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 blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1072                                          struct blk_mq_ctx *ctx,
1073                                          struct request *rq, struct bio *bio)
1074 {
1075         struct request_queue *q = hctx->queue;
1076
1077         if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1078                 blk_mq_bio_to_request(rq, bio);
1079                 spin_lock(&ctx->lock);
1080 insert_rq:
1081                 __blk_mq_insert_request(hctx, rq, false);
1082                 spin_unlock(&ctx->lock);
1083                 return false;
1084         } else {
1085                 spin_lock(&ctx->lock);
1086                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1087                         blk_mq_bio_to_request(rq, bio);
1088                         goto insert_rq;
1089                 }
1090
1091                 spin_unlock(&ctx->lock);
1092                 __blk_mq_free_request(hctx, ctx, rq);
1093                 return true;
1094         }
1095 }
1096
1097 struct blk_map_ctx {
1098         struct blk_mq_hw_ctx *hctx;
1099         struct blk_mq_ctx *ctx;
1100 };
1101
1102 static struct request *blk_mq_map_request(struct request_queue *q,
1103                                           struct bio *bio,
1104                                           struct blk_map_ctx *data)
1105 {
1106         struct blk_mq_hw_ctx *hctx;
1107         struct blk_mq_ctx *ctx;
1108         struct request *rq;
1109         int rw = bio_data_dir(bio);
1110         struct blk_mq_alloc_data alloc_data;
1111
1112         if (unlikely(blk_mq_queue_enter(q))) {
1113                 bio_endio(bio, -EIO);
1114                 return NULL;
1115         }
1116
1117         ctx = blk_mq_get_ctx(q);
1118         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1119
1120         if (rw_is_sync(bio->bi_rw))
1121                 rw |= REQ_SYNC;
1122
1123         trace_block_getrq(q, bio, rw);
1124         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1125                         hctx);
1126         rq = __blk_mq_alloc_request(&alloc_data, rw);
1127         if (unlikely(!rq)) {
1128                 __blk_mq_run_hw_queue(hctx);
1129                 blk_mq_put_ctx(ctx);
1130                 trace_block_sleeprq(q, bio, rw);
1131
1132                 ctx = blk_mq_get_ctx(q);
1133                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1134                 blk_mq_set_alloc_data(&alloc_data, q,
1135                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1136                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1137                 ctx = alloc_data.ctx;
1138                 hctx = alloc_data.hctx;
1139         }
1140
1141         hctx->queued++;
1142         data->hctx = hctx;
1143         data->ctx = ctx;
1144         return rq;
1145 }
1146
1147 /*
1148  * Multiple hardware queue variant. This will not use per-process plugs,
1149  * but will attempt to bypass the hctx queueing if we can go straight to
1150  * hardware for SYNC IO.
1151  */
1152 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1153 {
1154         const int is_sync = rw_is_sync(bio->bi_rw);
1155         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1156         struct blk_map_ctx data;
1157         struct request *rq;
1158
1159         blk_queue_bounce(q, &bio);
1160
1161         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1162                 bio_endio(bio, -EIO);
1163                 return;
1164         }
1165
1166         rq = blk_mq_map_request(q, bio, &data);
1167         if (unlikely(!rq))
1168                 return;
1169
1170         if (unlikely(is_flush_fua)) {
1171                 blk_mq_bio_to_request(rq, bio);
1172                 blk_insert_flush(rq);
1173                 goto run_queue;
1174         }
1175
1176         if (is_sync) {
1177                 int ret;
1178
1179                 blk_mq_bio_to_request(rq, bio);
1180                 blk_mq_start_request(rq, true);
1181
1182                 /*
1183                  * For OK queue, we are done. For error, kill it. Any other
1184                  * error (busy), just add it to our list as we previously
1185                  * would have done
1186                  */
1187                 ret = q->mq_ops->queue_rq(data.hctx, rq);
1188                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1189                         goto done;
1190                 else {
1191                         __blk_mq_requeue_request(rq);
1192
1193                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1194                                 rq->errors = -EIO;
1195                                 blk_mq_end_io(rq, rq->errors);
1196                                 goto done;
1197                         }
1198                 }
1199         }
1200
1201         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1202                 /*
1203                  * For a SYNC request, send it to the hardware immediately. For
1204                  * an ASYNC request, just ensure that we run it later on. The
1205                  * latter allows for merging opportunities and more efficient
1206                  * dispatching.
