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/mm.h>
  13 #include <linux/init.h>
  14 #include <linux/slab.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/smp.h>
  17 #include <linux/llist.h>
  18 #include <linux/list_sort.h>
  19 #include <linux/cpu.h>
  20 #include <linux/cache.h>
  21 #include <linux/sched/sysctl.h>
  22 #include <linux/delay.h>
  23 #include <linux/crash_dump.h>
  24
  25 #include <trace/events/block.h>
  26
  27 #include <linux/blk-mq.h>
  28 #include "blk.h"
  29 #include "blk-mq.h"
  30 #include "blk-mq-tag.h"
  31
  32 static DEFINE_MUTEX(all_q_mutex);
  33 static LIST_HEAD(all_q_list);
  34
  35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  36
  37 /*
  38  * Check if any of the ctx's have pending work in this hardware queue
  39  */
  40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  41 {
  42         unsigned int i;
  43
  44         for (i = 0; i < hctx->ctx_map.map_size; i++)
  45                 if (hctx->ctx_map.map[i].word)
  46                         return true;
  47
  48         return false;
  49 }
  50
  51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  52                                               struct blk_mq_ctx *ctx)
  53 {
  54         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  55 }
  56
  57 #define CTX_TO_BIT(hctx, ctx)   \
  58         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  59
  60 /*
  61  * Mark this ctx as having pending work in this hardware queue
  62  */
  63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  64                                      struct blk_mq_ctx *ctx)
  65 {
  66         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  67
  68         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  69                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  70 }
  71
  72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  73                                       struct blk_mq_ctx *ctx)
  74 {
  75         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  76
  77         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  78 }
  79
  80 static int blk_mq_queue_enter(struct request_queue *q)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 ret = wait_event_interruptible(q->mq_freeze_wq,
  89                                 !q->mq_freeze_depth || blk_queue_dying(q));
  90                 if (blk_queue_dying(q))
  91                         return -ENODEV;
  92                 if (ret)
  93                         return ret;
  94         }
  95 }
  96
  97 static void blk_mq_queue_exit(struct request_queue *q)
  98 {
  99         percpu_ref_put(&q->mq_usage_counter);
 100 }
 101
 102 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 103 {
 104         struct request_queue *q =
 105                 container_of(ref, struct request_queue, mq_usage_counter);
 106
 107         wake_up_all(&q->mq_freeze_wq);
 108 }
 109
 110 void blk_mq_freeze_queue_start(struct request_queue *q)
 111 {
 112         bool freeze;
 113
 114         spin_lock_irq(q->queue_lock);
 115         freeze = !q->mq_freeze_depth++;
 116         spin_unlock_irq(q->queue_lock);
 117
 118         if (freeze) {
 119                 percpu_ref_kill(&q->mq_usage_counter);
 120                 blk_mq_run_queues(q, false);
 121         }
 122 }
 123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 124
 125 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 126 {
 127         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 128 }
 129
 130 /*
 131  * Guarantee no request is in use, so we can change any data structure of
 132  * the queue afterward.
 133  */
 134 void blk_mq_freeze_queue(struct request_queue *q)
 135 {
 136         blk_mq_freeze_queue_start(q);
 137         blk_mq_freeze_queue_wait(q);
 138 }
 139
 140 void blk_mq_unfreeze_queue(struct request_queue *q)
 141 {
 142         bool wake;
 143
 144         spin_lock_irq(q->queue_lock);
 145         wake = !--q->mq_freeze_depth;
 146         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 147         spin_unlock_irq(q->queue_lock);
 148         if (wake) {
 149                 percpu_ref_reinit(&q->mq_usage_counter);
 150                 wake_up_all(&q->mq_freeze_wq);
 151         }
 152 }
 153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 154
 155 void blk_mq_wake_waiters(struct request_queue *q)
 156 {
 157         struct blk_mq_hw_ctx *hctx;
 158         unsigned int i;
 159
 160         queue_for_each_hw_ctx(q, hctx, i)
 161                 if (blk_mq_hw_queue_mapped(hctx))
 162                         blk_mq_tag_wakeup_all(hctx->tags, true);
 163
 164         /*
 165          * If we are called because the queue has now been marked as
 166          * dying, we need to ensure that processes currently waiting on
 167          * the queue are notified as well.
 168          */
 169         wake_up_all(&q->mq_freeze_wq);
 170 }
 171
 172 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 173 {
 174         return blk_mq_has_free_tags(hctx->tags);
 175 }
 176 EXPORT_SYMBOL(blk_mq_can_queue);
 177
 178 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 179                                struct request *rq, unsigned int rw_flags)
 180 {
 181         if (blk_queue_io_stat(q))
 182                 rw_flags |= REQ_IO_STAT;
 183
 184         INIT_LIST_HEAD(&rq->queuelist);
 185         /* csd/requeue_work/fifo_time is initialized before use */
 186         rq->q = q;
 187         rq->mq_ctx = ctx;
 188         rq->cmd_flags |= rw_flags;
 189         /* do not touch atomic flags, it needs atomic ops against the timer */
 190         rq->cpu = -1;
 191         INIT_HLIST_NODE(&rq->hash);
 192         RB_CLEAR_NODE(&rq->rb_node);
 193         rq->rq_disk = NULL;
 194         rq->part = NULL;
 195         rq->start_time = jiffies;
 196 #ifdef CONFIG_BLK_CGROUP
 197         rq->rl = NULL;
 198         set_start_time_ns(rq);
 199         rq->io_start_time_ns = 0;
 200 #endif
 201         rq->nr_phys_segments = 0;
 202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 203         rq->nr_integrity_segments = 0;
 204 #endif
 205         rq->special = NULL;
 206         /* tag was already set */
 207         rq->errors = 0;
 208
 209         rq->cmd = rq->__cmd;
 210
 211         rq->extra_len = 0;
 212         rq->sense_len = 0;
 213         rq->resid_len = 0;
 214         rq->sense = NULL;
 215
 216         INIT_LIST_HEAD(&rq->timeout_list);
 217         rq->timeout = 0;
 218
 219         rq->end_io = NULL;
 220         rq->end_io_data = NULL;
 221         rq->next_rq = NULL;
 222
 223         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 224 }
 225
 226 static struct request *
 227 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 228 {
 229         struct request *rq;
 230         unsigned int tag;
 231
 232         tag = blk_mq_get_tag(data);
 233         if (tag != BLK_MQ_TAG_FAIL) {
 234                 rq = data->hctx->tags->rqs[tag];
 235
 236                 if (blk_mq_tag_busy(data->hctx)) {
 237                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 238                         atomic_inc(&data->hctx->nr_active);
 239                 }
 240
 241                 rq->tag = tag;
 242                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 243                 return rq;
 244         }
 245
 246         return NULL;
 247 }
 248
 249 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 250                 bool reserved)
 251 {
 252         struct blk_mq_ctx *ctx;
 253         struct blk_mq_hw_ctx *hctx;
 254         struct request *rq;
 255         struct blk_mq_alloc_data alloc_data;
 256         int ret;
 257
 258         ret = blk_mq_queue_enter(q);
 259         if (ret)
 260                 return ERR_PTR(ret);
 261
 262         ctx = blk_mq_get_ctx(q);
 263         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 264         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 265                         reserved, ctx, hctx);
 266
 267         rq = __blk_mq_alloc_request(&alloc_data, rw);
 268         if (!rq && (gfp & __GFP_WAIT)) {
 269                 __blk_mq_run_hw_queue(hctx);
 270                 blk_mq_put_ctx(ctx);
 271
 272                 ctx = blk_mq_get_ctx(q);
 273                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 274                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 275                                 hctx);
 276                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 277                 ctx = alloc_data.ctx;
 278         }
 279         blk_mq_put_ctx(ctx);
 280         if (!rq) {
 281                 blk_mq_queue_exit(q);
 282                 return ERR_PTR(-EWOULDBLOCK);
 283         }
 284         return rq;
 285 }
 286 EXPORT_SYMBOL(blk_mq_alloc_request);
 287
 288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 289                                   struct blk_mq_ctx *ctx, struct request *rq)
 290 {
 291         const int tag = rq->tag;
 292         struct request_queue *q = rq->q;
 293
 294         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 295                 atomic_dec(&hctx->nr_active);
 296         rq->cmd_flags = 0;
 297
 298         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 299         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 300         blk_mq_queue_exit(q);
 301 }
 302
 303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 304 {
 305         struct blk_mq_ctx *ctx = rq->mq_ctx;
 306
 307         ctx->rq_completed[rq_is_sync(rq)]++;
 308         __blk_mq_free_request(hctx, ctx, rq);
 309
 310 }
 311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 312
 313 void blk_mq_free_request(struct request *rq)
 314 {
 315         struct blk_mq_hw_ctx *hctx;
 316         struct request_queue *q = rq->q;
 317
 318         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 319         blk_mq_free_hctx_request(hctx, rq);
