block/blk-mq.h

   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 typedef unsigned int __bitwise blk_insert_t;
  40 #define BLK_MQ_INSERT_AT_HEAD           ((__force blk_insert_t)0x01)
  41
  42 void blk_mq_submit_bio(struct bio *bio);
  43 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
  44                 unsigned int flags);
  45 void blk_mq_exit_queue(struct request_queue *q);
  46 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
  47 void blk_mq_wake_waiters(struct request_queue *q);
  48 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
  49                              unsigned int);
  50 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list);
  51 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
  52                                         struct blk_mq_ctx *start);
  53 void blk_mq_put_rq_ref(struct request *rq);
  54
  55 /*
  56  * Internal helpers for allocating/freeing the request map
  57  */
  58 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
  59                      unsigned int hctx_idx);
  60 void blk_mq_free_rq_map(struct blk_mq_tags *tags);
  61 struct blk_mq_tags *blk_mq_alloc_map_and_rqs(struct blk_mq_tag_set *set,
  62                                 unsigned int hctx_idx, unsigned int depth);
  63 void blk_mq_free_map_and_rqs(struct blk_mq_tag_set *set,
  64                              struct blk_mq_tags *tags,
  65                              unsigned int hctx_idx);
  66
  67 /*
  68  * CPU -> queue mappings
  69  */
  70 extern int blk_mq_hw_queue_to_node(struct blk_mq_queue_map *qmap, unsigned int);
  71
  72 /*
  73  * blk_mq_map_queue_type() - map (hctx_type,cpu) to hardware queue
  74  * @q: request queue
  75  * @type: the hctx type index
  76  * @cpu: CPU
  77  */
  78 static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *q,
  79                                                           enum hctx_type type,
  80                                                           unsigned int cpu)
  81 {
  82         return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]);
  83 }
  84
  85 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf)
  86 {
  87         enum hctx_type type = HCTX_TYPE_DEFAULT;
  88
  89         /*
  90          * The caller ensure that if REQ_POLLED, poll must be enabled.
  91          */
  92         if (opf & REQ_POLLED)
  93                 type = HCTX_TYPE_POLL;
  94         else if ((opf & REQ_OP_MASK) == REQ_OP_READ)
  95                 type = HCTX_TYPE_READ;
  96         return type;
  97 }
  98
  99 /*
 100  * blk_mq_map_queue() - map (cmd_flags,type) to hardware queue
 101  * @q: request queue
 102  * @opf: operation type (REQ_OP_*) and flags (e.g. REQ_POLLED).
 103  * @ctx: software queue cpu ctx
 104  */
 105 static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
 106                                                      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 request **cached_rq;
 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, int node, int alloc_policy);
 165 void blk_mq_free_tags(struct blk_mq_tags *tags);
 166 int blk_mq_init_bitmaps(struct sbitmap_queue *bitmap_tags,
 167                 struct sbitmap_queue *breserved_tags, unsigned int queue_depth,
 168                 unsigned int reserved, int node, int alloc_policy);
 169
 170 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data);
 171 unsigned long blk_mq_get_tags(struct blk_mq_alloc_data *data, int nr_tags,
 172                 unsigned int *offset);
 173 void blk_mq_put_tag(struct blk_mq_tags *tags, struct blk_mq_ctx *ctx,
 174                 unsigned int tag);
 175 void blk_mq_put_tags(struct blk_mq_tags *tags, int *tag_array, int nr_tags);
 176 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
 177                 struct blk_mq_tags **tags, unsigned int depth, bool can_grow);
 178 void blk_mq_tag_resize_shared_tags(struct blk_mq_tag_set *set,
 179                 unsigned int size);
 180 void blk_mq_tag_update_sched_shared_tags(struct request_queue *q);
 181
 182 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool);
 183 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
 184                 void *priv);
 185 void blk_mq_all_tag_iter(struct blk_mq_tags *tags, busy_tag_iter_fn *fn,
 186                 void *priv);
 187
 188 static inline struct sbq_wait_state *bt_wait_ptr(struct sbitmap_queue *bt,
 189                                                  struct blk_mq_hw_ctx *hctx)
 190 {
 191         if (!hctx)
 192                 return &bt->ws[0];
 193         return sbq_wait_ptr(bt, &hctx->wait_index);
 194 }
 195
 196 void __blk_mq_tag_busy(struct blk_mq_hw_ctx *);
 197 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *);
 198
 199 static inline void blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
 200 {
 201         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
 202                 __blk_mq_tag_busy(hctx);
 203 }
 204
 205 static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
 206 {
 207         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