1207                  */
1208 run_queue:
1209                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1210         }
1211 done:
1212         blk_mq_put_ctx(data.ctx);
1213 }
1214
1215 /*
1216  * Single hardware queue variant. This will attempt to use any per-process
1217  * plug for merging and IO deferral.
1218  */
1219 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1220 {
1221         const int is_sync = rw_is_sync(bio->bi_rw);
1222         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1223         unsigned int use_plug, request_count = 0;
1224         struct blk_map_ctx data;
1225         struct request *rq;
1226
1227         /*
1228          * If we have multiple hardware queues, just go directly to
1229          * one of those for sync IO.
1230          */
1231         use_plug = !is_flush_fua && !is_sync;
1232
1233         blk_queue_bounce(q, &bio);
1234
1235         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1236                 bio_endio(bio, -EIO);
1237                 return;
1238         }
1239
1240         if (use_plug && !blk_queue_nomerges(q) &&
1241             blk_attempt_plug_merge(q, bio, &request_count))
1242                 return;
1243
1244         rq = blk_mq_map_request(q, bio, &data);
1245         if (unlikely(!rq))
1246                 return;
1247
1248         if (unlikely(is_flush_fua)) {
1249                 blk_mq_bio_to_request(rq, bio);
1250                 blk_insert_flush(rq);
1251                 goto run_queue;
1252         }
1253
1254         /*
1255          * A task plug currently exists. Since this is completely lockless,
1256          * utilize that to temporarily store requests until the task is
1257          * either done or scheduled away.
1258          */
1259         if (use_plug) {
1260                 struct blk_plug *plug = current->plug;
1261
1262                 if (plug) {
1263                         blk_mq_bio_to_request(rq, bio);
1264                         if (list_empty(&plug->mq_list))
1265                                 trace_block_plug(q);
1266                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1267                                 blk_flush_plug_list(plug, false);
1268                                 trace_block_plug(q);
1269                         }
1270                         list_add_tail(&rq->queuelist, &plug->mq_list);
1271                         blk_mq_put_ctx(data.ctx);
1272                         return;
1273                 }
1274         }
1275
1276         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1277                 /*
1278                  * For a SYNC request, send it to the hardware immediately. For
1279                  * an ASYNC request, just ensure that we run it later on. The
1280                  * latter allows for merging opportunities and more efficient
1281                  * dispatching.
1282                  */
1283 run_queue:
1284                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1285         }
1286
1287         blk_mq_put_ctx(data.ctx);
1288 }
1289
1290 /*
1291  * Default mapping to a software queue, since we use one per CPU.
1292  */
1293 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1294 {
1295         return q->queue_hw_ctx[q->mq_map[cpu]];
1296 }
1297 EXPORT_SYMBOL(blk_mq_map_queue);
1298
1299 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1300                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1301 {
1302         struct page *page;
1303
1304         if (tags->rqs && set->ops->exit_request) {
1305                 int i;
1306
1307                 for (i = 0; i < tags->nr_tags; i++) {
1308                         if (!tags->rqs[i])
1309                                 continue;
1310                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1311                                                 hctx_idx, i);
1312                 }
1313         }
1314
1315         while (!list_empty(&tags->page_list)) {
1316                 page = list_first_entry(&tags->page_list, struct page, lru);
1317                 list_del_init(&page->lru);
1318                 __free_pages(page, page->private);
1319         }
1320
1321         kfree(tags->rqs);
1322
1323         blk_mq_free_tags(tags);
1324 }
1325
1326 static size_t order_to_size(unsigned int order)
1327 {
