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         int ret;
  82
  83         __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
  84         smp_wmb();
  85         /* we have problems to freeze the queue if it's initializing */
  86         if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
  87                 return 0;
  88
  89         __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
  90
  91         spin_lock_irq(q->queue_lock);
  92         ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
  93                 !blk_queue_bypass(q) || blk_queue_dying(q),
  94                 *q->queue_lock);
  95         /* inc usage with lock hold to avoid freeze_queue runs here */
  96         if (!ret && !blk_queue_dying(q))
  97                 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
  98         else if (blk_queue_dying(q))
  99                 ret = -ENODEV;
 100         spin_unlock_irq(q->queue_lock);
 101
 102         return ret;
 103 }
 104
 105 static void blk_mq_queue_exit(struct request_queue *q)
 106 {
 107         __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
 108 }
 109
 110 static void __blk_mq_drain_queue(struct request_queue *q)
 111 {
 112         while (true) {
 113                 s64 count;
 114
 115                 spin_lock_irq(q->queue_lock);
 116                 count = percpu_counter_sum(&q->mq_usage_counter);
 117                 spin_unlock_irq(q->queue_lock);
 118
 119                 if (count == 0)
 120                         break;
 121                 blk_mq_run_queues(q, false);
 122                 msleep(10);
 123         }
 124 }
 125
 126 /*
 127  * Guarantee no request is in use, so we can change any data structure of
 128  * the queue afterward.
 129  */
 130 static void blk_mq_freeze_queue(struct request_queue *q)
 131 {
 132         bool drain;
 133
 134         spin_lock_irq(q->queue_lock);
 135         drain = !q->bypass_depth++;
 136         queue_flag_set(QUEUE_FLAG_BYPASS, q);
 137         spin_unlock_irq(q->queue_lock);
 138
 139         if (drain)
 140                 __blk_mq_drain_queue(q);
 141 }
 142
 143 void blk_mq_drain_queue(struct request_queue *q)
 144 {
 145         __blk_mq_drain_queue(q);
 146 }
 147
 148 static void blk_mq_unfreeze_queue(struct request_queue *q)
 149 {
 150         bool wake = false;
 151
 152         spin_lock_irq(q->queue_lock);
 153         if (!--q->bypass_depth) {
 154                 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
 155                 wake = true;
 156         }
 157         WARN_ON_ONCE(q->bypass_depth < 0);
 158         spin_unlock_irq(q->queue_lock);
 159         if (wake)
 160                 wake_up_all(&q->mq_freeze_wq);
 161 }
 162
 163 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 164 {
 165         return blk_mq_has_free_tags(hctx->tags);
 166 }
 167 EXPORT_SYMBOL(blk_mq_can_queue);
 168
 169 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 170                                struct request *rq, unsigned int rw_flags)
 171 {
 172         if (blk_queue_io_stat(q))
 173                 rw_flags |= REQ_IO_STAT;
 174
 175         INIT_LIST_HEAD(&rq->queuelist);
 176         /* csd/requeue_work/fifo_time is initialized before use */
 177         rq->q = q;
 178         rq->mq_ctx = ctx;
 179         rq->cmd_flags |= rw_flags;
 180         /* do not touch atomic flags, it needs atomic ops against the timer */
 181         rq->cpu = -1;
 182         INIT_HLIST_NODE(&rq->hash);
 183         RB_CLEAR_NODE(&rq->rb_node);
 184         rq->rq_disk = NULL;
 185         rq->part = NULL;
 186 #ifdef CONFIG_BLK_CGROUP
 187         rq->rl = NULL;
 188         set_start_time_ns(rq);
 189         rq->io_start_time_ns = 0;
 190 #endif
 191         rq->nr_phys_segments = 0;
