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                 default:
 784                         continue;
 785                 }
 786
 787                 if (merged)
 788                         ctx->rq_merged++;
 789                 return merged;
 790         }
 791
 792         return false;
 793 }
 794
 795 struct flush_busy_ctx_data {
 796         struct blk_mq_hw_ctx *hctx;
 797         struct list_head *list;
 798 };
 799
 800 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 801 {
 802         struct flush_busy_ctx_data *flush_data = data;
 803         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 804         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 805
 806         sbitmap_clear_bit(sb, bitnr);
 807         spin_lock(&ctx->lock);
 808         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 809         spin_unlock(&ctx->lock);
 810         return true;
 811 }
 812
 813 /*
 814  * Process software queues that have been marked busy, splicing them
 815  * to the for-dispatch
 816  */
 817 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 818 {
 819         struct flush_busy_ctx_data data = {
 820                 .hctx = hctx,
 821                 .list = list,
 822         };
 823
 824         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 825 }
 826 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 827
 828 static inline unsigned int queued_to_index(unsigned int queued)
 829 {
 830         if (!queued)
 831                 return 0;
 832
 833         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 834 }
 835
 836 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 837                            bool wait)
 838 {
 839         struct blk_mq_alloc_data data = {
 840                 .q = rq->q,
 841                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 842                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 843         };
 844
 845         if (blk_mq_hctx_stopped(data.hctx))
 846                 return false;
 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 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 908 {
 909         struct request_queue *q = hctx->queue;
 910         struct request *rq;
 911         LIST_HEAD(driver_list);
 912         struct list_head *dptr;
 913         int queued, ret = BLK_MQ_RQ_QUEUE_OK;
 914
 915         /*
 916          * Start off with dptr being NULL, so we start the first request
 917          * immediately, even if we have more pending.
 918          */
 919         dptr = NULL;
 920
 921         /*
 922          * Now process all the entries, sending them to the driver.
 923          */
 924         queued = 0;
 925         while (!list_empty(list)) {
 926                 struct blk_mq_queue_data bd;
 927
 928                 rq = list_first_entry(list, struct request, queuelist);
 929                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
 930                         if (!queued && reorder_tags_to_front(list))
 931                                 continue;
 932
 933                         /*
 934                          * We failed getting a driver tag. Mark the queue(s)
 935                          * as needing a restart. Retry getting a tag again,
 936                          * in case the needed IO completed right before we
 937                          * marked the queue as needing a restart.
 938                          */
 939                         blk_mq_sched_mark_restart(hctx);
 940                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
 941                                 break;
 942                 }
 943                 list_del_init(&rq->queuelist);
 944
 945                 bd.rq = rq;
 946                 bd.list = dptr;
 947                 bd.last = list_empty(list);
 948
 949                 ret = q->mq_ops->queue_rq(hctx, &bd);
 950                 switch (ret) {
 951                 case BLK_MQ_RQ_QUEUE_OK:
 952                         queued++;
 953                         break;
 954                 case BLK_MQ_RQ_QUEUE_BUSY:
 955                         blk_mq_put_driver_tag(hctx, rq);
 956                         list_add(&rq->queuelist, list);
 957                         __blk_mq_requeue_request(rq);
 958                         break;
 959                 default:
 960                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 961                 case BLK_MQ_RQ_QUEUE_ERROR:
 962                         rq->errors = -EIO;
 963                         blk_mq_end_request(rq, rq->errors);
 964                         break;
 965                 }
 966
 967                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 968                         break;
 969
 970                 /*
 971                  * We've done the first request. If we have more than 1
 972                  * left in the list, set dptr to defer issue.
 973                  */
 974                 if (!dptr && list->next != list->prev)
 975                         dptr = &driver_list;
 976         }
 977
 978         hctx->dispatched[queued_to_index(queued)]++;
 979
 980         /*
 981          * Any items that need requeuing? Stuff them into hctx->dispatch,
 982          * that is where we will continue on next queue run.
 983          */
 984         if (!list_empty(list)) {
 985                 spin_lock(&hctx->lock);
 986                 list_splice_init(list, &hctx->dispatch);
 987                 spin_unlock(&hctx->lock);
 988
 989                 /*
 990                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 991                  * it's possible the queue is stopped and restarted again
 992                  * before this. Queue restart will dispatch requests. And since
 993                  * requests in rq_list aren't added into hctx->dispatch yet,
 994                  * the requests in rq_list might get lost.
 995                  *
 996                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 997                  *
 998                  * If RESTART is set, then let completion restart the queue
 999                  * instead of potentially looping here.
1000                  */
1001                 if (!blk_mq_sched_needs_restart(hctx))
1002                         blk_mq_run_hw_queue(hctx, true);
1003         }
1004
1005         return ret != BLK_MQ_RQ_QUEUE_BUSY;
1006 }
1007
1008 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1009 {
1010         int srcu_idx;
1011
1012         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1013                 cpu_online(hctx->next_cpu));
1014
1015         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1016                 rcu_read_lock();
1017                 blk_mq_sched_dispatch_requests(hctx);
1018                 rcu_read_unlock();
1019         } else {
1020                 srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu);
1021                 blk_mq_sched_dispatch_requests(hctx);
1022                 srcu_read_unlock(&hctx->queue_rq_srcu, srcu_idx);
1023         }
1024 }
1025
1026 /*
1027  * It'd be great if the workqueue API had a way to pass
1028  * in a mask and had some smarts for more clever placement.
1029  * For now we just round-robin here, switching for every
1030  * BLK_MQ_CPU_WORK_BATCH queued items.
1031  */
1032 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1033 {
1034         if (hctx->queue->nr_hw_queues == 1)
1035                 return WORK_CPU_UNBOUND;
1036
1037         if (--hctx->next_cpu_batch <= 0) {
1038                 int next_cpu;
1039
1040                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1041                 if (next_cpu >= nr_cpu_ids)
1042                         next_cpu = cpumask_first(hctx->cpumask);
1043
1044                 hctx->next_cpu = next_cpu;
1045                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1046         }
1047
1048         return hctx->next_cpu;
1049 }
1050
1051 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1052 {
1053         if (unlikely(blk_mq_hctx_stopped(hctx) ||
1054                      !blk_mq_hw_queue_mapped(hctx)))
1055                 return;
1056
1057         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1058                 int cpu = get_cpu();
1059                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1060                         __blk_mq_run_hw_queue(hctx);
1061                         put_cpu();
1062                         return;
1063                 }
1064
1065                 put_cpu();
1066         }
1067
1068         kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
1069 }
1070
1071 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1072 {
1073         struct blk_mq_hw_ctx *hctx;
1074         int i;
1075
1076         queue_for_each_hw_ctx(q, hctx, i) {
1077                 if (!blk_mq_hctx_has_pending(hctx) ||
1078                     blk_mq_hctx_stopped(hctx))
1079                         continue;
1080
1081                 blk_mq_run_hw_queue(hctx, async);
1082         }
1083 }
1084 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1085
1086 /**
1087  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1088  * @q: request queue.
1089  *
1090  * The caller is responsible for serializing this function against
1091  * blk_mq_{start,stop}_hw_queue().
