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
  24 #include <trace/events/block.h>
  25
  26 #include <linux/blk-mq.h>
  27 #include "blk.h"
  28 #include "blk-mq.h"
  29 #include "blk-mq-tag.h"
  30
  31 static DEFINE_MUTEX(all_q_mutex);
  32 static LIST_HEAD(all_q_list);
  33
  34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  35
  36 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
  37                                            unsigned int cpu)
  38 {
  39         return per_cpu_ptr(q->queue_ctx, cpu);
  40 }
  41
  42 /*
  43  * This assumes per-cpu software queueing queues. They could be per-node
  44  * as well, for instance. For now this is hardcoded as-is. Note that we don't
  45  * care about preemption, since we know the ctx's are persistent. This does
  46  * mean that we can't rely on ctx always matching the currently running CPU.
  47  */
  48 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
  49 {
  50         return __blk_mq_get_ctx(q, get_cpu());
  51 }
  52
  53 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
  54 {
  55         put_cpu();
  56 }
  57
  58 /*
  59  * Check if any of the ctx's have pending work in this hardware queue
  60  */
  61 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  62 {
  63         unsigned int i;
  64
  65         for (i = 0; i < hctx->ctx_map.map_size; i++)
  66                 if (hctx->ctx_map.map[i].word)
  67                         return true;
  68
  69         return false;
  70 }
  71
  72 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  73                                               struct blk_mq_ctx *ctx)
  74 {
  75         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  76 }
  77
  78 #define CTX_TO_BIT(hctx, ctx)   \
  79         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  80
  81 /*
  82  * Mark this ctx as having pending work in this hardware queue
  83  */
  84 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  85                                      struct blk_mq_ctx *ctx)
  86 {
  87         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  88
  89         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  90                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  91 }
  92
  93 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  94                                       struct blk_mq_ctx *ctx)
  95 {
  96         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  97
  98         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  99 }
 100
 101 static int blk_mq_queue_enter(struct request_queue *q)
 102 {
 103         int ret;
 104
 105         __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
 106         smp_wmb();
 107         /* we have problems to freeze the queue if it's initializing */
 108         if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
 109                 return 0;
 110
 111         __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
 112
 113         spin_lock_irq(q->queue_lock);
 114         ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
 115                 !blk_queue_bypass(q) || blk_queue_dying(q),
 116                 *q->queue_lock);
 117         /* inc usage with lock hold to avoid freeze_queue runs here */
 118         if (!ret && !blk_queue_dying(q))
 119                 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
 120         else if (blk_queue_dying(q))
 121                 ret = -ENODEV;
 122         spin_unlock_irq(q->queue_lock);
 123
 124         return ret;
 125 }
 126
 127 static void blk_mq_queue_exit(struct request_queue *q)
 128 {
 129         __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
 130 }
 131
 132 static void __blk_mq_drain_queue(struct request_queue *q)
 133 {
 134         while (true) {
 135                 s64 count;
 136
 137                 spin_lock_irq(q->queue_lock);
 138                 count = percpu_counter_sum(&q->mq_usage_counter);
 139                 spin_unlock_irq(q->queue_lock);
 140
 141                 if (count == 0)
 142                         break;
 143                 blk_mq_run_queues(q, false);
 144                 msleep(10);
 145         }
 146 }
 147
 148 /*
 149  * Guarantee no request is in use, so we can change any data structure of
 150  * the queue afterward.
 151  */
 152 static void blk_mq_freeze_queue(struct request_queue *q)
 153 {
 154         bool drain;
 155
 156         spin_lock_irq(q->queue_lock);
 157         drain = !q->bypass_depth++;
 158         queue_flag_set(QUEUE_FLAG_BYPASS, q);
 159         spin_unlock_irq(q->queue_lock);
 160
 161         if (drain)
 162                 __blk_mq_drain_queue(q);
 163 }
 164
 165 void blk_mq_drain_queue(struct request_queue *q)
 166 {
 167         __blk_mq_drain_queue(q);
 168 }
 169
 170 static void blk_mq_unfreeze_queue(struct request_queue *q)
 171 {
 172         bool wake = false;
 173
 174         spin_lock_irq(q->queue_lock);
 175         if (!--q->bypass_depth) {
 176                 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
 177                 wake = true;
 178         }
 179         WARN_ON_ONCE(q->bypass_depth < 0);
 180         spin_unlock_irq(q->queue_lock);
 181         if (wake)
 182                 wake_up_all(&q->mq_freeze_wq);
 183 }
 184
 185 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 186 {
 187         return blk_mq_has_free_tags(hctx->tags);
 188 }
 189 EXPORT_SYMBOL(blk_mq_can_queue);
 190
 191 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 192                                struct request *rq, unsigned int rw_flags)
 193 {
 194         if (blk_queue_io_stat(q))
 195                 rw_flags |= REQ_IO_STAT;
 196
 197         INIT_LIST_HEAD(&rq->queuelist);
 198         /* csd/requeue_work/fifo_time is initialized before use */
 199         rq->q = q;
 200         rq->mq_ctx = ctx;
 201         rq->cmd_flags |= rw_flags;
 202         /* do not touch atomic flags, it needs atomic ops against the timer */
 203         rq->cpu = -1;
 204         INIT_HLIST_NODE(&rq->hash);
 205         RB_CLEAR_NODE(&rq->rb_node);
 206         rq->rq_disk = NULL;
 207         rq->part = NULL;
 208 #ifdef CONFIG_BLK_CGROUP
 209         rq->rl = NULL;
 210         set_start_time_ns(rq);
 211         rq->io_start_time_ns = 0;
 212 #endif
 213         rq->nr_phys_segments = 0;
 214 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 215         rq->nr_integrity_segments = 0;
 216 #endif
 217         rq->special = NULL;
 218         /* tag was already set */
 219         rq->errors = 0;
 220
 221         rq->extra_len = 0;
 222         rq->sense_len = 0;
 223         rq->resid_len = 0;
 224         rq->sense = NULL;
 225
 226         INIT_LIST_HEAD(&rq->timeout_list);
 227         rq->end_io = NULL;
 228         rq->end_io_data = NULL;
 229         rq->next_rq = NULL;
 230
 231         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 232 }
 233
 234 static struct request *
 235 __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 236                 struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved)
 237 {
 238         struct request *rq;
 239         unsigned int tag;
 240
 241         tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
 242         if (tag != BLK_MQ_TAG_FAIL) {
 243                 rq = hctx->tags->rqs[tag];
 244
 245                 rq->cmd_flags = 0;
 246                 if (blk_mq_tag_busy(hctx)) {
 247                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 248                         atomic_inc(&hctx->nr_active);
 249                 }
 250
 251                 rq->tag = tag;
 252                 blk_mq_rq_ctx_init(q, ctx, rq, rw);
 253                 return rq;
 254         }
 255
 256         return NULL;
 257 }
 258
 259 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 260                 bool reserved)
 261 {
 262         struct blk_mq_ctx *ctx;
 263         struct blk_mq_hw_ctx *hctx;
 264         struct request *rq;
 265
 266         if (blk_mq_queue_enter(q))
 267                 return NULL;
 268
 269         ctx = blk_mq_get_ctx(q);
 270         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 271
 272         rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp & ~__GFP_WAIT,
 273                                     reserved);
 274         if (!rq && (gfp & __GFP_WAIT)) {
 275                 __blk_mq_run_hw_queue(hctx);
 276                 blk_mq_put_ctx(ctx);
 277
 278                 ctx = blk_mq_get_ctx(q);
 279                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 280                 rq =  __blk_mq_alloc_request(q, hctx, ctx, rw, gfp, reserved);
 281         }
 282         blk_mq_put_ctx(ctx);
 283         return rq;
 284 }
 285 EXPORT_SYMBOL(blk_mq_alloc_request);
 286
 287 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 288                                   struct blk_mq_ctx *ctx, struct request *rq)
 289 {
 290         const int tag = rq->tag;
 291         struct request_queue *q = rq->q;
 292
 293         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 294                 atomic_dec(&hctx->nr_active);
 295
 296         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 297         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 298         blk_mq_queue_exit(q);
 299 }
 300
 301 void blk_mq_free_request(struct request *rq)
 302 {
 303         struct blk_mq_ctx *ctx = rq->mq_ctx;
 304         struct blk_mq_hw_ctx *hctx;
 305         struct request_queue *q = rq->q;
 306
 307         ctx->rq_completed[rq_is_sync(rq)]++;
 308
 309         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 310         __blk_mq_free_request(hctx, ctx, rq);
 311 }
 312
 313 /*
 314  * Clone all relevant state from a request that has been put on hold in
 315  * the flush state machine into the preallocated flush request that hangs
 316  * off the request queue.
