block/blk-mq.c

   1 /*
   2  * Block multiqueue core code
   3  *
   4  * Copyright (C) 2013-2014 Jens Axboe
   5  * Copyright (C) 2013-2014 Christoph Hellwig
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/kmemleak.h>
  13 #include <linux/mm.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/smp.h>
  18 #include <linux/llist.h>
  19 #include <linux/list_sort.h>
  20 #include <linux/cpu.h>
  21 #include <linux/cache.h>
  22 #include <linux/sched/sysctl.h>
  23 #include <linux/sched/topology.h>
  24 #include <linux/sched/signal.h>
  25 #include <linux/delay.h>
  26 #include <linux/crash_dump.h>
  27 #include <linux/prefetch.h>
  28
  29 #include <trace/events/block.h>
  30
  31 #include <linux/blk-mq.h>
  32 #include "blk.h"
  33 #include "blk-mq.h"
  34 #include "blk-mq-debugfs.h"
  35 #include "blk-mq-tag.h"
  36 #include "blk-pm.h"
  37 #include "blk-stat.h"
  38 #include "blk-mq-sched.h"
  39 #include "blk-rq-qos.h"
  40
  41 static void blk_mq_poll_stats_start(struct request_queue *q);
  42 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  43
  44 static int blk_mq_poll_stats_bkt(const struct request *rq)
  45 {
  46         int ddir, bytes, bucket;
  47
  48         ddir = rq_data_dir(rq);
  49         bytes = blk_rq_bytes(rq);
  50
  51         bucket = ddir + 2*(ilog2(bytes) - 9);
  52
  53         if (bucket < 0)
  54                 return -1;
  55         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  56                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  57
  58         return bucket;
  59 }
  60
  61 /*
  62  * Check if any of the ctx's have pending work in this hardware queue
  63  */
  64 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  65 {
  66         return !list_empty_careful(&hctx->dispatch) ||
  67                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  68                         blk_mq_sched_has_work(hctx);
  69 }
  70
  71 /*
  72  * Mark this ctx as having pending work in this hardware queue
  73  */
  74 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  75                                      struct blk_mq_ctx *ctx)
  76 {
  77         const int bit = ctx->index_hw[hctx->type];
  78
  79         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
  80                 sbitmap_set_bit(&hctx->ctx_map, bit);
  81 }
  82
  83 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  84                                       struct blk_mq_ctx *ctx)
  85 {
  86         const int bit = ctx->index_hw[hctx->type];
  87
  88         sbitmap_clear_bit(&hctx->ctx_map, bit);
  89 }
  90
  91 struct mq_inflight {
  92         struct hd_struct *part;
  93         unsigned int *inflight;
  94 };
  95
  96 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  97                                   struct request *rq, void *priv,
  98                                   bool reserved)
  99 {
 100         struct mq_inflight *mi = priv;
 101
 102         /*
 103          * index[0] counts the specific partition that was asked for.
 104          */
 105         if (rq->part == mi->part)
 106                 mi->inflight[0]++;
 107
 108         return true;
 109 }
 110
 111 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
 112 {
 113         unsigned inflight[2];
 114         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 115
 116         inflight[0] = inflight[1] = 0;
 117         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 118
 119         return inflight[0];
 120 }
 121
 122 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
 123                                      struct request *rq, void *priv,
 124                                      bool reserved)
 125 {
 126         struct mq_inflight *mi = priv;
 127
 128         if (rq->part == mi->part)
 129                 mi->inflight[rq_data_dir(rq)]++;
 130
 131         return true;
 132 }
 133
 134 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 135                          unsigned int inflight[2])
 136 {
 137         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 138
 139         inflight[0] = inflight[1] = 0;
 140         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
 141 }
 142
 143 void blk_freeze_queue_start(struct request_queue *q)
 144 {
 145         int freeze_depth;
 146
 147         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 148         if (freeze_depth == 1) {
 149                 percpu_ref_kill(&q->q_usage_counter);
 150                 if (queue_is_mq(q))
 151                         blk_mq_run_hw_queues(q, false);
 152         }
 153 }
 154 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 155
 156 void blk_mq_freeze_queue_wait(struct request_queue *q)
 157 {
 158         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 159 }
 160 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 161
 162 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 163                                      unsigned long timeout)
 164 {
 165         return wait_event_timeout(q->mq_freeze_wq,
 166                                         percpu_ref_is_zero(&q->q_usage_counter),
 167                                         timeout);
 168 }
 169 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 170
 171 /*
 172  * Guarantee no request is in use, so we can change any data structure of
 173  * the queue afterward.
 174  */
 175 void blk_freeze_queue(struct request_queue *q)
 176 {
 177         /*
 178          * In the !blk_mq case we are only calling this to kill the
 179          * q_usage_counter, otherwise this increases the freeze depth
 180          * and waits for it to return to zero.  For this reason there is
 181          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 182          * exported to drivers as the only user for unfreeze is blk_mq.
 183          */
 184         blk_freeze_queue_start(q);
 185         blk_mq_freeze_queue_wait(q);
 186 }
 187
 188 void blk_mq_freeze_queue(struct request_queue *q)
 189 {
 190         /*
 191          * ...just an alias to keep freeze and unfreeze actions balanced
 192          * in the blk_mq_* namespace
 193          */
 194         blk_freeze_queue(q);
 195 }
 196 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 197
 198 void blk_mq_unfreeze_queue(struct request_queue *q)
 199 {
 200         int freeze_depth;
 201
 202         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 203         WARN_ON_ONCE(freeze_depth < 0);
 204         if (!freeze_depth) {
 205                 percpu_ref_resurrect(&q->q_usage_counter);
 206                 wake_up_all(&q->mq_freeze_wq);
 207         }
 208 }
 209 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 210
 211 /*
 212  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 213  * mpt3sas driver such that this function can be removed.
 214  */
 215 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 216 {
 217         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 218 }
 219 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 220
 221 /**
 222  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 223  * @q: request queue.
 224  *
 225  * Note: this function does not prevent that the struct request end_io()
 226  * callback function is invoked. Once this function is returned, we make
 227  * sure no dispatch can happen until the queue is unquiesced via
 228  * blk_mq_unquiesce_queue().
 229  */
 230 void blk_mq_quiesce_queue(struct request_queue *q)
 231 {
 232         struct blk_mq_hw_ctx *hctx;
 233         unsigned int i;
 234         bool rcu = false;
 235
 236         blk_mq_quiesce_queue_nowait(q);
 237
 238         queue_for_each_hw_ctx(q, hctx, i) {
 239                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 240                         synchronize_srcu(hctx->srcu);
 241                 else
 242                         rcu = true;
 243         }
 244         if (rcu)
 245                 synchronize_rcu();
 246 }
 247 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 248
 249 /*
 250  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 251  * @q: request queue.
 252  *
 253  * This function recovers queue into the state before quiescing
 254  * which is done by blk_mq_quiesce_queue.
 255  */
 256 void blk_mq_unquiesce_queue(struct request_queue *q)
 257 {
 258         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 259
 260         /* dispatch requests which are inserted during quiescing */
 261         blk_mq_run_hw_queues(q, true);
 262 }
 263 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 264
 265 void blk_mq_wake_waiters(struct request_queue *q)
 266 {
 267         struct blk_mq_hw_ctx *hctx;
 268         unsigned int i;
 269
 270         queue_for_each_hw_ctx(q, hctx, i)
 271                 if (blk_mq_hw_queue_mapped(hctx))
 272                         blk_mq_tag_wakeup_all(hctx->tags, true);
 273 }
 274
 275 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 276 {
 277         return blk_mq_has_free_tags(hctx->tags);
 278 }
 279 EXPORT_SYMBOL(blk_mq_can_queue);
 280
 281 /*
 282  * Only need start/end time stamping if we have stats enabled, or using
 283  * an IO scheduler.
 284  */
 285 static inline bool blk_mq_need_time_stamp(struct request *rq)
 286 {
 287         return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
 288 }
 289
 290 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 291                 unsigned int tag, unsigned int op)
 292 {
 293         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 294         struct request *rq = tags->static_rqs[tag];
 295         req_flags_t rq_flags = 0;
 296
 297         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 298                 rq->tag = -1;
 299                 rq->internal_tag = tag;
 300         } else {
 301                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
 302                         rq_flags = RQF_MQ_INFLIGHT;
 303                         atomic_inc(&data->hctx->nr_active);
 304                 }
 305                 rq->tag = tag;
 306                 rq->internal_tag = -1;
 307                 data->hctx->tags->rqs[rq->tag] = rq;
 308         }
 309
 310         /* csd/requeue_work/fifo_time is initialized before use */
 311         rq->q = data->q;
 312         rq->mq_ctx = data->ctx;
 313         rq->mq_hctx = data->hctx;
 314         rq->rq_flags = rq_flags;
 315         rq->cmd_flags = op;
 316         if (data->flags & BLK_MQ_REQ_PREEMPT)
 317                 rq->rq_flags |= RQF_PREEMPT;
 318         if (blk_queue_io_stat(data->q))
 319                 rq->rq_flags |= RQF_IO_STAT;
 320         INIT_LIST_HEAD(&rq->queuelist);
 321         INIT_HLIST_NODE(&rq->hash);
 322         RB_CLEAR_NODE(&rq->rb_node);
 323         rq->rq_disk = NULL;
 324         rq->part = NULL;
 325         if (blk_mq_need_time_stamp(rq))
 326                 rq->start_time_ns = ktime_get_ns();
 327         else
 328                 rq->start_time_ns = 0;
 329         rq->io_start_time_ns = 0;
 330         rq->nr_phys_segments = 0;
 331 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 332         rq->nr_integrity_segments = 0;
 333 #endif
 334         rq->special = NULL;
 335         /* tag was already set */
 336         rq->extra_len = 0;
 337         WRITE_ONCE(rq->deadline, 0);
 338
 339         rq->timeout = 0;
 340
 341         rq->end_io = NULL;
 342         rq->end_io_data = NULL;
 343         rq->next_rq = NULL;
 344
 345         data->ctx->rq_dispatched[op_is_sync(op)]++;
 346         refcount_set(&rq->ref, 1);
 347         return rq;
 348 }
 349
 350 static struct request *blk_mq_get_request(struct request_queue *q,
 351                                           struct bio *bio,
 352                                           struct blk_mq_alloc_data *data)
 353 {
 354         struct elevator_queue *e = q->elevator;
 355         struct request *rq;
 356         unsigned int tag;
 357         bool put_ctx_on_error = false;
 358
 359         blk_queue_enter_live(q);
 360         data->q = q;
 361         if (likely(!data->ctx)) {
 362                 data->ctx = blk_mq_get_ctx(q);
 363                 put_ctx_on_error = true;
 364         }
 365         if (likely(!data->hctx))
 366                 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
 367                                                 data->ctx->cpu);
 368         if (data->cmd_flags & REQ_NOWAIT)
 369                 data->flags |= BLK_MQ_REQ_NOWAIT;
 370
 371         if (e) {
 372                 data->flags |= BLK_MQ_REQ_INTERNAL;
 373
 374                 /*
 375                  * Flush requests are special and go directly to the
 376                  * dispatch list. Don't include reserved tags in the
 377                  * limiting, as it isn't useful.
 378                  */
 379                 if (!op_is_flush(data->cmd_flags) &&
 380                     e->type->ops.limit_depth &&
 381                     !(data->flags & BLK_MQ_REQ_RESERVED))
 382                         e->type->ops.limit_depth(data->cmd_flags, data);
 383         } else {
 384                 blk_mq_tag_busy(data->hctx);
 385         }
 386
 387         tag = blk_mq_get_tag(data);
 388         if (tag == BLK_MQ_TAG_FAIL) {
 389                 if (put_ctx_on_error) {
 390                         blk_mq_put_ctx(data->ctx);
 391                         data->ctx = NULL;
 392                 }
 393                 blk_queue_exit(q);
 394                 return NULL;
 395         }
 396
 397         rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
 398         if (!op_is_flush(data->cmd_flags)) {
 399                 rq->elv.icq = NULL;
 400                 if (e && e->type->ops.prepare_request) {
 401                         if (e->type->icq_cache)
 402                                 blk_mq_sched_assign_ioc(rq);
 403
 404                         e->type->ops.prepare_request(rq, bio);
 405                         rq->rq_flags |= RQF_ELVPRIV;
 406                 }
 407         }
 408         data->hctx->queued++;
 409         return rq;
 410 }
 411
 412 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 413                 blk_mq_req_flags_t flags)
 414 {
 415         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 416         struct request *rq;
 417         int ret;
 418
 419         ret = blk_queue_enter(q, flags);
 420         if (ret)
 421                 return ERR_PTR(ret);
 422
 423         rq = blk_mq_get_request(q, NULL, &alloc_data);
 424         blk_queue_exit(q);
 425
 426         if (!rq)
 427                 return ERR_PTR(-EWOULDBLOCK);
 428
 429         blk_mq_put_ctx(alloc_data.ctx);
 430
 431         rq->__data_len = 0;
 432         rq->__sector = (sector_t) -1;
 433         rq->bio = rq->biotail = NULL;
 434         return rq;
 435 }
 436 EXPORT_SYMBOL(blk_mq_alloc_request);
 437
 438 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 439         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 440 {
 441         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 442         struct request *rq;
 443         unsigned int cpu;
 444         int ret;
 445
 446         /*
 447          * If the tag allocator sleeps we could get an allocation for a
 448          * different hardware context.  No need to complicate the low level
 449          * allocator for this for the rare use case of a command tied to
 450          * a specific queue.
