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/delay.h>
  24 #include <linux/crash_dump.h>
  25 #include <linux/prefetch.h>
  26
  27 #include <trace/events/block.h>
  28
  29 #include <linux/blk-mq.h>
  30 #include "blk.h"
  31 #include "blk-mq.h"
  32 #include "blk-mq-tag.h"
  33 #include "blk-stat.h"
  34 #include "blk-wbt.h"
  35 #include "blk-mq-sched.h"
  36
  37 static DEFINE_MUTEX(all_q_mutex);
  38 static LIST_HEAD(all_q_list);
  39
  40 /*
  41  * Check if any of the ctx's have pending work in this hardware queue
  42  */
  43 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  44 {
  45         return sbitmap_any_bit_set(&hctx->ctx_map) ||
  46                         !list_empty_careful(&hctx->dispatch) ||
  47                         blk_mq_sched_has_work(hctx);
  48 }
  49
  50 /*
  51  * Mark this ctx as having pending work in this hardware queue
  52  */
  53 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  54                                      struct blk_mq_ctx *ctx)
  55 {
  56         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  57                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  58 }
  59
  60 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  61                                       struct blk_mq_ctx *ctx)
  62 {
  63         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  64 }
  65
  66 void blk_mq_freeze_queue_start(struct request_queue *q)
  67 {
  68         int freeze_depth;
  69
  70         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
  71         if (freeze_depth == 1) {
  72                 percpu_ref_kill(&q->q_usage_counter);
  73                 blk_mq_run_hw_queues(q, false);
  74         }
  75 }
  76 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
  77
  78 static void blk_mq_freeze_queue_wait(struct request_queue *q)
  79 {
  80         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
  81 }
  82
  83 /*
  84  * Guarantee no request is in use, so we can change any data structure of
  85  * the queue afterward.
  86  */
  87 void blk_freeze_queue(struct request_queue *q)
  88 {
  89         /*
  90          * In the !blk_mq case we are only calling this to kill the
  91          * q_usage_counter, otherwise this increases the freeze depth
  92          * and waits for it to return to zero.  For this reason there is
  93          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
  94          * exported to drivers as the only user for unfreeze is blk_mq.
  95          */
  96         blk_mq_freeze_queue_start(q);
  97         blk_mq_freeze_queue_wait(q);
  98 }
  99
 100 void blk_mq_freeze_queue(struct request_queue *q)
 101 {
 102         /*
 103          * ...just an alias to keep freeze and unfreeze actions balanced
 104          * in the blk_mq_* namespace
 105          */
 106         blk_freeze_queue(q);
 107 }
 108 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 109
 110 void blk_mq_unfreeze_queue(struct request_queue *q)
 111 {
 112         int freeze_depth;
 113
 114         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 115         WARN_ON_ONCE(freeze_depth < 0);
 116         if (!freeze_depth) {
 117                 percpu_ref_reinit(&q->q_usage_counter);
 118                 wake_up_all(&q->mq_freeze_wq);
 119         }
 120 }
 121 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 122
 123 /**
 124  * blk_mq_quiesce_queue() - wait until all ongoing queue_rq calls have finished
 125  * @q: request queue.
 126  *
 127  * Note: this function does not prevent that the struct request end_io()
 128  * callback function is invoked. Additionally, it is not prevented that
 129  * new queue_rq() calls occur unless the queue has been stopped first.
 130  */
 131 void blk_mq_quiesce_queue(struct request_queue *q)
 132 {
 133         struct blk_mq_hw_ctx *hctx;
 134         unsigned int i;
 135         bool rcu = false;
 136
 137         blk_mq_stop_hw_queues(q);
 138
 139         queue_for_each_hw_ctx(q, hctx, i) {
 140                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 141                         synchronize_srcu(&hctx->queue_rq_srcu);
 142                 else
 143                         rcu = true;
 144         }
 145         if (rcu)
 146                 synchronize_rcu();
 147 }
 148 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 149
 150 void blk_mq_wake_waiters(struct request_queue *q)
 151 {
 152         struct blk_mq_hw_ctx *hctx;
 153         unsigned int i;
 154
 155         queue_for_each_hw_ctx(q, hctx, i)
 156                 if (blk_mq_hw_queue_mapped(hctx))
 157                         blk_mq_tag_wakeup_all(hctx->tags, true);
 158
 159         /*
 160          * If we are called because the queue has now been marked as
 161          * dying, we need to ensure that processes currently waiting on
 162          * the queue are notified as well.
 163          */
 164         wake_up_all(&q->mq_freeze_wq);
 165 }
 166
 167 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 168 {
 169         return blk_mq_has_free_tags(hctx->tags);
 170 }
 171 EXPORT_SYMBOL(blk_mq_can_queue);
 172
 173 void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 174                         struct request *rq, unsigned int op)
 175 {
 176         INIT_LIST_HEAD(&rq->queuelist);
 177         /* csd/requeue_work/fifo_time is initialized before use */
 178         rq->q = q;
 179         rq->mq_ctx = ctx;
 180         rq->cmd_flags = op;
 181         if (blk_queue_io_stat(q))
 182                 rq->rq_flags |= RQF_IO_STAT;
 183         /* do not touch atomic flags, it needs atomic ops against the timer */
 184         rq->cpu = -1;
 185         INIT_HLIST_NODE(&rq->hash);
 186         RB_CLEAR_NODE(&rq->rb_node);
 187         rq->rq_disk = NULL;
 188         rq->part = NULL;
 189         rq->start_time = jiffies;
 190 #ifdef CONFIG_BLK_CGROUP
 191         rq->rl = NULL;
 192         set_start_time_ns(rq);
 193         rq->io_start_time_ns = 0;
 194 #endif
 195         rq->nr_phys_segments = 0;
 196 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 197         rq->nr_integrity_segments = 0;
 198 #endif
 199         rq->special = NULL;
 200         /* tag was already set */
 201         rq->errors = 0;
 202         rq->extra_len = 0;
 203
 204         INIT_LIST_HEAD(&rq->timeout_list);
 205         rq->timeout = 0;
 206
 207         rq->end_io = NULL;
 208         rq->end_io_data = NULL;
 209         rq->next_rq = NULL;
 210
 211         ctx->rq_dispatched[op_is_sync(op)]++;
 212 }
 213 EXPORT_SYMBOL_GPL(blk_mq_rq_ctx_init);
 214
 215 struct request *__blk_mq_alloc_request(struct blk_mq_alloc_data *data,
 216                                        unsigned int op)
 217 {
 218         struct request *rq;
 219         unsigned int tag;
 220
 221         tag = blk_mq_get_tag(data);
 222         if (tag != BLK_MQ_TAG_FAIL) {
 223                 struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 224
 225                 rq = tags->static_rqs[tag];
 226
 227                 if (data->flags & BLK_MQ_REQ_INTERNAL) {
 228                         rq->tag = -1;
 229                         rq->internal_tag = tag;
 230                 } else {
 231                         if (blk_mq_tag_busy(data->hctx)) {
 232                                 rq->rq_flags = RQF_MQ_INFLIGHT;
 233                                 atomic_inc(&data->hctx->nr_active);
 234                         }
 235                         rq->tag = tag;
 236                         rq->internal_tag = -1;
 237                 }
 238
 239                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op);
 240                 return rq;
 241         }
 242
 243         return NULL;
 244 }
 245 EXPORT_SYMBOL_GPL(__blk_mq_alloc_request);
 246
 247 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
 248                 unsigned int flags)
 249 {
 250         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 251         struct request *rq;
 252         int ret;
 253
 254         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 255         if (ret)
 256                 return ERR_PTR(ret);
 257
 258         rq = blk_mq_sched_get_request(q, NULL, rw, &alloc_data);
 259
 260         blk_mq_put_ctx(alloc_data.ctx);
 261         blk_queue_exit(q);
 262
 263         if (!rq)
 264                 return ERR_PTR(-EWOULDBLOCK);
 265
 266         rq->__data_len = 0;
 267         rq->__sector = (sector_t) -1;
 268         rq->bio = rq->biotail = NULL;
 269         return rq;
 270 }
 271 EXPORT_SYMBOL(blk_mq_alloc_request);
 272
 273 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
 274                 unsigned int flags, unsigned int hctx_idx)
 275 {
 276         struct blk_mq_hw_ctx *hctx;
 277         struct blk_mq_ctx *ctx;
 278         struct request *rq;
 279         struct blk_mq_alloc_data alloc_data;
 280         int ret;
 281
 282         /*
 283          * If the tag allocator sleeps we could get an allocation for a
 284          * different hardware context.  No need to complicate the low level
 285          * allocator for this for the rare use case of a command tied to
 286          * a specific queue.
 287          */
 288         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 289                 return ERR_PTR(-EINVAL);
 290
 291         if (hctx_idx >= q->nr_hw_queues)
 292                 return ERR_PTR(-EIO);
 293
 294         ret = blk_queue_enter(q, true);
 295         if (ret)
 296                 return ERR_PTR(ret);
 297
 298         /*
 299          * Check if the hardware context is actually mapped to anything.
 300          * If not tell the caller that it should skip this queue.