 317  *
 318  * For a driver the flush request should be invisible, that's why we are
 319  * impersonating the original request here.
 320  */
 321 void blk_mq_clone_flush_request(struct request *flush_rq,
 322                 struct request *orig_rq)
 323 {
 324         struct blk_mq_hw_ctx *hctx =
 325                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 326
 327         flush_rq->mq_ctx = orig_rq->mq_ctx;
 328         flush_rq->tag = orig_rq->tag;
 329         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 330                 hctx->cmd_size);
 331 }
 332
 333 inline void __blk_mq_end_io(struct request *rq, int error)
 334 {
 335         blk_account_io_done(rq);
 336
 337         if (rq->end_io) {
 338                 rq->end_io(rq, error);
 339         } else {
 340                 if (unlikely(blk_bidi_rq(rq)))
 341                         blk_mq_free_request(rq->next_rq);
 342                 blk_mq_free_request(rq);
 343         }
 344 }
 345 EXPORT_SYMBOL(__blk_mq_end_io);
 346
 347 void blk_mq_end_io(struct request *rq, int error)
 348 {
 349         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 350                 BUG();
 351         __blk_mq_end_io(rq, error);
 352 }
 353 EXPORT_SYMBOL(blk_mq_end_io);
 354
 355 static void __blk_mq_complete_request_remote(void *data)
 356 {
 357         struct request *rq = data;
 358
 359         rq->q->softirq_done_fn(rq);
 360 }
 361
 362 void __blk_mq_complete_request(struct request *rq)
 363 {
 364         struct blk_mq_ctx *ctx = rq->mq_ctx;
 365         bool shared = false;
 366         int cpu;
 367
 368         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 369                 rq->q->softirq_done_fn(rq);
 370                 return;
 371         }
 372
 373         cpu = get_cpu();
 374         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 375                 shared = cpus_share_cache(cpu, ctx->cpu);
 376
 377         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 378                 rq->csd.func = __blk_mq_complete_request_remote;
 379                 rq->csd.info = rq;
 380                 rq->csd.flags = 0;
 381                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 382         } else {
 383                 rq->q->softirq_done_fn(rq);
 384         }
 385         put_cpu();
 386 }
 387
 388 /**
 389  * blk_mq_complete_request - end I/O on a request
 390  * @rq:         the request being processed
 391  *
 392  * Description:
 393  *      Ends all I/O on a request. It does not handle partial completions.
 394  *      The actual completion happens out-of-order, through a IPI handler.
 395  **/
 396 void blk_mq_complete_request(struct request *rq)
 397 {
 398         struct request_queue *q = rq->q;
 399
 400         if (unlikely(blk_should_fake_timeout(q)))
 401                 return;
 402         if (!blk_mark_rq_complete(rq)) {
 403                 if (q->softirq_done_fn)
 404                         __blk_mq_complete_request(rq);
 405                 else
 406                         blk_mq_end_io(rq, rq->errors);
 407         }
 408 }
 409 EXPORT_SYMBOL(blk_mq_complete_request);
 410
 411 static void blk_mq_start_request(struct request *rq, bool last)
 412 {
 413         struct request_queue *q = rq->q;
 414
 415         trace_block_rq_issue(q, rq);
 416
 417         rq->resid_len = blk_rq_bytes(rq);
 418         if (unlikely(blk_bidi_rq(rq)))
 419                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 420
 421         /*
 422          * Just mark start time and set the started bit. Due to memory
 423          * ordering, we know we'll see the correct deadline as long as
 424          * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
 425          * unless one has been set in the request.
 426          */
 427         if (!rq->timeout)
 428                 rq->deadline = jiffies + q->rq_timeout;
 429         else
 430                 rq->deadline = jiffies + rq->timeout;
 431
 432         /*
 433          * Mark us as started and clear complete. Complete might have been
 434          * set if requeue raced with timeout, which then marked it as
 435          * complete. So be sure to clear complete again when we start
 436          * the request, otherwise we'll ignore the completion event.
 437          */
 438         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 439                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 440         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 441                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 442
 443         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 444                 /*
 445                  * Make sure space for the drain appears.  We know we can do
 446                  * this because max_hw_segments has been adjusted to be one
 447                  * fewer than the device can handle.
 448                  */
 449                 rq->nr_phys_segments++;
 450         }
 451
 452         /*
 453          * Flag the last request in the series so that drivers know when IO
 454          * should be kicked off, if they don't do it on a per-request basis.
 455          *
 456          * Note: the flag isn't the only condition drivers should do kick off.
 457          * If drive is busy, the last request might not have the bit set.
 458          */
 459         if (last)
 460                 rq->cmd_flags |= REQ_END;
 461 }
 462
 463 static void __blk_mq_requeue_request(struct request *rq)
 464 {
 465         struct request_queue *q = rq->q;
 466
 467         trace_block_rq_requeue(q, rq);
 468         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 469
 470         rq->cmd_flags &= ~REQ_END;
 471
 472         if (q->dma_drain_size && blk_rq_bytes(rq))
 473                 rq->nr_phys_segments--;
 474 }
 475
 476 void blk_mq_requeue_request(struct request *rq)
 477 {
 478         __blk_mq_requeue_request(rq);
 479         blk_clear_rq_complete(rq);
 480
 481         BUG_ON(blk_queued_rq(rq));
 482         blk_mq_add_to_requeue_list(rq, true);
 483 }
 484 EXPORT_SYMBOL(blk_mq_requeue_request);
 485
 486 static void blk_mq_requeue_work(struct work_struct *work)
 487 {
 488         struct request_queue *q =
 489                 container_of(work, struct request_queue, requeue_work);
 490         LIST_HEAD(rq_list);
 491         struct request *rq, *next;
 492         unsigned long flags;
 493
 494         spin_lock_irqsave(&q->requeue_lock, flags);
 495         list_splice_init(&q->requeue_list, &rq_list);
 496         spin_unlock_irqrestore(&q->requeue_lock, flags);
 497
 498         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 499                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 500                         continue;
 501
 502                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 503                 list_del_init(&rq->queuelist);
 504                 blk_mq_insert_request(rq, true, false, false);
 505         }
 506
 507         while (!list_empty(&rq_list)) {
 508                 rq = list_entry(rq_list.next, struct request, queuelist);
 509                 list_del_init(&rq->queuelist);
 510                 blk_mq_insert_request(rq, false, false, false);
 511         }
 512
 513         blk_mq_run_queues(q, false);
 514 }
 515
 516 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 517 {
 518         struct request_queue *q = rq->q;
 519         unsigned long flags;
 520
 521         /*
 522          * We abuse this flag that is otherwise used by the I/O scheduler to
 523          * request head insertation from the workqueue.