 451          */
 452         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 453                 return ERR_PTR(-EINVAL);
 454
 455         if (hctx_idx >= q->nr_hw_queues)
 456                 return ERR_PTR(-EIO);
 457
 458         ret = blk_queue_enter(q, flags);
 459         if (ret)
 460                 return ERR_PTR(ret);
 461
 462         /*
 463          * Check if the hardware context is actually mapped to anything.
 464          * If not tell the caller that it should skip this queue.
 465          */
 466         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 467         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 468                 blk_queue_exit(q);
 469                 return ERR_PTR(-EXDEV);
 470         }
 471         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 472         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 473
 474         rq = blk_mq_get_request(q, NULL, &alloc_data);
 475         blk_queue_exit(q);
 476
 477         if (!rq)
 478                 return ERR_PTR(-EWOULDBLOCK);
 479
 480         return rq;
 481 }
 482 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 483
 484 static void __blk_mq_free_request(struct request *rq)
 485 {
 486         struct request_queue *q = rq->q;
 487         struct blk_mq_ctx *ctx = rq->mq_ctx;
 488         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 489         const int sched_tag = rq->internal_tag;
 490
 491         blk_pm_mark_last_busy(rq);
 492         rq->mq_hctx = NULL;
 493         if (rq->tag != -1)
 494                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 495         if (sched_tag != -1)
 496                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 497         blk_mq_sched_restart(hctx);
 498         blk_queue_exit(q);
 499 }
 500
 501 void blk_mq_free_request(struct request *rq)
 502 {
 503         struct request_queue *q = rq->q;
 504         struct elevator_queue *e = q->elevator;
 505         struct blk_mq_ctx *ctx = rq->mq_ctx;
 506         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 507
 508         if (rq->rq_flags & RQF_ELVPRIV) {
 509                 if (e && e->type->ops.finish_request)
 510                         e->type->ops.finish_request(rq);
 511                 if (rq->elv.icq) {
 512                         put_io_context(rq->elv.icq->ioc);
 513                         rq->elv.icq = NULL;
 514                 }
 515         }
 516
 517         ctx->rq_completed[rq_is_sync(rq)]++;
 518         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 519                 atomic_dec(&hctx->nr_active);
 520
 521         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 522                 laptop_io_completion(q->backing_dev_info);
 523
 524         rq_qos_done(q, rq);
 525
 526         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 527         if (refcount_dec_and_test(&rq->ref))
 528                 __blk_mq_free_request(rq);
 529 }
 530 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 531
 532 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 533 {
 534         u64 now = 0;
 535
 536         if (blk_mq_need_time_stamp(rq))
 537                 now = ktime_get_ns();
 538
 539         if (rq->rq_flags & RQF_STATS) {
 540                 blk_mq_poll_stats_start(rq->q);
 541                 blk_stat_add(rq, now);
 542         }
 543
 544         if (rq->internal_tag != -1)
 545                 blk_mq_sched_completed_request(rq, now);
 546
 547         blk_account_io_done(rq, now);
 548
 549         if (rq->end_io) {
 550                 rq_qos_done(rq->q, rq);
 551                 rq->end_io(rq, error);
 552         } else {
 553                 if (unlikely(blk_bidi_rq(rq)))
 554                         blk_mq_free_request(rq->next_rq);
 555                 blk_mq_free_request(rq);
 556         }
 557 }
 558 EXPORT_SYMBOL(__blk_mq_end_request);
 559
 560 void blk_mq_end_request(struct request *rq, blk_status_t error)
 561 {
 562         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 563                 BUG();
 564         __blk_mq_end_request(rq, error);
 565 }
 566 EXPORT_SYMBOL(blk_mq_end_request);
 567
 568 static void __blk_mq_complete_request_remote(void *data)
 569 {
 570         struct request *rq = data;
 571         struct request_queue *q = rq->q;
 572
 573         q->mq_ops->complete(rq);
 574 }
 575
 576 static void __blk_mq_complete_request(struct request *rq)
 577 {
 578         struct blk_mq_ctx *ctx = rq->mq_ctx;
 579         struct request_queue *q = rq->q;
 580         bool shared = false;
 581         int cpu;
 582
 583         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 584         /*
 585          * Most of single queue controllers, there is only one irq vector
 586          * for handling IO completion, and the only irq's affinity is set
 587          * as all possible CPUs. On most of ARCHs, this affinity means the
 588          * irq is handled on one specific CPU.
 589          *
 590          * So complete IO reqeust in softirq context in case of single queue
 591          * for not degrading IO performance by irqsoff latency.
 592          */
 593         if (q->nr_hw_queues == 1) {
 594                 __blk_complete_request(rq);
 595                 return;
 596         }
 597
 598         /*
 599          * For a polled request, always complete locallly, it's pointless
 600          * to redirect the completion.
 601          */
 602         if ((rq->cmd_flags & REQ_HIPRI) ||
 603             !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
 604                 q->mq_ops->complete(rq);
 605                 return;
 606         }
 607
 608         cpu = get_cpu();
 609         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
 610                 shared = cpus_share_cache(cpu, ctx->cpu);
 611
 612         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 613                 rq->csd.func = __blk_mq_complete_request_remote;
 614                 rq->csd.info = rq;
 615                 rq->csd.flags = 0;
 616                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 617         } else {
 618                 q->mq_ops->complete(rq);
 619         }
 620         put_cpu();
 621 }
 622
 623 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 624         __releases(hctx->srcu)
 625 {
 626         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 627                 rcu_read_unlock();
 628         else
 629                 srcu_read_unlock(hctx->srcu, srcu_idx);
 630 }
 631
 632 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 633         __acquires(hctx->srcu)
 634 {
 635         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 636                 /* shut up gcc false positive */
 637                 *srcu_idx = 0;
 638                 rcu_read_lock();
 639         } else
 640                 *srcu_idx = srcu_read_lock(hctx->srcu);
 641 }
 642
 643 /**
 644  * blk_mq_complete_request - end I/O on a request
 645  * @rq:         the request being processed
 646  *
 647  * Description:
 648  *      Ends all I/O on a request. It does not handle partial completions.
 649  *      The actual completion happens out-of-order, through a IPI handler.
 650  **/
 651 bool blk_mq_complete_request(struct request *rq)
 652 {
 653         if (unlikely(blk_should_fake_timeout(rq->q)))
 654                 return false;
 655         __blk_mq_complete_request(rq);
 656         return true;
 657 }
 658 EXPORT_SYMBOL(blk_mq_complete_request);
 659
 660 int blk_mq_request_started(struct request *rq)
 661 {
 662         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 663 }
 664 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 665
 666 void blk_mq_start_request(struct request *rq)
 667 {
 668         struct request_queue *q = rq->q;
 669
 670         blk_mq_sched_started_request(rq);
 671
 672         trace_block_rq_issue(q, rq);
 673
 674         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 675                 rq->io_start_time_ns = ktime_get_ns();
 676 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 677                 rq->throtl_size = blk_rq_sectors(rq);
 678 #endif
 679                 rq->rq_flags |= RQF_STATS;
 680                 rq_qos_issue(q, rq);
 681         }
 682
 683         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 684
 685         blk_add_timer(rq);
 686         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 687
 688         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 689                 /*
 690                  * Make sure space for the drain appears.  We know we can do
 691                  * this because max_hw_segments has been adjusted to be one
 692                  * fewer than the device can handle.
 693                  */
 694                 rq->nr_phys_segments++;
 695         }
 696 }
 697 EXPORT_SYMBOL(blk_mq_start_request);
 698
 699 static void __blk_mq_requeue_request(struct request *rq)
 700 {
 701         struct request_queue *q = rq->q;
 702
 703         blk_mq_put_driver_tag(rq);
 704
 705         trace_block_rq_requeue(q, rq);
 706         rq_qos_requeue(q, rq);
 707
 708         if (blk_mq_request_started(rq)) {
 709                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 710                 rq->rq_flags &= ~RQF_TIMED_OUT;
 711                 if (q->dma_drain_size && blk_rq_bytes(rq))
 712                         rq->nr_phys_segments--;
 713         }
 714 }
 715
 716 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 717 {
 718         __blk_mq_requeue_request(rq);
 719
 720         /* this request will be re-inserted to io scheduler queue */
 721         blk_mq_sched_requeue_request(rq);
 722
 723         BUG_ON(!list_empty(&rq->queuelist));
 724         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 725 }
 726 EXPORT_SYMBOL(blk_mq_requeue_request);
 727
 728 static void blk_mq_requeue_work(struct work_struct *work)
 729 {
 730         struct request_queue *q =
 731                 container_of(work, struct request_queue, requeue_work.work);
 732         LIST_HEAD(rq_list);
 733         struct request *rq, *next;
 734
 735         spin_lock_irq(&q->requeue_lock);
 736         list_splice_init(&q->requeue_list, &rq_list);
 737         spin_unlock_irq(&q->requeue_lock);
 738
 739         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 740                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 741                         continue;
 742
 743                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 744                 list_del_init(&rq->queuelist);
 745                 blk_mq_sched_insert_request(rq, true, false, false);
 746         }
 747
 748         while (!list_empty(&rq_list)) {
 749                 rq = list_entry(rq_list.next, struct request, queuelist);
 750                 list_del_init(&rq->queuelist);
 751                 blk_mq_sched_insert_request(rq, false, false, false);
 752         }
 753
 754         blk_mq_run_hw_queues(q, false);
 755 }
 756
 757 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 758                                 bool kick_requeue_list)
 759 {
 760         struct request_queue *q = rq->q;
 761         unsigned long flags;
 762
 763         /*
 764          * We abuse this flag that is otherwise used by the I/O scheduler to
 765          * request head insertion from the workqueue.
 766          */
 767         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 768
 769         spin_lock_irqsave(&q->requeue_lock, flags);
 770         if (at_head) {
 771                 rq->rq_flags |= RQF_SOFTBARRIER;
 772                 list_add(&rq->queuelist, &q->requeue_list);
 773         } else {
 774                 list_add_tail(&rq->queuelist, &q->requeue_list);
 775         }
 776         spin_unlock_irqrestore(&q->requeue_lock, flags);
 777
 778         if (kick_requeue_list)
 779                 blk_mq_kick_requeue_list(q);
 780 }
 781 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 782
 783 void blk_mq_kick_requeue_list(struct request_queue *q)
 784 {
 785         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 786 }
 787 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 788
 789 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 790                                     unsigned long msecs)
 791 {
 792         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 793                                     msecs_to_jiffies(msecs));
 794 }
 795 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 796
 797 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 798 {
 799         if (tag < tags->nr_tags) {
 800                 prefetch(tags->rqs[tag]);
 801                 return tags->rqs[tag];
 802         }
 803
 804         return NULL;
 805 }
 806 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 807
 808 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 809                                void *priv, bool reserved)
 810 {
 811         /*
 812          * If we find a request that is inflight and the queue matches,
 813          * we know the queue is busy. Return false to stop the iteration.
 814          */
 815         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 816                 bool *busy = priv;
 817
 818                 *busy = true;
 819                 return false;
 820         }
 821
 822         return true;
 823 }
 824
 825 bool blk_mq_queue_inflight(struct request_queue *q)
 826 {
 827         bool busy = false;
 828
 829         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 830         return busy;
 831 }
 832 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 833
 834 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 835 {
 836         req->rq_flags |= RQF_TIMED_OUT;
 837         if (req->q->mq_ops->timeout) {
 838                 enum blk_eh_timer_return ret;
 839
 840                 ret = req->q->mq_ops->timeout(req, reserved);
 841                 if (ret == BLK_EH_DONE)
 842                         return;
 843                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 844         }
 845
 846         blk_add_timer(req);
 847 }
 848
 849 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 850 {
 851         unsigned long deadline;
 852
 853         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 854                 return false;
 855         if (rq->rq_flags & RQF_TIMED_OUT)
 856                 return false;
 857
 858         deadline = READ_ONCE(rq->deadline);
 859         if (time_after_eq(jiffies, deadline))
 860                 return true;
 861
 862         if (*next == 0)
 863                 *next = deadline;
 864         else if (time_after(*next, deadline))
 865                 *next = deadline;
 866         return false;
 867 }
 868
 869 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 870                 struct request *rq, void *priv, bool reserved)
 871 {
 872         unsigned long *next = priv;
 873
 874         /*
 875          * Just do a quick check if it is expired before locking the request in
 876          * so we're not unnecessarilly synchronizing across CPUs.
