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.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, gfp_t gfp)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 if (!(gfp & __GFP_WAIT))
  89                         return -EBUSY;
  90
  91                 ret = wait_event_interruptible(q->mq_freeze_wq,
  92                                 !q->mq_freeze_depth || blk_queue_dying(q));
  93                 if (blk_queue_dying(q))
  94                         return -ENODEV;
  95                 if (ret)
  96                         return ret;
  97         }
  98 }
  99
 100 static void blk_mq_queue_exit(struct request_queue *q)
 101 {
 102         percpu_ref_put(&q->mq_usage_counter);
 103 }
 104
 105 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 106 {
 107         struct request_queue *q =
 108                 container_of(ref, struct request_queue, mq_usage_counter);
 109
 110         wake_up_all(&q->mq_freeze_wq);
 111 }
 112
 113 void blk_mq_freeze_queue_start(struct request_queue *q)
 114 {
 115         bool freeze;
 116
 117         spin_lock_irq(q->queue_lock);
 118         freeze = !q->mq_freeze_depth++;
 119         spin_unlock_irq(q->queue_lock);
 120
 121         if (freeze) {
 122                 percpu_ref_kill(&q->mq_usage_counter);
 123                 blk_mq_run_hw_queues(q, false);
 124         }
 125 }
 126 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 127
 128 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 129 {
 130         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 131 }
 132
 133 /*
 134  * Guarantee no request is in use, so we can change any data structure of
 135  * the queue afterward.
 136  */
 137 void blk_mq_freeze_queue(struct request_queue *q)
 138 {
 139         blk_mq_freeze_queue_start(q);
 140         blk_mq_freeze_queue_wait(q);
 141 }
 142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 143
 144 void blk_mq_unfreeze_queue(struct request_queue *q)
 145 {
 146         bool wake;
 147
 148         spin_lock_irq(q->queue_lock);
 149         wake = !--q->mq_freeze_depth;
 150         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 151         spin_unlock_irq(q->queue_lock);
 152         if (wake) {
 153                 percpu_ref_reinit(&q->mq_usage_counter);
 154                 wake_up_all(&q->mq_freeze_wq);
 155         }
 156 }
 157 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 158
 159 void blk_mq_wake_waiters(struct request_queue *q)
 160 {
 161         struct blk_mq_hw_ctx *hctx;
 162         unsigned int i;
 163
 164         queue_for_each_hw_ctx(q, hctx, i)
 165                 if (blk_mq_hw_queue_mapped(hctx))
 166                         blk_mq_tag_wakeup_all(hctx->tags, true);
 167
 168         /*
 169          * If we are called because the queue has now been marked as
 170          * dying, we need to ensure that processes currently waiting on
 171          * the queue are notified as well.
 172          */
 173         wake_up_all(&q->mq_freeze_wq);
 174 }
 175
 176 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 177 {
 178         return blk_mq_has_free_tags(hctx->tags);
 179 }
 180 EXPORT_SYMBOL(blk_mq_can_queue);
 181
 182 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 183                                struct request *rq, unsigned int rw_flags)
 184 {
 185         if (blk_queue_io_stat(q))
 186                 rw_flags |= REQ_IO_STAT;
 187
 188         INIT_LIST_HEAD(&rq->queuelist);
 189         /* csd/requeue_work/fifo_time is initialized before use */
 190         rq->q = q;
 191         rq->mq_ctx = ctx;
 192         rq->cmd_flags |= rw_flags;
 193         /* do not touch atomic flags, it needs atomic ops against the timer */
 194         rq->cpu = -1;
 195         INIT_HLIST_NODE(&rq->hash);
 196         RB_CLEAR_NODE(&rq->rb_node);
 197         rq->rq_disk = NULL;
 198         rq->part = NULL;
 199         rq->start_time = jiffies;
 200 #ifdef CONFIG_BLK_CGROUP
 201         rq->rl = NULL;
 202         set_start_time_ns(rq);
 203         rq->io_start_time_ns = 0;
 204 #endif
 205         rq->nr_phys_segments = 0;
 206 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 207         rq->nr_integrity_segments = 0;
 208 #endif
 209         rq->special = NULL;
 210         /* tag was already set */
 211         rq->errors = 0;
 212
 213         rq->cmd = rq->__cmd;
 214
 215         rq->extra_len = 0;
 216         rq->sense_len = 0;
 217         rq->resid_len = 0;
 218         rq->sense = NULL;
 219
 220         INIT_LIST_HEAD(&rq->timeout_list);
 221         rq->timeout = 0;
 222
 223         rq->end_io = NULL;
 224         rq->end_io_data = NULL;
 225         rq->next_rq = NULL;
 226
 227         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 228 }
 229
 230 static struct request *
 231 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 232 {
 233         struct request *rq;
 234         unsigned int tag;
 235
 236         tag = blk_mq_get_tag(data);
 237         if (tag != BLK_MQ_TAG_FAIL) {
 238                 rq = data->hctx->tags->rqs[tag];
 239
 240                 if (blk_mq_tag_busy(data->hctx)) {
 241                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 242                         atomic_inc(&data->hctx->nr_active);
 243                 }
 244
 245                 rq->tag = tag;
 246                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 247                 return rq;
 248         }
 249
 250         return NULL;
 251 }
 252
 253 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 254                 bool reserved)
 255 {
 256         struct blk_mq_ctx *ctx;
 257         struct blk_mq_hw_ctx *hctx;
 258         struct request *rq;
 259         struct blk_mq_alloc_data alloc_data;
 260         int ret;
 261
 262         ret = blk_mq_queue_enter(q, gfp);
 263         if (ret)
 264                 return ERR_PTR(ret);
 265
 266         ctx = blk_mq_get_ctx(q);
 267         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 268         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 269                         reserved, ctx, hctx);
 270
 271         rq = __blk_mq_alloc_request(&alloc_data, rw);
 272         if (!rq && (gfp & __GFP_WAIT)) {
 273                 __blk_mq_run_hw_queue(hctx);
 274                 blk_mq_put_ctx(ctx);
 275
 276                 ctx = blk_mq_get_ctx(q);
 277                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 278                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 279                                 hctx);
 280                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 281                 ctx = alloc_data.ctx;
 282         }
 283         blk_mq_put_ctx(ctx);
 284         if (!rq) {
 285                 blk_mq_queue_exit(q);
 286                 return ERR_PTR(-EWOULDBLOCK);
 287         }
 288         return rq;
 289 }
 290 EXPORT_SYMBOL(blk_mq_alloc_request);
 291
 292 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 293                                   struct blk_mq_ctx *ctx, struct request *rq)
 294 {
 295         const int tag = rq->tag;
 296         struct request_queue *q = rq->q;
 297
 298         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 299                 atomic_dec(&hctx->nr_active);
 300         rq->cmd_flags = 0;
 301
 302         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 303         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 304         blk_mq_queue_exit(q);
 305 }
 306
 307 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 308 {
 309         struct blk_mq_ctx *ctx = rq->mq_ctx;
 310
 311         ctx->rq_completed[rq_is_sync(rq)]++;
 312         __blk_mq_free_request(hctx, ctx, rq);
 313
 314 }
 315 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 316
 317 void blk_mq_free_request(struct request *rq)
 318 {
 319         struct blk_mq_hw_ctx *hctx;
 320         struct request_queue *q = rq->q;
 321
 322         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 323         blk_mq_free_hctx_request(hctx, rq);
 324 }
 325 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 326
 327 inline void __blk_mq_end_request(struct request *rq, int error)
 328 {
 329         blk_account_io_done(rq);
 330
 331         if (rq->end_io) {
 332                 rq->end_io(rq, error);
 333         } else {
 334                 if (unlikely(blk_bidi_rq(rq)))
 335                         blk_mq_free_request(rq->next_rq);
 336                 blk_mq_free_request(rq);
 337         }
 338 }
 339 EXPORT_SYMBOL(__blk_mq_end_request);
 340
 341 void blk_mq_end_request(struct request *rq, int error)
 342 {
 343         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 344                 BUG();
 345         __blk_mq_end_request(rq, error);
 346 }
 347 EXPORT_SYMBOL(blk_mq_end_request);
 348
 349 static void __blk_mq_complete_request_remote(void *data)
 350 {
 351         struct request *rq = data;
 352
 353         rq->q->softirq_done_fn(rq);
 354 }
 355
 356 static void blk_mq_ipi_complete_request(struct request *rq)
 357 {
 358         struct blk_mq_ctx *ctx = rq->mq_ctx;
 359         bool shared = false;
 360         int cpu;
 361
 362         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 363                 rq->q->softirq_done_fn(rq);
 364                 return;
 365         }
 366
 367         cpu = get_cpu();
 368         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 369                 shared = cpus_share_cache(cpu, ctx->cpu);
 370
 371         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 372                 rq->csd.func = __blk_mq_complete_request_remote;
 373                 rq->csd.info = rq;
 374                 rq->csd.flags = 0;
 375                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 376         } else {
 377                 rq->q->softirq_done_fn(rq);
 378         }
 379         put_cpu();
 380 }
 381
 382 void __blk_mq_complete_request(struct request *rq)
 383 {
 384         struct request_queue *q = rq->q;
 385
 386         if (!q->softirq_done_fn)
 387                 blk_mq_end_request(rq, rq->errors);
 388         else
 389                 blk_mq_ipi_complete_request(rq);
 390 }
 391
 392 /**
 393  * blk_mq_complete_request - end I/O on a request
 394  * @rq:         the request being processed
 395  *
 396  * Description:
 397  *      Ends all I/O on a request. It does not handle partial completions.