 320 }
 321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 322
 323 inline void __blk_mq_end_request(struct request *rq, int error)
 324 {
 325         blk_account_io_done(rq);
 326
 327         if (rq->end_io) {
 328                 rq->end_io(rq, error);
 329         } else {
 330                 if (unlikely(blk_bidi_rq(rq)))
 331                         blk_mq_free_request(rq->next_rq);
 332                 blk_mq_free_request(rq);
 333         }
 334 }
 335 EXPORT_SYMBOL(__blk_mq_end_request);
 336
 337 void blk_mq_end_request(struct request *rq, int error)
 338 {
 339         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 340                 BUG();
 341         __blk_mq_end_request(rq, error);
 342 }
 343 EXPORT_SYMBOL(blk_mq_end_request);
 344
 345 static void __blk_mq_complete_request_remote(void *data)
 346 {
 347         struct request *rq = data;
 348
 349         rq->q->softirq_done_fn(rq);
 350 }
 351
 352 static void blk_mq_ipi_complete_request(struct request *rq)
 353 {
 354         struct blk_mq_ctx *ctx = rq->mq_ctx;
 355         bool shared = false;
 356         int cpu;
 357
 358         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 359                 rq->q->softirq_done_fn(rq);
 360                 return;
 361         }
 362
 363         cpu = get_cpu();
 364         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 365                 shared = cpus_share_cache(cpu, ctx->cpu);
 366
 367         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 368                 rq->csd.func = __blk_mq_complete_request_remote;
 369                 rq->csd.info = rq;
 370                 rq->csd.flags = 0;
 371                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 372         } else {
 373                 rq->q->softirq_done_fn(rq);
 374         }
 375         put_cpu();
 376 }
 377
 378 void __blk_mq_complete_request(struct request *rq)
 379 {
 380         struct request_queue *q = rq->q;
 381
 382         if (!q->softirq_done_fn)
 383                 blk_mq_end_request(rq, rq->errors);
 384         else
 385                 blk_mq_ipi_complete_request(rq);
 386 }
 387
 388 /**
 389  * blk_mq_complete_request - end I/O on a request
 390  * @rq:         the request being processed
 391  *
 392  * Description:
 393  *      Ends all I/O on a request. It does not handle partial completions.
 394  *      The actual completion happens out-of-order, through a IPI handler.
 395  **/
 396 void blk_mq_complete_request(struct request *rq)
 397 {
 398         struct request_queue *q = rq->q;
 399
 400         if (unlikely(blk_should_fake_timeout(q)))
 401                 return;
 402         if (!blk_mark_rq_complete(rq))
 403                 __blk_mq_complete_request(rq);
 404 }
 405 EXPORT_SYMBOL(blk_mq_complete_request);
 406
 407 int blk_mq_request_started(struct request *rq)
 408 {
 409         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 410 }
 411 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 412
 413 void blk_mq_start_request(struct request *rq)
 414 {
 415         struct request_queue *q = rq->q;
 416
 417         trace_block_rq_issue(q, rq);
 418
 419         rq->resid_len = blk_rq_bytes(rq);
 420         if (unlikely(blk_bidi_rq(rq)))
 421                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 422
 423         blk_add_timer(rq);
 424
 425         /*
 426          * Ensure that ->deadline is visible before set the started
 427          * flag and clear the completed flag.
 428          */
 429         smp_mb__before_atomic();
 430
 431         /*
 432          * Mark us as started and clear complete. Complete might have been
 433          * set if requeue raced with timeout, which then marked it as
 434          * complete. So be sure to clear complete again when we start
 435          * the request, otherwise we'll ignore the completion event.
 436          */
 437         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 438                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 439         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 440                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 441
 442         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 443                 /*
 444                  * Make sure space for the drain appears.  We know we can do
 445                  * this because max_hw_segments has been adjusted to be one
 446                  * fewer than the device can handle.
 447                  */
 448                 rq->nr_phys_segments++;
 449         }
 450 }
 451 EXPORT_SYMBOL(blk_mq_start_request);
 452
 453 static void __blk_mq_requeue_request(struct request *rq)
 454 {
 455         struct request_queue *q = rq->q;
 456
 457         trace_block_rq_requeue(q, rq);
 458
 459         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 460                 if (q->dma_drain_size && blk_rq_bytes(rq))
 461                         rq->nr_phys_segments--;
 462         }
 463 }
 464
 465 void blk_mq_requeue_request(struct request *rq)
 466 {
 467         __blk_mq_requeue_request(rq);
 468
 469         BUG_ON(blk_queued_rq(rq));
 470         blk_mq_add_to_requeue_list(rq, true);
 471 }
 472 EXPORT_SYMBOL(blk_mq_requeue_request);
 473
 474 static void blk_mq_requeue_work(struct work_struct *work)
 475 {
 476         struct request_queue *q =
 477                 container_of(work, struct request_queue, requeue_work);
 478         LIST_HEAD(rq_list);
 479         struct request *rq, *next;
 480         unsigned long flags;
 481
 482         spin_lock_irqsave(&q->requeue_lock, flags);
 483         list_splice_init(&q->requeue_list, &rq_list);
 484         spin_unlock_irqrestore(&q->requeue_lock, flags);
 485
 486         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 487                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 488                         continue;
 489
 490                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 491                 list_del_init(&rq->queuelist);
 492                 blk_mq_insert_request(rq, true, false, false);
 493         }
 494
 495         while (!list_empty(&rq_list)) {
 496                 rq = list_entry(rq_list.next, struct request, queuelist);
 497                 list_del_init(&rq->queuelist);
 498                 blk_mq_insert_request(rq, false, false, false);
 499         }
 500
 501         /*
 502          * Use the start variant of queue running here, so that running
 503          * the requeue work will kick stopped queues.
 504          */
 505         blk_mq_start_hw_queues(q);
 506 }
 507
 508 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 509 {
 510         struct request_queue *q = rq->q;
 511         unsigned long flags;
 512
 513         /*
 514          * We abuse this flag that is otherwise used by the I/O scheduler to
 515          * request head insertation from the workqueue.
 516          */
 517         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 518
 519         spin_lock_irqsave(&q->requeue_lock, flags);
 520         if (at_head) {
 521                 rq->cmd_flags |= REQ_SOFTBARRIER;
 522                 list_add(&rq->queuelist, &q->requeue_list);
 523         } else {
 524                 list_add_tail(&rq->queuelist, &q->requeue_list);
 525         }
 526         spin_unlock_irqrestore(&q->requeue_lock, flags);
 527 }
 528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 529
 530 void blk_mq_cancel_requeue_work(struct request_queue *q)
 531 {
 532         cancel_work_sync(&q->requeue_work);
 533 }
 534 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 535
 536 void blk_mq_kick_requeue_list(struct request_queue *q)
 537 {
 538         kblockd_schedule_work(&q->requeue_work);
 539 }
 540 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 541
 542 void blk_mq_abort_requeue_list(struct request_queue *q)
 543 {
 544         unsigned long flags;
 545         LIST_HEAD(rq_list);
 546
 547         spin_lock_irqsave(&q->requeue_lock, flags);
 548         list_splice_init(&q->requeue_list, &rq_list);
 549         spin_unlock_irqrestore(&q->requeue_lock, flags);
 550
 551         while (!list_empty(&rq_list)) {
 552                 struct request *rq;
 553
 554                 rq = list_first_entry(&rq_list, struct request, queuelist);
 555                 list_del_init(&rq->queuelist);
 556                 rq->errors = -EIO;
 557                 blk_mq_end_request(rq, rq->errors);
 558         }
 559 }
 560 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 561
 562 static inline bool is_flush_request(struct request *rq,
 563                 struct blk_flush_queue *fq, unsigned int tag)
 564 {
 565         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 566                         fq->flush_rq->tag == tag);
 567 }
 568
 569 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 570 {
 571         struct request *rq = tags->rqs[tag];
 572         /* mq_ctx of flush rq is always cloned from the corresponding req */
 573         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 574
 575         if (!is_flush_request(rq, fq, tag))
 576                 return rq;
 577
 578         return fq->flush_rq;
 579 }
 580 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 581
 582 struct blk_mq_timeout_data {
 583         unsigned long next;
 584         unsigned int next_set;
 585 };
 586
 587 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 588 {
 589         struct blk_mq_ops *ops = req->q->mq_ops;
 590         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 591
 592         /*
 593          * We know that complete is set at this point. If STARTED isn't set
 594          * anymore, then the request isn't active and the "timeout" should
 595          * just be ignored. This can happen due to the bitflag ordering.