 877          */
 878         if (!blk_mq_req_expired(rq, next))
 879                 return true;
 880
 881         /*
 882          * We have reason to believe the request may be expired. Take a
 883          * reference on the request to lock this request lifetime into its
 884          * currently allocated context to prevent it from being reallocated in
 885          * the event the completion by-passes this timeout handler.
 886          *
 887          * If the reference was already released, then the driver beat the
 888          * timeout handler to posting a natural completion.
 889          */
 890         if (!refcount_inc_not_zero(&rq->ref))
 891                 return true;
 892
 893         /*
 894          * The request is now locked and cannot be reallocated underneath the
 895          * timeout handler's processing. Re-verify this exact request is truly
 896          * expired; if it is not expired, then the request was completed and
 897          * reallocated as a new request.
 898          */
 899         if (blk_mq_req_expired(rq, next))
 900                 blk_mq_rq_timed_out(rq, reserved);
 901         if (refcount_dec_and_test(&rq->ref))
 902                 __blk_mq_free_request(rq);
 903
 904         return true;
 905 }
 906
 907 static void blk_mq_timeout_work(struct work_struct *work)
 908 {
 909         struct request_queue *q =
 910                 container_of(work, struct request_queue, timeout_work);
 911         unsigned long next = 0;
 912         struct blk_mq_hw_ctx *hctx;
 913         int i;
 914
 915         /* A deadlock might occur if a request is stuck requiring a
 916          * timeout at the same time a queue freeze is waiting
 917          * completion, since the timeout code would not be able to
 918          * acquire the queue reference here.
 919          *
 920          * That's why we don't use blk_queue_enter here; instead, we use
 921          * percpu_ref_tryget directly, because we need to be able to
 922          * obtain a reference even in the short window between the queue
 923          * starting to freeze, by dropping the first reference in
 924          * blk_freeze_queue_start, and the moment the last request is
 925          * consumed, marked by the instant q_usage_counter reaches
 926          * zero.
 927          */
 928         if (!percpu_ref_tryget(&q->q_usage_counter))
 929                 return;
 930
 931         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 932
 933         if (next != 0) {
 934                 mod_timer(&q->timeout, next);
 935         } else {
 936                 /*
 937                  * Request timeouts are handled as a forward rolling timer. If
 938                  * we end up here it means that no requests are pending and
 939                  * also that no request has been pending for a while. Mark
 940                  * each hctx as idle.
 941                  */
 942                 queue_for_each_hw_ctx(q, hctx, i) {
 943                         /* the hctx may be unmapped, so check it here */
 944                         if (blk_mq_hw_queue_mapped(hctx))
 945                                 blk_mq_tag_idle(hctx);
 946                 }
 947         }
 948         blk_queue_exit(q);
 949 }
 950
 951 struct flush_busy_ctx_data {
 952         struct blk_mq_hw_ctx *hctx;
 953         struct list_head *list;
 954 };
 955
 956 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 957 {
 958         struct flush_busy_ctx_data *flush_data = data;
 959         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 960         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 961         enum hctx_type type = hctx->type;
 962
 963         spin_lock(&ctx->lock);
 964         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 965         sbitmap_clear_bit(sb, bitnr);
 966         spin_unlock(&ctx->lock);
 967         return true;
 968 }
 969
 970 /*
 971  * Process software queues that have been marked busy, splicing them
 972  * to the for-dispatch
 973  */
 974 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 975 {
 976         struct flush_busy_ctx_data data = {
 977                 .hctx = hctx,
 978                 .list = list,
 979         };
 980
 981         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 982 }
 983 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 984
 985 struct dispatch_rq_data {
 986         struct blk_mq_hw_ctx *hctx;
 987         struct request *rq;
 988 };
 989
 990 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
 991                 void *data)
 992 {
 993         struct dispatch_rq_data *dispatch_data = data;
 994         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
 995         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 996         enum hctx_type type = hctx->type;
 997
 998         spin_lock(&ctx->lock);
 999         if (!list_empty(&ctx->rq_lists[type])) {
1000                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1001                 list_del_init(&dispatch_data->rq->queuelist);
1002                 if (list_empty(&ctx->rq_lists[type]))
1003                         sbitmap_clear_bit(sb, bitnr);
1004         }
1005         spin_unlock(&ctx->lock);
1006
1007         return !dispatch_data->rq;
1008 }
1009
1010 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1011                                         struct blk_mq_ctx *start)
1012 {
1013         unsigned off = start ? start->index_hw[hctx->type] : 0;
1014         struct dispatch_rq_data data = {
1015                 .hctx = hctx,
1016                 .rq   = NULL,
1017         };
1018
1019         __sbitmap_for_each_set(&hctx->ctx_map, off,
1020                                dispatch_rq_from_ctx, &data);
1021
1022         return data.rq;
1023 }
1024
1025 static inline unsigned int queued_to_index(unsigned int queued)
1026 {
1027         if (!queued)
1028                 return 0;
1029
1030         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1031 }
1032
1033 bool blk_mq_get_driver_tag(struct request *rq)
1034 {
1035         struct blk_mq_alloc_data data = {
1036                 .q = rq->q,
1037                 .hctx = rq->mq_hctx,
1038                 .flags = BLK_MQ_REQ_NOWAIT,
1039                 .cmd_flags = rq->cmd_flags,
1040         };
1041         bool shared;
1042
1043         if (rq->tag != -1)
1044                 goto done;
1045
1046         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1047                 data.flags |= BLK_MQ_REQ_RESERVED;
1048
1049         shared = blk_mq_tag_busy(data.hctx);
1050         rq->tag = blk_mq_get_tag(&data);
1051         if (rq->tag >= 0) {
1052                 if (shared) {
1053                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1054                         atomic_inc(&data.hctx->nr_active);
1055                 }
1056                 data.hctx->tags->rqs[rq->tag] = rq;
1057         }
1058
1059 done:
1060         return rq->tag != -1;
1061 }
1062
1063 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1064                                 int flags, void *key)
1065 {
1066         struct blk_mq_hw_ctx *hctx;
1067
1068         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1069
1070         spin_lock(&hctx->dispatch_wait_lock);
1071         list_del_init(&wait->entry);
1072         spin_unlock(&hctx->dispatch_wait_lock);
1073
1074         blk_mq_run_hw_queue(hctx, true);
1075         return 1;
1076 }
1077
1078 /*
1079  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1080  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1081  * restart. For both cases, take care to check the condition again after
1082  * marking us as waiting.
1083  */
1084 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1085                                  struct request *rq)
1086 {
1087         struct wait_queue_head *wq;
1088         wait_queue_entry_t *wait;
1089         bool ret;
1090
1091         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1092                 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
1093                         set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
1094
1095                 /*
1096                  * It's possible that a tag was freed in the window between the
1097                  * allocation failure and adding the hardware queue to the wait
1098                  * queue.
1099                  *
1100                  * Don't clear RESTART here, someone else could have set it.
1101                  * At most this will cost an extra queue run.
1102                  */
1103                 return blk_mq_get_driver_tag(rq);
1104         }
1105
1106         wait = &hctx->dispatch_wait;
1107         if (!list_empty_careful(&wait->entry))
1108                 return false;
1109
1110         wq = &bt_wait_ptr(&hctx->tags->bitmap_tags, hctx)->wait;
1111
1112         spin_lock_irq(&wq->lock);
1113         spin_lock(&hctx->dispatch_wait_lock);
1114         if (!list_empty(&wait->entry)) {
1115                 spin_unlock(&hctx->dispatch_wait_lock);
1116                 spin_unlock_irq(&wq->lock);
1117                 return false;
1118         }
1119
1120         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1121         __add_wait_queue(wq, wait);
1122
1123         /*
1124          * It's possible that a tag was freed in the window between the
1125          * allocation failure and adding the hardware queue to the wait
1126          * queue.
1127          */
1128         ret = blk_mq_get_driver_tag(rq);
1129         if (!ret) {
1130                 spin_unlock(&hctx->dispatch_wait_lock);
1131                 spin_unlock_irq(&wq->lock);
1132                 return false;
1133         }
1134
1135         /*
1136          * We got a tag, remove ourselves from the wait queue to ensure
1137          * someone else gets the wakeup.
1138          */
1139         list_del_init(&wait->entry);
1140         spin_unlock(&hctx->dispatch_wait_lock);
1141         spin_unlock_irq(&wq->lock);
1142
1143         return true;
1144 }
1145
1146 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1147 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1148 /*
1149  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1150  * - EWMA is one simple way to compute running average value
1151  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1152  * - take 4 as factor for avoiding to get too small(0) result, and this
1153  *   factor doesn't matter because EWMA decreases exponentially
1154  */
1155 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1156 {
1157         unsigned int ewma;
1158
1159         if (hctx->queue->elevator)
1160                 return;
1161
1162         ewma = hctx->dispatch_busy;
1163
1164         if (!ewma && !busy)
1165                 return;
1166
1167         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1168         if (busy)
1169                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1170         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1171
1172         hctx->dispatch_busy = ewma;
1173 }
1174
1175 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1176
1177 /*
1178  * Returns true if we did some work AND can potentially do more.
1179  */
1180 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1181                              bool got_budget)
1182 {
1183         struct blk_mq_hw_ctx *hctx;
1184         struct request *rq, *nxt;
1185         bool no_tag = false;
1186         int errors, queued;
1187         blk_status_t ret = BLK_STS_OK;
1188
1189         if (list_empty(list))
1190                 return false;
1191
1192         WARN_ON(!list_is_singular(list) && got_budget);
1193
1194         /*
1195          * Now process all the entries, sending them to the driver.
1196          */
1197         errors = queued = 0;
1198         do {
1199                 struct blk_mq_queue_data bd;
1200
1201                 rq = list_first_entry(list, struct request, queuelist);
1202
1203                 hctx = rq->mq_hctx;
1204                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1205                         break;
1206
1207                 if (!blk_mq_get_driver_tag(rq)) {
1208                         /*
1209                          * The initial allocation attempt failed, so we need to
1210                          * rerun the hardware queue when a tag is freed. The
1211                          * waitqueue takes care of that. If the queue is run
1212                          * before we add this entry back on the dispatch list,
1213                          * we'll re-run it below.
1214                          */
1215                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1216                                 blk_mq_put_dispatch_budget(hctx);
1217                                 /*
1218                                  * For non-shared tags, the RESTART check
1219                                  * will suffice.
1220                                  */
1221                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1222                                         no_tag = true;
1223                                 break;
1224                         }
1225                 }
1226
1227                 list_del_init(&rq->queuelist);
1228
1229                 bd.rq = rq;
1230
1231                 /*
1232                  * Flag last if we have no more requests, or if we have more
1233                  * but can't assign a driver tag to it.
1234                  */
1235                 if (list_empty(list))
1236                         bd.last = true;
1237                 else {
1238                         nxt = list_first_entry(list, struct request, queuelist);
1239                         bd.last = !blk_mq_get_driver_tag(nxt);
1240                 }
1241
1242                 ret = q->mq_ops->queue_rq(hctx, &bd);
1243                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1244                         /*
1245                          * If an I/O scheduler has been configured and we got a
1246                          * driver tag for the next request already, free it
1247                          * again.
1248                          */
1249                         if (!list_empty(list)) {
1250                                 nxt = list_first_entry(list, struct request, queuelist);
1251                                 blk_mq_put_driver_tag(nxt);
1252                         }
1253                         list_add(&rq->queuelist, list);
1254                         __blk_mq_requeue_request(rq);
1255                         break;
1256                 }
1257
1258                 if (unlikely(ret != BLK_STS_OK)) {
1259                         errors++;
1260                         blk_mq_end_request(rq, BLK_STS_IOERR);
1261                         continue;
1262                 }
1263
1264                 queued++;
1265         } while (!list_empty(list));
1266
1267         hctx->dispatched[queued_to_index(queued)]++;
1268
1269         /*
1270          * Any items that need requeuing? Stuff them into hctx->dispatch,
1271          * that is where we will continue on next queue run.
1272          */
1273         if (!list_empty(list)) {
1274                 bool needs_restart;
1275
1276                 /*
1277                  * If we didn't flush the entire list, we could have told
1278                  * the driver there was more coming, but that turned out to
1279                  * be a lie.
1280                  */
1281                 if (q->mq_ops->commit_rqs)
1282                         q->mq_ops->commit_rqs(hctx);
1283
1284                 spin_lock(&hctx->lock);
1285                 list_splice_init(list, &hctx->dispatch);
1286                 spin_unlock(&hctx->lock);
1287
1288                 /*
1289                  * If SCHED_RESTART was set by the caller of this function and
1290                  * it is no longer set that means that it was cleared by another
1291                  * thread and hence that a queue rerun is needed.
1292                  *
1293                  * If 'no_tag' is set, that means that we failed getting
1294                  * a driver tag with an I/O scheduler attached. If our dispatch
1295                  * waitqueue is no longer active, ensure that we run the queue
1296                  * AFTER adding our entries back to the list.
1297                  *
1298                  * If no I/O scheduler has been configured it is possible that
1299                  * the hardware queue got stopped and restarted before requests
1300                  * were pushed back onto the dispatch list. Rerun the queue to
1301                  * avoid starvation. Notes:
1302                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1303                  *   been stopped before rerunning a queue.