 398  *      The actual completion happens out-of-order, through a IPI handler.
 399  **/
 400 void blk_mq_complete_request(struct request *rq)
 401 {
 402         struct request_queue *q = rq->q;
 403
 404         if (unlikely(blk_should_fake_timeout(q)))
 405                 return;
 406         if (!blk_mark_rq_complete(rq))
 407                 __blk_mq_complete_request(rq);
 408 }
 409 EXPORT_SYMBOL(blk_mq_complete_request);
 410
 411 int blk_mq_request_started(struct request *rq)
 412 {
 413         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 414 }
 415 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 416
 417 void blk_mq_start_request(struct request *rq)
 418 {
 419         struct request_queue *q = rq->q;
 420
 421         trace_block_rq_issue(q, rq);
 422
 423         rq->resid_len = blk_rq_bytes(rq);
 424         if (unlikely(blk_bidi_rq(rq)))
 425                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 426
 427         blk_add_timer(rq);
 428
 429         /*
 430          * Ensure that ->deadline is visible before set the started
 431          * flag and clear the completed flag.
 432          */
 433         smp_mb__before_atomic();
 434
 435         /*
 436          * Mark us as started and clear complete. Complete might have been
 437          * set if requeue raced with timeout, which then marked it as
 438          * complete. So be sure to clear complete again when we start
 439          * the request, otherwise we'll ignore the completion event.
 440          */
 441         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 442                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 443         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 444                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 445
 446         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 447                 /*
 448                  * Make sure space for the drain appears.  We know we can do
 449                  * this because max_hw_segments has been adjusted to be one
 450                  * fewer than the device can handle.
 451                  */
 452                 rq->nr_phys_segments++;
 453         }
 454 }
 455 EXPORT_SYMBOL(blk_mq_start_request);
 456
 457 static void __blk_mq_requeue_request(struct request *rq)
 458 {
 459         struct request_queue *q = rq->q;
 460
 461         trace_block_rq_requeue(q, rq);
 462
 463         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 464                 if (q->dma_drain_size && blk_rq_bytes(rq))
 465                         rq->nr_phys_segments--;
 466         }
 467 }
 468
 469 void blk_mq_requeue_request(struct request *rq)
 470 {
 471         __blk_mq_requeue_request(rq);
 472
 473         BUG_ON(blk_queued_rq(rq));
 474         blk_mq_add_to_requeue_list(rq, true);
 475 }
 476 EXPORT_SYMBOL(blk_mq_requeue_request);
 477
 478 static void blk_mq_requeue_work(struct work_struct *work)
 479 {
 480         struct request_queue *q =
 481                 container_of(work, struct request_queue, requeue_work);
 482         LIST_HEAD(rq_list);
 483         struct request *rq, *next;
 484         unsigned long flags;
 485
 486         spin_lock_irqsave(&q->requeue_lock, flags);
 487         list_splice_init(&q->requeue_list, &rq_list);
 488         spin_unlock_irqrestore(&q->requeue_lock, flags);
 489
 490         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 491                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 492                         continue;
 493
 494                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 495                 list_del_init(&rq->queuelist);
 496                 blk_mq_insert_request(rq, true, false, false);
 497         }
 498
 499         while (!list_empty(&rq_list)) {
 500                 rq = list_entry(rq_list.next, struct request, queuelist);
 501                 list_del_init(&rq->queuelist);
 502                 blk_mq_insert_request(rq, false, false, false);
 503         }
 504
 505         /*
 506          * Use the start variant of queue running here, so that running
 507          * the requeue work will kick stopped queues.
 508          */
 509         blk_mq_start_hw_queues(q);
 510 }
 511
 512 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 513 {
 514         struct request_queue *q = rq->q;
 515         unsigned long flags;
 516
 517         /*
 518          * We abuse this flag that is otherwise used by the I/O scheduler to
 519          * request head insertation from the workqueue.
 520          */
 521         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 522
 523         spin_lock_irqsave(&q->requeue_lock, flags);
 524         if (at_head) {
 525                 rq->cmd_flags |= REQ_SOFTBARRIER;
 526                 list_add(&rq->queuelist, &q->requeue_list);
 527         } else {
 528                 list_add_tail(&rq->queuelist, &q->requeue_list);
 529         }
 530         spin_unlock_irqrestore(&q->requeue_lock, flags);
 531 }
 532 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 533
 534 void blk_mq_cancel_requeue_work(struct request_queue *q)
 535 {
 536         cancel_work_sync(&q->requeue_work);
 537 }
 538 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 539
 540 void blk_mq_kick_requeue_list(struct request_queue *q)
 541 {
 542         kblockd_schedule_work(&q->requeue_work);
 543 }
 544 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 545
 546 void blk_mq_abort_requeue_list(struct request_queue *q)
 547 {
 548         unsigned long flags;
 549         LIST_HEAD(rq_list);
 550
 551         spin_lock_irqsave(&q->requeue_lock, flags);
 552         list_splice_init(&q->requeue_list, &rq_list);
 553         spin_unlock_irqrestore(&q->requeue_lock, flags);
 554
 555         while (!list_empty(&rq_list)) {
 556                 struct request *rq;
 557
 558                 rq = list_first_entry(&rq_list, struct request, queuelist);
 559                 list_del_init(&rq->queuelist);
 560                 rq->errors = -EIO;
 561                 blk_mq_end_request(rq, rq->errors);
 562         }
 563 }
 564 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 565
 566 static inline bool is_flush_request(struct request *rq,
 567                 struct blk_flush_queue *fq, unsigned int tag)
 568 {
 569         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 570                         fq->flush_rq->tag == tag);
 571 }
 572
 573 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 574 {
 575         struct request *rq = tags->rqs[tag];
 576         /* mq_ctx of flush rq is always cloned from the corresponding req */
 577         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 578
 579         if (!is_flush_request(rq, fq, tag))
 580                 return rq;
 581
 582         return fq->flush_rq;
 583 }
 584 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 585
 586 struct blk_mq_timeout_data {
 587         unsigned long next;
 588         unsigned int next_set;
 589 };
 590
 591 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 592 {
 593         struct blk_mq_ops *ops = req->q->mq_ops;
 594         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 595
 596         /*
 597          * We know that complete is set at this point. If STARTED isn't set
 598          * anymore, then the request isn't active and the "timeout" should
 599          * just be ignored. This can happen due to the bitflag ordering.
