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-tag.h"
  35 #include "blk-stat.h"
  36 #include "blk-wbt.h"
  37 #include "blk-mq-sched.h"
  38
  39 static DEFINE_MUTEX(all_q_mutex);
  40 static LIST_HEAD(all_q_list);
  41
  42 /*
  43  * Check if any of the ctx's have pending work in this hardware queue
  44  */
  45 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  46 {
  47         return sbitmap_any_bit_set(&hctx->ctx_map) ||
  48                         !list_empty_careful(&hctx->dispatch) ||
  49                         blk_mq_sched_has_work(hctx);
  50 }
  51
  52 /*
  53  * Mark this ctx as having pending work in this hardware queue
  54  */
  55 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  56                                      struct blk_mq_ctx *ctx)
  57 {
  58         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  59                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  60 }
  61
  62 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  63                                       struct blk_mq_ctx *ctx)
  64 {
  65         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  66 }
  67
  68 void blk_mq_freeze_queue_start(struct request_queue *q)
  69 {
  70         int freeze_depth;
  71
  72         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
  73         if (freeze_depth == 1) {
  74                 percpu_ref_kill(&q->q_usage_counter);
  75                 blk_mq_run_hw_queues(q, false);
  76         }
  77 }
  78 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
  79
  80 void blk_mq_freeze_queue_wait(struct request_queue *q)
  81 {
  82         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
  83 }
  84 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
  85
  86 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
  87                                      unsigned long timeout)
  88 {
  89         return wait_event_timeout(q->mq_freeze_wq,
  90                                         percpu_ref_is_zero(&q->q_usage_counter),
  91                                         timeout);
  92 }
  93 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
  94
  95 /*
  96  * Guarantee no request is in use, so we can change any data structure of
  97  * the queue afterward.
  98  */
  99 void blk_freeze_queue(struct request_queue *q)
 100 {
 101         /*
 102          * In the !blk_mq case we are only calling this to kill the
 103          * q_usage_counter, otherwise this increases the freeze depth
 104          * and waits for it to return to zero.  For this reason there is
 105          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 106          * exported to drivers as the only user for unfreeze is blk_mq.
 107          */
 108         blk_mq_freeze_queue_start(q);
 109         blk_mq_freeze_queue_wait(q);
 110 }
 111
 112 void blk_mq_freeze_queue(struct request_queue *q)
 113 {
 114         /*
 115          * ...just an alias to keep freeze and unfreeze actions balanced
 116          * in the blk_mq_* namespace
 117          */
 118         blk_freeze_queue(q);
 119 }
 120 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 121
 122 void blk_mq_unfreeze_queue(struct request_queue *q)
 123 {
 124         int freeze_depth;
 125
 126         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 127         WARN_ON_ONCE(freeze_depth < 0);
 128         if (!freeze_depth) {
 129                 percpu_ref_reinit(&q->q_usage_counter);
 130                 wake_up_all(&q->mq_freeze_wq);
 131         }
 132 }
 133 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 134
 135 /**
 136  * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
 137  * @q: request queue.
 138  *
 139  * Note: this function does not prevent that the struct request end_io()
 140  * callback function is invoked. Additionally, it is not prevented that
 141  * new queue_rq() calls occur unless the queue has been stopped first.
 142  */
 143 void blk_mq_quiesce_queue(struct request_queue *q)
 144 {
 145         struct blk_mq_hw_ctx *hctx;
 146         unsigned int i;
 147         bool rcu = false;
 148
 149         blk_mq_stop_hw_queues(q);
 150
 151         queue_for_each_hw_ctx(q, hctx, i) {
 152                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 153                         synchronize_srcu(&hctx->queue_rq_srcu);
 154                 else
 155                         rcu = true;
 156         }
 157         if (rcu)
 158                 synchronize_rcu();
 159 }
 160 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 161
 162 void blk_mq_wake_waiters(struct request_queue *q)
 163 {
 164         struct blk_mq_hw_ctx *hctx;
 165         unsigned int i;
 166
 167         queue_for_each_hw_ctx(q, hctx, i)
 168                 if (blk_mq_hw_queue_mapped(hctx))
 169                         blk_mq_tag_wakeup_all(hctx->tags, true);
 170
 171         /*
 172          * If we are called because the queue has now been marked as
 173          * dying, we need to ensure that processes currently waiting on
 174          * the queue are notified as well.
 175          */
 176         wake_up_all(&q->mq_freeze_wq);
 177 }
 178
 179 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 180 {
 181         return blk_mq_has_free_tags(hctx->tags);
 182 }
 183 EXPORT_SYMBOL(blk_mq_can_queue);
 184
 185 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 186                         struct request *rq, unsigned int op)
 187 {
 188         INIT_LIST_HEAD(&rq->queuelist);
 189         /* csd/requeue_work/fifo_time is initialized before use */
 190         rq->q = q;
 191         rq->mq_ctx = ctx;
 192         rq->cmd_flags = op;
 193         if (blk_queue_io_stat(q))
 194                 rq->rq_flags |= RQF_IO_STAT;
 195         /* do not touch atomic flags, it needs atomic ops against the timer */
 196         rq->cpu = -1;
 197         INIT_HLIST_NODE(&rq->hash);
 198         RB_CLEAR_NODE(&rq->rb_node);
 199         rq->rq_disk = NULL;
 200         rq->part = NULL;
 201         rq->start_time = jiffies;
 202 #ifdef CONFIG_BLK_CGROUP
 203         rq->rl = NULL;
 204         set_start_time_ns(rq);
 205         rq->io_start_time_ns = 0;
 206 #endif
 207         rq->nr_phys_segments = 0;
 208 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 209         rq->nr_integrity_segments = 0;
 210 #endif
 211         rq->special = NULL;
 212         /* tag was already set */
 213         rq->errors = 0;
 214         rq->extra_len = 0;
 215
 216         INIT_LIST_HEAD(&rq->timeout_list);
 217         rq->timeout = 0;
 218
 219         rq->end_io = NULL;
 220         rq->end_io_data = NULL;
 221         rq->next_rq = NULL;
 222
 223         ctx->rq_dispatched[op_is_sync(op)]++;
 224 }
 225 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
 226
 227 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
 228                                        unsigned int op)
 229 {
 230         struct request *rq;
 231         unsigned int tag;
 232
 233         tag = blk_mq_get_tag(data);
 234         if (tag != BLK_MQ_TAG_FAIL) {
 235                 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 236
 237                 rq = tags->static_rqs[tag];
 238
 239                 if (data->flags & BLK_MQ_REQ_INTERNAL) {
 240                         rq->tag = -1;
 241                         rq->internal_tag = tag;
 242                 } else {
 243                         if (blk_mq_tag_busy(data->hctx)) {
 244                                 rq->rq_flags = RQF_MQ_INFLIGHT;
 245                                 atomic_inc(&data->hctx->nr_active);
 246                         }
 247                         rq->tag = tag;
 248                         rq->internal_tag = -1;
 249                         data->hctx->tags->rqs[rq->tag] = rq;
 250                 }
 251
 252                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
 253                 return rq;
 254         }
 255
 256         return NULL;
 257 }
 258 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
 259
 260 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
 261                 unsigned int flags)
 262 {
 263         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 264         struct request *rq;
 265         int ret;
 266
 267         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 268         if (ret)
 269                 return ERR_PTR(ret);
 270
 271         rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
 272
 273         blk_mq_put_ctx(alloc_data.ctx);
 274         blk_queue_exit(q);
 275
 276         if (!rq)
 277                 return ERR_PTR(-EWOULDBLOCK);
 278
 279         rq->__data_len = 0;
 280         rq->__sector = (sector_t) -1;
 281         rq->bio = rq->biotail = NULL;
 282         return rq;
 283 }
 284 EXPORT_SYMBOL(blk_mq_alloc_request);
 285
 286 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
 287                 unsigned int flags, unsigned int hctx_idx)
 288 {
 289         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 290         struct request *rq;
 291         unsigned int cpu;
 292         int ret;
 293
 294         /*
 295          * If the tag allocator sleeps we could get an allocation for a
 296          * different hardware context.  No need to complicate the low level
 297          * allocator for this for the rare use case of a command tied to
 298          * a specific queue.
 299          */
 300         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 301                 return ERR_PTR(-EINVAL);
 302
 303         if (hctx_idx >= q->nr_hw_queues)
 304                 return ERR_PTR(-EIO);
 305
 306         ret = blk_queue_enter(q, true);
 307         if (ret)
 308                 return ERR_PTR(ret);
 309
 310         /*
 311          * Check if the hardware context is actually mapped to anything.
 312          * If not tell the caller that it should skip this queue.
 313          */
 314         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 315         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 316                 blk_queue_exit(q);
 317                 return ERR_PTR(-EXDEV);
 318         }
 319         cpu = cpumask_first(alloc_data.hctx->cpumask);
 320         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 321
 322         rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
 323
 324         blk_queue_exit(q);
 325
 326         if (!rq)
 327                 return ERR_PTR(-EWOULDBLOCK);
 328
 329         return rq;
 330 }
 331 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 332
 333 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
 334                              struct request *rq)
 335 {
 336         const int sched_tag = rq->internal_tag;
 337         struct request_queue *q = rq->q;
 338
 339         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 340                 atomic_dec(&hctx->nr_active);
 341
 342         wbt_done(q->rq_wb, &rq->issue_stat);
 343         rq->rq_flags = 0;
 344
 345         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 346         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 347         if (rq->tag != -1)
 348                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 349         if (sched_tag != -1)
 350                 blk_mq_sched_completed_request(hctx, rq);
 351         blk_mq_sched_restart_queues(hctx);
 352         blk_queue_exit(q);
 353 }
 354
 355 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
 356                                      struct request *rq)
 357 {
 358         struct blk_mq_ctx *ctx = rq->mq_ctx;
 359
 360         ctx->rq_completed[rq_is_sync(rq)]++;
 361         __blk_mq_finish_request(hctx, ctx, rq);
 362 }
 363
 364 void blk_mq_finish_request(struct request *rq)
 365 {
 366         blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
 367 }
 368
 369 void blk_mq_free_request(struct request *rq)
 370 {
 371         blk_mq_sched_put_request(rq);
 372 }
 373 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 374
 375 inline void __blk_mq_end_request(struct request *rq, int error)
 376 {
 377         blk_account_io_done(rq);
 378
 379         if (rq->end_io) {
 380                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 381                 rq->end_io(rq, error);
 382         } else {
 383                 if (unlikely(blk_bidi_rq(rq)))
 384                         blk_mq_free_request(rq->next_rq);
 385                 blk_mq_free_request(rq);
 386         }
 387 }
 388 EXPORT_SYMBOL(__blk_mq_end_request);
 389
 390 void blk_mq_end_request(struct request *rq, int error)
 391 {
 392         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 393                 BUG();
 394         __blk_mq_end_request(rq, error);
 395 }
 396 EXPORT_SYMBOL(blk_mq_end_request);
 397
 398 static void __blk_mq_complete_request_remote(void *data)
 399 {
 400         struct request *rq = data;
 401
 402         rq->q->softirq_done_fn(rq);
 403 }
 404
 405 static void blk_mq_ipi_complete_request(struct request *rq)
 406 {
 407         struct blk_mq_ctx *ctx = rq->mq_ctx;
 408         bool shared = false;
 409         int cpu;
 410
 411         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 412                 rq->q->softirq_done_fn(rq);
 413                 return;
 414         }
 415
 416         cpu = get_cpu();
 417         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 418                 shared = cpus_share_cache(cpu, ctx->cpu);
 419
 420         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 421                 rq->csd.func = __blk_mq_complete_request_remote;
 422                 rq->csd.info = rq;
 423                 rq->csd.flags = 0;
 424                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 425         } else {
 426                 rq->q->softirq_done_fn(rq);
 427         }
 428         put_cpu();
 429 }
 430
 431 static void blk_mq_stat_add(struct request *rq)
 432 {
 433         if (rq->rq_flags & RQF_STATS) {
 434                 /*
 435                  * We could rq->mq_ctx here, but there's less of a risk
 436                  * of races if we have the completion event add the stats
 437                  * to the local software queue.