1092  */
1093 bool blk_mq_queue_stopped(struct request_queue *q)
1094 {
1095         struct blk_mq_hw_ctx *hctx;
1096         int i;
1097
1098         queue_for_each_hw_ctx(q, hctx, i)
1099                 if (blk_mq_hctx_stopped(hctx))
1100                         return true;
1101
1102         return false;
1103 }
1104 EXPORT_SYMBOL(blk_mq_queue_stopped);
1105
1106 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1107 {
1108         cancel_work(&hctx->run_work);
1109         cancel_delayed_work(&hctx->delay_work);
1110         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1111 }
1112 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1113
1114 void blk_mq_stop_hw_queues(struct request_queue *q)
1115 {
1116         struct blk_mq_hw_ctx *hctx;
1117         int i;
1118
1119         queue_for_each_hw_ctx(q, hctx, i)
1120                 blk_mq_stop_hw_queue(hctx);
1121 }
1122 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1123
1124 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1125 {
1126         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1127
1128         blk_mq_run_hw_queue(hctx, false);
1129 }
1130 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1131
1132 void blk_mq_start_hw_queues(struct request_queue *q)
1133 {
1134         struct blk_mq_hw_ctx *hctx;
1135         int i;
1136
1137         queue_for_each_hw_ctx(q, hctx, i)
1138                 blk_mq_start_hw_queue(hctx);
1139 }
1140 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1141
1142 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1143 {
1144         if (!blk_mq_hctx_stopped(hctx))
1145                 return;
1146
1147         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1148         blk_mq_run_hw_queue(hctx, async);
1149 }
1150 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1151
1152 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1153 {
1154         struct blk_mq_hw_ctx *hctx;
1155         int i;
1156
1157         queue_for_each_hw_ctx(q, hctx, i)
1158                 blk_mq_start_stopped_hw_queue(hctx, async);
1159 }
1160 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1161
1162 static void blk_mq_run_work_fn(struct work_struct *work)
1163 {
1164         struct blk_mq_hw_ctx *hctx;
1165
1166         hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1167
1168         __blk_mq_run_hw_queue(hctx);
1169 }
1170
1171 static void blk_mq_delay_work_fn(struct work_struct *work)
1172 {
1173         struct blk_mq_hw_ctx *hctx;
1174
1175         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1176
1177         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1178                 __blk_mq_run_hw_queue(hctx);
1179 }
1180
1181 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1182 {
1183         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1184                 return;
1185
1186         blk_mq_stop_hw_queue(hctx);
1187         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1188                         &hctx->delay_work, msecs_to_jiffies(msecs));
1189 }
1190 EXPORT_SYMBOL(blk_mq_delay_queue);
1191
1192 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1193                                             struct request *rq,
1194                                             bool at_head)
1195 {
1196         struct blk_mq_ctx *ctx = rq->mq_ctx;
1197
1198         trace_block_rq_insert(hctx->queue, rq);
1199
1200         if (at_head)
1201                 list_add(&rq->queuelist, &ctx->rq_list);
1202         else
1203                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1204 }
1205
1206 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1207                              bool at_head)
1208 {
1209         struct blk_mq_ctx *ctx = rq->mq_ctx;
1210
1211         __blk_mq_insert_req_list(hctx, rq, at_head);
1212         blk_mq_hctx_mark_pending(hctx, ctx);
1213 }
1214
1215 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1216                             struct list_head *list)
1217
1218 {
1219         /*
1220          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1221          * offline now
1222          */
1223         spin_lock(&ctx->lock);
1224         while (!list_empty(list)) {
1225                 struct request *rq;
1226
1227                 rq = list_first_entry(list, struct request, queuelist);
1228                 BUG_ON(rq->mq_ctx != ctx);
1229                 list_del_init(&rq->queuelist);
1230                 __blk_mq_insert_req_list(hctx, rq, false);
1231         }
1232         blk_mq_hctx_mark_pending(hctx, ctx);
1233         spin_unlock(&ctx->lock);
1234 }
1235
1236 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1237 {
1238         struct request *rqa = container_of(a, struct request, queuelist);
1239         struct request *rqb = container_of(b, struct request, queuelist);
1240
1241         return !(rqa->mq_ctx < rqb->mq_ctx ||
1242                  (rqa->mq_ctx == rqb->mq_ctx &&
1243                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1244 }
1245
1246 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1247 {
1248         struct blk_mq_ctx *this_ctx;
1249         struct request_queue *this_q;
1250         struct request *rq;
1251         LIST_HEAD(list);
1252         LIST_HEAD(ctx_list);
1253         unsigned int depth;
1254
1255         list_splice_init(&plug->mq_list, &list);
1256
1257         list_sort(NULL, &list, plug_ctx_cmp);
1258
1259         this_q = NULL;
1260         this_ctx = NULL;
1261         depth = 0;
1262
1263         while (!list_empty(&list)) {
1264                 rq = list_entry_rq(list.next);
1265                 list_del_init(&rq->queuelist);
1266                 BUG_ON(!rq->q);
1267                 if (rq->mq_ctx != this_ctx) {
1268                         if (this_ctx) {
1269                                 trace_block_unplug(this_q, depth, from_schedule);
1270                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1271                                                                 &ctx_list,
1272                                                                 from_schedule);
1273                         }
1274
1275                         this_ctx = rq->mq_ctx;
1276                         this_q = rq->q;
1277                         depth = 0;
1278                 }
1279
1280                 depth++;
1281                 list_add_tail(&rq->queuelist, &ctx_list);
1282         }
1283
1284         /*
1285          * If 'this_ctx' is set, we know we have entries to complete
1286          * on 'ctx_list'. Do those.
1287          */
1288         if (this_ctx) {
1289                 trace_block_unplug(this_q, depth, from_schedule);
1290                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1291                                                 from_schedule);
1292         }
1293 }
1294
1295 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1296 {
1297         init_request_from_bio(rq, bio);
1298
1299         blk_account_io_start(rq, true);
1300 }
1301
1302 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1303 {
1304         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1305                 !blk_queue_nomerges(hctx->queue);
1306 }
1307
1308 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1309                                          struct blk_mq_ctx *ctx,
1310                                          struct request *rq, struct bio *bio)
1311 {
1312         if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1313                 blk_mq_bio_to_request(rq, bio);
1314                 spin_lock(&ctx->lock);
1315 insert_rq:
1316                 __blk_mq_insert_request(hctx, rq, false);
1317                 spin_unlock(&ctx->lock);
1318                 return false;
1319         } else {
1320                 struct request_queue *q = hctx->queue;
1321
1322                 spin_lock(&ctx->lock);
1323                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1324                         blk_mq_bio_to_request(rq, bio);
1325                         goto insert_rq;
1326                 }
1327
1328                 spin_unlock(&ctx->lock);
1329                 __blk_mq_finish_request(hctx, ctx, rq);
1330                 return true;
1331         }
1332 }
1333
1334 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1335 {
1336         if (rq->tag != -1)
1337                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1338
1339         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1340 }
1341
1342 static void blk_mq_try_issue_directly(struct request *rq, blk_qc_t *cookie)
1343 {
1344         struct request_queue *q = rq->q;
1345         struct blk_mq_queue_data bd = {
1346                 .rq = rq,
1347                 .list = NULL,
1348                 .last = 1
1349         };
1350         struct blk_mq_hw_ctx *hctx;
1351         blk_qc_t new_cookie;
1352         int ret;
1353
1354         if (q->elevator)
1355                 goto insert;
1356
1357         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1358                 goto insert;
1359
1360         new_cookie = request_to_qc_t(hctx, rq);
1361
1362         /*
1363          * For OK queue, we are done. For error, kill it. Any other
1364          * error (busy), just add it to our list as we previously
1365          * would have done
1366          */
1367         ret = q->mq_ops->queue_rq(hctx, &bd);
1368         if (ret == BLK_MQ_RQ_QUEUE_OK) {
1369                 *cookie = new_cookie;
1370                 return;
1371         }
1372
1373         __blk_mq_requeue_request(rq);
1374
1375         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1376                 *cookie = BLK_QC_T_NONE;
1377                 rq->errors = -EIO;
1378                 blk_mq_end_request(rq, rq->errors);
1379                 return;
1380         }
1381
1382 insert:
1383         blk_mq_sched_insert_request(rq, false, true, true, false);
1384 }
1385
1386 /*
1387  * Multiple hardware queue variant. This will not use per-process plugs,
1388  * but will attempt to bypass the hctx queueing if we can go straight to
1389  * hardware for SYNC IO.