 600          * Timeout first checks if STARTED is set, and if it is, assumes
 601          * the request is active. But if we race with completion, then
 602          * we both flags will get cleared. So check here again, and ignore
 603          * a timeout event with a request that isn't active.
 604          */
 605         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 606                 return;
 607
 608         if (ops->timeout)
 609                 ret = ops->timeout(req, reserved);
 610
 611         switch (ret) {
 612         case BLK_EH_HANDLED:
 613                 __blk_mq_complete_request(req);
 614                 break;
 615         case BLK_EH_RESET_TIMER:
 616                 blk_add_timer(req);
 617                 blk_clear_rq_complete(req);
 618                 break;
 619         case BLK_EH_NOT_HANDLED:
 620                 break;
 621         default:
 622                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 623                 break;
 624         }
 625 }
 626
 627 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 628                 struct request *rq, void *priv, bool reserved)
 629 {
 630         struct blk_mq_timeout_data *data = priv;
 631
 632         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 633                 /*
 634                  * If a request wasn't started before the queue was
 635                  * marked dying, kill it here or it'll go unnoticed.
 636                  */
 637                 if (unlikely(blk_queue_dying(rq->q))) {
 638                         rq->errors = -EIO;
 639                         blk_mq_complete_request(rq);
 640                 }
 641                 return;
 642         }
 643         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 644                 return;
 645
 646         if (time_after_eq(jiffies, rq->deadline)) {
 647                 if (!blk_mark_rq_complete(rq))
 648                         blk_mq_rq_timed_out(rq, reserved);
 649         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 650                 data->next = rq->deadline;
 651                 data->next_set = 1;
 652         }
 653 }
 654
 655 static void blk_mq_rq_timer(unsigned long priv)
 656 {
 657         struct request_queue *q = (struct request_queue *)priv;
 658         struct blk_mq_timeout_data data = {
 659                 .next           = 0,
 660                 .next_set       = 0,
 661         };
 662         struct blk_mq_hw_ctx *hctx;
 663         int i;
 664
 665         queue_for_each_hw_ctx(q, hctx, i) {
 666                 /*
 667                  * If not software queues are currently mapped to this
 668                  * hardware queue, there's nothing to check
 669                  */
 670                 if (!blk_mq_hw_queue_mapped(hctx))
 671                         continue;
 672
 673                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 674         }
 675
 676         if (data.next_set) {
 677                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 678                 mod_timer(&q->timeout, data.next);
 679         } else {
 680                 queue_for_each_hw_ctx(q, hctx, i) {
 681                         /* the hctx may be unmapped, so check it here */
 682                         if (blk_mq_hw_queue_mapped(hctx))
 683                                 blk_mq_tag_idle(hctx);
 684                 }
 685         }
 686 }
 687
 688 /*
 689  * Reverse check our software queue for entries that we could potentially
 690  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 691  * too much time checking for merges.
 692  */
 693 static bool blk_mq_attempt_merge(struct request_queue *q,
 694                                  struct blk_mq_ctx *ctx, struct bio *bio)
 695 {
 696         struct request *rq;
 697         int checked = 8;
 698
 699         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 700                 int el_ret;
 701
 702                 if (!checked--)
 703                         break;
 704
 705                 if (!blk_rq_merge_ok(rq, bio))
 706                         continue;
 707
 708                 el_ret = blk_try_merge(rq, bio);
 709                 if (el_ret == ELEVATOR_BACK_MERGE) {
 710                         if (bio_attempt_back_merge(q, rq, bio)) {
 711                                 ctx->rq_merged++;
 712                                 return true;
 713                         }
 714                         break;
 715                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 716                         if (bio_attempt_front_merge(q, rq, bio)) {
 717                                 ctx->rq_merged++;
 718                                 return true;
 719                         }
 720                         break;
 721                 }
 722         }
 723
 724         return false;
 725 }
 726
 727 /*
 728  * Process software queues that have been marked busy, splicing them
 729  * to the for-dispatch
 730  */
 731 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 732 {
 733         struct blk_mq_ctx *ctx;
 734         int i;
 735
 736         for (i = 0; i < hctx->ctx_map.size; i++) {
 737                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 738                 unsigned int off, bit;
 739
 740                 if (!bm->word)
 741                         continue;
 742
 743                 bit = 0;
 744                 off = i * hctx->ctx_map.bits_per_word;
 745                 do {
 746                         bit = find_next_bit(&bm->word, bm->depth, bit);
 747                         if (bit >= bm->depth)
 748                                 break;
 749
 750                         ctx = hctx->ctxs[bit + off];
 751                         clear_bit(bit, &bm->word);
 752                         spin_lock(&ctx->lock);
 753                         list_splice_tail_init(&ctx->rq_list, list);
 754                         spin_unlock(&ctx->lock);
 755
 756                         bit++;
 757                 } while (1);
 758         }
 759 }
 760
 761 /*
 762  * Run this hardware queue, pulling any software queues mapped to it in.
 763  * Note that this function currently has various problems around ordering
 764  * of IO. In particular, we'd like FIFO behaviour on handling existing
 765  * items on the hctx->dispatch list. Ignore that for now.
 766  */
 767 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 768 {
 769         struct request_queue *q = hctx->queue;
 770         struct request *rq;
 771         LIST_HEAD(rq_list);
 772         LIST_HEAD(driver_list);
 773         struct list_head *dptr;
 774         int queued;
 775
 776         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 777
 778         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 779                 return;
 780
 781         hctx->run++;
 782
 783         /*
 784          * Touch any software queue that has pending entries.
 785          */
 786         flush_busy_ctxs(hctx, &rq_list);
 787
 788         /*
 789          * If we have previous entries on our dispatch list, grab them
 790          * and stuff them at the front for more fair dispatch.
 791          */
 792         if (!list_empty_careful(&hctx->dispatch)) {
 793                 spin_lock(&hctx->lock);
 794                 if (!list_empty(&hctx->dispatch))
 795                         list_splice_init(&hctx->dispatch, &rq_list);
 796                 spin_unlock(&hctx->lock);
 797         }
 798
 799         /*
 800          * Start off with dptr being NULL, so we start the first request
 801          * immediately, even if we have more pending.
 802          */
 803         dptr = NULL;
 804
 805         /*
 806          * Now process all the entries, sending them to the driver.
 807          */
 808         queued = 0;
 809         while (!list_empty(&rq_list)) {
 810                 struct blk_mq_queue_data bd;
 811                 int ret;
 812
 813                 rq = list_first_entry(&rq_list, struct request, queuelist);
 814                 list_del_init(&rq->queuelist);
 815
 816                 bd.rq = rq;
 817                 bd.list = dptr;
 818                 bd.last = list_empty(&rq_list);
 819
 820                 ret = q->mq_ops->queue_rq(hctx, &bd);
 821                 switch (ret) {
 822                 case BLK_MQ_RQ_QUEUE_OK:
 823                         queued++;
 824                         continue;
 825                 case BLK_MQ_RQ_QUEUE_BUSY:
 826                         list_add(&rq->queuelist, &rq_list);
 827                         __blk_mq_requeue_request(rq);
 828                         break;
 829                 default:
 830                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 831                 case BLK_MQ_RQ_QUEUE_ERROR:
 832                         rq->errors = -EIO;
 833                         blk_mq_end_request(rq, rq->errors);
 834                         break;
 835                 }
 836
 837                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 838                         break;
 839
 840                 /*
 841                  * We've done the first request. If we have more than 1
 842                  * left in the list, set dptr to defer issue.