 438                  */
 439                 struct blk_mq_ctx *ctx;
 440
 441                 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
 442                 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
 443         }
 444 }
 445
 446 static void __blk_mq_complete_request(struct request *rq)
 447 {
 448         struct request_queue *q = rq->q;
 449
 450         blk_mq_stat_add(rq);
 451
 452         if (!q->softirq_done_fn)
 453                 blk_mq_end_request(rq, rq->errors);
 454         else
 455                 blk_mq_ipi_complete_request(rq);
 456 }
 457
 458 /**
 459  * blk_mq_complete_request - end I/O on a request
 460  * @rq:         the request being processed
 461  *
 462  * Description:
 463  *      Ends all I/O on a request. It does not handle partial completions.
 464  *      The actual completion happens out-of-order, through a IPI handler.
 465  **/
 466 void blk_mq_complete_request(struct request *rq, int error)
 467 {
 468         struct request_queue *q = rq->q;
 469
 470         if (unlikely(blk_should_fake_timeout(q)))
 471                 return;
 472         if (!blk_mark_rq_complete(rq)) {
 473                 rq->errors = error;
 474                 __blk_mq_complete_request(rq);
 475         }
 476 }
 477 EXPORT_SYMBOL(blk_mq_complete_request);
 478
 479 int blk_mq_request_started(struct request *rq)
 480 {
 481         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 482 }
 483 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 484
 485 void blk_mq_start_request(struct request *rq)
 486 {
 487         struct request_queue *q = rq->q;
 488
 489         blk_mq_sched_started_request(rq);
 490
 491         trace_block_rq_issue(q, rq);
 492
 493         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 494                 blk_stat_set_issue_time(&rq->issue_stat);
 495                 rq->rq_flags |= RQF_STATS;
 496                 wbt_issue(q->rq_wb, &rq->issue_stat);
 497         }
 498
 499         blk_add_timer(rq);
 500
 501         /*
 502          * Ensure that ->deadline is visible before set the started
 503          * flag and clear the completed flag.
 504          */
 505         smp_mb__before_atomic();
 506
 507         /*
 508          * Mark us as started and clear complete. Complete might have been
 509          * set if requeue raced with timeout, which then marked it as
 510          * complete. So be sure to clear complete again when we start
 511          * the request, otherwise we'll ignore the completion event.
 512          */
 513         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 514                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 515         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 516                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 517
 518         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 519                 /*
 520                  * Make sure space for the drain appears.  We know we can do
 521                  * this because max_hw_segments has been adjusted to be one
 522                  * fewer than the device can handle.
 523                  */
 524                 rq->nr_phys_segments++;
 525         }
 526 }
 527 EXPORT_SYMBOL(blk_mq_start_request);
 528
 529 static void __blk_mq_requeue_request(struct request *rq)
 530 {
 531         struct request_queue *q = rq->q;
 532
 533         trace_block_rq_requeue(q, rq);
 534         wbt_requeue(q->rq_wb, &rq->issue_stat);
 535         blk_mq_sched_requeue_request(rq);
 536
 537         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 538                 if (q->dma_drain_size && blk_rq_bytes(rq))
 539                         rq->nr_phys_segments--;
 540         }
 541 }
 542
 543 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 544 {
 545         __blk_mq_requeue_request(rq);
 546
 547         BUG_ON(blk_queued_rq(rq));
 548         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 549 }
 550 EXPORT_SYMBOL(blk_mq_requeue_request);
 551
 552 static void blk_mq_requeue_work(struct work_struct *work)
 553 {
 554         struct request_queue *q =
 555                 container_of(work, struct request_queue, requeue_work.work);
 556         LIST_HEAD(rq_list);
 557         struct request *rq, *next;
 558         unsigned long flags;
 559
 560         spin_lock_irqsave(&q->requeue_lock, flags);
 561         list_splice_init(&q->requeue_list, &rq_list);
 562         spin_unlock_irqrestore(&q->requeue_lock, flags);
 563
 564         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 565                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 566                         continue;
 567
 568                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 569                 list_del_init(&rq->queuelist);
 570                 blk_mq_sched_insert_request(rq, true, false, false, true);
 571         }
 572
 573         while (!list_empty(&rq_list)) {
 574                 rq = list_entry(rq_list.next, struct request, queuelist);
 575                 list_del_init(&rq->queuelist);
 576                 blk_mq_sched_insert_request(rq, false, false, false, true);
 577         }
 578
 579         blk_mq_run_hw_queues(q, false);
 580 }
 581
 582 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 583                                 bool kick_requeue_list)
 584 {
 585         struct request_queue *q = rq->q;
 586         unsigned long flags;
 587
 588         /*
 589          * We abuse this flag that is otherwise used by the I/O scheduler to
 590          * request head insertation from the workqueue.
 591          */
 592         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 593
 594         spin_lock_irqsave(&q->requeue_lock, flags);
 595         if (at_head) {
 596                 rq->rq_flags |= RQF_SOFTBARRIER;
 597                 list_add(&rq->queuelist, &q->requeue_list);
 598         } else {
 599                 list_add_tail(&rq->queuelist, &q->requeue_list);
 600         }
 601         spin_unlock_irqrestore(&q->requeue_lock, flags);
 602
 603         if (kick_requeue_list)
 604                 blk_mq_kick_requeue_list(q);
 605 }
 606 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 607
 608 void blk_mq_kick_requeue_list(struct request_queue *q)
 609 {
 610         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 611 }
 612 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 613
 614 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 615                                     unsigned long msecs)
 616 {
 617         kblockd_schedule_delayed_work(&q->requeue_work,
 618                                       msecs_to_jiffies(msecs));
 619 }
 620 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 621
 622 void blk_mq_abort_requeue_list(struct request_queue *q)
 623 {
 624         unsigned long flags;
 625         LIST_HEAD(rq_list);
 626
 627         spin_lock_irqsave(&q->requeue_lock, flags);
 628         list_splice_init(&q->requeue_list, &rq_list);
 629         spin_unlock_irqrestore(&q->requeue_lock, flags);
 630
 631         while (!list_empty(&rq_list)) {
 632                 struct request *rq;
 633
 634                 rq = list_first_entry(&rq_list, struct request, queuelist);
 635                 list_del_init(&rq->queuelist);
 636                 rq->errors = -EIO;
 637                 blk_mq_end_request(rq, rq->errors);
 638         }
 639 }
 640 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 641
 642 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 643 {
 644         if (tag < tags->nr_tags) {
 645                 prefetch(tags->rqs[tag]);
 646                 return tags->rqs[tag];
 647         }
 648
 649         return NULL;
 650 }
 651 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 652
 653 struct blk_mq_timeout_data {
 654         unsigned long next;
 655         unsigned int next_set;
 656 };
 657
 658 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 659 {
 660         const struct blk_mq_ops *ops = req->q->mq_ops;
 661         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 662
 663         /*
 664          * We know that complete is set at this point. If STARTED isn't set
 665          * anymore, then the request isn't active and the "timeout" should
 666          * just be ignored. This can happen due to the bitflag ordering.
 667          * Timeout first checks if STARTED is set, and if it is, assumes
 668          * the request is active. But if we race with completion, then
 669          * we both flags will get cleared. So check here again, and ignore
 670          * a timeout event with a request that isn't active.
 671          */
 672         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 673                 return;
 674
 675         if (ops->timeout)
 676                 ret = ops->timeout(req, reserved);
 677
 678         switch (ret) {
 679         case BLK_EH_HANDLED:
 680                 __blk_mq_complete_request(req);
 681                 break;
 682         case BLK_EH_RESET_TIMER:
 683                 blk_add_timer(req);
 684                 blk_clear_rq_complete(req);
 685                 break;
 686         case BLK_EH_NOT_HANDLED:
 687                 break;
 688         default:
 689                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 690                 break;
 691         }
 692 }
 693
 694 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 695                 struct request *rq, void *priv, bool reserved)
 696 {
 697         struct blk_mq_timeout_data *data = priv;
 698
 699         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 700                 return;
 701
 702         if (time_after_eq(jiffies, rq->deadline)) {
 703                 if (!blk_mark_rq_complete(rq))
 704                         blk_mq_rq_timed_out(rq, reserved);
 705         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 706                 data->next = rq->deadline;
 707                 data->next_set = 1;
 708         }
 709 }
 710
 711 static void blk_mq_timeout_work(struct work_struct *work)
 712 {
 713         struct request_queue *q =
 714                 container_of(work, struct request_queue, timeout_work);
 715         struct blk_mq_timeout_data data = {
 716                 .next           = 0,
 717                 .next_set       = 0,
 718         };
 719         int i;
 720
 721         /* A deadlock might occur if a request is stuck requiring a
 722          * timeout at the same time a queue freeze is waiting
 723          * completion, since the timeout code would not be able to
 724          * acquire the queue reference here.
 725          *
 726          * That's why we don't use blk_queue_enter here; instead, we use
 727          * percpu_ref_tryget directly, because we need to be able to
 728          * obtain a reference even in the short window between the queue
 729          * starting to freeze, by dropping the first reference in
 730          * blk_mq_freeze_queue_start, and the moment the last request is
 731          * consumed, marked by the instant q_usage_counter reaches
 732          * zero.