 524          */
 525         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 526
 527         spin_lock_irqsave(&q->requeue_lock, flags);
 528         if (at_head) {
 529                 rq->cmd_flags |= REQ_SOFTBARRIER;
 530                 list_add(&rq->queuelist, &q->requeue_list);
 531         } else {
 532                 list_add_tail(&rq->queuelist, &q->requeue_list);
 533         }
 534         spin_unlock_irqrestore(&q->requeue_lock, flags);
 535 }
 536 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 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 struct request *blk_mq_tag_to_rq(struct blk_mq_hw_ctx *hctx, unsigned int tag)
 545 {
 546         struct request_queue *q = hctx->queue;
 547
 548         if ((q->flush_rq->cmd_flags & REQ_FLUSH_SEQ) &&
 549             q->flush_rq->tag == tag)
 550                 return q->flush_rq;
 551
 552         return hctx->tags->rqs[tag];
 553 }
 554 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 555
 556 struct blk_mq_timeout_data {
 557         struct blk_mq_hw_ctx *hctx;
 558         unsigned long *next;
 559         unsigned int *next_set;
 560 };
 561
 562 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
 563 {
 564         struct blk_mq_timeout_data *data = __data;
 565         struct blk_mq_hw_ctx *hctx = data->hctx;
 566         unsigned int tag;
 567
 568          /* It may not be in flight yet (this is where
 569          * the REQ_ATOMIC_STARTED flag comes in). The requests are
 570          * statically allocated, so we know it's always safe to access the
 571          * memory associated with a bit offset into ->rqs[].
 572          */
 573         tag = 0;
 574         do {
 575                 struct request *rq;
 576
 577                 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
 578                 if (tag >= hctx->tags->nr_tags)
 579                         break;
 580
 581                 rq = blk_mq_tag_to_rq(hctx, tag++);
 582                 if (rq->q != hctx->queue)
 583                         continue;
 584                 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 585                         continue;
 586
 587                 blk_rq_check_expired(rq, data->next, data->next_set);
 588         } while (1);
 589 }
 590
 591 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
 592                                         unsigned long *next,
 593                                         unsigned int *next_set)
 594 {
 595         struct blk_mq_timeout_data data = {
 596                 .hctx           = hctx,
 597                 .next           = next,
 598                 .next_set       = next_set,
 599         };
 600
 601         /*
 602          * Ask the tagging code to iterate busy requests, so we can
 603          * check them for timeout.
 604          */
 605         blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
 606 }
 607
 608 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
 609 {
 610         struct request_queue *q = rq->q;
 611
 612         /*
 613          * We know that complete is set at this point. If STARTED isn't set
 614          * anymore, then the request isn't active and the "timeout" should
 615          * just be ignored. This can happen due to the bitflag ordering.
 616          * Timeout first checks if STARTED is set, and if it is, assumes
 617          * the request is active. But if we race with completion, then
 618          * we both flags will get cleared. So check here again, and ignore
 619          * a timeout event with a request that isn't active.
 620          */
 621         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 622                 return BLK_EH_NOT_HANDLED;
 623
 624         if (!q->mq_ops->timeout)
 625                 return BLK_EH_RESET_TIMER;
 626
 627         return q->mq_ops->timeout(rq);
 628 }
 629
 630 static void blk_mq_rq_timer(unsigned long data)
 631 {
 632         struct request_queue *q = (struct request_queue *) data;
 633         struct blk_mq_hw_ctx *hctx;
 634         unsigned long next = 0;
 635         int i, next_set = 0;
 636
 637         queue_for_each_hw_ctx(q, hctx, i) {
 638                 /*
 639                  * If not software queues are currently mapped to this
 640                  * hardware queue, there's nothing to check
 641                  */
 642                 if (!hctx->nr_ctx || !hctx->tags)
 643                         continue;
 644
 645                 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
 646         }
 647
 648         if (next_set) {
 649                 next = blk_rq_timeout(round_jiffies_up(next));
 650                 mod_timer(&q->timeout, next);
 651         } else {
 652                 queue_for_each_hw_ctx(q, hctx, i)
 653                         blk_mq_tag_idle(hctx);
 654         }
 655 }
 656
 657 /*
 658  * Reverse check our software queue for entries that we could potentially
 659  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 660  * too much time checking for merges.
 661  */
 662 static bool blk_mq_attempt_merge(struct request_queue *q,
 663                                  struct blk_mq_ctx *ctx, struct bio *bio)
 664 {
 665         struct request *rq;
 666         int checked = 8;
 667
 668         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 669                 int el_ret;
 670
 671                 if (!checked--)
 672                         break;
 673
 674                 if (!blk_rq_merge_ok(rq, bio))
 675                         continue;
 676
 677                 el_ret = blk_try_merge(rq, bio);
 678                 if (el_ret == ELEVATOR_BACK_MERGE) {
 679                         if (bio_attempt_back_merge(q, rq, bio)) {
 680                                 ctx->rq_merged++;
 681                                 return true;
 682                         }
 683                         break;
 684                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 685                         if (bio_attempt_front_merge(q, rq, bio)) {
 686                                 ctx->rq_merged++;
 687                                 return true;
 688                         }
 689                         break;
 690                 }
 691         }
 692
 693         return false;
 694 }
 695
 696 /*
 697  * Process software queues that have been marked busy, splicing them
 698  * to the for-dispatch
 699  */
 700 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 701 {
 702         struct blk_mq_ctx *ctx;
 703         int i;
 704
 705         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 706                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 707                 unsigned int off, bit;
 708
 709                 if (!bm->word)
 710                         continue;
 711
 712                 bit = 0;
 713                 off = i * hctx->ctx_map.bits_per_word;
 714                 do {
 715                         bit = find_next_bit(&bm->word, bm->depth, bit);
 716                         if (bit >= bm->depth)
 717                                 break;
 718
 719                         ctx = hctx->ctxs[bit + off];
 720                         clear_bit(bit, &bm->word);
 721                         spin_lock(&ctx->lock);
 722                         list_splice_tail_init(&ctx->rq_list, list);
 723                         spin_unlock(&ctx->lock);
 724
 725                         bit++;
 726                 } while (1);
 727         }
 728 }
 729
 730 /*
 731  * Run this hardware queue, pulling any software queues mapped to it in.
 732  * Note that this function currently has various problems around ordering
 733  * of IO. In particular, we'd like FIFO behaviour on handling existing
 734  * items on the hctx->dispatch list. Ignore that for now.
 735  */
 736 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 737 {
 738         struct request_queue *q = hctx->queue;
 739         struct request *rq;
 740         LIST_HEAD(rq_list);
 741         int queued;
 742
 743         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 744
 745         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 746                 return;
 747
 748         hctx->run++;
 749
 750         /*
 751          * Touch any software queue that has pending entries.
 752          */
 753         flush_busy_ctxs(hctx, &rq_list);
 754
 755         /*
 756          * If we have previous entries on our dispatch list, grab them
 757          * and stuff them at the front for more fair dispatch.
 758          */
 759         if (!list_empty_careful(&hctx->dispatch)) {
 760                 spin_lock(&hctx->lock);
 761                 if (!list_empty(&hctx->dispatch))
 762                         list_splice_init(&hctx->dispatch, &rq_list);
 763                 spin_unlock(&hctx->lock);
 764         }
 765
 766         /*
 767          * Now process all the entries, sending them to the driver.