1304                  * - Some but not all block drivers stop a queue before
1305                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1306                  *   and dm-rq.
1307                  *
1308                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1309                  * bit is set, run queue after a delay to avoid IO stalls
1310                  * that could otherwise occur if the queue is idle.
1311                  */
1312                 needs_restart = blk_mq_sched_needs_restart(hctx);
1313                 if (!needs_restart ||
1314                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1315                         blk_mq_run_hw_queue(hctx, true);
1316                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1317                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1318
1319                 blk_mq_update_dispatch_busy(hctx, true);
1320                 return false;
1321         } else
1322                 blk_mq_update_dispatch_busy(hctx, false);
1323
1324         /*
1325          * If the host/device is unable to accept more work, inform the
1326          * caller of that.
1327          */
1328         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1329                 return false;
1330
1331         return (queued + errors) != 0;
1332 }
1333
1334 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1335 {
1336         int srcu_idx;
1337
1338         /*
1339          * We should be running this queue from one of the CPUs that
1340          * are mapped to it.
1341          *
1342          * There are at least two related races now between setting
1343          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1344          * __blk_mq_run_hw_queue():
1345          *
1346          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1347          *   but later it becomes online, then this warning is harmless
1348          *   at all
1349          *
1350          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1351          *   but later it becomes offline, then the warning can't be
1352          *   triggered, and we depend on blk-mq timeout handler to
1353          *   handle dispatched requests to this hctx
1354          */
1355         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1356                 cpu_online(hctx->next_cpu)) {
1357                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1358                         raw_smp_processor_id(),
1359                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1360                 dump_stack();
1361         }
1362
1363         /*
1364          * We can't run the queue inline with ints disabled. Ensure that
1365          * we catch bad users of this early.
1366          */
1367         WARN_ON_ONCE(in_interrupt());
1368
1369         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1370
1371         hctx_lock(hctx, &srcu_idx);
1372         blk_mq_sched_dispatch_requests(hctx);
1373         hctx_unlock(hctx, srcu_idx);
1374 }
1375
1376 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1377 {
1378         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1379
1380         if (cpu >= nr_cpu_ids)
1381                 cpu = cpumask_first(hctx->cpumask);
1382         return cpu;
1383 }
1384
1385 /*
1386  * It'd be great if the workqueue API had a way to pass
1387  * in a mask and had some smarts for more clever placement.
1388  * For now we just round-robin here, switching for every
1389  * BLK_MQ_CPU_WORK_BATCH queued items.
1390  */
1391 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1392 {
1393         bool tried = false;
1394         int next_cpu = hctx->next_cpu;
1395
1396         if (hctx->queue->nr_hw_queues == 1)
1397                 return WORK_CPU_UNBOUND;
1398
1399         if (--hctx->next_cpu_batch <= 0) {
1400 select_cpu:
1401                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1402                                 cpu_online_mask);
1403                 if (next_cpu >= nr_cpu_ids)
1404                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1405                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1406         }
1407
1408         /*
1409          * Do unbound schedule if we can't find a online CPU for this hctx,
1410          * and it should only happen in the path of handling CPU DEAD.
1411          */
1412         if (!cpu_online(next_cpu)) {
1413                 if (!tried) {
1414                         tried = true;
1415                         goto select_cpu;
1416                 }
1417
1418                 /*
1419                  * Make sure to re-select CPU next time once after CPUs
1420                  * in hctx->cpumask become online again.
1421                  */
1422                 hctx->next_cpu = next_cpu;
1423                 hctx->next_cpu_batch = 1;
1424                 return WORK_CPU_UNBOUND;
1425         }
1426
1427         hctx->next_cpu = next_cpu;
1428         return next_cpu;
1429 }
1430
1431 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1432                                         unsigned long msecs)
1433 {
1434         if (unlikely(blk_mq_hctx_stopped(hctx)))
1435                 return;
1436
1437         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1438                 int cpu = get_cpu();
1439                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1440                         __blk_mq_run_hw_queue(hctx);
1441                         put_cpu();
1442                         return;
1443                 }
1444
1445                 put_cpu();
1446         }
1447
1448         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1449                                     msecs_to_jiffies(msecs));
1450 }
1451
1452 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1453 {
1454         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1455 }
1456 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1457
1458 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1459 {
1460         int srcu_idx;
1461         bool need_run;
1462
1463         /*
1464          * When queue is quiesced, we may be switching io scheduler, or
1465          * updating nr_hw_queues, or other things, and we can't run queue
1466          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1467          *
1468          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1469          * quiesced.
1470          */
1471         hctx_lock(hctx, &srcu_idx);
1472         need_run = !blk_queue_quiesced(hctx->queue) &&
1473                 blk_mq_hctx_has_pending(hctx);
1474         hctx_unlock(hctx, srcu_idx);
1475
1476         if (need_run) {
1477                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1478                 return true;
1479         }
1480
1481         return false;
1482 }
1483 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1484
1485 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1486 {
1487         struct blk_mq_hw_ctx *hctx;
1488         int i;
1489
1490         queue_for_each_hw_ctx(q, hctx, i) {
1491                 if (blk_mq_hctx_stopped(hctx))
1492                         continue;
1493
1494                 blk_mq_run_hw_queue(hctx, async);
1495         }
1496 }
1497 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1498
1499 /**
1500  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1501  * @q: request queue.
1502  *
1503  * The caller is responsible for serializing this function against
1504  * blk_mq_{start,stop}_hw_queue().
1505  */
1506 bool blk_mq_queue_stopped(struct request_queue *q)
1507 {
1508         struct blk_mq_hw_ctx *hctx;
1509         int i;
1510
1511         queue_for_each_hw_ctx(q, hctx, i)
1512                 if (blk_mq_hctx_stopped(hctx))
1513                         return true;
1514
1515         return false;
1516 }
1517 EXPORT_SYMBOL(blk_mq_queue_stopped);
1518
1519 /*
1520  * This function is often used for pausing .queue_rq() by driver when
1521  * there isn't enough resource or some conditions aren't satisfied, and
1522  * BLK_STS_RESOURCE is usually returned.
1523  *
1524  * We do not guarantee that dispatch can be drained or blocked
1525  * after blk_mq_stop_hw_queue() returns. Please use
1526  * blk_mq_quiesce_queue() for that requirement.
1527  */
1528 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1529 {
1530         cancel_delayed_work(&hctx->run_work);
1531
1532         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1533 }
1534 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1535
1536 /*
1537  * This function is often used for pausing .queue_rq() by driver when
1538  * there isn't enough resource or some conditions aren't satisfied, and
1539  * BLK_STS_RESOURCE is usually returned.
1540  *
1541  * We do not guarantee that dispatch can be drained or blocked
1542  * after blk_mq_stop_hw_queues() returns. Please use
1543  * blk_mq_quiesce_queue() for that requirement.
1544  */
1545 void blk_mq_stop_hw_queues(struct request_queue *q)
1546 {
1547         struct blk_mq_hw_ctx *hctx;
1548         int i;
1549
1550         queue_for_each_hw_ctx(q, hctx, i)
1551                 blk_mq_stop_hw_queue(hctx);
1552 }
1553 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1554
1555 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1556 {
1557         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1558
1559         blk_mq_run_hw_queue(hctx, false);
1560 }
1561 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1562
1563 void blk_mq_start_hw_queues(struct request_queue *q)
1564 {
1565         struct blk_mq_hw_ctx *hctx;
1566         int i;
1567
1568         queue_for_each_hw_ctx(q, hctx, i)
1569                 blk_mq_start_hw_queue(hctx);
1570 }
1571 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1572
1573 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1574 {
1575         if (!blk_mq_hctx_stopped(hctx))
1576                 return;
1577
1578         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1579         blk_mq_run_hw_queue(hctx, async);
1580 }
1581 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1582
1583 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1584 {
1585         struct blk_mq_hw_ctx *hctx;
1586         int i;
1587
1588         queue_for_each_hw_ctx(q, hctx, i)
1589                 blk_mq_start_stopped_hw_queue(hctx, async);
1590 }
1591 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1592
1593 static void blk_mq_run_work_fn(struct work_struct *work)
1594 {
1595         struct blk_mq_hw_ctx *hctx;
1596
1597         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1598
1599         /*
1600          * If we are stopped, don't run the queue.
1601          */
1602         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1603                 return;
1604
1605         __blk_mq_run_hw_queue(hctx);
1606 }
1607
1608 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1609                                             struct request *rq,
1610                                             bool at_head)
1611 {
1612         struct blk_mq_ctx *ctx = rq->mq_ctx;
1613         enum hctx_type type = hctx->type;
1614
1615         lockdep_assert_held(&ctx->lock);
1616
1617         trace_block_rq_insert(hctx->queue, rq);
1618
1619         if (at_head)
1620                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1621         else
1622                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1623 }
1624
1625 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1626                              bool at_head)
1627 {
1628         struct blk_mq_ctx *ctx = rq->mq_ctx;
1629
1630         lockdep_assert_held(&ctx->lock);
1631
1632         __blk_mq_insert_req_list(hctx, rq, at_head);
1633         blk_mq_hctx_mark_pending(hctx, ctx);
1634 }
1635
1636 /*
1637  * Should only be used carefully, when the caller knows we want to
1638  * bypass a potential IO scheduler on the target device.
1639  */
1640 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1641 {
1642         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1643
1644         spin_lock(&hctx->lock);
1645         list_add_tail(&rq->queuelist, &hctx->dispatch);
1646         spin_unlock(&hctx->lock);
1647
1648         if (run_queue)
1649                 blk_mq_run_hw_queue(hctx, false);
1650 }
1651
1652 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1653                             struct list_head *list)
1654
1655 {
1656         struct request *rq;
1657         enum hctx_type type = hctx->type;
1658
1659         /*
1660          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1661          * offline now
1662          */
1663         list_for_each_entry(rq, list, queuelist) {
1664                 BUG_ON(rq->mq_ctx != ctx);
1665                 trace_block_rq_insert(hctx->queue, rq);
1666         }
1667
1668         spin_lock(&ctx->lock);
1669         list_splice_tail_init(list, &ctx->rq_lists[type]);
1670         blk_mq_hctx_mark_pending(hctx, ctx);
1671         spin_unlock(&ctx->lock);
1672 }
1673
1674 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1675 {
1676         struct request *rqa = container_of(a, struct request, queuelist);
1677         struct request *rqb = container_of(b, struct request, queuelist);
1678
1679         if (rqa->mq_ctx < rqb->mq_ctx)
1680                 return -1;
1681         else if (rqa->mq_ctx > rqb->mq_ctx)
1682                 return 1;
1683         else if (rqa->mq_hctx < rqb->mq_hctx)
1684                 return -1;
1685         else if (rqa->mq_hctx > rqb->mq_hctx)
1686                 return 1;
1687
1688         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1689 }
1690
1691 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1692 {
1693         struct blk_mq_hw_ctx *this_hctx;
1694         struct blk_mq_ctx *this_ctx;
1695         struct request_queue *this_q;
1696         struct request *rq;
1697         LIST_HEAD(list);
1698         LIST_HEAD(rq_list);
1699         unsigned int depth;
1700
1701         list_splice_init(&plug->mq_list, &list);
1702         plug->rq_count = 0;
1703
1704         if (plug->rq_count > 2 && plug->multiple_queues)
1705                 list_sort(NULL, &list, plug_rq_cmp);
1706
1707         this_q = NULL;
1708         this_hctx = NULL;
1709         this_ctx = NULL;
1710         depth = 0;
1711
1712         while (!list_empty(&list)) {
1713                 rq = list_entry_rq(list.next);
1714                 list_del_init(&rq->queuelist);
1715                 BUG_ON(!rq->q);
1716                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1717                         if (this_hctx) {
1718                                 trace_block_unplug(this_q, depth, !from_schedule);
1719                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1720                                                                 &rq_list,
1721                                                                 from_schedule);
1722                         }
1723
1724                         this_q = rq->q;
1725                         this_ctx = rq->mq_ctx;
1726                         this_hctx = rq->mq_hctx;
1727                         depth = 0;
1728                 }
1729
1730                 depth++;
1731                 list_add_tail(&rq->queuelist, &rq_list);
1732         }
1733
1734         /*
1735          * If 'this_hctx' is set, we know we have entries to complete
1736          * on 'rq_list'. Do those.
1737          */
1738         if (this_hctx) {
1739                 trace_block_unplug(this_q, depth, !from_schedule);
1740                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1741                                                 from_schedule);
1742         }
1743 }
1744
1745 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1746 {
1747         blk_init_request_from_bio(rq, bio);
1748
1749         blk_account_io_start(rq, true);
1750 }
1751
1752 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1753                                             struct request *rq,
1754                                             blk_qc_t *cookie, bool last)
1755 {
1756         struct request_queue *q = rq->q;
1757         struct blk_mq_queue_data bd = {
1758                 .rq = rq,
1759                 .last = last,
1760         };
1761         blk_qc_t new_cookie;
1762         blk_status_t ret;
1763
1764         new_cookie = request_to_qc_t(hctx, rq);
1765
1766         /*
1767          * For OK queue, we are done. For error, caller may kill it.
