 66 struct blk_mq_tags *tags, 67 unsigned int hctx_idx); 68 69 /* 70 * CPU -> queue mappings 71 */ 72 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int); 73 74 /* 75 * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue 76 * @q: request queue 77 * @type: the hctx type index 78 * @cpu: CPU 79 */ 80 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q, 81 enum hctx_type type, 82 unsigned int cpu) 83 { 84 return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]); 85 } 86 87 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf) 88 { 89 enum hctx_type type = HCTX_TYPE_DEFAULT; 90 91 /* 92 * The caller ensure that if REQ_POLLED, poll must be enabled. 93 */ 94 if (opf & REQ_POLLED) 95 type = HCTX_TYPE_POLL; 96 else if ((opf & REQ_OP_MASK) == REQ_OP_READ) 97 type = HCTX_TYPE_READ; 98 return type; 99 } 100 101 /* 102 * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue 103 * @opf: operation type (REQ_OP_*) and flags (e.g. REQ_POLLED). 104 * @ctx: software queue cpu ctx 105 */ 106 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(blk_opf_t opf, 107 struct blk_mq_ctx *ctx) 108 { 109 return ctx->hctxs[blk_mq_get_hctx_type(opf)]; 110 } 111 112 /* 113 * sysfs helpers 114 */ 115 extern void blk_mq_sysfs_init(struct request_queue *q); 116 extern void blk_mq_sysfs_deinit(struct request_queue *q); 117 int blk_mq_sysfs_register(struct gendisk *disk); 118 void blk_mq_sysfs_unregister(struct gendisk *disk); 119 int blk_mq_sysfs_register_hctxs(struct request_queue *q); 120 void blk_mq_sysfs_unregister_hctxs(struct request_queue *q); 121 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx); 122 void blk_mq_free_plug_rqs(struct blk_plug *plug); 123 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule); 124 125 void blk_mq_cancel_work_sync(struct request_queue *q); 126 127 void blk_mq_release(struct request_queue *q); 128 129 static 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 */ 141 static inline struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q) 142 { 143 return __blk_mq_get_ctx(q, raw_smp_processor_id()); 144 } 145 146 struct blk_mq_alloc_data { 147 /* input parameter */ 148 struct request_queue *q; 149 blk_mq_req_flags_t flags; 150 unsigned int shallow_depth; 151 blk_opf_t cmd_flags; 152 req_flags_t rq_flags; 153 154 /* allocate multiple requests/tags in one go */ 155 unsigned int nr_tags; 156 struct rq_list *cached_rqs; 157 158 /* input & output parameter */ 159 struct blk_mq_ctx *ctx; 160 struct blk_mq_hw_ctx *hctx; 161 }; 162 163 struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags, 164 unsigned int reserved_tags, unsigned int flags, int node); 165 void blk_mq_free_tags(struct blk_mq_tags *tags); 166 167 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data); 168 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags, 169 unsigned int *offset); 170 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx, 171 unsigned int tag); 172 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags); 173 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx, 174 struct blk_mq_tags **tags, unsigned int depth, bool can_grow); 175 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set, 176 unsigned int size); 177 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q); 178 179 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool); 180 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn, 181 void *priv); 182 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn, 183 void *priv); 184 185 static inline struct sbq_wait_state *bt_wait_ptr(struct sbitmap_queue *bt, 186 struct blk_mq_hw_ctx *hctx) 187 { 188 if (!hctx) 189 return &bt->ws[0]; 190 return sbq_wait_ptr(bt, &hctx->wait_index); 191 } 192 193 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *); 194 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *); 195 196 static inline void blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx) 197 { 198 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 199 __blk_mq_tag_busy(hctx); 200 } 201 202 static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx) 203 { 204 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 205 __blk_mq_tag_idle(hctx); 206 } 207 208 static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags, 209 unsigned int tag) 210 { 211 return tag < tags->nr_reserved_tags; 212 } 213 214 static inline bool blk_mq_is_shared_tags(unsigned int flags) 215 { 216 return flags & BLK_MQ_F_TAG_HCTX_SHARED; 217 } 218 219 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data) 220 { 221 if (data->rq_flags & RQF_SCHED_TAGS) 222 return data->hctx->sched_tags; 223 return data->hctx->tags; 224 } 225 226 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx) 227 { 228 /* Fast path: hardware queue is not stopped most of the time. */ 229 if (likely(!test_bit(BLK_MQ_S_STOPPED, &hctx->state))) 230 return false; 231 232 /* 233 * This barrier is used to order adding of dispatch list before and 234 * the test of BLK_MQ_S_STOPPED below. Pairs with the memory barrier 235 * in blk_mq_start_stopped_hw_queue() so that dispatch code could 236 * either see BLK_MQ_S_STOPPED is cleared or dispatch list is not 237 * empty to avoid missing dispatching requests. 