 301          */
 302         hctx = q->queue_hw_ctx[hctx_idx];
 303         if (!blk_mq_hw_queue_mapped(hctx)) {
 304                 ret = -EXDEV;
 305                 goto out_queue_exit;
 306         }
 307         ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
 308
 309         blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
 310         rq = __blk_mq_alloc_request(&alloc_data, rw);
 311         if (!rq) {
 312                 ret = -EWOULDBLOCK;
 313                 goto out_queue_exit;
 314         }
 315
 316         return rq;
 317
 318 out_queue_exit:
 319         blk_queue_exit(q);
 320         return ERR_PTR(ret);
 321 }
 322 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 323
 324 void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
 325                              struct request *rq)
 326 {
 327         const int sched_tag = rq->internal_tag;
 328         struct request_queue *q = rq->q;
 329
 330         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 331                 atomic_dec(&hctx->nr_active);
 332
 333         wbt_done(q->rq_wb, &rq->issue_stat);
 334         rq->rq_flags = 0;
 335
 336         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 337         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 338         if (rq->tag != -1)
 339                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 340         if (sched_tag != -1)
 341                 blk_mq_sched_completed_request(hctx, rq);
 342         blk_mq_sched_restart_queues(hctx);
 343         blk_queue_exit(q);
 344 }
 345
 346 static void blk_mq_finish_hctx_request(struct blk_mq_hw_ctx *hctx,
 347                                      struct request *rq)
 348 {
 349         struct blk_mq_ctx *ctx = rq->mq_ctx;
 350
 351         ctx->rq_completed[rq_is_sync(rq)]++;
 352         __blk_mq_finish_request(hctx, ctx, rq);
 353 }
 354
 355 void blk_mq_finish_request(struct request *rq)
 356 {
 357         blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
 358 }
 359
 360 void blk_mq_free_request(struct request *rq)
 361 {
 362         blk_mq_sched_put_request(rq);
 363 }
 364 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 365
 366 inline void __blk_mq_end_request(struct request *rq, int error)
 367 {
 368         blk_account_io_done(rq);
 369
 370         if (rq->end_io) {
 371                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 372                 rq->end_io(rq, error);
 373         } else {
 374                 if (unlikely(blk_bidi_rq(rq)))
 375                         blk_mq_free_request(rq->next_rq);
 376                 blk_mq_free_request(rq);
 377         }
 378 }
 379 EXPORT_SYMBOL(__blk_mq_end_request);
 380
 381 void blk_mq_end_request(struct request *rq, int error)
 382 {
 383         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 384                 BUG();
 385         __blk_mq_end_request(rq, error);
 386 }
 387 EXPORT_SYMBOL(blk_mq_end_request);
 388
 389 static void __blk_mq_complete_request_remote(void *data)
 390 {
 391         struct request *rq = data;
 392
 393         rq->q->softirq_done_fn(rq);
 394 }
 395
 396 static void blk_mq_ipi_complete_request(struct request *rq)
 397 {
 398         struct blk_mq_ctx *ctx = rq->mq_ctx;
 399         bool shared = false;
 400         int cpu;
 401
 402         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 403                 rq->q->softirq_done_fn(rq);
 404                 return;
 405         }
 406
 407         cpu = get_cpu();
 408         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 409                 shared = cpus_share_cache(cpu, ctx->cpu);
 410
 411         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 412                 rq->csd.func = __blk_mq_complete_request_remote;
 413                 rq->csd.info = rq;
 414                 rq->csd.flags = 0;
 415                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 416         } else {
 417                 rq->q->softirq_done_fn(rq);
 418         }
 419         put_cpu();
 420 }
 421
 422 static void blk_mq_stat_add(struct request *rq)
 423 {
 424         if (rq->rq_flags & RQF_STATS) {
 425                 /*
 426                  * We could rq->mq_ctx here, but there's less of a risk
 427                  * of races if we have the completion event add the stats
 428                  * to the local software queue.
 429                  */
 430                 struct blk_mq_ctx *ctx;
 431
 432                 ctx = __blk_mq_get_ctx(rq->q, raw_smp_processor_id());
 433                 blk_stat_add(&ctx->stat[rq_data_dir(rq)], rq);
 434         }
 435 }
 436
 437 static void __blk_mq_complete_request(struct request *rq)
 438 {
 439         struct request_queue *q = rq->q;
 440
 441         blk_mq_stat_add(rq);
 442
 443         if (!q->softirq_done_fn)
 444                 blk_mq_end_request(rq, rq->errors);
 445         else
 446                 blk_mq_ipi_complete_request(rq);
 447 }
 448
 449 /**
 450  * blk_mq_complete_request - end I/O on a request
 451  * @rq:         the request being processed
 452  *
 453  * Description:
 454  *      Ends all I/O on a request. It does not handle partial completions.
 455  *      The actual completion happens out-of-order, through a IPI handler.
 456  **/
 457 void blk_mq_complete_request(struct request *rq, int error)
 458 {
 459         struct request_queue *q = rq->q;
 460
 461         if (unlikely(blk_should_fake_timeout(q)))
 462                 return;
 463         if (!blk_mark_rq_complete(rq)) {
 464                 rq->errors = error;
 465                 __blk_mq_complete_request(rq);
 466         }
 467 }
 468 EXPORT_SYMBOL(blk_mq_complete_request);
 469
 470 int blk_mq_request_started(struct request *rq)
 471 {
 472         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 473 }
 474 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 475
 476 void blk_mq_start_request(struct request *rq)
 477 {
 478         struct request_queue *q = rq->q;
 479
 480         blk_mq_sched_started_request(rq);
 481
 482         trace_block_rq_issue(q, rq);
 483
 484         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 485                 blk_stat_set_issue_time(&rq->issue_stat);
 486                 rq->rq_flags |= RQF_STATS;
 487                 wbt_issue(q->rq_wb, &rq->issue_stat);
 488         }
 489
 490         blk_add_timer(rq);
 491
 492         /*
 493          * Ensure that ->deadline is visible before set the started
 494          * flag and clear the completed flag.
 495          */
 496         smp_mb__before_atomic();
 497
 498         /*
 499          * Mark us as started and clear complete. Complete might have been
 500          * set if requeue raced with timeout, which then marked it as
 501          * complete. So be sure to clear complete again when we start
 502          * the request, otherwise we'll ignore the completion event.
 503          */
 504         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 505                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 506         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 507                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 508
 509         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 510                 /*
 511                  * Make sure space for the drain appears.  We know we can do
 512                  * this because max_hw_segments has been adjusted to be one
 513                  * fewer than the device can handle.
 514                  */
 515                 rq->nr_phys_segments++;
 516         }
 517 }
 518 EXPORT_SYMBOL(blk_mq_start_request);
 519
 520 static void __blk_mq_requeue_request(struct request *rq)
 521 {
 522         struct request_queue *q = rq->q;
 523
 524         trace_block_rq_requeue(q, rq);
 525         wbt_requeue(q->rq_wb, &rq->issue_stat);
 526         blk_mq_sched_requeue_request(rq);
 527
 528         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 529                 if (q->dma_drain_size && blk_rq_bytes(rq))
 530                         rq->nr_phys_segments--;
 531         }
 532 }
 533
 534 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 535 {
 536         __blk_mq_requeue_request(rq);
 537
 538         BUG_ON(blk_queued_rq(rq));
 539         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 540 }
 541 EXPORT_SYMBOL(blk_mq_requeue_request);
 542
 543 static void blk_mq_requeue_work(struct work_struct *work)
 544 {
 545         struct request_queue *q =
 546                 container_of(work, struct request_queue, requeue_work.work);
 547         LIST_HEAD(rq_list);
 548         struct request *rq, *next;
 549         unsigned long flags;
 550
 551         spin_lock_irqsave(&q->requeue_lock, flags);
 552         list_splice_init(&q->requeue_list, &rq_list);
 553         spin_unlock_irqrestore(&q->requeue_lock, flags);
 554
 555         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 556                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 557                         continue;
 558
 559                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 560                 list_del_init(&rq->queuelist);
 561                 blk_mq_sched_insert_request(rq, true, false, false, true);
 562         }
 563
 564         while (!list_empty(&rq_list)) {
 565                 rq = list_entry(rq_list.next, struct request, queuelist);
 566                 list_del_init(&rq->queuelist);
 567                 blk_mq_sched_insert_request(rq, false, false, false, true);
 568         }
 569
 570         blk_mq_run_hw_queues(q, false);
 571 }
 572
 573 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 574                                 bool kick_requeue_list)
 575 {
 576         struct request_queue *q = rq->q;
 577         unsigned long flags;
 578
 579         /*
 580          * We abuse this flag that is otherwise used by the I/O scheduler to
 581          * request head insertation from the workqueue.
 582          */
 583         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 584
 585         spin_lock_irqsave(&q->requeue_lock, flags);
 586         if (at_head) {
 587                 rq->rq_flags |= RQF_SOFTBARRIER;
 588                 list_add(&rq->queuelist, &q->requeue_list);
 589         } else {
 590                 list_add_tail(&rq->queuelist, &q->requeue_list);
 591         }
 592         spin_unlock_irqrestore(&q->requeue_lock, flags);
 593
 594         if (kick_requeue_list)
 595                 blk_mq_kick_requeue_list(q);
 596 }
 597 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 598
 599 void blk_mq_kick_requeue_list(struct request_queue *q)
 600 {
 601         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 602 }
 603 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 604
 605 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 606                                     unsigned long msecs)
 607 {
 608         kblockd_schedule_delayed_work(&q->requeue_work,
 609                                       msecs_to_jiffies(msecs));
 610 }
 611 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 612
 613 void blk_mq_abort_requeue_list(struct request_queue *q)
 614 {
 615         unsigned long flags;
 616         LIST_HEAD(rq_list);
 617
 618         spin_lock_irqsave(&q->requeue_lock, flags);
 619         list_splice_init(&q->requeue_list, &rq_list);
 620         spin_unlock_irqrestore(&q->requeue_lock, flags);
 621
 622         while (!list_empty(&rq_list)) {
 623                 struct request *rq;
 624
 625                 rq = list_first_entry(&rq_list, struct request, queuelist);
 626                 list_del_init(&rq->queuelist);
 627                 rq->errors = -EIO;
 628                 blk_mq_end_request(rq, rq->errors);
 629         }
 630 }
 631 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 632
 633 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 634 {
 635         if (tag < tags->nr_tags) {
 636                 prefetch(tags->rqs[tag]);
 637                 return tags->rqs[tag];
 638         }
 639
 640         return NULL;
 641 }
 642 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 643
 644 struct blk_mq_timeout_data {
 645         unsigned long next;
 646         unsigned int next_set;
 647 };
 648
 649 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 650 {
 651         const struct blk_mq_ops *ops = req->q->mq_ops;
 652         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 653
 654         /*
 655          * We know that complete is set at this point. If STARTED isn't set
 656          * anymore, then the request isn't active and the "timeout" should
 657          * just be ignored. This can happen due to the bitflag ordering.
 658          * Timeout first checks if STARTED is set, and if it is, assumes
 659          * the request is active. But if we race with completion, then
 660          * we both flags will get cleared. So check here again, and ignore
 661          * a timeout event with a request that isn't active.
