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