 768          */
 769         queued = 0;
 770         while (!list_empty(&rq_list)) {
 771                 int ret;
 772
 773                 rq = list_first_entry(&rq_list, struct request, queuelist);
 774                 list_del_init(&rq->queuelist);
 775
 776                 blk_mq_start_request(rq, list_empty(&rq_list));
 777
 778                 ret = q->mq_ops->queue_rq(hctx, rq);
 779                 switch (ret) {
 780                 case BLK_MQ_RQ_QUEUE_OK:
 781                         queued++;
 782                         continue;
 783                 case BLK_MQ_RQ_QUEUE_BUSY:
 784                         list_add(&rq->queuelist, &rq_list);
 785                         __blk_mq_requeue_request(rq);
 786                         break;
 787                 default:
 788                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 789                 case BLK_MQ_RQ_QUEUE_ERROR:
 790                         rq->errors = -EIO;
 791                         blk_mq_end_io(rq, rq->errors);
 792                         break;
 793                 }
 794
 795                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 796                         break;
 797         }
 798
 799         if (!queued)
 800                 hctx->dispatched[0]++;
 801         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 802                 hctx->dispatched[ilog2(queued) + 1]++;
 803
 804         /*
 805          * Any items that need requeuing? Stuff them into hctx->dispatch,
 806          * that is where we will continue on next queue run.
 807          */
 808         if (!list_empty(&rq_list)) {
 809                 spin_lock(&hctx->lock);
 810                 list_splice(&rq_list, &hctx->dispatch);
 811                 spin_unlock(&hctx->lock);
 812         }
 813 }
 814
 815 /*
 816  * It'd be great if the workqueue API had a way to pass
 817  * in a mask and had some smarts for more clever placement.
 818  * For now we just round-robin here, switching for every
 819  * BLK_MQ_CPU_WORK_BATCH queued items.
 820  */
 821 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 822 {
 823         int cpu = hctx->next_cpu;
 824
 825         if (--hctx->next_cpu_batch <= 0) {
 826                 int next_cpu;
 827
 828                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 829                 if (next_cpu >= nr_cpu_ids)
 830                         next_cpu = cpumask_first(hctx->cpumask);
 831
 832                 hctx->next_cpu = next_cpu;
 833                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 834         }
 835
 836         return cpu;
 837 }
 838
 839 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 840 {
 841         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 842                 return;
 843
 844         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 845                 __blk_mq_run_hw_queue(hctx);
 846         else if (hctx->queue->nr_hw_queues == 1)
 847                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 848         else {
 849                 unsigned int cpu;
 850
 851                 cpu = blk_mq_hctx_next_cpu(hctx);
 852                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 853         }
 854 }
 855
 856 void blk_mq_run_queues(struct request_queue *q, bool async)
 857 {
 858         struct blk_mq_hw_ctx *hctx;
 859         int i;
 860
 861         queue_for_each_hw_ctx(q, hctx, i) {
 862                 if ((!blk_mq_hctx_has_pending(hctx) &&
 863                     list_empty_careful(&hctx->dispatch)) ||
 864                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 865                         continue;
 866
 867                 preempt_disable();
 868                 blk_mq_run_hw_queue(hctx, async);
 869                 preempt_enable();
 870         }
 871 }
 872 EXPORT_SYMBOL(blk_mq_run_queues);
 873
 874 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 875 {
 876         cancel_delayed_work(&hctx->run_work);
 877         cancel_delayed_work(&hctx->delay_work);
 878         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 879 }
 880 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 881
 882 void blk_mq_stop_hw_queues(struct request_queue *q)
 883 {
 884         struct blk_mq_hw_ctx *hctx;
 885         int i;
 886
 887         queue_for_each_hw_ctx(q, hctx, i)
 888                 blk_mq_stop_hw_queue(hctx);
 889 }
 890 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 891
 892 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 893 {
 894         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 895
 896         preempt_disable();
 897         __blk_mq_run_hw_queue(hctx);
 898         preempt_enable();
 899 }
 900 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 901
 902 void blk_mq_start_hw_queues(struct request_queue *q)
 903 {
 904         struct blk_mq_hw_ctx *hctx;
 905         int i;
 906
 907         queue_for_each_hw_ctx(q, hctx, i)
 908                 blk_mq_start_hw_queue(hctx);
 909 }
 910 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 911
 912
 913 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 914 {
 915         struct blk_mq_hw_ctx *hctx;
 916         int i;
 917
 918         queue_for_each_hw_ctx(q, hctx, i) {
 919                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 920                         continue;
 921
 922                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 923                 preempt_disable();
 924                 blk_mq_run_hw_queue(hctx, async);
 925                 preempt_enable();
 926         }
 927 }
 928 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 929
 930 static void blk_mq_run_work_fn(struct work_struct *work)
 931 {
 932         struct blk_mq_hw_ctx *hctx;
 933
 934         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 935
 936         __blk_mq_run_hw_queue(hctx);
 937 }
 938
 939 static void blk_mq_delay_work_fn(struct work_struct *work)
 940 {
 941         struct blk_mq_hw_ctx *hctx;
 942
 943         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 944
 945         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 946                 __blk_mq_run_hw_queue(hctx);
 947 }
 948
 949 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 950 {
 951         unsigned long tmo = msecs_to_jiffies(msecs);
 952
 953         if (hctx->queue->nr_hw_queues == 1)
 954                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 955         else {
 956                 unsigned int cpu;
 957
 958                 cpu = blk_mq_hctx_next_cpu(hctx);
 959                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 960         }
 961 }
 962 EXPORT_SYMBOL(blk_mq_delay_queue);
 963
 964 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 965                                     struct request *rq, bool at_head)
 966 {
 967         struct blk_mq_ctx *ctx = rq->mq_ctx;
 968
 969         trace_block_rq_insert(hctx->queue, rq);
 970
 971         if (at_head)
 972                 list_add(&rq->queuelist, &ctx->rq_list);
 973         else
 974                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 975
 976         blk_mq_hctx_mark_pending(hctx, ctx);
 977
 978         /*
 979          * We do this early, to ensure we are on the right CPU.