 192 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 193         rq->nr_integrity_segments = 0;
 194 #endif
 195         rq->special = NULL;
 196         /* tag was already set */
 197         rq->errors = 0;
 198
 199         rq->extra_len = 0;
 200         rq->sense_len = 0;
 201         rq->resid_len = 0;
 202         rq->sense = NULL;
 203
 204         INIT_LIST_HEAD(&rq->timeout_list);
 205         rq->end_io = NULL;
 206         rq->end_io_data = NULL;
 207         rq->next_rq = NULL;
 208
 209         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 210 }
 211
 212 static struct request *
 213 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 214 {
 215         struct request *rq;
 216         unsigned int tag;
 217
 218         tag = blk_mq_get_tag(data);
 219         if (tag != BLK_MQ_TAG_FAIL) {
 220                 rq = data->hctx->tags->rqs[tag];
 221
 222                 rq->cmd_flags = 0;
 223                 if (blk_mq_tag_busy(data->hctx)) {
 224                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 225                         atomic_inc(&data->hctx->nr_active);
 226                 }
 227
 228                 rq->tag = tag;
 229                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 230                 return rq;
 231         }
 232
 233         return NULL;
 234 }
 235
 236 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 237                 bool reserved)
 238 {
 239         struct blk_mq_ctx *ctx;
 240         struct blk_mq_hw_ctx *hctx;
 241         struct request *rq;
 242         struct blk_mq_alloc_data alloc_data;
 243
 244         if (blk_mq_queue_enter(q))
 245                 return NULL;
 246
 247         ctx = blk_mq_get_ctx(q);
 248         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 249         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 250                         reserved, ctx, hctx);
 251
 252         rq = __blk_mq_alloc_request(&alloc_data, rw);
 253         if (!rq && (gfp & __GFP_WAIT)) {
 254                 __blk_mq_run_hw_queue(hctx);
 255                 blk_mq_put_ctx(ctx);
 256
 257                 ctx = blk_mq_get_ctx(q);
 258                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 259                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 260                                 hctx);
 261                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 262                 ctx = alloc_data.ctx;
 263         }
 264         blk_mq_put_ctx(ctx);
 265         return rq;
 266 }
 267 EXPORT_SYMBOL(blk_mq_alloc_request);
 268
 269 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 270                                   struct blk_mq_ctx *ctx, struct request *rq)
 271 {
 272         const int tag = rq->tag;
 273         struct request_queue *q = rq->q;
 274
 275         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 276                 atomic_dec(&hctx->nr_active);
 277
 278         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 279         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 280         blk_mq_queue_exit(q);
 281 }
 282
 283 void blk_mq_free_request(struct request *rq)
 284 {
 285         struct blk_mq_ctx *ctx = rq->mq_ctx;
 286         struct blk_mq_hw_ctx *hctx;
 287         struct request_queue *q = rq->q;
 288
 289         ctx->rq_completed[rq_is_sync(rq)]++;
 290
 291         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 292         __blk_mq_free_request(hctx, ctx, rq);
 293 }
 294
 295 /*
 296  * Clone all relevant state from a request that has been put on hold in
 297  * the flush state machine into the preallocated flush request that hangs
 298  * off the request queue.
 299  *
 300  * For a driver the flush request should be invisible, that's why we are
 301  * impersonating the original request here.