 208                 __blk_mq_tag_idle(hctx);
 209 }
 210
 211 static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags,
 212                                           unsigned int tag)
 213 {
 214         return tag < tags->nr_reserved_tags;
 215 }
 216
 217 static inline bool blk_mq_is_shared_tags(unsigned int flags)
 218 {
 219         return flags & BLK_MQ_F_TAG_HCTX_SHARED;
 220 }
 221
 222 static inline struct blk_mq_tags *blk_mq_tags_from_data(struct blk_mq_alloc_data *data)
 223 {
 224         if (data->rq_flags & RQF_SCHED_TAGS)
 225                 return data->hctx->sched_tags;
 226         return data->hctx->tags;
 227 }
 228
 229 static inline bool blk_mq_hctx_stopped(struct blk_mq_hw_ctx *hctx)
 230 {
 231         return test_bit(BLK_MQ_S_STOPPED, &hctx->state);
 232 }
 233
 234 static inline bool blk_mq_hw_queue_mapped(struct blk_mq_hw_ctx *hctx)
 235 {
 236         return hctx->nr_ctx && hctx->tags;
 237 }
 238
 239 unsigned int blk_mq_in_flight(struct request_queue *q,
 240                 struct block_device *part);
 241 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
 242                 unsigned int inflight[2]);
 243
 244 static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
 245                                               int budget_token)
 246 {
 247         if (q->mq_ops->put_budget)
 248                 q->mq_ops->put_budget(q, budget_token);
 249 }
 250
 251 static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
 252 {
 253         if (q->mq_ops->get_budget)
 254                 return q->mq_ops->get_budget(q);
 255         return 0;
 256 }
 257
 258 static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
 259 {
 260         if (token < 0)
 261                 return;
 262
 263         if (rq->q->mq_ops->set_rq_budget_token)
 264                 rq->q->mq_ops->set_rq_budget_token(rq, token);
 265 }
 266
 267 static inline int blk_mq_get_rq_budget_token(struct request *rq)
 268 {
 269         if (rq->q->mq_ops->get_rq_budget_token)
 270                 return rq->q->mq_ops->get_rq_budget_token(rq);
 271         return -1;
 272 }
 273
 274 static inline void __blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx,
 275                                                 int val)
 276 {
 277         if (blk_mq_is_shared_tags(hctx->flags))
 278                 atomic_add(val, &hctx->queue->nr_active_requests_shared_tags);
 279         else
 280                 atomic_add(val, &hctx->nr_active);
 281 }
 282
 283 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
 284 {
 285         __blk_mq_add_active_requests(hctx, 1);
 286 }
 287
 288 static inline void __blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
 289                 int val)
 290 {
 291         if (blk_mq_is_shared_tags(hctx->flags))
 292                 atomic_sub(val, &hctx->queue->nr_active_requests_shared_tags);
 293         else
 294                 atomic_sub(val, &hctx->nr_active);
 295 }
 296
 297 static inline void __blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
 298 {
 299         __blk_mq_sub_active_requests(hctx, 1);
 300 }
 301
 302 static inline void blk_mq_add_active_requests(struct blk_mq_hw_ctx *hctx,
 303                                               int val)
 304 {
 305         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
 306                 __blk_mq_add_active_requests(hctx, val);
 307 }
 308
 309 static inline void blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)
 310 {
 311         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
 312                 __blk_mq_inc_active_requests(hctx);
 313 }
 314
 315 static inline void blk_mq_sub_active_requests(struct blk_mq_hw_ctx *hctx,
 316                                               int val)
 317 {
 318         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
 319                 __blk_mq_sub_active_requests(hctx, val);
 320 }
 321
 322 static inline void blk_mq_dec_active_requests(struct blk_mq_hw_ctx *hctx)
 323 {
 324         if (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)
 325                 __blk_mq_dec_active_requests(hctx);
 326 }
 327
 328 static inline int __blk_mq_active_requests(struct blk_mq_hw_ctx *hctx)
 329 {
 330         if (blk_mq_is_shared_tags(hctx->flags))
 331                 return atomic_read(&hctx->queue->nr_active_requests_shared_tags);
 332         return atomic_read(&hctx->nr_active);
 333 }
 334 static inline void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 335                                            struct request *rq)
 336 {
 337         blk_mq_dec_active_requests(hctx);
 338         blk_mq_put_tag(hctx->tags, rq->mq_ctx, rq->tag);
 339         rq->tag = BLK_MQ_NO_TAG;
 340 }
 341
 342 static inline void blk_mq_put_driver_tag(struct request *rq)
 343 {
 344         if (rq->tag == BLK_MQ_NO_TAG || rq->internal_tag == BLK_MQ_NO_TAG)
 345                 return;
 346
 347         __blk_mq_put_driver_tag(rq->mq_hctx, rq);
 348 }
 349
 350 bool __blk_mq_alloc_driver_tag(struct request *rq);