 733          */
 734         if (!percpu_ref_tryget(&q->q_usage_counter))
 735                 return;
 736
 737         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 738
 739         if (data.next_set) {
 740                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 741                 mod_timer(&q->timeout, data.next);
 742         } else {
 743                 struct blk_mq_hw_ctx *hctx;
 744
 745                 queue_for_each_hw_ctx(q, hctx, i) {
 746                         /* the hctx may be unmapped, so check it here */
 747                         if (blk_mq_hw_queue_mapped(hctx))
 748                                 blk_mq_tag_idle(hctx);
 749                 }
 750         }
 751         blk_queue_exit(q);
 752 }
 753
 754 /*
 755  * Reverse check our software queue for entries that we could potentially
 756  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 757  * too much time checking for merges.
 758  */
 759 static bool blk_mq_attempt_merge(struct request_queue *q,
 760                                  struct blk_mq_ctx *ctx, struct bio *bio)
 761 {
 762         struct request *rq;
 763         int checked = 8;
 764
 765         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 766                 bool merged = false;
 767
 768                 if (!checked--)
 769                         break;
 770
 771                 if (!blk_rq_merge_ok(rq, bio))
 772                         continue;
 773
 774                 switch (blk_try_merge(rq, bio)) {
 775                 case ELEVATOR_BACK_MERGE:
 776                         if (blk_mq_sched_allow_merge(q, rq, bio))
 777                                 merged = bio_attempt_back_merge(q, rq, bio);
 778                         break;
 779                 case ELEVATOR_FRONT_MERGE:
 780                         if (blk_mq_sched_allow_merge(q, rq, bio))
 781                                 merged = bio_attempt_front_merge(q, rq, bio);
 782                         break;
 783                 case ELEVATOR_DISCARD_MERGE:
 784                         merged = bio_attempt_discard_merge(q, rq, bio);
 785                         break;
 786                 default:
 787                         continue;
 788                 }
 789
 790                 if (merged)
 791                         ctx->rq_merged++;
 792                 return merged;
 793         }
 794
 795         return false;
 796 }
 797
 798 struct flush_busy_ctx_data {
 799         struct blk_mq_hw_ctx *hctx;
 800         struct list_head *list;
 801 };
 802
 803 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 804 {
 805         struct flush_busy_ctx_data *flush_data = data;
 806         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 807         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 808
 809         sbitmap_clear_bit(sb, bitnr);
 810         spin_lock(&ctx->lock);
 811         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 812         spin_unlock(&ctx->lock);
 813         return true;
 814 }
 815
 816 /*
 817  * Process software queues that have been marked busy, splicing them
 818  * to the for-dispatch
 819  */
 820 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 821 {
 822         struct flush_busy_ctx_data data = {
 823                 .hctx = hctx,
 824                 .list = list,
 825         };
 826
 827         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 828 }
 829 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 830
 831 static inline unsigned int queued_to_index(unsigned int queued)
 832 {
 833         if (!queued)
 834                 return 0;
 835
 836         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 837 }
 838
 839 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 840                            bool wait)
 841 {
 842         struct blk_mq_alloc_data data = {
 843                 .q = rq->q,
 844                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 845                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 846         };
 847
 848         if (rq->tag != -1)
 849                 goto done;
 850
 851         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 852                 data.flags |= BLK_MQ_REQ_RESERVED;
 853
 854         rq->tag = blk_mq_get_tag(&data);
 855         if (rq->tag >= 0) {
 856                 if (blk_mq_tag_busy(data.hctx)) {
 857                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 858                         atomic_inc(&data.hctx->nr_active);
 859                 }
 860                 data.hctx->tags->rqs[rq->tag] = rq;
 861         }
 862
 863 done:
 864         if (hctx)
 865                 *hctx = data.hctx;
 866         return rq->tag != -1;
 867 }
 868
 869 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 870                                     struct request *rq)
 871 {
 872         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
 873         rq->tag = -1;
 874
 875         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 876                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
 877                 atomic_dec(&hctx->nr_active);
 878         }
 879 }
 880
 881 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
 882                                        struct request *rq)
 883 {
 884         if (rq->tag == -1 || rq->internal_tag == -1)
 885                 return;
 886
 887         __blk_mq_put_driver_tag(hctx, rq);
 888 }
 889
 890 static void blk_mq_put_driver_tag(struct request *rq)
 891 {
 892         struct blk_mq_hw_ctx *hctx;
 893
 894         if (rq->tag == -1 || rq->internal_tag == -1)
 895                 return;
 896
 897         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
 898         __blk_mq_put_driver_tag(hctx, rq);
 899 }
 900
 901 /*
 902  * If we fail getting a driver tag because all the driver tags are already
 903  * assigned and on the dispatch list, BUT the first entry does not have a
 904  * tag, then we could deadlock. For that case, move entries with assigned
 905  * driver tags to the front, leaving the set of tagged requests in the
 906  * same order, and the untagged set in the same order.
 907  */
 908 static bool reorder_tags_to_front(struct list_head *list)
 909 {
 910         struct request *rq, *tmp, *first = NULL;
 911
 912         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
 913                 if (rq == first)
 914                         break;
 915                 if (rq->tag != -1) {
 916                         list_move(&rq->queuelist, list);
 917                         if (!first)
 918                                 first = rq;
 919                 }
 920         }
 921
 922         return first != NULL;
 923 }
 924
 925 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
 926                                 void *key)
 927 {
 928         struct blk_mq_hw_ctx *hctx;
 929
 930         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
 931
 932         list_del(&wait->task_list);
 933         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
 934         blk_mq_run_hw_queue(hctx, true);
 935         return 1;
 936 }
 937
 938 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
 939 {
 940         struct sbq_wait_state *ws;
 941
 942         /*
 943          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
 944          * The thread which wins the race to grab this bit adds the hardware
 945          * queue to the wait queue.
 946          */
 947         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
 948             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
 949                 return false;
 950
 951         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
 952         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
 953
 954         /*
 955          * As soon as this returns, it's no longer safe to fiddle with
 956          * hctx->dispatch_wait, since a completion can wake up the wait queue
 957          * and unlock the bit.
 958          */
 959         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
 960         return true;
 961 }
 962
 963 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
 964 {
 965         struct blk_mq_hw_ctx *hctx;
 966         struct request *rq;
 967         LIST_HEAD(driver_list);
 968         struct list_head *dptr;
 969         int errors, queued, ret = BLK_MQ_RQ_QUEUE_OK;
 970
 971         if (list_empty(list))
 972                 return false;
 973
 974         /*
 975          * Start off with dptr being NULL, so we start the first request
 976          * immediately, even if we have more pending.
 977          */
 978         dptr = NULL;
 979
 980         /*
 981          * Now process all the entries, sending them to the driver.
 982          */
 983         errors = queued = 0;
 984         do {
 985                 struct blk_mq_queue_data bd;
 986
 987                 rq = list_first_entry(list, struct request, queuelist);
 988                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
 989                         if (!queued && reorder_tags_to_front(list))
 990                                 continue;
 991
 992                         /*
 993                          * The initial allocation attempt failed, so we need to
 994                          * rerun the hardware queue when a tag is freed.
 995                          */
 996                         if (blk_mq_dispatch_wait_add(hctx)) {
 997                                 /*
 998                                  * It's possible that a tag was freed in the
 999                                  * window between the allocation failure and
1000                                  * adding the hardware queue to the wait queue.
1001                                  */
1002                                 if (!blk_mq_get_driver_tag(rq, &hctx, false))
1003                                         break;
1004                         } else {
1005                                 break;
1006                         }
1007                 }
1008
1009                 list_del_init(&rq->queuelist);
1010
1011                 bd.rq = rq;
1012                 bd.list = dptr;
1013
1014                 /*
1015                  * Flag last if we have no more requests, or if we have more
1016                  * but can't assign a driver tag to it.
1017                  */
1018                 if (list_empty(list))
1019                         bd.last = true;
1020                 else {
1021                         struct request *nxt;
1022
1023                         nxt = list_first_entry(list, struct request, queuelist);
1024                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1025                 }
1026
1027                 ret = q->mq_ops->queue_rq(hctx, &bd);
1028                 switch (ret) {
1029                 case BLK_MQ_RQ_QUEUE_OK:
1030                         queued++;
1031                         break;
1032                 case BLK_MQ_RQ_QUEUE_BUSY:
1033                         blk_mq_put_driver_tag_hctx(hctx, rq);
1034                         list_add(&rq->queuelist, list);
1035                         __blk_mq_requeue_request(rq);
1036                         break;
1037                 default:
1038                         pr_err("blk-mq: bad return on queue: %d\n", ret);
1039                 case BLK_MQ_RQ_QUEUE_ERROR:
1040                         errors++;
1041                         rq->errors = -EIO;
1042                         blk_mq_end_request(rq, rq->errors);
1043                         break;
1044                 }
1045
1046                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1047                         break;
1048
1049                 /*
1050                  * We've done the first request. If we have more than 1
1051                  * left in the list, set dptr to defer issue.
1052                  */
1053                 if (!dptr && list->next != list->prev)
1054                         dptr = &driver_list;
1055         } while (!list_empty(list));
1056
1057         hctx->dispatched[queued_to_index(queued)]++;
1058
1059         /*
1060          * Any items that need requeuing? Stuff them into hctx->dispatch,
1061          * that is where we will continue on next queue run.
1062          */
1063         if (!list_empty(list)) {
1064                 /*
1065                  * If we got a driver tag for the next request already,
1066                  * free it again.
1067                  */
1068                 rq = list_first_entry(list, struct request, queuelist);
1069                 blk_mq_put_driver_tag(rq);
1070
1071                 spin_lock(&hctx->lock);
1072                 list_splice_init(list, &hctx->dispatch);
1073                 spin_unlock(&hctx->lock);
1074
1075                 /*
1076                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1077                  * it's possible the queue is stopped and restarted again
1078                  * before this. Queue restart will dispatch requests. And since
1079                  * requests in rq_list aren't added into hctx->dispatch yet,
1080                  * the requests in rq_list might get lost.
1081                  *
1082                  * blk_mq_run_hw_queue() already checks the STOPPED bit
1083                  *
1084                  * If RESTART or TAG_WAITING is set, then let completion restart
1085                  * the queue instead of potentially looping here.
1086                  */
1087                 if (!blk_mq_sched_needs_restart(hctx) &&
1088                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1089                         blk_mq_run_hw_queue(hctx, true);
1090         }
1091
1092         return (queued + errors) != 0;
1093 }
1094
1095 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1096 {
1097         int srcu_idx;
1098
1099         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1100                 cpu_online(hctx->next_cpu));
1101
1102         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1103                 rcu_read_lock();
1104                 blk_mq_sched_dispatch_requests(hctx);
1105                 rcu_read_unlock();
1106         } else {
1107                 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1108                 blk_mq_sched_dispatch_requests(hctx);
1109                 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1110         }
1111 }
1112
1113 /*
1114  * It'd be great if the workqueue API had a way to pass
1115  * in a mask and had some smarts for more clever placement.