1390  */
1391 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1392 {
1393         const int is_sync = op_is_sync(bio->bi_opf);
1394         const int is_flush_fua = op_is_flush(bio->bi_opf);
1395         struct blk_mq_alloc_data data = { .flags = 0 };
1396         struct request *rq;
1397         unsigned int request_count = 0, srcu_idx;
1398         struct blk_plug *plug;
1399         struct request *same_queue_rq = NULL;
1400         blk_qc_t cookie;
1401         unsigned int wb_acct;
1402
1403         blk_queue_bounce(q, &bio);
1404
1405         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1406                 bio_io_error(bio);
1407                 return BLK_QC_T_NONE;
1408         }
1409
1410         blk_queue_split(q, &bio, q->bio_split);
1411
1412         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1413             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1414                 return BLK_QC_T_NONE;
1415
1416         if (blk_mq_sched_bio_merge(q, bio))
1417                 return BLK_QC_T_NONE;
1418
1419         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1420
1421         trace_block_getrq(q, bio, bio->bi_opf);
1422
1423         rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1424         if (unlikely(!rq)) {
1425                 __wbt_done(q->rq_wb, wb_acct);
1426                 return BLK_QC_T_NONE;
1427         }
1428
1429         wbt_track(&rq->issue_stat, wb_acct);
1430
1431         cookie = request_to_qc_t(data.hctx, rq);
1432
1433         if (unlikely(is_flush_fua)) {
1434                 blk_mq_put_ctx(data.ctx);
1435                 blk_mq_bio_to_request(rq, bio);
1436                 blk_mq_get_driver_tag(rq, NULL, true);
1437                 blk_insert_flush(rq);
1438                 blk_mq_run_hw_queue(data.hctx, true);
1439                 goto done;
1440         }
1441
1442         plug = current->plug;
1443         /*
1444          * If the driver supports defer issued based on 'last', then
1445          * queue it up like normal since we can potentially save some
1446          * CPU this way.
1447          */
1448         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1449             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1450                 struct request *old_rq = NULL;
1451
1452                 blk_mq_bio_to_request(rq, bio);
1453
1454                 /*
1455                  * We do limited plugging. If the bio can be merged, do that.
1456                  * Otherwise the existing request in the plug list will be
1457                  * issued. So the plug list will have one request at most
1458                  */
1459                 if (plug) {
1460                         /*
1461                          * The plug list might get flushed before this. If that
1462                          * happens, same_queue_rq is invalid and plug list is
1463                          * empty
1464                          */
1465                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1466                                 old_rq = same_queue_rq;
1467                                 list_del_init(&old_rq->queuelist);
1468                         }
1469                         list_add_tail(&rq->queuelist, &plug->mq_list);
1470                 } else /* is_sync */
1471                         old_rq = rq;
1472                 blk_mq_put_ctx(data.ctx);
1473                 if (!old_rq)
1474                         goto done;
1475
1476                 if (!(data.hctx->flags & BLK_MQ_F_BLOCKING)) {
1477                         rcu_read_lock();
1478                         blk_mq_try_issue_directly(old_rq, &cookie);
1479                         rcu_read_unlock();
1480                 } else {
1481                         srcu_idx = srcu_read_lock(&data.hctx->queue_rq_srcu);
1482                         blk_mq_try_issue_directly(old_rq, &cookie);
1483                         srcu_read_unlock(&data.hctx->queue_rq_srcu, srcu_idx);
1484                 }
1485                 goto done;
1486         }
1487
1488         if (q->elevator) {
1489                 blk_mq_put_ctx(data.ctx);
1490                 blk_mq_bio_to_request(rq, bio);
1491                 blk_mq_sched_insert_request(rq, false, true,
1492                                                 !is_sync || is_flush_fua, true);
1493                 goto done;
1494         }
1495         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1496                 /*
1497                  * For a SYNC request, send it to the hardware immediately. For
1498                  * an ASYNC request, just ensure that we run it later on. The
1499                  * latter allows for merging opportunities and more efficient
1500                  * dispatching.
1501                  */
1502                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1503         }
1504         blk_mq_put_ctx(data.ctx);
1505 done:
1506         return cookie;
1507 }
1508
1509 /*
1510  * Single hardware queue variant. This will attempt to use any per-process
1511  * plug for merging and IO deferral.
1512  */
1513 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1514 {
1515         const int is_sync = op_is_sync(bio->bi_opf);
1516         const int is_flush_fua = op_is_flush(bio->bi_opf);
1517         struct blk_plug *plug;
1518         unsigned int request_count = 0;
1519         struct blk_mq_alloc_data data = { .flags = 0 };
1520         struct request *rq;
1521         blk_qc_t cookie;
1522         unsigned int wb_acct;
1523
1524         blk_queue_bounce(q, &bio);
1525
1526         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1527                 bio_io_error(bio);
1528                 return BLK_QC_T_NONE;
1529         }
1530
1531         blk_queue_split(q, &bio, q->bio_split);
1532
1533         if (!is_flush_fua && !blk_queue_nomerges(q)) {
1534                 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1535                         return BLK_QC_T_NONE;
1536         } else
1537                 request_count = blk_plug_queued_count(q);
1538
1539         if (blk_mq_sched_bio_merge(q, bio))
1540                 return BLK_QC_T_NONE;
1541
1542         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1543
1544         trace_block_getrq(q, bio, bio->bi_opf);
1545
1546         rq = blk_mq_sched_get_request(q, bio, bio->bi_opf, &data);
1547         if (unlikely(!rq)) {
1548                 __wbt_done(q->rq_wb, wb_acct);
1549                 return BLK_QC_T_NONE;
1550         }
1551
1552         wbt_track(&rq->issue_stat, wb_acct);
1553
1554         cookie = request_to_qc_t(data.hctx, rq);
1555
1556         if (unlikely(is_flush_fua)) {
1557                 blk_mq_put_ctx(data.ctx);
1558                 blk_mq_bio_to_request(rq, bio);
1559                 blk_mq_get_driver_tag(rq, NULL, true);
1560                 blk_insert_flush(rq);
1561                 blk_mq_run_hw_queue(data.hctx, true);
1562                 goto done;
1563         }
1564
1565         /*
1566          * A task plug currently exists. Since this is completely lockless,
1567          * utilize that to temporarily store requests until the task is
1568          * either done or scheduled away.
1569          */
1570         plug = current->plug;
1571         if (plug) {
1572                 struct request *last = NULL;
1573
1574                 blk_mq_bio_to_request(rq, bio);
1575
1576                 /*
1577                  * @request_count may become stale because of schedule
1578                  * out, so check the list again.
