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-stat.h"
  37 #include "blk-wbt.h"
  38 #include "blk-mq-sched.h"
  39
  40 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie);
  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         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  78                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  79 }
  80
  81 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  82                                       struct blk_mq_ctx *ctx)
  83 {
  84         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  85 }
  86
  87 struct mq_inflight {
  88         struct hd_struct *part;
  89         unsigned int *inflight;
  90 };
  91
  92 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  93                                   struct request *rq, void *priv,
  94                                   bool reserved)
  95 {
  96         struct mq_inflight *mi = priv;
  97
  98         if (blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) {
  99                 /*
 100                  * index[0] counts the specific partition that was asked
 101                  * for. index[1] counts the ones that are active on the
 102                  * whole device, so increment that if mi->part is indeed
 103                  * a partition, and not a whole device.
 104                  */
 105                 if (rq->part == mi->part)
 106                         mi->inflight[0]++;
 107                 if (mi->part->partno)
 108                         mi->inflight[1]++;
 109         }
 110 }
 111
 112 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
 113                       unsigned int inflight[2])
 114 {
 115         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 116
 117         inflight[0] = inflight[1] = 0;
 118         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 119 }
 120
 121 void blk_freeze_queue_start(struct request_queue *q)
 122 {
 123         int freeze_depth;
 124
 125         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 126         if (freeze_depth == 1) {
 127                 percpu_ref_kill(&q->q_usage_counter);
 128                 if (q->mq_ops)
 129                         blk_mq_run_hw_queues(q, false);
 130         }
 131 }
 132 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 133
 134 void blk_mq_freeze_queue_wait(struct request_queue *q)
 135 {
 136         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 137 }
 138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 139
 140 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 141                                      unsigned long timeout)
 142 {
 143         return wait_event_timeout(q->mq_freeze_wq,
 144                                         percpu_ref_is_zero(&q->q_usage_counter),
 145                                         timeout);
 146 }
 147 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 148
 149 /*
 150  * Guarantee no request is in use, so we can change any data structure of
 151  * the queue afterward.
 152  */
 153 void blk_freeze_queue(struct request_queue *q)
 154 {
 155         /*
 156          * In the !blk_mq case we are only calling this to kill the
 157          * q_usage_counter, otherwise this increases the freeze depth
 158          * and waits for it to return to zero.  For this reason there is
 159          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 160          * exported to drivers as the only user for unfreeze is blk_mq.
 161          */
 162         blk_freeze_queue_start(q);
 163         if (!q->mq_ops)
 164                 blk_drain_queue(q);
 165         blk_mq_freeze_queue_wait(q);
 166 }
 167
 168 void blk_mq_freeze_queue(struct request_queue *q)
 169 {
 170         /*
 171          * ...just an alias to keep freeze and unfreeze actions balanced
 172          * in the blk_mq_* namespace
 173          */
 174         blk_freeze_queue(q);
 175 }
 176 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 177
 178 void blk_mq_unfreeze_queue(struct request_queue *q)
 179 {
 180         int freeze_depth;
 181
 182         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 183         WARN_ON_ONCE(freeze_depth < 0);
 184         if (!freeze_depth) {
 185                 percpu_ref_reinit(&q->q_usage_counter);
 186                 wake_up_all(&q->mq_freeze_wq);
 187         }
 188 }
 189 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 190
 191 /*
 192  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 193  * mpt3sas driver such that this function can be removed.
 194  */
 195 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 196 {
 197         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 198 }
 199 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 200
 201 /**
 202  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 203  * @q: request queue.
 204  *
 205  * Note: this function does not prevent that the struct request end_io()
 206  * callback function is invoked. Once this function is returned, we make
 207  * sure no dispatch can happen until the queue is unquiesced via
 208  * blk_mq_unquiesce_queue().
 209  */
 210 void blk_mq_quiesce_queue(struct request_queue *q)
 211 {
 212         struct blk_mq_hw_ctx *hctx;
 213         unsigned int i;
 214         bool rcu = false;
 215
 216         blk_mq_quiesce_queue_nowait(q);
 217
 218         queue_for_each_hw_ctx(q, hctx, i) {
 219                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 220                         synchronize_srcu(hctx->srcu);
 221                 else
 222                         rcu = true;
 223         }
 224         if (rcu)
 225                 synchronize_rcu();
 226 }
 227 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 228
 229 /*
 230  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 231  * @q: request queue.
 232  *
 233  * This function recovers queue into the state before quiescing
 234  * which is done by blk_mq_quiesce_queue.
 235  */
 236 void blk_mq_unquiesce_queue(struct request_queue *q)
 237 {
 238         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 239
 240         /* dispatch requests which are inserted during quiescing */
 241         blk_mq_run_hw_queues(q, true);
 242 }
 243 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 244
 245 void blk_mq_wake_waiters(struct request_queue *q)
 246 {
 247         struct blk_mq_hw_ctx *hctx;
 248         unsigned int i;
 249
 250         queue_for_each_hw_ctx(q, hctx, i)
 251                 if (blk_mq_hw_queue_mapped(hctx))
 252                         blk_mq_tag_wakeup_all(hctx->tags, true);
 253 }
 254
 255 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 256 {
 257         return blk_mq_has_free_tags(hctx->tags);
 258 }
 259 EXPORT_SYMBOL(blk_mq_can_queue);
 260
 261 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 262                 unsigned int tag, unsigned int op)
 263 {
 264         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 265         struct request *rq = tags->static_rqs[tag];
 266         req_flags_t rq_flags = 0;
 267
 268         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 269                 rq->tag = -1;
 270                 rq->internal_tag = tag;
 271         } else {
 272                 if (blk_mq_tag_busy(data->hctx)) {
 273                         rq_flags = RQF_MQ_INFLIGHT;
 274                         atomic_inc(&data->hctx->nr_active);
 275                 }
 276                 rq->tag = tag;
 277                 rq->internal_tag = -1;
 278                 data->hctx->tags->rqs[rq->tag] = rq;
 279         }
 280
 281         /* csd/requeue_work/fifo_time is initialized before use */
 282         rq->q = data->q;
 283         rq->mq_ctx = data->ctx;
 284         rq->rq_flags = rq_flags;
 285         rq->cpu = -1;
 286         rq->cmd_flags = op;
 287         if (data->flags & BLK_MQ_REQ_PREEMPT)
 288                 rq->rq_flags |= RQF_PREEMPT;
 289         if (blk_queue_io_stat(data->q))
 290                 rq->rq_flags |= RQF_IO_STAT;
 291         INIT_LIST_HEAD(&rq->queuelist);
 292         INIT_HLIST_NODE(&rq->hash);
 293         RB_CLEAR_NODE(&rq->rb_node);
 294         rq->rq_disk = NULL;
 295         rq->part = NULL;
 296         rq->start_time = jiffies;
 297         rq->nr_phys_segments = 0;
 298 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 299         rq->nr_integrity_segments = 0;
 300 #endif
 301         rq->special = NULL;
 302         /* tag was already set */
 303         rq->extra_len = 0;
 304         rq->__deadline = 0;
 305
 306         INIT_LIST_HEAD(&rq->timeout_list);
 307         rq->timeout = 0;
 308
 309         rq->end_io = NULL;
 310         rq->end_io_data = NULL;
 311         rq->next_rq = NULL;
 312
 313 #ifdef CONFIG_BLK_CGROUP
 314         rq->rl = NULL;
 315         set_start_time_ns(rq);
 316         rq->io_start_time_ns = 0;
 317 #endif
 318
 319         data->ctx->rq_dispatched[op_is_sync(op)]++;
 320         return rq;
 321 }
 322
 323 static struct request *blk_mq_get_request(struct request_queue *q,
 324                 struct bio *bio, unsigned int op,
 325                 struct blk_mq_alloc_data *data)
 326 {
 327         struct elevator_queue *e = q->elevator;
 328         struct request *rq;
 329         unsigned int tag;
 330         bool put_ctx_on_error = false;
 331
 332         blk_queue_enter_live(q);
 333         data->q = q;
 334         if (likely(!data->ctx)) {
 335                 data->ctx = blk_mq_get_ctx(q);
 336                 put_ctx_on_error = true;
 337         }
 338         if (likely(!data->hctx))
 339                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 340         if (op & REQ_NOWAIT)
 341                 data->flags |= BLK_MQ_REQ_NOWAIT;
 342
 343         if (e) {
 344                 data->flags |= BLK_MQ_REQ_INTERNAL;
 345
 346                 /*
 347                  * Flush requests are special and go directly to the
 348                  * dispatch list.
 349                  */
 350                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
 351                         e->type->ops.mq.limit_depth(op, data);
 352         }
 353
 354         tag = blk_mq_get_tag(data);
 355         if (tag == BLK_MQ_TAG_FAIL) {
 356                 if (put_ctx_on_error) {
 357                         blk_mq_put_ctx(data->ctx);
 358                         data->ctx = NULL;
 359                 }
 360                 blk_queue_exit(q);
 361                 return NULL;
 362         }
 363
 364         rq = blk_mq_rq_ctx_init(data, tag, op);
 365         if (!op_is_flush(op)) {
 366                 rq->elv.icq = NULL;
 367                 if (e && e->type->ops.mq.prepare_request) {
 368                         if (e->type->icq_cache && rq_ioc(bio))
 369                                 blk_mq_sched_assign_ioc(rq, bio);
 370
 371                         e->type->ops.mq.prepare_request(rq, bio);
 372                         rq->rq_flags |= RQF_ELVPRIV;
 373                 }
 374         }
 375         data->hctx->queued++;
 376         return rq;
 377 }
 378
 379 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 380                 blk_mq_req_flags_t flags)
 381 {
 382         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 383         struct request *rq;
 384         int ret;
 385
 386         ret = blk_queue_enter(q, flags);
 387         if (ret)
 388                 return ERR_PTR(ret);
 389
 390         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 391         blk_queue_exit(q);
 392
 393         if (!rq)
 394                 return ERR_PTR(-EWOULDBLOCK);
 395
 396         blk_mq_put_ctx(alloc_data.ctx);
 397
 398         rq->__data_len = 0;
 399         rq->__sector = (sector_t) -1;
 400         rq->bio = rq->biotail = NULL;
 401         return rq;
 402 }
 403 EXPORT_SYMBOL(blk_mq_alloc_request);
 404
 405 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 406         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 407 {
 408         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 409         struct request *rq;
 410         unsigned int cpu;
 411         int ret;
 412
 413         /*
 414          * If the tag allocator sleeps we could get an allocation for a
 415          * different hardware context.  No need to complicate the low level
 416          * allocator for this for the rare use case of a command tied to
 417          * a specific queue.
 418          */
 419         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 420                 return ERR_PTR(-EINVAL);
 421
 422         if (hctx_idx >= q->nr_hw_queues)
 423                 return ERR_PTR(-EIO);
 424
 425         ret = blk_queue_enter(q, flags);
 426         if (ret)
 427                 return ERR_PTR(ret);
 428
 429         /*
 430          * Check if the hardware context is actually mapped to anything.
 431          * If not tell the caller that it should skip this queue.