 662          */
 663         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 664                 return;
 665
 666         if (ops->timeout)
 667                 ret = ops->timeout(req, reserved);
 668
 669         switch (ret) {
 670         case BLK_EH_HANDLED:
 671                 __blk_mq_complete_request(req);
 672                 break;
 673         case BLK_EH_RESET_TIMER:
 674                 blk_add_timer(req);
 675                 blk_clear_rq_complete(req);
 676                 break;
 677         case BLK_EH_NOT_HANDLED:
 678                 break;
 679         default:
 680                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 681                 break;
 682         }
 683 }
 684
 685 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 686                 struct request *rq, void *priv, bool reserved)
 687 {
 688         struct blk_mq_timeout_data *data = priv;
 689
 690         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 691                 /*
 692                  * If a request wasn't started before the queue was
 693                  * marked dying, kill it here or it'll go unnoticed.
 694                  */
 695                 if (unlikely(blk_queue_dying(rq->q))) {
 696                         rq->errors = -EIO;
 697                         blk_mq_end_request(rq, rq->errors);
 698                 }
 699                 return;
 700         }
 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 done:
 850                 if (hctx)
 851                         *hctx = data.hctx;
 852                 return true;
 853         }
 854
 855         rq->tag = blk_mq_get_tag(&data);
 856         if (rq->tag >= 0) {
 857                 if (blk_mq_tag_busy(data.hctx)) {
 858                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 859                         atomic_inc(&data.hctx->nr_active);
 860                 }
 861                 data.hctx->tags->rqs[rq->tag] = rq;
 862                 goto done;
 863         }
 864
 865         return false;
 866 }
 867
 868 static void blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 869                                   struct request *rq)
 870 {
 871         if (rq->tag == -1 || rq->internal_tag == -1)
 872                 return;
 873
 874         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
 875         rq->tag = -1;
 876
 877         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 878                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
 879                 atomic_dec(&hctx->nr_active);
 880         }
 881 }
 882
 883 /*
 884  * If we fail getting a driver tag because all the driver tags are already
 885  * assigned and on the dispatch list, BUT the first entry does not have a
 886  * tag, then we could deadlock. For that case, move entries with assigned
 887  * driver tags to the front, leaving the set of tagged requests in the
 888  * same order, and the untagged set in the same order.
 889  */
 890 static bool reorder_tags_to_front(struct list_head *list)
 891 {
 892         struct request *rq, *tmp, *first = NULL;
 893
 894         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
 895                 if (rq == first)
 896                         break;
 897                 if (rq->tag != -1) {
 898                         list_move(&rq->queuelist, list);
 899                         if (!first)
 900                                 first = rq;
 901                 }
 902         }
 903
 904         return first != NULL;
 905 }
 906
 907 static int blk_mq_dispatch_wake(wait_queue_t *wait, unsigned mode, int flags,
 908                                 void *key)
 909 {
 910         struct blk_mq_hw_ctx *hctx;
 911
 912         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
 913
 914         list_del(&wait->task_list);
 915         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
 916         blk_mq_run_hw_queue(hctx, true);
 917         return 1;
 918 }
 919
 920 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
 921 {
 922         struct sbq_wait_state *ws;
 923
 924         /*
 925          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
 926          * The thread which wins the race to grab this bit adds the hardware
 927          * queue to the wait queue.
 928          */
 929         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
 930             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
 931                 return false;
 932
 933         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
 934         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
 935
 936         /*
 937          * As soon as this returns, it's no longer safe to fiddle with
 938          * hctx->dispatch_wait, since a completion can wake up the wait queue
 939          * and unlock the bit.
 940          */
 941         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
 942         return true;
 943 }
 944
 945 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 946 {
 947         struct request_queue *q = hctx->queue;
 948         struct request *rq;
 949         LIST_HEAD(driver_list);
 950         struct list_head *dptr;
 951         int queued, ret = BLK_MQ_RQ_QUEUE_OK;
 952
 953         /*
 954          * Start off with dptr being NULL, so we start the first request
 955          * immediately, even if we have more pending.
 956          */
 957         dptr = NULL;
 958
 959         /*
 960          * Now process all the entries, sending them to the driver.
 961          */
 962         queued = 0;
 963         while (!list_empty(list)) {
 964                 struct blk_mq_queue_data bd;
 965
 966                 rq = list_first_entry(list, struct request, queuelist);
 967                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
 968                         if (!queued && reorder_tags_to_front(list))
 969                                 continue;
 970
 971                         /*
 972                          * The initial allocation attempt failed, so we need to
 973                          * rerun the hardware queue when a tag is freed.
 974                          */
 975                         if (blk_mq_dispatch_wait_add(hctx)) {
 976                                 /*
 977                                  * It's possible that a tag was freed in the
 978                                  * window between the allocation failure and
 979                                  * adding the hardware queue to the wait queue.
 980                                  */
 981                                 if (!blk_mq_get_driver_tag(rq, &hctx, false))
 982                                         break;
 983                         } else {
 984                                 break;
 985                         }
 986                 }
 987
 988                 list_del_init(&rq->queuelist);
 989
 990                 bd.rq = rq;
 991                 bd.list = dptr;
 992                 bd.last = list_empty(list);
 993
 994                 ret = q->mq_ops->queue_rq(hctx, &bd);
 995                 switch (ret) {
 996                 case BLK_MQ_RQ_QUEUE_OK:
 997                         queued++;
 998                         break;
 999                 case BLK_MQ_RQ_QUEUE_BUSY:
1000                         blk_mq_put_driver_tag(hctx, rq);
1001                         list_add(&rq->queuelist, list);
1002                         __blk_mq_requeue_request(rq);
1003                         break;
1004                 default:
1005                         pr_err("blk-mq: bad return on queue: %d\n", ret);
1006                 case BLK_MQ_RQ_QUEUE_ERROR:
1007                         rq->errors = -EIO;
1008                         blk_mq_end_request(rq, rq->errors);
1009                         break;
1010                 }
1011
1012                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
1013                         break;
1014
1015                 /*
1016                  * We've done the first request. If we have more than 1
1017                  * left in the list, set dptr to defer issue.
1018                  */
1019                 if (!dptr && list->next != list->prev)
1020                         dptr = &driver_list;
1021         }
1022
1023         hctx->dispatched[queued_to_index(queued)]++;
1024
1025         /*
1026          * Any items that need requeuing? Stuff them into hctx->dispatch,
1027          * that is where we will continue on next queue run.
1028          */
1029         if (!list_empty(list)) {
1030                 spin_lock(&hctx->lock);
1031                 list_splice_init(list, &hctx->dispatch);
1032                 spin_unlock(&hctx->lock);
1033
1034                 /*
1035                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
1036                  * it's possible the queue is stopped and restarted again
1037                  * before this. Queue restart will dispatch requests. And since
1038                  * requests in rq_list aren't added into hctx->dispatch yet,
1039                  * the requests in rq_list might get lost.
1040                  *
1041                  * blk_mq_run_hw_queue() already checks the STOPPED bit
1042                  *
1043                  * If RESTART or TAG_WAITING is set, then let completion restart
1044                  * the queue instead of potentially looping here.
1045                  */
1046                 if (!blk_mq_sched_needs_restart(hctx) &&
1047                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1048                         blk_mq_run_hw_queue(hctx, true);
1049         }
1050
1051         return queued != 0;
1052 }
1053
1054 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1055 {
1056         int srcu_idx;
1057
1058         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1059                 cpu_online(hctx->next_cpu));
1060
1061         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1062                 rcu_read_lock();
1063                 blk_mq_sched_dispatch_requests(hctx);
1064                 rcu_read_unlock();
1065         } else {
1066                 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1067                 blk_mq_sched_dispatch_requests(hctx);
1068                 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1069         }
1070 }
1071
1072 /*
1073  * It'd be great if the workqueue API had a way to pass
1074  * in a mask and had some smarts for more clever placement.
1075  * For now we just round-robin here, switching for every
1076  * BLK_MQ_CPU_WORK_BATCH queued items.
1077  */
1078 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1079 {
1080         if (hctx->queue->nr_hw_queues == 1)
1081                 return WORK_CPU_UNBOUND;
1082
1083         if (--hctx->next_cpu_batch <= 0) {
1084                 int next_cpu;
1085
1086                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1087                 if (next_cpu >= nr_cpu_ids)
1088                         next_cpu = cpumask_first(hctx->cpumask);
1089
1090                 hctx->next_cpu = next_cpu;
1091                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1092         }
1093
1094         return hctx->next_cpu;
1095 }
1096
1097 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1098 {
1099         if (unlikely(blk_mq_hctx_stopped(hctx) ||
1100                      !blk_mq_hw_queue_mapped(hctx)))
1101                 return;
1102
1103         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1104                 int cpu = get_cpu();
1105                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1106                         __blk_mq_run_hw_queue(hctx);
1107                         put_cpu();
1108                         return;
1109                 }
1110
1111                 put_cpu();
1112         }
1113
1114         kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1115 }
1116
1117 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1118 {
1119         struct blk_mq_hw_ctx *hctx;
1120         int i;
1121
1122         queue_for_each_hw_ctx(q, hctx, i) {
1123                 if (!blk_mq_hctx_has_pending(hctx) ||
1124                     blk_mq_hctx_stopped(hctx))
1125                         continue;
1126
1127                 blk_mq_run_hw_queue(hctx, async);
1128         }
1129 }
1130 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1131
1132 /**
1133  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1134  * @q: request queue.
1135  *
1136  * The caller is responsible for serializing this function against
1137  * blk_mq_{start,stop}_hw_queue().