1328         return (size_t)PAGE_SIZE << order;
1329 }
1330
1331 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1332                 unsigned int hctx_idx)
1333 {
1334         struct blk_mq_tags *tags;
1335         unsigned int i, j, entries_per_page, max_order = 4;
1336         size_t rq_size, left;
1337
1338         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1339                                 set->numa_node);
1340         if (!tags)
1341                 return NULL;
1342
1343         INIT_LIST_HEAD(&tags->page_list);
1344
1345         tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1346                                         GFP_KERNEL, set->numa_node);
1347         if (!tags->rqs) {
1348                 blk_mq_free_tags(tags);
1349                 return NULL;
1350         }
1351
1352         /*
1353          * rq_size is the size of the request plus driver payload, rounded
1354          * to the cacheline size
1355          */
1356         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1357                                 cache_line_size());
1358         left = rq_size * set->queue_depth;
1359
1360         for (i = 0; i < set->queue_depth; ) {
1361                 int this_order = max_order;
1362                 struct page *page;
1363                 int to_do;
1364                 void *p;
1365
1366                 while (left < order_to_size(this_order - 1) && this_order)
1367                         this_order--;
1368
1369                 do {
1370                         page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1371                                                 this_order);
1372                         if (page)
1373                                 break;
1374                         if (!this_order--)
1375                                 break;
1376                         if (order_to_size(this_order) < rq_size)
1377                                 break;
1378                 } while (1);
1379
1380                 if (!page)
1381                         goto fail;
1382
1383                 page->private = this_order;
1384                 list_add_tail(&page->lru, &tags->page_list);
1385
1386                 p = page_address(page);
1387                 entries_per_page = order_to_size(this_order) / rq_size;
1388                 to_do = min(entries_per_page, set->queue_depth - i);
1389                 left -= to_do * rq_size;
1390                 for (j = 0; j < to_do; j++) {
1391                         tags->rqs[i] = p;
1392                         if (set->ops->init_request) {
1393                                 if (set->ops->init_request(set->driver_data,
1394                                                 tags->rqs[i], hctx_idx, i,
1395                                                 set->numa_node))
1396                                         goto fail;
1397                         }
1398
1399                         p += rq_size;
1400                         i++;
1401                 }
1402         }
1403
1404         return tags;
1405
1406 fail:
1407         pr_warn("%s: failed to allocate requests\n", __func__);
1408         blk_mq_free_rq_map(set, tags, hctx_idx);
1409         return NULL;
1410 }
1411
1412 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1413 {
1414         kfree(bitmap->map);
1415 }
1416
1417 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1418 {
1419         unsigned int bpw = 8, total, num_maps, i;
1420
1421         bitmap->bits_per_word = bpw;
1422
1423         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1424         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1425                                         GFP_KERNEL, node);
1426         if (!bitmap->map)
1427                 return -ENOMEM;
1428
1429         bitmap->map_size = num_maps;
1430
1431         total = nr_cpu_ids;
1432         for (i = 0; i < num_maps; i++) {
1433                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1434                 total -= bitmap->map[i].depth;
1435         }
1436
1437         return 0;
1438 }
1439
1440 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1441 {
1442         struct request_queue *q = hctx->queue;
1443         struct blk_mq_ctx *ctx;
1444         LIST_HEAD(tmp);
1445
1446         /*
1447          * Move ctx entries to new CPU, if this one is going away.