 302  */
 303 void blk_mq_clone_flush_request(struct request *flush_rq,
 304                 struct request *orig_rq)
 305 {
 306         struct blk_mq_hw_ctx *hctx =
 307                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 308
 309         flush_rq->mq_ctx = orig_rq->mq_ctx;
 310         flush_rq->tag = orig_rq->tag;
 311         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 312                 hctx->cmd_size);
 313 }
 314
 315 inline void __blk_mq_end_io(struct request *rq, int error)
 316 {
 317         blk_account_io_done(rq);
 318
 319         if (rq->end_io) {
 320                 rq->end_io(rq, error);
 321         } else {
 322                 if (unlikely(blk_bidi_rq(rq)))
 323                         blk_mq_free_request(rq->next_rq);
 324                 blk_mq_free_request(rq);
 325         }
 326 }
 327 EXPORT_SYMBOL(__blk_mq_end_io);
 328
 329 void blk_mq_end_io(struct request *rq, int error)
 330 {
 331         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 332                 BUG();
 333         __blk_mq_end_io(rq, error);
 334 }
 335 EXPORT_SYMBOL(blk_mq_end_io);
 336
 337 static void __blk_mq_complete_request_remote(void *data)
 338 {
 339         struct request *rq = data;
 340
 341         rq->q->softirq_done_fn(rq);
 342 }
 343
 344 static void blk_mq_ipi_complete_request(struct request *rq)
 345 {
 346         struct blk_mq_ctx *ctx = rq->mq_ctx;
 347         bool shared = false;
 348         int cpu;
 349
 350         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 351                 rq->q->softirq_done_fn(rq);
 352                 return;
 353         }
 354
 355         cpu = get_cpu();
 356         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 357                 shared = cpus_share_cache(cpu, ctx->cpu);
 358
 359         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 360                 rq->csd.func = __blk_mq_complete_request_remote;
 361                 rq->csd.info = rq;
 362                 rq->csd.flags = 0;
 363                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 364         } else {
 365                 rq->q->softirq_done_fn(rq);
 366         }
 367         put_cpu();
 368 }
 369
 370 void __blk_mq_complete_request(struct request *rq)
 371 {
 372         struct request_queue *q = rq->q;
 373
 374         if (!q->softirq_done_fn)
 375                 blk_mq_end_io(rq, rq->errors);
 376         else
 377                 blk_mq_ipi_complete_request(rq);
 378 }
 379
 380 /**
 381  * blk_mq_complete_request - end I/O on a request
 382  * @rq:         the request being processed
 383  *
 384  * Description:
 385  *      Ends all I/O on a request. It does not handle partial completions.
 386  *      The actual completion happens out-of-order, through a IPI handler.
 387  **/
 388 void blk_mq_complete_request(struct request *rq)
 389 {
 390         struct request_queue *q = rq->q;
 391
 392         if (unlikely(blk_should_fake_timeout(q)))
 393                 return;
 394         if (!blk_mark_rq_complete(rq))
 395                 __blk_mq_complete_request(rq);
 396 }
 397 EXPORT_SYMBOL(blk_mq_complete_request);
 398
 399 static void blk_mq_start_request(struct request *rq, bool last)
 400 {
 401         struct request_queue *q = rq->q;
 402
 403         trace_block_rq_issue(q, rq);
 404
 405         rq->resid_len = blk_rq_bytes(rq);
 406         if (unlikely(blk_bidi_rq(rq)))
 407                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 408
 409         /*
 410          * Just mark start time and set the started bit. Due to memory
 411          * ordering, we know we'll see the correct deadline as long as
 412          * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
 413          * unless one has been set in the request.
 414          */
 415         if (!rq->timeout)
 416                 rq->deadline = jiffies + q->rq_timeout;
 417         else
 418                 rq->deadline = jiffies + rq->timeout;
 419
 420         /*
 421          * Mark us as started and clear complete. Complete might have been
 422          * set if requeue raced with timeout, which then marked it as
 423          * complete. So be sure to clear complete again when we start
 424          * the request, otherwise we'll ignore the completion event.
 425          */
 426         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 427                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 428         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 429                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 430
 431         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 432                 /*
 433                  * Make sure space for the drain appears.  We know we can do
 434                  * this because max_hw_segments has been adjusted to be one
 435                  * fewer than the device can handle.
 436                  */
 437                 rq->nr_phys_segments++;
 438         }
 439
 440         /*
 441          * Flag the last request in the series so that drivers know when IO
 442          * should be kicked off, if they don't do it on a per-request basis.
 443          *
 444          * Note: the flag isn't the only condition drivers should do kick off.
 445          * If drive is busy, the last request might not have the bit set.