 432          */
 433         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 434         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 435                 blk_queue_exit(q);
 436                 return ERR_PTR(-EXDEV);
 437         }
 438         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 439         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 440
 441         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 442         blk_queue_exit(q);
 443
 444         if (!rq)
 445                 return ERR_PTR(-EWOULDBLOCK);
 446
 447         return rq;
 448 }
 449 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 450
 451 void blk_mq_free_request(struct request *rq)
 452 {
 453         struct request_queue *q = rq->q;
 454         struct elevator_queue *e = q->elevator;
 455         struct blk_mq_ctx *ctx = rq->mq_ctx;
 456         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 457         const int sched_tag = rq->internal_tag;
 458
 459         if (rq->rq_flags & RQF_ELVPRIV) {
 460                 if (e && e->type->ops.mq.finish_request)
 461                         e->type->ops.mq.finish_request(rq);
 462                 if (rq->elv.icq) {
 463                         put_io_context(rq->elv.icq->ioc);
 464                         rq->elv.icq = NULL;
 465                 }
 466         }
 467
 468         ctx->rq_completed[rq_is_sync(rq)]++;
 469         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 470                 atomic_dec(&hctx->nr_active);
 471
 472         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 473                 laptop_io_completion(q->backing_dev_info);
 474
 475         wbt_done(q->rq_wb, &rq->issue_stat);
 476
 477         if (blk_rq_rl(rq))
 478                 blk_put_rl(blk_rq_rl(rq));
 479
 480         blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
 481         if (rq->tag != -1)
 482                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 483         if (sched_tag != -1)
 484                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 485         blk_mq_sched_restart(hctx);
 486         blk_queue_exit(q);
 487 }
 488 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 489
 490 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 491 {
 492         blk_account_io_done(rq);
 493
 494         if (rq->end_io) {
 495                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 496                 rq->end_io(rq, error);
 497         } else {
 498                 if (unlikely(blk_bidi_rq(rq)))
 499                         blk_mq_free_request(rq->next_rq);
 500                 blk_mq_free_request(rq);
 501         }
 502 }
 503 EXPORT_SYMBOL(__blk_mq_end_request);
 504
 505 void blk_mq_end_request(struct request *rq, blk_status_t error)
 506 {
 507         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 508                 BUG();
 509         __blk_mq_end_request(rq, error);
 510 }
 511 EXPORT_SYMBOL(blk_mq_end_request);
 512
 513 static void __blk_mq_complete_request_remote(void *data)
 514 {
 515         struct request *rq = data;
 516
 517         rq->q->softirq_done_fn(rq);
 518 }
 519
 520 static void __blk_mq_complete_request(struct request *rq)
 521 {
 522         struct blk_mq_ctx *ctx = rq->mq_ctx;
 523         bool shared = false;
 524         int cpu;
 525
 526         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
 527         blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
 528
 529         if (rq->internal_tag != -1)
 530                 blk_mq_sched_completed_request(rq);
 531         if (rq->rq_flags & RQF_STATS) {
 532                 blk_mq_poll_stats_start(rq->q);
 533                 blk_stat_add(rq);
 534         }
 535
 536         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 537                 rq->q->softirq_done_fn(rq);
 538                 return;
 539         }
 540
 541         cpu = get_cpu();
 542         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 543                 shared = cpus_share_cache(cpu, ctx->cpu);
 544
 545         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 546                 rq->csd.func = __blk_mq_complete_request_remote;
 547                 rq->csd.info = rq;
 548                 rq->csd.flags = 0;
 549                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 550         } else {
 551                 rq->q->softirq_done_fn(rq);
 552         }
 553         put_cpu();
 554 }
 555
 556 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 557         __releases(hctx->srcu)
 558 {
 559         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 560                 rcu_read_unlock();
 561         else
 562                 srcu_read_unlock(hctx->srcu, srcu_idx);
 563 }
 564
 565 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 566         __acquires(hctx->srcu)
 567 {
 568         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 569                 /* shut up gcc false positive */
 570                 *srcu_idx = 0;
 571                 rcu_read_lock();
 572         } else
 573                 *srcu_idx = srcu_read_lock(hctx->srcu);
 574 }
 575
 576 static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
 577 {
 578         unsigned long flags;
 579
 580         /*
 581          * blk_mq_rq_aborted_gstate() is used from the completion path and
 582          * can thus be called from irq context.  u64_stats_fetch in the
 583          * middle of update on the same CPU leads to lockup.  Disable irq
 584          * while updating.
 585          */
 586         local_irq_save(flags);
 587         u64_stats_update_begin(&rq->aborted_gstate_sync);
 588         rq->aborted_gstate = gstate;
 589         u64_stats_update_end(&rq->aborted_gstate_sync);
 590         local_irq_restore(flags);
 591 }
 592
 593 static u64 blk_mq_rq_aborted_gstate(struct request *rq)
 594 {
 595         unsigned int start;
 596         u64 aborted_gstate;
 597
 598         do {
 599                 start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
 600                 aborted_gstate = rq->aborted_gstate;
 601         } while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
 602
 603         return aborted_gstate;
 604 }
 605
 606 /**
 607  * blk_mq_complete_request - end I/O on a request
 608  * @rq:         the request being processed
 609  *
 610  * Description:
 611  *      Ends all I/O on a request. It does not handle partial completions.
 612  *      The actual completion happens out-of-order, through a IPI handler.
 613  **/
 614 void blk_mq_complete_request(struct request *rq)
 615 {
 616         struct request_queue *q = rq->q;
 617         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
 618         int srcu_idx;
 619
 620         if (unlikely(blk_should_fake_timeout(q)))
 621                 return;
 622
 623         /*
 624          * If @rq->aborted_gstate equals the current instance, timeout is
 625          * claiming @rq and we lost.  This is synchronized through
 626          * hctx_lock().  See blk_mq_timeout_work() for details.
 627          *
 628          * Completion path never blocks and we can directly use RCU here
 629          * instead of hctx_lock() which can be either RCU or SRCU.
 630          * However, that would complicate paths which want to synchronize
 631          * against us.  Let stay in sync with the issue path so that
 632          * hctx_lock() covers both issue and completion paths.
 633          */
 634         hctx_lock(hctx, &srcu_idx);
 635         if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
 636                 __blk_mq_complete_request(rq);
 637         hctx_unlock(hctx, srcu_idx);
 638 }
 639 EXPORT_SYMBOL(blk_mq_complete_request);
 640
 641 int blk_mq_request_started(struct request *rq)
 642 {
 643         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 644 }
 645 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 646
 647 void blk_mq_start_request(struct request *rq)
 648 {
 649         struct request_queue *q = rq->q;
 650
 651         blk_mq_sched_started_request(rq);
 652
 653         trace_block_rq_issue(q, rq);
 654
 655         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 656                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
 657                 rq->rq_flags |= RQF_STATS;
 658                 wbt_issue(q->rq_wb, &rq->issue_stat);
 659         }
 660
 661         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 662
 663         /*
 664          * Mark @rq in-flight which also advances the generation number,
 665          * and register for timeout.  Protect with a seqcount to allow the
 666          * timeout path to read both @rq->gstate and @rq->deadline
 667          * coherently.
 668          *
 669          * This is the only place where a request is marked in-flight.  If
 670          * the timeout path reads an in-flight @rq->gstate, the
 671          * @rq->deadline it reads together under @rq->gstate_seq is
 672          * guaranteed to be the matching one.
 673          */
 674         preempt_disable();
 675         write_seqcount_begin(&rq->gstate_seq);
 676
 677         blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
 678         blk_add_timer(rq);
 679
 680         write_seqcount_end(&rq->gstate_seq);
 681         preempt_enable();
 682
 683         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 684                 /*
 685                  * Make sure space for the drain appears.  We know we can do
 686                  * this because max_hw_segments has been adjusted to be one
 687                  * fewer than the device can handle.
 688                  */
 689                 rq->nr_phys_segments++;
 690         }
 691 }
 692 EXPORT_SYMBOL(blk_mq_start_request);
 693
 694 /*
 695  * When we reach here because queue is busy, it's safe to change the state
 696  * to IDLE without checking @rq->aborted_gstate because we should still be
 697  * holding the RCU read lock and thus protected against timeout.
 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         wbt_requeue(q->rq_wb, &rq->issue_stat);
 707
 708         if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
 709                 blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
 710                 if (q->dma_drain_size && blk_rq_bytes(rq))
 711                         rq->nr_phys_segments--;
 712         }
 713 }
 714
 715 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 716 {
 717         __blk_mq_requeue_request(rq);
 718
 719         /* this request will be re-inserted to io scheduler queue */
 720         blk_mq_sched_requeue_request(rq);
 721
 722         BUG_ON(blk_queued_rq(rq));
 723         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 724 }
 725 EXPORT_SYMBOL(blk_mq_requeue_request);
 726
 727 static void blk_mq_requeue_work(struct work_struct *work)
 728 {
 729         struct request_queue *q =
 730                 container_of(work, struct request_queue, requeue_work.work);
 731         LIST_HEAD(rq_list);
 732         struct request *rq, *next;
 733
 734         spin_lock_irq(&q->requeue_lock);
 735         list_splice_init(&q->requeue_list, &rq_list);
 736         spin_unlock_irq(&q->requeue_lock);
 737
 738         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 739                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 740                         continue;
 741
 742                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 743                 list_del_init(&rq->queuelist);
 744                 blk_mq_sched_insert_request(rq, true, false, false);
 745         }
 746
 747         while (!list_empty(&rq_list)) {
 748                 rq = list_entry(rq_list.next, struct request, queuelist);
 749                 list_del_init(&rq->queuelist);
 750                 blk_mq_sched_insert_request(rq, false, false, false);
 751         }
 752
 753         blk_mq_run_hw_queues(q, false);
 754 }
 755
 756 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 757                                 bool kick_requeue_list)
 758 {
 759         struct request_queue *q = rq->q;
 760         unsigned long flags;
 761
 762         /*
 763          * We abuse this flag that is otherwise used by the I/O scheduler to
 764          * request head insertion from the workqueue.
 765          */
 766         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 767
 768         spin_lock_irqsave(&q->requeue_lock, flags);
 769         if (at_head) {
 770                 rq->rq_flags |= RQF_SOFTBARRIER;
 771                 list_add(&rq->queuelist, &q->requeue_list);
 772         } else {
 773                 list_add_tail(&rq->queuelist, &q->requeue_list);
 774         }
 775         spin_unlock_irqrestore(&q->requeue_lock, flags);
 776
 777         if (kick_requeue_list)
 778                 blk_mq_kick_requeue_list(q);
 779 }
 780 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 781
 782 void blk_mq_kick_requeue_list(struct request_queue *q)
 783 {
 784         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 785 }
 786 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 787
 788 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 789                                     unsigned long msecs)
 790 {
 791         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 792                                     msecs_to_jiffies(msecs));
 793 }
 794 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 795
 796 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 797 {
 798         if (tag < tags->nr_tags) {
 799                 prefetch(tags->rqs[tag]);
 800                 return tags->rqs[tag];
 801         }
 802
 803         return NULL;
 804 }
 805 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 806
 807 struct blk_mq_timeout_data {
 808         unsigned long next;
 809         unsigned int next_set;
 810         unsigned int nr_expired;
 811 };
 812
 813 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 814 {
 815         const struct blk_mq_ops *ops = req->q->mq_ops;
 816         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 817
 818         req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
 819
 820         if (ops->timeout)
 821                 ret = ops->timeout(req, reserved);
 822
 823         switch (ret) {
 824         case BLK_EH_HANDLED:
 825                 __blk_mq_complete_request(req);
 826                 break;
 827         case BLK_EH_RESET_TIMER:
 828                 /*
 829                  * As nothing prevents from completion happening while
 830                  * ->aborted_gstate is set, this may lead to ignored
 831                  * completions and further spurious timeouts.