1138  */
1139 bool blk_mq_queue_stopped(struct request_queue *q)
1140 {
1141         struct blk_mq_hw_ctx *hctx;
1142         int i;
1143
1144         queue_for_each_hw_ctx(q, hctx, i)
1145                 if (blk_mq_hctx_stopped(hctx))
1146                         return true;
1147
1148         return false;
1149 }
1150 EXPORT_SYMBOL(blk_mq_queue_stopped);
1151
1152 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1153 {
1154         cancel_work(&hctx->run_work);
1155         cancel_delayed_work(&hctx->delay_work);
1156         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1157 }
1158 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1159
1160 void blk_mq_stop_hw_queues(struct request_queue *q)
1161 {
1162         struct blk_mq_hw_ctx *hctx;
1163         int i;
1164
1165         queue_for_each_hw_ctx(q, hctx, i)
1166                 blk_mq_stop_hw_queue(hctx);
1167 }
1168 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1169
1170 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1171 {
1172         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1173
1174         blk_mq_run_hw_queue(hctx, false);
1175 }
1176 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1177
1178 void blk_mq_start_hw_queues(struct request_queue *q)
1179 {
1180         struct blk_mq_hw_ctx *hctx;
1181         int i;
1182
1183         queue_for_each_hw_ctx(q, hctx, i)
1184                 blk_mq_start_hw_queue(hctx);
1185 }
1186 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1187
1188 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1189 {
1190         if (!blk_mq_hctx_stopped(hctx))
1191                 return;
1192
1193         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1194         blk_mq_run_hw_queue(hctx, async);
1195 }
1196 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1197
1198 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1199 {
1200         struct blk_mq_hw_ctx *hctx;
1201         int i;
1202
1203         queue_for_each_hw_ctx(q, hctx, i)
1204                 blk_mq_start_stopped_hw_queue(hctx, async);
1205 }
1206 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1207
1208 static void blk_mq_run_work_fn(struct work_struct *work)
1209 {
1210         struct blk_mq_hw_ctx *hctx;
1211
1212         hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1213
1214         __blk_mq_run_hw_queue(hctx);
1215 }
1216
1217 static void blk_mq_delay_work_fn(struct work_struct *work)
1218 {
1219         struct blk_mq_hw_ctx *hctx;
1220
1221         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1222
1223         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1224                 __blk_mq_run_hw_queue(hctx);
1225 }
1226
1227 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1228 {
1229         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1230                 return;
1231
1232         blk_mq_stop_hw_queue(hctx);
1233         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1234                         &hctx->delay_work, msecs_to_jiffies(msecs));
1235 }
1236 EXPORT_SYMBOL(blk_mq_delay_queue);
1237
1238 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1239                                             struct request *rq,
1240                                             bool at_head)
1241 {
1242         struct blk_mq_ctx *ctx = rq->mq_ctx;
1243
1244         trace_block_rq_insert(hctx->queue, rq);
1245
1246         if (at_head)
1247                 list_add(&rq->queuelist, &ctx->rq_list);
1248         else
1249                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1250 }
1251
1252 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1253                              bool at_head)
1254 {
1255         struct blk_mq_ctx *ctx = rq->mq_ctx;
1256
1257         __blk_mq_insert_req_list(hctx, rq, at_head);
1258         blk_mq_hctx_mark_pending(hctx, ctx);
1259 }
1260
1261 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1262                             struct list_head *list)
1263
1264 {
1265         /*
1266          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1267          * offline now
1268          */
1269         spin_lock(&ctx->lock);
1270         while (!list_empty(list)) {
1271                 struct request *rq;
1272
1273                 rq = list_first_entry(list, struct request, queuelist);
1274                 BUG_ON(rq->mq_ctx != ctx);
1275                 list_del_init(&rq->queuelist);
1276                 __blk_mq_insert_req_list(hctx, rq, false);
1277         }
1278         blk_mq_hctx_mark_pending(hctx, ctx);
1279         spin_unlock(&ctx->lock);
1280 }
1281
1282 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1283 {
1284         struct request *rqa = container_of(a, struct request, queuelist);
1285         struct request *rqb = container_of(b, struct request, queuelist);
1286
1287         return !(rqa->mq_ctx < rqb->mq_ctx ||
1288                  (rqa->mq_ctx == rqb->mq_ctx &&
1289                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1290 }
1291
1292 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1293 {
1294         struct blk_mq_ctx *this_ctx;
1295         struct request_queue *this_q;
1296         struct request *rq;
1297         LIST_HEAD(list);
1298         LIST_HEAD(ctx_list);
1299         unsigned int depth;
1300
1301         list_splice_init(&plug->mq_list, &list);
1302
1303         list_sort(NULL, &list, plug_ctx_cmp);
1304
1305         this_q = NULL;
1306         this_ctx = NULL;
1307         depth = 0;
1308
1309         while (!list_empty(&list)) {
1310                 rq = list_entry_rq(list.next);
1311                 list_del_init(&rq->queuelist);
1312                 BUG_ON(!rq->q);
1313                 if (rq->mq_ctx != this_ctx) {
1314                         if (this_ctx) {
1315                                 trace_block_unplug(this_q, depth, from_schedule);
1316                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1317                                                                 &ctx_list,
1318                                                                 from_schedule);
1319                         }
1320
1321                         this_ctx = rq->mq_ctx;
1322                         this_q = rq->q;
1323                         depth = 0;
1324                 }
1325
1326                 depth++;
1327                 list_add_tail(&rq->queuelist, &ctx_list);
1328         }
1329
1330         /*
1331          * If 'this_ctx' is set, we know we have entries to complete
1332          * on 'ctx_list'. Do those.
1333          */
1334         if (this_ctx) {
1335                 trace_block_unplug(this_q, depth, from_schedule);
1336                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1337                                                 from_schedule);
1338         }
1339 }
1340
1341 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1342 {
1343         init_request_from_bio(rq, bio);
1344
1345         blk_account_io_start(rq, true);
1346 }
1347
1348 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1349 {
1350         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1351                 !blk_queue_nomerges(hctx->queue);
1352 }
1353
1354 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1355                                          struct blk_mq_ctx *ctx,
1356                                          struct request *rq, struct bio *bio)
1357 {
1358         if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1359                 blk_mq_bio_to_request(rq, bio);
1360                 spin_lock(&ctx->lock);
1361 insert_rq:
1362                 __blk_mq_insert_request(hctx, rq, false);
1363                 spin_unlock(&ctx->lock);
1364                 return false;
1365         } else {
1366                 struct request_queue *q = hctx->queue;
1367
1368                 spin_lock(&ctx->lock);
1369                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1370                         blk_mq_bio_to_request(rq, bio);
1371                         goto insert_rq;
1372                 }
1373
1374                 spin_unlock(&ctx->lock);
1375                 __blk_mq_finish_request(hctx, ctx, rq);
1376                 return true;
1377         }
1378 }
1379
1380 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1381 {
1382         if (rq->tag != -1)
1383                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1384
1385         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1386 }
1387
1388 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie)
1389 {
1390         struct request_queue *q = rq->q;
1391         struct blk_mq_queue_data bd = {
1392                 .rq = rq,
1393                 .list = NULL,
1394                 .last = 1
1395         };
1396         struct blk_mq_hw_ctx *hctx;
1397         blk_qc_t new_cookie;
1398         int ret;
1399
1400         if (q->elevator)
1401                 goto insert;
1402
1403         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1404                 goto insert;
1405
1406         new_cookie = request_to_qc_t(hctx, rq);
1407
1408         /*
1409          * For OK queue, we are done. For error, kill it. Any other
1410          * error (busy), just add it to our list as we previously
1411          * would have done
1412          */
1413         ret = q->mq_ops->queue_rq(hctx, &bd);
1414         if (ret == BLK_MQ_RQ_QUEUE_OK) {
1415                 *cookie = new_cookie;
1416                 return;
1417         }
1418
1419         __blk_mq_requeue_request(rq);
1420
1421         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1422                 *cookie = BLK_QC_T_NONE;
1423                 rq->errors = -EIO;
1424                 blk_mq_end_request(rq, rq->errors);
1425                 return;
1426         }
1427
1428 insert:
1429         blk_mq_sched_insert_request(rq, false, true, true, false);
1430 }
1431
1432 /*
1433  * Multiple hardware queue variant. This will not use per-process plugs,
1434  * but will attempt to bypass the hctx queueing if we can go straight to
1435  * hardware for SYNC IO.
1436  */
1437 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1438 {
1439         const int is_sync = op_is_sync(bio->bi_opf);
1440         const int is_flush_fua = op_is_flush(bio->bi_opf);
1441         struct blk_mq_alloc_data data = { .flags = 0 };
1442         struct request *rq;
1443         unsigned int request_count = 0, srcu_idx;
1444         struct blk_plug *plug;
1445         struct request *same_queue_rq = NULL;
1446         blk_qc_t cookie;
1447         unsigned int wb_acct;
1448
1449         blk_queue_bounce(q, &bio);
1450
1451         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1452                 bio_io_error(bio);
1453                 return BLK_QC_T_NONE;
1454         }
1455
1456         blk_queue_split(q, &bio, q->bio_split);
1457
1458         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1459             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1460                 return BLK_QC_T_NONE;
1461
1462         if (blk_mq_sched_bio_merge(q, bio))
1463                 return BLK_QC_T_NONE;
1464
1465         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1466
1467         trace_block_getrq(q, bio, bio->bi_opf);
1468
1469         rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1470         if (unlikely(!rq)) {
1471                 __wbt_done(q->rq_wb, wb_acct);
1472                 return BLK_QC_T_NONE;
1473         }
1474
1475         wbt_track(&rq->issue_stat, wb_acct);
1476
1477         cookie = request_to_qc_t(data.hctx, rq);
1478
1479         if (unlikely(is_flush_fua)) {
1480                 if (q->elevator)
1481                         goto elv_insert;
1482                 blk_mq_bio_to_request(rq, bio);
1483                 blk_insert_flush(rq);
1484                 goto run_queue;
1485         }
1486
1487         plug = current->plug;
1488         /*
1489          * If the driver supports defer issued based on 'last', then
1490          * queue it up like normal since we can potentially save some
1491          * CPU this way.
1492          */
1493         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1494             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1495                 struct request *old_rq = NULL;
1496
1497                 blk_mq_bio_to_request(rq, bio);
1498
1499                 /*
1500                  * We do limited plugging. If the bio can be merged, do that.