1579                  */
1580                 if (list_empty(&plug->mq_list))
1581                         request_count = 0;
1582                 if (!request_count)
1583                         trace_block_plug(q);
1584                 else
1585                         last = list_entry_rq(plug->mq_list.prev);
1586
1587                 blk_mq_put_ctx(data.ctx);
1588
1589                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1590                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1591                         blk_flush_plug_list(plug, false);
1592                         trace_block_plug(q);
1593                 }
1594
1595                 list_add_tail(&rq->queuelist, &plug->mq_list);
1596                 return cookie;
1597         }
1598
1599         if (q->elevator) {
1600                 blk_mq_put_ctx(data.ctx);
1601                 blk_mq_bio_to_request(rq, bio);
1602                 blk_mq_sched_insert_request(rq, false, true,
1603                                                 !is_sync || is_flush_fua, true);
1604                 goto done;
1605         }
1606         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1607                 /*
1608                  * For a SYNC request, send it to the hardware immediately. For
1609                  * an ASYNC request, just ensure that we run it later on. The
1610                  * latter allows for merging opportunities and more efficient
1611                  * dispatching.
1612                  */
1613                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1614         }
1615
1616         blk_mq_put_ctx(data.ctx);
1617 done:
1618         return cookie;
1619 }
1620
1621 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1622                      unsigned int hctx_idx)
1623 {
1624         struct page *page;
1625
1626         if (tags->rqs && set->ops->exit_request) {
1627                 int i;
1628
1629                 for (i = 0; i < tags->nr_tags; i++) {
1630                         struct request *rq = tags->static_rqs[i];
1631
1632                         if (!rq)
1633                                 continue;
1634                         set->ops->exit_request(set->driver_data, rq,
1635                                                 hctx_idx, i);
1636                         tags->static_rqs[i] = NULL;
1637                 }
1638         }
1639
1640         while (!list_empty(&tags->page_list)) {
1641                 page = list_first_entry(&tags->page_list, struct page, lru);
1642                 list_del_init(&page->lru);
1643                 /*
1644                  * Remove kmemleak object previously allocated in
1645                  * blk_mq_init_rq_map().
1646                  */
1647                 kmemleak_free(page_address(page));
1648                 __free_pages(page, page->private);
1649         }
1650 }
1651
1652 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1653 {
1654         kfree(tags->rqs);
1655         tags->rqs = NULL;
1656         kfree(tags->static_rqs);
1657         tags->static_rqs = NULL;
1658
1659         blk_mq_free_tags(tags);
1660 }
1661
1662 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1663                                         unsigned int hctx_idx,
1664                                         unsigned int nr_tags,
1665                                         unsigned int reserved_tags)
1666 {
1667         struct blk_mq_tags *tags;
1668
1669         tags = blk_mq_init_tags(nr_tags, reserved_tags,
1670                                 set->numa_node,
1671                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1672         if (!tags)
1673                 return NULL;
1674
1675         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1676                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1677                                  set->numa_node);
1678         if (!tags->rqs) {
1679                 blk_mq_free_tags(tags);
1680                 return NULL;
1681         }
1682
1683         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1684                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1685                                  set->numa_node);
1686         if (!tags->static_rqs) {
1687                 kfree(tags->rqs);
1688                 blk_mq_free_tags(tags);
1689                 return NULL;
1690         }
1691
1692         return tags;
1693 }
1694
1695 static size_t order_to_size(unsigned int order)
1696 {
1697         return (size_t)PAGE_SIZE << order;
1698 }
1699
1700 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1701                      unsigned int hctx_idx, unsigned int depth)
1702 {
1703         unsigned int i, j, entries_per_page, max_order = 4;
1704         size_t rq_size, left;
1705
1706         INIT_LIST_HEAD(&tags->page_list);
1707
1708         /*
1709          * rq_size is the size of the request plus driver payload, rounded
1710          * to the cacheline size
1711          */
1712         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1713                                 cache_line_size());
1714         left = rq_size * depth;
1715
1716         for (i = 0; i < depth; ) {
1717                 int this_order = max_order;
1718                 struct page *page;
1719                 int to_do;
1720                 void *p;
1721
1722                 while (this_order && left < order_to_size(this_order - 1))
1723                         this_order--;
1724
1725                 do {
1726                         page = alloc_pages_node(set->numa_node,
1727                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1728                                 this_order);
1729                         if (page)
1730                                 break;
1731                         if (!this_order--)
1732                                 break;
1733                         if (order_to_size(this_order) < rq_size)
1734                                 break;
1735                 } while (1);
1736
1737                 if (!page)
1738                         goto fail;
1739
1740                 page->private = this_order;
1741                 list_add_tail(&page->lru, &tags->page_list);
1742
1743                 p = page_address(page);
1744                 /*
1745                  * Allow kmemleak to scan these pages as they contain pointers
1746                  * to additional allocations like via ops->init_request().
1747                  */
1748                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1749                 entries_per_page = order_to_size(this_order) / rq_size;
1750                 to_do = min(entries_per_page, depth - i);
1751                 left -= to_do * rq_size;
1752                 for (j = 0; j < to_do; j++) {
1753                         struct request *rq = p;
1754
1755                         tags->static_rqs[i] = rq;
1756                         if (set->ops->init_request) {
1757                                 if (set->ops->init_request(set->driver_data,
1758                                                 rq, hctx_idx, i,
1759                                                 set->numa_node)) {
1760                                         tags->static_rqs[i] = NULL;
1761                                         goto fail;
1762                                 }
1763                         }
1764
1765                         p += rq_size;
1766                         i++;
1767                 }
1768         }
1769         return 0;
1770
1771 fail:
1772         blk_mq_free_rqs(set, tags, hctx_idx);
1773         return -ENOMEM;
1774 }
1775
1776 /*
1777  * 'cpu' is going away. splice any existing rq_list entries from this
1778  * software queue to the hw queue dispatch list, and ensure that it
1779  * gets run.