 832                  */
 833                 blk_mq_rq_update_aborted_gstate(req, 0);
 834                 blk_add_timer(req);
 835                 break;
 836         case BLK_EH_NOT_HANDLED:
 837                 break;
 838         default:
 839                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 840                 break;
 841         }
 842 }
 843
 844 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 845                 struct request *rq, void *priv, bool reserved)
 846 {
 847         struct blk_mq_timeout_data *data = priv;
 848         unsigned long gstate, deadline;
 849         int start;
 850
 851         might_sleep();
 852
 853         if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
 854                 return;
 855
 856         /* read coherent snapshots of @rq->state_gen and @rq->deadline */
 857         while (true) {
 858                 start = read_seqcount_begin(&rq->gstate_seq);
 859                 gstate = READ_ONCE(rq->gstate);
 860                 deadline = blk_rq_deadline(rq);
 861                 if (!read_seqcount_retry(&rq->gstate_seq, start))
 862                         break;
 863                 cond_resched();
 864         }
 865
 866         /* if in-flight && overdue, mark for abortion */
 867         if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
 868             time_after_eq(jiffies, deadline)) {
 869                 blk_mq_rq_update_aborted_gstate(rq, gstate);
 870                 data->nr_expired++;
 871                 hctx->nr_expired++;
 872         } else if (!data->next_set || time_after(data->next, deadline)) {
 873                 data->next = deadline;
 874                 data->next_set = 1;
 875         }
 876 }
 877
 878 static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
 879                 struct request *rq, void *priv, bool reserved)
 880 {
 881         /*
 882          * We marked @rq->aborted_gstate and waited for RCU.  If there were
 883          * completions that we lost to, they would have finished and
 884          * updated @rq->gstate by now; otherwise, the completion path is
 885          * now guaranteed to see @rq->aborted_gstate and yield.  If
 886          * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
 887          */
 888         if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
 889             READ_ONCE(rq->gstate) == rq->aborted_gstate)
 890                 blk_mq_rq_timed_out(rq, reserved);
 891 }
 892
 893 static void blk_mq_timeout_work(struct work_struct *work)
 894 {
 895         struct request_queue *q =
 896                 container_of(work, struct request_queue, timeout_work);
 897         struct blk_mq_timeout_data data = {
 898                 .next           = 0,
 899                 .next_set       = 0,
 900                 .nr_expired     = 0,
 901         };
 902         struct blk_mq_hw_ctx *hctx;
 903         int i;
 904
 905         /* A deadlock might occur if a request is stuck requiring a
 906          * timeout at the same time a queue freeze is waiting
 907          * completion, since the timeout code would not be able to
 908          * acquire the queue reference here.
 909          *
 910          * That's why we don't use blk_queue_enter here; instead, we use
 911          * percpu_ref_tryget directly, because we need to be able to
 912          * obtain a reference even in the short window between the queue
 913          * starting to freeze, by dropping the first reference in
 914          * blk_freeze_queue_start, and the moment the last request is
 915          * consumed, marked by the instant q_usage_counter reaches
 916          * zero.
 917          */
 918         if (!percpu_ref_tryget(&q->q_usage_counter))
 919                 return;
 920
 921         /* scan for the expired ones and set their ->aborted_gstate */
 922         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 923
 924         if (data.nr_expired) {
 925                 bool has_rcu = false;
 926
 927                 /*
 928                  * Wait till everyone sees ->aborted_gstate.  The
 929                  * sequential waits for SRCUs aren't ideal.  If this ever
 930                  * becomes a problem, we can add per-hw_ctx rcu_head and
 931                  * wait in parallel.
 932                  */
 933                 queue_for_each_hw_ctx(q, hctx, i) {
 934                         if (!hctx->nr_expired)
 935                                 continue;
 936
 937                         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 938                                 has_rcu = true;
 939                         else
 940                                 synchronize_srcu(hctx->srcu);
 941
 942                         hctx->nr_expired = 0;
 943                 }
 944                 if (has_rcu)
 945                         synchronize_rcu();
 946
 947                 /* terminate the ones we won */
 948                 blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
 949         }
 950
 951         if (data.next_set) {
 952                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 953                 mod_timer(&q->timeout, data.next);
 954         } else {
 955                 /*
 956                  * Request timeouts are handled as a forward rolling timer. If
 957                  * we end up here it means that no requests are pending and
 958                  * also that no request has been pending for a while. Mark
 959                  * each hctx as idle.
 960                  */
 961                 queue_for_each_hw_ctx(q, hctx, i) {
 962                         /* the hctx may be unmapped, so check it here */
 963                         if (blk_mq_hw_queue_mapped(hctx))
 964                                 blk_mq_tag_idle(hctx);
 965                 }
 966         }
 967         blk_queue_exit(q);
 968 }
 969
 970 struct flush_busy_ctx_data {
 971         struct blk_mq_hw_ctx *hctx;
 972         struct list_head *list;
 973 };
 974
 975 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 976 {
 977         struct flush_busy_ctx_data *flush_data = data;
 978         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 979         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 980
 981         spin_lock(&ctx->lock);
 982         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 983         sbitmap_clear_bit(sb, bitnr);
 984         spin_unlock(&ctx->lock);
 985         return true;
 986 }
 987
 988 /*
 989  * Process software queues that have been marked busy, splicing them
 990  * to the for-dispatch
 991  */
 992 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 993 {
 994         struct flush_busy_ctx_data data = {
 995                 .hctx = hctx,
 996                 .list = list,
 997         };
 998
 999         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1000 }
1001 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1002
1003 struct dispatch_rq_data {
1004         struct blk_mq_hw_ctx *hctx;
1005         struct request *rq;
1006 };
1007
1008 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1009                 void *data)
1010 {
1011         struct dispatch_rq_data *dispatch_data = data;
1012         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1013         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1014
1015         spin_lock(&ctx->lock);
1016         if (unlikely(!list_empty(&ctx->rq_list))) {
1017                 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
1018                 list_del_init(&dispatch_data->rq->queuelist);
1019                 if (list_empty(&ctx->rq_list))
1020                         sbitmap_clear_bit(sb, bitnr);
1021         }
1022         spin_unlock(&ctx->lock);
1023
1024         return !dispatch_data->rq;
1025 }
1026
1027 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1028                                         struct blk_mq_ctx *start)
1029 {
1030         unsigned off = start ? start->index_hw : 0;
1031         struct dispatch_rq_data data = {
1032                 .hctx = hctx,
1033                 .rq   = NULL,
1034         };
1035
1036         __sbitmap_for_each_set(&hctx->ctx_map, off,
1037                                dispatch_rq_from_ctx, &data);
1038
1039         return data.rq;
1040 }
1041
1042 static inline unsigned int queued_to_index(unsigned int queued)
1043 {
1044         if (!queued)
1045                 return 0;
1046
1047         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1048 }
1049
1050 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
1051                            bool wait)
1052 {
1053         struct blk_mq_alloc_data data = {
1054                 .q = rq->q,
1055                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
1056                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
1057         };
1058
1059         might_sleep_if(wait);
1060
1061         if (rq->tag != -1)
1062                 goto done;
1063
1064         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1065                 data.flags |= BLK_MQ_REQ_RESERVED;
1066
1067         rq->tag = blk_mq_get_tag(&data);
1068         if (rq->tag >= 0) {
1069                 if (blk_mq_tag_busy(data.hctx)) {
1070                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1071                         atomic_inc(&data.hctx->nr_active);
1072                 }
1073                 data.hctx->tags->rqs[rq->tag] = rq;
1074         }
1075
1076 done:
1077         if (hctx)
1078                 *hctx = data.hctx;
1079         return rq->tag != -1;
1080 }
1081
1082 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1083                                 int flags, void *key)
1084 {
1085         struct blk_mq_hw_ctx *hctx;
1086
1087         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1088
1089         list_del_init(&wait->entry);
1090         blk_mq_run_hw_queue(hctx, true);
1091         return 1;
1092 }
1093
1094 /*
1095  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1096  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1097  * restart. For both cases, take care to check the condition again after
1098  * marking us as waiting.
1099  */
1100 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1101                                  struct request *rq)
1102 {
1103         struct blk_mq_hw_ctx *this_hctx = *hctx;
1104         struct sbq_wait_state *ws;
1105         wait_queue_entry_t *wait;
1106         bool ret;
1107
1108         if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1109                 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1110                         set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1111
1112                 /*
1113                  * It's possible that a tag was freed in the window between the
1114                  * allocation failure and adding the hardware queue to the wait
1115                  * queue.
1116                  *
1117                  * Don't clear RESTART here, someone else could have set it.
1118                  * At most this will cost an extra queue run.
1119                  */
1120                 return blk_mq_get_driver_tag(rq, hctx, false);
1121         }
1122
1123         wait = &this_hctx->dispatch_wait;
1124         if (!list_empty_careful(&wait->entry))
1125                 return false;
1126
1127         spin_lock(&this_hctx->lock);
1128         if (!list_empty(&wait->entry)) {
1129                 spin_unlock(&this_hctx->lock);
1130                 return false;
1131         }
1132
1133         ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1134         add_wait_queue(&ws->wait, wait);
1135
1136         /*
1137          * It's possible that a tag was freed in the window between the
1138          * allocation failure and adding the hardware queue to the wait
1139          * queue.
1140          */
1141         ret = blk_mq_get_driver_tag(rq, hctx, false);
1142         if (!ret) {
1143                 spin_unlock(&this_hctx->lock);
1144                 return false;
1145         }
1146
1147         /*
1148          * We got a tag, remove ourselves from the wait queue to ensure
1149          * someone else gets the wakeup.
1150          */
1151         spin_lock_irq(&ws->wait.lock);
1152         list_del_init(&wait->entry);
1153         spin_unlock_irq(&ws->wait.lock);
1154         spin_unlock(&this_hctx->lock);
1155
1156         return true;
1157 }
1158
1159 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1160
1161 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1162                              bool got_budget)
1163 {
1164         struct blk_mq_hw_ctx *hctx;
1165         struct request *rq, *nxt;
1166         bool no_tag = false;
1167         int errors, queued;
1168         blk_status_t ret = BLK_STS_OK;
1169
1170         if (list_empty(list))
1171                 return false;
1172
1173         WARN_ON(!list_is_singular(list) && got_budget);
1174
1175         /*
1176          * Now process all the entries, sending them to the driver.
1177          */
1178         errors = queued = 0;
1179         do {
1180                 struct blk_mq_queue_data bd;
1181
1182                 rq = list_first_entry(list, struct request, queuelist);
1183                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1184                         /*
1185                          * The initial allocation attempt failed, so we need to
1186                          * rerun the hardware queue when a tag is freed. The
1187                          * waitqueue takes care of that. If the queue is run
1188                          * before we add this entry back on the dispatch list,
1189                          * we'll re-run it below.
1190                          */
1191                         if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1192                                 if (got_budget)
1193                                         blk_mq_put_dispatch_budget(hctx);
1194                                 /*
1195                                  * For non-shared tags, the RESTART check
1196                                  * will suffice.
1197                                  */
1198                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1199                                         no_tag = true;
1200                                 break;
1201                         }
1202                 }
1203
1204                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1205                         blk_mq_put_driver_tag(rq);
1206                         break;
1207                 }
1208
1209                 list_del_init(&rq->queuelist);
1210
1211                 bd.rq = rq;
1212
1213                 /*
1214                  * Flag last if we have no more requests, or if we have more
1215                  * but can't assign a driver tag to it.
1216                  */
1217                 if (list_empty(list))
1218                         bd.last = true;
1219                 else {
1220                         nxt = list_first_entry(list, struct request, queuelist);
1221                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1222                 }
1223
1224                 ret = q->mq_ops->queue_rq(hctx, &bd);
1225                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1226                         /*
1227                          * If an I/O scheduler has been configured and we got a
1228                          * driver tag for the next request already, free it
1229                          * again.