 980          */
 981         blk_add_timer(rq);
 982 }
 983
 984 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 985                 bool async)
 986 {
 987         struct request_queue *q = rq->q;
 988         struct blk_mq_hw_ctx *hctx;
 989         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 990
 991         current_ctx = blk_mq_get_ctx(q);
 992         if (!cpu_online(ctx->cpu))
 993                 rq->mq_ctx = ctx = current_ctx;
 994
 995         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 996
 997         if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
 998             !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
 999                 blk_insert_flush(rq);
1000         } else {
1001                 spin_lock(&ctx->lock);
1002                 __blk_mq_insert_request(hctx, rq, at_head);
1003                 spin_unlock(&ctx->lock);
1004         }
1005
1006         if (run_queue)
1007                 blk_mq_run_hw_queue(hctx, async);
1008
1009         blk_mq_put_ctx(current_ctx);
1010 }
1011
1012 static void blk_mq_insert_requests(struct request_queue *q,
1013                                      struct blk_mq_ctx *ctx,
1014                                      struct list_head *list,
1015                                      int depth,
1016                                      bool from_schedule)
1017
1018 {
1019         struct blk_mq_hw_ctx *hctx;
1020         struct blk_mq_ctx *current_ctx;
1021
1022         trace_block_unplug(q, depth, !from_schedule);
1023
1024         current_ctx = blk_mq_get_ctx(q);
1025
1026         if (!cpu_online(ctx->cpu))
1027                 ctx = current_ctx;
1028         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1029
1030         /*
1031          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1032          * offline now
1033          */
1034         spin_lock(&ctx->lock);
1035         while (!list_empty(list)) {
1036                 struct request *rq;
1037
1038                 rq = list_first_entry(list, struct request, queuelist);
1039                 list_del_init(&rq->queuelist);
1040                 rq->mq_ctx = ctx;
1041                 __blk_mq_insert_request(hctx, rq, false);
1042         }
1043         spin_unlock(&ctx->lock);
1044
1045         blk_mq_run_hw_queue(hctx, from_schedule);
1046         blk_mq_put_ctx(current_ctx);
1047 }
1048
1049 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1050 {
1051         struct request *rqa = container_of(a, struct request, queuelist);
1052         struct request *rqb = container_of(b, struct request, queuelist);
1053
1054         return !(rqa->mq_ctx < rqb->mq_ctx ||
1055                  (rqa->mq_ctx == rqb->mq_ctx &&
1056                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1057 }
1058
1059 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1060 {
1061         struct blk_mq_ctx *this_ctx;
1062         struct request_queue *this_q;
1063         struct request *rq;
1064         LIST_HEAD(list);
1065         LIST_HEAD(ctx_list);
1066         unsigned int depth;
1067
1068         list_splice_init(&plug->mq_list, &list);
1069
1070         list_sort(NULL, &list, plug_ctx_cmp);
1071
1072         this_q = NULL;
1073         this_ctx = NULL;
1074         depth = 0;
1075
1076         while (!list_empty(&list)) {
1077                 rq = list_entry_rq(list.next);
1078                 list_del_init(&rq->queuelist);
1079                 BUG_ON(!rq->q);
1080                 if (rq->mq_ctx != this_ctx) {
1081                         if (this_ctx) {
1082                                 blk_mq_insert_requests(this_q, this_ctx,
1083                                                         &ctx_list, depth,
1084                                                         from_schedule);
1085                         }
1086
1087                         this_ctx = rq->mq_ctx;
1088                         this_q = rq->q;
1089                         depth = 0;
1090                 }
1091
1092                 depth++;
1093                 list_add_tail(&rq->queuelist, &ctx_list);
1094         }
1095
1096         /*
1097          * If 'this_ctx' is set, we know we have entries to complete
1098          * on 'ctx_list'. Do those.
1099          */
1100         if (this_ctx) {
1101                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1102                                        from_schedule);
1103         }
1104 }
1105
1106 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1107 {
1108         init_request_from_bio(rq, bio);
1109
1110         if (blk_do_io_stat(rq)) {
1111                 rq->start_time = jiffies;
1112                 blk_account_io_start(rq, 1);
1113         }
1114 }
1115
1116 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1117                                          struct blk_mq_ctx *ctx,
1118                                          struct request *rq, struct bio *bio)
1119 {
1120         struct request_queue *q = hctx->queue;
1121
1122         if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1123                 blk_mq_bio_to_request(rq, bio);
1124                 spin_lock(&ctx->lock);
1125 insert_rq:
1126                 __blk_mq_insert_request(hctx, rq, false);
1127                 spin_unlock(&ctx->lock);
1128                 return false;
1129         } else {
1130                 spin_lock(&ctx->lock);
1131                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1132                         blk_mq_bio_to_request(rq, bio);
1133                         goto insert_rq;
1134                 }
1135
1136                 spin_unlock(&ctx->lock);
1137                 __blk_mq_free_request(hctx, ctx, rq);
1138                 return true;
1139         }
1140 }
1141
1142 struct blk_map_ctx {
1143         struct blk_mq_hw_ctx *hctx;
1144         struct blk_mq_ctx *ctx;
1145 };
1146
1147 static struct request *blk_mq_map_request(struct request_queue *q,
1148                                           struct bio *bio,
1149                                           struct blk_map_ctx *data)
1150 {
1151         struct blk_mq_hw_ctx *hctx;
1152         struct blk_mq_ctx *ctx;
1153         struct request *rq;
1154         int rw = bio_data_dir(bio);
1155
1156         if (unlikely(blk_mq_queue_enter(q))) {
1157                 bio_endio(bio, -EIO);
1158                 return NULL;
1159         }
1160
1161         ctx = blk_mq_get_ctx(q);
1162         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1163
1164         if (rw_is_sync(bio->bi_rw))
1165                 rw |= REQ_SYNC;
1166
1167         trace_block_getrq(q, bio, rw);
1168         rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
1169         if (unlikely(!rq)) {
1170                 __blk_mq_run_hw_queue(hctx);
1171                 blk_mq_put_ctx(ctx);
1172                 trace_block_sleeprq(q, bio, rw);
1173
1174                 ctx = blk_mq_get_ctx(q);
1175                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1176                 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1177                                             __GFP_WAIT|GFP_ATOMIC, false);
1178         }
1179
1180         hctx->queued++;
1181         data->hctx = hctx;
1182         data->ctx = ctx;
1183         return rq;
1184 }
1185
1186 /*
1187  * Multiple hardware queue variant. This will not use per-process plugs,
1188  * but will attempt to bypass the hctx queueing if we can go straight to
1189  * hardware for SYNC IO.
1190  */
1191 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1192 {
1193         const int is_sync = rw_is_sync(bio->bi_rw);
1194         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1195         struct blk_map_ctx data;
1196         struct request *rq;
1197
1198         blk_queue_bounce(q, &bio);
1199
1200         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1201                 bio_endio(bio, -EIO);
1202                 return;
1203         }
1204
1205         rq = blk_mq_map_request(q, bio, &data);
1206         if (unlikely(!rq))
1207                 return;
1208
1209         if (unlikely(is_flush_fua)) {
1210                 blk_mq_bio_to_request(rq, bio);
1211                 blk_insert_flush(rq);
1212                 goto run_queue;
1213         }
1214
1215         if (is_sync) {
1216                 int ret;
1217
1218                 blk_mq_bio_to_request(rq, bio);
1219                 blk_mq_start_request(rq, true);
1220                 blk_add_timer(rq);
1221
1222                 /*
1223                  * For OK queue, we are done. For error, kill it. Any other
1224                  * error (busy), just add it to our list as we previously
1225                  * would have done
1226                  */
1227                 ret = q->mq_ops->queue_rq(data.hctx, rq);
1228                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1229                         goto done;
1230                 else {
1231                         __blk_mq_requeue_request(rq);
1232
1233                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1234                                 rq->errors = -EIO;
1235                                 blk_mq_end_io(rq, rq->errors);
1236                                 goto done;
1237                         }
1238                 }
1239         }
1240
1241         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1242                 /*
1243                  * For a SYNC request, send it to the hardware immediately. For
1244                  * an ASYNC request, just ensure that we run it later on. The
1245                  * latter allows for merging opportunities and more efficient
1246                  * dispatching.
1247                  */
1248 run_queue:
1249                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1250         }
1251 done:
1252         blk_mq_put_ctx(data.ctx);
1253 }
1254
1255 /*
1256  * Single hardware queue variant. This will attempt to use any per-process
1257  * plug for merging and IO deferral.
1258  */
1259 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1260 {
1261         const int is_sync = rw_is_sync(bio->bi_rw);
1262         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1263         unsigned int use_plug, request_count = 0;
1264         struct blk_map_ctx data;
1265         struct request *rq;
1266
1267         /*
1268          * If we have multiple hardware queues, just go directly to
1269          * one of those for sync IO.
1270          */
1271         use_plug = !is_flush_fua && !is_sync;
1272
1273         blk_queue_bounce(q, &bio);
1274
1275         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1276                 bio_endio(bio, -EIO);
1277                 return;
1278         }
1279
1280         if (use_plug && !blk_queue_nomerges(q) &&
1281             blk_attempt_plug_merge(q, bio, &request_count))
1282                 return;
1283
1284         rq = blk_mq_map_request(q, bio, &data);
1285
1286         if (unlikely(is_flush_fua)) {
1287                 blk_mq_bio_to_request(rq, bio);
1288                 blk_insert_flush(rq);
1289                 goto run_queue;
1290         }
1291
1292         /*
1293          * A task plug currently exists. Since this is completely lockless,
1294          * utilize that to temporarily store requests until the task is
1295          * either done or scheduled away.