 596          * Timeout first checks if STARTED is set, and if it is, assumes
 597          * the request is active. But if we race with completion, then
 598          * we both flags will get cleared. So check here again, and ignore
 599          * a timeout event with a request that isn't active.
 600          */
 601         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 602                 return;
 603
 604         if (ops->timeout)
 605                 ret = ops->timeout(req, reserved);
 606
 607         switch (ret) {
 608         case BLK_EH_HANDLED:
 609                 __blk_mq_complete_request(req);
 610                 break;
 611         case BLK_EH_RESET_TIMER:
 612                 blk_add_timer(req);
 613                 blk_clear_rq_complete(req);
 614                 break;
 615         case BLK_EH_NOT_HANDLED:
 616                 break;
 617         default:
 618                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 619                 break;
 620         }
 621 }
 622
 623 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 624                 struct request *rq, void *priv, bool reserved)
 625 {
 626         struct blk_mq_timeout_data *data = priv;
 627
 628         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 629                 /*
 630                  * If a request wasn't started before the queue was
 631                  * marked dying, kill it here or it'll go unnoticed.
 632                  */
 633                 if (unlikely(blk_queue_dying(rq->q))) {
 634                         rq->errors = -EIO;
 635                         blk_mq_complete_request(rq);
 636                 }
 637                 return;
 638         }
 639         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 640                 return;
 641
 642         if (time_after_eq(jiffies, rq->deadline)) {
 643                 if (!blk_mark_rq_complete(rq))
 644                         blk_mq_rq_timed_out(rq, reserved);
 645         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 646                 data->next = rq->deadline;
 647                 data->next_set = 1;
 648         }
 649 }
 650
 651 static void blk_mq_rq_timer(unsigned long priv)
 652 {
 653         struct request_queue *q = (struct request_queue *)priv;
 654         struct blk_mq_timeout_data data = {
 655                 .next           = 0,
 656                 .next_set       = 0,
 657         };
 658         struct blk_mq_hw_ctx *hctx;
 659         int i;
 660
 661         queue_for_each_hw_ctx(q, hctx, i) {
 662                 /*
 663                  * If not software queues are currently mapped to this
 664                  * hardware queue, there's nothing to check
 665                  */
 666                 if (!blk_mq_hw_queue_mapped(hctx))
 667                         continue;
 668
 669                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 670         }
 671
 672         if (data.next_set) {
 673                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 674                 mod_timer(&q->timeout, data.next);
 675         } else {
 676                 queue_for_each_hw_ctx(q, hctx, i)
 677                         blk_mq_tag_idle(hctx);
 678         }
 679 }
 680
 681 /*
 682  * Reverse check our software queue for entries that we could potentially
 683  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 684  * too much time checking for merges.
 685  */
 686 static bool blk_mq_attempt_merge(struct request_queue *q,
 687                                  struct blk_mq_ctx *ctx, struct bio *bio)
 688 {
 689         struct request *rq;
 690         int checked = 8;
 691
 692         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 693                 int el_ret;
 694
 695                 if (!checked--)
 696                         break;
 697
 698                 if (!blk_rq_merge_ok(rq, bio))
 699                         continue;
 700
 701                 el_ret = blk_try_merge(rq, bio);
 702                 if (el_ret == ELEVATOR_BACK_MERGE) {
 703                         if (bio_attempt_back_merge(q, rq, bio)) {
 704                                 ctx->rq_merged++;
 705                                 return true;
 706                         }
 707                         break;
 708                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 709                         if (bio_attempt_front_merge(q, rq, bio)) {
 710                                 ctx->rq_merged++;
 711                                 return true;
 712                         }
 713                         break;
 714                 }
 715         }
 716
 717         return false;
 718 }
 719
 720 /*
 721  * Process software queues that have been marked busy, splicing them
 722  * to the for-dispatch
 723  */
 724 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 725 {
 726         struct blk_mq_ctx *ctx;
 727         int i;
 728
 729         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 730                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 731                 unsigned int off, bit;
 732
 733                 if (!bm->word)
 734                         continue;
 735
 736                 bit = 0;
 737                 off = i * hctx->ctx_map.bits_per_word;
 738                 do {
 739                         bit = find_next_bit(&bm->word, bm->depth, bit);
 740                         if (bit >= bm->depth)
 741                                 break;
 742
 743                         ctx = hctx->ctxs[bit + off];
 744                         clear_bit(bit, &bm->word);
 745                         spin_lock(&ctx->lock);
 746                         list_splice_tail_init(&ctx->rq_list, list);
 747                         spin_unlock(&ctx->lock);
 748
 749                         bit++;
 750                 } while (1);
 751         }
 752 }
 753
 754 /*
 755  * Run this hardware queue, pulling any software queues mapped to it in.
 756  * Note that this function currently has various problems around ordering
 757  * of IO. In particular, we'd like FIFO behaviour on handling existing
 758  * items on the hctx->dispatch list. Ignore that for now.
 759  */
 760 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 761 {
 762         struct request_queue *q = hctx->queue;
 763         struct request *rq;
 764         LIST_HEAD(rq_list);
 765         LIST_HEAD(driver_list);
 766         struct list_head *dptr;
 767         int queued;
 768
 769         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 770
 771         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 772                 return;
 773
 774         hctx->run++;
 775
 776         /*
 777          * Touch any software queue that has pending entries.
 778          */
 779         flush_busy_ctxs(hctx, &rq_list);
 780
 781         /*
 782          * If we have previous entries on our dispatch list, grab them
 783          * and stuff them at the front for more fair dispatch.
 784          */
 785         if (!list_empty_careful(&hctx->dispatch)) {
 786                 spin_lock(&hctx->lock);
 787                 if (!list_empty(&hctx->dispatch))
 788                         list_splice_init(&hctx->dispatch, &rq_list);
 789                 spin_unlock(&hctx->lock);
 790         }
 791
 792         /*
 793          * Start off with dptr being NULL, so we start the first request
 794          * immediately, even if we have more pending.
 795          */
 796         dptr = NULL;
 797
 798         /*
 799          * Now process all the entries, sending them to the driver.
 800          */
 801         queued = 0;
 802         while (!list_empty(&rq_list)) {
 803                 struct blk_mq_queue_data bd;
 804                 int ret;
 805
 806                 rq = list_first_entry(&rq_list, struct request, queuelist);
 807                 list_del_init(&rq->queuelist);
 808
 809                 bd.rq = rq;
 810                 bd.list = dptr;
 811                 bd.last = list_empty(&rq_list);
 812
 813                 ret = q->mq_ops->queue_rq(hctx, &bd);
 814                 switch (ret) {
 815                 case BLK_MQ_RQ_QUEUE_OK:
 816                         queued++;
 817                         continue;
 818                 case BLK_MQ_RQ_QUEUE_BUSY:
 819                         list_add(&rq->queuelist, &rq_list);
 820                         __blk_mq_requeue_request(rq);
 821                         break;
 822                 default:
 823                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 824                 case BLK_MQ_RQ_QUEUE_ERROR:
 825                         rq->errors = -EIO;
 826                         blk_mq_end_request(rq, rq->errors);
 827                         break;
 828                 }
 829
 830                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 831                         break;
 832
 833                 /*
 834                  * We've done the first request. If we have more than 1
 835                  * left in the list, set dptr to defer issue.
 836                  */
 837                 if (!dptr && rq_list.next != rq_list.prev)
 838                         dptr = &driver_list;
 839         }
 840
 841         if (!queued)
 842                 hctx->dispatched[0]++;
 843         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 844                 hctx->dispatched[ilog2(queued) + 1]++;
 845
 846         /*
 847          * Any items that need requeuing? Stuff them into hctx->dispatch,
 848          * that is where we will continue on next queue run.