1230                          */
1231                         if (!list_empty(list)) {
1232                                 nxt = list_first_entry(list, struct request, queuelist);
1233                                 blk_mq_put_driver_tag(nxt);
1234                         }
1235                         list_add(&rq->queuelist, list);
1236                         __blk_mq_requeue_request(rq);
1237                         break;
1238                 }
1239
1240                 if (unlikely(ret != BLK_STS_OK)) {
1241                         errors++;
1242                         blk_mq_end_request(rq, BLK_STS_IOERR);
1243                         continue;
1244                 }
1245
1246                 queued++;
1247         } while (!list_empty(list));
1248
1249         hctx->dispatched[queued_to_index(queued)]++;
1250
1251         /*
1252          * Any items that need requeuing? Stuff them into hctx->dispatch,
1253          * that is where we will continue on next queue run.
1254          */
1255         if (!list_empty(list)) {
1256                 bool needs_restart;
1257
1258                 spin_lock(&hctx->lock);
1259                 list_splice_init(list, &hctx->dispatch);
1260                 spin_unlock(&hctx->lock);
1261
1262                 /*
1263                  * If SCHED_RESTART was set by the caller of this function and
1264                  * it is no longer set that means that it was cleared by another
1265                  * thread and hence that a queue rerun is needed.
1266                  *
1267                  * If 'no_tag' is set, that means that we failed getting
1268                  * a driver tag with an I/O scheduler attached. If our dispatch
1269                  * waitqueue is no longer active, ensure that we run the queue
1270                  * AFTER adding our entries back to the list.
1271                  *
1272                  * If no I/O scheduler has been configured it is possible that
1273                  * the hardware queue got stopped and restarted before requests
1274                  * were pushed back onto the dispatch list. Rerun the queue to
1275                  * avoid starvation. Notes:
1276                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1277                  *   been stopped before rerunning a queue.
1278                  * - Some but not all block drivers stop a queue before
1279                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1280                  *   and dm-rq.
1281                  *
1282                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1283                  * bit is set, run queue after a delay to avoid IO stalls
1284                  * that could otherwise occur if the queue is idle.
1285                  */
1286                 needs_restart = blk_mq_sched_needs_restart(hctx);
1287                 if (!needs_restart ||
1288                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1289                         blk_mq_run_hw_queue(hctx, true);
1290                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1291                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1292         }
1293
1294         return (queued + errors) != 0;
1295 }
1296
1297 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1298 {
1299         int srcu_idx;
1300
1301         /*
1302          * We should be running this queue from one of the CPUs that
1303          * are mapped to it.
1304          *
1305          * There are at least two related races now between setting
1306          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1307          * __blk_mq_run_hw_queue():
1308          *
1309          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1310          *   but later it becomes online, then this warning is harmless
1311          *   at all
1312          *
1313          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1314          *   but later it becomes offline, then the warning can't be
1315          *   triggered, and we depend on blk-mq timeout handler to
1316          *   handle dispatched requests to this hctx
1317          */
1318         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1319                 cpu_online(hctx->next_cpu)) {
1320                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1321                         raw_smp_processor_id(),
1322                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1323                 dump_stack();
1324         }
1325
1326         /*
1327          * We can't run the queue inline with ints disabled. Ensure that
1328          * we catch bad users of this early.
1329          */
1330         WARN_ON_ONCE(in_interrupt());
1331
1332         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1333
1334         hctx_lock(hctx, &srcu_idx);
1335         blk_mq_sched_dispatch_requests(hctx);
1336         hctx_unlock(hctx, srcu_idx);
1337 }
1338
1339 /*
1340  * It'd be great if the workqueue API had a way to pass
1341  * in a mask and had some smarts for more clever placement.
1342  * For now we just round-robin here, switching for every
1343  * BLK_MQ_CPU_WORK_BATCH queued items.
1344  */
1345 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1346 {
1347         bool tried = false;
1348
1349         if (hctx->queue->nr_hw_queues == 1)
1350                 return WORK_CPU_UNBOUND;
1351
1352         if (--hctx->next_cpu_batch <= 0) {
1353                 int next_cpu;
1354 select_cpu:
1355                 next_cpu = cpumask_next_and(hctx->next_cpu, hctx->cpumask,
1356                                 cpu_online_mask);
1357                 if (next_cpu >= nr_cpu_ids)
1358                         next_cpu = cpumask_first_and(hctx->cpumask,cpu_online_mask);
1359
1360                 /*
1361                  * No online CPU is found, so have to make sure hctx->next_cpu
1362                  * is set correctly for not breaking workqueue.
1363                  */
1364                 if (next_cpu >= nr_cpu_ids)
1365                         hctx->next_cpu = cpumask_first(hctx->cpumask);
1366                 else
1367                         hctx->next_cpu = next_cpu;
1368                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1369         }
1370
1371         /*
1372          * Do unbound schedule if we can't find a online CPU for this hctx,
1373          * and it should only happen in the path of handling CPU DEAD.
1374          */
1375         if (!cpu_online(hctx->next_cpu)) {
1376                 if (!tried) {
1377                         tried = true;
1378                         goto select_cpu;
1379                 }
1380
1381                 /*
1382                  * Make sure to re-select CPU next time once after CPUs
1383                  * in hctx->cpumask become online again.
1384                  */
1385                 hctx->next_cpu_batch = 1;
1386                 return WORK_CPU_UNBOUND;
1387         }
1388         return hctx->next_cpu;
1389 }
1390
1391 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1392                                         unsigned long msecs)
1393 {
1394         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1395                 return;
1396
1397         if (unlikely(blk_mq_hctx_stopped(hctx)))
1398                 return;
1399
1400         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1401                 int cpu = get_cpu();
1402                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1403                         __blk_mq_run_hw_queue(hctx);
1404                         put_cpu();
1405                         return;
1406                 }
1407
1408                 put_cpu();
1409         }
1410
1411         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1412                                     msecs_to_jiffies(msecs));
1413 }
1414
1415 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1416 {
1417         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1418 }
1419 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1420
1421 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1422 {
1423         int srcu_idx;
1424         bool need_run;
1425
1426         /*
1427          * When queue is quiesced, we may be switching io scheduler, or
1428          * updating nr_hw_queues, or other things, and we can't run queue
1429          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1430          *
1431          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1432          * quiesced.
1433          */
1434         hctx_lock(hctx, &srcu_idx);
1435         need_run = !blk_queue_quiesced(hctx->queue) &&
1436                 blk_mq_hctx_has_pending(hctx);
1437         hctx_unlock(hctx, srcu_idx);
1438
1439         if (need_run) {
1440                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1441                 return true;
1442         }
1443
1444         return false;
1445 }
1446 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1447
1448 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1449 {
1450         struct blk_mq_hw_ctx *hctx;
1451         int i;
1452
1453         queue_for_each_hw_ctx(q, hctx, i) {
1454                 if (blk_mq_hctx_stopped(hctx))
1455                         continue;
1456
1457                 blk_mq_run_hw_queue(hctx, async);
1458         }
1459 }
1460 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1461
1462 /**
1463  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1464  * @q: request queue.
1465  *
1466  * The caller is responsible for serializing this function against
1467  * blk_mq_{start,stop}_hw_queue().
1468  */
1469 bool blk_mq_queue_stopped(struct request_queue *q)
1470 {
1471         struct blk_mq_hw_ctx *hctx;
1472         int i;
1473
1474         queue_for_each_hw_ctx(q, hctx, i)
1475                 if (blk_mq_hctx_stopped(hctx))
1476                         return true;
1477
1478         return false;
1479 }
1480 EXPORT_SYMBOL(blk_mq_queue_stopped);
1481
1482 /*
1483  * This function is often used for pausing .queue_rq() by driver when
1484  * there isn't enough resource or some conditions aren't satisfied, and
1485  * BLK_STS_RESOURCE is usually returned.
1486  *
1487  * We do not guarantee that dispatch can be drained or blocked
1488  * after blk_mq_stop_hw_queue() returns. Please use
1489  * blk_mq_quiesce_queue() for that requirement.
1490  */
1491 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1492 {
1493         cancel_delayed_work(&hctx->run_work);
1494
1495         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1496 }
1497 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1498
1499 /*
1500  * This function is often used for pausing .queue_rq() by driver when
1501  * there isn't enough resource or some conditions aren't satisfied, and
1502  * BLK_STS_RESOURCE is usually returned.
1503  *
1504  * We do not guarantee that dispatch can be drained or blocked
1505  * after blk_mq_stop_hw_queues() returns. Please use
1506  * blk_mq_quiesce_queue() for that requirement.
1507  */
1508 void blk_mq_stop_hw_queues(struct request_queue *q)
1509 {
1510         struct blk_mq_hw_ctx *hctx;
1511         int i;
1512
1513         queue_for_each_hw_ctx(q, hctx, i)
1514                 blk_mq_stop_hw_queue(hctx);
1515 }
1516 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1517
1518 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1519 {
1520         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1521
1522         blk_mq_run_hw_queue(hctx, false);
1523 }
1524 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1525
1526 void blk_mq_start_hw_queues(struct request_queue *q)
1527 {
1528         struct blk_mq_hw_ctx *hctx;
1529         int i;
1530
1531         queue_for_each_hw_ctx(q, hctx, i)
1532                 blk_mq_start_hw_queue(hctx);
1533 }
1534 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1535
1536 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1537 {
1538         if (!blk_mq_hctx_stopped(hctx))
1539                 return;
1540
1541         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1542         blk_mq_run_hw_queue(hctx, async);
1543 }
1544 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1545
1546 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1547 {
1548         struct blk_mq_hw_ctx *hctx;
1549         int i;
1550
1551         queue_for_each_hw_ctx(q, hctx, i)
1552                 blk_mq_start_stopped_hw_queue(hctx, async);
1553 }
1554 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1555
1556 static void blk_mq_run_work_fn(struct work_struct *work)
1557 {
1558         struct blk_mq_hw_ctx *hctx;
1559
1560         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1561
1562         /*
1563          * If we are stopped, don't run the queue. The exception is if
1564          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1565          * the STOPPED bit and run it.
1566          */
1567         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1568                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1569                         return;
1570
1571                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1572                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1573         }
1574
1575         __blk_mq_run_hw_queue(hctx);
1576 }
1577
1578
1579 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1580 {
1581         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1582                 return;
1583
1584         /*
1585          * Stop the hw queue, then modify currently delayed work.
1586          * This should prevent us from running the queue prematurely.
1587          * Mark the queue as auto-clearing STOPPED when it runs.
1588          */
1589         blk_mq_stop_hw_queue(hctx);
1590         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1591         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1592                                         &hctx->run_work,
1593                                         msecs_to_jiffies(msecs));
1594 }
1595 EXPORT_SYMBOL(blk_mq_delay_queue);
1596
1597 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1598                                             struct request *rq,
1599                                             bool at_head)
1600 {
1601         struct blk_mq_ctx *ctx = rq->mq_ctx;
1602
1603         lockdep_assert_held(&ctx->lock);
1604
1605         trace_block_rq_insert(hctx->queue, rq);
1606
1607         if (at_head)
1608                 list_add(&rq->queuelist, &ctx->rq_list);
1609         else
1610                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1611 }
1612
1613 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1614                              bool at_head)
1615 {
1616         struct blk_mq_ctx *ctx = rq->mq_ctx;
1617
1618         lockdep_assert_held(&ctx->lock);
1619
1620         __blk_mq_insert_req_list(hctx, rq, at_head);
1621         blk_mq_hctx_mark_pending(hctx, ctx);
1622 }
1623
1624 /*
1625  * Should only be used carefully, when the caller knows we want to
1626  * bypass a potential IO scheduler on the target device.