1768          * Any other error (busy), just add it to our list as we
1769          * previously would have done.
1770          */
1771         ret = q->mq_ops->queue_rq(hctx, &bd);
1772         switch (ret) {
1773         case BLK_STS_OK:
1774                 blk_mq_update_dispatch_busy(hctx, false);
1775                 *cookie = new_cookie;
1776                 break;
1777         case BLK_STS_RESOURCE:
1778         case BLK_STS_DEV_RESOURCE:
1779                 blk_mq_update_dispatch_busy(hctx, true);
1780                 __blk_mq_requeue_request(rq);
1781                 break;
1782         default:
1783                 blk_mq_update_dispatch_busy(hctx, false);
1784                 *cookie = BLK_QC_T_NONE;
1785                 break;
1786         }
1787
1788         return ret;
1789 }
1790
1791 blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1792                                                 struct request *rq,
1793                                                 blk_qc_t *cookie,
1794                                                 bool bypass, bool last)
1795 {
1796         struct request_queue *q = rq->q;
1797         bool run_queue = true;
1798         blk_status_t ret = BLK_STS_RESOURCE;
1799         int srcu_idx;
1800         bool force = false;
1801
1802         hctx_lock(hctx, &srcu_idx);
1803         /*
1804          * hctx_lock is needed before checking quiesced flag.
1805          *
1806          * When queue is stopped or quiesced, ignore 'bypass', insert
1807          * and return BLK_STS_OK to caller, and avoid driver to try to
1808          * dispatch again.
1809          */
1810         if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q))) {
1811                 run_queue = false;
1812                 bypass = false;
1813                 goto out_unlock;
1814         }
1815
1816         if (unlikely(q->elevator && !bypass))
1817                 goto out_unlock;
1818
1819         if (!blk_mq_get_dispatch_budget(hctx))
1820                 goto out_unlock;
1821
1822         if (!blk_mq_get_driver_tag(rq)) {
1823                 blk_mq_put_dispatch_budget(hctx);
1824                 goto out_unlock;
1825         }
1826
1827         /*
1828          * Always add a request that has been through
1829          *.queue_rq() to the hardware dispatch list.
1830          */
1831         force = true;
1832         ret = __blk_mq_issue_directly(hctx, rq, cookie, last);
1833 out_unlock:
1834         hctx_unlock(hctx, srcu_idx);
1835         switch (ret) {
1836         case BLK_STS_OK:
1837                 break;
1838         case BLK_STS_DEV_RESOURCE:
1839         case BLK_STS_RESOURCE:
1840                 if (force) {
1841                         blk_mq_request_bypass_insert(rq, run_queue);
1842                         /*
1843                          * We have to return BLK_STS_OK for the DM
1844                          * to avoid livelock. Otherwise, we return
1845                          * the real result to indicate whether the
1846                          * request is direct-issued successfully.
1847                          */
1848                         ret = bypass ? BLK_STS_OK : ret;
1849                 } else if (!bypass) {
1850                         blk_mq_sched_insert_request(rq, false,
1851                                                     run_queue, false);
1852                 }
1853                 break;
1854         default:
1855                 if (!bypass)
1856                         blk_mq_end_request(rq, ret);
1857                 break;
1858         }
1859
1860         return ret;
1861 }
1862
1863 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1864                 struct list_head *list)
1865 {
1866         blk_qc_t unused;
1867         blk_status_t ret = BLK_STS_OK;
1868
1869         while (!list_empty(list)) {
1870                 struct request *rq = list_first_entry(list, struct request,
1871                                 queuelist);
1872
1873                 list_del_init(&rq->queuelist);
1874                 if (ret == BLK_STS_OK)
1875                         ret = blk_mq_try_issue_directly(hctx, rq, &unused,
1876                                                         false,
1877                                                         list_empty(list));
1878                 else
1879                         blk_mq_sched_insert_request(rq, false, true, false);
1880         }
1881
1882         /*
1883          * If we didn't flush the entire list, we could have told
1884          * the driver there was more coming, but that turned out to
1885          * be a lie.
1886          */
1887         if (ret != BLK_STS_OK && hctx->queue->mq_ops->commit_rqs)
1888                 hctx->queue->mq_ops->commit_rqs(hctx);
1889 }
1890
1891 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1892 {
1893         list_add_tail(&rq->queuelist, &plug->mq_list);
1894         plug->rq_count++;
1895         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1896                 struct request *tmp;
1897
1898                 tmp = list_first_entry(&plug->mq_list, struct request,
1899                                                 queuelist);
1900                 if (tmp->q != rq->q)
1901                         plug->multiple_queues = true;
1902         }
1903 }
1904
1905 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1906 {
1907         const int is_sync = op_is_sync(bio->bi_opf);
1908         const int is_flush_fua = op_is_flush(bio->bi_opf);
1909         struct blk_mq_alloc_data data = { .flags = 0};
1910         struct request *rq;
1911         struct blk_plug *plug;
1912         struct request *same_queue_rq = NULL;
1913         blk_qc_t cookie;
1914
1915         blk_queue_bounce(q, &bio);
1916
1917         blk_queue_split(q, &bio);
1918
1919         if (!bio_integrity_prep(bio))
1920                 return BLK_QC_T_NONE;
1921
1922         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1923             blk_attempt_plug_merge(q, bio, &same_queue_rq))
1924                 return BLK_QC_T_NONE;
1925
1926         if (blk_mq_sched_bio_merge(q, bio))
1927                 return BLK_QC_T_NONE;
1928
1929         rq_qos_throttle(q, bio);
1930
1931         data.cmd_flags = bio->bi_opf;
1932         rq = blk_mq_get_request(q, bio, &data);
1933         if (unlikely(!rq)) {
1934                 rq_qos_cleanup(q, bio);
1935                 if (bio->bi_opf & REQ_NOWAIT)
1936                         bio_wouldblock_error(bio);
1937                 return BLK_QC_T_NONE;
1938         }
1939
1940         trace_block_getrq(q, bio, bio->bi_opf);
1941
1942         rq_qos_track(q, rq, bio);
1943
1944         cookie = request_to_qc_t(data.hctx, rq);
1945
1946         plug = current->plug;
1947         if (unlikely(is_flush_fua)) {
1948                 blk_mq_put_ctx(data.ctx);
1949                 blk_mq_bio_to_request(rq, bio);
1950
1951                 /* bypass scheduler for flush rq */
1952                 blk_insert_flush(rq);
1953                 blk_mq_run_hw_queue(data.hctx, true);
1954         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1955                 /*
1956                  * Use plugging if we have a ->commit_rqs() hook as well, as
1957                  * we know the driver uses bd->last in a smart fashion.
1958                  */
1959                 unsigned int request_count = plug->rq_count;
1960                 struct request *last = NULL;
1961
1962                 blk_mq_put_ctx(data.ctx);
1963                 blk_mq_bio_to_request(rq, bio);
1964
1965                 if (!request_count)
1966                         trace_block_plug(q);
1967                 else
1968                         last = list_entry_rq(plug->mq_list.prev);
1969
1970                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1971                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1972                         blk_flush_plug_list(plug, false);
1973                         trace_block_plug(q);
1974                 }
1975
1976                 blk_add_rq_to_plug(plug, rq);
1977         } else if (plug && !blk_queue_nomerges(q)) {
1978                 blk_mq_bio_to_request(rq, bio);
1979
1980                 /*
1981                  * We do limited plugging. If the bio can be merged, do that.
1982                  * Otherwise the existing request in the plug list will be
1983                  * issued. So the plug list will have one request at most
1984                  * The plug list might get flushed before this. If that happens,
1985                  * the plug list is empty, and same_queue_rq is invalid.
1986                  */
1987                 if (list_empty(&plug->mq_list))
1988                         same_queue_rq = NULL;
1989                 if (same_queue_rq) {
1990                         list_del_init(&same_queue_rq->queuelist);
1991                         plug->rq_count--;
1992                 }
1993                 blk_add_rq_to_plug(plug, rq);
1994
1995                 blk_mq_put_ctx(data.ctx);
1996
1997                 if (same_queue_rq) {
1998                         data.hctx = same_queue_rq->mq_hctx;
1999                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2000                                         &cookie, false, true);
2001                 }
2002         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2003                         !data.hctx->dispatch_busy)) {
2004                 blk_mq_put_ctx(data.ctx);
2005                 blk_mq_bio_to_request(rq, bio);
2006                 blk_mq_try_issue_directly(data.hctx, rq, &cookie, false, true);
2007         } else {
2008                 blk_mq_put_ctx(data.ctx);
2009                 blk_mq_bio_to_request(rq, bio);
2010                 blk_mq_sched_insert_request(rq, false, true, true);
2011         }
2012
2013         return cookie;
2014 }
2015
2016 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2017                      unsigned int hctx_idx)
2018 {
2019         struct page *page;
2020
2021         if (tags->rqs && set->ops->exit_request) {
2022                 int i;
2023
2024                 for (i = 0; i < tags->nr_tags; i++) {
2025                         struct request *rq = tags->static_rqs[i];
2026
2027                         if (!rq)
2028                                 continue;
2029                         set->ops->exit_request(set, rq, hctx_idx);
2030                         tags->static_rqs[i] = NULL;
2031                 }
2032         }
2033
2034         while (!list_empty(&tags->page_list)) {
2035                 page = list_first_entry(&tags->page_list, struct page, lru);
2036                 list_del_init(&page->lru);
2037                 /*
2038                  * Remove kmemleak object previously allocated in
2039                  * blk_mq_init_rq_map().
2040                  */
2041                 kmemleak_free(page_address(page));
2042                 __free_pages(page, page->private);
2043         }
2044 }
2045
2046 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2047 {
2048         kfree(tags->rqs);
2049         tags->rqs = NULL;
2050         kfree(tags->static_rqs);
2051         tags->static_rqs = NULL;
2052
2053         blk_mq_free_tags(tags);
2054 }
2055
2056 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2057                                         unsigned int hctx_idx,
2058                                         unsigned int nr_tags,
2059                                         unsigned int reserved_tags)
2060 {
2061         struct blk_mq_tags *tags;
2062         int node;
2063
2064         node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx);
2065         if (node == NUMA_NO_NODE)
2066                 node = set->numa_node;
2067
2068         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2069                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2070         if (!tags)
2071                 return NULL;
2072
2073         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2074                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2075                                  node);
2076         if (!tags->rqs) {
2077                 blk_mq_free_tags(tags);
2078                 return NULL;
2079         }
2080
2081         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2082                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2083                                         node);
2084         if (!tags->static_rqs) {
2085                 kfree(tags->rqs);
2086                 blk_mq_free_tags(tags);
2087                 return NULL;
2088         }
2089
2090         return tags;
2091 }
2092
2093 static size_t order_to_size(unsigned int order)
2094 {
2095         return (size_t)PAGE_SIZE << order;
2096 }
2097
2098 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2099                                unsigned int hctx_idx, int node)
2100 {
2101         int ret;
2102
2103         if (set->ops->init_request) {
2104                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2105                 if (ret)
2106                         return ret;
2107         }
2108
2109         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2110         return 0;
2111 }
2112
2113 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2114                      unsigned int hctx_idx, unsigned int depth)
2115 {
2116         unsigned int i, j, entries_per_page, max_order = 4;
2117         size_t rq_size, left;
2118         int node;
2119
2120         node = blk_mq_hw_queue_to_node(&set->map[0], hctx_idx);
2121         if (node == NUMA_NO_NODE)
2122                 node = set->numa_node;
2123
2124         INIT_LIST_HEAD(&tags->page_list);
2125
2126         /*
2127          * rq_size is the size of the request plus driver payload, rounded
2128          * to the cacheline size
2129          */
2130         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2131                                 cache_line_size());
2132         left = rq_size * depth;
2133
2134         for (i = 0; i < depth; ) {
2135                 int this_order = max_order;
2136                 struct page *page;
2137                 int to_do;
2138                 void *p;
2139
2140                 while (this_order && left < order_to_size(this_order - 1))
2141                         this_order--;
2142
2143                 do {
2144                         page = alloc_pages_node(node,
2145                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2146                                 this_order);
2147                         if (page)
2148                                 break;
2149                         if (!this_order--)
2150                                 break;
2151                         if (order_to_size(this_order) < rq_size)
2152                                 break;
2153                 } while (1);
2154
2155                 if (!page)
2156                         goto fail;
2157
2158                 page->private = this_order;
2159                 list_add_tail(&page->lru, &tags->page_list);
2160
2161                 p = page_address(page);
2162                 /*
2163                  * Allow kmemleak to scan these pages as they contain pointers
2164                  * to additional allocations like via ops->init_request().