1780  */
1781 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1782 {
1783         struct blk_mq_hw_ctx *hctx;
1784         struct blk_mq_ctx *ctx;
1785         LIST_HEAD(tmp);
1786
1787         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1788         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1789
1790         spin_lock(&ctx->lock);
1791         if (!list_empty(&ctx->rq_list)) {
1792                 list_splice_init(&ctx->rq_list, &tmp);
1793                 blk_mq_hctx_clear_pending(hctx, ctx);
1794         }
1795         spin_unlock(&ctx->lock);
1796
1797         if (list_empty(&tmp))
1798                 return 0;
1799
1800         spin_lock(&hctx->lock);
1801         list_splice_tail_init(&tmp, &hctx->dispatch);
1802         spin_unlock(&hctx->lock);
1803
1804         blk_mq_run_hw_queue(hctx, true);
1805         return 0;
1806 }
1807
1808 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1809 {
1810         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1811                                             &hctx->cpuhp_dead);
1812 }
1813
1814 /* hctx->ctxs will be freed in queue's release handler */
1815 static void blk_mq_exit_hctx(struct request_queue *q,
1816                 struct blk_mq_tag_set *set,
1817                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1818 {
1819         unsigned flush_start_tag = set->queue_depth;
1820
1821         blk_mq_tag_idle(hctx);
1822
1823         if (set->ops->exit_request)
1824                 set->ops->exit_request(set->driver_data,
1825                                        hctx->fq->flush_rq, hctx_idx,
1826                                        flush_start_tag + hctx_idx);
1827
1828         if (set->ops->exit_hctx)
1829                 set->ops->exit_hctx(hctx, hctx_idx);
1830
1831         if (hctx->flags & BLK_MQ_F_BLOCKING)
1832                 cleanup_srcu_struct(&hctx->queue_rq_srcu);
1833
1834         blk_mq_remove_cpuhp(hctx);
1835         blk_free_flush_queue(hctx->fq);
1836         sbitmap_free(&hctx->ctx_map);
1837 }
1838
1839 static void blk_mq_exit_hw_queues(struct request_queue *q,
1840                 struct blk_mq_tag_set *set, int nr_queue)
1841 {
1842         struct blk_mq_hw_ctx *hctx;
1843         unsigned int i;
1844
1845         queue_for_each_hw_ctx(q, hctx, i) {
1846                 if (i == nr_queue)
1847                         break;
1848                 blk_mq_exit_hctx(q, set, hctx, i);
1849         }
1850 }
1851
1852 static void blk_mq_free_hw_queues(struct request_queue *q,
1853                 struct blk_mq_tag_set *set)
1854 {
1855         struct blk_mq_hw_ctx *hctx;
1856         unsigned int i;
1857
1858         queue_for_each_hw_ctx(q, hctx, i)
1859                 free_cpumask_var(hctx->cpumask);
1860 }
1861
1862 static int blk_mq_init_hctx(struct request_queue *q,
1863                 struct blk_mq_tag_set *set,
1864                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1865 {
1866         int node;
1867         unsigned flush_start_tag = set->queue_depth;
1868
1869         node = hctx->numa_node;
1870         if (node == NUMA_NO_NODE)
1871                 node = hctx->numa_node = set->numa_node;
1872
1873         INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1874         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1875         spin_lock_init(&hctx->lock);
1876         INIT_LIST_HEAD(&hctx->dispatch);
1877         hctx->queue = q;
1878         hctx->queue_num = hctx_idx;
1879         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1880
1881         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1882
1883         hctx->tags = set->tags[hctx_idx];
1884
1885         /*
1886          * Allocate space for all possible cpus to avoid allocation at
1887          * runtime
1888          */
1889         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1890                                         GFP_KERNEL, node);
1891         if (!hctx->ctxs)
1892                 goto unregister_cpu_notifier;
1893
1894         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1895                               node))
1896                 goto free_ctxs;
1897
1898         hctx->nr_ctx = 0;
1899
1900         if (set->ops->init_hctx &&
1901             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1902                 goto free_bitmap;
1903
1904         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1905         if (!hctx->fq)
1906                 goto exit_hctx;
1907
1908         if (set->ops->init_request &&
1909             set->ops->init_request(set->driver_data,
1910                                    hctx->fq->flush_rq, hctx_idx,
1911                                    flush_start_tag + hctx_idx, node))
1912                 goto free_fq;
1913
1914         if (hctx->flags & BLK_MQ_F_BLOCKING)
1915                 init_srcu_struct(&hctx->queue_rq_srcu);
1916
1917         return 0;
1918
1919  free_fq:
1920         kfree(hctx->fq);
1921  exit_hctx:
1922         if (set->ops->exit_hctx)
1923                 set->ops->exit_hctx(hctx, hctx_idx);
1924  free_bitmap:
1925         sbitmap_free(&hctx->ctx_map);
1926  free_ctxs:
1927         kfree(hctx->ctxs);
1928  unregister_cpu_notifier:
1929         blk_mq_remove_cpuhp(hctx);
1930         return -1;
1931 }
1932
1933 static void blk_mq_init_cpu_queues(struct request_queue *q,
1934                                    unsigned int nr_hw_queues)
1935 {
1936         unsigned int i;
1937
1938         for_each_possible_cpu(i) {
1939                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1940                 struct blk_mq_hw_ctx *hctx;
1941
1942                 memset(__ctx, 0, sizeof(*__ctx));
1943                 __ctx->cpu = i;
1944                 spin_lock_init(&__ctx->lock);
1945                 INIT_LIST_HEAD(&__ctx->rq_list);
1946                 __ctx->queue = q;
1947                 blk_stat_init(&__ctx->stat[BLK_STAT_READ]);
1948                 blk_stat_init(&__ctx->stat[BLK_STAT_WRITE]);
1949
1950                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1951                 if (!cpu_online(i))
1952                         continue;
1953
1954                 hctx = blk_mq_map_queue(q, i);
1955
1956                 /*
1957                  * Set local node, IFF we have more than one hw queue. If
1958                  * not, we remain on the home node of the device
1959                  */
1960                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1961                         hctx->numa_node = local_memory_node(cpu_to_node(i));
1962         }
1963 }
1964
1965 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
1966 {
1967         int ret = 0;
1968
1969         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
1970                                         set->queue_depth, set->reserved_tags);
1971         if (!set->tags[hctx_idx])
1972                 return false;
1973
1974         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
1975                                 set->queue_depth);
1976         if (!ret)
1977                 return true;
1978
1979         blk_mq_free_rq_map(set->tags[hctx_idx]);
1980         set->tags[hctx_idx] = NULL;
1981         return false;
1982 }
1983
1984 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
1985                                          unsigned int hctx_idx)
1986 {
1987         if (set->tags[hctx_idx]) {
1988                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
1989                 blk_mq_free_rq_map(set->tags[hctx_idx]);
1990                 set->tags[hctx_idx] = NULL;
1991         }
1992 }
1993
1994 static void blk_mq_map_swqueue(struct request_queue *q,
1995                                const struct cpumask *online_mask)
1996 {
1997         unsigned int i, hctx_idx;
1998         struct blk_mq_hw_ctx *hctx;
1999         struct blk_mq_ctx *ctx;
2000         struct blk_mq_tag_set *set = q->tag_set;
2001
2002         /*
2003          * Avoid others reading imcomplete hctx->cpumask through sysfs
2004          */
2005         mutex_lock(&q->sysfs_lock);
2006
2007         queue_for_each_hw_ctx(q, hctx, i) {
2008                 cpumask_clear(hctx->cpumask);
2009                 hctx->nr_ctx = 0;
2010         }
2011
2012         /*
2013          * Map software to hardware queues
2014          */
2015         for_each_possible_cpu(i) {
2016                 /* If the cpu isn't online, the cpu is mapped to first hctx */
2017                 if (!cpumask_test_cpu(i, online_mask))
2018                         continue;
2019
2020                 hctx_idx = q->mq_map[i];
2021                 /* unmapped hw queue can be remapped after CPU topo changed */
2022                 if (!set->tags[hctx_idx] &&
2023                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2024                         /*
2025                          * If tags initialization fail for some hctx,
2026                          * that hctx won't be brought online.  In this
2027                          * case, remap the current ctx to hctx[0] which
2028                          * is guaranteed to always have tags allocated
2029                          */
2030                         q->mq_map[i] = 0;
2031                 }
2032
2033                 ctx = per_cpu_ptr(q->queue_ctx, i);
2034                 hctx = blk_mq_map_queue(q, i);
2035
2036                 cpumask_set_cpu(i, hctx->cpumask);
2037                 ctx->index_hw = hctx->nr_ctx;
2038                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2039         }
2040
2041         mutex_unlock(&q->sysfs_lock);
2042
2043         queue_for_each_hw_ctx(q, hctx, i) {
2044                 /*
2045                  * If no software queues are mapped to this hardware queue,
2046                  * disable it and free the request entries.
2047                  */
2048                 if (!hctx->nr_ctx) {
2049                         /* Never unmap queue 0.  We need it as a
2050                          * fallback in case of a new remap fails
2051                          * allocation
2052                          */
2053                         if (i && set->tags[i])
2054                                 blk_mq_free_map_and_requests(set, i);
2055
2056                         hctx->tags = NULL;
2057                         continue;
2058                 }
2059
2060                 hctx->tags = set->tags[i];
2061                 WARN_ON(!hctx->tags);
2062
2063                 /*
2064                  * Set the map size to the number of mapped software queues.
2065                  * This is more accurate and more efficient than looping
2066                  * over all possibly mapped software queues.