 849          */
 850         if (!list_empty(&rq_list)) {
 851                 spin_lock(&hctx->lock);
 852                 list_splice(&rq_list, &hctx->dispatch);
 853                 spin_unlock(&hctx->lock);
 854         }
 855 }
 856
 857 /*
 858  * It'd be great if the workqueue API had a way to pass
 859  * in a mask and had some smarts for more clever placement.
 860  * For now we just round-robin here, switching for every
 861  * BLK_MQ_CPU_WORK_BATCH queued items.
 862  */
 863 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 864 {
 865         if (hctx->queue->nr_hw_queues == 1)
 866                 return WORK_CPU_UNBOUND;
 867
 868         if (--hctx->next_cpu_batch <= 0) {
 869                 int cpu = hctx->next_cpu, next_cpu;
 870
 871                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 872                 if (next_cpu >= nr_cpu_ids)
 873                         next_cpu = cpumask_first(hctx->cpumask);
 874
 875                 hctx->next_cpu = next_cpu;
 876                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 877
 878                 return cpu;
 879         }
 880
 881         return hctx->next_cpu;
 882 }
 883
 884 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 885 {
 886         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 887             !blk_mq_hw_queue_mapped(hctx)))
 888                 return;
 889
 890         if (!async) {
 891                 int cpu = get_cpu();
 892                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 893                         __blk_mq_run_hw_queue(hctx);
 894                         put_cpu();
 895                         return;
 896                 }
 897
 898                 put_cpu();
 899         }
 900
 901         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 902                         &hctx->run_work, 0);
 903 }
 904
 905 void blk_mq_run_queues(struct request_queue *q, bool async)
 906 {
 907         struct blk_mq_hw_ctx *hctx;
 908         int i;
 909
 910         queue_for_each_hw_ctx(q, hctx, i) {
 911                 if ((!blk_mq_hctx_has_pending(hctx) &&
 912                     list_empty_careful(&hctx->dispatch)) ||
 913                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 914                         continue;
 915
 916                 blk_mq_run_hw_queue(hctx, async);
 917         }
 918 }
 919 EXPORT_SYMBOL(blk_mq_run_queues);
 920
 921 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 922 {
 923         cancel_delayed_work(&hctx->run_work);
 924         cancel_delayed_work(&hctx->delay_work);
 925         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 926 }
 927 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 928
 929 void blk_mq_stop_hw_queues(struct request_queue *q)
 930 {
 931         struct blk_mq_hw_ctx *hctx;
 932         int i;
 933
 934         queue_for_each_hw_ctx(q, hctx, i)
 935                 blk_mq_stop_hw_queue(hctx);
 936 }
 937 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 938
 939 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 940 {
 941         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 942
 943         blk_mq_run_hw_queue(hctx, false);
 944 }
 945 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 946
 947 void blk_mq_start_hw_queues(struct request_queue *q)
 948 {
 949         struct blk_mq_hw_ctx *hctx;
 950         int i;
 951
 952         queue_for_each_hw_ctx(q, hctx, i)
 953                 blk_mq_start_hw_queue(hctx);
 954 }
 955 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 956
 957
 958 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 959 {
 960         struct blk_mq_hw_ctx *hctx;
 961         int i;
 962
 963         queue_for_each_hw_ctx(q, hctx, i) {
 964                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 965                         continue;
 966
 967                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 968                 blk_mq_run_hw_queue(hctx, async);
 969         }
 970 }
 971 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 972
 973 static void blk_mq_run_work_fn(struct work_struct *work)
 974 {
 975         struct blk_mq_hw_ctx *hctx;
 976
 977         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 978
 979         __blk_mq_run_hw_queue(hctx);
 980 }
 981
 982 static void blk_mq_delay_work_fn(struct work_struct *work)
 983 {
 984         struct blk_mq_hw_ctx *hctx;
 985
 986         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 987
 988         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 989                 __blk_mq_run_hw_queue(hctx);
 990 }
 991
 992 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 993 {
 994         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
 995                 return;
 996
 997         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 998                         &hctx->delay_work, msecs_to_jiffies(msecs));
 999 }
1000 EXPORT_SYMBOL(blk_mq_delay_queue);
1001
1002 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1003                                     struct request *rq, bool at_head)
1004 {
1005         struct blk_mq_ctx *ctx = rq->mq_ctx;
1006
1007         trace_block_rq_insert(hctx->queue, rq);
1008
1009         if (at_head)
1010                 list_add(&rq->queuelist, &ctx->rq_list);
1011         else
1012                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1013
1014         blk_mq_hctx_mark_pending(hctx, ctx);
1015 }
1016
1017 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1018                 bool async)
1019 {
1020         struct request_queue *q = rq->q;
1021         struct blk_mq_hw_ctx *hctx;
1022         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1023
1024         current_ctx = blk_mq_get_ctx(q);
1025         if (!cpu_online(ctx->cpu))
1026                 rq->mq_ctx = ctx = current_ctx;
1027
1028         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1029
1030         spin_lock(&ctx->lock);
1031         __blk_mq_insert_request(hctx, rq, at_head);
1032         spin_unlock(&ctx->lock);
1033
1034         if (run_queue)
1035                 blk_mq_run_hw_queue(hctx, async);
1036
1037         blk_mq_put_ctx(current_ctx);
1038 }
1039
1040 static void blk_mq_insert_requests(struct request_queue *q,
1041                                      struct blk_mq_ctx *ctx,
1042                                      struct list_head *list,
1043                                      int depth,
1044                                      bool from_schedule)
1045
1046 {
1047         struct blk_mq_hw_ctx *hctx;
1048         struct blk_mq_ctx *current_ctx;
1049
1050         trace_block_unplug(q, depth, !from_schedule);
1051
1052         current_ctx = blk_mq_get_ctx(q);
1053
1054         if (!cpu_online(ctx->cpu))
1055                 ctx = current_ctx;
1056         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1057
1058         /*
1059          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1060          * offline now
1061          */
1062         spin_lock(&ctx->lock);
1063         while (!list_empty(list)) {
1064                 struct request *rq;
1065
1066                 rq = list_first_entry(list, struct request, queuelist);
1067                 list_del_init(&rq->queuelist);
1068                 rq->mq_ctx = ctx;
1069                 __blk_mq_insert_request(hctx, rq, false);
1070         }
1071         spin_unlock(&ctx->lock);
1072
1073         blk_mq_run_hw_queue(hctx, from_schedule);
1074         blk_mq_put_ctx(current_ctx);
1075 }
1076
1077 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1078 {
1079         struct request *rqa = container_of(a, struct request, queuelist);
1080         struct request *rqb = container_of(b, struct request, queuelist);
1081
1082         return !(rqa->mq_ctx < rqb->mq_ctx ||
1083                  (rqa->mq_ctx == rqb->mq_ctx &&
1084                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1085 }
1086
1087 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1088 {
1089         struct blk_mq_ctx *this_ctx;
1090         struct request_queue *this_q;
1091         struct request *rq;
1092         LIST_HEAD(list);
1093         LIST_HEAD(ctx_list);
1094         unsigned int depth;
1095
1096         list_splice_init(&plug->mq_list, &list);
1097
1098         list_sort(NULL, &list, plug_ctx_cmp);
1099
1100         this_q = NULL;
1101         this_ctx = NULL;
1102         depth = 0;
1103
1104         while (!list_empty(&list)) {
1105                 rq = list_entry_rq(list.next);
1106                 list_del_init(&rq->queuelist);
1107                 BUG_ON(!rq->q);
1108                 if (rq->mq_ctx != this_ctx) {
1109                         if (this_ctx) {
1110                                 blk_mq_insert_requests(this_q, this_ctx,
1111                                                         &ctx_list, depth,
1112                                                         from_schedule);
1113                         }
1114
1115                         this_ctx = rq->mq_ctx;
1116                         this_q = rq->q;
1117                         depth = 0;
1118                 }
1119
1120                 depth++;
1121                 list_add_tail(&rq->queuelist, &ctx_list);
1122         }
1123
1124         /*
1125          * If 'this_ctx' is set, we know we have entries to complete
1126          * on 'ctx_list'. Do those.