1116  * For now we just round-robin here, switching for every
1117  * BLK_MQ_CPU_WORK_BATCH queued items.
1118  */
1119 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1120 {
1121         if (hctx->queue->nr_hw_queues == 1)
1122                 return WORK_CPU_UNBOUND;
1123
1124         if (--hctx->next_cpu_batch <= 0) {
1125                 int next_cpu;
1126
1127                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1128                 if (next_cpu >= nr_cpu_ids)
1129                         next_cpu = cpumask_first(hctx->cpumask);
1130
1131                 hctx->next_cpu = next_cpu;
1132                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1133         }
1134
1135         return hctx->next_cpu;
1136 }
1137
1138 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1139 {
1140         if (unlikely(blk_mq_hctx_stopped(hctx) ||
1141                      !blk_mq_hw_queue_mapped(hctx)))
1142                 return;
1143
1144         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1145                 int cpu = get_cpu();
1146                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1147                         __blk_mq_run_hw_queue(hctx);
1148                         put_cpu();
1149                         return;
1150                 }
1151
1152                 put_cpu();
1153         }
1154
1155         kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1156 }
1157
1158 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1159 {
1160         struct blk_mq_hw_ctx *hctx;
1161         int i;
1162
1163         queue_for_each_hw_ctx(q, hctx, i) {
1164                 if (!blk_mq_hctx_has_pending(hctx) ||
1165                     blk_mq_hctx_stopped(hctx))
1166                         continue;
1167
1168                 blk_mq_run_hw_queue(hctx, async);
1169         }
1170 }
1171 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1172
1173 /**
1174  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1175  * @q: request queue.
1176  *
1177  * The caller is responsible for serializing this function against
1178  * blk_mq_{start,stop}_hw_queue().
1179  */
1180 bool blk_mq_queue_stopped(struct request_queue *q)
1181 {
1182         struct blk_mq_hw_ctx *hctx;
1183         int i;
1184
1185         queue_for_each_hw_ctx(q, hctx, i)
1186                 if (blk_mq_hctx_stopped(hctx))
1187                         return true;
1188
1189         return false;
1190 }
1191 EXPORT_SYMBOL(blk_mq_queue_stopped);
1192
1193 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1194 {
1195         cancel_work(&hctx->run_work);
1196         cancel_delayed_work(&hctx->delay_work);
1197         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1198 }
1199 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1200
1201 void blk_mq_stop_hw_queues(struct request_queue *q)
1202 {
1203         struct blk_mq_hw_ctx *hctx;
1204         int i;
1205
1206         queue_for_each_hw_ctx(q, hctx, i)
1207                 blk_mq_stop_hw_queue(hctx);
1208 }
1209 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1210
1211 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1212 {
1213         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1214
1215         blk_mq_run_hw_queue(hctx, false);
1216 }
1217 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1218
1219 void blk_mq_start_hw_queues(struct request_queue *q)
1220 {
1221         struct blk_mq_hw_ctx *hctx;
1222         int i;
1223
1224         queue_for_each_hw_ctx(q, hctx, i)
1225                 blk_mq_start_hw_queue(hctx);
1226 }
1227 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1228
1229 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1230 {
1231         if (!blk_mq_hctx_stopped(hctx))
1232                 return;
1233
1234         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1235         blk_mq_run_hw_queue(hctx, async);
1236 }
1237 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1238
1239 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1240 {
1241         struct blk_mq_hw_ctx *hctx;
1242         int i;
1243
1244         queue_for_each_hw_ctx(q, hctx, i)
1245                 blk_mq_start_stopped_hw_queue(hctx, async);
1246 }
1247 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1248
1249 static void blk_mq_run_work_fn(struct work_struct *work)
1250 {
1251         struct blk_mq_hw_ctx *hctx;
1252
1253         hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1254
1255         __blk_mq_run_hw_queue(hctx);
1256 }
1257
1258 static void blk_mq_delay_work_fn(struct work_struct *work)
1259 {
1260         struct blk_mq_hw_ctx *hctx;
1261
1262         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1263
1264         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1265                 __blk_mq_run_hw_queue(hctx);
1266 }
1267
1268 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1269 {
1270         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1271                 return;
1272
1273         blk_mq_stop_hw_queue(hctx);
1274         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1275                         &hctx->delay_work, msecs_to_jiffies(msecs));
1276 }
1277 EXPORT_SYMBOL(blk_mq_delay_queue);
1278
1279 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1280                                             struct request *rq,
1281                                             bool at_head)
1282 {
1283         struct blk_mq_ctx *ctx = rq->mq_ctx;
1284
1285         trace_block_rq_insert(hctx->queue, rq);
1286
1287         if (at_head)
1288                 list_add(&rq->queuelist, &ctx->rq_list);
1289         else
1290                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1291 }
1292
1293 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1294                              bool at_head)
1295 {
1296         struct blk_mq_ctx *ctx = rq->mq_ctx;
1297
1298         __blk_mq_insert_req_list(hctx, rq, at_head);
1299         blk_mq_hctx_mark_pending(hctx, ctx);
1300 }
1301
1302 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1303                             struct list_head *list)
1304
1305 {
1306         /*
1307          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1308          * offline now
1309          */
1310         spin_lock(&ctx->lock);
1311         while (!list_empty(list)) {
1312                 struct request *rq;
1313
1314                 rq = list_first_entry(list, struct request, queuelist);
1315                 BUG_ON(rq->mq_ctx != ctx);
1316                 list_del_init(&rq->queuelist);
1317                 __blk_mq_insert_req_list(hctx, rq, false);
1318         }
1319         blk_mq_hctx_mark_pending(hctx, ctx);
1320         spin_unlock(&ctx->lock);
1321 }
1322
1323 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1324 {
1325         struct request *rqa = container_of(a, struct request, queuelist);
1326         struct request *rqb = container_of(b, struct request, queuelist);
1327
1328         return !(rqa->mq_ctx < rqb->mq_ctx ||
1329                  (rqa->mq_ctx == rqb->mq_ctx &&
1330                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1331 }
1332
1333 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1334 {
1335         struct blk_mq_ctx *this_ctx;
1336         struct request_queue *this_q;
1337         struct request *rq;
1338         LIST_HEAD(list);
1339         LIST_HEAD(ctx_list);
1340         unsigned int depth;
1341
1342         list_splice_init(&plug->mq_list, &list);
1343
1344         list_sort(NULL, &list, plug_ctx_cmp);
1345
1346         this_q = NULL;
1347         this_ctx = NULL;
1348         depth = 0;
1349
1350         while (!list_empty(&list)) {
1351                 rq = list_entry_rq(list.next);
1352                 list_del_init(&rq->queuelist);
1353                 BUG_ON(!rq->q);
1354                 if (rq->mq_ctx != this_ctx) {
1355                         if (this_ctx) {
1356                                 trace_block_unplug(this_q, depth, from_schedule);
1357                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1358                                                                 &ctx_list,
1359                                                                 from_schedule);
1360                         }
1361
1362                         this_ctx = rq->mq_ctx;
1363                         this_q = rq->q;
1364                         depth = 0;
1365                 }
1366
1367                 depth++;
1368                 list_add_tail(&rq->queuelist, &ctx_list);
1369         }
1370
1371         /*
1372          * If 'this_ctx' is set, we know we have entries to complete
1373          * on 'ctx_list'. Do those.
1374          */
1375         if (this_ctx) {
1376                 trace_block_unplug(this_q, depth, from_schedule);
1377                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1378                                                 from_schedule);
1379         }
1380 }
1381
1382 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1383 {
1384         init_request_from_bio(rq, bio);
1385
1386         blk_account_io_start(rq, true);
1387 }
1388
1389 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1390 {
1391         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1392                 !blk_queue_nomerges(hctx->queue);
1393 }
1394
1395 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1396                                          struct blk_mq_ctx *ctx,
1397                                          struct request *rq, struct bio *bio)
1398 {
1399         if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1400                 blk_mq_bio_to_request(rq, bio);
1401                 spin_lock(&ctx->lock);
1402 insert_rq:
1403                 __blk_mq_insert_request(hctx, rq, false);
1404                 spin_unlock(&ctx->lock);
1405                 return false;
1406         } else {
1407                 struct request_queue *q = hctx->queue;
1408
1409                 spin_lock(&ctx->lock);
1410                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1411                         blk_mq_bio_to_request(rq, bio);
1412                         goto insert_rq;
1413                 }
1414
1415                 spin_unlock(&ctx->lock);
1416                 __blk_mq_finish_request(hctx, ctx, rq);
1417                 return true;
1418         }
1419 }
1420
1421 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1422 {
1423         if (rq->tag != -1)
1424                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1425
1426         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1427 }
1428
1429 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie,
1430                                       bool may_sleep)
1431 {
1432         struct request_queue *q = rq->q;
1433         struct blk_mq_queue_data bd = {
1434                 .rq = rq,
1435                 .list = NULL,
1436                 .last = 1
1437         };
1438         struct blk_mq_hw_ctx *hctx;
1439         blk_qc_t new_cookie;
1440         int ret;
1441
1442         if (q->elevator)
1443                 goto insert;
1444
1445         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1446                 goto insert;
1447
1448         new_cookie = request_to_qc_t(hctx, rq);
1449
1450         /*
1451          * For OK queue, we are done. For error, kill it. Any other
1452          * error (busy), just add it to our list as we previously
1453          * would have done
1454          */
1455         ret = q->mq_ops->queue_rq(hctx, &bd);
1456         if (ret == BLK_MQ_RQ_QUEUE_OK) {
1457                 *cookie = new_cookie;
1458                 return;
1459         }
1460
1461         __blk_mq_requeue_request(rq);
1462
1463         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1464                 *cookie = BLK_QC_T_NONE;
1465                 rq->errors = -EIO;
1466                 blk_mq_end_request(rq, rq->errors);
1467                 return;
1468         }
1469
1470 insert:
1471         blk_mq_sched_insert_request(rq, false, true, false, may_sleep);
1472 }
1473
1474 /*
1475  * Multiple hardware queue variant. This will not use per-process plugs,
1476  * but will attempt to bypass the hctx queueing if we can go straight to
1477  * hardware for SYNC IO.