1448          */
1449         ctx = __blk_mq_get_ctx(q, cpu);
1450
1451         spin_lock(&ctx->lock);
1452         if (!list_empty(&ctx->rq_list)) {
1453                 list_splice_init(&ctx->rq_list, &tmp);
1454                 blk_mq_hctx_clear_pending(hctx, ctx);
1455         }
1456         spin_unlock(&ctx->lock);
1457
1458         if (list_empty(&tmp))
1459                 return NOTIFY_OK;
1460
1461         ctx = blk_mq_get_ctx(q);
1462         spin_lock(&ctx->lock);
1463
1464         while (!list_empty(&tmp)) {
1465                 struct request *rq;
1466
1467                 rq = list_first_entry(&tmp, struct request, queuelist);
1468                 rq->mq_ctx = ctx;
1469                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1470         }
1471
1472         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1473         blk_mq_hctx_mark_pending(hctx, ctx);
1474
1475         spin_unlock(&ctx->lock);
1476
1477         blk_mq_run_hw_queue(hctx, true);
1478         blk_mq_put_ctx(ctx);
1479         return NOTIFY_OK;
1480 }
1481
1482 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1483 {
1484         struct request_queue *q = hctx->queue;
1485         struct blk_mq_tag_set *set = q->tag_set;
1486
1487         if (set->tags[hctx->queue_num])
1488                 return NOTIFY_OK;
1489
1490         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1491         if (!set->tags[hctx->queue_num])
1492                 return NOTIFY_STOP;
1493
1494         hctx->tags = set->tags[hctx->queue_num];
1495         return NOTIFY_OK;
1496 }
1497
1498 static int blk_mq_hctx_notify(void *data, unsigned long action,
1499                               unsigned int cpu)
1500 {
1501         struct blk_mq_hw_ctx *hctx = data;
1502
1503         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1504                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1505         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1506                 return blk_mq_hctx_cpu_online(hctx, cpu);
1507
1508         return NOTIFY_OK;
1509 }
1510
1511 static void blk_mq_exit_hw_queues(struct request_queue *q,
1512                 struct blk_mq_tag_set *set, int nr_queue)
1513 {
1514         struct blk_mq_hw_ctx *hctx;
1515         unsigned int i;
1516
1517         queue_for_each_hw_ctx(q, hctx, i) {
1518                 if (i == nr_queue)
1519                         break;
1520
1521                 blk_mq_tag_idle(hctx);
1522
1523                 if (set->ops->exit_hctx)
1524                         set->ops->exit_hctx(hctx, i);
1525
1526                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1527                 kfree(hctx->ctxs);
1528                 blk_mq_free_bitmap(&hctx->ctx_map);
1529         }
1530
1531 }
1532
1533 static void blk_mq_free_hw_queues(struct request_queue *q,
1534                 struct blk_mq_tag_set *set)
1535 {
1536         struct blk_mq_hw_ctx *hctx;
1537         unsigned int i;
1538
1539         queue_for_each_hw_ctx(q, hctx, i) {
1540                 free_cpumask_var(hctx->cpumask);
1541                 kfree(hctx);
1542         }
1543 }
1544
1545 static int blk_mq_init_hw_queues(struct request_queue *q,
1546                 struct blk_mq_tag_set *set)
1547 {
1548         struct blk_mq_hw_ctx *hctx;
1549         unsigned int i;
1550
1551         /*
1552          * Initialize hardware queues
1553          */
1554         queue_for_each_hw_ctx(q, hctx, i) {
1555                 int node;
1556
1557                 node = hctx->numa_node;
1558                 if (node == NUMA_NO_NODE)
1559                         node = hctx->numa_node = set->numa_node;
1560
1561                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1562                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1563                 spin_lock_init(&hctx->lock);
1564                 INIT_LIST_HEAD(&hctx->dispatch);
1565                 hctx->queue = q;
1566                 hctx->queue_num = i;
1567                 hctx->flags = set->flags;
1568                 hctx->cmd_size = set->cmd_size;
1569
1570                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1571                                                 blk_mq_hctx_notify, hctx);
1572                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1573
1574                 hctx->tags = set->tags[i];
1575
1576                 /*
1577                  * Allocate space for all possible cpus to avoid allocation in
1578                  * runtime
1579                  */
1580                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1581                                                 GFP_KERNEL, node);
1582                 if (!hctx->ctxs)
1583                         break;
1584
1585                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1586                         break;
1587
1588                 hctx->nr_ctx = 0;
1589
1590                 if (set->ops->init_hctx &&
1591                     set->ops->init_hctx(hctx, set->driver_data, i))
1592                         break;
1593         }
1594
1595         if (i == q->nr_hw_queues)
1596                 return 0;