1501                  * Otherwise the existing request in the plug list will be
1502                  * issued. So the plug list will have one request at most
1503                  */
1504                 if (plug) {
1505                         /*
1506                          * The plug list might get flushed before this. If that
1507                          * happens, same_queue_rq is invalid and plug list is
1508                          * empty
1509                          */
1510                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1511                                 old_rq = same_queue_rq;
1512                                 list_del_init(&old_rq->queuelist);
1513                         }
1514                         list_add_tail(&rq->queuelist, &plug->mq_list);
1515                 } else /* is_sync */
1516                         old_rq = rq;
1517                 blk_mq_put_ctx(data.ctx);
1518                 if (!old_rq)
1519                         goto done;
1520
1521                 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1522                         rcu_read_lock();
1523                         blk_mq_try_issue_directly(old_rq, &cookie);
1524                         rcu_read_unlock();
1525                 } else {
1526                         srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1527                         blk_mq_try_issue_directly(old_rq, &cookie);
1528                         srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1529                 }
1530                 goto done;
1531         }
1532
1533         if (q->elevator) {
1534 elv_insert:
1535                 blk_mq_put_ctx(data.ctx);
1536                 blk_mq_bio_to_request(rq, bio);
1537                 blk_mq_sched_insert_request(rq, false, true,
1538                                                 !is_sync || is_flush_fua, true);
1539                 goto done;
1540         }
1541         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1542                 /*
1543                  * For a SYNC request, send it to the hardware immediately. For
1544                  * an ASYNC request, just ensure that we run it later on. The
1545                  * latter allows for merging opportunities and more efficient
1546                  * dispatching.
1547                  */
1548 run_queue:
1549                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1550         }
1551         blk_mq_put_ctx(data.ctx);
1552 done:
1553         return cookie;
1554 }
1555
1556 /*
1557  * Single hardware queue variant. This will attempt to use any per-process
1558  * plug for merging and IO deferral.
1559  */
1560 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1561 {
1562         const int is_sync = op_is_sync(bio->bi_opf);
1563         const int is_flush_fua = op_is_flush(bio->bi_opf);
1564         struct blk_plug *plug;
1565         unsigned int request_count = 0;
1566         struct blk_mq_alloc_data data = { .flags = 0 };
1567         struct request *rq;
1568         blk_qc_t cookie;
1569         unsigned int wb_acct;
1570
1571         blk_queue_bounce(q, &bio);
1572
1573         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1574                 bio_io_error(bio);
1575                 return BLK_QC_T_NONE;
1576         }
1577
1578         blk_queue_split(q, &bio, q->bio_split);
1579
1580         if (!is_flush_fua && !blk_queue_nomerges(q)) {
1581                 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1582                         return BLK_QC_T_NONE;
1583         } else
1584                 request_count = blk_plug_queued_count(q);
1585
1586         if (blk_mq_sched_bio_merge(q, bio))
1587                 return BLK_QC_T_NONE;
1588
1589         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1590
1591         trace_block_getrq(q, bio, bio->bi_opf);
1592
1593         rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1594         if (unlikely(!rq)) {
1595                 __wbt_done(q->rq_wb, wb_acct);
1596                 return BLK_QC_T_NONE;
1597         }
1598
1599         wbt_track(&rq->issue_stat, wb_acct);
1600
1601         cookie = request_to_qc_t(data.hctx, rq);
1602
1603         if (unlikely(is_flush_fua)) {
1604                 if (q->elevator)
1605                         goto elv_insert;
1606                 blk_mq_bio_to_request(rq, bio);
1607                 blk_insert_flush(rq);
1608                 goto run_queue;
1609         }
1610
1611         /*
1612          * A task plug currently exists. Since this is completely lockless,
1613          * utilize that to temporarily store requests until the task is
1614          * either done or scheduled away.
1615          */
1616         plug = current->plug;
1617         if (plug) {
1618                 struct request *last = NULL;
1619
1620                 blk_mq_bio_to_request(rq, bio);
1621
1622                 /*
1623                  * @request_count may become stale because of schedule
1624                  * out, so check the list again.
1625                  */
1626                 if (list_empty(&plug->mq_list))
1627                         request_count = 0;
1628                 if (!request_count)
1629                         trace_block_plug(q);
1630                 else
1631                         last = list_entry_rq(plug->mq_list.prev);
1632
1633                 blk_mq_put_ctx(data.ctx);
1634
1635                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1636                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1637                         blk_flush_plug_list(plug, false);
1638                         trace_block_plug(q);
1639                 }
1640
1641                 list_add_tail(&rq->queuelist, &plug->mq_list);
1642                 return cookie;
1643         }
1644
1645         if (q->elevator) {
1646 elv_insert:
1647                 blk_mq_put_ctx(data.ctx);
1648                 blk_mq_bio_to_request(rq, bio);
1649                 blk_mq_sched_insert_request(rq, false, true,
1650                                                 !is_sync || is_flush_fua, true);
1651                 goto done;
1652         }
1653         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1654                 /*
1655                  * For a SYNC request, send it to the hardware immediately. For
1656                  * an ASYNC request, just ensure that we run it later on. The
1657                  * latter allows for merging opportunities and more efficient
1658                  * dispatching.
1659                  */
1660 run_queue:
1661                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1662         }
1663
1664         blk_mq_put_ctx(data.ctx);
1665 done:
1666         return cookie;
1667 }
1668
1669 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1670                      unsigned int hctx_idx)
1671 {
1672         struct page *page;
1673
1674         if (tags->rqs && set->ops->exit_request) {
1675                 int i;
1676
1677                 for (i = 0; i < tags->nr_tags; i++) {
1678                         struct request *rq = tags->static_rqs[i];
1679
1680                         if (!rq)
1681                                 continue;
1682                         set->ops->exit_request(set->driver_data, rq,
1683                                                 hctx_idx, i);
1684                         tags->static_rqs[i] = NULL;
1685                 }
1686         }
1687
1688         while (!list_empty(&tags->page_list)) {
1689                 page = list_first_entry(&tags->page_list, struct page, lru);
1690                 list_del_init(&page->lru);
1691                 /*
1692                  * Remove kmemleak object previously allocated in
1693                  * blk_mq_init_rq_map().
1694                  */
1695                 kmemleak_free(page_address(page));
1696                 __free_pages(page, page->private);
1697         }
1698 }
1699
1700 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1701 {
1702         kfree(tags->rqs);
1703         tags->rqs = NULL;
1704         kfree(tags->static_rqs);
1705         tags->static_rqs = NULL;
1706
1707         blk_mq_free_tags(tags);
1708 }
1709
1710 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1711                                         unsigned int hctx_idx,
1712                                         unsigned int nr_tags,
1713                                         unsigned int reserved_tags)
1714 {
1715         struct blk_mq_tags *tags;
1716
1717         tags = blk_mq_init_tags(nr_tags, reserved_tags,
1718                                 set->numa_node,
1719                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1720         if (!tags)
1721                 return NULL;
1722
1723         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1724                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1725                                  set->numa_node);
1726         if (!tags->rqs) {
1727                 blk_mq_free_tags(tags);
1728                 return NULL;
1729         }
1730
1731         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1732                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1733                                  set->numa_node);
1734         if (!tags->static_rqs) {
1735                 kfree(tags->rqs);
1736                 blk_mq_free_tags(tags);
1737                 return NULL;
1738         }
1739
1740         return tags;
1741 }
1742
1743 static size_t order_to_size(unsigned int order)
1744 {
1745         return (size_t)PAGE_SIZE << order;
1746 }
1747
1748 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1749                      unsigned int hctx_idx, unsigned int depth)
1750 {
1751         unsigned int i, j, entries_per_page, max_order = 4;
1752         size_t rq_size, left;
1753
1754         INIT_LIST_HEAD(&tags->page_list);
1755
1756         /*
1757          * rq_size is the size of the request plus driver payload, rounded
1758          * to the cacheline size
1759          */
1760         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1761                                 cache_line_size());
1762         left = rq_size * depth;
1763
1764         for (i = 0; i < depth; ) {
1765                 int this_order = max_order;
1766                 struct page *page;
1767                 int to_do;
1768                 void *p;
1769
1770                 while (this_order && left < order_to_size(this_order - 1))
1771                         this_order--;
1772
1773                 do {
1774                         page = alloc_pages_node(set->numa_node,
1775                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1776                                 this_order);
1777                         if (page)
1778                                 break;
1779                         if (!this_order--)
1780                                 break;
1781                         if (order_to_size(this_order) < rq_size)
1782                                 break;
1783                 } while (1);
1784
1785                 if (!page)
1786                         goto fail;
1787
1788                 page->private = this_order;
1789                 list_add_tail(&page->lru, &tags->page_list);
1790
1791                 p = page_address(page);
1792                 /*
1793                  * Allow kmemleak to scan these pages as they contain pointers
1794                  * to additional allocations like via ops->init_request().
1795                  */
1796                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1797                 entries_per_page = order_to_size(this_order) / rq_size;
1798                 to_do = min(entries_per_page, depth - i);
1799                 left -= to_do * rq_size;
1800                 for (j = 0; j < to_do; j++) {
1801                         struct request *rq = p;
1802
1803                         tags->static_rqs[i] = rq;
1804                         if (set->ops->init_request) {
1805                                 if (set->ops->init_request(set->driver_data,
1806                                                 rq, hctx_idx, i,
1807                                                 set->numa_node)) {
1808                                         tags->static_rqs[i] = NULL;
1809                                         goto fail;
1810                                 }
1811                         }
1812
1813                         p += rq_size;
1814                         i++;
1815                 }
1816         }
1817         return 0;
1818
1819 fail:
1820         blk_mq_free_rqs(set, tags, hctx_idx);
1821         return -ENOMEM;
1822 }
1823
1824 /*
1825  * 'cpu' is going away. splice any existing rq_list entries from this
1826  * software queue to the hw queue dispatch list, and ensure that it
1827  * gets run.
