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