 843                  */
 844                 if (!dptr && rq_list.next != rq_list.prev)
 845                         dptr = &driver_list;
 846         }
 847
 848         if (!queued)
 849                 hctx->dispatched[0]++;
 850         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 851                 hctx->dispatched[ilog2(queued) + 1]++;
 852
 853         /*
 854          * Any items that need requeuing? Stuff them into hctx->dispatch,
 855          * that is where we will continue on next queue run.
 856          */
 857         if (!list_empty(&rq_list)) {
 858                 spin_lock(&hctx->lock);
 859                 list_splice(&rq_list, &hctx->dispatch);
 860                 spin_unlock(&hctx->lock);
 861                 /*
 862                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 863                  * it's possible the queue is stopped and restarted again
 864                  * before this. Queue restart will dispatch requests. And since
 865                  * requests in rq_list aren't added into hctx->dispatch yet,
 866                  * the requests in rq_list might get lost.
 867                  *
 868                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 869                  **/
 870                 blk_mq_run_hw_queue(hctx, true);
 871         }
 872 }
 873
 874 /*
 875  * It'd be great if the workqueue API had a way to pass
 876  * in a mask and had some smarts for more clever placement.
 877  * For now we just round-robin here, switching for every
 878  * BLK_MQ_CPU_WORK_BATCH queued items.
 879  */
 880 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 881 {
 882         if (hctx->queue->nr_hw_queues == 1)
 883                 return WORK_CPU_UNBOUND;
 884
 885         if (--hctx->next_cpu_batch <= 0) {
 886                 int cpu = hctx->next_cpu, next_cpu;
 887
 888                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 889                 if (next_cpu >= nr_cpu_ids)
 890                         next_cpu = cpumask_first(hctx->cpumask);
 891
 892                 hctx->next_cpu = next_cpu;
 893                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 894
 895                 return cpu;
 896         }
 897
 898         return hctx->next_cpu;
 899 }
 900
 901 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 902 {
 903         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 904             !blk_mq_hw_queue_mapped(hctx)))
 905                 return;
 906
 907         if (!async) {
 908                 int cpu = get_cpu();
 909                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 910                         __blk_mq_run_hw_queue(hctx);
 911                         put_cpu();
 912                         return;
 913                 }
 914
 915                 put_cpu();
 916         }
 917
 918         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 919                         &hctx->run_work, 0);
 920 }
 921
 922 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 923 {
 924         struct blk_mq_hw_ctx *hctx;
 925         int i;
 926
 927         queue_for_each_hw_ctx(q, hctx, i) {
 928                 if ((!blk_mq_hctx_has_pending(hctx) &&
 929                     list_empty_careful(&hctx->dispatch)) ||
 930                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 931                         continue;
 932
 933                 blk_mq_run_hw_queue(hctx, async);
 934         }
 935 }
 936 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 937
 938 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 939 {
 940         cancel_delayed_work(&hctx->run_work);
 941         cancel_delayed_work(&hctx->delay_work);
 942         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 943 }
 944 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 945
 946 void blk_mq_stop_hw_queues(struct request_queue *q)
 947 {
 948         struct blk_mq_hw_ctx *hctx;
 949         int i;
 950
 951         queue_for_each_hw_ctx(q, hctx, i)
 952                 blk_mq_stop_hw_queue(hctx);
 953 }
 954 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 955
 956 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 957 {
 958         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 959
 960         blk_mq_run_hw_queue(hctx, false);
 961 }
 962 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 963
 964 void blk_mq_start_hw_queues(struct request_queue *q)
 965 {
 966         struct blk_mq_hw_ctx *hctx;
 967         int i;
 968
 969         queue_for_each_hw_ctx(q, hctx, i)
 970                 blk_mq_start_hw_queue(hctx);
 971 }
 972 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 973
 974 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 975 {
 976         struct blk_mq_hw_ctx *hctx;
 977         int i;
 978
 979         queue_for_each_hw_ctx(q, hctx, i) {
 980                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 981                         continue;
 982
 983                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 984                 blk_mq_run_hw_queue(hctx, async);
 985         }
 986 }
 987 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 988
 989 static void blk_mq_run_work_fn(struct work_struct *work)
 990 {
 991         struct blk_mq_hw_ctx *hctx;
 992
 993         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 994
 995         __blk_mq_run_hw_queue(hctx);
 996 }
 997
 998 static void blk_mq_delay_work_fn(struct work_struct *work)
 999 {
1000         struct blk_mq_hw_ctx *hctx;
1001
1002         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1003
1004         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1005                 __blk_mq_run_hw_queue(hctx);
1006 }
1007
1008 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1009 {
1010         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1011                 return;
1012
1013         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1014                         &hctx->delay_work, msecs_to_jiffies(msecs));
1015 }
1016 EXPORT_SYMBOL(blk_mq_delay_queue);
1017
1018 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1019                                     struct request *rq, bool at_head)
1020 {
1021         struct blk_mq_ctx *ctx = rq->mq_ctx;
1022
1023         trace_block_rq_insert(hctx->queue, rq);
1024
1025         if (at_head)
1026                 list_add(&rq->queuelist, &ctx->rq_list);
1027         else
1028                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1029
1030         blk_mq_hctx_mark_pending(hctx, ctx);
1031 }
1032
1033 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1034                 bool async)
1035 {
1036         struct request_queue *q = rq->q;
1037         struct blk_mq_hw_ctx *hctx;
1038         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1039
1040         current_ctx = blk_mq_get_ctx(q);
1041         if (!cpu_online(ctx->cpu))
1042                 rq->mq_ctx = ctx = current_ctx;
1043
1044         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1045
1046         spin_lock(&ctx->lock);
1047         __blk_mq_insert_request(hctx, rq, at_head);
1048         spin_unlock(&ctx->lock);
1049
1050         if (run_queue)
1051                 blk_mq_run_hw_queue(hctx, async);
1052
1053         blk_mq_put_ctx(current_ctx);
1054 }
1055
1056 static void blk_mq_insert_requests(struct request_queue *q,
1057                                      struct blk_mq_ctx *ctx,
1058                                      struct list_head *list,
1059                                      int depth,
1060                                      bool from_schedule)
1061
1062 {
1063         struct blk_mq_hw_ctx *hctx;
1064         struct blk_mq_ctx *current_ctx;
1065
1066         trace_block_unplug(q, depth, !from_schedule);
1067
1068         current_ctx = blk_mq_get_ctx(q);
1069
1070         if (!cpu_online(ctx->cpu))
1071                 ctx = current_ctx;
1072         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1073
1074         /*
1075          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1076          * offline now
1077          */
1078         spin_lock(&ctx->lock);
1079         while (!list_empty(list)) {
1080                 struct request *rq;
1081
1082                 rq = list_first_entry(list, struct request, queuelist);
1083                 list_del_init(&rq->queuelist);
1084                 rq->mq_ctx = ctx;
1085                 __blk_mq_insert_request(hctx, rq, false);
1086         }
1087         spin_unlock(&ctx->lock);
1088
1089         blk_mq_run_hw_queue(hctx, from_schedule);
1090         blk_mq_put_ctx(current_ctx);
1091 }
1092
1093 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1094 {
1095         struct request *rqa = container_of(a, struct request, queuelist);
1096         struct request *rqb = container_of(b, struct request, queuelist);
1097
1098         return !(rqa->mq_ctx < rqb->mq_ctx ||
1099                  (rqa->mq_ctx == rqb->mq_ctx &&
1100                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1101 }
1102
1103 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1104 {
1105         struct blk_mq_ctx *this_ctx;
1106         struct request_queue *this_q;
1107         struct request *rq;
1108         LIST_HEAD(list);
1109         LIST_HEAD(ctx_list);
1110         unsigned int depth;
1111
1112         list_splice_init(&plug->mq_list, &list);
1113
1114         list_sort(NULL, &list, plug_ctx_cmp);
1115
1116         this_q = NULL;
1117         this_ctx = NULL;
1118         depth = 0;
1119
1120         while (!list_empty(&list)) {
1121                 rq = list_entry_rq(list.next);
1122                 list_del_init(&rq->queuelist);
1123                 BUG_ON(!rq->q);
1124                 if (rq->mq_ctx != this_ctx) {
1125                         if (this_ctx) {
1126                                 blk_mq_insert_requests(this_q, this_ctx,
1127                                                         &ctx_list, depth,
1128                                                         from_schedule);
1129                         }
1130
1131                         this_ctx = rq->mq_ctx;
1132                         this_q = rq->q;
1133                         depth = 0;
1134                 }
1135
1136                 depth++;
1137                 list_add_tail(&rq->queuelist, &ctx_list);
1138         }
1139
1140         /*
1141          * If 'this_ctx' is set, we know we have entries to complete
1142          * on 'ctx_list'. Do those.