1627  */
1628 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1629 {
1630         struct blk_mq_ctx *ctx = rq->mq_ctx;
1631         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1632
1633         spin_lock(&hctx->lock);
1634         list_add_tail(&rq->queuelist, &hctx->dispatch);
1635         spin_unlock(&hctx->lock);
1636
1637         if (run_queue)
1638                 blk_mq_run_hw_queue(hctx, false);
1639 }
1640
1641 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1642                             struct list_head *list)
1643
1644 {
1645         /*
1646          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1647          * offline now
1648          */
1649         spin_lock(&ctx->lock);
1650         while (!list_empty(list)) {
1651                 struct request *rq;
1652
1653                 rq = list_first_entry(list, struct request, queuelist);
1654                 BUG_ON(rq->mq_ctx != ctx);
1655                 list_del_init(&rq->queuelist);
1656                 __blk_mq_insert_req_list(hctx, rq, false);
1657         }
1658         blk_mq_hctx_mark_pending(hctx, ctx);
1659         spin_unlock(&ctx->lock);
1660 }
1661
1662 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1663 {
1664         struct request *rqa = container_of(a, struct request, queuelist);
1665         struct request *rqb = container_of(b, struct request, queuelist);
1666
1667         return !(rqa->mq_ctx < rqb->mq_ctx ||
1668                  (rqa->mq_ctx == rqb->mq_ctx &&
1669                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1670 }
1671
1672 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1673 {
1674         struct blk_mq_ctx *this_ctx;
1675         struct request_queue *this_q;
1676         struct request *rq;
1677         LIST_HEAD(list);
1678         LIST_HEAD(ctx_list);
1679         unsigned int depth;
1680
1681         list_splice_init(&plug->mq_list, &list);
1682
1683         list_sort(NULL, &list, plug_ctx_cmp);
1684
1685         this_q = NULL;
1686         this_ctx = NULL;
1687         depth = 0;
1688
1689         while (!list_empty(&list)) {
1690                 rq = list_entry_rq(list.next);
1691                 list_del_init(&rq->queuelist);
1692                 BUG_ON(!rq->q);
1693                 if (rq->mq_ctx != this_ctx) {
1694                         if (this_ctx) {
1695                                 trace_block_unplug(this_q, depth, from_schedule);
1696                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1697                                                                 &ctx_list,
1698                                                                 from_schedule);
1699                         }
1700
1701                         this_ctx = rq->mq_ctx;
1702                         this_q = rq->q;
1703                         depth = 0;
1704                 }
1705
1706                 depth++;
1707                 list_add_tail(&rq->queuelist, &ctx_list);
1708         }
1709
1710         /*
1711          * If 'this_ctx' is set, we know we have entries to complete
1712          * on 'ctx_list'. Do those.
1713          */
1714         if (this_ctx) {
1715                 trace_block_unplug(this_q, depth, from_schedule);
1716                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1717                                                 from_schedule);
1718         }
1719 }
1720
1721 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1722 {
1723         blk_init_request_from_bio(rq, bio);
1724
1725         blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1726
1727         blk_account_io_start(rq, true);
1728 }
1729
1730 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1731                                    struct blk_mq_ctx *ctx,
1732                                    struct request *rq)
1733 {
1734         spin_lock(&ctx->lock);
1735         __blk_mq_insert_request(hctx, rq, false);
1736         spin_unlock(&ctx->lock);
1737 }
1738
1739 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1740 {
1741         if (rq->tag != -1)
1742                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1743
1744         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1745 }
1746
1747 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1748                                             struct request *rq,
1749                                             blk_qc_t *cookie)
1750 {
1751         struct request_queue *q = rq->q;
1752         struct blk_mq_queue_data bd = {
1753                 .rq = rq,
1754                 .last = true,
1755         };
1756         blk_qc_t new_cookie;
1757         blk_status_t ret;
1758
1759         new_cookie = request_to_qc_t(hctx, rq);
1760
1761         /*
1762          * For OK queue, we are done. For error, caller may kill it.
1763          * Any other error (busy), just add it to our list as we
1764          * previously would have done.
1765          */
1766         ret = q->mq_ops->queue_rq(hctx, &bd);
1767         switch (ret) {
1768         case BLK_STS_OK:
1769                 *cookie = new_cookie;
1770                 break;
1771         case BLK_STS_RESOURCE:
1772         case BLK_STS_DEV_RESOURCE:
1773                 __blk_mq_requeue_request(rq);
1774                 break;
1775         default:
1776                 *cookie = BLK_QC_T_NONE;
1777                 break;
1778         }
1779
1780         return ret;
1781 }
1782
1783 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1784                                                 struct request *rq,
1785                                                 blk_qc_t *cookie,
1786                                                 bool bypass_insert)
1787 {
1788         struct request_queue *q = rq->q;
1789         bool run_queue = true;
1790
1791         /*
1792          * RCU or SRCU read lock is needed before checking quiesced flag.
1793          *
1794          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1795          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1796          * and avoid driver to try to dispatch again.
1797          */
1798         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1799                 run_queue = false;
1800                 bypass_insert = false;
1801                 goto insert;
1802         }
1803
1804         if (q->elevator && !bypass_insert)
1805                 goto insert;
1806
1807         if (!blk_mq_get_driver_tag(rq, NULL, false))
1808                 goto insert;
1809
1810         if (!blk_mq_get_dispatch_budget(hctx)) {
1811                 blk_mq_put_driver_tag(rq);
1812                 goto insert;
1813         }
1814
1815         return __blk_mq_issue_directly(hctx, rq, cookie);
1816 insert:
1817         if (bypass_insert)
1818                 return BLK_STS_RESOURCE;
1819
1820         blk_mq_sched_insert_request(rq, false, run_queue, false);
1821         return BLK_STS_OK;
1822 }
1823
1824 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1825                 struct request *rq, blk_qc_t *cookie)
1826 {
1827         blk_status_t ret;
1828         int srcu_idx;
1829
1830         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1831
1832         hctx_lock(hctx, &srcu_idx);
1833
1834         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1835         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1836                 blk_mq_sched_insert_request(rq, false, true, false);
1837         else if (ret != BLK_STS_OK)
1838                 blk_mq_end_request(rq, ret);
1839
1840         hctx_unlock(hctx, srcu_idx);
1841 }
1842
1843 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1844 {
1845         blk_status_t ret;
1846         int srcu_idx;
1847         blk_qc_t unused_cookie;
1848         struct blk_mq_ctx *ctx = rq->mq_ctx;
1849         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1850
1851         hctx_lock(hctx, &srcu_idx);
1852         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1853         hctx_unlock(hctx, srcu_idx);
1854
1855         return ret;
1856 }
1857
1858 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1859 {
1860         const int is_sync = op_is_sync(bio->bi_opf);
1861         const int is_flush_fua = op_is_flush(bio->bi_opf);
1862         struct blk_mq_alloc_data data = { .flags = 0 };
1863         struct request *rq;
1864         unsigned int request_count = 0;
1865         struct blk_plug *plug;
1866         struct request *same_queue_rq = NULL;
1867         blk_qc_t cookie;
1868         unsigned int wb_acct;
1869
1870         blk_queue_bounce(q, &bio);
1871
1872         blk_queue_split(q, &bio);
1873
1874         if (!bio_integrity_prep(bio))
1875                 return BLK_QC_T_NONE;
1876
1877         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1878             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1879                 return BLK_QC_T_NONE;
1880
1881         if (blk_mq_sched_bio_merge(q, bio))
1882                 return BLK_QC_T_NONE;
1883
1884         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1885
1886         trace_block_getrq(q, bio, bio->bi_opf);
1887
1888         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1889         if (unlikely(!rq)) {
1890                 __wbt_done(q->rq_wb, wb_acct);
1891                 if (bio->bi_opf & REQ_NOWAIT)
1892                         bio_wouldblock_error(bio);
1893                 return BLK_QC_T_NONE;
1894         }
1895
1896         wbt_track(&rq->issue_stat, wb_acct);
1897
1898         cookie = request_to_qc_t(data.hctx, rq);
1899
1900         plug = current->plug;
1901         if (unlikely(is_flush_fua)) {
1902                 blk_mq_put_ctx(data.ctx);
1903                 blk_mq_bio_to_request(rq, bio);
1904
1905                 /* bypass scheduler for flush rq */
1906                 blk_insert_flush(rq);
1907                 blk_mq_run_hw_queue(data.hctx, true);
1908         } else if (plug && q->nr_hw_queues == 1) {
1909                 struct request *last = NULL;
1910
1911                 blk_mq_put_ctx(data.ctx);
1912                 blk_mq_bio_to_request(rq, bio);
1913
1914                 /*
1915                  * @request_count may become stale because of schedule
1916                  * out, so check the list again.
1917                  */
1918                 if (list_empty(&plug->mq_list))
1919                         request_count = 0;
1920                 else if (blk_queue_nomerges(q))
1921                         request_count = blk_plug_queued_count(q);
1922
1923                 if (!request_count)
1924                         trace_block_plug(q);
1925                 else
1926                         last = list_entry_rq(plug->mq_list.prev);
1927
1928                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1929                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1930                         blk_flush_plug_list(plug, false);
1931                         trace_block_plug(q);
1932                 }
1933
1934                 list_add_tail(&rq->queuelist, &plug->mq_list);
1935         } else if (plug && !blk_queue_nomerges(q)) {
1936                 blk_mq_bio_to_request(rq, bio);
1937
1938                 /*
1939                  * We do limited plugging. If the bio can be merged, do that.
1940                  * Otherwise the existing request in the plug list will be
1941                  * issued. So the plug list will have one request at most
1942                  * The plug list might get flushed before this. If that happens,
1943                  * the plug list is empty, and same_queue_rq is invalid.
1944                  */
1945                 if (list_empty(&plug->mq_list))
1946                         same_queue_rq = NULL;
1947                 if (same_queue_rq)
1948                         list_del_init(&same_queue_rq->queuelist);
1949                 list_add_tail(&rq->queuelist, &plug->mq_list);
1950
1951                 blk_mq_put_ctx(data.ctx);
1952
1953                 if (same_queue_rq) {
1954                         data.hctx = blk_mq_map_queue(q,
1955                                         same_queue_rq->mq_ctx->cpu);
1956                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1957                                         &cookie);
1958                 }
1959         } else if (q->nr_hw_queues > 1 && is_sync) {
1960                 blk_mq_put_ctx(data.ctx);
1961                 blk_mq_bio_to_request(rq, bio);
1962                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1963         } else if (q->elevator) {
1964                 blk_mq_put_ctx(data.ctx);
1965                 blk_mq_bio_to_request(rq, bio);
1966                 blk_mq_sched_insert_request(rq, false, true, true);
1967         } else {
1968                 blk_mq_put_ctx(data.ctx);
1969                 blk_mq_bio_to_request(rq, bio);
1970                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1971                 blk_mq_run_hw_queue(data.hctx, true);
1972         }
1973
1974         return cookie;
1975 }
1976
1977 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1978                      unsigned int hctx_idx)
1979 {
1980         struct page *page;
1981
1982         if (tags->rqs && set->ops->exit_request) {
1983                 int i;
1984
1985                 for (i = 0; i < tags->nr_tags; i++) {
1986                         struct request *rq = tags->static_rqs[i];
1987
1988                         if (!rq)
1989                                 continue;
1990                         set->ops->exit_request(set, rq, hctx_idx);
1991                         tags->static_rqs[i] = NULL;
1992                 }
1993         }
1994
1995         while (!list_empty(&tags->page_list)) {
1996                 page = list_first_entry(&tags->page_list, struct page, lru);
1997                 list_del_init(&page->lru);
1998                 /*
1999                  * Remove kmemleak object previously allocated in
2000                  * blk_mq_init_rq_map().