2165                  */
2166                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2167                 entries_per_page = order_to_size(this_order) / rq_size;
2168                 to_do = min(entries_per_page, depth - i);
2169                 left -= to_do * rq_size;
2170                 for (j = 0; j < to_do; j++) {
2171                         struct request *rq = p;
2172
2173                         tags->static_rqs[i] = rq;
2174                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2175                                 tags->static_rqs[i] = NULL;
2176                                 goto fail;
2177                         }
2178
2179                         p += rq_size;
2180                         i++;
2181                 }
2182         }
2183         return 0;
2184
2185 fail:
2186         blk_mq_free_rqs(set, tags, hctx_idx);
2187         return -ENOMEM;
2188 }
2189
2190 /*
2191  * 'cpu' is going away. splice any existing rq_list entries from this
2192  * software queue to the hw queue dispatch list, and ensure that it
2193  * gets run.
2194  */
2195 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2196 {
2197         struct blk_mq_hw_ctx *hctx;
2198         struct blk_mq_ctx *ctx;
2199         LIST_HEAD(tmp);
2200         enum hctx_type type;
2201
2202         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2203         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2204         type = hctx->type;
2205
2206         spin_lock(&ctx->lock);
2207         if (!list_empty(&ctx->rq_lists[type])) {
2208                 list_splice_init(&ctx->rq_lists[type], &tmp);
2209                 blk_mq_hctx_clear_pending(hctx, ctx);
2210         }
2211         spin_unlock(&ctx->lock);
2212
2213         if (list_empty(&tmp))
2214                 return 0;
2215
2216         spin_lock(&hctx->lock);
2217         list_splice_tail_init(&tmp, &hctx->dispatch);
2218         spin_unlock(&hctx->lock);
2219
2220         blk_mq_run_hw_queue(hctx, true);
2221         return 0;
2222 }
2223
2224 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2225 {
2226         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2227                                             &hctx->cpuhp_dead);
2228 }
2229
2230 /* hctx->ctxs will be freed in queue's release handler */
2231 static void blk_mq_exit_hctx(struct request_queue *q,
2232                 struct blk_mq_tag_set *set,
2233                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2234 {
2235         if (blk_mq_hw_queue_mapped(hctx))
2236                 blk_mq_tag_idle(hctx);
2237
2238         if (set->ops->exit_request)
2239                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2240
2241         if (set->ops->exit_hctx)
2242                 set->ops->exit_hctx(hctx, hctx_idx);
2243
2244         if (hctx->flags & BLK_MQ_F_BLOCKING)
2245                 cleanup_srcu_struct(hctx->srcu);
2246
2247         blk_mq_remove_cpuhp(hctx);
2248         blk_free_flush_queue(hctx->fq);
2249         sbitmap_free(&hctx->ctx_map);
2250 }
2251
2252 static void blk_mq_exit_hw_queues(struct request_queue *q,
2253                 struct blk_mq_tag_set *set, int nr_queue)
2254 {
2255         struct blk_mq_hw_ctx *hctx;
2256         unsigned int i;
2257
2258         queue_for_each_hw_ctx(q, hctx, i) {
2259                 if (i == nr_queue)
2260                         break;
2261                 blk_mq_debugfs_unregister_hctx(hctx);
2262                 blk_mq_exit_hctx(q, set, hctx, i);
2263         }
2264 }
2265
2266 static int blk_mq_init_hctx(struct request_queue *q,
2267                 struct blk_mq_tag_set *set,
2268                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2269 {
2270         int node;
2271
2272         node = hctx->numa_node;
2273         if (node == NUMA_NO_NODE)
2274                 node = hctx->numa_node = set->numa_node;
2275
2276         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2277         spin_lock_init(&hctx->lock);
2278         INIT_LIST_HEAD(&hctx->dispatch);
2279         hctx->queue = q;
2280         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2281
2282         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2283
2284         hctx->tags = set->tags[hctx_idx];
2285
2286         /*
2287          * Allocate space for all possible cpus to avoid allocation at
2288          * runtime
2289          */
2290         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2291                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node);
2292         if (!hctx->ctxs)
2293                 goto unregister_cpu_notifier;
2294
2295         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2296                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY, node))
2297                 goto free_ctxs;
2298
2299         hctx->nr_ctx = 0;
2300
2301         spin_lock_init(&hctx->dispatch_wait_lock);
2302         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2303         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2304
2305         if (set->ops->init_hctx &&
2306             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2307                 goto free_bitmap;
2308
2309         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2310                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
2311         if (!hctx->fq)
2312                 goto exit_hctx;
2313
2314         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2315                 goto free_fq;
2316
2317         if (hctx->flags & BLK_MQ_F_BLOCKING)
2318                 init_srcu_struct(hctx->srcu);
2319
2320         return 0;
2321
2322  free_fq:
2323         kfree(hctx->fq);
2324  exit_hctx:
2325         if (set->ops->exit_hctx)
2326                 set->ops->exit_hctx(hctx, hctx_idx);
2327  free_bitmap:
2328         sbitmap_free(&hctx->ctx_map);
2329  free_ctxs:
2330         kfree(hctx->ctxs);
2331  unregister_cpu_notifier:
2332         blk_mq_remove_cpuhp(hctx);
2333         return -1;
2334 }
2335
2336 static void blk_mq_init_cpu_queues(struct request_queue *q,
2337                                    unsigned int nr_hw_queues)
2338 {
2339         struct blk_mq_tag_set *set = q->tag_set;
2340         unsigned int i, j;
2341
2342         for_each_possible_cpu(i) {
2343                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2344                 struct blk_mq_hw_ctx *hctx;
2345                 int k;
2346
2347                 __ctx->cpu = i;
2348                 spin_lock_init(&__ctx->lock);
2349                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2350                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2351
2352                 __ctx->queue = q;
2353
2354                 /*
2355                  * Set local node, IFF we have more than one hw queue. If
2356                  * not, we remain on the home node of the device
2357                  */
2358                 for (j = 0; j < set->nr_maps; j++) {
2359                         hctx = blk_mq_map_queue_type(q, j, i);
2360                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2361                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2362                 }
2363         }
2364 }
2365
2366 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2367 {
2368         int ret = 0;
2369
2370         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2371                                         set->queue_depth, set->reserved_tags);
2372         if (!set->tags[hctx_idx])
2373                 return false;
2374
2375         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2376                                 set->queue_depth);
2377         if (!ret)
2378                 return true;
2379
2380         blk_mq_free_rq_map(set->tags[hctx_idx]);
2381         set->tags[hctx_idx] = NULL;
2382         return false;
2383 }
2384
2385 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2386                                          unsigned int hctx_idx)
2387 {
2388         if (set->tags && set->tags[hctx_idx]) {
2389                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2390                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2391                 set->tags[hctx_idx] = NULL;
2392         }
2393 }
2394
2395 static void blk_mq_map_swqueue(struct request_queue *q)
2396 {
2397         unsigned int i, j, hctx_idx;
2398         struct blk_mq_hw_ctx *hctx;
2399         struct blk_mq_ctx *ctx;
2400         struct blk_mq_tag_set *set = q->tag_set;
2401
2402         /*
2403          * Avoid others reading imcomplete hctx->cpumask through sysfs
2404          */
2405         mutex_lock(&q->sysfs_lock);
2406
2407         queue_for_each_hw_ctx(q, hctx, i) {
2408                 cpumask_clear(hctx->cpumask);
2409                 hctx->nr_ctx = 0;
2410                 hctx->dispatch_from = NULL;
2411         }
2412
2413         /*
2414          * Map software to hardware queues.
2415          *
2416          * If the cpu isn't present, the cpu is mapped to first hctx.
2417          */
2418         for_each_possible_cpu(i) {
2419                 hctx_idx = set->map[0].mq_map[i];
2420                 /* unmapped hw queue can be remapped after CPU topo changed */
2421                 if (!set->tags[hctx_idx] &&
2422                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2423                         /*
2424                          * If tags initialization fail for some hctx,
2425                          * that hctx won't be brought online.  In this
2426                          * case, remap the current ctx to hctx[0] which
2427                          * is guaranteed to always have tags allocated
2428                          */
2429                         set->map[0].mq_map[i] = 0;
2430                 }
2431
2432                 ctx = per_cpu_ptr(q->queue_ctx, i);
2433                 for (j = 0; j < set->nr_maps; j++) {
2434                         if (!set->map[j].nr_queues)
2435                                 continue;
2436
2437                         hctx = blk_mq_map_queue_type(q, j, i);
2438
2439                         /*
2440                          * If the CPU is already set in the mask, then we've
2441                          * mapped this one already. This can happen if
2442                          * devices share queues across queue maps.
2443                          */
2444                         if (cpumask_test_cpu(i, hctx->cpumask))
2445                                 continue;
2446
2447                         cpumask_set_cpu(i, hctx->cpumask);
2448                         hctx->type = j;
2449                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2450                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2451
2452                         /*
2453                          * If the nr_ctx type overflows, we have exceeded the
2454                          * amount of sw queues we can support.
2455                          */
2456                         BUG_ON(!hctx->nr_ctx);
2457                 }
2458         }
2459
2460         mutex_unlock(&q->sysfs_lock);
2461
2462         queue_for_each_hw_ctx(q, hctx, i) {
2463                 /*
2464                  * If no software queues are mapped to this hardware queue,
2465                  * disable it and free the request entries.
2466                  */
2467                 if (!hctx->nr_ctx) {
2468                         /* Never unmap queue 0.  We need it as a
2469                          * fallback in case of a new remap fails
2470                          * allocation
2471                          */
2472                         if (i && set->tags[i])
2473                                 blk_mq_free_map_and_requests(set, i);
2474
2475                         hctx->tags = NULL;
2476                         continue;
2477                 }
2478
2479                 hctx->tags = set->tags[i];
2480                 WARN_ON(!hctx->tags);
2481
2482                 /*
2483                  * Set the map size to the number of mapped software queues.
2484                  * This is more accurate and more efficient than looping
2485                  * over all possibly mapped software queues.
2486                  */
2487                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2488
2489                 /*
2490                  * Initialize batch roundrobin counts
2491                  */
2492                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2493                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2494         }
2495 }
2496
2497 /*
2498  * Caller needs to ensure that we're either frozen/quiesced, or that
2499  * the queue isn't live yet.
2500  */
2501 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2502 {
2503         struct blk_mq_hw_ctx *hctx;
2504         int i;
2505
2506         queue_for_each_hw_ctx(q, hctx, i) {
2507                 if (shared)
2508                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2509                 else
2510                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2511         }
2512 }
2513
2514 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2515                                         bool shared)
2516 {
2517         struct request_queue *q;
2518
2519         lockdep_assert_held(&set->tag_list_lock);
2520
2521         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2522                 blk_mq_freeze_queue(q);
2523                 queue_set_hctx_shared(q, shared);
2524                 blk_mq_unfreeze_queue(q);
2525         }
2526 }
2527
2528 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2529 {
2530         struct blk_mq_tag_set *set = q->tag_set;
2531
2532         mutex_lock(&set->tag_list_lock);
2533         list_del_rcu(&q->tag_set_list);
2534         if (list_is_singular(&set->tag_list)) {
2535                 /* just transitioned to unshared */
2536                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2537                 /* update existing queue */
2538                 blk_mq_update_tag_set_depth(set, false);
2539         }
2540         mutex_unlock(&set->tag_list_lock);
2541         INIT_LIST_HEAD(&q->tag_set_list);
2542 }
2543
2544 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2545                                      struct request_queue *q)
2546 {
2547         mutex_lock(&set->tag_list_lock);
2548
2549         /*
2550          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2551          */
2552         if (!list_empty(&set->tag_list) &&
2553             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2554                 set->flags |= BLK_MQ_F_TAG_SHARED;
2555                 /* update existing queue */
2556                 blk_mq_update_tag_set_depth(set, true);
2557         }
2558         if (set->flags & BLK_MQ_F_TAG_SHARED)
2559                 queue_set_hctx_shared(q, true);
2560         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2561
2562         mutex_unlock(&set->tag_list_lock);
2563 }
2564
2565 /* All allocations will be freed in release handler of q->mq_kobj */
2566 static int blk_mq_alloc_ctxs(struct request_queue *q)
2567 {
2568         struct blk_mq_ctxs *ctxs;
2569         int cpu;
2570
2571         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2572         if (!ctxs)
2573                 return -ENOMEM;
2574
2575         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2576         if (!ctxs->queue_ctx)
2577                 goto fail;
2578
2579         for_each_possible_cpu(cpu) {
2580                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2581                 ctx->ctxs = ctxs;
2582         }
2583
2584         q->mq_kobj = &ctxs->kobj;
2585         q->queue_ctx = ctxs->queue_ctx;
2586
2587         return 0;
2588  fail:
2589         kfree(ctxs);
2590         return -ENOMEM;
2591 }
2592
2593 /*
2594  * It is the actual release handler for mq, but we do it from
2595  * request queue's release handler for avoiding use-after-free
2596  * and headache because q->mq_kobj shouldn't have been introduced,
2597  * but we can't group ctx/kctx kobj without it.
2598  */
2599 void blk_mq_release(struct request_queue *q)
2600 {
2601         struct blk_mq_hw_ctx *hctx;
2602         unsigned int i;
2603
2604         /* hctx kobj stays in hctx */
2605         queue_for_each_hw_ctx(q, hctx, i) {
2606                 if (!hctx)
2607                         continue;
2608                 kobject_put(&hctx->kobj);
2609         }
2610
2611         kfree(q->queue_hw_ctx);
2612
2613         /*
2614          * release .mq_kobj and sw queue's kobject now because
2615          * both share lifetime with request queue.