1478  */
1479 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1480 {
1481         const int is_sync = op_is_sync(bio->bi_opf);
1482         const int is_flush_fua = op_is_flush(bio->bi_opf);
1483         struct blk_mq_alloc_data data = { .flags = 0 };
1484         struct request *rq;
1485         unsigned int request_count = 0, srcu_idx;
1486         struct blk_plug *plug;
1487         struct request *same_queue_rq = NULL;
1488         blk_qc_t cookie;
1489         unsigned int wb_acct;
1490
1491         blk_queue_bounce(q, &bio);
1492
1493         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1494                 bio_io_error(bio);
1495                 return BLK_QC_T_NONE;
1496         }
1497
1498         blk_queue_split(q, &bio, q->bio_split);
1499
1500         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1501             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1502                 return BLK_QC_T_NONE;
1503
1504         if (blk_mq_sched_bio_merge(q, bio))
1505                 return BLK_QC_T_NONE;
1506
1507         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1508
1509         trace_block_getrq(q, bio, bio->bi_opf);
1510
1511         rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1512         if (unlikely(!rq)) {
1513                 __wbt_done(q->rq_wb, wb_acct);
1514                 return BLK_QC_T_NONE;
1515         }
1516
1517         wbt_track(&rq->issue_stat, wb_acct);
1518
1519         cookie = request_to_qc_t(data.hctx, rq);
1520
1521         if (unlikely(is_flush_fua)) {
1522                 if (q->elevator)
1523                         goto elv_insert;
1524                 blk_mq_bio_to_request(rq, bio);
1525                 blk_insert_flush(rq);
1526                 goto run_queue;
1527         }
1528
1529         plug = current->plug;
1530         /*
1531          * If the driver supports defer issued based on 'last', then
1532          * queue it up like normal since we can potentially save some
1533          * CPU this way.
1534          */
1535         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1536             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1537                 struct request *old_rq = NULL;
1538
1539                 blk_mq_bio_to_request(rq, bio);
1540
1541                 /*
1542                  * We do limited plugging. If the bio can be merged, do that.
1543                  * Otherwise the existing request in the plug list will be
1544                  * issued. So the plug list will have one request at most
1545                  */
1546                 if (plug) {
1547                         /*
1548                          * The plug list might get flushed before this. If that
1549                          * happens, same_queue_rq is invalid and plug list is
1550                          * empty
1551                          */
1552                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1553                                 old_rq = same_queue_rq;
1554                                 list_del_init(&old_rq->queuelist);
1555                         }
1556                         list_add_tail(&rq->queuelist, &plug->mq_list);
1557                 } else /* is_sync */
1558                         old_rq = rq;
1559                 blk_mq_put_ctx(data.ctx);
1560                 if (!old_rq)
1561                         goto done;
1562
1563                 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1564                         rcu_read_lock();
1565                         blk_mq_try_issue_directly(old_rq, &cookie, false);
1566                         rcu_read_unlock();
1567                 } else {
1568                         srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1569                         blk_mq_try_issue_directly(old_rq, &cookie, true);
1570                         srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1571                 }
1572                 goto done;
1573         }
1574
1575         if (q->elevator) {
1576 elv_insert:
1577                 blk_mq_put_ctx(data.ctx);
1578                 blk_mq_bio_to_request(rq, bio);
1579                 blk_mq_sched_insert_request(rq, false, true,
1580                                                 !is_sync || is_flush_fua, true);
1581                 goto done;
1582         }
1583         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1584                 /*
1585                  * For a SYNC request, send it to the hardware immediately. For
1586                  * an ASYNC request, just ensure that we run it later on. The
1587                  * latter allows for merging opportunities and more efficient
1588                  * dispatching.
1589                  */
1590 run_queue:
1591                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1592         }
1593         blk_mq_put_ctx(data.ctx);
1594 done:
1595         return cookie;
1596 }
1597
1598 /*
1599  * Single hardware queue variant. This will attempt to use any per-process
1600  * plug for merging and IO deferral.
1601  */
1602 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1603 {
1604         const int is_sync = op_is_sync(bio->bi_opf);
1605         const int is_flush_fua = op_is_flush(bio->bi_opf);
1606         struct blk_plug *plug;
1607         unsigned int request_count = 0;
1608         struct blk_mq_alloc_data data = { .flags = 0 };
1609         struct request *rq;
1610         blk_qc_t cookie;
1611         unsigned int wb_acct;
1612
1613         blk_queue_bounce(q, &bio);
1614
1615         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1616                 bio_io_error(bio);
1617                 return BLK_QC_T_NONE;
1618         }
1619
1620         blk_queue_split(q, &bio, q->bio_split);
1621
1622         if (!is_flush_fua && !blk_queue_nomerges(q)) {
1623                 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1624                         return BLK_QC_T_NONE;
1625         } else
1626                 request_count = blk_plug_queued_count(q);
1627
1628         if (blk_mq_sched_bio_merge(q, bio))
1629                 return BLK_QC_T_NONE;
1630
1631         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1632
1633         trace_block_getrq(q, bio, bio->bi_opf);
1634
1635         rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1636         if (unlikely(!rq)) {
1637                 __wbt_done(q->rq_wb, wb_acct);
1638                 return BLK_QC_T_NONE;
1639         }
1640
1641         wbt_track(&rq->issue_stat, wb_acct);
1642
1643         cookie = request_to_qc_t(data.hctx, rq);
1644
1645         if (unlikely(is_flush_fua)) {
1646                 if (q->elevator)
1647                         goto elv_insert;
1648                 blk_mq_bio_to_request(rq, bio);
1649                 blk_insert_flush(rq);
1650                 goto run_queue;
1651         }
1652
1653         /*
1654          * A task plug currently exists. Since this is completely lockless,
1655          * utilize that to temporarily store requests until the task is
1656          * either done or scheduled away.
1657          */
1658         plug = current->plug;
1659         if (plug) {
1660                 struct request *last = NULL;
1661
1662                 blk_mq_bio_to_request(rq, bio);
1663
1664                 /*
1665                  * @request_count may become stale because of schedule
1666                  * out, so check the list again.
1667                  */
1668                 if (list_empty(&plug->mq_list))
1669                         request_count = 0;
1670                 if (!request_count)
1671                         trace_block_plug(q);
1672                 else
1673                         last = list_entry_rq(plug->mq_list.prev);
1674
1675                 blk_mq_put_ctx(data.ctx);
1676
1677                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1678                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1679                         blk_flush_plug_list(plug, false);
1680                         trace_block_plug(q);
1681                 }
1682
1683                 list_add_tail(&rq->queuelist, &plug->mq_list);
1684                 return cookie;
1685         }
1686
1687         if (q->elevator) {
1688 elv_insert:
1689                 blk_mq_put_ctx(data.ctx);
1690                 blk_mq_bio_to_request(rq, bio);
1691                 blk_mq_sched_insert_request(rq, false, true,
1692                                                 !is_sync || is_flush_fua, true);
1693                 goto done;
1694         }
1695         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1696                 /*
1697                  * For a SYNC request, send it to the hardware immediately. For
1698                  * an ASYNC request, just ensure that we run it later on. The
1699                  * latter allows for merging opportunities and more efficient
1700                  * dispatching.
1701                  */
1702 run_queue:
1703                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1704         }
1705
1706         blk_mq_put_ctx(data.ctx);
1707 done:
1708         return cookie;
1709 }
1710
1711 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1712                      unsigned int hctx_idx)
1713 {
1714         struct page *page;
1715
1716         if (tags->rqs && set->ops->exit_request) {
1717                 int i;
1718
1719                 for (i = 0; i < tags->nr_tags; i++) {
1720                         struct request *rq = tags->static_rqs[i];
1721
1722                         if (!rq)
1723                                 continue;
1724                         set->ops->exit_request(set->driver_data, rq,
1725                                                 hctx_idx, i);
1726                         tags->static_rqs[i] = NULL;
1727                 }
1728         }
1729
1730         while (!list_empty(&tags->page_list)) {
1731                 page = list_first_entry(&tags->page_list, struct page, lru);
1732                 list_del_init(&page->lru);
1733                 /*
1734                  * Remove kmemleak object previously allocated in
1735                  * blk_mq_init_rq_map().
1736                  */
1737                 kmemleak_free(page_address(page));
1738                 __free_pages(page, page->private);
1739         }
1740 }
1741
1742 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1743 {
1744         kfree(tags->rqs);
1745         tags->rqs = NULL;
1746         kfree(tags->static_rqs);
1747         tags->static_rqs = NULL;
1748
1749         blk_mq_free_tags(tags);
1750 }
1751
1752 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1753                                         unsigned int hctx_idx,
1754                                         unsigned int nr_tags,
1755                                         unsigned int reserved_tags)
1756 {
1757         struct blk_mq_tags *tags;
1758         int node;
1759
1760         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1761         if (node == NUMA_NO_NODE)
1762                 node = set->numa_node;
1763
1764         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1765                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1766         if (!tags)
1767                 return NULL;
1768
1769         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1770                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1771                                  node);
1772         if (!tags->rqs) {
1773                 blk_mq_free_tags(tags);
1774                 return NULL;
1775         }
1776
1777         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1778                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1779                                  node);
1780         if (!tags->static_rqs) {
1781                 kfree(tags->rqs);
1782                 blk_mq_free_tags(tags);
1783                 return NULL;
1784         }
1785
1786         return tags;
1787 }
1788
1789 static size_t order_to_size(unsigned int order)
1790 {
1791         return (size_t)PAGE_SIZE << order;
1792 }
1793
1794 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1795                      unsigned int hctx_idx, unsigned int depth)
1796 {
1797         unsigned int i, j, entries_per_page, max_order = 4;
1798         size_t rq_size, left;
1799         int node;
1800
1801         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1802         if (node == NUMA_NO_NODE)
1803                 node = set->numa_node;
1804
1805         INIT_LIST_HEAD(&tags->page_list);
1806
1807         /*
1808          * rq_size is the size of the request plus driver payload, rounded
1809          * to the cacheline size
1810          */
1811         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1812                                 cache_line_size());
1813         left = rq_size * depth;
1814
1815         for (i = 0; i < depth; ) {
1816                 int this_order = max_order;
1817                 struct page *page;
1818                 int to_do;
1819                 void *p;
1820
1821                 while (this_order && left < order_to_size(this_order - 1))
1822                         this_order--;
1823
1824                 do {
1825                         page = alloc_pages_node(node,
1826                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1827                                 this_order);
1828                         if (page)
1829                                 break;
1830                         if (!this_order--)
1831                                 break;
1832                         if (order_to_size(this_order) < rq_size)
1833                                 break;
1834                 } while (1);
1835
1836                 if (!page)
1837                         goto fail;
1838
1839                 page->private = this_order;
1840                 list_add_tail(&page->lru, &tags->page_list);
1841
1842                 p = page_address(page);
1843                 /*
1844                  * Allow kmemleak to scan these pages as they contain pointers
1845                  * to additional allocations like via ops->init_request().