 446          */
 447         if (last)
 448                 rq->cmd_flags |= REQ_END;
 449 }
 450
 451 static void __blk_mq_requeue_request(struct request *rq)
 452 {
 453         struct request_queue *q = rq->q;
 454
 455         trace_block_rq_requeue(q, rq);
 456         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 457
 458         rq->cmd_flags &= ~REQ_END;
 459
 460         if (q->dma_drain_size && blk_rq_bytes(rq))
 461                 rq->nr_phys_segments--;
 462 }
 463
 464 void blk_mq_requeue_request(struct request *rq)
 465 {
 466         __blk_mq_requeue_request(rq);
 467         blk_clear_rq_complete(rq);
 468
 469         BUG_ON(blk_queued_rq(rq));
 470         blk_mq_add_to_requeue_list(rq, true);
 471 }
 472 EXPORT_SYMBOL(blk_mq_requeue_request);
 473
 474 static void blk_mq_requeue_work(struct work_struct *work)
 475 {
 476         struct request_queue *q =
 477                 container_of(work, struct request_queue, requeue_work);
 478         LIST_HEAD(rq_list);
 479         struct request *rq, *next;
 480         unsigned long flags;
 481
 482         spin_lock_irqsave(&q->requeue_lock, flags);
 483         list_splice_init(&q->requeue_list, &rq_list);
 484         spin_unlock_irqrestore(&q->requeue_lock, flags);
 485
 486         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 487                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 488                         continue;
 489
 490                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 491                 list_del_init(&rq->queuelist);
 492                 blk_mq_insert_request(rq, true, false, false);
 493         }
 494
 495         while (!list_empty(&rq_list)) {
 496                 rq = list_entry(rq_list.next, struct request, queuelist);
 497                 list_del_init(&rq->queuelist);
 498                 blk_mq_insert_request(rq, false, false, false);
 499         }
 500
 501         blk_mq_run_queues(q, false);
 502 }
 503
 504 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 505 {
 506         struct request_queue *q = rq->q;
 507         unsigned long flags;
 508
 509         /*
 510          * We abuse this flag that is otherwise used by the I/O scheduler to
 511          * request head insertation from the workqueue.
 512          */
 513         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 514
 515         spin_lock_irqsave(&q->requeue_lock, flags);
 516         if (at_head) {
 517                 rq->cmd_flags |= REQ_SOFTBARRIER;
 518                 list_add(&rq->queuelist, &q->requeue_list);
 519         } else {
 520                 list_add_tail(&rq->queuelist, &q->requeue_list);
 521         }
 522         spin_unlock_irqrestore(&q->requeue_lock, flags);
 523 }
 524 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 525
 526 void blk_mq_kick_requeue_list(struct request_queue *q)
 527 {
 528         kblockd_schedule_work(&q->requeue_work);
 529 }
 530 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 531
 532 static inline bool is_flush_request(struct request *rq, unsigned int tag)
 533 {
 534         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 535                         rq->q->flush_rq->tag == tag);
 536 }
 537
 538 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 539 {
 540         struct request *rq = tags->rqs[tag];
 541
 542         if (!is_flush_request(rq, tag))
 543                 return rq;
 544
 545         return rq->q->flush_rq;
 546 }
 547 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 548
 549 struct blk_mq_timeout_data {
 550         struct blk_mq_hw_ctx *hctx;
 551         unsigned long *next;
 552         unsigned int *next_set;
 553 };
 554
 555 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
 556 {
 557         struct blk_mq_timeout_data *data = __data;
 558         struct blk_mq_hw_ctx *hctx = data->hctx;
 559         unsigned int tag;
 560
 561          /* It may not be in flight yet (this is where
 562          * the REQ_ATOMIC_STARTED flag comes in). The requests are
 563          * statically allocated, so we know it's always safe to access the
 564          * memory associated with a bit offset into ->rqs[].