2067                  */
2068                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2069
2070                 /*
2071                  * Initialize batch roundrobin counts
2072                  */
2073                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2074                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2075         }
2076 }
2077
2078 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2079 {
2080         struct blk_mq_hw_ctx *hctx;
2081         int i;
2082
2083         queue_for_each_hw_ctx(q, hctx, i) {
2084                 if (shared)
2085                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2086                 else
2087                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2088         }
2089 }
2090
2091 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
2092 {
2093         struct request_queue *q;
2094
2095         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2096                 blk_mq_freeze_queue(q);
2097                 queue_set_hctx_shared(q, shared);
2098                 blk_mq_unfreeze_queue(q);
2099         }
2100 }
2101
2102 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2103 {
2104         struct blk_mq_tag_set *set = q->tag_set;
2105
2106         mutex_lock(&set->tag_list_lock);
2107         list_del_init(&q->tag_set_list);
2108         if (list_is_singular(&set->tag_list)) {
2109                 /* just transitioned to unshared */
2110                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2111                 /* update existing queue */
2112                 blk_mq_update_tag_set_depth(set, false);
2113         }
2114         mutex_unlock(&set->tag_list_lock);
2115 }
2116
2117 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2118                                      struct request_queue *q)
2119 {
2120         q->tag_set = set;
2121
2122         mutex_lock(&set->tag_list_lock);
2123
2124         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2125         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2126                 set->flags |= BLK_MQ_F_TAG_SHARED;
2127                 /* update existing queue */
2128                 blk_mq_update_tag_set_depth(set, true);
2129         }
2130         if (set->flags & BLK_MQ_F_TAG_SHARED)
2131                 queue_set_hctx_shared(q, true);
2132         list_add_tail(&q->tag_set_list, &set->tag_list);
2133
2134         mutex_unlock(&set->tag_list_lock);
2135 }
2136
2137 /*
2138  * It is the actual release handler for mq, but we do it from
2139  * request queue's release handler for avoiding use-after-free
2140  * and headache because q->mq_kobj shouldn't have been introduced,
2141  * but we can't group ctx/kctx kobj without it.
2142  */
2143 void blk_mq_release(struct request_queue *q)
2144 {
2145         struct blk_mq_hw_ctx *hctx;
2146         unsigned int i;
2147
2148         blk_mq_sched_teardown(q);
2149
2150         /* hctx kobj stays in hctx */
2151         queue_for_each_hw_ctx(q, hctx, i) {
2152                 if (!hctx)
2153                         continue;
2154                 kfree(hctx->ctxs);
2155                 kfree(hctx);
2156         }
2157
2158         q->mq_map = NULL;
2159
2160         kfree(q->queue_hw_ctx);
2161
2162         /* ctx kobj stays in queue_ctx */
2163         free_percpu(q->queue_ctx);
2164 }
2165
2166 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2167 {
2168         struct request_queue *uninit_q, *q;
2169
2170         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2171         if (!uninit_q)
2172                 return ERR_PTR(-ENOMEM);
2173
2174         q = blk_mq_init_allocated_queue(set, uninit_q);
2175         if (IS_ERR(q))
2176                 blk_cleanup_queue(uninit_q);
2177
2178         return q;
2179 }
2180 EXPORT_SYMBOL(blk_mq_init_queue);
2181
2182 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2183                                                 struct request_queue *q)
2184 {
2185         int i, j;
2186         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2187
2188         blk_mq_sysfs_unregister(q);
2189         for (i = 0; i < set->nr_hw_queues; i++) {
2190                 int node;
2191
2192                 if (hctxs[i])
2193                         continue;
2194
2195                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2196                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
2197                                         GFP_KERNEL, node);
2198                 if (!hctxs[i])
2199                         break;
2200
2201                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2202                                                 node)) {
2203                         kfree(hctxs[i]);
2204                         hctxs[i] = NULL;
2205                         break;
2206                 }
2207
2208                 atomic_set(&hctxs[i]->nr_active, 0);
2209                 hctxs[i]->numa_node = node;
2210                 hctxs[i]->queue_num = i;
2211
2212                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2213                         free_cpumask_var(hctxs[i]->cpumask);
2214                         kfree(hctxs[i]);
2215                         hctxs[i] = NULL;
2216                         break;
2217                 }
2218                 blk_mq_hctx_kobj_init(hctxs[i]);
2219         }
2220         for (j = i; j < q->nr_hw_queues; j++) {
2221                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2222
2223                 if (hctx) {
2224                         if (hctx->tags)
2225                                 blk_mq_free_map_and_requests(set, j);
2226                         blk_mq_exit_hctx(q, set, hctx, j);
2227                         free_cpumask_var(hctx->cpumask);
2228                         kobject_put(&hctx->kobj);
2229                         kfree(hctx->ctxs);
2230                         kfree(hctx);
2231                         hctxs[j] = NULL;
2232
2233                 }
2234         }
2235         q->nr_hw_queues = i;
2236         blk_mq_sysfs_register(q);
2237 }
2238
2239 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2240                                                   struct request_queue *q)
2241 {
2242         /* mark the queue as mq asap */
2243         q->mq_ops = set->ops;
2244
2245         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2246         if (!q->queue_ctx)
2247                 goto err_exit;
2248
2249         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2250                                                 GFP_KERNEL, set->numa_node);
2251         if (!q->queue_hw_ctx)
2252                 goto err_percpu;
2253
2254         q->mq_map = set->mq_map;
2255
2256         blk_mq_realloc_hw_ctxs(set, q);
2257         if (!q->nr_hw_queues)
2258                 goto err_hctxs;
2259
2260         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2261         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2262
2263         q->nr_queues = nr_cpu_ids;
2264
2265         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2266
2267         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2268                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2269
2270         q->sg_reserved_size = INT_MAX;
2271
2272         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2273         INIT_LIST_HEAD(&q->requeue_list);
2274         spin_lock_init(&q->requeue_lock);
2275
2276         if (q->nr_hw_queues > 1)
2277                 blk_queue_make_request(q, blk_mq_make_request);
2278         else
2279                 blk_queue_make_request(q, blk_sq_make_request);
2280
2281         /*
2282          * Do this after blk_queue_make_request() overrides it...
2283          */
2284         q->nr_requests = set->queue_depth;
2285
2286         /*
2287          * Default to classic polling
2288          */
2289         q->poll_nsec = -1;
2290
2291         if (set->ops->complete)
2292                 blk_queue_softirq_done(q, set->ops->complete);
2293
2294         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2295
2296         get_online_cpus();
2297         mutex_lock(&all_q_mutex);
2298
2299         list_add_tail(&q->all_q_node, &all_q_list);
2300         blk_mq_add_queue_tag_set(set, q);
2301         blk_mq_map_swqueue(q, cpu_online_mask);
2302
2303         mutex_unlock(&all_q_mutex);
2304         put_online_cpus();
2305
2306         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2307                 int ret;
2308
2309                 ret = blk_mq_sched_init(q);
2310                 if (ret)
2311                         return ERR_PTR(ret);
2312         }
2313
2314         return q;
2315
2316 err_hctxs:
2317         kfree(q->queue_hw_ctx);
2318 err_percpu:
2319         free_percpu(q->queue_ctx);
2320 err_exit:
2321         q->mq_ops = NULL;
2322         return ERR_PTR(-ENOMEM);
2323 }
2324 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2325
2326 void blk_mq_free_queue(struct request_queue *q)
2327 {
2328         struct blk_mq_tag_set   *set = q->tag_set;
2329
2330         mutex_lock(&all_q_mutex);
2331         list_del_init(&q->all_q_node);
2332         mutex_unlock(&all_q_mutex);
2333
2334         wbt_exit(q);
2335
2336         blk_mq_del_queue_tag_set(q);
2337
2338         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2339         blk_mq_free_hw_queues(q, set);
2340 }
2341
2342 /* Basically redo blk_mq_init_queue with queue frozen */
2343 static void blk_mq_queue_reinit(struct request_queue *q,
2344                                 const struct cpumask *online_mask)
2345 {
2346         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2347
2348         blk_mq_sysfs_unregister(q);
2349
2350         /*
2351          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2352          * we should change hctx numa_node according to new topology (this
2353          * involves free and re-allocate memory, worthy doing?)