1127          */
1128         if (this_ctx) {
1129                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1130                                        from_schedule);
1131         }
1132 }
1133
1134 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1135 {
1136         init_request_from_bio(rq, bio);
1137
1138         if (blk_do_io_stat(rq))
1139                 blk_account_io_start(rq, 1);
1140 }
1141
1142 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1143 {
1144         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1145                 !blk_queue_nomerges(hctx->queue);
1146 }
1147
1148 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1149                                          struct blk_mq_ctx *ctx,
1150                                          struct request *rq, struct bio *bio)
1151 {
1152         if (!hctx_allow_merges(hctx)) {
1153                 blk_mq_bio_to_request(rq, bio);
1154                 spin_lock(&ctx->lock);
1155 insert_rq:
1156                 __blk_mq_insert_request(hctx, rq, false);
1157                 spin_unlock(&ctx->lock);
1158                 return false;
1159         } else {
1160                 struct request_queue *q = hctx->queue;
1161
1162                 spin_lock(&ctx->lock);
1163                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1164                         blk_mq_bio_to_request(rq, bio);
1165                         goto insert_rq;
1166                 }
1167
1168                 spin_unlock(&ctx->lock);
1169                 __blk_mq_free_request(hctx, ctx, rq);
1170                 return true;
1171         }
1172 }
1173
1174 struct blk_map_ctx {
1175         struct blk_mq_hw_ctx *hctx;
1176         struct blk_mq_ctx *ctx;
1177 };
1178
1179 static struct request *blk_mq_map_request(struct request_queue *q,
1180                                           struct bio *bio,
1181                                           struct blk_map_ctx *data)
1182 {
1183         struct blk_mq_hw_ctx *hctx;
1184         struct blk_mq_ctx *ctx;
1185         struct request *rq;
1186         int rw = bio_data_dir(bio);
1187         struct blk_mq_alloc_data alloc_data;
1188
1189         if (unlikely(blk_mq_queue_enter(q))) {
1190                 bio_endio(bio, -EIO);
1191                 return NULL;
1192         }
1193
1194         ctx = blk_mq_get_ctx(q);
1195         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1196
1197         if (rw_is_sync(bio->bi_rw))
1198                 rw |= REQ_SYNC;
1199
1200         trace_block_getrq(q, bio, rw);
1201         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1202                         hctx);
1203         rq = __blk_mq_alloc_request(&alloc_data, rw);
1204         if (unlikely(!rq)) {
1205                 __blk_mq_run_hw_queue(hctx);
1206                 blk_mq_put_ctx(ctx);
1207                 trace_block_sleeprq(q, bio, rw);
1208
1209                 ctx = blk_mq_get_ctx(q);
1210                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1211                 blk_mq_set_alloc_data(&alloc_data, q,
1212                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1213                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1214                 ctx = alloc_data.ctx;
1215                 hctx = alloc_data.hctx;
1216         }
1217
1218         hctx->queued++;
1219         data->hctx = hctx;
1220         data->ctx = ctx;
1221         return rq;
1222 }
1223
1224 /*
1225  * Multiple hardware queue variant. This will not use per-process plugs,
1226  * but will attempt to bypass the hctx queueing if we can go straight to
1227  * hardware for SYNC IO.
1228  */
1229 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1230 {
1231         const int is_sync = rw_is_sync(bio->bi_rw);
1232         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1233         struct blk_map_ctx data;
1234         struct request *rq;
1235
1236         blk_queue_bounce(q, &bio);
1237
1238         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1239                 bio_endio(bio, -EIO);
1240                 return;
1241         }
1242
1243         rq = blk_mq_map_request(q, bio, &data);
1244         if (unlikely(!rq))
1245                 return;
1246
1247         if (unlikely(is_flush_fua)) {
1248                 blk_mq_bio_to_request(rq, bio);
1249                 blk_insert_flush(rq);
1250                 goto run_queue;
1251         }
1252
1253         /*
1254          * If the driver supports defer issued based on 'last', then
1255          * queue it up like normal since we can potentially save some
1256          * CPU this way.
1257          */
1258         if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1259                 struct blk_mq_queue_data bd = {
1260                         .rq = rq,
1261                         .list = NULL,
1262                         .last = 1
1263                 };
1264                 int ret;
1265
1266                 blk_mq_bio_to_request(rq, bio);
1267
1268                 /*
1269                  * For OK queue, we are done. For error, kill it. Any other
1270                  * error (busy), just add it to our list as we previously
1271                  * would have done
1272                  */
1273                 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1274                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1275                         goto done;
1276                 else {
1277                         __blk_mq_requeue_request(rq);
1278
1279                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1280                                 rq->errors = -EIO;
1281                                 blk_mq_end_request(rq, rq->errors);
1282                                 goto done;
1283                         }
1284                 }
1285         }
1286
1287         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1288                 /*
1289                  * For a SYNC request, send it to the hardware immediately. For
1290                  * an ASYNC request, just ensure that we run it later on. The
1291                  * latter allows for merging opportunities and more efficient
1292                  * dispatching.
1293                  */
1294 run_queue:
1295                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1296         }
1297 done:
1298         blk_mq_put_ctx(data.ctx);
1299 }
1300
1301 /*
1302  * Single hardware queue variant. This will attempt to use any per-process
1303  * plug for merging and IO deferral.
1304  */
1305 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1306 {
1307         const int is_sync = rw_is_sync(bio->bi_rw);
1308         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1309         unsigned int use_plug, request_count = 0;
1310         struct blk_map_ctx data;
1311         struct request *rq;
1312
1313         /*
1314          * If we have multiple hardware queues, just go directly to
1315          * one of those for sync IO.
1316          */
1317         use_plug = !is_flush_fua && !is_sync;
1318
1319         blk_queue_bounce(q, &bio);
1320
1321         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1322                 bio_endio(bio, -EIO);
1323                 return;
1324         }
1325
1326         if (use_plug && !blk_queue_nomerges(q) &&
1327             blk_attempt_plug_merge(q, bio, &request_count))
1328                 return;
1329
1330         rq = blk_mq_map_request(q, bio, &data);
1331         if (unlikely(!rq))
1332                 return;
1333
1334         if (unlikely(is_flush_fua)) {
1335                 blk_mq_bio_to_request(rq, bio);
1336                 blk_insert_flush(rq);
1337                 goto run_queue;
1338         }
1339
1340         /*
1341          * A task plug currently exists. Since this is completely lockless,
1342          * utilize that to temporarily store requests until the task is
1343          * either done or scheduled away.
1344          */
1345         if (use_plug) {
1346                 struct blk_plug *plug = current->plug;
1347
1348                 if (plug) {
1349                         blk_mq_bio_to_request(rq, bio);
1350                         if (list_empty(&plug->mq_list))
1351                                 trace_block_plug(q);
1352                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1353                                 blk_flush_plug_list(plug, false);
1354                                 trace_block_plug(q);
1355                         }
1356                         list_add_tail(&rq->queuelist, &plug->mq_list);
1357                         blk_mq_put_ctx(data.ctx);
1358                         return;
1359                 }
1360         }
1361
1362         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1363                 /*
1364                  * For a SYNC request, send it to the hardware immediately. For
1365                  * an ASYNC request, just ensure that we run it later on. The
1366                  * latter allows for merging opportunities and more efficient
1367                  * dispatching.
1368                  */
1369 run_queue:
1370                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1371         }
1372
1373         blk_mq_put_ctx(data.ctx);
1374 }
1375
1376 /*
1377  * Default mapping to a software queue, since we use one per CPU.