1597
1598         /*
1599          * Init failed
1600          */
1601         blk_mq_exit_hw_queues(q, set, i);
1602
1603         return 1;
1604 }
1605
1606 static void blk_mq_init_cpu_queues(struct request_queue *q,
1607                                    unsigned int nr_hw_queues)
1608 {
1609         unsigned int i;
1610
1611         for_each_possible_cpu(i) {
1612                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1613                 struct blk_mq_hw_ctx *hctx;
1614
1615                 memset(__ctx, 0, sizeof(*__ctx));
1616                 __ctx->cpu = i;
1617                 spin_lock_init(&__ctx->lock);
1618                 INIT_LIST_HEAD(&__ctx->rq_list);
1619                 __ctx->queue = q;
1620
1621                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1622                 if (!cpu_online(i))
1623                         continue;
1624
1625                 hctx = q->mq_ops->map_queue(q, i);
1626                 cpumask_set_cpu(i, hctx->cpumask);
1627                 hctx->nr_ctx++;
1628
1629                 /*
1630                  * Set local node, IFF we have more than one hw queue. If
1631                  * not, we remain on the home node of the device
1632                  */
1633                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1634                         hctx->numa_node = cpu_to_node(i);
1635         }
1636 }
1637
1638 static void blk_mq_map_swqueue(struct request_queue *q)
1639 {
1640         unsigned int i;
1641         struct blk_mq_hw_ctx *hctx;
1642         struct blk_mq_ctx *ctx;
1643
1644         queue_for_each_hw_ctx(q, hctx, i) {
1645                 cpumask_clear(hctx->cpumask);
1646                 hctx->nr_ctx = 0;
1647         }
1648
1649         /*
1650          * Map software to hardware queues
1651          */
1652         queue_for_each_ctx(q, ctx, i) {
1653                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1654                 if (!cpu_online(i))
1655                         continue;
1656
1657                 hctx = q->mq_ops->map_queue(q, i);
1658                 cpumask_set_cpu(i, hctx->cpumask);
1659                 ctx->index_hw = hctx->nr_ctx;
1660                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1661         }
1662
1663         queue_for_each_hw_ctx(q, hctx, i) {
1664                 /*
1665                  * If not software queues are mapped to this hardware queue,
1666                  * disable it and free the request entries
1667                  */
1668                 if (!hctx->nr_ctx) {
1669                         struct blk_mq_tag_set *set = q->tag_set;
1670
1671                         if (set->tags[i]) {
1672                                 blk_mq_free_rq_map(set, set->tags[i], i);
1673                                 set->tags[i] = NULL;
1674                                 hctx->tags = NULL;
1675                         }
1676                         continue;
1677                 }
1678
1679                 /*
1680                  * Initialize batch roundrobin counts
1681                  */
1682                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1683                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1684         }
1685 }
1686
1687 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1688 {
1689         struct blk_mq_hw_ctx *hctx;
1690         struct request_queue *q;
1691         bool shared;
1692         int i;
1693
1694         if (set->tag_list.next == set->tag_list.prev)
1695                 shared = false;
1696         else
1697                 shared = true;
1698
1699         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1700                 blk_mq_freeze_queue(q);
1701
1702                 queue_for_each_hw_ctx(q, hctx, i) {
1703                         if (shared)
1704                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1705                         else
1706                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1707                 }
1708                 blk_mq_unfreeze_queue(q);
1709         }
1710 }
1711
1712 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1713 {
1714         struct blk_mq_tag_set *set = q->tag_set;
1715
1716         blk_mq_freeze_queue(q);
1717
1718         mutex_lock(&set->tag_list_lock);
1719         list_del_init(&q->tag_set_list);
1720         blk_mq_update_tag_set_depth(set);
1721         mutex_unlock(&set->tag_list_lock);
1722
1723         blk_mq_unfreeze_queue(q);
1724 }
1725
1726 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1727                                      struct request_queue *q)
1728 {
1729         q->tag_set = set;
1730
1731         mutex_lock(&set->tag_list_lock);
1732         list_add_tail(&q->tag_set_list, &set->tag_list);
1733         blk_mq_update_tag_set_depth(set);
1734         mutex_unlock(&set->tag_list_lock);
1735 }
1736
1737 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1738 {
1739         struct blk_mq_hw_ctx **hctxs;
1740         struct blk_mq_ctx __percpu *ctx;
1741         struct request_queue *q;
1742         unsigned int *map;
1743         int i;
1744
1745         ctx = alloc_percpu(struct blk_mq_ctx);