1296          */
1297         if (use_plug) {
1298                 struct blk_plug *plug = current->plug;
1299
1300                 if (plug) {
1301                         blk_mq_bio_to_request(rq, bio);
1302                         if (list_empty(&plug->mq_list))
1303                                 trace_block_plug(q);
1304                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1305                                 blk_flush_plug_list(plug, false);
1306                                 trace_block_plug(q);
1307                         }
1308                         list_add_tail(&rq->queuelist, &plug->mq_list);
1309                         blk_mq_put_ctx(data.ctx);
1310                         return;
1311                 }
1312         }
1313
1314         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1315                 /*
1316                  * For a SYNC request, send it to the hardware immediately. For
1317                  * an ASYNC request, just ensure that we run it later on. The
1318                  * latter allows for merging opportunities and more efficient
1319                  * dispatching.
1320                  */
1321 run_queue:
1322                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1323         }
1324
1325         blk_mq_put_ctx(data.ctx);
1326 }
1327
1328 /*
1329  * Default mapping to a software queue, since we use one per CPU.
1330  */
1331 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1332 {
1333         return q->queue_hw_ctx[q->mq_map[cpu]];
1334 }
1335 EXPORT_SYMBOL(blk_mq_map_queue);
1336
1337 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1338                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1339 {
1340         struct page *page;
1341
1342         if (tags->rqs && set->ops->exit_request) {
1343                 int i;
1344
1345                 for (i = 0; i < tags->nr_tags; i++) {
1346                         if (!tags->rqs[i])
1347                                 continue;
1348                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1349                                                 hctx_idx, i);
1350                 }
1351         }
1352
1353         while (!list_empty(&tags->page_list)) {
1354                 page = list_first_entry(&tags->page_list, struct page, lru);
1355                 list_del_init(&page->lru);
1356                 __free_pages(page, page->private);
1357         }
1358
1359         kfree(tags->rqs);
1360
1361         blk_mq_free_tags(tags);
1362 }
1363
1364 static size_t order_to_size(unsigned int order)
1365 {
1366         return (size_t)PAGE_SIZE << order;
1367 }
1368
1369 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1370                 unsigned int hctx_idx)
1371 {
1372         struct blk_mq_tags *tags;
1373         unsigned int i, j, entries_per_page, max_order = 4;
1374         size_t rq_size, left;
1375
1376         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1377                                 set->numa_node);
1378         if (!tags)
1379                 return NULL;
1380
1381         INIT_LIST_HEAD(&tags->page_list);
1382
1383         tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1384                                         GFP_KERNEL, set->numa_node);
1385         if (!tags->rqs) {
1386                 blk_mq_free_tags(tags);
1387                 return NULL;
1388         }
1389
1390         /*
1391          * rq_size is the size of the request plus driver payload, rounded
1392          * to the cacheline size
1393          */
1394         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1395                                 cache_line_size());
1396         left = rq_size * set->queue_depth;
1397
1398         for (i = 0; i < set->queue_depth; ) {
1399                 int this_order = max_order;
1400                 struct page *page;
1401                 int to_do;
1402                 void *p;
1403
1404                 while (left < order_to_size(this_order - 1) && this_order)
1405                         this_order--;
1406
1407                 do {
1408                         page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1409                                                 this_order);
1410                         if (page)
1411                                 break;
1412                         if (!this_order--)
1413                                 break;
1414                         if (order_to_size(this_order) < rq_size)
1415                                 break;
1416                 } while (1);
1417
1418                 if (!page)
1419                         goto fail;
1420
1421                 page->private = this_order;
1422                 list_add_tail(&page->lru, &tags->page_list);
1423
1424                 p = page_address(page);
1425                 entries_per_page = order_to_size(this_order) / rq_size;
1426                 to_do = min(entries_per_page, set->queue_depth - i);
1427                 left -= to_do * rq_size;
1428                 for (j = 0; j < to_do; j++) {
1429                         tags->rqs[i] = p;
1430                         if (set->ops->init_request) {
1431                                 if (set->ops->init_request(set->driver_data,
1432                                                 tags->rqs[i], hctx_idx, i,
1433                                                 set->numa_node))
1434                                         goto fail;
1435                         }
1436
1437                         p += rq_size;
1438                         i++;
1439                 }
1440         }
1441
1442         return tags;
1443
1444 fail:
1445         pr_warn("%s: failed to allocate requests\n", __func__);
1446         blk_mq_free_rq_map(set, tags, hctx_idx);
1447         return NULL;
1448 }
1449
1450 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1451 {
1452         kfree(bitmap->map);
1453 }
1454
1455 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1456 {
1457         unsigned int bpw = 8, total, num_maps, i;
1458
1459         bitmap->bits_per_word = bpw;
1460
1461         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1462         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1463                                         GFP_KERNEL, node);
1464         if (!bitmap->map)
1465                 return -ENOMEM;
1466
1467         bitmap->map_size = num_maps;
1468
1469         total = nr_cpu_ids;
1470         for (i = 0; i < num_maps; i++) {
1471                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1472                 total -= bitmap->map[i].depth;
1473         }
1474
1475         return 0;
1476 }
1477
1478 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1479 {
1480         struct request_queue *q = hctx->queue;
1481         struct blk_mq_ctx *ctx;
1482         LIST_HEAD(tmp);
1483
1484         /*
1485          * Move ctx entries to new CPU, if this one is going away.