1378  */
1379 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1380 {
1381         return q->queue_hw_ctx[q->mq_map[cpu]];
1382 }
1383 EXPORT_SYMBOL(blk_mq_map_queue);
1384
1385 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1386                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1387 {
1388         struct page *page;
1389
1390         if (tags->rqs && set->ops->exit_request) {
1391                 int i;
1392
1393                 for (i = 0; i < tags->nr_tags; i++) {
1394                         if (!tags->rqs[i])
1395                                 continue;
1396                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1397                                                 hctx_idx, i);
1398                         tags->rqs[i] = NULL;
1399                 }
1400         }
1401
1402         while (!list_empty(&tags->page_list)) {
1403                 page = list_first_entry(&tags->page_list, struct page, lru);
1404                 list_del_init(&page->lru);
1405                 __free_pages(page, page->private);
1406         }
1407
1408         kfree(tags->rqs);
1409
1410         blk_mq_free_tags(tags);
1411 }
1412
1413 static size_t order_to_size(unsigned int order)
1414 {
1415         return (size_t)PAGE_SIZE << order;
1416 }
1417
1418 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1419                 unsigned int hctx_idx)
1420 {
1421         struct blk_mq_tags *tags;
1422         unsigned int i, j, entries_per_page, max_order = 4;
1423         size_t rq_size, left;
1424
1425         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1426                                 set->numa_node);
1427         if (!tags)
1428                 return NULL;
1429
1430         INIT_LIST_HEAD(&tags->page_list);
1431
1432         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1433                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1434                                  set->numa_node);
1435         if (!tags->rqs) {
1436                 blk_mq_free_tags(tags);
1437                 return NULL;
1438         }
1439
1440         /*
1441          * rq_size is the size of the request plus driver payload, rounded
1442          * to the cacheline size
1443          */
1444         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1445                                 cache_line_size());
1446         left = rq_size * set->queue_depth;
1447
1448         for (i = 0; i < set->queue_depth; ) {
1449                 int this_order = max_order;
1450                 struct page *page;
1451                 int to_do;
1452                 void *p;
1453
1454                 while (left < order_to_size(this_order - 1) && this_order)
1455                         this_order--;
1456
1457                 do {
1458                         page = alloc_pages_node(set->numa_node,
1459                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1460                                 this_order);
1461                         if (page)
1462                                 break;
1463                         if (!this_order--)
1464                                 break;
1465                         if (order_to_size(this_order) < rq_size)
1466                                 break;
1467                 } while (1);
1468
1469                 if (!page)
1470                         goto fail;
1471
1472                 page->private = this_order;
1473                 list_add_tail(&page->lru, &tags->page_list);
1474
1475                 p = page_address(page);
1476                 entries_per_page = order_to_size(this_order) / rq_size;
1477                 to_do = min(entries_per_page, set->queue_depth - i);
1478                 left -= to_do * rq_size;
1479                 for (j = 0; j < to_do; j++) {
1480                         tags->rqs[i] = p;
1481                         tags->rqs[i]->atomic_flags = 0;
1482                         tags->rqs[i]->cmd_flags = 0;
1483                         if (set->ops->init_request) {
1484                                 if (set->ops->init_request(set->driver_data,
1485                                                 tags->rqs[i], hctx_idx, i,
1486                                                 set->numa_node)) {
1487                                         tags->rqs[i] = NULL;
1488                                         goto fail;
1489                                 }
1490                         }
1491
1492                         p += rq_size;
1493                         i++;
1494                 }
1495         }
1496
1497         return tags;
1498
1499 fail:
1500         blk_mq_free_rq_map(set, tags, hctx_idx);
1501         return NULL;
1502 }
1503
1504 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1505 {
1506         kfree(bitmap->map);
1507 }
1508
1509 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1510 {
1511         unsigned int bpw = 8, total, num_maps, i;
1512
1513         bitmap->bits_per_word = bpw;
1514
1515         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1516         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1517                                         GFP_KERNEL, node);
1518         if (!bitmap->map)
1519                 return -ENOMEM;
1520
1521         bitmap->map_size = num_maps;
1522
1523         total = nr_cpu_ids;
1524         for (i = 0; i < num_maps; i++) {
1525                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1526                 total -= bitmap->map[i].depth;
1527         }
1528
1529         return 0;
1530 }
1531
1532 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1533 {
1534         struct request_queue *q = hctx->queue;
1535         struct blk_mq_ctx *ctx;
1536         LIST_HEAD(tmp);
1537
1538         /*
1539          * Move ctx entries to new CPU, if this one is going away.
1540          */
1541         ctx = __blk_mq_get_ctx(q, cpu);
1542
1543         spin_lock(&ctx->lock);
1544         if (!list_empty(&ctx->rq_list)) {
1545                 list_splice_init(&ctx->rq_list, &tmp);
1546                 blk_mq_hctx_clear_pending(hctx, ctx);
1547         }
1548         spin_unlock(&ctx->lock);
1549
1550         if (list_empty(&tmp))
1551                 return NOTIFY_OK;
1552
1553         ctx = blk_mq_get_ctx(q);
1554         spin_lock(&ctx->lock);
1555
1556         while (!list_empty(&tmp)) {
1557                 struct request *rq;
1558
1559                 rq = list_first_entry(&tmp, struct request, queuelist);
1560                 rq->mq_ctx = ctx;
1561                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1562         }
1563
1564         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1565         blk_mq_hctx_mark_pending(hctx, ctx);
1566
1567         spin_unlock(&ctx->lock);
1568
1569         blk_mq_run_hw_queue(hctx, true);
1570         blk_mq_put_ctx(ctx);
1571         return NOTIFY_OK;
1572 }
1573
1574 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1575 {
1576         struct request_queue *q = hctx->queue;
1577         struct blk_mq_tag_set *set = q->tag_set;
1578
1579         if (set->tags[hctx->queue_num])
1580                 return NOTIFY_OK;
1581
1582         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1583         if (!set->tags[hctx->queue_num])
1584                 return NOTIFY_STOP;
1585
1586         hctx->tags = set->tags[hctx->queue_num];
1587         return NOTIFY_OK;
1588 }
1589
1590 static int blk_mq_hctx_notify(void *data, unsigned long action,
1591                               unsigned int cpu)
1592 {
1593         struct blk_mq_hw_ctx *hctx = data;
1594
1595         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1596                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1597         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1598                 return blk_mq_hctx_cpu_online(hctx, cpu);
1599
1600         return NOTIFY_OK;
1601 }
1602
1603 static void blk_mq_exit_hctx(struct request_queue *q,
1604                 struct blk_mq_tag_set *set,
1605                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1606 {
1607         unsigned flush_start_tag = set->queue_depth;
1608
1609         blk_mq_tag_idle(hctx);
1610
1611         if (set->ops->exit_request)
1612                 set->ops->exit_request(set->driver_data,
1613                                        hctx->fq->flush_rq, hctx_idx,
1614                                        flush_start_tag + hctx_idx);
1615
1616         if (set->ops->exit_hctx)
1617                 set->ops->exit_hctx(hctx, hctx_idx);
1618
1619         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1620         blk_free_flush_queue(hctx->fq);
1621         kfree(hctx->ctxs);
1622         blk_mq_free_bitmap(&hctx->ctx_map);
1623 }
1624
1625 static void blk_mq_exit_hw_queues(struct request_queue *q,
1626                 struct blk_mq_tag_set *set, int nr_queue)
1627 {
1628         struct blk_mq_hw_ctx *hctx;
1629         unsigned int i;
1630
1631         queue_for_each_hw_ctx(q, hctx, i) {
1632                 if (i == nr_queue)
1633                         break;
1634                 blk_mq_exit_hctx(q, set, hctx, i);
1635         }
1636 }
1637
1638 static void blk_mq_free_hw_queues(struct request_queue *q,
1639                 struct blk_mq_tag_set *set)
1640 {
1641         struct blk_mq_hw_ctx *hctx;
1642         unsigned int i;
1643
1644         queue_for_each_hw_ctx(q, hctx, i)
1645                 free_cpumask_var(hctx->cpumask);
1646 }
1647
1648 static int blk_mq_init_hctx(struct request_queue *q,
1649                 struct blk_mq_tag_set *set,
1650                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1651 {
1652         int node;
1653         unsigned flush_start_tag = set->queue_depth;
1654
1655         node = hctx->numa_node;
1656         if (node == NUMA_NO_NODE)
1657                 node = hctx->numa_node = set->numa_node;
1658
1659         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1660         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1661         spin_lock_init(&hctx->lock);
1662         INIT_LIST_HEAD(&hctx->dispatch);
1663         hctx->queue = q;
1664         hctx->queue_num = hctx_idx;
1665         hctx->flags = set->flags;
1666
1667         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1668                                         blk_mq_hctx_notify, hctx);
1669         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1670
1671         hctx->tags = set->tags[hctx_idx];