2616          */
2617         blk_mq_sysfs_deinit(q);
2618 }
2619
2620 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2621 {
2622         struct request_queue *uninit_q, *q;
2623
2624         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2625         if (!uninit_q)
2626                 return ERR_PTR(-ENOMEM);
2627
2628         q = blk_mq_init_allocated_queue(set, uninit_q);
2629         if (IS_ERR(q))
2630                 blk_cleanup_queue(uninit_q);
2631
2632         return q;
2633 }
2634 EXPORT_SYMBOL(blk_mq_init_queue);
2635
2636 /*
2637  * Helper for setting up a queue with mq ops, given queue depth, and
2638  * the passed in mq ops flags.
2639  */
2640 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2641                                            const struct blk_mq_ops *ops,
2642                                            unsigned int queue_depth,
2643                                            unsigned int set_flags)
2644 {
2645         struct request_queue *q;
2646         int ret;
2647
2648         memset(set, 0, sizeof(*set));
2649         set->ops = ops;
2650         set->nr_hw_queues = 1;
2651         set->nr_maps = 1;
2652         set->queue_depth = queue_depth;
2653         set->numa_node = NUMA_NO_NODE;
2654         set->flags = set_flags;
2655
2656         ret = blk_mq_alloc_tag_set(set);
2657         if (ret)
2658                 return ERR_PTR(ret);
2659
2660         q = blk_mq_init_queue(set);
2661         if (IS_ERR(q)) {
2662                 blk_mq_free_tag_set(set);
2663                 return q;
2664         }
2665
2666         return q;
2667 }
2668 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2669
2670 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2671 {
2672         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2673
2674         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2675                            __alignof__(struct blk_mq_hw_ctx)) !=
2676                      sizeof(struct blk_mq_hw_ctx));
2677
2678         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2679                 hw_ctx_size += sizeof(struct srcu_struct);
2680
2681         return hw_ctx_size;
2682 }
2683
2684 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2685                 struct blk_mq_tag_set *set, struct request_queue *q,
2686                 int hctx_idx, int node)
2687 {
2688         struct blk_mq_hw_ctx *hctx;
2689
2690         hctx = kzalloc_node(blk_mq_hw_ctx_size(set),
2691                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2692                         node);
2693         if (!hctx)
2694                 return NULL;
2695
2696         if (!zalloc_cpumask_var_node(&hctx->cpumask,
2697                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2698                                 node)) {
2699                 kfree(hctx);
2700                 return NULL;
2701         }
2702
2703         atomic_set(&hctx->nr_active, 0);
2704         hctx->numa_node = node;
2705         hctx->queue_num = hctx_idx;
2706
2707         if (blk_mq_init_hctx(q, set, hctx, hctx_idx)) {
2708                 free_cpumask_var(hctx->cpumask);
2709                 kfree(hctx);
2710                 return NULL;
2711         }
2712         blk_mq_hctx_kobj_init(hctx);
2713
2714         return hctx;
2715 }
2716
2717 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2718                                                 struct request_queue *q)
2719 {
2720         int i, j, end;
2721         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2722
2723         /* protect against switching io scheduler  */
2724         mutex_lock(&q->sysfs_lock);
2725         for (i = 0; i < set->nr_hw_queues; i++) {
2726                 int node;
2727                 struct blk_mq_hw_ctx *hctx;
2728
2729                 node = blk_mq_hw_queue_to_node(&set->map[0], i);
2730                 /*
2731                  * If the hw queue has been mapped to another numa node,
2732                  * we need to realloc the hctx. If allocation fails, fallback
2733                  * to use the previous one.
2734                  */
2735                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2736                         continue;
2737
2738                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2739                 if (hctx) {
2740                         if (hctxs[i]) {
2741                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2742                                 kobject_put(&hctxs[i]->kobj);
2743                         }
2744                         hctxs[i] = hctx;
2745                 } else {
2746                         if (hctxs[i])
2747                                 pr_warn("Allocate new hctx on node %d fails,\
2748                                                 fallback to previous one on node %d\n",
2749                                                 node, hctxs[i]->numa_node);
2750                         else
2751                                 break;
2752                 }
2753         }
2754         /*
2755          * Increasing nr_hw_queues fails. Free the newly allocated
2756          * hctxs and keep the previous q->nr_hw_queues.
2757          */
2758         if (i != set->nr_hw_queues) {
2759                 j = q->nr_hw_queues;
2760                 end = i;
2761         } else {
2762                 j = i;
2763                 end = q->nr_hw_queues;
2764                 q->nr_hw_queues = set->nr_hw_queues;
2765         }
2766
2767         for (; j < end; j++) {
2768                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2769
2770                 if (hctx) {
2771                         if (hctx->tags)
2772                                 blk_mq_free_map_and_requests(set, j);
2773                         blk_mq_exit_hctx(q, set, hctx, j);
2774                         kobject_put(&hctx->kobj);
2775                         hctxs[j] = NULL;
2776
2777                 }
2778         }
2779         mutex_unlock(&q->sysfs_lock);
2780 }
2781
2782 /*
2783  * Maximum number of hardware queues we support. For single sets, we'll never
2784  * have more than the CPUs (software queues). For multiple sets, the tag_set
2785  * user may have set ->nr_hw_queues larger.
2786  */
2787 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2788 {
2789         if (set->nr_maps == 1)
2790                 return nr_cpu_ids;
2791
2792         return max(set->nr_hw_queues, nr_cpu_ids);
2793 }
2794
2795 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2796                                                   struct request_queue *q)
2797 {
2798         /* mark the queue as mq asap */
2799         q->mq_ops = set->ops;
2800
2801         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2802                                              blk_mq_poll_stats_bkt,
2803                                              BLK_MQ_POLL_STATS_BKTS, q);
2804         if (!q->poll_cb)
2805                 goto err_exit;
2806
2807         if (blk_mq_alloc_ctxs(q))
2808                 goto err_exit;
2809
2810         /* init q->mq_kobj and sw queues' kobjects */
2811         blk_mq_sysfs_init(q);
2812
2813         q->nr_queues = nr_hw_queues(set);
2814         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2815                                                 GFP_KERNEL, set->numa_node);
2816         if (!q->queue_hw_ctx)
2817                 goto err_sys_init;
2818
2819         blk_mq_realloc_hw_ctxs(set, q);
2820         if (!q->nr_hw_queues)
2821                 goto err_hctxs;
2822
2823         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2824         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2825
2826         q->tag_set = set;
2827
2828         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2829         if (set->nr_maps > HCTX_TYPE_POLL &&
2830             set->map[HCTX_TYPE_POLL].nr_queues)
2831                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2832
2833         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2834                 blk_queue_flag_set(QUEUE_FLAG_NO_SG_MERGE, q);
2835
2836         q->sg_reserved_size = INT_MAX;
2837
2838         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2839         INIT_LIST_HEAD(&q->requeue_list);
2840         spin_lock_init(&q->requeue_lock);
2841
2842         blk_queue_make_request(q, blk_mq_make_request);
2843
2844         /*
2845          * Do this after blk_queue_make_request() overrides it...
2846          */
2847         q->nr_requests = set->queue_depth;
2848
2849         /*
2850          * Default to classic polling
2851          */
2852         q->poll_nsec = -1;
2853
2854         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2855         blk_mq_add_queue_tag_set(set, q);
2856         blk_mq_map_swqueue(q);
2857
2858         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2859                 int ret;
2860
2861                 ret = elevator_init_mq(q);
2862                 if (ret)
2863                         return ERR_PTR(ret);
2864         }
2865
2866         return q;
2867
2868 err_hctxs:
2869         kfree(q->queue_hw_ctx);
2870 err_sys_init:
2871         blk_mq_sysfs_deinit(q);
2872 err_exit:
2873         q->mq_ops = NULL;
2874         return ERR_PTR(-ENOMEM);
2875 }
2876 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2877
2878 void blk_mq_free_queue(struct request_queue *q)
2879 {
2880         struct blk_mq_tag_set   *set = q->tag_set;
2881
2882         blk_mq_del_queue_tag_set(q);
2883         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2884 }
2885
2886 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2887 {
2888         int i;
2889
2890         for (i = 0; i < set->nr_hw_queues; i++)
2891                 if (!__blk_mq_alloc_rq_map(set, i))
2892                         goto out_unwind;
2893
2894         return 0;
2895
2896 out_unwind:
2897         while (--i >= 0)
2898                 blk_mq_free_rq_map(set->tags[i]);
2899
2900         return -ENOMEM;
2901 }
2902
2903 /*
2904  * Allocate the request maps associated with this tag_set. Note that this
2905  * may reduce the depth asked for, if memory is tight. set->queue_depth
2906  * will be updated to reflect the allocated depth.
2907  */
2908 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2909 {
2910         unsigned int depth;
2911         int err;
2912
2913         depth = set->queue_depth;
2914         do {
2915                 err = __blk_mq_alloc_rq_maps(set);
2916                 if (!err)
2917                         break;
2918
2919                 set->queue_depth >>= 1;
2920                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2921                         err = -ENOMEM;
2922                         break;
2923                 }
2924         } while (set->queue_depth);
2925
2926         if (!set->queue_depth || err) {
2927                 pr_err("blk-mq: failed to allocate request map\n");
2928                 return -ENOMEM;
2929         }
2930
2931         if (depth != set->queue_depth)
2932                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2933                                                 depth, set->queue_depth);
2934
2935         return 0;
2936 }
2937
2938 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2939 {
2940         if (set->ops->map_queues && !is_kdump_kernel()) {
2941                 int i;
2942
2943                 /*
2944                  * transport .map_queues is usually done in the following
2945                  * way:
2946                  *
2947                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2948                  *      mask = get_cpu_mask(queue)
2949                  *      for_each_cpu(cpu, mask)
2950                  *              set->map[x].mq_map[cpu] = queue;
2951                  * }
2952                  *
2953                  * When we need to remap, the table has to be cleared for
2954                  * killing stale mapping since one CPU may not be mapped
2955                  * to any hw queue.
2956                  */
2957                 for (i = 0; i < set->nr_maps; i++)
2958                         blk_mq_clear_mq_map(&set->map[i]);
2959
2960                 return set->ops->map_queues(set);
2961         } else {
2962                 BUG_ON(set->nr_maps > 1);
2963                 return blk_mq_map_queues(&set->map[0]);
2964         }
2965 }
2966
2967 /*
2968  * Alloc a tag set to be associated with one or more request queues.
2969  * May fail with EINVAL for various error conditions. May adjust the
2970  * requested depth down, if it's too large. In that case, the set
2971  * value will be stored in set->queue_depth.
2972  */
2973 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2974 {
2975         int i, ret;
2976
2977         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2978
2979         if (!set->nr_hw_queues)
2980                 return -EINVAL;
2981         if (!set->queue_depth)
2982                 return -EINVAL;
2983         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2984                 return -EINVAL;
2985
2986         if (!set->ops->queue_rq)
2987                 return -EINVAL;
2988
2989         if (!set->ops->get_budget ^ !set->ops->put_budget)
2990                 return -EINVAL;
2991
2992         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2993                 pr_info("blk-mq: reduced tag depth to %u\n",
2994                         BLK_MQ_MAX_DEPTH);
2995                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2996         }
2997
2998         if (!set->nr_maps)
2999                 set->nr_maps = 1;
3000         else if (set->nr_maps > HCTX_MAX_TYPES)
3001                 return -EINVAL;
3002
3003         /*
3004          * If a crashdump is active, then we are potentially in a very
3005          * memory constrained environment. Limit us to 1 queue and
3006          * 64 tags to prevent using too much memory.