 565          */
 566         tag = 0;
 567         do {
 568                 struct request *rq;
 569
 570                 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
 571                 if (tag >= hctx->tags->nr_tags)
 572                         break;
 573
 574                 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
 575                 if (rq->q != hctx->queue)
 576                         continue;
 577                 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 578                         continue;
 579
 580                 blk_rq_check_expired(rq, data->next, data->next_set);
 581         } while (1);
 582 }
 583
 584 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
 585                                         unsigned long *next,
 586                                         unsigned int *next_set)
 587 {
 588         struct blk_mq_timeout_data data = {
 589                 .hctx           = hctx,
 590                 .next           = next,
 591                 .next_set       = next_set,
 592         };
 593
 594         /*
 595          * Ask the tagging code to iterate busy requests, so we can
 596          * check them for timeout.
 597          */
 598         blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
 599 }
 600
 601 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
 602 {
 603         struct request_queue *q = rq->q;
 604
 605         /*
 606          * We know that complete is set at this point. If STARTED isn't set
 607          * anymore, then the request isn't active and the "timeout" should
 608          * just be ignored. This can happen due to the bitflag ordering.
 609          * Timeout first checks if STARTED is set, and if it is, assumes
 610          * the request is active. But if we race with completion, then
 611          * we both flags will get cleared. So check here again, and ignore
 612          * a timeout event with a request that isn't active.
 613          */
 614         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 615                 return BLK_EH_NOT_HANDLED;
 616
 617         if (!q->mq_ops->timeout)
 618                 return BLK_EH_RESET_TIMER;
 619
 620         return q->mq_ops->timeout(rq);
 621 }
 622
 623 static void blk_mq_rq_timer(unsigned long data)
 624 {
 625         struct request_queue *q = (struct request_queue *) data;
 626         struct blk_mq_hw_ctx *hctx;
 627         unsigned long next = 0;
 628         int i, next_set = 0;
 629
 630         queue_for_each_hw_ctx(q, hctx, i) {
 631                 /*
 632                  * If not software queues are currently mapped to this
 633                  * hardware queue, there's nothing to check
 634                  */
 635                 if (!hctx->nr_ctx || !hctx->tags)
 636                         continue;
 637
 638                 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
 639         }
 640
 641         if (next_set) {
 642                 next = blk_rq_timeout(round_jiffies_up(next));
 643                 mod_timer(&q->timeout, next);
 644         } else {
 645                 queue_for_each_hw_ctx(q, hctx, i)
 646                         blk_mq_tag_idle(hctx);
 647         }
 648 }
 649
 650 /*
 651  * Reverse check our software queue for entries that we could potentially
 652  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 653  * too much time checking for merges.
 654  */
 655 static bool blk_mq_attempt_merge(struct request_queue *q,
 656                                  struct blk_mq_ctx *ctx, struct bio *bio)
 657 {
 658         struct request *rq;
 659         int checked = 8;
 660
 661         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 662                 int el_ret;
 663
 664                 if (!checked--)
 665                         break;
 666
 667                 if (!blk_rq_merge_ok(rq, bio))
 668                         continue;
 669
 670                 el_ret = blk_try_merge(rq, bio);
 671                 if (el_ret == ELEVATOR_BACK_MERGE) {
 672                         if (bio_attempt_back_merge(q, rq, bio)) {
 673                                 ctx->rq_merged++;
 674                                 return true;
 675                         }
 676                         break;
 677                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 678                         if (bio_attempt_front_merge(q, rq, bio)) {
 679                                 ctx->rq_merged++;
 680                                 return true;
 681                         }
 682                         break;
 683                 }
 684         }
 685
 686         return false;
 687 }
 688
 689 /*
 690  * Process software queues that have been marked busy, splicing them
 691  * to the for-dispatch
 692  */
 693 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 694 {
 695         struct blk_mq_ctx *ctx;
 696         int i;
 697
 698         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 699                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 700                 unsigned int off, bit;
 701
 702                 if (!bm->word)
 703                         continue;
 704
 705                 bit = 0;
 706                 off = i * hctx->ctx_map.bits_per_word;
 707                 do {
 708                         bit = find_next_bit(&bm->word, bm->depth, bit);
 709                         if (bit >= bm->depth)
 710                                 break;
 711
 712                         ctx = hctx->ctxs[bit + off];
 713                         clear_bit(bit, &bm->word);
 714                         spin_lock(&ctx->lock);
 715                         list_splice_tail_init(&ctx->rq_list, list);
 716                         spin_unlock(&ctx->lock);
 717
 718                         bit++;
 719                 } while (1);
 720         }
 721 }
 722
 723 /*
 724  * Run this hardware queue, pulling any software queues mapped to it in.