1672
1673         /*
1674          * Allocate space for all possible cpus to avoid allocation at
1675          * runtime
1676          */
1677         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1678                                         GFP_KERNEL, node);
1679         if (!hctx->ctxs)
1680                 goto unregister_cpu_notifier;
1681
1682         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1683                 goto free_ctxs;
1684
1685         hctx->nr_ctx = 0;
1686
1687         if (set->ops->init_hctx &&
1688             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1689                 goto free_bitmap;
1690
1691         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1692         if (!hctx->fq)
1693                 goto exit_hctx;
1694
1695         if (set->ops->init_request &&
1696             set->ops->init_request(set->driver_data,
1697                                    hctx->fq->flush_rq, hctx_idx,
1698                                    flush_start_tag + hctx_idx, node))
1699                 goto free_fq;
1700
1701         return 0;
1702
1703  free_fq:
1704         kfree(hctx->fq);
1705  exit_hctx:
1706         if (set->ops->exit_hctx)
1707                 set->ops->exit_hctx(hctx, hctx_idx);
1708  free_bitmap:
1709         blk_mq_free_bitmap(&hctx->ctx_map);
1710  free_ctxs:
1711         kfree(hctx->ctxs);
1712  unregister_cpu_notifier:
1713         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1714
1715         return -1;
1716 }
1717
1718 static int blk_mq_init_hw_queues(struct request_queue *q,
1719                 struct blk_mq_tag_set *set)
1720 {
1721         struct blk_mq_hw_ctx *hctx;
1722         unsigned int i;
1723
1724         /*
1725          * Initialize hardware queues
1726          */
1727         queue_for_each_hw_ctx(q, hctx, i) {
1728                 if (blk_mq_init_hctx(q, set, hctx, i))
1729                         break;
1730         }
1731
1732         if (i == q->nr_hw_queues)
1733                 return 0;
1734
1735         /*
1736          * Init failed
1737          */
1738         blk_mq_exit_hw_queues(q, set, i);
1739
1740         return 1;
1741 }
1742
1743 static void blk_mq_init_cpu_queues(struct request_queue *q,
1744                                    unsigned int nr_hw_queues)
1745 {
1746         unsigned int i;
1747
1748         for_each_possible_cpu(i) {
1749                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1750                 struct blk_mq_hw_ctx *hctx;
1751
1752                 memset(__ctx, 0, sizeof(*__ctx));
1753                 __ctx->cpu = i;
1754                 spin_lock_init(&__ctx->lock);
1755                 INIT_LIST_HEAD(&__ctx->rq_list);
1756                 __ctx->queue = q;
1757
1758                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1759                 if (!cpu_online(i))
1760                         continue;
1761
1762                 hctx = q->mq_ops->map_queue(q, i);
1763                 cpumask_set_cpu(i, hctx->cpumask);
1764                 hctx->nr_ctx++;
1765
1766                 /*
1767                  * Set local node, IFF we have more than one hw queue. If
1768                  * not, we remain on the home node of the device
1769                  */
1770                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1771                         hctx->numa_node = cpu_to_node(i);
1772         }
1773 }
1774
1775 static void blk_mq_map_swqueue(struct request_queue *q)
1776 {
1777         unsigned int i;
1778         struct blk_mq_hw_ctx *hctx;
1779         struct blk_mq_ctx *ctx;
1780
1781         queue_for_each_hw_ctx(q, hctx, i) {
1782                 cpumask_clear(hctx->cpumask);
1783                 hctx->nr_ctx = 0;
1784         }
1785
1786         /*
1787          * Map software to hardware queues
1788          */
1789         queue_for_each_ctx(q, ctx, i) {
1790                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1791                 if (!cpu_online(i))
1792                         continue;
1793
1794                 hctx = q->mq_ops->map_queue(q, i);
1795                 cpumask_set_cpu(i, hctx->cpumask);
1796                 ctx->index_hw = hctx->nr_ctx;
1797                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1798         }
1799
1800         queue_for_each_hw_ctx(q, hctx, i) {
1801                 /*
1802                  * If no software queues are mapped to this hardware queue,
1803                  * disable it and free the request entries.
1804                  */
1805                 if (!hctx->nr_ctx) {
1806                         struct blk_mq_tag_set *set = q->tag_set;
1807
1808                         if (set->tags[i]) {
1809                                 blk_mq_free_rq_map(set, set->tags[i], i);
1810                                 set->tags[i] = NULL;
1811                                 hctx->tags = NULL;
1812                         }
1813                         continue;
1814                 }
1815
1816                 /*
1817                  * Initialize batch roundrobin counts
1818                  */
1819                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1820                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1821         }
1822 }
1823
1824 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1825 {
1826         struct blk_mq_hw_ctx *hctx;
1827         struct request_queue *q;
1828         bool shared;
1829         int i;
1830
1831         if (set->tag_list.next == set->tag_list.prev)
1832                 shared = false;
1833         else
1834                 shared = true;
1835
1836         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1837                 blk_mq_freeze_queue(q);
1838
1839                 queue_for_each_hw_ctx(q, hctx, i) {
1840                         if (shared)
1841                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1842                         else
1843                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1844                 }
1845                 blk_mq_unfreeze_queue(q);
1846         }
1847 }
1848
1849 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1850 {
1851         struct blk_mq_tag_set *set = q->tag_set;
1852
1853         mutex_lock(&set->tag_list_lock);
1854         list_del_init(&q->tag_set_list);
1855         blk_mq_update_tag_set_depth(set);
1856         mutex_unlock(&set->tag_list_lock);
1857 }
1858
1859 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1860                                      struct request_queue *q)
1861 {
1862         q->tag_set = set;
1863
1864         mutex_lock(&set->tag_list_lock);
1865         list_add_tail(&q->tag_set_list, &set->tag_list);
1866         blk_mq_update_tag_set_depth(set);
1867         mutex_unlock(&set->tag_list_lock);
1868 }
1869
1870 /*
1871  * It is the actual release handler for mq, but we do it from
1872  * request queue's release handler for avoiding use-after-free
1873  * and headache because q->mq_kobj shouldn't have been introduced,
1874  * but we can't group ctx/kctx kobj without it.
1875  */
1876 void blk_mq_release(struct request_queue *q)
1877 {
1878         struct blk_mq_hw_ctx *hctx;
1879         unsigned int i;
1880
1881         /* hctx kobj stays in hctx */
1882         queue_for_each_hw_ctx(q, hctx, i)
1883                 kfree(hctx);
1884
1885         kfree(q->queue_hw_ctx);
1886
1887         /* ctx kobj stays in queue_ctx */
1888         free_percpu(q->queue_ctx);
1889 }
1890
1891 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1892 {
1893         struct blk_mq_hw_ctx **hctxs;
1894         struct blk_mq_ctx __percpu *ctx;
1895         struct request_queue *q;
1896         unsigned int *map;
1897         int i;
1898
1899         ctx = alloc_percpu(struct blk_mq_ctx);
1900         if (!ctx)
1901                 return ERR_PTR(-ENOMEM);
1902
1903         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1904                         set->numa_node);
1905
1906         if (!hctxs)
1907                 goto err_percpu;
1908
1909         map = blk_mq_make_queue_map(set);
1910         if (!map)
1911                 goto err_map;
1912
1913         for (i = 0; i < set->nr_hw_queues; i++) {
1914                 int node = blk_mq_hw_queue_to_node(map, i);
1915
1916                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1917                                         GFP_KERNEL, node);
1918                 if (!hctxs[i])
1919                         goto err_hctxs;
1920
1921                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1922                                                 node))
1923                         goto err_hctxs;
1924
1925                 atomic_set(&hctxs[i]->nr_active, 0);
1926                 hctxs[i]->numa_node = node;
1927                 hctxs[i]->queue_num = i;
1928         }
1929
1930         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1931         if (!q)
1932                 goto err_hctxs;
1933
1934         /*
1935          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1936          * See blk_register_queue() for details.
1937          */
1938         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1939                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1940                 goto err_map;
1941
1942         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1943         blk_queue_rq_timeout(q, 30000);
1944
1945         q->nr_queues = nr_cpu_ids;
1946         q->nr_hw_queues = set->nr_hw_queues;
1947         q->mq_map = map;
1948
1949         q->queue_ctx = ctx;
1950         q->queue_hw_ctx = hctxs;
1951
1952         q->mq_ops = set->ops;
1953         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1954
1955         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1956                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1957
1958         q->sg_reserved_size = INT_MAX;
1959
1960         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1961         INIT_LIST_HEAD(&q->requeue_list);
1962         spin_lock_init(&q->requeue_lock);
1963
1964         if (q->nr_hw_queues > 1)
1965                 blk_queue_make_request(q, blk_mq_make_request);
1966         else
1967                 blk_queue_make_request(q, blk_sq_make_request);
1968
1969         if (set->timeout)
1970                 blk_queue_rq_timeout(q, set->timeout);
1971
1972         /*
1973          * Do this after blk_queue_make_request() overrides it...