1143          */
1144         if (this_ctx) {
1145                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1146                                        from_schedule);
1147         }
1148 }
1149
1150 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1151 {
1152         init_request_from_bio(rq, bio);
1153
1154         if (blk_do_io_stat(rq))
1155                 blk_account_io_start(rq, 1);
1156 }
1157
1158 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1159 {
1160         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1161                 !blk_queue_nomerges(hctx->queue);
1162 }
1163
1164 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1165                                          struct blk_mq_ctx *ctx,
1166                                          struct request *rq, struct bio *bio)
1167 {
1168         if (!hctx_allow_merges(hctx)) {
1169                 blk_mq_bio_to_request(rq, bio);
1170                 spin_lock(&ctx->lock);
1171 insert_rq:
1172                 __blk_mq_insert_request(hctx, rq, false);
1173                 spin_unlock(&ctx->lock);
1174                 return false;
1175         } else {
1176                 struct request_queue *q = hctx->queue;
1177
1178                 spin_lock(&ctx->lock);
1179                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1180                         blk_mq_bio_to_request(rq, bio);
1181                         goto insert_rq;
1182                 }
1183
1184                 spin_unlock(&ctx->lock);
1185                 __blk_mq_free_request(hctx, ctx, rq);
1186                 return true;
1187         }
1188 }
1189
1190 struct blk_map_ctx {
1191         struct blk_mq_hw_ctx *hctx;
1192         struct blk_mq_ctx *ctx;
1193 };
1194
1195 static struct request *blk_mq_map_request(struct request_queue *q,
1196                                           struct bio *bio,
1197                                           struct blk_map_ctx *data)
1198 {
1199         struct blk_mq_hw_ctx *hctx;
1200         struct blk_mq_ctx *ctx;
1201         struct request *rq;
1202         int rw = bio_data_dir(bio);
1203         struct blk_mq_alloc_data alloc_data;
1204
1205         if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1206                 bio_endio(bio, -EIO);
1207                 return NULL;
1208         }
1209
1210         ctx = blk_mq_get_ctx(q);
1211         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1212
1213         if (rw_is_sync(bio->bi_rw))
1214                 rw |= REQ_SYNC;
1215
1216         trace_block_getrq(q, bio, rw);
1217         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1218                         hctx);
1219         rq = __blk_mq_alloc_request(&alloc_data, rw);
1220         if (unlikely(!rq)) {
1221                 __blk_mq_run_hw_queue(hctx);
1222                 blk_mq_put_ctx(ctx);
1223                 trace_block_sleeprq(q, bio, rw);
1224
1225                 ctx = blk_mq_get_ctx(q);
1226                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1227                 blk_mq_set_alloc_data(&alloc_data, q,
1228                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1229                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1230                 ctx = alloc_data.ctx;
1231                 hctx = alloc_data.hctx;
1232         }
1233
1234         hctx->queued++;
1235         data->hctx = hctx;
1236         data->ctx = ctx;
1237         return rq;
1238 }
1239
1240 /*
1241  * Multiple hardware queue variant. This will not use per-process plugs,
1242  * but will attempt to bypass the hctx queueing if we can go straight to
1243  * hardware for SYNC IO.
1244  */
1245 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1246 {
1247         const int is_sync = rw_is_sync(bio->bi_rw);
1248         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1249         struct blk_map_ctx data;
1250         struct request *rq;
1251
1252         blk_queue_bounce(q, &bio);
1253
1254         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1255                 bio_endio(bio, -EIO);
1256                 return;
1257         }
1258
1259         rq = blk_mq_map_request(q, bio, &data);
1260         if (unlikely(!rq))
1261                 return;
1262
1263         if (unlikely(is_flush_fua)) {
1264                 blk_mq_bio_to_request(rq, bio);
1265                 blk_insert_flush(rq);
1266                 goto run_queue;
1267         }
1268
1269         /*
1270          * If the driver supports defer issued based on 'last', then
1271          * queue it up like normal since we can potentially save some
1272          * CPU this way.
1273          */
1274         if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1275                 struct blk_mq_queue_data bd = {
1276                         .rq = rq,
1277                         .list = NULL,
1278                         .last = 1
1279                 };
1280                 int ret;
1281
1282                 blk_mq_bio_to_request(rq, bio);
1283
1284                 /*
1285                  * For OK queue, we are done. For error, kill it. Any other
1286                  * error (busy), just add it to our list as we previously
1287                  * would have done
1288                  */
1289                 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1290                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1291                         goto done;
1292                 else {
1293                         __blk_mq_requeue_request(rq);
1294
1295                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1296                                 rq->errors = -EIO;
1297                                 blk_mq_end_request(rq, rq->errors);
1298                                 goto done;
1299                         }
1300                 }
1301         }
1302
1303         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1304                 /*
1305                  * For a SYNC request, send it to the hardware immediately. For
1306                  * an ASYNC request, just ensure that we run it later on. The
1307                  * latter allows for merging opportunities and more efficient
1308                  * dispatching.
1309                  */
1310 run_queue:
1311                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1312         }
1313 done:
1314         blk_mq_put_ctx(data.ctx);
1315 }
1316
1317 /*
1318  * Single hardware queue variant. This will attempt to use any per-process
1319  * plug for merging and IO deferral.
1320  */
1321 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1322 {
1323         const int is_sync = rw_is_sync(bio->bi_rw);
1324         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1325         unsigned int use_plug, request_count = 0;
1326         struct blk_map_ctx data;
1327         struct request *rq;
1328
1329         /*
1330          * If we have multiple hardware queues, just go directly to
1331          * one of those for sync IO.
1332          */
1333         use_plug = !is_flush_fua && !is_sync;
1334
1335         blk_queue_bounce(q, &bio);
1336
1337         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1338                 bio_endio(bio, -EIO);
1339                 return;
1340         }
1341
1342         if (use_plug && !blk_queue_nomerges(q) &&
1343             blk_attempt_plug_merge(q, bio, &request_count))
1344                 return;
1345
1346         rq = blk_mq_map_request(q, bio, &data);
1347         if (unlikely(!rq))
1348                 return;
1349
1350         if (unlikely(is_flush_fua)) {
1351                 blk_mq_bio_to_request(rq, bio);
1352                 blk_insert_flush(rq);
1353                 goto run_queue;
1354         }
1355
1356         /*
1357          * A task plug currently exists. Since this is completely lockless,
1358          * utilize that to temporarily store requests until the task is
1359          * either done or scheduled away.