2001                  */
2002                 kmemleak_free(page_address(page));
2003                 __free_pages(page, page->private);
2004         }
2005 }
2006
2007 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2008 {
2009         kfree(tags->rqs);
2010         tags->rqs = NULL;
2011         kfree(tags->static_rqs);
2012         tags->static_rqs = NULL;
2013
2014         blk_mq_free_tags(tags);
2015 }
2016
2017 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2018                                         unsigned int hctx_idx,
2019                                         unsigned int nr_tags,
2020                                         unsigned int reserved_tags)
2021 {
2022         struct blk_mq_tags *tags;
2023         int node;
2024
2025         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2026         if (node == NUMA_NO_NODE)
2027                 node = set->numa_node;
2028
2029         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2030                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2031         if (!tags)
2032                 return NULL;
2033
2034         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2035                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2036                                  node);
2037         if (!tags->rqs) {
2038                 blk_mq_free_tags(tags);
2039                 return NULL;
2040         }
2041
2042         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
2043                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2044                                  node);
2045         if (!tags->static_rqs) {
2046                 kfree(tags->rqs);
2047                 blk_mq_free_tags(tags);
2048                 return NULL;
2049         }
2050
2051         return tags;
2052 }
2053
2054 static size_t order_to_size(unsigned int order)
2055 {
2056         return (size_t)PAGE_SIZE << order;
2057 }
2058
2059 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2060                                unsigned int hctx_idx, int node)
2061 {
2062         int ret;
2063
2064         if (set->ops->init_request) {
2065                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2066                 if (ret)
2067                         return ret;
2068         }
2069
2070         seqcount_init(&rq->gstate_seq);
2071         u64_stats_init(&rq->aborted_gstate_sync);
2072         return 0;
2073 }
2074
2075 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2076                      unsigned int hctx_idx, unsigned int depth)
2077 {
2078         unsigned int i, j, entries_per_page, max_order = 4;
2079         size_t rq_size, left;
2080         int node;
2081
2082         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
2083         if (node == NUMA_NO_NODE)
2084                 node = set->numa_node;
2085
2086         INIT_LIST_HEAD(&tags->page_list);
2087
2088         /*
2089          * rq_size is the size of the request plus driver payload, rounded
2090          * to the cacheline size
2091          */
2092         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2093                                 cache_line_size());
2094         left = rq_size * depth;
2095
2096         for (i = 0; i < depth; ) {
2097                 int this_order = max_order;
2098                 struct page *page;
2099                 int to_do;
2100                 void *p;
2101
2102                 while (this_order && left < order_to_size(this_order - 1))
2103                         this_order--;
2104
2105                 do {
2106                         page = alloc_pages_node(node,
2107                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2108                                 this_order);
2109                         if (page)
2110                                 break;
2111                         if (!this_order--)
2112                                 break;
2113                         if (order_to_size(this_order) < rq_size)
2114                                 break;
2115                 } while (1);
2116
2117                 if (!page)
2118                         goto fail;
2119
2120                 page->private = this_order;
2121                 list_add_tail(&page->lru, &tags->page_list);
2122
2123                 p = page_address(page);
2124                 /*
2125                  * Allow kmemleak to scan these pages as they contain pointers
2126                  * to additional allocations like via ops->init_request().
2127                  */
2128                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2129                 entries_per_page = order_to_size(this_order) / rq_size;
2130                 to_do = min(entries_per_page, depth - i);
2131                 left -= to_do * rq_size;
2132                 for (j = 0; j < to_do; j++) {
2133                         struct request *rq = p;
2134
2135                         tags->static_rqs[i] = rq;
2136                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2137                                 tags->static_rqs[i] = NULL;
2138                                 goto fail;
2139                         }
2140
2141                         p += rq_size;
2142                         i++;
2143                 }
2144         }
2145         return 0;
2146
2147 fail:
2148         blk_mq_free_rqs(set, tags, hctx_idx);
2149         return -ENOMEM;
2150 }
2151
2152 /*
2153  * 'cpu' is going away. splice any existing rq_list entries from this
2154  * software queue to the hw queue dispatch list, and ensure that it
2155  * gets run.
2156  */
2157 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2158 {
2159         struct blk_mq_hw_ctx *hctx;
2160         struct blk_mq_ctx *ctx;
2161         LIST_HEAD(tmp);
2162
2163         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2164         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2165
2166         spin_lock(&ctx->lock);
2167         if (!list_empty(&ctx->rq_list)) {
2168                 list_splice_init(&ctx->rq_list, &tmp);
2169                 blk_mq_hctx_clear_pending(hctx, ctx);
2170         }
2171         spin_unlock(&ctx->lock);
2172
2173         if (list_empty(&tmp))
2174                 return 0;
2175
2176         spin_lock(&hctx->lock);
2177         list_splice_tail_init(&tmp, &hctx->dispatch);
2178         spin_unlock(&hctx->lock);
2179
2180         blk_mq_run_hw_queue(hctx, true);
2181         return 0;
2182 }
2183
2184 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2185 {
2186         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2187                                             &hctx->cpuhp_dead);
2188 }
2189
2190 /* hctx->ctxs will be freed in queue's release handler */
2191 static void blk_mq_exit_hctx(struct request_queue *q,
2192                 struct blk_mq_tag_set *set,
2193                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2194 {
2195         blk_mq_debugfs_unregister_hctx(hctx);
2196
2197         if (blk_mq_hw_queue_mapped(hctx))
2198                 blk_mq_tag_idle(hctx);
2199
2200         if (set->ops->exit_request)
2201                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2202
2203         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2204
2205         if (set->ops->exit_hctx)
2206                 set->ops->exit_hctx(hctx, hctx_idx);
2207
2208         if (hctx->flags & BLK_MQ_F_BLOCKING)
2209                 cleanup_srcu_struct(hctx->srcu);
2210
2211         blk_mq_remove_cpuhp(hctx);
2212         blk_free_flush_queue(hctx->fq);
2213         sbitmap_free(&hctx->ctx_map);
2214 }
2215
2216 static void blk_mq_exit_hw_queues(struct request_queue *q,
2217                 struct blk_mq_tag_set *set, int nr_queue)
2218 {
2219         struct blk_mq_hw_ctx *hctx;
2220         unsigned int i;
2221
2222         queue_for_each_hw_ctx(q, hctx, i) {
2223                 if (i == nr_queue)
2224                         break;
2225                 blk_mq_exit_hctx(q, set, hctx, i);
2226         }
2227 }
2228
2229 static int blk_mq_init_hctx(struct request_queue *q,
2230                 struct blk_mq_tag_set *set,
2231                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2232 {
2233         int node;
2234
2235         node = hctx->numa_node;
2236         if (node == NUMA_NO_NODE)
2237                 node = hctx->numa_node = set->numa_node;
2238
2239         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2240         spin_lock_init(&hctx->lock);
2241         INIT_LIST_HEAD(&hctx->dispatch);
2242         hctx->queue = q;
2243         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2244
2245         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2246
2247         hctx->tags = set->tags[hctx_idx];
2248
2249         /*
2250          * Allocate space for all possible cpus to avoid allocation at
2251          * runtime
2252          */
2253         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2254                                         GFP_KERNEL, node);
2255         if (!hctx->ctxs)
2256                 goto unregister_cpu_notifier;
2257
2258         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2259                               node))
2260                 goto free_ctxs;
2261
2262         hctx->nr_ctx = 0;
2263
2264         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2265         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2266
2267         if (set->ops->init_hctx &&
2268             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2269                 goto free_bitmap;
2270
2271         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2272                 goto exit_hctx;
2273
2274         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2275         if (!hctx->fq)
2276                 goto sched_exit_hctx;
2277
2278         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2279                 goto free_fq;
2280
2281         if (hctx->flags & BLK_MQ_F_BLOCKING)
2282                 init_srcu_struct(hctx->srcu);
2283
2284         blk_mq_debugfs_register_hctx(q, hctx);
2285
2286         return 0;
2287
2288  free_fq:
2289         kfree(hctx->fq);
2290  sched_exit_hctx:
2291         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2292  exit_hctx:
2293         if (set->ops->exit_hctx)
2294                 set->ops->exit_hctx(hctx, hctx_idx);
2295  free_bitmap:
2296         sbitmap_free(&hctx->ctx_map);
2297  free_ctxs:
2298         kfree(hctx->ctxs);
2299  unregister_cpu_notifier:
2300         blk_mq_remove_cpuhp(hctx);
2301         return -1;
2302 }
2303
2304 static void blk_mq_init_cpu_queues(struct request_queue *q,
2305                                    unsigned int nr_hw_queues)
2306 {
2307         unsigned int i;
2308
2309         for_each_possible_cpu(i) {
2310                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2311                 struct blk_mq_hw_ctx *hctx;
2312
2313                 __ctx->cpu = i;
2314                 spin_lock_init(&__ctx->lock);
2315                 INIT_LIST_HEAD(&__ctx->rq_list);
2316                 __ctx->queue = q;
2317
2318                 /*
2319                  * Set local node, IFF we have more than one hw queue. If
2320                  * not, we remain on the home node of the device
2321                  */
2322                 hctx = blk_mq_map_queue(q, i);
2323                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2324                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2325         }
2326 }
2327
2328 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2329 {
2330         int ret = 0;
2331
2332         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2333                                         set->queue_depth, set->reserved_tags);
2334         if (!set->tags[hctx_idx])
2335                 return false;
2336
2337         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2338                                 set->queue_depth);
2339         if (!ret)
2340                 return true;
2341
2342         blk_mq_free_rq_map(set->tags[hctx_idx]);
2343         set->tags[hctx_idx] = NULL;
2344         return false;
2345 }
2346
2347 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2348                                          unsigned int hctx_idx)
2349 {
2350         if (set->tags[hctx_idx]) {
2351                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2352                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2353                 set->tags[hctx_idx] = NULL;
2354         }
2355 }
2356
2357 static void blk_mq_map_swqueue(struct request_queue *q)
2358 {
2359         unsigned int i, hctx_idx;
2360         struct blk_mq_hw_ctx *hctx;
2361         struct blk_mq_ctx *ctx;
2362         struct blk_mq_tag_set *set = q->tag_set;
2363
2364         /*
2365          * Avoid others reading imcomplete hctx->cpumask through sysfs
2366          */
2367         mutex_lock(&q->sysfs_lock);
2368
2369         queue_for_each_hw_ctx(q, hctx, i) {
2370                 cpumask_clear(hctx->cpumask);
2371                 hctx->nr_ctx = 0;
2372         }
2373
2374         /*
2375          * Map software to hardware queues.
2376          *
2377          * If the cpu isn't present, the cpu is mapped to first hctx.
2378          */
2379         for_each_possible_cpu(i) {
2380                 hctx_idx = q->mq_map[i];
2381                 /* unmapped hw queue can be remapped after CPU topo changed */
2382                 if (!set->tags[hctx_idx] &&
2383                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2384                         /*
2385                          * If tags initialization fail for some hctx,
2386                          * that hctx won't be brought online.  In this
2387                          * case, remap the current ctx to hctx[0] which
2388                          * is guaranteed to always have tags allocated
2389                          */
2390                         q->mq_map[i] = 0;
2391                 }
2392
2393                 ctx = per_cpu_ptr(q->queue_ctx, i);
2394                 hctx = blk_mq_map_queue(q, i);
2395
2396                 cpumask_set_cpu(i, hctx->cpumask);
2397                 ctx->index_hw = hctx->nr_ctx;
2398                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2399         }
2400
2401         mutex_unlock(&q->sysfs_lock);
2402
2403         queue_for_each_hw_ctx(q, hctx, i) {
2404                 /*
2405                  * If no software queues are mapped to this hardware queue,
2406                  * disable it and free the request entries.