1828  */
1829 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1830 {
1831         struct blk_mq_hw_ctx *hctx;
1832         struct blk_mq_ctx *ctx;
1833         LIST_HEAD(tmp);
1834
1835         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1836         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1837
1838         spin_lock(&ctx->lock);
1839         if (!list_empty(&ctx->rq_list)) {
1840                 list_splice_init(&ctx->rq_list, &tmp);
1841                 blk_mq_hctx_clear_pending(hctx, ctx);
1842         }
1843         spin_unlock(&ctx->lock);
1844
1845         if (list_empty(&tmp))
1846                 return 0;
1847
1848         spin_lock(&hctx->lock);
1849         list_splice_tail_init(&tmp, &hctx->dispatch);
1850         spin_unlock(&hctx->lock);
1851
1852         blk_mq_run_hw_queue(hctx, true);
1853         return 0;
1854 }
1855
1856 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1857 {
1858         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1859                                             &hctx->cpuhp_dead);
1860 }
1861
1862 /* hctx->ctxs will be freed in queue's release handler */
1863 static void blk_mq_exit_hctx(struct request_queue *q,
1864                 struct blk_mq_tag_set *set,
1865                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1866 {
1867         unsigned flush_start_tag = set->queue_depth;
1868
1869         blk_mq_tag_idle(hctx);
1870
1871         if (set->ops->exit_request)
1872                 set->ops->exit_request(set->driver_data,
1873                                        hctx->fq->flush_rq, hctx_idx,
1874                                        flush_start_tag + hctx_idx);
1875
1876         if (set->ops->exit_hctx)
1877                 set->ops->exit_hctx(hctx, hctx_idx);
1878
1879         if (hctx->flags & BLK_MQ_F_BLOCKING)
1880                 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1881
1882         blk_mq_remove_cpuhp(hctx);
1883         blk_free_flush_queue(hctx->fq);
1884         sbitmap_free(&hctx->ctx_map);
1885 }
1886
1887 static void blk_mq_exit_hw_queues(struct request_queue *q,
1888                 struct blk_mq_tag_set *set, int nr_queue)
1889 {
1890         struct blk_mq_hw_ctx *hctx;
1891         unsigned int i;
1892
1893         queue_for_each_hw_ctx(q, hctx, i) {
1894                 if (i == nr_queue)
1895                         break;
1896                 blk_mq_exit_hctx(q, set, hctx, i);
1897         }
1898 }
1899
1900 static void blk_mq_free_hw_queues(struct request_queue *q,
1901                 struct blk_mq_tag_set *set)
1902 {
1903         struct blk_mq_hw_ctx *hctx;
1904         unsigned int i;
1905
1906         queue_for_each_hw_ctx(q, hctx, i)
1907                 free_cpumask_var(hctx->cpumask);
1908 }
1909
1910 static int blk_mq_init_hctx(struct request_queue *q,
1911                 struct blk_mq_tag_set *set,
1912                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1913 {
1914         int node;
1915         unsigned flush_start_tag = set->queue_depth;
1916
1917         node = hctx->numa_node;
1918         if (node == NUMA_NO_NODE)
1919                 node = hctx->numa_node = set->numa_node;
1920
1921         INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1922         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1923         spin_lock_init(&hctx->lock);
1924         INIT_LIST_HEAD(&hctx->dispatch);
1925         hctx->queue = q;
1926         hctx->queue_num = hctx_idx;
1927         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1928
1929         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1930
1931         hctx->tags = set->tags[hctx_idx];
1932
1933         /*
1934          * Allocate space for all possible cpus to avoid allocation at
1935          * runtime
1936          */
1937         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1938                                         GFP_KERNEL, node);
1939         if (!hctx->ctxs)
1940                 goto unregister_cpu_notifier;
1941
1942         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1943                               node))
1944                 goto free_ctxs;
1945
1946         hctx->nr_ctx = 0;
1947
1948         if (set->ops->init_hctx &&
1949             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1950                 goto free_bitmap;
1951
1952         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1953         if (!hctx->fq)
1954                 goto exit_hctx;
1955
1956         if (set->ops->init_request &&
1957             set->ops->init_request(set->driver_data,
1958                                    hctx->fq->flush_rq, hctx_idx,
1959                                    flush_start_tag + hctx_idx, node))
1960                 goto free_fq;
1961
1962         if (hctx->flags & BLK_MQ_F_BLOCKING)
1963                 init_srcu_struct(&hctx->queue_rq_srcu);
1964
1965         return 0;
1966
1967  free_fq:
1968         kfree(hctx->fq);
1969  exit_hctx:
1970         if (set->ops->exit_hctx)
1971                 set->ops->exit_hctx(hctx, hctx_idx);
1972  free_bitmap:
1973         sbitmap_free(&hctx->ctx_map);
1974  free_ctxs:
1975         kfree(hctx->ctxs);
1976  unregister_cpu_notifier:
1977         blk_mq_remove_cpuhp(hctx);
1978         return -1;
1979 }
1980
1981 static void blk_mq_init_cpu_queues(struct request_queue *q,
1982                                    unsigned int nr_hw_queues)
1983 {
1984         unsigned int i;
1985
1986         for_each_possible_cpu(i) {
1987                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1988                 struct blk_mq_hw_ctx *hctx;
1989
1990                 memset(__ctx, 0, sizeof(*__ctx));
1991                 __ctx->cpu = i;
1992                 spin_lock_init(&__ctx->lock);
1993                 INIT_LIST_HEAD(&__ctx->rq_list);
1994                 __ctx->queue = q;
1995                 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
1996                 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
1997
1998                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1999                 if (!cpu_online(i))
2000                         continue;
2001
2002                 hctx = blk_mq_map_queue(q, i);
2003
2004                 /*
2005                  * Set local node, IFF we have more than one hw queue. If
2006                  * not, we remain on the home node of the device
2007                  */
2008                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2009                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2010         }
2011 }
2012
2013 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2014 {
2015         int ret = 0;
2016
2017         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2018                                         set->queue_depth, set->reserved_tags);
2019         if (!set->tags[hctx_idx])
2020                 return false;
2021
2022         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2023                                 set->queue_depth);
2024         if (!ret)
2025                 return true;
2026
2027         blk_mq_free_rq_map(set->tags[hctx_idx]);
2028         set->tags[hctx_idx] = NULL;
2029         return false;
2030 }
2031
2032 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2033                                          unsigned int hctx_idx)
2034 {
2035         if (set->tags[hctx_idx]) {
2036                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2037                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2038                 set->tags[hctx_idx] = NULL;
2039         }
2040 }
2041
2042 static void blk_mq_map_swqueue(struct request_queue *q,
2043                                const struct cpumask *online_mask)
2044 {
2045         unsigned int i, hctx_idx;
2046         struct blk_mq_hw_ctx *hctx;
2047         struct blk_mq_ctx *ctx;
2048         struct blk_mq_tag_set *set = q->tag_set;
2049
2050         /*
2051          * Avoid others reading imcomplete hctx->cpumask through sysfs
2052          */
2053         mutex_lock(&q->sysfs_lock);
2054
2055         queue_for_each_hw_ctx(q, hctx, i) {
2056                 cpumask_clear(hctx->cpumask);
2057                 hctx->nr_ctx = 0;
2058         }
2059
2060         /*
2061          * Map software to hardware queues
2062          */
2063         for_each_possible_cpu(i) {
2064                 /* If the cpu isn't online, the cpu is mapped to first hctx */
2065                 if (!cpumask_test_cpu(i, online_mask))
2066                         continue;
2067
2068                 hctx_idx = q->mq_map[i];
2069                 /* unmapped hw queue can be remapped after CPU topo changed */
2070                 if (!set->tags[hctx_idx] &&
2071                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2072                         /*
2073                          * If tags initialization fail for some hctx,
2074                          * that hctx won't be brought online.  In this
2075                          * case, remap the current ctx to hctx[0] which
2076                          * is guaranteed to always have tags allocated
2077                          */
2078                         q->mq_map[i] = 0;
2079                 }
2080
2081                 ctx = per_cpu_ptr(q->queue_ctx, i);
2082                 hctx = blk_mq_map_queue(q, i);
2083
2084                 cpumask_set_cpu(i, hctx->cpumask);
2085                 ctx->index_hw = hctx->nr_ctx;
2086                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2087         }
2088
2089         mutex_unlock(&q->sysfs_lock);
2090
2091         queue_for_each_hw_ctx(q, hctx, i) {
2092                 /*
2093                  * If no software queues are mapped to this hardware queue,
2094                  * disable it and free the request entries.
2095                  */
2096                 if (!hctx->nr_ctx) {
2097                         /* Never unmap queue 0.  We need it as a
2098                          * fallback in case of a new remap fails
2099                          * allocation
2100                          */
2101                         if (i && set->tags[i])
2102                                 blk_mq_free_map_and_requests(set, i);
2103
2104                         hctx->tags = NULL;
2105                         continue;
2106                 }
2107
2108                 hctx->tags = set->tags[i];
2109                 WARN_ON(!hctx->tags);
2110
2111                 /*
2112                  * Set the map size to the number of mapped software queues.
2113                  * This is more accurate and more efficient than looping
2114                  * over all possibly mapped software queues.
2115                  */
2116                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2117
2118                 /*
2119                  * Initialize batch roundrobin counts
2120                  */
2121                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2122                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2123         }
2124 }
2125
2126 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2127 {
2128         struct blk_mq_hw_ctx *hctx;
2129         int i;
2130
2131         queue_for_each_hw_ctx(q, hctx, i) {
2132                 if (shared)
2133                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2134                 else
2135                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2136         }
2137 }
2138
2139 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2140 {
2141         struct request_queue *q;
2142
2143         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2144                 blk_mq_freeze_queue(q);
2145                 queue_set_hctx_shared(q, shared);
2146                 blk_mq_unfreeze_queue(q);
2147         }
2148 }
2149
2150 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2151 {
2152         struct blk_mq_tag_set *set = q->tag_set;
2153
2154         mutex_lock(&set->tag_list_lock);
2155         list_del_init(&q->tag_set_list);
2156         if (list_is_singular(&set->tag_list)) {
2157                 /* just transitioned to unshared */
2158                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2159                 /* update existing queue */
2160                 blk_mq_update_tag_set_depth(set, false);
2161         }
2162         mutex_unlock(&set->tag_list_lock);
2163 }
2164
2165 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2166                                      struct request_queue *q)
2167 {
2168         q->tag_set = set;
2169
2170         mutex_lock(&set->tag_list_lock);
2171
2172         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2173         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2174                 set->flags |= BLK_MQ_F_TAG_SHARED;
2175                 /* update existing queue */
2176                 blk_mq_update_tag_set_depth(set, true);
2177         }
2178         if (set->flags & BLK_MQ_F_TAG_SHARED)
2179                 queue_set_hctx_shared(q, true);
2180         list_add_tail(&q->tag_set_list, &set->tag_list);
2181
2182         mutex_unlock(&set->tag_list_lock);
2183 }
2184
2185 /*
2186  * It is the actual release handler for mq, but we do it from
2187  * request queue's release handler for avoiding use-after-free
2188  * and headache because q->mq_kobj shouldn't have been introduced,
2189  * but we can't group ctx/kctx kobj without it.