1360          */
1361         if (use_plug) {
1362                 struct blk_plug *plug = current->plug;
1363
1364                 if (plug) {
1365                         blk_mq_bio_to_request(rq, bio);
1366                         if (list_empty(&plug->mq_list))
1367                                 trace_block_plug(q);
1368                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1369                                 blk_flush_plug_list(plug, false);
1370                                 trace_block_plug(q);
1371                         }
1372                         list_add_tail(&rq->queuelist, &plug->mq_list);
1373                         blk_mq_put_ctx(data.ctx);
1374                         return;
1375                 }
1376         }
1377
1378         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1379                 /*
1380                  * For a SYNC request, send it to the hardware immediately. For
1381                  * an ASYNC request, just ensure that we run it later on. The
1382                  * latter allows for merging opportunities and more efficient
1383                  * dispatching.
1384                  */
1385 run_queue:
1386                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1387         }
1388
1389         blk_mq_put_ctx(data.ctx);
1390 }
1391
1392 /*
1393  * Default mapping to a software queue, since we use one per CPU.
1394  */
1395 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1396 {
1397         return q->queue_hw_ctx[q->mq_map[cpu]];
1398 }
1399 EXPORT_SYMBOL(blk_mq_map_queue);
1400
1401 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1402                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1403 {
1404         struct page *page;
1405
1406         if (tags->rqs && set->ops->exit_request) {
1407                 int i;
1408
1409                 for (i = 0; i < tags->nr_tags; i++) {
1410                         if (!tags->rqs[i])
1411                                 continue;
1412                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1413                                                 hctx_idx, i);
1414                         tags->rqs[i] = NULL;
1415                 }
1416         }
1417
1418         while (!list_empty(&tags->page_list)) {
1419                 page = list_first_entry(&tags->page_list, struct page, lru);
1420                 list_del_init(&page->lru);
1421                 __free_pages(page, page->private);
1422         }
1423
1424         kfree(tags->rqs);
1425
1426         blk_mq_free_tags(tags);
1427 }
1428
1429 static size_t order_to_size(unsigned int order)
1430 {
1431         return (size_t)PAGE_SIZE << order;
1432 }
1433
1434 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1435                 unsigned int hctx_idx)
1436 {
1437         struct blk_mq_tags *tags;
1438         unsigned int i, j, entries_per_page, max_order = 4;
1439         size_t rq_size, left;
1440
1441         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1442                                 set->numa_node,
1443                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1444         if (!tags)
1445                 return NULL;
1446
1447         INIT_LIST_HEAD(&tags->page_list);
1448
1449         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1450                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1451                                  set->numa_node);
1452         if (!tags->rqs) {
1453                 blk_mq_free_tags(tags);
1454                 return NULL;
1455         }
1456
1457         /*
1458          * rq_size is the size of the request plus driver payload, rounded
1459          * to the cacheline size
1460          */
1461         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1462                                 cache_line_size());
1463         left = rq_size * set->queue_depth;
1464
1465         for (i = 0; i < set->queue_depth; ) {
1466                 int this_order = max_order;
1467                 struct page *page;
1468                 int to_do;
1469                 void *p;
1470
1471                 while (left < order_to_size(this_order - 1) && this_order)
1472                         this_order--;
1473
1474                 do {
1475                         page = alloc_pages_node(set->numa_node,
1476                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1477                                 this_order);
1478                         if (page)
1479                                 break;
1480                         if (!this_order--)
1481                                 break;
1482                         if (order_to_size(this_order) < rq_size)
1483                                 break;
1484                 } while (1);
1485
1486                 if (!page)
1487                         goto fail;
1488
1489                 page->private = this_order;
1490                 list_add_tail(&page->lru, &tags->page_list);
1491
1492                 p = page_address(page);
1493                 entries_per_page = order_to_size(this_order) / rq_size;
1494                 to_do = min(entries_per_page, set->queue_depth - i);
1495                 left -= to_do * rq_size;
1496                 for (j = 0; j < to_do; j++) {
1497                         tags->rqs[i] = p;
1498                         if (set->ops->init_request) {
1499                                 if (set->ops->init_request(set->driver_data,
1500                                                 tags->rqs[i], hctx_idx, i,
1501                                                 set->numa_node)) {
1502                                         tags->rqs[i] = NULL;
1503                                         goto fail;
1504                                 }
1505                         }
1506
1507                         p += rq_size;
1508                         i++;
1509                 }
1510         }
1511
1512         return tags;
1513
1514 fail:
1515         blk_mq_free_rq_map(set, tags, hctx_idx);
1516         return NULL;
1517 }
1518
1519 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1520 {
1521         kfree(bitmap->map);
1522 }
1523
1524 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1525 {
1526         unsigned int bpw = 8, total, num_maps, i;
1527
1528         bitmap->bits_per_word = bpw;
1529
1530         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1531         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1532                                         GFP_KERNEL, node);
1533         if (!bitmap->map)
1534                 return -ENOMEM;
1535
1536         total = nr_cpu_ids;
1537         for (i = 0; i < num_maps; i++) {
1538                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1539                 total -= bitmap->map[i].depth;
1540         }
1541
1542         return 0;
1543 }
1544
1545 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1546 {
1547         struct request_queue *q = hctx->queue;
1548         struct blk_mq_ctx *ctx;
1549         LIST_HEAD(tmp);
1550
1551         /*
1552          * Move ctx entries to new CPU, if this one is going away.
1553          */
1554         ctx = __blk_mq_get_ctx(q, cpu);
1555
1556         spin_lock(&ctx->lock);
1557         if (!list_empty(&ctx->rq_list)) {
1558                 list_splice_init(&ctx->rq_list, &tmp);
1559                 blk_mq_hctx_clear_pending(hctx, ctx);
1560         }
1561         spin_unlock(&ctx->lock);
1562
1563         if (list_empty(&tmp))
1564                 return NOTIFY_OK;
1565
1566         ctx = blk_mq_get_ctx(q);
1567         spin_lock(&ctx->lock);
1568
1569         while (!list_empty(&tmp)) {
1570                 struct request *rq;
1571
1572                 rq = list_first_entry(&tmp, struct request, queuelist);
1573                 rq->mq_ctx = ctx;
1574                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1575         }
1576
1577         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1578         blk_mq_hctx_mark_pending(hctx, ctx);
1579
1580         spin_unlock(&ctx->lock);
1581
1582         blk_mq_run_hw_queue(hctx, true);
1583         blk_mq_put_ctx(ctx);
1584         return NOTIFY_OK;
1585 }
1586
1587 static int blk_mq_hctx_notify(void *data, unsigned long action,
1588                               unsigned int cpu)
1589 {
1590         struct blk_mq_hw_ctx *hctx = data;
1591
1592         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1593                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1594
1595         /*
1596          * In case of CPU online, tags may be reallocated
1597          * in blk_mq_map_swqueue() after mapping is updated.