2354          */
2355
2356         blk_mq_map_swqueue(q, online_mask);
2357
2358         blk_mq_sysfs_register(q);
2359 }
2360
2361 /*
2362  * New online cpumask which is going to be set in this hotplug event.
2363  * Declare this cpumasks as global as cpu-hotplug operation is invoked
2364  * one-by-one and dynamically allocating this could result in a failure.
2365  */
2366 static struct cpumask cpuhp_online_new;
2367
2368 static void blk_mq_queue_reinit_work(void)
2369 {
2370         struct request_queue *q;
2371
2372         mutex_lock(&all_q_mutex);
2373         /*
2374          * We need to freeze and reinit all existing queues.  Freezing
2375          * involves synchronous wait for an RCU grace period and doing it
2376          * one by one may take a long time.  Start freezing all queues in
2377          * one swoop and then wait for the completions so that freezing can
2378          * take place in parallel.
2379          */
2380         list_for_each_entry(q, &all_q_list, all_q_node)
2381                 blk_mq_freeze_queue_start(q);
2382         list_for_each_entry(q, &all_q_list, all_q_node)
2383                 blk_mq_freeze_queue_wait(q);
2384
2385         list_for_each_entry(q, &all_q_list, all_q_node)
2386                 blk_mq_queue_reinit(q, &cpuhp_online_new);
2387
2388         list_for_each_entry(q, &all_q_list, all_q_node)
2389                 blk_mq_unfreeze_queue(q);
2390
2391         mutex_unlock(&all_q_mutex);
2392 }
2393
2394 static int blk_mq_queue_reinit_dead(unsigned int cpu)
2395 {
2396         cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2397         blk_mq_queue_reinit_work();
2398         return 0;
2399 }
2400
2401 /*
2402  * Before hotadded cpu starts handling requests, new mappings must be
2403  * established.  Otherwise, these requests in hw queue might never be
2404  * dispatched.
2405  *
2406  * For example, there is a single hw queue (hctx) and two CPU queues (ctx0
2407  * for CPU0, and ctx1 for CPU1).
2408  *
2409  * Now CPU1 is just onlined and a request is inserted into ctx1->rq_list
2410  * and set bit0 in pending bitmap as ctx1->index_hw is still zero.
2411  *
2412  * And then while running hw queue, blk_mq_flush_busy_ctxs() finds bit0 is set
2413  * in pending bitmap and tries to retrieve requests in hctx->ctxs[0]->rq_list.
2414  * But htx->ctxs[0] is a pointer to ctx0, so the request in ctx1->rq_list is
2415  * ignored.
2416  */
2417 static int blk_mq_queue_reinit_prepare(unsigned int cpu)
2418 {
2419         cpumask_copy(&cpuhp_online_new, cpu_online_mask);
2420         cpumask_set_cpu(cpu, &cpuhp_online_new);
2421         blk_mq_queue_reinit_work();
2422         return 0;
2423 }
2424
2425 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2426 {
2427         int i;
2428
2429         for (i = 0; i < set->nr_hw_queues; i++)
2430                 if (!__blk_mq_alloc_rq_map(set, i))
2431                         goto out_unwind;
2432
2433         return 0;
2434
2435 out_unwind:
2436         while (--i >= 0)
2437                 blk_mq_free_rq_map(set->tags[i]);
2438
2439         return -ENOMEM;
2440 }
2441
2442 /*
2443  * Allocate the request maps associated with this tag_set. Note that this
2444  * may reduce the depth asked for, if memory is tight. set->queue_depth
2445  * will be updated to reflect the allocated depth.
2446  */
2447 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2448 {
2449         unsigned int depth;
2450         int err;
2451
2452         depth = set->queue_depth;
2453         do {
2454                 err = __blk_mq_alloc_rq_maps(set);
2455                 if (!err)
2456                         break;
2457
2458                 set->queue_depth >>= 1;
2459                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2460                         err = -ENOMEM;
2461                         break;
2462                 }
2463         } while (set->queue_depth);
2464
2465         if (!set->queue_depth || err) {
2466                 pr_err("blk-mq: failed to allocate request map\n");
2467                 return -ENOMEM;
2468         }
2469
2470         if (depth != set->queue_depth)
2471                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2472                                                 depth, set->queue_depth);
2473
2474         return 0;
2475 }
2476
2477 /*
2478  * Alloc a tag set to be associated with one or more request queues.
2479  * May fail with EINVAL for various error conditions. May adjust the
2480  * requested depth down, if if it too large. In that case, the set
2481  * value will be stored in set->queue_depth.
2482  */
2483 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2484 {
2485         int ret;
2486
2487         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2488
2489         if (!set->nr_hw_queues)
2490                 return -EINVAL;
2491         if (!set->queue_depth)
2492                 return -EINVAL;
2493         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2494                 return -EINVAL;
2495
2496         if (!set->ops->queue_rq)
2497                 return -EINVAL;
2498
2499         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2500                 pr_info("blk-mq: reduced tag depth to %u\n",
2501                         BLK_MQ_MAX_DEPTH);
2502                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2503         }
2504
2505         /*
2506          * If a crashdump is active, then we are potentially in a very
2507          * memory constrained environment. Limit us to 1 queue and
2508          * 64 tags to prevent using too much memory.
2509          */
2510         if (is_kdump_kernel()) {
2511                 set->nr_hw_queues = 1;
2512                 set->queue_depth = min(64U, set->queue_depth);
2513         }
2514         /*
2515          * There is no use for more h/w queues than cpus.
2516          */
2517         if (set->nr_hw_queues > nr_cpu_ids)
2518                 set->nr_hw_queues = nr_cpu_ids;
2519
2520         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2521                                  GFP_KERNEL, set->numa_node);
2522         if (!set->tags)
2523                 return -ENOMEM;
2524
2525         ret = -ENOMEM;
2526         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2527                         GFP_KERNEL, set->numa_node);
2528         if (!set->mq_map)
2529                 goto out_free_tags;
2530
2531         if (set->ops->map_queues)
2532                 ret = set->ops->map_queues(set);
2533         else
2534                 ret = blk_mq_map_queues(set);
2535         if (ret)
2536                 goto out_free_mq_map;
2537
2538         ret = blk_mq_alloc_rq_maps(set);
2539         if (ret)
2540                 goto out_free_mq_map;
2541
2542         mutex_init(&set->tag_list_lock);
2543         INIT_LIST_HEAD(&set->tag_list);
2544
2545         return 0;
2546
2547 out_free_mq_map:
2548         kfree(set->mq_map);
2549         set->mq_map = NULL;
2550 out_free_tags:
2551         kfree(set->tags);
2552         set->tags = NULL;
2553         return ret;
2554 }
2555 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2556
2557 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2558 {
2559         int i;
2560
2561         for (i = 0; i < nr_cpu_ids; i++)
2562                 blk_mq_free_map_and_requests(set, i);
2563
2564         kfree(set->mq_map);
2565         set->mq_map = NULL;
2566
2567         kfree(set->tags);
2568         set->tags = NULL;
2569 }
2570 EXPORT_SYMBOL(blk_mq_free_tag_set);
2571
2572 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2573 {
2574         struct blk_mq_tag_set *set = q->tag_set;
2575         struct blk_mq_hw_ctx *hctx;
2576         int i, ret;
2577
2578         if (!set)
2579                 return -EINVAL;
2580
2581         blk_mq_freeze_queue(q);
2582         blk_mq_quiesce_queue(q);
2583
2584         ret = 0;
2585         queue_for_each_hw_ctx(q, hctx, i) {
2586                 if (!hctx->tags)
2587                         continue;
2588                 /*
2589                  * If we're using an MQ scheduler, just update the scheduler
2590                  * queue depth. This is similar to what the old code would do.