2190  */
2191 void blk_mq_release(struct request_queue *q)
2192 {
2193         struct blk_mq_hw_ctx *hctx;
2194         unsigned int i;
2195
2196         blk_mq_sched_teardown(q);
2197
2198         /* hctx kobj stays in hctx */
2199         queue_for_each_hw_ctx(q, hctx, i) {
2200                 if (!hctx)
2201                         continue;
2202                 kfree(hctx->ctxs);
2203                 kfree(hctx);
2204         }
2205
2206         q->mq_map = NULL;
2207
2208         kfree(q->queue_hw_ctx);
2209
2210         /* ctx kobj stays in queue_ctx */
2211         free_percpu(q->queue_ctx);
2212 }
2213
2214 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2215 {
2216         struct request_queue *uninit_q, *q;
2217
2218         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2219         if (!uninit_q)
2220                 return ERR_PTR(-ENOMEM);
2221
2222         q = blk_mq_init_allocated_queue(set, uninit_q);
2223         if (IS_ERR(q))
2224                 blk_cleanup_queue(uninit_q);
2225
2226         return q;
2227 }
2228 EXPORT_SYMBOL(blk_mq_init_queue);
2229
2230 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2231                                                 struct request_queue *q)
2232 {
2233         int i, j;
2234         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2235
2236         blk_mq_sysfs_unregister(q);
2237         for (i = 0; i < set->nr_hw_queues; i++) {
2238                 int node;
2239
2240                 if (hctxs[i])
2241                         continue;
2242
2243                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2244                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2245                                         GFP_KERNEL, node);
2246                 if (!hctxs[i])
2247                         break;
2248
2249                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2250                                                 node)) {
2251                         kfree(hctxs[i]);
2252                         hctxs[i] = NULL;
2253                         break;
2254                 }
2255
2256                 atomic_set(&hctxs[i]->nr_active, 0);
2257                 hctxs[i]->numa_node = node;
2258                 hctxs[i]->queue_num = i;
2259
2260                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2261                         free_cpumask_var(hctxs[i]->cpumask);
2262                         kfree(hctxs[i]);
2263                         hctxs[i] = NULL;
2264                         break;
2265                 }
2266                 blk_mq_hctx_kobj_init(hctxs[i]);
2267         }
2268         for (j = i; j < q->nr_hw_queues; j++) {
2269                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2270
2271                 if (hctx) {
2272                         if (hctx->tags)
2273                                 blk_mq_free_map_and_requests(set, j);
2274                         blk_mq_exit_hctx(q, set, hctx, j);
2275                         free_cpumask_var(hctx->cpumask);
2276                         kobject_put(&hctx->kobj);
2277                         kfree(hctx->ctxs);
2278                         kfree(hctx);
2279                         hctxs[j] = NULL;
2280
2281                 }
2282         }
2283         q->nr_hw_queues = i;
2284         blk_mq_sysfs_register(q);
2285 }
2286
2287 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2288                                                   struct request_queue *q)
2289 {
2290         /* mark the queue as mq asap */
2291         q->mq_ops = set->ops;
2292
2293         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2294         if (!q->queue_ctx)
2295                 goto err_exit;
2296
2297         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2298                                                 GFP_KERNEL, set->numa_node);
2299         if (!q->queue_hw_ctx)
2300                 goto err_percpu;
2301
2302         q->mq_map = set->mq_map;
2303
2304         blk_mq_realloc_hw_ctxs(set, q);
2305         if (!q->nr_hw_queues)
2306                 goto err_hctxs;
2307
2308         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2309         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2310
2311         q->nr_queues = nr_cpu_ids;
2312
2313         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2314
2315         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2316                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2317
2318         q->sg_reserved_size = INT_MAX;
2319
2320         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2321         INIT_LIST_HEAD(&q->requeue_list);
2322         spin_lock_init(&q->requeue_lock);
2323
2324         if (q->nr_hw_queues > 1)
2325                 blk_queue_make_request(q, blk_mq_make_request);
2326         else
2327                 blk_queue_make_request(q, blk_sq_make_request);
2328
2329         /*
2330          * Do this after blk_queue_make_request() overrides it...
2331          */
2332         q->nr_requests = set->queue_depth;
2333
2334         /*
2335          * Default to classic polling
2336          */
2337         q->poll_nsec = -1;
2338
2339         if (set->ops->complete)
2340                 blk_queue_softirq_done(q, set->ops->complete);
2341
2342         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2343
2344         get_online_cpus();
2345         mutex_lock(&all_q_mutex);
2346
2347         list_add_tail(&q->all_q_node, &all_q_list);
2348         blk_mq_add_queue_tag_set(set, q);
2349         blk_mq_map_swqueue(q, cpu_online_mask);
2350
2351         mutex_unlock(&all_q_mutex);
2352         put_online_cpus();
2353
2354         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2355                 int ret;
2356
2357                 ret = blk_mq_sched_init(q);
2358                 if (ret)
2359                         return ERR_PTR(ret);
2360         }
2361
2362         return q;
2363
2364 err_hctxs:
2365         kfree(q->queue_hw_ctx);
2366 err_percpu:
2367         free_percpu(q->queue_ctx);
2368 err_exit:
2369         q->mq_ops = NULL;
2370         return ERR_PTR(-ENOMEM);
2371 }
2372 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2373
2374 void blk_mq_free_queue(struct request_queue *q)
2375 {
2376         struct blk_mq_tag_set   *set = q->tag_set;
2377
2378         mutex_lock(&all_q_mutex);
2379         list_del_init(&q->all_q_node);
2380         mutex_unlock(&all_q_mutex);
2381
2382         wbt_exit(q);
2383
2384         blk_mq_del_queue_tag_set(q);
2385
2386         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2387         blk_mq_free_hw_queues(q, set);
2388 }
2389
2390 /* Basically redo blk_mq_init_queue with queue frozen */
2391 static void blk_mq_queue_reinit(struct request_queue *q,
2392                                 const struct cpumask *online_mask)
2393 {
2394         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2395
2396         blk_mq_sysfs_unregister(q);
2397
2398         /*
2399          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2400          * we should change hctx numa_node according to new topology (this
2401          * involves free and re-allocate memory, worthy doing?)
2402          */
2403
2404         blk_mq_map_swqueue(q, online_mask);
2405
2406         blk_mq_sysfs_register(q);
2407 }
2408
2409 /*
2410  * New online cpumask which is going to be set in this hotplug event.
2411  * Declare this cpumasks as global as cpu-hotplug operation is invoked
2412  * one-by-one and dynamically allocating this could result in a failure.
2413  */
2414 static struct cpumask cpuhp_online_new;
2415
2416 static void blk_mq_queue_reinit_work(void)
2417 {
2418         struct request_queue *q;
2419
2420         mutex_lock(&all_q_mutex);
2421         /*
2422          * We need to freeze and reinit all existing queues.  Freezing
2423          * involves synchronous wait for an RCU grace period and doing it
2424          * one by one may take a long time.  Start freezing all queues in
2425          * one swoop and then wait for the completions so that freezing can
2426          * take place in parallel.
2427          */
2428         list_for_each_entry(q, &all_q_list, all_q_node)
2429                 blk_mq_freeze_queue_start(q);
2430         list_for_each_entry(q, &all_q_list, all_q_node)
2431                 blk_mq_freeze_queue_wait(q);
2432
2433         list_for_each_entry(q, &all_q_list, all_q_node)
2434                 blk_mq_queue_reinit(q, &cpuhp_online_new);
2435
2436         list_for_each_entry(q, &all_q_list, all_q_node)
2437                 blk_mq_unfreeze_queue(q);
2438
2439         mutex_unlock(&all_q_mutex);
2440 }
2441
2442 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2443 {
2444         cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2445         blk_mq_queue_reinit_work();
2446         return 0;
2447 }
2448
2449 /*
2450  * Before hotadded cpu starts handling requests, new mappings must be
2451  * established.  Otherwise, these requests in hw queue might never be
2452  * dispatched.
2453  *
2454  * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2455  * for CPU0, and ctx1 for CPU1).
2456  *
2457  * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2458  * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2459  *
2460  * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2461  * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2462  * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2463  * ignored.
2464  */
2465 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2466 {
2467         cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2468         cpumask_set_cpu(cpu, &cpuhp_online_new);
2469         blk_mq_queue_reinit_work();
2470         return 0;
2471 }
2472
2473 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2474 {
2475         int i;
2476
2477         for (i = 0; i < set->nr_hw_queues; i++)
2478                 if (!__blk_mq_alloc_rq_map(set, i))
2479                         goto out_unwind;
2480
2481         return 0;
2482
2483 out_unwind:
2484         while (--i >= 0)
2485                 blk_mq_free_rq_map(set->tags[i]);
2486
2487         return -ENOMEM;
2488 }
2489
2490 /*
2491  * Allocate the request maps associated with this tag_set. Note that this
2492  * may reduce the depth asked for, if memory is tight. set->queue_depth
2493  * will be updated to reflect the allocated depth.
2494  */
2495 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2496 {
2497         unsigned int depth;
2498         int err;
2499
2500         depth = set->queue_depth;
2501         do {
2502                 err = __blk_mq_alloc_rq_maps(set);
2503                 if (!err)
2504                         break;
2505
2506                 set->queue_depth >>= 1;
2507                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2508                         err = -ENOMEM;
2509                         break;
2510                 }
2511         } while (set->queue_depth);
2512
2513         if (!set->queue_depth || err) {
2514                 pr_err("blk-mq: failed to allocate request map\n");
2515                 return -ENOMEM;
2516         }
2517
2518         if (depth != set->queue_depth)
2519                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2520                                                 depth, set->queue_depth);
2521
2522         return 0;
2523 }
2524
2525 /*
2526  * Alloc a tag set to be associated with one or more request queues.