1598          */
1599
1600         return NOTIFY_OK;
1601 }
1602
1603 /* hctx->ctxs will be freed in queue's release handler */
1604 static void blk_mq_exit_hctx(struct request_queue *q,
1605                 struct blk_mq_tag_set *set,
1606                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1607 {
1608         unsigned flush_start_tag = set->queue_depth;
1609
1610         blk_mq_tag_idle(hctx);
1611
1612         if (set->ops->exit_request)
1613                 set->ops->exit_request(set->driver_data,
1614                                        hctx->fq->flush_rq, hctx_idx,
1615                                        flush_start_tag + hctx_idx);
1616
1617         if (set->ops->exit_hctx)
1618                 set->ops->exit_hctx(hctx, hctx_idx);
1619
1620         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1621         blk_free_flush_queue(hctx->fq);
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
1764                 /*
1765                  * Set local node, IFF we have more than one hw queue. If
1766                  * not, we remain on the home node of the device
1767                  */
1768                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1769                         hctx->numa_node = cpu_to_node(i);
1770         }
1771 }
1772
1773 static void blk_mq_map_swqueue(struct request_queue *q)
1774 {
1775         unsigned int i;
1776         struct blk_mq_hw_ctx *hctx;
1777         struct blk_mq_ctx *ctx;
1778         struct blk_mq_tag_set *set = q->tag_set;
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                 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1801
1802                 /*
1803                  * If no software queues are mapped to this hardware queue,
1804                  * disable it and free the request entries.
1805                  */
1806                 if (!hctx->nr_ctx) {
1807                         if (set->tags[i]) {
1808                                 blk_mq_free_rq_map(set, set->tags[i], i);
1809                                 set->tags[i] = NULL;
1810                         }
1811                         hctx->tags = NULL;
1812                         continue;
1813                 }
1814
1815                 /* unmapped hw queue can be remapped after CPU topo changed */
1816                 if (!set->tags[i])
1817                         set->tags[i] = blk_mq_init_rq_map(set, i);
1818                 hctx->tags = set->tags[i];
1819                 WARN_ON(!hctx->tags);
1820
1821                 /*
1822                  * Set the map size to the number of mapped software queues.
1823                  * This is more accurate and more efficient than looping
1824                  * over all possibly mapped software queues.
1825                  */
1826                 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1827
1828                 /*
1829                  * Initialize batch roundrobin counts
1830                  */
1831                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1832                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1833         }
1834 }
1835
1836 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1837 {
1838         struct blk_mq_hw_ctx *hctx;
1839         struct request_queue *q;
1840         bool shared;
1841         int i;
1842
1843         if (set->tag_list.next == set->tag_list.prev)
1844                 shared = false;
1845         else
1846                 shared = true;
1847
1848         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1849                 blk_mq_freeze_queue(q);
1850
1851                 queue_for_each_hw_ctx(q, hctx, i) {
1852                         if (shared)
1853                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1854                         else
1855                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1856                 }
1857                 blk_mq_unfreeze_queue(q);
1858         }
1859 }
1860
1861 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1862 {
1863         struct blk_mq_tag_set *set = q->tag_set;
1864
1865         mutex_lock(&set->tag_list_lock);
1866         list_del_init(&q->tag_set_list);
1867         blk_mq_update_tag_set_depth(set);
1868         mutex_unlock(&set->tag_list_lock);
1869 }
1870
1871 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1872                                      struct request_queue *q)
1873 {
1874         q->tag_set = set;
1875
1876         mutex_lock(&set->tag_list_lock);
1877         list_add_tail(&q->tag_set_list, &set->tag_list);
1878         blk_mq_update_tag_set_depth(set);
1879         mutex_unlock(&set->tag_list_lock);
1880 }
1881
1882 /*
1883  * It is the actual release handler for mq, but we do it from
1884  * request queue's release handler for avoiding use-after-free
1885  * and headache because q->mq_kobj shouldn't have been introduced,
1886  * but we can't group ctx/kctx kobj without it.
1887  */
1888 void blk_mq_release(struct request_queue *q)
1889 {
1890         struct blk_mq_hw_ctx *hctx;
1891         unsigned int i;
1892
1893         /* hctx kobj stays in hctx */
1894         queue_for_each_hw_ctx(q, hctx, i) {
1895                 if (!hctx)
1896                         continue;
1897                 kfree(hctx->ctxs);
1898                 kfree(hctx);
1899         }
1900
1901         kfree(q->queue_hw_ctx);
1902
1903         /* ctx kobj stays in queue_ctx */
1904         free_percpu(q->queue_ctx);
1905 }
1906
1907 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1908 {
1909         struct request_queue *uninit_q, *q;
1910
1911         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1912         if (!uninit_q)
1913                 return ERR_PTR(-ENOMEM);
1914
1915         q = blk_mq_init_allocated_queue(set, uninit_q);
1916         if (IS_ERR(q))
1917                 blk_cleanup_queue(uninit_q);
1918
1919         return q;
1920 }
1921 EXPORT_SYMBOL(blk_mq_init_queue);
1922
1923 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1924                                                   struct request_queue *q)
1925 {
1926         struct blk_mq_hw_ctx **hctxs;
1927         struct blk_mq_ctx __percpu *ctx;
1928         unsigned int *map;
1929         int i;
1930
1931         ctx = alloc_percpu(struct blk_mq_ctx);
1932         if (!ctx)
1933                 return ERR_PTR(-ENOMEM);
1934
1935         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1936                         set->numa_node);
1937
1938         if (!hctxs)
1939                 goto err_percpu;
1940
1941         map = blk_mq_make_queue_map(set);
1942         if (!map)
1943                 goto err_map;
1944
1945         for (i = 0; i < set->nr_hw_queues; i++) {
1946                 int node = blk_mq_hw_queue_to_node(map, i);
1947
1948                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1949                                         GFP_KERNEL, node);
1950                 if (!hctxs[i])
1951                         goto err_hctxs;
1952
1953                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1954                                                 node))
1955                         goto err_hctxs;
1956
1957                 atomic_set(&hctxs[i]->nr_active, 0);
1958                 hctxs[i]->numa_node = node;
1959                 hctxs[i]->queue_num = i;
1960         }
1961
1962         /*
1963          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1964          * See blk_register_queue() for details.
1965          */
1966         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1967                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1968                 goto err_hctxs;
1969
1970         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1971         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
1972
1973         q->nr_queues = nr_cpu_ids;
1974         q->nr_hw_queues = set->nr_hw_queues;
1975         q->mq_map = map;
1976
1977         q->queue_ctx = ctx;
1978         q->queue_hw_ctx = hctxs;
1979
1980         q->mq_ops = set->ops;
1981         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1982
1983         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1984                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1985
1986         q->sg_reserved_size = INT_MAX;
1987
1988         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1989         INIT_LIST_HEAD(&q->requeue_list);
1990         spin_lock_init(&q->requeue_lock);
1991
1992         if (q->nr_hw_queues > 1)
1993                 blk_queue_make_request(q, blk_mq_make_request);
1994         else
1995                 blk_queue_make_request(q, blk_sq_make_request);
1996
1997         /*
1998          * Do this after blk_queue_make_request() overrides it...