 351
 352 static inline bool blk_mq_get_driver_tag(struct request *rq)
 353 {
 354         if (rq->tag == BLK_MQ_NO_TAG && !__blk_mq_alloc_driver_tag(rq))
 355                 return false;
 356
 357         return true;
 358 }
 359
 360 static inline void blk_mq_clear_mq_map(struct blk_mq_queue_map *qmap)
 361 {
 362         int cpu;
 363
 364         for_each_possible_cpu(cpu)
 365                 qmap->mq_map[cpu] = 0;
 366 }
 367
 368 /*
 369  * blk_mq_plug() - Get caller context plug
 370  * @bio : the bio being submitted by the caller context
 371  *
 372  * Plugging, by design, may delay the insertion of BIOs into the elevator in
 373  * order to increase BIO merging opportunities. This however can cause BIO
 374  * insertion order to change from the order in which submit_bio() is being
 375  * executed in the case of multiple contexts concurrently issuing BIOs to a
 376  * device, even if these context are synchronized to tightly control BIO issuing
 377  * order. While this is not a problem with regular block devices, this ordering
 378  * change can cause write BIO failures with zoned block devices as these
 379  * require sequential write patterns to zones. Prevent this from happening by
 380  * ignoring the plug state of a BIO issuing context if it is for a zoned block
 381  * device and the BIO to plug is a write operation.
 382  *
 383  * Return current->plug if the bio can be plugged and NULL otherwise
 384  */
 385 static inline struct blk_plug *blk_mq_plug( struct bio *bio)
 386 {
 387         /* Zoned block device write operation case: do not plug the BIO */
 388         if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
 389             bdev_op_is_zoned_write(bio->bi_bdev, bio_op(bio)))
 390                 return NULL;
 391
 392         /*
 393          * For regular block devices or read operations, use the context plug
 394          * which may be NULL if blk_start_plug() was not executed.
 395          */
 396         return current->plug;
 397 }
 398
 399 /* Free all requests on the list */
 400 static inline void blk_mq_free_requests(struct list_head *list)
 401 {
 402         while (!list_empty(list)) {
 403                 struct request *rq = list_entry_rq(list->next);
 404
 405                 list_del_init(&rq->queuelist);
 406                 blk_mq_free_request(rq);
 407         }
 408 }
 409
 410 /*
 411  * For shared tag users, we track the number of currently active users
 412  * and attempt to provide a fair share of the tag depth for each of them.
 413  */
 414 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
 415                                   struct sbitmap_queue *bt)
 416 {
 417         unsigned int depth, users;
 418
 419         if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED))
 420                 return true;
 421
 422         /*
 423          * Don't try dividing an ant
 424          */
 425         if (bt->sb.depth == 1)
 426                 return true;
 427
 428         if (blk_mq_is_shared_tags(hctx->flags)) {
 429                 struct request_queue *q = hctx->queue;
 430
 431                 if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
 432                         return true;
 433         } else {
 434                 if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
 435                         return true;
 436         }
 437
 438         users = READ_ONCE(hctx->tags->active_queues);
 439         if (!users)
 440                 return true;
 441
 442         /*
 443          * Allow at least some tags
 444          */
 445         depth = max((bt->sb.depth + users - 1) / users, 4U);
 446         return __blk_mq_active_requests(hctx) < depth;
 447 }
 448
 449 /* run the code block in @dispatch_ops with rcu/srcu read lock held */
 450 #define __blk_mq_run_dispatch_ops(q, check_sleep, dispatch_ops) \
 451 do {                                                            \
 452         if ((q)->tag_set->flags & BLK_MQ_F_BLOCKING) {          \
 453                 struct blk_mq_tag_set *__tag_set = (q)->tag_set; \
 454                 int srcu_idx;                                   \
 455                                                                 \
 456                 might_sleep_if(check_sleep);                    \
 457                 srcu_idx = srcu_read_lock(__tag_set->srcu);     \
 458                 (dispatch_ops);                                 \
 459                 srcu_read_unlock(__tag_set->srcu, srcu_idx);    \
 460         } else {                                                \
 461                 rcu_read_lock();                                \
 462                 (dispatch_ops);                                 \
 463                 rcu_read_unlock();                              \
 464         }                                                       \
 465 } while (0)
 466
 467 #define blk_mq_run_dispatch_ops(q, dispatch_ops)                \
 468         __blk_mq_run_dispatch_ops(q, true, dispatch_ops)        \
 469
 470 #endif