2527  * May fail with EINVAL for various error conditions. May adjust the
2528  * requested depth down, if if it too large. In that case, the set
2529  * value will be stored in set->queue_depth.
2530  */
2531 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2532 {
2533         int ret;
2534
2535         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2536
2537         if (!set->nr_hw_queues)
2538                 return -EINVAL;
2539         if (!set->queue_depth)
2540                 return -EINVAL;
2541         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2542                 return -EINVAL;
2543
2544         if (!set->ops->queue_rq)
2545                 return -EINVAL;
2546
2547         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2548                 pr_info("blk-mq: reduced tag depth to %u\n",
2549                         BLK_MQ_MAX_DEPTH);
2550                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2551         }
2552
2553         /*
2554          * If a crashdump is active, then we are potentially in a very
2555          * memory constrained environment. Limit us to 1 queue and
2556          * 64 tags to prevent using too much memory.
2557          */
2558         if (is_kdump_kernel()) {
2559                 set->nr_hw_queues = 1;
2560                 set->queue_depth = min(64U, set->queue_depth);
2561         }
2562         /*
2563          * There is no use for more h/w queues than cpus.
2564          */
2565         if (set->nr_hw_queues > nr_cpu_ids)
2566                 set->nr_hw_queues = nr_cpu_ids;
2567
2568         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2569                                  GFP_KERNEL, set->numa_node);
2570         if (!set->tags)
2571                 return -ENOMEM;
2572
2573         ret = -ENOMEM;
2574         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2575                         GFP_KERNEL, set->numa_node);
2576         if (!set->mq_map)
2577                 goto out_free_tags;
2578
2579         if (set->ops->map_queues)
2580                 ret = set->ops->map_queues(set);
2581         else
2582                 ret = blk_mq_map_queues(set);
2583         if (ret)
2584                 goto out_free_mq_map;
2585
2586         ret = blk_mq_alloc_rq_maps(set);
2587         if (ret)
2588                 goto out_free_mq_map;
2589
2590         mutex_init(&set->tag_list_lock);
2591         INIT_LIST_HEAD(&set->tag_list);
2592
2593         return 0;
2594
2595 out_free_mq_map:
2596         kfree(set->mq_map);
2597         set->mq_map = NULL;
2598 out_free_tags:
2599         kfree(set->tags);
2600         set->tags = NULL;
2601         return ret;
2602 }
2603 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2604
2605 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2606 {
2607         int i;
2608
2609         for (i = 0; i < nr_cpu_ids; i++)
2610                 blk_mq_free_map_and_requests(set, i);
2611
2612         kfree(set->mq_map);
2613         set->mq_map = NULL;
2614
2615         kfree(set->tags);
2616         set->tags = NULL;
2617 }
2618 EXPORT_SYMBOL(blk_mq_free_tag_set);
2619
2620 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2621 {
2622         struct blk_mq_tag_set *set = q->tag_set;
2623         struct blk_mq_hw_ctx *hctx;
2624         int i, ret;
2625
2626         if (!set)
2627                 return -EINVAL;
2628
2629         blk_mq_freeze_queue(q);
2630         blk_mq_quiesce_queue(q);
2631
2632         ret = 0;
2633         queue_for_each_hw_ctx(q, hctx, i) {
2634                 if (!hctx->tags)
2635                         continue;
2636                 /*
2637                  * If we're using an MQ scheduler, just update the scheduler
2638                  * queue depth. This is similar to what the old code would do.
2639                  */
2640                 if (!hctx->sched_tags) {
2641                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2642                                                         min(nr, set->queue_depth),
2643                                                         false);
2644                 } else {
2645                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2646                                                         nr, true);
2647                 }
2648                 if (ret)
2649                         break;
2650         }
2651
2652         if (!ret)
2653                 q->nr_requests = nr;
2654
2655         blk_mq_unfreeze_queue(q);
2656         blk_mq_start_stopped_hw_queues(q, true);
2657
2658         return ret;
2659 }
2660
2661 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2662 {
2663         struct request_queue *q;
2664
2665         if (nr_hw_queues > nr_cpu_ids)
2666                 nr_hw_queues = nr_cpu_ids;
2667         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2668                 return;
2669
2670         list_for_each_entry(q, &set->tag_list, tag_set_list)
2671                 blk_mq_freeze_queue(q);
2672
2673         set->nr_hw_queues = nr_hw_queues;
2674         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2675                 blk_mq_realloc_hw_ctxs(set, q);
2676
2677                 /*
2678                  * Manually set the make_request_fn as blk_queue_make_request
2679                  * resets a lot of the queue settings.
2680                  */
2681                 if (q->nr_hw_queues > 1)
2682                         q->make_request_fn = blk_mq_make_request;
2683                 else
2684                         q->make_request_fn = blk_sq_make_request;
2685
2686                 blk_mq_queue_reinit(q, cpu_online_mask);
2687         }
2688
2689         list_for_each_entry(q, &set->tag_list, tag_set_list)
2690                 blk_mq_unfreeze_queue(q);
2691 }
2692 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2693
2694 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2695                                        struct blk_mq_hw_ctx *hctx,
2696                                        struct request *rq)
2697 {
2698         struct blk_rq_stat stat[2];
2699         unsigned long ret = 0;
2700
2701         /*
2702          * If stats collection isn't on, don't sleep but turn it on for
2703          * future users
2704          */
2705         if (!blk_stat_enable(q))
2706                 return 0;
2707
2708         /*
2709          * We don't have to do this once per IO, should optimize this
2710          * to just use the current window of stats until it changes
2711          */
2712         memset(&stat, 0, sizeof(stat));
2713         blk_hctx_stat_get(hctx, stat);
2714
2715         /*
2716          * As an optimistic guess, use half of the mean service time
2717          * for this type of request. We can (and should) make this smarter.
2718          * For instance, if the completion latencies are tight, we can
2719          * get closer than just half the mean. This is especially
2720          * important on devices where the completion latencies are longer
2721          * than ~10 usec.
2722          */
2723         if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2724                 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2725         else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2726                 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2727
2728         return ret;
2729 }
2730
2731 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2732                                      struct blk_mq_hw_ctx *hctx,
2733                                      struct request *rq)
2734 {
2735         struct hrtimer_sleeper hs;
2736         enum hrtimer_mode mode;
2737         unsigned int nsecs;
2738         ktime_t kt;
2739
2740         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2741                 return false;
2742
2743         /*
2744          * poll_nsec can be:
2745          *
2746          * -1:  don't ever hybrid sleep
2747          *  0:  use half of prev avg
2748          * >0:  use this specific value
2749          */
2750         if (q->poll_nsec == -1)
2751                 return false;
2752         else if (q->poll_nsec > 0)
2753                 nsecs = q->poll_nsec;
2754         else
2755                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2756
2757         if (!nsecs)
2758                 return false;
2759
2760         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2761
2762         /*
2763          * This will be replaced with the stats tracking code, using
2764          * 'avg_completion_time / 2' as the pre-sleep target.
2765          */
2766         kt = nsecs;
2767
2768         mode = HRTIMER_MODE_REL;
2769         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2770         hrtimer_set_expires(&hs.timer, kt);
2771
2772         hrtimer_init_sleeper(&hs, current);
2773         do {
2774                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2775                         break;
2776                 set_current_state(TASK_UNINTERRUPTIBLE);
2777                 hrtimer_start_expires(&hs.timer, mode);
2778                 if (hs.task)
2779                         io_schedule();
2780                 hrtimer_cancel(&hs.timer);
2781                 mode = HRTIMER_MODE_ABS;
2782         } while (hs.task && !signal_pending(current));
2783
2784         __set_current_state(TASK_RUNNING);
2785         destroy_hrtimer_on_stack(&hs.timer);
2786         return true;
2787 }
2788
2789 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2790 {
2791         struct request_queue *q = hctx->queue;
2792         long state;
2793
2794         /*
2795          * If we sleep, have the caller restart the poll loop to reset
2796          * the state. Like for the other success return cases, the
2797          * caller is responsible for checking if the IO completed. If
2798          * the IO isn't complete, we'll get called again and will go
2799          * straight to the busy poll loop.
2800          */
2801         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2802                 return true;
2803
2804         hctx->poll_considered++;
2805
2806         state = current->state;
2807         while (!need_resched()) {
2808                 int ret;
2809
2810                 hctx->poll_invoked++;
2811
2812                 ret = q->mq_ops->poll(hctx, rq->tag);
2813                 if (ret > 0) {
2814                         hctx->poll_success++;
2815                         set_current_state(TASK_RUNNING);
2816                         return true;
2817                 }
2818
2819                 if (signal_pending_state(state, current))
2820                         set_current_state(TASK_RUNNING);
2821
2822                 if (current->state == TASK_RUNNING)
2823                         return true;
2824                 if (ret < 0)
2825                         break;
2826                 cpu_relax();
2827         }
2828
2829         return false;
2830 }
2831
2832 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2833 {
2834         struct blk_mq_hw_ctx *hctx;
2835         struct blk_plug *plug;
2836         struct request *rq;
2837
2838         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2839             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2840                 return false;
2841
2842         plug = current->plug;
2843         if (plug)
2844                 blk_flush_plug_list(plug, false);
2845
2846         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2847         if (!blk_qc_t_is_internal(cookie))
2848                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2849         else
2850                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2851
2852         return __blk_mq_poll(hctx, rq);
2853 }
2854 EXPORT_SYMBOL_GPL(blk_mq_poll);
2855
2856 void blk_mq_disable_hotplug(void)
2857 {
2858         mutex_lock(&all_q_mutex);
2859 }
2860
2861 void blk_mq_enable_hotplug(void)
2862 {
2863         mutex_unlock(&all_q_mutex);
2864 }
2865
2866 static int __init blk_mq_init(void)
2867 {
2868         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2869                                 blk_mq_hctx_notify_dead);
2870
2871         cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2872                                   blk_mq_queue_reinit_prepare,
2873                                   blk_mq_queue_reinit_dead);
2874         return 0;
2875 }
2876 subsys_initcall(blk_mq_init);