2591                  */
2592                 if (!hctx->sched_tags) {
2593                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2594                                                         min(nr, set->queue_depth),
2595                                                         false);
2596                 } else {
2597                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2598                                                         nr, true);
2599                 }
2600                 if (ret)
2601                         break;
2602         }
2603
2604         if (!ret)
2605                 q->nr_requests = nr;
2606
2607         blk_mq_unfreeze_queue(q);
2608         blk_mq_start_stopped_hw_queues(q, true);
2609
2610         return ret;
2611 }
2612
2613 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2614 {
2615         struct request_queue *q;
2616
2617         if (nr_hw_queues > nr_cpu_ids)
2618                 nr_hw_queues = nr_cpu_ids;
2619         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2620                 return;
2621
2622         list_for_each_entry(q, &set->tag_list, tag_set_list)
2623                 blk_mq_freeze_queue(q);
2624
2625         set->nr_hw_queues = nr_hw_queues;
2626         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2627                 blk_mq_realloc_hw_ctxs(set, q);
2628
2629                 if (q->nr_hw_queues > 1)
2630                         blk_queue_make_request(q, blk_mq_make_request);
2631                 else
2632                         blk_queue_make_request(q, blk_sq_make_request);
2633
2634                 blk_mq_queue_reinit(q, cpu_online_mask);
2635         }
2636
2637         list_for_each_entry(q, &set->tag_list, tag_set_list)
2638                 blk_mq_unfreeze_queue(q);
2639 }
2640 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2641
2642 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2643                                        struct blk_mq_hw_ctx *hctx,
2644                                        struct request *rq)
2645 {
2646         struct blk_rq_stat stat[2];
2647         unsigned long ret = 0;
2648
2649         /*
2650          * If stats collection isn't on, don't sleep but turn it on for
2651          * future users
2652          */
2653         if (!blk_stat_enable(q))
2654                 return 0;
2655
2656         /*
2657          * We don't have to do this once per IO, should optimize this
2658          * to just use the current window of stats until it changes
2659          */
2660         memset(&stat, 0, sizeof(stat));
2661         blk_hctx_stat_get(hctx, stat);
2662
2663         /*
2664          * As an optimistic guess, use half of the mean service time
2665          * for this type of request. We can (and should) make this smarter.
2666          * For instance, if the completion latencies are tight, we can
2667          * get closer than just half the mean. This is especially
2668          * important on devices where the completion latencies are longer
2669          * than ~10 usec.
2670          */
2671         if (req_op(rq) == REQ_OP_READ && stat[BLK_STAT_READ].nr_samples)
2672                 ret = (stat[BLK_STAT_READ].mean + 1) / 2;
2673         else if (req_op(rq) == REQ_OP_WRITE && stat[BLK_STAT_WRITE].nr_samples)
2674                 ret = (stat[BLK_STAT_WRITE].mean + 1) / 2;
2675
2676         return ret;
2677 }
2678
2679 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2680                                      struct blk_mq_hw_ctx *hctx,
2681                                      struct request *rq)
2682 {
2683         struct hrtimer_sleeper hs;
2684         enum hrtimer_mode mode;
2685         unsigned int nsecs;
2686         ktime_t kt;
2687
2688         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2689                 return false;
2690
2691         /*
2692          * poll_nsec can be:
2693          *
2694          * -1:  don't ever hybrid sleep
2695          *  0:  use half of prev avg
2696          * >0:  use this specific value
2697          */
2698         if (q->poll_nsec == -1)
2699                 return false;
2700         else if (q->poll_nsec > 0)
2701                 nsecs = q->poll_nsec;
2702         else
2703                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2704
2705         if (!nsecs)
2706                 return false;
2707
2708         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2709
2710         /*
2711          * This will be replaced with the stats tracking code, using
2712          * 'avg_completion_time / 2' as the pre-sleep target.
2713          */
2714         kt = nsecs;
2715
2716         mode = HRTIMER_MODE_REL;
2717         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2718         hrtimer_set_expires(&hs.timer, kt);
2719
2720         hrtimer_init_sleeper(&hs, current);
2721         do {
2722                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2723                         break;
2724                 set_current_state(TASK_UNINTERRUPTIBLE);
2725                 hrtimer_start_expires(&hs.timer, mode);
2726                 if (hs.task)
2727                         io_schedule();
2728                 hrtimer_cancel(&hs.timer);
2729                 mode = HRTIMER_MODE_ABS;
2730         } while (hs.task && !signal_pending(current));
2731
2732         __set_current_state(TASK_RUNNING);
2733         destroy_hrtimer_on_stack(&hs.timer);
2734         return true;
2735 }
2736
2737 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2738 {
2739         struct request_queue *q = hctx->queue;
2740         long state;
2741
2742         /*
2743          * If we sleep, have the caller restart the poll loop to reset
2744          * the state. Like for the other success return cases, the
2745          * caller is responsible for checking if the IO completed. If
2746          * the IO isn't complete, we'll get called again and will go
2747          * straight to the busy poll loop.
2748          */
2749         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2750                 return true;
2751
2752         hctx->poll_considered++;
2753
2754         state = current->state;
2755         while (!need_resched()) {
2756                 int ret;
2757
2758                 hctx->poll_invoked++;
2759
2760                 ret = q->mq_ops->poll(hctx, rq->tag);
2761                 if (ret > 0) {
2762                         hctx->poll_success++;
2763                         set_current_state(TASK_RUNNING);
2764                         return true;
2765                 }
2766
2767                 if (signal_pending_state(state, current))
2768                         set_current_state(TASK_RUNNING);
2769
2770                 if (current->state == TASK_RUNNING)
2771                         return true;
2772                 if (ret < 0)
2773                         break;
2774                 cpu_relax();
2775         }
2776
2777         return false;
2778 }
2779
2780 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2781 {
2782         struct blk_mq_hw_ctx *hctx;
2783         struct blk_plug *plug;
2784         struct request *rq;
2785
2786         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2787             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2788                 return false;
2789
2790         plug = current->plug;
2791         if (plug)
2792                 blk_flush_plug_list(plug, false);
2793
2794         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2795         if (!blk_qc_t_is_internal(cookie))
2796                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2797         else
2798                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2799
2800         return __blk_mq_poll(hctx, rq);
2801 }
2802 EXPORT_SYMBOL_GPL(blk_mq_poll);
2803
2804 void blk_mq_disable_hotplug(void)
2805 {
2806         mutex_lock(&all_q_mutex);
2807 }
2808
2809 void blk_mq_enable_hotplug(void)
2810 {
2811         mutex_unlock(&all_q_mutex);
2812 }
2813
2814 static int __init blk_mq_init(void)
2815 {
2816         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2817                                 blk_mq_hctx_notify_dead);
2818
2819         cpuhp_setup_state_nocalls(CPUHP_BLK_MQ_PREPARE, "block/mq:prepare",
2820                                   blk_mq_queue_reinit_prepare,
2821                                   blk_mq_queue_reinit_dead);
2822         return 0;
2823 }
2824 subsys_initcall(blk_mq_init);