1999          */
2000         q->nr_requests = set->queue_depth;
2001
2002         if (set->ops->complete)
2003                 blk_queue_softirq_done(q, set->ops->complete);
2004
2005         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2006
2007         if (blk_mq_init_hw_queues(q, set))
2008                 goto err_hctxs;
2009
2010         mutex_lock(&all_q_mutex);
2011         list_add_tail(&q->all_q_node, &all_q_list);
2012         mutex_unlock(&all_q_mutex);
2013
2014         blk_mq_add_queue_tag_set(set, q);
2015
2016         blk_mq_map_swqueue(q);
2017
2018         return q;
2019
2020 err_hctxs:
2021         kfree(map);
2022         for (i = 0; i < set->nr_hw_queues; i++) {
2023                 if (!hctxs[i])
2024                         break;
2025                 free_cpumask_var(hctxs[i]->cpumask);
2026                 kfree(hctxs[i]);
2027         }
2028 err_map:
2029         kfree(hctxs);
2030 err_percpu:
2031         free_percpu(ctx);
2032         return ERR_PTR(-ENOMEM);
2033 }
2034 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2035
2036 void blk_mq_free_queue(struct request_queue *q)
2037 {
2038         struct blk_mq_tag_set   *set = q->tag_set;
2039
2040         blk_mq_del_queue_tag_set(q);
2041
2042         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2043         blk_mq_free_hw_queues(q, set);
2044
2045         percpu_ref_exit(&q->mq_usage_counter);
2046
2047         kfree(q->mq_map);
2048
2049         q->mq_map = NULL;
2050
2051         mutex_lock(&all_q_mutex);
2052         list_del_init(&q->all_q_node);
2053         mutex_unlock(&all_q_mutex);
2054 }
2055
2056 /* Basically redo blk_mq_init_queue with queue frozen */
2057 static void blk_mq_queue_reinit(struct request_queue *q)
2058 {
2059         WARN_ON_ONCE(!q->mq_freeze_depth);
2060
2061         blk_mq_sysfs_unregister(q);
2062
2063         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2064
2065         /*
2066          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2067          * we should change hctx numa_node according to new topology (this
2068          * involves free and re-allocate memory, worthy doing?)
2069          */
2070
2071         blk_mq_map_swqueue(q);
2072
2073         blk_mq_sysfs_register(q);
2074 }
2075
2076 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2077                                       unsigned long action, void *hcpu)
2078 {
2079         struct request_queue *q;
2080
2081         /*
2082          * Before new mappings are established, hotadded cpu might already
2083          * start handling requests. This doesn't break anything as we map
2084          * offline CPUs to first hardware queue. We will re-init the queue
2085          * below to get optimal settings.
2086          */
2087         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2088             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2089                 return NOTIFY_OK;
2090
2091         mutex_lock(&all_q_mutex);
2092
2093         /*
2094          * We need to freeze and reinit all existing queues.  Freezing
2095          * involves synchronous wait for an RCU grace period and doing it
2096          * one by one may take a long time.  Start freezing all queues in
2097          * one swoop and then wait for the completions so that freezing can
2098          * take place in parallel.
2099          */
2100         list_for_each_entry(q, &all_q_list, all_q_node)
2101                 blk_mq_freeze_queue_start(q);
2102         list_for_each_entry(q, &all_q_list, all_q_node) {
2103                 blk_mq_freeze_queue_wait(q);
2104
2105                 /*
2106                  * timeout handler can't touch hw queue during the
2107                  * reinitialization
2108                  */
2109                 del_timer_sync(&q->timeout);
2110         }
2111
2112         list_for_each_entry(q, &all_q_list, all_q_node)
2113                 blk_mq_queue_reinit(q);
2114
2115         list_for_each_entry(q, &all_q_list, all_q_node)
2116                 blk_mq_unfreeze_queue(q);
2117
2118         mutex_unlock(&all_q_mutex);
2119         return NOTIFY_OK;
2120 }
2121
2122 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2123 {
2124         int i;
2125
2126         for (i = 0; i < set->nr_hw_queues; i++) {
2127                 set->tags[i] = blk_mq_init_rq_map(set, i);
2128                 if (!set->tags[i])
2129                         goto out_unwind;
2130         }
2131
2132         return 0;
2133
2134 out_unwind:
2135         while (--i >= 0)
2136                 blk_mq_free_rq_map(set, set->tags[i], i);
2137
2138         return -ENOMEM;
2139 }
2140
2141 /*
2142  * Allocate the request maps associated with this tag_set. Note that this
2143  * may reduce the depth asked for, if memory is tight. set->queue_depth
2144  * will be updated to reflect the allocated depth.
2145  */
2146 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2147 {
2148         unsigned int depth;
2149         int err;
2150
2151         depth = set->queue_depth;
2152         do {
2153                 err = __blk_mq_alloc_rq_maps(set);
2154                 if (!err)
2155                         break;
2156
2157                 set->queue_depth >>= 1;
2158                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2159                         err = -ENOMEM;
2160                         break;
2161                 }
2162         } while (set->queue_depth);
2163
2164         if (!set->queue_depth || err) {
2165                 pr_err("blk-mq: failed to allocate request map\n");
2166                 return -ENOMEM;
2167         }
2168
2169         if (depth != set->queue_depth)
2170                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2171                                                 depth, set->queue_depth);
2172
2173         return 0;
2174 }
2175
2176 /*
2177  * Alloc a tag set to be associated with one or more request queues.
2178  * May fail with EINVAL for various error conditions. May adjust the
2179  * requested depth down, if if it too large. In that case, the set
2180  * value will be stored in set->queue_depth.
2181  */
2182 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2183 {
2184         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2185
2186         if (!set->nr_hw_queues)
2187                 return -EINVAL;
2188         if (!set->queue_depth)
2189                 return -EINVAL;
2190         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2191                 return -EINVAL;
2192
2193         if (!set->ops->queue_rq || !set->ops->map_queue)
2194                 return -EINVAL;
2195
2196         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2197                 pr_info("blk-mq: reduced tag depth to %u\n",
2198                         BLK_MQ_MAX_DEPTH);
2199                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2200         }
2201
2202         /*
2203          * If a crashdump is active, then we are potentially in a very
2204          * memory constrained environment. Limit us to 1 queue and
2205          * 64 tags to prevent using too much memory.
2206          */
2207         if (is_kdump_kernel()) {
2208                 set->nr_hw_queues = 1;
2209                 set->queue_depth = min(64U, set->queue_depth);
2210         }
2211
2212         set->tags = kmalloc_node(set->nr_hw_queues *
2213                                  sizeof(struct blk_mq_tags *),
2214                                  GFP_KERNEL, set->numa_node);
2215         if (!set->tags)
2216                 return -ENOMEM;
2217
2218         if (blk_mq_alloc_rq_maps(set))
2219                 goto enomem;
2220
2221         mutex_init(&set->tag_list_lock);
2222         INIT_LIST_HEAD(&set->tag_list);
2223
2224         return 0;
2225 enomem:
2226         kfree(set->tags);
2227         set->tags = NULL;
2228         return -ENOMEM;
2229 }
2230 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2231
2232 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2233 {
2234         int i;
2235
2236         for (i = 0; i < set->nr_hw_queues; i++) {
2237                 if (set->tags[i])
2238                         blk_mq_free_rq_map(set, set->tags[i], i);
2239         }
2240
2241         kfree(set->tags);
2242         set->tags = NULL;
2243 }
2244 EXPORT_SYMBOL(blk_mq_free_tag_set);
2245
2246 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2247 {
2248         struct blk_mq_tag_set *set = q->tag_set;
2249         struct blk_mq_hw_ctx *hctx;
2250         int i, ret;
2251
2252         if (!set || nr > set->queue_depth)
2253                 return -EINVAL;
2254
2255         ret = 0;
2256         queue_for_each_hw_ctx(q, hctx, i) {
2257                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2258                 if (ret)
2259                         break;
2260         }
2261
2262         if (!ret)
2263                 q->nr_requests = nr;
2264
2265         return ret;
2266 }
2267
2268 void blk_mq_disable_hotplug(void)
2269 {
2270         mutex_lock(&all_q_mutex);
2271 }
2272
2273 void blk_mq_enable_hotplug(void)
2274 {
2275         mutex_unlock(&all_q_mutex);
2276 }
2277
2278 static int __init blk_mq_init(void)
2279 {
2280         blk_mq_cpu_init();
2281
2282         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2283
2284         return 0;
2285 }
2286 subsys_initcall(blk_mq_init);