1746         if (!ctx)
1747                 return ERR_PTR(-ENOMEM);
1748
1749         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1750                         set->numa_node);
1751
1752         if (!hctxs)
1753                 goto err_percpu;
1754
1755         map = blk_mq_make_queue_map(set);
1756         if (!map)
1757                 goto err_map;
1758
1759         for (i = 0; i < set->nr_hw_queues; i++) {
1760                 int node = blk_mq_hw_queue_to_node(map, i);
1761
1762                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1763                                         GFP_KERNEL, node);
1764                 if (!hctxs[i])
1765                         goto err_hctxs;
1766
1767                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1768                         goto err_hctxs;
1769
1770                 atomic_set(&hctxs[i]->nr_active, 0);
1771                 hctxs[i]->numa_node = node;
1772                 hctxs[i]->queue_num = i;
1773         }
1774
1775         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1776         if (!q)
1777                 goto err_hctxs;
1778
1779         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1780                 goto err_map;
1781
1782         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1783         blk_queue_rq_timeout(q, 30000);
1784
1785         q->nr_queues = nr_cpu_ids;
1786         q->nr_hw_queues = set->nr_hw_queues;
1787         q->mq_map = map;
1788
1789         q->queue_ctx = ctx;
1790         q->queue_hw_ctx = hctxs;
1791
1792         q->mq_ops = set->ops;
1793         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1794
1795         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1796                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1797
1798         q->sg_reserved_size = INT_MAX;
1799
1800         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1801         INIT_LIST_HEAD(&q->requeue_list);
1802         spin_lock_init(&q->requeue_lock);
1803
1804         if (q->nr_hw_queues > 1)
1805                 blk_queue_make_request(q, blk_mq_make_request);
1806         else
1807                 blk_queue_make_request(q, blk_sq_make_request);
1808
1809         blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1810         if (set->timeout)
1811                 blk_queue_rq_timeout(q, set->timeout);
1812
1813         /*
1814          * Do this after blk_queue_make_request() overrides it...
1815          */
1816         q->nr_requests = set->queue_depth;
1817
1818         if (set->ops->complete)
1819                 blk_queue_softirq_done(q, set->ops->complete);
1820
1821         blk_mq_init_flush(q);
1822         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1823
1824         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1825                                 set->cmd_size, cache_line_size()),
1826                                 GFP_KERNEL);
1827         if (!q->flush_rq)
1828                 goto err_hw;
1829
1830         if (blk_mq_init_hw_queues(q, set))
1831                 goto err_flush_rq;
1832
1833         mutex_lock(&all_q_mutex);
1834         list_add_tail(&q->all_q_node, &all_q_list);
1835         mutex_unlock(&all_q_mutex);
1836
1837         blk_mq_add_queue_tag_set(set, q);
1838
1839         blk_mq_map_swqueue(q);
1840
1841         return q;
1842
1843 err_flush_rq:
1844         kfree(q->flush_rq);
1845 err_hw:
1846         blk_cleanup_queue(q);
1847 err_hctxs:
1848         kfree(map);
1849         for (i = 0; i < set->nr_hw_queues; i++) {
1850                 if (!hctxs[i])
1851                         break;
1852                 free_cpumask_var(hctxs[i]->cpumask);
1853                 kfree(hctxs[i]);
1854         }
1855 err_map:
1856         kfree(hctxs);
1857 err_percpu:
1858         free_percpu(ctx);
1859         return ERR_PTR(-ENOMEM);
1860 }
1861 EXPORT_SYMBOL(blk_mq_init_queue);
1862
1863 void blk_mq_free_queue(struct request_queue *q)
1864 {
1865         struct blk_mq_tag_set   *set = q->tag_set;
1866
1867         blk_mq_del_queue_tag_set(q);
1868
1869         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1870         blk_mq_free_hw_queues(q, set);
1871
1872         percpu_ref_exit(&q->mq_usage_counter);
1873
1874         free_percpu(q->queue_ctx);
1875         kfree(q->queue_hw_ctx);
1876         kfree(q->mq_map);
1877
1878         q->queue_ctx = NULL;
1879         q->queue_hw_ctx = NULL;
1880         q->mq_map = NULL;
1881
1882         mutex_lock(&all_q_mutex);
1883         list_del_init(&q->all_q_node);
1884         mutex_unlock(&all_q_mutex);
1885 }
1886
1887 /* Basically redo blk_mq_init_queue with queue frozen */
1888 static void blk_mq_queue_reinit(struct request_queue *q)
1889 {
1890         blk_mq_freeze_queue(q);
1891
1892         blk_mq_sysfs_unregister(q);
1893
1894         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1895
1896         /*
1897          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1898          * we should change hctx numa_node according to new topology (this
1899          * involves free and re-allocate memory, worthy doing?)