1974          */
1975         q->nr_requests = set->queue_depth;
1976
1977         if (set->ops->complete)
1978                 blk_queue_softirq_done(q, set->ops->complete);
1979
1980         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1981
1982         if (blk_mq_init_hw_queues(q, set))
1983                 goto err_hw;
1984
1985         mutex_lock(&all_q_mutex);
1986         list_add_tail(&q->all_q_node, &all_q_list);
1987         mutex_unlock(&all_q_mutex);
1988
1989         blk_mq_add_queue_tag_set(set, q);
1990
1991         blk_mq_map_swqueue(q);
1992
1993         return q;
1994
1995 err_hw:
1996         blk_cleanup_queue(q);
1997 err_hctxs:
1998         kfree(map);
1999         for (i = 0; i < set->nr_hw_queues; i++) {
2000                 if (!hctxs[i])
2001                         break;
2002                 free_cpumask_var(hctxs[i]->cpumask);
2003                 kfree(hctxs[i]);
2004         }
2005 err_map:
2006         kfree(hctxs);
2007 err_percpu:
2008         free_percpu(ctx);
2009         return ERR_PTR(-ENOMEM);
2010 }
2011 EXPORT_SYMBOL(blk_mq_init_queue);
2012
2013 void blk_mq_free_queue(struct request_queue *q)
2014 {
2015         struct blk_mq_tag_set   *set = q->tag_set;
2016
2017         blk_mq_del_queue_tag_set(q);
2018
2019         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2020         blk_mq_free_hw_queues(q, set);
2021
2022         percpu_ref_exit(&q->mq_usage_counter);
2023
2024         kfree(q->mq_map);
2025
2026         q->mq_map = NULL;
2027
2028         mutex_lock(&all_q_mutex);
2029         list_del_init(&q->all_q_node);
2030         mutex_unlock(&all_q_mutex);
2031 }
2032
2033 /* Basically redo blk_mq_init_queue with queue frozen */
2034 static void blk_mq_queue_reinit(struct request_queue *q)
2035 {
2036         WARN_ON_ONCE(!q->mq_freeze_depth);
2037
2038         blk_mq_sysfs_unregister(q);
2039
2040         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2041
2042         /*
2043          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2044          * we should change hctx numa_node according to new topology (this
2045          * involves free and re-allocate memory, worthy doing?)
2046          */
2047
2048         blk_mq_map_swqueue(q);
2049
2050         blk_mq_sysfs_register(q);
2051 }
2052
2053 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2054                                       unsigned long action, void *hcpu)
2055 {
2056         struct request_queue *q;
2057
2058         /*
2059          * Before new mappings are established, hotadded cpu might already
2060          * start handling requests. This doesn't break anything as we map
2061          * offline CPUs to first hardware queue. We will re-init the queue
2062          * below to get optimal settings.
2063          */
2064         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2065             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2066                 return NOTIFY_OK;
2067
2068         mutex_lock(&all_q_mutex);
2069
2070         /*
2071          * We need to freeze and reinit all existing queues.  Freezing
2072          * involves synchronous wait for an RCU grace period and doing it
2073          * one by one may take a long time.  Start freezing all queues in
2074          * one swoop and then wait for the completions so that freezing can
2075          * take place in parallel.
2076          */
2077         list_for_each_entry(q, &all_q_list, all_q_node)
2078                 blk_mq_freeze_queue_start(q);
2079         list_for_each_entry(q, &all_q_list, all_q_node)
2080                 blk_mq_freeze_queue_wait(q);
2081
2082         list_for_each_entry(q, &all_q_list, all_q_node)
2083                 blk_mq_queue_reinit(q);
2084
2085         list_for_each_entry(q, &all_q_list, all_q_node)
2086                 blk_mq_unfreeze_queue(q);
2087
2088         mutex_unlock(&all_q_mutex);
2089         return NOTIFY_OK;
2090 }
2091
2092 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2093 {
2094         int i;
2095
2096         for (i = 0; i < set->nr_hw_queues; i++) {
2097                 set->tags[i] = blk_mq_init_rq_map(set, i);
2098                 if (!set->tags[i])
2099                         goto out_unwind;
2100         }
2101
2102         return 0;
2103
2104 out_unwind:
2105         while (--i >= 0)
2106                 blk_mq_free_rq_map(set, set->tags[i], i);
2107
2108         return -ENOMEM;
2109 }
2110
2111 /*
2112  * Allocate the request maps associated with this tag_set. Note that this
2113  * may reduce the depth asked for, if memory is tight. set->queue_depth
2114  * will be updated to reflect the allocated depth.
2115  */
2116 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2117 {
2118         unsigned int depth;
2119         int err;
2120
2121         depth = set->queue_depth;
2122         do {
2123                 err = __blk_mq_alloc_rq_maps(set);
2124                 if (!err)
2125                         break;
2126
2127                 set->queue_depth >>= 1;
2128                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2129                         err = -ENOMEM;
2130                         break;
2131                 }
2132         } while (set->queue_depth);
2133
2134         if (!set->queue_depth || err) {
2135                 pr_err("blk-mq: failed to allocate request map\n");
2136                 return -ENOMEM;
2137         }
2138
2139         if (depth != set->queue_depth)
2140                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2141                                                 depth, set->queue_depth);
2142
2143         return 0;
2144 }
2145
2146 /*
2147  * Alloc a tag set to be associated with one or more request queues.
2148  * May fail with EINVAL for various error conditions. May adjust the
2149  * requested depth down, if if it too large. In that case, the set
2150  * value will be stored in set->queue_depth.
2151  */
2152 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2153 {
2154         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2155
2156         if (!set->nr_hw_queues)
2157                 return -EINVAL;
2158         if (!set->queue_depth)
2159                 return -EINVAL;
2160         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2161                 return -EINVAL;
2162
2163         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2164                 return -EINVAL;
2165
2166         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2167                 pr_info("blk-mq: reduced tag depth to %u\n",
2168                         BLK_MQ_MAX_DEPTH);
2169                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2170         }
2171
2172         /*
2173          * If a crashdump is active, then we are potentially in a very
2174          * memory constrained environment. Limit us to 1 queue and
2175          * 64 tags to prevent using too much memory.
2176          */
2177         if (is_kdump_kernel()) {
2178                 set->nr_hw_queues = 1;
2179                 set->queue_depth = min(64U, set->queue_depth);
2180         }
2181
2182         set->tags = kmalloc_node(set->nr_hw_queues *
2183                                  sizeof(struct blk_mq_tags *),
2184                                  GFP_KERNEL, set->numa_node);
2185         if (!set->tags)
2186                 return -ENOMEM;
2187
2188         if (blk_mq_alloc_rq_maps(set))
2189                 goto enomem;
2190
2191         mutex_init(&set->tag_list_lock);
2192         INIT_LIST_HEAD(&set->tag_list);
2193
2194         return 0;
2195 enomem:
2196         kfree(set->tags);
2197         set->tags = NULL;
2198         return -ENOMEM;
2199 }
2200 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2201
2202 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2203 {
2204         int i;
2205
2206         for (i = 0; i < set->nr_hw_queues; i++) {
2207                 if (set->tags[i])
2208                         blk_mq_free_rq_map(set, set->tags[i], i);
2209         }
2210
2211         kfree(set->tags);
2212         set->tags = NULL;
2213 }
2214 EXPORT_SYMBOL(blk_mq_free_tag_set);
2215
2216 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2217 {
2218         struct blk_mq_tag_set *set = q->tag_set;
2219         struct blk_mq_hw_ctx *hctx;
2220         int i, ret;
2221
2222         if (!set || nr > set->queue_depth)
2223                 return -EINVAL;
2224
2225         ret = 0;
2226         queue_for_each_hw_ctx(q, hctx, i) {
2227                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2228                 if (ret)
2229                         break;
2230         }
2231
2232         if (!ret)
2233                 q->nr_requests = nr;
2234
2235         return ret;
2236 }
2237
2238 void blk_mq_disable_hotplug(void)
2239 {
2240         mutex_lock(&all_q_mutex);
2241 }
2242
2243 void blk_mq_enable_hotplug(void)
2244 {
2245         mutex_unlock(&all_q_mutex);
2246 }
2247
2248 static int __init blk_mq_init(void)
2249 {
2250         blk_mq_cpu_init();
2251
2252         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2253
2254         return 0;
2255 }
2256 subsys_initcall(blk_mq_init);