238 */ 239 smp_mb(); 240 241 return test_bit(BLK_MQ_S_STOPPED, &hctx->state); 242 } 243 244 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx) 245 { 246 return hctx->nr_ctx && hctx->tags; 247 } 248 249 unsigned int blk_mq_in_flight(struct request_queue *q, 250 struct block_device *part); 251 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part, 252 unsigned int inflight[2]); 253 254 static inline void blk_mq_put_dispatch_budget(struct request_queue *q, 255 int budget_token) 256 { 257 if (q->mq_ops->put_budget) 258 q->mq_ops->put_budget(q, budget_token); 259 } 260 261 static inline int blk_mq_get_dispatch_budget(struct request_queue *q) 262 { 263 if (q->mq_ops->get_budget) 264 return q->mq_ops->get_budget(q); 265 return 0; 266 } 267 268 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token) 269 { 270 if (token < 0) 271 return; 272 273 if (rq->q->mq_ops->set_rq_budget_token) 274 rq->q->mq_ops->set_rq_budget_token(rq, token); 275 } 276 277 static inline int blk_mq_get_rq_budget_token(struct request *rq) 278 { 279 if (rq->q->mq_ops->get_rq_budget_token) 280 return rq->q->mq_ops->get_rq_budget_token(rq); 281 return -1; 282 } 283 284 static inline void __blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx, 285 int val) 286 { 287 if (blk_mq_is_shared_tags(hctx->flags)) 288 atomic_add(val, &hctx->queue->nr_active_requests_shared_tags); 289 else 290 atomic_add(val, &hctx->nr_active); 291 } 292 293 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx) 294 { 295 __blk_mq_add_active_requests(hctx, 1); 296 } 297 298 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx, 299 int val) 300 { 301 if (blk_mq_is_shared_tags(hctx->flags)) 302 atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags); 303 else 304 atomic_sub(val, &hctx->nr_active); 305 } 306 307 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx) 308 { 309 __blk_mq_sub_active_requests(hctx, 1); 310 } 311 312 static inline void blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx, 313 int val) 314 { 315 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 316 __blk_mq_add_active_requests(hctx, val); 317 } 318 319 static inline void blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx) 320 { 321 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 322 __blk_mq_inc_active_requests(hctx); 323 } 324 325 static inline void blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx, 326 int val) 327 { 328 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 329 __blk_mq_sub_active_requests(hctx, val); 330 } 331 332 static inline void blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx) 333 { 334 if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED) 335 __blk_mq_dec_active_requests(hctx); 336 } 337 338 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx) 339 { 340 if (blk_mq_is_shared_tags(hctx->flags)) 341 return atomic_read(&hctx->queue->nr_active_requests_shared_tags); 342 return atomic_read(&hctx->nr_active); 343 } 344 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx, 345 struct request *rq) 346 { 347 blk_mq_dec_active_requests(hctx); 348 blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag); 349 rq->tag = BLK_MQ_NO_TAG; 350 } 351 352 static inline void blk_mq_put_driver_tag(struct request *rq) 353 { 354 if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG) 355 return; 356 357 __blk_mq_put_driver_tag(rq->mq_hctx, rq); 358 } 359 360 bool __blk_mq_alloc_driver_tag(struct request *rq); 361 362 static inline bool blk_mq_get_driver_tag(struct request *rq) 363 { 364 if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq)) 365 return false; 366 367 return true; 368 } 369 370 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap) 371 { 372 int cpu; 373 374 for_each_possible_cpu(cpu) 375 qmap->mq_map[cpu] = 0; 376 } 377 378 /* Free all requests on the list */ 379 static inline void blk_mq_free_requests(struct list_head *list) 380 { 381 while (!list_empty(list)) { 382 struct request *rq = list_entry_rq(list->next); 383 384 list_del_init(&rq->queuelist); 385 blk_mq_free_request(rq); 386 } 387 } 388 389 /* 390 * For shared tag users, we track the number of currently active users 391 * and attempt to provide a fair share of the tag depth for each of them. 392 */ 393 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, 394 struct sbitmap_queue *bt) 395 { 396 unsigned int depth, users; 397 398 if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)) 399 return true; 400 401 /* 402 * Don't try dividing an ant 403 */ 404 if (bt->sb.depth == 1) 405 return true; 406 407 if (blk_mq_is_shared_tags(hctx->flags)) { 408 struct request_queue *q = hctx->queue; 409 410 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags)) 411 return true; 412 } else { 413 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state)) 414 return true; 415 } 416 417 users = READ_ONCE(hctx->tags->active_queues); 418 if (!users) 419 return true; 420 421 /* 422 * Allow at least some tags 423 */ 424 depth = max((bt->sb.depth + users - 1) / users, 4U); 425 return __blk_mq_active_requests(hctx) < depth; 426 } 427 428 /* run the code block in @dispatch_ops with rcu/srcu read lock held */ 429 #define __blk_mq_run_dispatch_ops(q, check_sleep, dispatch_ops) \ 430 do { \ 431 if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) { \ 432 struct blk_mq_tag_set *__tag_set = (q)->tag_set; \ 433 int srcu_idx; \ 434 \ 435 might_sleep_if(check_sleep); \ 436 srcu_idx = srcu_read_lock(__tag_set->srcu); \ 437 (dispatch_ops); \ 438 srcu_read_unlock(__tag_set->srcu, srcu_idx); \ 439 } else { \ 440 rcu_read_lock(); \ 441 (dispatch_ops); \ 442 rcu_read_unlock(); \ 443 } \ 444 } while (0) 445 446 #define blk_mq_run_dispatch_ops(q, dispatch_ops) \ 447 __blk_mq_run_dispatch_ops(q, true, dispatch_ops) \ 448 449 static inline bool blk_mq_can_poll(struct request_queue *q) 450 { 451 return (q->limits.features & BLK_FEAT_POLL) && 452 q->tag_set->map[HCTX_TYPE_POLL].nr_queues; 453 } 454 455 #endif 456