1900          */
1901
1902         blk_mq_map_swqueue(q);
1903
1904         blk_mq_sysfs_register(q);
1905
1906         blk_mq_unfreeze_queue(q);
1907 }
1908
1909 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1910                                       unsigned long action, void *hcpu)
1911 {
1912         struct request_queue *q;
1913
1914         /*
1915          * Before new mappings are established, hotadded cpu might already
1916          * start handling requests. This doesn't break anything as we map
1917          * offline CPUs to first hardware queue. We will re-init the queue
1918          * below to get optimal settings.
1919          */
1920         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1921             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1922                 return NOTIFY_OK;
1923
1924         mutex_lock(&all_q_mutex);
1925         list_for_each_entry(q, &all_q_list, all_q_node)
1926                 blk_mq_queue_reinit(q);
1927         mutex_unlock(&all_q_mutex);
1928         return NOTIFY_OK;
1929 }
1930
1931 /*
1932  * Alloc a tag set to be associated with one or more request queues.
1933  * May fail with EINVAL for various error conditions. May adjust the
1934  * requested depth down, if if it too large. In that case, the set
1935  * value will be stored in set->queue_depth.
1936  */
1937 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1938 {
1939         int i;
1940
1941         if (!set->nr_hw_queues)
1942                 return -EINVAL;
1943         if (!set->queue_depth)
1944                 return -EINVAL;
1945         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1946                 return -EINVAL;
1947
1948         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1949                 return -EINVAL;
1950
1951         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
1952                 pr_info("blk-mq: reduced tag depth to %u\n",
1953                         BLK_MQ_MAX_DEPTH);
1954                 set->queue_depth = BLK_MQ_MAX_DEPTH;
1955         }
1956
1957         set->tags = kmalloc_node(set->nr_hw_queues *
1958                                  sizeof(struct blk_mq_tags *),
1959                                  GFP_KERNEL, set->numa_node);
1960         if (!set->tags)
1961                 goto out;
1962
1963         for (i = 0; i < set->nr_hw_queues; i++) {
1964                 set->tags[i] = blk_mq_init_rq_map(set, i);
1965                 if (!set->tags[i])
1966                         goto out_unwind;
1967         }
1968
1969         mutex_init(&set->tag_list_lock);
1970         INIT_LIST_HEAD(&set->tag_list);
1971
1972         return 0;
1973
1974 out_unwind:
1975         while (--i >= 0)
1976                 blk_mq_free_rq_map(set, set->tags[i], i);
1977 out:
1978         return -ENOMEM;
1979 }
1980 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
1981
1982 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
1983 {
1984         int i;
1985
1986         for (i = 0; i < set->nr_hw_queues; i++) {
1987                 if (set->tags[i])
1988                         blk_mq_free_rq_map(set, set->tags[i], i);
1989         }
1990
1991         kfree(set->tags);
1992 }
1993 EXPORT_SYMBOL(blk_mq_free_tag_set);
1994
1995 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
1996 {
1997         struct blk_mq_tag_set *set = q->tag_set;
1998         struct blk_mq_hw_ctx *hctx;
1999         int i, ret;
2000
2001         if (!set || nr > set->queue_depth)
2002                 return -EINVAL;
2003
2004         ret = 0;
2005         queue_for_each_hw_ctx(q, hctx, i) {
2006                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2007                 if (ret)
2008                         break;
2009         }
2010
2011         if (!ret)
2012                 q->nr_requests = nr;
2013
2014         return ret;
2015 }
2016
2017 void blk_mq_disable_hotplug(void)
2018 {
2019         mutex_lock(&all_q_mutex);
2020 }
2021
2022 void blk_mq_enable_hotplug(void)
2023 {
2024         mutex_unlock(&all_q_mutex);
2025 }
2026
2027 static int __init blk_mq_init(void)
2028 {
2029         blk_mq_cpu_init();
2030
2031         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2032
2033         return 0;
2034 }
2035 subsys_initcall(blk_mq_init);