1846                  */
1847                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1848                 entries_per_page = order_to_size(this_order) / rq_size;
1849                 to_do = min(entries_per_page, depth - i);
1850                 left -= to_do * rq_size;
1851                 for (j = 0; j < to_do; j++) {
1852                         struct request *rq = p;
1853
1854                         tags->static_rqs[i] = rq;
1855                         if (set->ops->init_request) {
1856                                 if (set->ops->init_request(set->driver_data,
1857                                                 rq, hctx_idx, i,
1858                                                 node)) {
1859                                         tags->static_rqs[i] = NULL;
1860                                         goto fail;
1861                                 }
1862                         }
1863
1864                         p += rq_size;
1865                         i++;
1866                 }
1867         }
1868         return 0;
1869
1870 fail:
1871         blk_mq_free_rqs(set, tags, hctx_idx);
1872         return -ENOMEM;
1873 }
1874
1875 /*
1876  * 'cpu' is going away. splice any existing rq_list entries from this
1877  * software queue to the hw queue dispatch list, and ensure that it
1878  * gets run.
1879  */
1880 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1881 {
1882         struct blk_mq_hw_ctx *hctx;
1883         struct blk_mq_ctx *ctx;
1884         LIST_HEAD(tmp);
1885
1886         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1887         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1888
1889         spin_lock(&ctx->lock);
1890         if (!list_empty(&ctx->rq_list)) {
1891                 list_splice_init(&ctx->rq_list, &tmp);
1892                 blk_mq_hctx_clear_pending(hctx, ctx);
1893         }
1894         spin_unlock(&ctx->lock);
1895
1896         if (list_empty(&tmp))
1897                 return 0;
1898
1899         spin_lock(&hctx->lock);
1900         list_splice_tail_init(&tmp, &hctx->dispatch);
1901         spin_unlock(&hctx->lock);
1902
1903         blk_mq_run_hw_queue(hctx, true);
1904         return 0;
1905 }
1906
1907 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1908 {
1909         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1910                                             &hctx->cpuhp_dead);
1911 }
1912
1913 /* hctx->ctxs will be freed in queue's release handler */
1914 static void blk_mq_exit_hctx(struct request_queue *q,
1915                 struct blk_mq_tag_set *set,
1916                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1917 {
1918         unsigned flush_start_tag = set->queue_depth;
1919
1920         blk_mq_tag_idle(hctx);
1921
1922         if (set->ops->exit_request)
1923                 set->ops->exit_request(set->driver_data,
1924                                        hctx->fq->flush_rq, hctx_idx,
1925                                        flush_start_tag + hctx_idx);
1926
1927         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1928
1929         if (set->ops->exit_hctx)
1930                 set->ops->exit_hctx(hctx, hctx_idx);
1931
1932         if (hctx->flags & BLK_MQ_F_BLOCKING)
1933                 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1934
1935         blk_mq_remove_cpuhp(hctx);
1936         blk_free_flush_queue(hctx->fq);
1937         sbitmap_free(&hctx->ctx_map);
1938 }
1939
1940 static void blk_mq_exit_hw_queues(struct request_queue *q,
1941                 struct blk_mq_tag_set *set, int nr_queue)
1942 {
1943         struct blk_mq_hw_ctx *hctx;
1944         unsigned int i;
1945
1946         queue_for_each_hw_ctx(q, hctx, i) {
1947                 if (i == nr_queue)
1948                         break;
1949                 blk_mq_exit_hctx(q, set, hctx, i);
1950         }
1951 }
1952
1953 static int blk_mq_init_hctx(struct request_queue *q,
1954                 struct blk_mq_tag_set *set,
1955                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1956 {
1957         int node;
1958         unsigned flush_start_tag = set->queue_depth;
1959
1960         node = hctx->numa_node;
1961         if (node == NUMA_NO_NODE)
1962                 node = hctx->numa_node = set->numa_node;
1963
1964         INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1965         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1966         spin_lock_init(&hctx->lock);
1967         INIT_LIST_HEAD(&hctx->dispatch);
1968         hctx->queue = q;
1969         hctx->queue_num = hctx_idx;
1970         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1971
1972         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1973
1974         hctx->tags = set->tags[hctx_idx];
1975
1976         /*
1977          * Allocate space for all possible cpus to avoid allocation at
1978          * runtime
1979          */
1980         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1981                                         GFP_KERNEL, node);
1982         if (!hctx->ctxs)
1983                 goto unregister_cpu_notifier;
1984
1985         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1986                               node))
1987                 goto free_ctxs;
1988
1989         hctx->nr_ctx = 0;
1990
1991         if (set->ops->init_hctx &&
1992             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1993                 goto free_bitmap;
1994
1995         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1996                 goto exit_hctx;
1997
1998         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1999         if (!hctx->fq)
2000                 goto sched_exit_hctx;
2001
2002         if (set->ops->init_request &&
2003             set->ops->init_request(set->driver_data,
2004                                    hctx->fq->flush_rq, hctx_idx,
2005                                    flush_start_tag + hctx_idx, node))
2006                 goto free_fq;
2007
2008         if (hctx->flags & BLK_MQ_F_BLOCKING)
2009                 init_srcu_struct(&hctx->queue_rq_srcu);
2010
2011         return 0;
2012
2013  free_fq:
2014         kfree(hctx->fq);
2015  sched_exit_hctx:
2016         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2017  exit_hctx:
2018         if (set->ops->exit_hctx)
2019                 set->ops->exit_hctx(hctx, hctx_idx);
2020  free_bitmap:
2021         sbitmap_free(&hctx->ctx_map);
2022  free_ctxs:
2023         kfree(hctx->ctxs);
2024  unregister_cpu_notifier:
2025         blk_mq_remove_cpuhp(hctx);
2026         return -1;
2027 }
2028
2029 static void blk_mq_init_cpu_queues(struct request_queue *q,
2030                                    unsigned int nr_hw_queues)
2031 {
2032         unsigned int i;
2033
2034         for_each_possible_cpu(i) {
2035                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2036                 struct blk_mq_hw_ctx *hctx;
2037
2038                 __ctx->cpu = i;
2039                 spin_lock_init(&__ctx->lock);
2040                 INIT_LIST_HEAD(&__ctx->rq_list);
2041                 __ctx->queue = q;
2042                 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
2043                 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
2044
2045                 /* If the cpu isn't online, the cpu is mapped to first hctx */
2046                 if (!cpu_online(i))
2047                         continue;
2048
2049                 hctx = blk_mq_map_queue(q, i);
2050
2051                 /*
2052                  * Set local node, IFF we have more than one hw queue. If
2053                  * not, we remain on the home node of the device
2054                  */
2055                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2056                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2057         }
2058 }
2059
2060 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2061 {
2062         int ret = 0;
2063
2064         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2065                                         set->queue_depth, set->reserved_tags);
2066         if (!set->tags[hctx_idx])
2067                 return false;
2068
2069         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2070                                 set->queue_depth);
2071         if (!ret)
2072                 return true;
2073
2074         blk_mq_free_rq_map(set->tags[hctx_idx]);
2075         set->tags[hctx_idx] = NULL;
2076         return false;
2077 }
2078
2079 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2080                                          unsigned int hctx_idx)
2081 {
2082         if (set->tags[hctx_idx]) {
2083                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2084                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2085                 set->tags[hctx_idx] = NULL;
2086         }
2087 }
2088
2089 static void blk_mq_map_swqueue(struct request_queue *q,
2090                                const struct cpumask *online_mask)
2091 {
2092         unsigned int i, hctx_idx;
2093         struct blk_mq_hw_ctx *hctx;
2094         struct blk_mq_ctx *ctx;
2095         struct blk_mq_tag_set *set = q->tag_set;
2096
2097         /*
2098          * Avoid others reading imcomplete hctx->cpumask through sysfs
2099          */
2100         mutex_lock(&q->sysfs_lock);
2101
2102         queue_for_each_hw_ctx(q, hctx, i) {
2103                 cpumask_clear(hctx->cpumask);
2104                 hctx->nr_ctx = 0;
2105         }
2106
2107         /*
2108          * Map software to hardware queues
2109          */
2110         for_each_possible_cpu(i) {
2111                 /* If the cpu isn't online, the cpu is mapped to first hctx */
2112                 if (!cpumask_test_cpu(i, online_mask))
2113                         continue;
2114
2115                 hctx_idx = q->mq_map[i];
2116                 /* unmapped hw queue can be remapped after CPU topo changed */
2117                 if (!set->tags[hctx_idx] &&
2118                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2119                         /*
2120                          * If tags initialization fail for some hctx,
2121                          * that hctx won't be brought online.  In this
2122                          * case, remap the current ctx to hctx[0] which
2123                          * is guaranteed to always have tags allocated
2124                          */
2125                         q->mq_map[i] = 0;
2126                 }
2127
2128                 ctx = per_cpu_ptr(q->queue_ctx, i);
2129                 hctx = blk_mq_map_queue(q, i);
2130
2131                 cpumask_set_cpu(i, hctx->cpumask);
2132                 ctx->index_hw = hctx->nr_ctx;
2133                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2134         }
2135
2136         mutex_unlock(&q->sysfs_lock);
2137
2138         queue_for_each_hw_ctx(q, hctx, i) {
2139                 /*
2140                  * If no software queues are mapped to this hardware queue,
2141                  * disable it and free the request entries.
2142                  */
2143                 if (!hctx->nr_ctx) {
2144                         /* Never unmap queue 0.  We need it as a
2145                          * fallback in case of a new remap fails
2146                          * allocation
2147                          */
2148                         if (i && set->tags[i])
2149                                 blk_mq_free_map_and_requests(set, i);
2150
2151                         hctx->tags = NULL;
2152                         continue;
2153                 }
2154
2155                 hctx->tags = set->tags[i];
2156                 WARN_ON(!hctx->tags);
2157
2158                 /*
2159                  * Set the map size to the number of mapped software queues.
2160                  * This is more accurate and more efficient than looping
2161                  * over all possibly mapped software queues.