1486          */
1487         ctx = __blk_mq_get_ctx(q, cpu);
1488
1489         spin_lock(&ctx->lock);
1490         if (!list_empty(&ctx->rq_list)) {
1491                 list_splice_init(&ctx->rq_list, &tmp);
1492                 blk_mq_hctx_clear_pending(hctx, ctx);
1493         }
1494         spin_unlock(&ctx->lock);
1495
1496         if (list_empty(&tmp))
1497                 return NOTIFY_OK;
1498
1499         ctx = blk_mq_get_ctx(q);
1500         spin_lock(&ctx->lock);
1501
1502         while (!list_empty(&tmp)) {
1503                 struct request *rq;
1504
1505                 rq = list_first_entry(&tmp, struct request, queuelist);
1506                 rq->mq_ctx = ctx;
1507                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1508         }
1509
1510         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1511         blk_mq_hctx_mark_pending(hctx, ctx);
1512
1513         spin_unlock(&ctx->lock);
1514
1515         blk_mq_run_hw_queue(hctx, true);
1516         blk_mq_put_ctx(ctx);
1517         return NOTIFY_OK;
1518 }
1519
1520 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1521 {
1522         struct request_queue *q = hctx->queue;
1523         struct blk_mq_tag_set *set = q->tag_set;
1524
1525         if (set->tags[hctx->queue_num])
1526                 return NOTIFY_OK;
1527
1528         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1529         if (!set->tags[hctx->queue_num])
1530                 return NOTIFY_STOP;
1531
1532         hctx->tags = set->tags[hctx->queue_num];
1533         return NOTIFY_OK;
1534 }
1535
1536 static int blk_mq_hctx_notify(void *data, unsigned long action,
1537                               unsigned int cpu)
1538 {
1539         struct blk_mq_hw_ctx *hctx = data;
1540
1541         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1542                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1543         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1544                 return blk_mq_hctx_cpu_online(hctx, cpu);
1545
1546         return NOTIFY_OK;
1547 }
1548
1549 static void blk_mq_exit_hw_queues(struct request_queue *q,
1550                 struct blk_mq_tag_set *set, int nr_queue)
1551 {
1552         struct blk_mq_hw_ctx *hctx;
1553         unsigned int i;
1554
1555         queue_for_each_hw_ctx(q, hctx, i) {
1556                 if (i == nr_queue)
1557                         break;
1558
1559                 if (set->ops->exit_hctx)
1560                         set->ops->exit_hctx(hctx, i);
1561
1562                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1563                 kfree(hctx->ctxs);
1564                 blk_mq_free_bitmap(&hctx->ctx_map);
1565         }
1566
1567 }
1568
1569 static void blk_mq_free_hw_queues(struct request_queue *q,
1570                 struct blk_mq_tag_set *set)
1571 {
1572         struct blk_mq_hw_ctx *hctx;
1573         unsigned int i;
1574
1575         queue_for_each_hw_ctx(q, hctx, i) {
1576                 free_cpumask_var(hctx->cpumask);
1577                 kfree(hctx);
1578         }
1579 }
1580
1581 static int blk_mq_init_hw_queues(struct request_queue *q,
1582                 struct blk_mq_tag_set *set)
1583 {
1584         struct blk_mq_hw_ctx *hctx;
1585         unsigned int i;
1586
1587         /*
1588          * Initialize hardware queues
1589          */
1590         queue_for_each_hw_ctx(q, hctx, i) {
1591                 int node;
1592
1593                 node = hctx->numa_node;
1594                 if (node == NUMA_NO_NODE)
1595                         node = hctx->numa_node = set->numa_node;
1596
1597                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1598                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1599                 spin_lock_init(&hctx->lock);
1600                 INIT_LIST_HEAD(&hctx->dispatch);
1601                 hctx->queue = q;
1602                 hctx->queue_num = i;
1603                 hctx->flags = set->flags;
1604                 hctx->cmd_size = set->cmd_size;
1605
1606                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1607                                                 blk_mq_hctx_notify, hctx);
1608                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1609
1610                 hctx->tags = set->tags[i];
1611
1612                 /*
1613                  * Allocate space for all possible cpus to avoid allocation in
1614                  * runtime
1615                  */
1616                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1617                                                 GFP_KERNEL, node);
1618                 if (!hctx->ctxs)
1619                         break;
1620
1621                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1622                         break;
1623
1624                 hctx->nr_ctx = 0;
1625
1626                 if (set->ops->init_hctx &&
1627                     set->ops->init_hctx(hctx, set->driver_data, i))
1628                         break;
1629         }
1630
1631         if (i == q->nr_hw_queues)
1632                 return 0;
1633
1634         /*
1635          * Init failed
1636          */
1637         blk_mq_exit_hw_queues(q, set, i);
1638
1639         return 1;
1640 }
1641
1642 static void blk_mq_init_cpu_queues(struct request_queue *q,
1643                                    unsigned int nr_hw_queues)
1644 {
1645         unsigned int i;
1646
1647         for_each_possible_cpu(i) {
1648                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1649                 struct blk_mq_hw_ctx *hctx;
1650
1651                 memset(__ctx, 0, sizeof(*__ctx));
1652                 __ctx->cpu = i;
1653                 spin_lock_init(&__ctx->lock);
1654                 INIT_LIST_HEAD(&__ctx->rq_list);
1655                 __ctx->queue = q;
1656
1657                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1658                 if (!cpu_online(i))
1659                         continue;
1660
1661                 hctx = q->mq_ops->map_queue(q, i);
1662                 cpumask_set_cpu(i, hctx->cpumask);
1663                 hctx->nr_ctx++;
1664
1665                 /*
1666                  * Set local node, IFF we have more than one hw queue. If
1667                  * not, we remain on the home node of the device
1668                  */
1669                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1670                         hctx->numa_node = cpu_to_node(i);
1671         }
1672 }
1673
1674 static void blk_mq_map_swqueue(struct request_queue *q)
1675 {
1676         unsigned int i;
1677         struct blk_mq_hw_ctx *hctx;
1678         struct blk_mq_ctx *ctx;
1679
1680         queue_for_each_hw_ctx(q, hctx, i) {
1681                 cpumask_clear(hctx->cpumask);
1682                 hctx->nr_ctx = 0;
1683         }
1684
1685         /*
1686          * Map software to hardware queues
1687          */
1688         queue_for_each_ctx(q, ctx, i) {
1689                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1690                 if (!cpu_online(i))
1691                         continue;
1692
1693                 hctx = q->mq_ops->map_queue(q, i);
1694                 cpumask_set_cpu(i, hctx->cpumask);
1695                 ctx->index_hw = hctx->nr_ctx;
1696                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1697         }
1698
1699         queue_for_each_hw_ctx(q, hctx, i) {
1700                 /*
1701                  * If not software queues are mapped to this hardware queue,
1702                  * disable it and free the request entries
1703                  */
1704                 if (!hctx->nr_ctx) {
1705                         struct blk_mq_tag_set *set = q->tag_set;
1706
1707                         if (set->tags[i]) {
1708                                 blk_mq_free_rq_map(set, set->tags[i], i);
1709                                 set->tags[i] = NULL;
1710                                 hctx->tags = NULL;
1711                         }
1712                         continue;
1713                 }
1714
1715                 /*
1716                  * Initialize batch roundrobin counts
1717                  */
1718                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1719                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1720         }
1721 }
1722
1723 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1724 {
1725         struct blk_mq_hw_ctx *hctx;
1726         struct request_queue *q;
1727         bool shared;
1728         int i;
1729
1730         if (set->tag_list.next == set->tag_list.prev)
1731                 shared = false;
1732         else
1733                 shared = true;
1734
1735         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1736                 blk_mq_freeze_queue(q);
1737
1738                 queue_for_each_hw_ctx(q, hctx, i) {
1739                         if (shared)
1740                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1741                         else
1742                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1743                 }
1744                 blk_mq_unfreeze_queue(q);
1745         }
1746 }
1747
1748 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1749 {
1750         struct blk_mq_tag_set *set = q->tag_set;
1751
1752         blk_mq_freeze_queue(q);
1753
1754         mutex_lock(&set->tag_list_lock);
1755         list_del_init(&q->tag_set_list);
1756         blk_mq_update_tag_set_depth(set);
1757         mutex_unlock(&set->tag_list_lock);
1758
1759         blk_mq_unfreeze_queue(q);
1760 }
1761
1762 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1763                                      struct request_queue *q)
1764 {
1765         q->tag_set = set;
1766
1767         mutex_lock(&set->tag_list_lock);
1768         list_add_tail(&q->tag_set_list, &set->tag_list);
1769         blk_mq_update_tag_set_depth(set);
1770         mutex_unlock(&set->tag_list_lock);
1771 }
1772
1773 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1774 {
1775         struct blk_mq_hw_ctx **hctxs;
1776         struct blk_mq_ctx *ctx;
1777         struct request_queue *q;
1778         unsigned int *map;
1779         int i;
1780
1781         ctx = alloc_percpu(struct blk_mq_ctx);
1782         if (!ctx)
1783                 return ERR_PTR(-ENOMEM);
1784
1785         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1786                         set->numa_node);
1787
1788         if (!hctxs)
1789                 goto err_percpu;
1790
1791         map = blk_mq_make_queue_map(set);
1792         if (!map)
1793                 goto err_map;
1794
1795         for (i = 0; i < set->nr_hw_queues; i++) {
1796                 int node = blk_mq_hw_queue_to_node(map, i);
1797
1798                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1799                                         GFP_KERNEL, node);
1800                 if (!hctxs[i])
1801                         goto err_hctxs;
1802
1803                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1804                         goto err_hctxs;
1805
1806                 atomic_set(&hctxs[i]->nr_active, 0);
1807                 hctxs[i]->numa_node = node;
1808                 hctxs[i]->queue_num = i;
1809         }
1810
1811         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1812         if (!q)
1813                 goto err_hctxs;
1814
1815         if (percpu_counter_init(&q->mq_usage_counter, 0))
1816                 goto err_map;
1817
1818         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1819         blk_queue_rq_timeout(q, 30000);
1820
1821         q->nr_queues = nr_cpu_ids;
1822         q->nr_hw_queues = set->nr_hw_queues;
1823         q->mq_map = map;
1824
1825         q->queue_ctx = ctx;
1826         q->queue_hw_ctx = hctxs;
1827
1828         q->mq_ops = set->ops;
1829         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1830
1831         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1832                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1833
1834         q->sg_reserved_size = INT_MAX;
1835
1836         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1837         INIT_LIST_HEAD(&q->requeue_list);
1838         spin_lock_init(&q->requeue_lock);
1839
1840         if (q->nr_hw_queues > 1)
1841                 blk_queue_make_request(q, blk_mq_make_request);
1842         else
1843                 blk_queue_make_request(q, blk_sq_make_request);
1844
1845         blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1846         if (set->timeout)
1847                 blk_queue_rq_timeout(q, set->timeout);
1848
1849         /*
1850          * Do this after blk_queue_make_request() overrides it...