2407                  */
2408                 if (!hctx->nr_ctx) {
2409                         /* Never unmap queue 0.  We need it as a
2410                          * fallback in case of a new remap fails
2411                          * allocation
2412                          */
2413                         if (i && set->tags[i])
2414                                 blk_mq_free_map_and_requests(set, i);
2415
2416                         hctx->tags = NULL;
2417                         continue;
2418                 }
2419
2420                 hctx->tags = set->tags[i];
2421                 WARN_ON(!hctx->tags);
2422
2423                 /*
2424                  * Set the map size to the number of mapped software queues.
2425                  * This is more accurate and more efficient than looping
2426                  * over all possibly mapped software queues.
2427                  */
2428                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2429
2430                 /*
2431                  * Initialize batch roundrobin counts
2432                  */
2433                 hctx->next_cpu = cpumask_first_and(hctx->cpumask,
2434                                 cpu_online_mask);
2435                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2436         }
2437 }
2438
2439 /*
2440  * Caller needs to ensure that we're either frozen/quiesced, or that
2441  * the queue isn't live yet.
2442  */
2443 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2444 {
2445         struct blk_mq_hw_ctx *hctx;
2446         int i;
2447
2448         queue_for_each_hw_ctx(q, hctx, i) {
2449                 if (shared) {
2450                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2451                                 atomic_inc(&q->shared_hctx_restart);
2452                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2453                 } else {
2454                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2455                                 atomic_dec(&q->shared_hctx_restart);
2456                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2457                 }
2458         }
2459 }
2460
2461 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2462                                         bool shared)
2463 {
2464         struct request_queue *q;
2465
2466         lockdep_assert_held(&set->tag_list_lock);
2467
2468         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2469                 blk_mq_freeze_queue(q);
2470                 queue_set_hctx_shared(q, shared);
2471                 blk_mq_unfreeze_queue(q);
2472         }
2473 }
2474
2475 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2476 {
2477         struct blk_mq_tag_set *set = q->tag_set;
2478
2479         mutex_lock(&set->tag_list_lock);
2480         list_del_rcu(&q->tag_set_list);
2481         INIT_LIST_HEAD(&q->tag_set_list);
2482         if (list_is_singular(&set->tag_list)) {
2483                 /* just transitioned to unshared */
2484                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2485                 /* update existing queue */
2486                 blk_mq_update_tag_set_depth(set, false);
2487         }
2488         mutex_unlock(&set->tag_list_lock);
2489
2490         synchronize_rcu();
2491 }
2492
2493 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2494                                      struct request_queue *q)
2495 {
2496         q->tag_set = set;
2497
2498         mutex_lock(&set->tag_list_lock);
2499
2500         /*
2501          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2502          */
2503         if (!list_empty(&set->tag_list) &&
2504             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2505                 set->flags |= BLK_MQ_F_TAG_SHARED;
2506                 /* update existing queue */
2507                 blk_mq_update_tag_set_depth(set, true);
2508         }
2509         if (set->flags & BLK_MQ_F_TAG_SHARED)
2510                 queue_set_hctx_shared(q, true);
2511         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2512
2513         mutex_unlock(&set->tag_list_lock);
2514 }
2515
2516 /*
2517  * It is the actual release handler for mq, but we do it from
2518  * request queue's release handler for avoiding use-after-free
2519  * and headache because q->mq_kobj shouldn't have been introduced,
2520  * but we can't group ctx/kctx kobj without it.
2521  */
2522 void blk_mq_release(struct request_queue *q)
2523 {
2524         struct blk_mq_hw_ctx *hctx;
2525         unsigned int i;
2526
2527         /* hctx kobj stays in hctx */
2528         queue_for_each_hw_ctx(q, hctx, i) {
2529                 if (!hctx)
2530                         continue;
2531                 kobject_put(&hctx->kobj);
2532         }
2533
2534         q->mq_map = NULL;
2535
2536         kfree(q->queue_hw_ctx);
2537
2538         /*
2539          * release .mq_kobj and sw queue's kobject now because
2540          * both share lifetime with request queue.
2541          */
2542         blk_mq_sysfs_deinit(q);
2543
2544         free_percpu(q->queue_ctx);
2545 }
2546
2547 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2548 {
2549         struct request_queue *uninit_q, *q;
2550
2551         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2552         if (!uninit_q)
2553                 return ERR_PTR(-ENOMEM);
2554
2555         q = blk_mq_init_allocated_queue(set, uninit_q);
2556         if (IS_ERR(q))
2557                 blk_cleanup_queue(uninit_q);
2558
2559         return q;
2560 }
2561 EXPORT_SYMBOL(blk_mq_init_queue);
2562
2563 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2564 {
2565         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2566
2567         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2568                            __alignof__(struct blk_mq_hw_ctx)) !=
2569                      sizeof(struct blk_mq_hw_ctx));
2570
2571         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2572                 hw_ctx_size += sizeof(struct srcu_struct);
2573
2574         return hw_ctx_size;
2575 }
2576
2577 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2578                                                 struct request_queue *q)
2579 {
2580         int i, j;
2581         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2582
2583         blk_mq_sysfs_unregister(q);
2584
2585         /* protect against switching io scheduler  */
2586         mutex_lock(&q->sysfs_lock);
2587         for (i = 0; i < set->nr_hw_queues; i++) {
2588                 int node;
2589
2590                 if (hctxs[i])
2591                         continue;
2592
2593                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2594                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2595                                         GFP_KERNEL, node);
2596                 if (!hctxs[i])
2597                         break;
2598
2599                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2600                                                 node)) {
2601                         kfree(hctxs[i]);
2602                         hctxs[i] = NULL;
2603                         break;
2604                 }
2605
2606                 atomic_set(&hctxs[i]->nr_active, 0);
2607                 hctxs[i]->numa_node = node;
2608                 hctxs[i]->queue_num = i;
2609
2610                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2611                         free_cpumask_var(hctxs[i]->cpumask);
2612                         kfree(hctxs[i]);
2613                         hctxs[i] = NULL;
2614                         break;
2615                 }
2616                 blk_mq_hctx_kobj_init(hctxs[i]);
2617         }
2618         for (j = i; j < q->nr_hw_queues; j++) {
2619                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2620
2621                 if (hctx) {
2622                         if (hctx->tags)
2623                                 blk_mq_free_map_and_requests(set, j);
2624                         blk_mq_exit_hctx(q, set, hctx, j);
2625                         kobject_put(&hctx->kobj);
2626                         hctxs[j] = NULL;
2627
2628                 }
2629         }
2630         q->nr_hw_queues = i;
2631         mutex_unlock(&q->sysfs_lock);
2632         blk_mq_sysfs_register(q);
2633 }
2634
2635 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2636                                                   struct request_queue *q)
2637 {
2638         /* mark the queue as mq asap */
2639         q->mq_ops = set->ops;
2640
2641         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2642                                              blk_mq_poll_stats_bkt,
2643                                              BLK_MQ_POLL_STATS_BKTS, q);
2644         if (!q->poll_cb)
2645                 goto err_exit;
2646
2647         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2648         if (!q->queue_ctx)
2649                 goto err_exit;
2650
2651         /* init q->mq_kobj and sw queues' kobjects */
2652         blk_mq_sysfs_init(q);
2653
2654         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2655                                                 GFP_KERNEL, set->numa_node);
2656         if (!q->queue_hw_ctx)
2657                 goto err_percpu;
2658
2659         q->mq_map = set->mq_map;
2660
2661         blk_mq_realloc_hw_ctxs(set, q);
2662         if (!q->nr_hw_queues)
2663                 goto err_hctxs;
2664
2665         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2666         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2667
2668         q->nr_queues = nr_cpu_ids;
2669
2670         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2671
2672         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2673                 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2674
2675         q->sg_reserved_size = INT_MAX;
2676
2677         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2678         INIT_LIST_HEAD(&q->requeue_list);
2679         spin_lock_init(&q->requeue_lock);
2680
2681         blk_queue_make_request(q, blk_mq_make_request);
2682         if (q->mq_ops->poll)
2683                 q->poll_fn = blk_mq_poll;
2684
2685         /*
2686          * Do this after blk_queue_make_request() overrides it...
2687          */
2688         q->nr_requests = set->queue_depth;
2689
2690         /*
2691          * Default to classic polling
2692          */
2693         q->poll_nsec = -1;
2694
2695         if (set->ops->complete)
2696                 blk_queue_softirq_done(q, set->ops->complete);
2697
2698         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2699         blk_mq_add_queue_tag_set(set, q);
2700         blk_mq_map_swqueue(q);
2701
2702         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2703                 int ret;
2704
2705                 ret = blk_mq_sched_init(q);
2706                 if (ret)
2707                         return ERR_PTR(ret);
2708         }
2709
2710         return q;
2711
2712 err_hctxs:
2713         kfree(q->queue_hw_ctx);
2714 err_percpu:
2715         free_percpu(q->queue_ctx);
2716 err_exit:
2717         q->mq_ops = NULL;
2718         return ERR_PTR(-ENOMEM);
2719 }
2720 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2721
2722 void blk_mq_free_queue(struct request_queue *q)
2723 {
2724         struct blk_mq_tag_set   *set = q->tag_set;
2725
2726         blk_mq_del_queue_tag_set(q);
2727         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2728 }
2729
2730 /* Basically redo blk_mq_init_queue with queue frozen */
2731 static void blk_mq_queue_reinit(struct request_queue *q)
2732 {
2733         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2734
2735         blk_mq_debugfs_unregister_hctxs(q);
2736         blk_mq_sysfs_unregister(q);
2737
2738         /*
2739          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2740          * we should change hctx numa_node according to the new topology (this
2741          * involves freeing and re-allocating memory, worth doing?)
2742          */
2743         blk_mq_map_swqueue(q);
2744
2745         blk_mq_sysfs_register(q);
2746         blk_mq_debugfs_register_hctxs(q);
2747 }
2748
2749 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2750 {
2751         int i;
2752
2753         for (i = 0; i < set->nr_hw_queues; i++)
2754                 if (!__blk_mq_alloc_rq_map(set, i))
2755                         goto out_unwind;
2756
2757         return 0;
2758
2759 out_unwind:
2760         while (--i >= 0)
2761                 blk_mq_free_rq_map(set->tags[i]);
2762
2763         return -ENOMEM;
2764 }
2765
2766 /*
2767  * Allocate the request maps associated with this tag_set. Note that this
2768  * may reduce the depth asked for, if memory is tight. set->queue_depth
2769  * will be updated to reflect the allocated depth.