 725  * Note that this function currently has various problems around ordering
 726  * of IO. In particular, we'd like FIFO behaviour on handling existing
 727  * items on the hctx->dispatch list. Ignore that for now.
 728  */
 729 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 730 {
 731         struct request_queue *q = hctx->queue;
 732         struct request *rq;
 733         LIST_HEAD(rq_list);
 734         int queued;
 735
 736         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 737
 738         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 739                 return;
 740
 741         hctx->run++;
 742
 743         /*
 744          * Touch any software queue that has pending entries.
 745          */
 746         flush_busy_ctxs(hctx, &rq_list);
 747
 748         /*
 749          * If we have previous entries on our dispatch list, grab them
 750          * and stuff them at the front for more fair dispatch.
 751          */
 752         if (!list_empty_careful(&hctx->dispatch)) {
 753                 spin_lock(&hctx->lock);
 754                 if (!list_empty(&hctx->dispatch))
 755                         list_splice_init(&hctx->dispatch, &rq_list);
 756                 spin_unlock(&hctx->lock);
 757         }
 758
 759         /*
 760          * Now process all the entries, sending them to the driver.
 761          */
 762         queued = 0;
 763         while (!list_empty(&rq_list)) {
 764                 int ret;
 765
 766                 rq = list_first_entry(&rq_list, struct request, queuelist);
 767                 list_del_init(&rq->queuelist);
 768
 769                 blk_mq_start_request(rq, list_empty(&rq_list));
 770
 771                 ret = q->mq_ops->queue_rq(hctx, rq);
 772                 switch (ret) {
 773                 case BLK_MQ_RQ_QUEUE_OK:
 774                         queued++;
 775                         continue;
 776                 case BLK_MQ_RQ_QUEUE_BUSY:
 777                         list_add(&rq->queuelist, &rq_list);
 778                         __blk_mq_requeue_request(rq);
 779                         break;
 780                 default:
 781                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 782                 case BLK_MQ_RQ_QUEUE_ERROR:
 783                         rq->errors = -EIO;
 784                         blk_mq_end_io(rq, rq->errors);
 785                         break;
 786                 }
 787
 788                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 789                         break;
 790         }
 791
 792         if (!queued)
 793                 hctx->dispatched[0]++;
 794         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 795                 hctx->dispatched[ilog2(queued) + 1]++;
 796
 797         /*
 798          * Any items that need requeuing? Stuff them into hctx->dispatch,
 799          * that is where we will continue on next queue run.
 800          */
 801         if (!list_empty(&rq_list)) {
 802                 spin_lock(&hctx->lock);
 803                 list_splice(&rq_list, &hctx->dispatch);
 804                 spin_unlock(&hctx->lock);
 805         }
 806 }
 807
 808 /*
 809  * It'd be great if the workqueue API had a way to pass
 810  * in a mask and had some smarts for more clever placement.
 811  * For now we just round-robin here, switching for every
 812  * BLK_MQ_CPU_WORK_BATCH queued items.