2162                  */
2163                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2164
2165                 /*
2166                  * Initialize batch roundrobin counts
2167                  */
2168                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2169                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2170         }
2171 }
2172
2173 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2174 {
2175         struct blk_mq_hw_ctx *hctx;
2176         int i;
2177
2178         queue_for_each_hw_ctx(q, hctx, i) {
2179                 if (shared)
2180                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2181                 else
2182                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2183         }
2184 }
2185
2186 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2187 {
2188         struct request_queue *q;
2189
2190         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2191                 blk_mq_freeze_queue(q);
2192                 queue_set_hctx_shared(q, shared);
2193                 blk_mq_unfreeze_queue(q);
2194         }
2195 }
2196
2197 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2198 {
2199         struct blk_mq_tag_set *set = q->tag_set;
2200
2201         mutex_lock(&set->tag_list_lock);
2202         list_del_init(&q->tag_set_list);
2203         if (list_is_singular(&set->tag_list)) {
2204                 /* just transitioned to unshared */
2205                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2206                 /* update existing queue */
2207                 blk_mq_update_tag_set_depth(set, false);
2208         }
2209         mutex_unlock(&set->tag_list_lock);
2210 }
2211
2212 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2213                                      struct request_queue *q)
2214 {
2215         q->tag_set = set;
2216
2217         mutex_lock(&set->tag_list_lock);
2218
2219         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2220         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2221                 set->flags |= BLK_MQ_F_TAG_SHARED;
2222                 /* update existing queue */
2223                 blk_mq_update_tag_set_depth(set, true);
2224         }
2225         if (set->flags & BLK_MQ_F_TAG_SHARED)
2226                 queue_set_hctx_shared(q, true);
2227         list_add_tail(&q->tag_set_list, &set->tag_list);
2228
2229         mutex_unlock(&set->tag_list_lock);
2230 }
2231
2232 /*
2233  * It is the actual release handler for mq, but we do it from
2234  * request queue's release handler for avoiding use-after-free
2235  * and headache because q->mq_kobj shouldn't have been introduced,
2236  * but we can't group ctx/kctx kobj without it.
2237  */
2238 void blk_mq_release(struct request_queue *q)
2239 {
2240         struct blk_mq_hw_ctx *hctx;
2241         unsigned int i;
2242
2243         /* hctx kobj stays in hctx */
2244         queue_for_each_hw_ctx(q, hctx, i) {
2245                 if (!hctx)
2246                         continue;
2247                 kobject_put(&hctx->kobj);
2248         }
2249
2250         q->mq_map = NULL;
2251
2252         kfree(q->queue_hw_ctx);
2253
2254         /*
2255          * release .mq_kobj and sw queue's kobject now because
2256          * both share lifetime with request queue.
2257          */
2258         blk_mq_sysfs_deinit(q);
2259
2260         free_percpu(q->queue_ctx);
2261 }
2262
2263 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2264 {
2265         struct request_queue *uninit_q, *q;
2266
2267         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2268         if (!uninit_q)
2269                 return ERR_PTR(-ENOMEM);
2270
2271         q = blk_mq_init_allocated_queue(set, uninit_q);
2272         if (IS_ERR(q))
2273                 blk_cleanup_queue(uninit_q);
2274
2275         return q;
2276 }
2277 EXPORT_SYMBOL(blk_mq_init_queue);
2278
2279 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2280                                                 struct request_queue *q)
2281 {
2282         int i, j;
2283         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2284
2285         blk_mq_sysfs_unregister(q);
2286         for (i = 0; i < set->nr_hw_queues; i++) {
2287                 int node;
2288
2289                 if (hctxs[i])
2290                         continue;
2291
2292                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2293                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2294                                         GFP_KERNEL, node);
2295                 if (!hctxs[i])
2296                         break;
2297
2298                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2299                                                 node)) {
2300                         kfree(hctxs[i]);
2301                         hctxs[i] = NULL;
2302                         break;
2303                 }
2304
2305                 atomic_set(&hctxs[i]->nr_active, 0);
2306                 hctxs[i]->numa_node = node;
2307                 hctxs[i]->queue_num = i;
2308
2309                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2310                         free_cpumask_var(hctxs[i]->cpumask);
2311                         kfree(hctxs[i]);
2312                         hctxs[i] = NULL;
2313                         break;
2314                 }
2315                 blk_mq_hctx_kobj_init(hctxs[i]);
2316         }
2317         for (j = i; j < q->nr_hw_queues; j++) {
2318                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2319
2320                 if (hctx) {
2321                         if (hctx->tags)
2322                                 blk_mq_free_map_and_requests(set, j);
2323                         blk_mq_exit_hctx(q, set, hctx, j);
2324                         kobject_put(&hctx->kobj);
2325                         hctxs[j] = NULL;
2326
2327                 }
2328         }
2329         q->nr_hw_queues = i;
2330         blk_mq_sysfs_register(q);
2331 }
2332
2333 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2334                                                   struct request_queue *q)
2335 {
2336         /* mark the queue as mq asap */
2337         q->mq_ops = set->ops;
2338
2339         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2340         if (!q->queue_ctx)
2341                 goto err_exit;
2342
2343         /* init q->mq_kobj and sw queues' kobjects */
2344         blk_mq_sysfs_init(q);
2345
2346         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2347                                                 GFP_KERNEL, set->numa_node);
2348         if (!q->queue_hw_ctx)
2349                 goto err_percpu;
2350
2351         q->mq_map = set->mq_map;
2352
2353         blk_mq_realloc_hw_ctxs(set, q);
2354         if (!q->nr_hw_queues)
2355                 goto err_hctxs;
2356
2357         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2358         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2359
2360         q->nr_queues = nr_cpu_ids;
2361
2362         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2363
2364         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2365                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2366
2367         q->sg_reserved_size = INT_MAX;
2368
2369         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2370         INIT_LIST_HEAD(&q->requeue_list);
2371         spin_lock_init(&q->requeue_lock);
2372
2373         if (q->nr_hw_queues > 1)
2374                 blk_queue_make_request(q, blk_mq_make_request);
2375         else
2376                 blk_queue_make_request(q, blk_sq_make_request);
2377
2378         /*
2379          * Do this after blk_queue_make_request() overrides it...
2380          */
2381         q->nr_requests = set->queue_depth;
2382
2383         /*
2384          * Default to classic polling
2385          */
2386         q->poll_nsec = -1;
2387
2388         if (set->ops->complete)
2389                 blk_queue_softirq_done(q, set->ops->complete);
2390
2391         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2392
2393         get_online_cpus();
2394         mutex_lock(&all_q_mutex);
2395
2396         list_add_tail(&q->all_q_node, &all_q_list);
2397         blk_mq_add_queue_tag_set(set, q);
2398         blk_mq_map_swqueue(q, cpu_online_mask);
2399
2400         mutex_unlock(&all_q_mutex);
2401         put_online_cpus();
2402
2403         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2404                 int ret;
2405
2406                 ret = blk_mq_sched_init(q);
2407                 if (ret)
2408                         return ERR_PTR(ret);
2409         }
2410
2411         return q;
2412
2413 err_hctxs:
2414         kfree(q->queue_hw_ctx);
2415 err_percpu:
2416         free_percpu(q->queue_ctx);
2417 err_exit:
2418         q->mq_ops = NULL;
2419         return ERR_PTR(-ENOMEM);
2420 }
2421 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2422
2423 void blk_mq_free_queue(struct request_queue *q)
2424 {
2425         struct blk_mq_tag_set   *set = q->tag_set;
2426
2427         mutex_lock(&all_q_mutex);
2428         list_del_init(&q->all_q_node);
2429         mutex_unlock(&all_q_mutex);
2430
2431         wbt_exit(q);
2432
2433         blk_mq_del_queue_tag_set(q);
2434
2435         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2436 }
2437
2438 /* Basically redo blk_mq_init_queue with queue frozen */
2439 static void blk_mq_queue_reinit(struct request_queue *q,
2440                                 const struct cpumask *online_mask)
2441 {
2442         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2443
2444         blk_mq_sysfs_unregister(q);
2445
2446         /*
2447          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2448          * we should change hctx numa_node according to new topology (this
2449          * involves free and re-allocate memory, worthy doing?)
2450          */
2451
2452         blk_mq_map_swqueue(q, online_mask);
2453
2454         blk_mq_sysfs_register(q);
2455 }
2456
2457 /*
2458  * New online cpumask which is going to be set in this hotplug event.
2459  * Declare this cpumasks as global as cpu-hotplug operation is invoked
2460  * one-by-one and dynamically allocating this could result in a failure.
2461  */
2462 static struct cpumask cpuhp_online_new;
2463
2464 static void blk_mq_queue_reinit_work(void)
2465 {
2466         struct request_queue *q;
2467
2468         mutex_lock(&all_q_mutex);
2469         /*
2470          * We need to freeze and reinit all existing queues.  Freezing
2471          * involves synchronous wait for an RCU grace period and doing it
2472          * one by one may take a long time.  Start freezing all queues in
2473          * one swoop and then wait for the completions so that freezing can
2474          * take place in parallel.
2475          */
2476         list_for_each_entry(q, &all_q_list, all_q_node)
2477                 blk_mq_freeze_queue_start(q);
2478         list_for_each_entry(q, &all_q_list, all_q_node)
2479                 blk_mq_freeze_queue_wait(q);
2480
2481         list_for_each_entry(q, &all_q_list, all_q_node)
2482                 blk_mq_queue_reinit(q, &cpuhp_online_new);
2483
2484         list_for_each_entry(q, &all_q_list, all_q_node)
2485                 blk_mq_unfreeze_queue(q);
2486
2487         mutex_unlock(&all_q_mutex);
2488 }
2489
2490 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2491 {
2492         cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2493         blk_mq_queue_reinit_work();
2494         return 0;
2495 }
2496
2497 /*
2498  * Before hotadded cpu starts handling requests, new mappings must be
2499  * established.  Otherwise, these requests in hw queue might never be
2500  * dispatched.
2501  *
2502  * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2503  * for CPU0, and ctx1 for CPU1).
2504  *
2505  * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2506  * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2507  *
2508  * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2509  * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2510  * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2511  * ignored.
2512  */
2513 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2514 {
2515         cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2516         cpumask_set_cpu(cpu, &cpuhp_online_new);
2517         blk_mq_queue_reinit_work();
2518         return 0;
2519 }
2520
2521 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2522 {
2523         int i;
2524
2525         for (i = 0; i < set->nr_hw_queues; i++)
2526                 if (!__blk_mq_alloc_rq_map(set, i))
2527                         goto out_unwind;
2528
2529         return 0;
2530
2531 out_unwind:
2532         while (--i >= 0)
2533                 blk_mq_free_rq_map(set->tags[i]);
2534
2535         return -ENOMEM;
2536 }
2537
2538 /*
2539  * Allocate the request maps associated with this tag_set. Note that this
2540  * may reduce the depth asked for, if memory is tight. set->queue_depth
2541  * will be updated to reflect the allocated depth.