3007          */
3008         if (is_kdump_kernel()) {
3009                 set->nr_hw_queues = 1;
3010                 set->nr_maps = 1;
3011                 set->queue_depth = min(64U, set->queue_depth);
3012         }
3013         /*
3014          * There is no use for more h/w queues than cpus if we just have
3015          * a single map
3016          */
3017         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3018                 set->nr_hw_queues = nr_cpu_ids;
3019
3020         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3021                                  GFP_KERNEL, set->numa_node);
3022         if (!set->tags)
3023                 return -ENOMEM;
3024
3025         ret = -ENOMEM;
3026         for (i = 0; i < set->nr_maps; i++) {
3027                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3028                                                   sizeof(set->map[i].mq_map[0]),
3029                                                   GFP_KERNEL, set->numa_node);
3030                 if (!set->map[i].mq_map)
3031                         goto out_free_mq_map;
3032                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3033         }
3034
3035         ret = blk_mq_update_queue_map(set);
3036         if (ret)
3037                 goto out_free_mq_map;
3038
3039         ret = blk_mq_alloc_rq_maps(set);
3040         if (ret)
3041                 goto out_free_mq_map;
3042
3043         mutex_init(&set->tag_list_lock);
3044         INIT_LIST_HEAD(&set->tag_list);
3045
3046         return 0;
3047
3048 out_free_mq_map:
3049         for (i = 0; i < set->nr_maps; i++) {
3050                 kfree(set->map[i].mq_map);
3051                 set->map[i].mq_map = NULL;
3052         }
3053         kfree(set->tags);
3054         set->tags = NULL;
3055         return ret;
3056 }
3057 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3058
3059 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3060 {
3061         int i, j;
3062
3063         for (i = 0; i < nr_hw_queues(set); i++)
3064                 blk_mq_free_map_and_requests(set, i);
3065
3066         for (j = 0; j < set->nr_maps; j++) {
3067                 kfree(set->map[j].mq_map);
3068                 set->map[j].mq_map = NULL;
3069         }
3070
3071         kfree(set->tags);
3072         set->tags = NULL;
3073 }
3074 EXPORT_SYMBOL(blk_mq_free_tag_set);
3075
3076 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3077 {
3078         struct blk_mq_tag_set *set = q->tag_set;
3079         struct blk_mq_hw_ctx *hctx;
3080         int i, ret;
3081
3082         if (!set)
3083                 return -EINVAL;
3084
3085         blk_mq_freeze_queue(q);
3086         blk_mq_quiesce_queue(q);
3087
3088         ret = 0;
3089         queue_for_each_hw_ctx(q, hctx, i) {
3090                 if (!hctx->tags)
3091                         continue;
3092                 /*
3093                  * If we're using an MQ scheduler, just update the scheduler
3094                  * queue depth. This is similar to what the old code would do.
3095                  */
3096                 if (!hctx->sched_tags) {
3097                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3098                                                         false);
3099                 } else {
3100                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3101                                                         nr, true);
3102                 }
3103                 if (ret)
3104                         break;
3105         }
3106
3107         if (!ret)
3108                 q->nr_requests = nr;
3109
3110         blk_mq_unquiesce_queue(q);
3111         blk_mq_unfreeze_queue(q);
3112
3113         return ret;
3114 }
3115
3116 /*
3117  * request_queue and elevator_type pair.
3118  * It is just used by __blk_mq_update_nr_hw_queues to cache
3119  * the elevator_type associated with a request_queue.
3120  */
3121 struct blk_mq_qe_pair {
3122         struct list_head node;
3123         struct request_queue *q;
3124         struct elevator_type *type;
3125 };
3126
3127 /*
3128  * Cache the elevator_type in qe pair list and switch the
3129  * io scheduler to 'none'
3130  */
3131 static bool blk_mq_elv_switch_none(struct list_head *head,
3132                 struct request_queue *q)
3133 {
3134         struct blk_mq_qe_pair *qe;
3135
3136         if (!q->elevator)
3137                 return true;
3138
3139         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3140         if (!qe)
3141                 return false;
3142
3143         INIT_LIST_HEAD(&qe->node);
3144         qe->q = q;
3145         qe->type = q->elevator->type;
3146         list_add(&qe->node, head);
3147
3148         mutex_lock(&q->sysfs_lock);
3149         /*
3150          * After elevator_switch_mq, the previous elevator_queue will be
3151          * released by elevator_release. The reference of the io scheduler
3152          * module get by elevator_get will also be put. So we need to get
3153          * a reference of the io scheduler module here to prevent it to be
3154          * removed.
3155          */
3156         __module_get(qe->type->elevator_owner);
3157         elevator_switch_mq(q, NULL);
3158         mutex_unlock(&q->sysfs_lock);
3159
3160         return true;
3161 }
3162
3163 static void blk_mq_elv_switch_back(struct list_head *head,
3164                 struct request_queue *q)
3165 {
3166         struct blk_mq_qe_pair *qe;
3167         struct elevator_type *t = NULL;
3168
3169         list_for_each_entry(qe, head, node)
3170                 if (qe->q == q) {
3171                         t = qe->type;
3172                         break;
3173                 }
3174
3175         if (!t)
3176                 return;
3177
3178         list_del(&qe->node);
3179         kfree(qe);
3180
3181         mutex_lock(&q->sysfs_lock);
3182         elevator_switch_mq(q, t);
3183         mutex_unlock(&q->sysfs_lock);
3184 }
3185
3186 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3187                                                         int nr_hw_queues)
3188 {
3189         struct request_queue *q;
3190         LIST_HEAD(head);
3191         int prev_nr_hw_queues;
3192
3193         lockdep_assert_held(&set->tag_list_lock);
3194
3195         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3196                 nr_hw_queues = nr_cpu_ids;
3197         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3198                 return;
3199
3200         list_for_each_entry(q, &set->tag_list, tag_set_list)
3201                 blk_mq_freeze_queue(q);
3202         /*
3203          * Sync with blk_mq_queue_tag_busy_iter.
3204          */
3205         synchronize_rcu();
3206         /*
3207          * Switch IO scheduler to 'none', cleaning up the data associated
3208          * with the previous scheduler. We will switch back once we are done
3209          * updating the new sw to hw queue mappings.
3210          */
3211         list_for_each_entry(q, &set->tag_list, tag_set_list)
3212                 if (!blk_mq_elv_switch_none(&head, q))
3213                         goto switch_back;
3214
3215         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3216                 blk_mq_debugfs_unregister_hctxs(q);
3217                 blk_mq_sysfs_unregister(q);
3218         }
3219
3220         prev_nr_hw_queues = set->nr_hw_queues;
3221         set->nr_hw_queues = nr_hw_queues;
3222         blk_mq_update_queue_map(set);
3223 fallback:
3224         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3225                 blk_mq_realloc_hw_ctxs(set, q);
3226                 if (q->nr_hw_queues != set->nr_hw_queues) {
3227                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3228                                         nr_hw_queues, prev_nr_hw_queues);
3229                         set->nr_hw_queues = prev_nr_hw_queues;
3230                         blk_mq_map_queues(&set->map[0]);
3231                         goto fallback;
3232                 }
3233                 blk_mq_map_swqueue(q);
3234         }
3235
3236         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3237                 blk_mq_sysfs_register(q);
3238                 blk_mq_debugfs_register_hctxs(q);
3239         }
3240
3241 switch_back:
3242         list_for_each_entry(q, &set->tag_list, tag_set_list)
3243                 blk_mq_elv_switch_back(&head, q);
3244
3245         list_for_each_entry(q, &set->tag_list, tag_set_list)
3246                 blk_mq_unfreeze_queue(q);
3247 }
3248
3249 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3250 {
3251         mutex_lock(&set->tag_list_lock);
3252         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3253         mutex_unlock(&set->tag_list_lock);
3254 }
3255 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3256
3257 /* Enable polling stats and return whether they were already enabled. */
3258 static bool blk_poll_stats_enable(struct request_queue *q)
3259 {
3260         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3261             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3262                 return true;
3263         blk_stat_add_callback(q, q->poll_cb);
3264         return false;
3265 }
3266
3267 static void blk_mq_poll_stats_start(struct request_queue *q)
3268 {
3269         /*
3270          * We don't arm the callback if polling stats are not enabled or the
3271          * callback is already active.
3272          */
3273         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3274             blk_stat_is_active(q->poll_cb))
3275                 return;
3276
3277         blk_stat_activate_msecs(q->poll_cb, 100);
3278 }
3279
3280 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3281 {
3282         struct request_queue *q = cb->data;
3283         int bucket;
3284
3285         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3286                 if (cb->stat[bucket].nr_samples)
3287                         q->poll_stat[bucket] = cb->stat[bucket];
3288         }
3289 }
3290
3291 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3292                                        struct blk_mq_hw_ctx *hctx,
3293                                        struct request *rq)
3294 {
3295         unsigned long ret = 0;
3296         int bucket;
3297
3298         /*
3299          * If stats collection isn't on, don't sleep but turn it on for
3300          * future users
3301          */
3302         if (!blk_poll_stats_enable(q))
3303                 return 0;
3304
3305         /*
3306          * As an optimistic guess, use half of the mean service time
3307          * for this type of request. We can (and should) make this smarter.
3308          * For instance, if the completion latencies are tight, we can
3309          * get closer than just half the mean. This is especially
3310          * important on devices where the completion latencies are longer
3311          * than ~10 usec. We do use the stats for the relevant IO size
3312          * if available which does lead to better estimates.
3313          */
3314         bucket = blk_mq_poll_stats_bkt(rq);
3315         if (bucket < 0)
3316                 return ret;
3317
3318         if (q->poll_stat[bucket].nr_samples)
3319                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3320
3321         return ret;
3322 }
3323
3324 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3325                                      struct blk_mq_hw_ctx *hctx,
3326                                      struct request *rq)
3327 {
3328         struct hrtimer_sleeper hs;
3329         enum hrtimer_mode mode;
3330         unsigned int nsecs;
3331         ktime_t kt;
3332
3333         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3334                 return false;
3335
3336         /*
3337          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3338          *
3339          *  0:  use half of prev avg
3340          * >0:  use this specific value
3341          */
3342         if (q->poll_nsec > 0)
3343                 nsecs = q->poll_nsec;
3344         else
3345                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3346
3347         if (!nsecs)
3348                 return false;
3349
3350         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3351
3352         /*
3353          * This will be replaced with the stats tracking code, using
3354          * 'avg_completion_time / 2' as the pre-sleep target.
3355          */
3356         kt = nsecs;
3357
3358         mode = HRTIMER_MODE_REL;
3359         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3360         hrtimer_set_expires(&hs.timer, kt);
3361
3362         hrtimer_init_sleeper(&hs, current);
3363         do {
3364                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3365                         break;
3366                 set_current_state(TASK_UNINTERRUPTIBLE);
3367                 hrtimer_start_expires(&hs.timer, mode);
3368                 if (hs.task)
3369                         io_schedule();
3370                 hrtimer_cancel(&hs.timer);
3371                 mode = HRTIMER_MODE_ABS;
3372         } while (hs.task && !signal_pending(current));
3373
3374         __set_current_state(TASK_RUNNING);
3375         destroy_hrtimer_on_stack(&hs.timer);
3376         return true;
3377 }
3378
3379 static bool blk_mq_poll_hybrid(struct request_queue *q,
3380                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3381 {
3382         struct request *rq;
3383
3384         if (q->poll_nsec == -1)
3385                 return false;
3386
3387         if (!blk_qc_t_is_internal(cookie))
3388                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3389         else {
3390                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3391                 /*
3392                  * With scheduling, if the request has completed, we'll
3393                  * get a NULL return here, as we clear the sched tag when
3394                  * that happens. The request still remains valid, like always,
3395                  * so we should be safe with just the NULL check.
3396                  */
3397                 if (!rq)
3398                         return false;
3399         }
3400
3401         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3402 }
3403
3404 /**
3405  * blk_poll - poll for IO completions
3406  * @q:  the queue
3407  * @cookie: cookie passed back at IO submission time
3408  * @spin: whether to spin for completions
3409  *
3410  * Description:
3411  *    Poll for completions on the passed in queue. Returns number of
3412  *    completed entries found. If @spin is true, then blk_poll will continue
3413  *    looping until at least one completion is found, unless the task is
3414  *    otherwise marked running (or we need to reschedule).
3415  */
3416 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3417 {
3418         struct blk_mq_hw_ctx *hctx;
3419         long state;
3420
3421         if (!blk_qc_t_valid(cookie) ||
3422             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3423                 return 0;
3424
3425         if (current->plug)
3426                 blk_flush_plug_list(current->plug, false);
3427
3428         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3429
3430         /*
3431          * If we sleep, have the caller restart the poll loop to reset
3432          * the state. Like for the other success return cases, the
3433          * caller is responsible for checking if the IO completed. If
3434          * the IO isn't complete, we'll get called again and will go
3435          * straight to the busy poll loop.
3436          */
3437         if (blk_mq_poll_hybrid(q, hctx, cookie))
3438                 return 1;
3439
3440         hctx->poll_considered++;
3441
3442         state = current->state;
3443         do {
3444                 int ret;
3445
3446                 hctx->poll_invoked++;
3447
3448                 ret = q->mq_ops->poll(hctx);
3449                 if (ret > 0) {
3450                         hctx->poll_success++;
3451                         __set_current_state(TASK_RUNNING);
3452                         return ret;
3453                 }
3454
3455                 if (signal_pending_state(state, current))
3456                         __set_current_state(TASK_RUNNING);
3457
3458                 if (current->state == TASK_RUNNING)
3459                         return 1;
3460                 if (ret < 0 || !spin)
3461                         break;
3462                 cpu_relax();
3463         } while (!need_resched());
3464
3465         __set_current_state(TASK_RUNNING);
3466         return 0;
3467 }
3468 EXPORT_SYMBOL_GPL(blk_poll);
3469
3470 unsigned int blk_mq_rq_cpu(struct request *rq)
3471 {
3472         return rq->mq_ctx->cpu;
3473 }
3474 EXPORT_SYMBOL(blk_mq_rq_cpu);
3475
3476 static int __init blk_mq_init(void)
3477 {
3478         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3479                                 blk_mq_hctx_notify_dead);
3480         return 0;
3481 }
3482 subsys_initcall(blk_mq_init);