2770  */
2771 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2772 {
2773         unsigned int depth;
2774         int err;
2775
2776         depth = set->queue_depth;
2777         do {
2778                 err = __blk_mq_alloc_rq_maps(set);
2779                 if (!err)
2780                         break;
2781
2782                 set->queue_depth >>= 1;
2783                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2784                         err = -ENOMEM;
2785                         break;
2786                 }
2787         } while (set->queue_depth);
2788
2789         if (!set->queue_depth || err) {
2790                 pr_err("blk-mq: failed to allocate request map\n");
2791                 return -ENOMEM;
2792         }
2793
2794         if (depth != set->queue_depth)
2795                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2796                                                 depth, set->queue_depth);
2797
2798         return 0;
2799 }
2800
2801 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2802 {
2803         if (set->ops->map_queues) {
2804                 int cpu;
2805                 /*
2806                  * transport .map_queues is usually done in the following
2807                  * way:
2808                  *
2809                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2810                  *      mask = get_cpu_mask(queue)
2811                  *      for_each_cpu(cpu, mask)
2812                  *              set->mq_map[cpu] = queue;
2813                  * }
2814                  *
2815                  * When we need to remap, the table has to be cleared for
2816                  * killing stale mapping since one CPU may not be mapped
2817                  * to any hw queue.
2818                  */
2819                 for_each_possible_cpu(cpu)
2820                         set->mq_map[cpu] = 0;
2821
2822                 return set->ops->map_queues(set);
2823         } else
2824                 return blk_mq_map_queues(set);
2825 }
2826
2827 /*
2828  * Alloc a tag set to be associated with one or more request queues.
2829  * May fail with EINVAL for various error conditions. May adjust the
2830  * requested depth down, if if it too large. In that case, the set
2831  * value will be stored in set->queue_depth.
2832  */
2833 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2834 {
2835         int ret;
2836
2837         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2838
2839         if (!set->nr_hw_queues)
2840                 return -EINVAL;
2841         if (!set->queue_depth)
2842                 return -EINVAL;
2843         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2844                 return -EINVAL;
2845
2846         if (!set->ops->queue_rq)
2847                 return -EINVAL;
2848
2849         if (!set->ops->get_budget ^ !set->ops->put_budget)
2850                 return -EINVAL;
2851
2852         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2853                 pr_info("blk-mq: reduced tag depth to %u\n",
2854                         BLK_MQ_MAX_DEPTH);
2855                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2856         }
2857
2858         /*
2859          * If a crashdump is active, then we are potentially in a very
2860          * memory constrained environment. Limit us to 1 queue and
2861          * 64 tags to prevent using too much memory.
2862          */
2863         if (is_kdump_kernel()) {
2864                 set->nr_hw_queues = 1;
2865                 set->queue_depth = min(64U, set->queue_depth);
2866         }
2867         /*
2868          * There is no use for more h/w queues than cpus.
2869          */
2870         if (set->nr_hw_queues > nr_cpu_ids)
2871                 set->nr_hw_queues = nr_cpu_ids;
2872
2873         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2874                                  GFP_KERNEL, set->numa_node);
2875         if (!set->tags)
2876                 return -ENOMEM;
2877
2878         ret = -ENOMEM;
2879         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2880                         GFP_KERNEL, set->numa_node);
2881         if (!set->mq_map)
2882                 goto out_free_tags;
2883
2884         ret = blk_mq_update_queue_map(set);
2885         if (ret)
2886                 goto out_free_mq_map;
2887
2888         ret = blk_mq_alloc_rq_maps(set);
2889         if (ret)
2890                 goto out_free_mq_map;
2891
2892         mutex_init(&set->tag_list_lock);
2893         INIT_LIST_HEAD(&set->tag_list);
2894
2895         return 0;
2896
2897 out_free_mq_map:
2898         kfree(set->mq_map);
2899         set->mq_map = NULL;
2900 out_free_tags:
2901         kfree(set->tags);
2902         set->tags = NULL;
2903         return ret;
2904 }
2905 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2906
2907 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2908 {
2909         int i;
2910
2911         for (i = 0; i < nr_cpu_ids; i++)
2912                 blk_mq_free_map_and_requests(set, i);
2913
2914         kfree(set->mq_map);
2915         set->mq_map = NULL;
2916
2917         kfree(set->tags);
2918         set->tags = NULL;
2919 }
2920 EXPORT_SYMBOL(blk_mq_free_tag_set);
2921
2922 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2923 {
2924         struct blk_mq_tag_set *set = q->tag_set;
2925         struct blk_mq_hw_ctx *hctx;
2926         int i, ret;
2927
2928         if (!set)
2929                 return -EINVAL;
2930
2931         blk_mq_freeze_queue(q);
2932         blk_mq_quiesce_queue(q);
2933
2934         ret = 0;
2935         queue_for_each_hw_ctx(q, hctx, i) {
2936                 if (!hctx->tags)
2937                         continue;
2938                 /*
2939                  * If we're using an MQ scheduler, just update the scheduler
2940                  * queue depth. This is similar to what the old code would do.
2941                  */
2942                 if (!hctx->sched_tags) {
2943                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2944                                                         false);
2945                 } else {
2946                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2947                                                         nr, true);
2948                 }
2949                 if (ret)
2950                         break;
2951         }
2952
2953         if (!ret)
2954                 q->nr_requests = nr;
2955
2956         blk_mq_unquiesce_queue(q);
2957         blk_mq_unfreeze_queue(q);
2958
2959         return ret;
2960 }
2961
2962 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2963                                                         int nr_hw_queues)
2964 {
2965         struct request_queue *q;
2966
2967         lockdep_assert_held(&set->tag_list_lock);
2968
2969         if (nr_hw_queues > nr_cpu_ids)
2970                 nr_hw_queues = nr_cpu_ids;
2971         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2972                 return;
2973
2974         list_for_each_entry(q, &set->tag_list, tag_set_list)
2975                 blk_mq_freeze_queue(q);
2976
2977         set->nr_hw_queues = nr_hw_queues;
2978         blk_mq_update_queue_map(set);
2979         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2980                 blk_mq_realloc_hw_ctxs(set, q);
2981                 blk_mq_queue_reinit(q);
2982         }
2983
2984         list_for_each_entry(q, &set->tag_list, tag_set_list)
2985                 blk_mq_unfreeze_queue(q);
2986 }
2987
2988 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2989 {
2990         mutex_lock(&set->tag_list_lock);
2991         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2992         mutex_unlock(&set->tag_list_lock);
2993 }
2994 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2995
2996 /* Enable polling stats and return whether they were already enabled. */
2997 static bool blk_poll_stats_enable(struct request_queue *q)
2998 {
2999         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3000             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3001                 return true;
3002         blk_stat_add_callback(q, q->poll_cb);
3003         return false;
3004 }
3005
3006 static void blk_mq_poll_stats_start(struct request_queue *q)
3007 {
3008         /*
3009          * We don't arm the callback if polling stats are not enabled or the
3010          * callback is already active.
3011          */
3012         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3013             blk_stat_is_active(q->poll_cb))
3014                 return;
3015
3016         blk_stat_activate_msecs(q->poll_cb, 100);
3017 }
3018
3019 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3020 {
3021         struct request_queue *q = cb->data;
3022         int bucket;
3023
3024         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3025                 if (cb->stat[bucket].nr_samples)
3026                         q->poll_stat[bucket] = cb->stat[bucket];
3027         }
3028 }
3029
3030 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3031                                        struct blk_mq_hw_ctx *hctx,
3032                                        struct request *rq)
3033 {
3034         unsigned long ret = 0;
3035         int bucket;
3036
3037         /*
3038          * If stats collection isn't on, don't sleep but turn it on for
3039          * future users
3040          */
3041         if (!blk_poll_stats_enable(q))
3042                 return 0;
3043
3044         /*
3045          * As an optimistic guess, use half of the mean service time
3046          * for this type of request. We can (and should) make this smarter.
3047          * For instance, if the completion latencies are tight, we can
3048          * get closer than just half the mean. This is especially
3049          * important on devices where the completion latencies are longer
3050          * than ~10 usec. We do use the stats for the relevant IO size
3051          * if available which does lead to better estimates.
3052          */
3053         bucket = blk_mq_poll_stats_bkt(rq);
3054         if (bucket < 0)
3055                 return ret;
3056
3057         if (q->poll_stat[bucket].nr_samples)
3058                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3059
3060         return ret;
3061 }
3062
3063 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3064                                      struct blk_mq_hw_ctx *hctx,
3065                                      struct request *rq)
3066 {
3067         struct hrtimer_sleeper hs;
3068         enum hrtimer_mode mode;
3069         unsigned int nsecs;
3070         ktime_t kt;
3071
3072         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3073                 return false;
3074
3075         /*
3076          * poll_nsec can be:
3077          *
3078          * -1:  don't ever hybrid sleep
3079          *  0:  use half of prev avg
3080          * >0:  use this specific value
3081          */
3082         if (q->poll_nsec == -1)
3083                 return false;
3084         else if (q->poll_nsec > 0)
3085                 nsecs = q->poll_nsec;
3086         else
3087                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3088
3089         if (!nsecs)
3090                 return false;
3091
3092         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3093
3094         /*
3095          * This will be replaced with the stats tracking code, using
3096          * 'avg_completion_time / 2' as the pre-sleep target.
3097          */
3098         kt = nsecs;
3099
3100         mode = HRTIMER_MODE_REL;
3101         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3102         hrtimer_set_expires(&hs.timer, kt);
3103
3104         hrtimer_init_sleeper(&hs, current);
3105         do {
3106                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3107                         break;
3108                 set_current_state(TASK_UNINTERRUPTIBLE);
3109                 hrtimer_start_expires(&hs.timer, mode);
3110                 if (hs.task)
3111                         io_schedule();
3112                 hrtimer_cancel(&hs.timer);
3113                 mode = HRTIMER_MODE_ABS;
3114         } while (hs.task && !signal_pending(current));
3115
3116         __set_current_state(TASK_RUNNING);
3117         destroy_hrtimer_on_stack(&hs.timer);
3118         return true;
3119 }
3120
3121 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3122 {
3123         struct request_queue *q = hctx->queue;
3124         long state;
3125
3126         /*
3127          * If we sleep, have the caller restart the poll loop to reset
3128          * the state. Like for the other success return cases, the
3129          * caller is responsible for checking if the IO completed. If
3130          * the IO isn't complete, we'll get called again and will go
3131          * straight to the busy poll loop.
3132          */
3133         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3134                 return true;
3135
3136         hctx->poll_considered++;
3137
3138         state = current->state;
3139         while (!need_resched()) {
3140                 int ret;
3141
3142                 hctx->poll_invoked++;
3143
3144                 ret = q->mq_ops->poll(hctx, rq->tag);
3145                 if (ret > 0) {
3146                         hctx->poll_success++;
3147                         set_current_state(TASK_RUNNING);
3148                         return true;
3149                 }
3150
3151                 if (signal_pending_state(state, current))
3152                         set_current_state(TASK_RUNNING);
3153
3154                 if (current->state == TASK_RUNNING)
3155                         return true;
3156                 if (ret < 0)
3157                         break;
3158                 cpu_relax();
3159         }
3160
3161         __set_current_state(TASK_RUNNING);
3162         return false;
3163 }
3164
3165 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3166 {
3167         struct blk_mq_hw_ctx *hctx;
3168         struct request *rq;
3169
3170         if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3171                 return false;
3172
3173         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3174         if (!blk_qc_t_is_internal(cookie))
3175                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3176         else {
3177                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3178                 /*
3179                  * With scheduling, if the request has completed, we'll
3180                  * get a NULL return here, as we clear the sched tag when
3181                  * that happens. The request still remains valid, like always,
3182                  * so we should be safe with just the NULL check.
3183                  */
3184                 if (!rq)
3185                         return false;
3186         }
3187
3188         return __blk_mq_poll(hctx, rq);
3189 }
3190
3191 static int __init blk_mq_init(void)
3192 {
3193         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3194                                 blk_mq_hctx_notify_dead);
3195         return 0;
3196 }
3197 subsys_initcall(blk_mq_init);