2542  */
2543 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2544 {
2545         unsigned int depth;
2546         int err;
2547
2548         depth = set->queue_depth;
2549         do {
2550                 err = __blk_mq_alloc_rq_maps(set);
2551                 if (!err)
2552                         break;
2553
2554                 set->queue_depth >>= 1;
2555                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2556                         err = -ENOMEM;
2557                         break;
2558                 }
2559         } while (set->queue_depth);
2560
2561         if (!set->queue_depth || err) {
2562                 pr_err("blk-mq: failed to allocate request map\n");
2563                 return -ENOMEM;
2564         }
2565
2566         if (depth != set->queue_depth)
2567                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2568                                                 depth, set->queue_depth);
2569
2570         return 0;
2571 }
2572
2573 /*
2574  * Alloc a tag set to be associated with one or more request queues.
2575  * May fail with EINVAL for various error conditions. May adjust the
2576  * requested depth down, if if it too large. In that case, the set
2577  * value will be stored in set->queue_depth.
2578  */
2579 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2580 {
2581         int ret;
2582
2583         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2584
2585         if (!set->nr_hw_queues)
2586                 return -EINVAL;
2587         if (!set->queue_depth)
2588                 return -EINVAL;
2589         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2590                 return -EINVAL;
2591
2592         if (!set->ops->queue_rq)
2593                 return -EINVAL;
2594
2595         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2596                 pr_info("blk-mq: reduced tag depth to %u\n",
2597                         BLK_MQ_MAX_DEPTH);
2598                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2599         }
2600
2601         /*
2602          * If a crashdump is active, then we are potentially in a very
2603          * memory constrained environment. Limit us to 1 queue and
2604          * 64 tags to prevent using too much memory.
2605          */
2606         if (is_kdump_kernel()) {
2607                 set->nr_hw_queues = 1;
2608                 set->queue_depth = min(64U, set->queue_depth);
2609         }
2610         /*
2611          * There is no use for more h/w queues than cpus.
2612          */
2613         if (set->nr_hw_queues > nr_cpu_ids)
2614                 set->nr_hw_queues = nr_cpu_ids;
2615
2616         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2617                                  GFP_KERNEL, set->numa_node);
2618         if (!set->tags)
2619                 return -ENOMEM;
2620
2621         ret = -ENOMEM;
2622         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2623                         GFP_KERNEL, set->numa_node);
2624         if (!set->mq_map)
2625                 goto out_free_tags;
2626
2627         if (set->ops->map_queues)
2628                 ret = set->ops->map_queues(set);
2629         else
2630                 ret = blk_mq_map_queues(set);
2631         if (ret)
2632                 goto out_free_mq_map;
2633
2634         ret = blk_mq_alloc_rq_maps(set);
2635         if (ret)
2636                 goto out_free_mq_map;
2637
2638         mutex_init(&set->tag_list_lock);
2639         INIT_LIST_HEAD(&set->tag_list);
2640
2641         return 0;
2642
2643 out_free_mq_map:
2644         kfree(set->mq_map);
2645         set->mq_map = NULL;
2646 out_free_tags:
2647         kfree(set->tags);
2648         set->tags = NULL;
2649         return ret;
2650 }
2651 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2652
2653 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2654 {
2655         int i;
2656
2657         for (i = 0; i < nr_cpu_ids; i++)
2658                 blk_mq_free_map_and_requests(set, i);
2659
2660         kfree(set->mq_map);
2661         set->mq_map = NULL;
2662
2663         kfree(set->tags);
2664         set->tags = NULL;
2665 }
2666 EXPORT_SYMBOL(blk_mq_free_tag_set);
2667
2668 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2669 {
2670         struct blk_mq_tag_set *set = q->tag_set;
2671         struct blk_mq_hw_ctx *hctx;
2672         int i, ret;
2673
2674         if (!set)
2675                 return -EINVAL;
2676
2677         blk_mq_freeze_queue(q);
2678         blk_mq_quiesce_queue(q);
2679
2680         ret = 0;
2681         queue_for_each_hw_ctx(q, hctx, i) {
2682                 if (!hctx->tags)
2683                         continue;
2684                 /*
2685                  * If we're using an MQ scheduler, just update the scheduler
2686                  * queue depth. This is similar to what the old code would do.
2687                  */
2688                 if (!hctx->sched_tags) {
2689                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2690                                                         min(nr, set->queue_depth),
2691                                                         false);
2692                 } else {
2693                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2694                                                         nr, true);
2695                 }
2696                 if (ret)
2697                         break;
2698         }
2699
2700         if (!ret)
2701                 q->nr_requests = nr;
2702
2703         blk_mq_unfreeze_queue(q);
2704         blk_mq_start_stopped_hw_queues(q, true);
2705
2706         return ret;
2707 }
2708
2709 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2710 {
2711         struct request_queue *q;
2712
2713         if (nr_hw_queues > nr_cpu_ids)
2714                 nr_hw_queues = nr_cpu_ids;
2715         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2716                 return;
2717
2718         list_for_each_entry(q, &set->tag_list, tag_set_list)
2719                 blk_mq_freeze_queue(q);
2720
2721         set->nr_hw_queues = nr_hw_queues;
2722         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2723                 blk_mq_realloc_hw_ctxs(set, q);
2724
2725                 /*
2726                  * Manually set the make_request_fn as blk_queue_make_request
2727                  * resets a lot of the queue settings.
2728                  */
2729                 if (q->nr_hw_queues > 1)
2730                         q->make_request_fn = blk_mq_make_request;
2731                 else
2732                         q->make_request_fn = blk_sq_make_request;
2733
2734                 blk_mq_queue_reinit(q, cpu_online_mask);
2735         }
2736
2737         list_for_each_entry(q, &set->tag_list, tag_set_list)
2738                 blk_mq_unfreeze_queue(q);
2739 }
2740 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2741
2742 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2743                                        struct blk_mq_hw_ctx *hctx,
2744                                        struct request *rq)
2745 {
2746         struct blk_rq_stat stat[2];
2747         unsigned long ret = 0;
2748
2749         /*
2750          * If stats collection isn't on, don't sleep but turn it on for
2751          * future users
2752          */
2753         if (!blk_stat_enable(q))
2754                 return 0;
2755
2756         /*
2757          * We don't have to do this once per IO, should optimize this
2758          * to just use the current window of stats until it changes
2759          */
2760         memset(&stat, 0, sizeof(stat));
2761         blk_hctx_stat_get(hctx, stat);
2762
2763         /*
2764          * As an optimistic guess, use half of the mean service time
2765          * for this type of request. We can (and should) make this smarter.
2766          * For instance, if the completion latencies are tight, we can
2767          * get closer than just half the mean. This is especially
2768          * important on devices where the completion latencies are longer
2769          * than ~10 usec.
2770          */
2771         if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2772                 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2773         else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2774                 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2775
2776         return ret;
2777 }
2778
2779 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2780                                      struct blk_mq_hw_ctx *hctx,
2781                                      struct request *rq)
2782 {
2783         struct hrtimer_sleeper hs;
2784         enum hrtimer_mode mode;
2785         unsigned int nsecs;
2786         ktime_t kt;
2787
2788         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2789                 return false;
2790
2791         /*
2792          * poll_nsec can be:
2793          *
2794          * -1:  don't ever hybrid sleep
2795          *  0:  use half of prev avg
2796          * >0:  use this specific value
2797          */
2798         if (q->poll_nsec == -1)
2799                 return false;
2800         else if (q->poll_nsec > 0)
2801                 nsecs = q->poll_nsec;
2802         else
2803                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2804
2805         if (!nsecs)
2806                 return false;
2807
2808         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2809
2810         /*
2811          * This will be replaced with the stats tracking code, using
2812          * 'avg_completion_time / 2' as the pre-sleep target.
2813          */
2814         kt = nsecs;
2815
2816         mode = HRTIMER_MODE_REL;
2817         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2818         hrtimer_set_expires(&hs.timer, kt);
2819
2820         hrtimer_init_sleeper(&hs, current);
2821         do {
2822                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2823                         break;
2824                 set_current_state(TASK_UNINTERRUPTIBLE);
2825                 hrtimer_start_expires(&hs.timer, mode);
2826                 if (hs.task)
2827                         io_schedule();
2828                 hrtimer_cancel(&hs.timer);
2829                 mode = HRTIMER_MODE_ABS;
2830         } while (hs.task && !signal_pending(current));
2831
2832         __set_current_state(TASK_RUNNING);
2833         destroy_hrtimer_on_stack(&hs.timer);
2834         return true;
2835 }
2836
2837 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2838 {
2839         struct request_queue *q = hctx->queue;
2840         long state;
2841
2842         /*
2843          * If we sleep, have the caller restart the poll loop to reset
2844          * the state. Like for the other success return cases, the
2845          * caller is responsible for checking if the IO completed. If
2846          * the IO isn't complete, we'll get called again and will go
2847          * straight to the busy poll loop.
2848          */
2849         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2850                 return true;
2851
2852         hctx->poll_considered++;
2853
2854         state = current->state;
2855         while (!need_resched()) {
2856                 int ret;
2857
2858                 hctx->poll_invoked++;
2859
2860                 ret = q->mq_ops->poll(hctx, rq->tag);
2861                 if (ret > 0) {
2862                         hctx->poll_success++;
2863                         set_current_state(TASK_RUNNING);
2864                         return true;
2865                 }
2866
2867                 if (signal_pending_state(state, current))
2868                         set_current_state(TASK_RUNNING);
2869
2870                 if (current->state == TASK_RUNNING)
2871                         return true;
2872                 if (ret < 0)
2873                         break;
2874                 cpu_relax();
2875         }
2876
2877         return false;
2878 }
2879
2880 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2881 {
2882         struct blk_mq_hw_ctx *hctx;
2883         struct blk_plug *plug;
2884         struct request *rq;
2885
2886         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2887             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2888                 return false;
2889
2890         plug = current->plug;
2891         if (plug)
2892                 blk_flush_plug_list(plug, false);
2893
2894         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2895         if (!blk_qc_t_is_internal(cookie))
2896                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2897         else
2898                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2899
2900         return __blk_mq_poll(hctx, rq);
2901 }
2902 EXPORT_SYMBOL_GPL(blk_mq_poll);
2903
2904 void blk_mq_disable_hotplug(void)
2905 {
2906         mutex_lock(&all_q_mutex);
2907 }
2908
2909 void blk_mq_enable_hotplug(void)
2910 {
2911         mutex_unlock(&all_q_mutex);
2912 }
2913
2914 static int __init blk_mq_init(void)
2915 {
2916         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2917                                 blk_mq_hctx_notify_dead);
2918
2919         cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2920                                   blk_mq_queue_reinit_prepare,
2921                                   blk_mq_queue_reinit_dead);
2922         return 0;
2923 }
2924 subsys_initcall(blk_mq_init);