1851          */
1852         q->nr_requests = set->queue_depth;
1853
1854         if (set->ops->complete)
1855                 blk_queue_softirq_done(q, set->ops->complete);
1856
1857         blk_mq_init_flush(q);
1858         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1859
1860         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1861                                 set->cmd_size, cache_line_size()),
1862                                 GFP_KERNEL);
1863         if (!q->flush_rq)
1864                 goto err_hw;
1865
1866         if (blk_mq_init_hw_queues(q, set))
1867                 goto err_flush_rq;
1868
1869         mutex_lock(&all_q_mutex);
1870         list_add_tail(&q->all_q_node, &all_q_list);
1871         mutex_unlock(&all_q_mutex);
1872
1873         blk_mq_add_queue_tag_set(set, q);
1874
1875         blk_mq_map_swqueue(q);
1876
1877         return q;
1878
1879 err_flush_rq:
1880         kfree(q->flush_rq);
1881 err_hw:
1882         blk_cleanup_queue(q);
1883 err_hctxs:
1884         kfree(map);
1885         for (i = 0; i < set->nr_hw_queues; i++) {
1886                 if (!hctxs[i])
1887                         break;
1888                 free_cpumask_var(hctxs[i]->cpumask);
1889                 kfree(hctxs[i]);
1890         }
1891 err_map:
1892         kfree(hctxs);
1893 err_percpu:
1894         free_percpu(ctx);
1895         return ERR_PTR(-ENOMEM);
1896 }
1897 EXPORT_SYMBOL(blk_mq_init_queue);
1898
1899 void blk_mq_free_queue(struct request_queue *q)
1900 {
1901         struct blk_mq_tag_set   *set = q->tag_set;
1902
1903         blk_mq_del_queue_tag_set(q);
1904
1905         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1906         blk_mq_free_hw_queues(q, set);
1907
1908         percpu_counter_destroy(&q->mq_usage_counter);
1909
1910         free_percpu(q->queue_ctx);
1911         kfree(q->queue_hw_ctx);
1912         kfree(q->mq_map);
1913
1914         q->queue_ctx = NULL;
1915         q->queue_hw_ctx = NULL;
1916         q->mq_map = NULL;
1917
1918         mutex_lock(&all_q_mutex);
1919         list_del_init(&q->all_q_node);
1920         mutex_unlock(&all_q_mutex);
1921 }
1922
1923 /* Basically redo blk_mq_init_queue with queue frozen */
1924 static void blk_mq_queue_reinit(struct request_queue *q)
1925 {
1926         blk_mq_freeze_queue(q);
1927
1928         blk_mq_sysfs_unregister(q);
1929
1930         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1931
1932         /*
1933          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1934          * we should change hctx numa_node according to new topology (this
1935          * involves free and re-allocate memory, worthy doing?)
1936          */
1937
1938         blk_mq_map_swqueue(q);
1939
1940         blk_mq_sysfs_register(q);
1941
1942         blk_mq_unfreeze_queue(q);
1943 }
1944
1945 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1946                                       unsigned long action, void *hcpu)
1947 {
1948         struct request_queue *q;
1949
1950         /*
1951          * Before new mappings are established, hotadded cpu might already
1952          * start handling requests. This doesn't break anything as we map
1953          * offline CPUs to first hardware queue. We will re-init the queue
1954          * below to get optimal settings.
1955          */
1956         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1957             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1958                 return NOTIFY_OK;
1959
1960         mutex_lock(&all_q_mutex);
1961         list_for_each_entry(q, &all_q_list, all_q_node)
1962                 blk_mq_queue_reinit(q);
1963         mutex_unlock(&all_q_mutex);
1964         return NOTIFY_OK;
1965 }
1966
1967 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1968 {
1969         int i;
1970
1971         if (!set->nr_hw_queues)
1972                 return -EINVAL;
1973         if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1974                 return -EINVAL;
1975         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1976                 return -EINVAL;
1977
1978         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1979                 return -EINVAL;
1980
1981
1982         set->tags = kmalloc_node(set->nr_hw_queues *
1983                                  sizeof(struct blk_mq_tags *),
1984                                  GFP_KERNEL, set->numa_node);
1985         if (!set->tags)
1986                 goto out;
1987
1988         for (i = 0; i < set->nr_hw_queues; i++) {
1989                 set->tags[i] = blk_mq_init_rq_map(set, i);
1990                 if (!set->tags[i])
1991                         goto out_unwind;
1992         }
1993
1994         mutex_init(&set->tag_list_lock);
1995         INIT_LIST_HEAD(&set->tag_list);
1996
1997         return 0;
1998
1999 out_unwind:
2000         while (--i >= 0)
2001                 blk_mq_free_rq_map(set, set->tags[i], i);
2002 out:
2003         return -ENOMEM;
2004 }
2005 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2006
2007 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2008 {
2009         int i;
2010
2011         for (i = 0; i < set->nr_hw_queues; i++) {
2012                 if (set->tags[i])
2013                         blk_mq_free_rq_map(set, set->tags[i], i);
2014         }
2015
2016         kfree(set->tags);
2017 }
2018 EXPORT_SYMBOL(blk_mq_free_tag_set);
2019
2020 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2021 {
2022         struct blk_mq_tag_set *set = q->tag_set;
2023         struct blk_mq_hw_ctx *hctx;
2024         int i, ret;
2025
2026         if (!set || nr > set->queue_depth)
2027                 return -EINVAL;
2028
2029         ret = 0;
2030         queue_for_each_hw_ctx(q, hctx, i) {
2031                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2032                 if (ret)
2033                         break;
2034         }
2035
2036         if (!ret)
2037                 q->nr_requests = nr;
2038
2039         return ret;
2040 }
2041
2042 void blk_mq_disable_hotplug(void)
2043 {
2044         mutex_lock(&all_q_mutex);
2045 }
2046
2047 void blk_mq_enable_hotplug(void)
2048 {
2049         mutex_unlock(&all_q_mutex);
2050 }
2051
2052 static int __init blk_mq_init(void)
2053 {
2054         blk_mq_cpu_init();
2055
2056         /* Must be called after percpu_counter_hotcpu_callback() */
2057         hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2058
2059         return 0;
2060 }
2061 subsys_initcall(blk_mq_init);