 813  */
 814 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 815 {
 816         int cpu = hctx->next_cpu;
 817
 818         if (--hctx->next_cpu_batch <= 0) {
 819                 int next_cpu;
 820
 821                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 822                 if (next_cpu >= nr_cpu_ids)
 823                         next_cpu = cpumask_first(hctx->cpumask);
 824
 825                 hctx->next_cpu = next_cpu;
 826                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 827         }
 828
 829         return cpu;
 830 }
 831
 832 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 833 {
 834         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 835                 return;
 836
 837         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 838                 __blk_mq_run_hw_queue(hctx);
 839         else if (hctx->queue->nr_hw_queues == 1)
 840                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 841         else {
 842                 unsigned int cpu;
 843
 844                 cpu = blk_mq_hctx_next_cpu(hctx);
 845                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 846         }
 847 }
 848
 849 void blk_mq_run_queues(struct request_queue *q, bool async)
 850 {
 851         struct blk_mq_hw_ctx *hctx;
 852         int i;
 853
 854         queue_for_each_hw_ctx(q, hctx, i) {
 855                 if ((!blk_mq_hctx_has_pending(hctx) &&
 856                     list_empty_careful(&hctx->dispatch)) ||
 857                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 858                         continue;
 859
 860                 preempt_disable();
 861                 blk_mq_run_hw_queue(hctx, async);
 862                 preempt_enable();
 863         }
 864 }
 865 EXPORT_SYMBOL(blk_mq_run_queues);
 866
 867 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 868 {
 869         cancel_delayed_work(&hctx->run_work);
 870         cancel_delayed_work(&hctx->delay_work);
 871         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 872 }
 873 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 874
 875 void blk_mq_stop_hw_queues(struct request_queue *q)
 876 {
 877         struct blk_mq_hw_ctx *hctx;
 878         int i;
 879
 880         queue_for_each_hw_ctx(q, hctx, i)
 881                 blk_mq_stop_hw_queue(hctx);
 882 }
 883 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 884
 885 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 886 {
 887         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 888
 889         preempt_disable();
 890         __blk_mq_run_hw_queue(hctx);
 891         preempt_enable();
 892 }
 893 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 894
 895 void blk_mq_start_hw_queues(struct request_queue *q)
 896 {
 897         struct blk_mq_hw_ctx *hctx;
 898         int i;
 899
 900         queue_for_each_hw_ctx(q, hctx, i)
 901                 blk_mq_start_hw_queue(hctx);
 902 }
 903 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 904
 905
 906 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 907 {
 908         struct blk_mq_hw_ctx *hctx;
 909         int i;
 910
 911         queue_for_each_hw_ctx(q, hctx, i) {
 912                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 913                         continue;
 914
 915                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 916                 preempt_disable();
 917                 blk_mq_run_hw_queue(hctx, async);
 918                 preempt_enable();
 919         }
 920 }
 921 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 922
 923 static void blk_mq_run_work_fn(struct work_struct *work)
 924 {
 925         struct blk_mq_hw_ctx *hctx;
 926
 927         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 928
 929         __blk_mq_run_hw_queue(hctx);
 930 }
 931
 932 static void blk_mq_delay_work_fn(struct work_struct *work)
 933 {
 934         struct blk_mq_hw_ctx *hctx;
 935
 936         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 937
 938         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 939                 __blk_mq_run_hw_queue(hctx);
 940 }
 941
 942 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 943 {
 944         unsigned long tmo = msecs_to_jiffies(msecs);
 945
 946         if (hctx->queue->nr_hw_queues == 1)
 947                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 948         else {
 949                 unsigned int cpu;
 950
 951                 cpu = blk_mq_hctx_next_cpu(hctx);
 952                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 953         }
 954 }
 955 EXPORT_SYMBOL(blk_mq_delay_queue);
 956
 957 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 958                                     struct request *rq, bool at_head)
 959 {
 960         struct blk_mq_ctx *ctx = rq->mq_ctx;
 961
 962         trace_block_rq_insert(hctx->queue, rq);
 963
 964         if (at_head)
 965                 list_add(&rq->queuelist, &ctx->rq_list);
 966         else
 967                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 968
 969         blk_mq_hctx_mark_pending(hctx, ctx);
 970
 971         /*
 972          * We do this early, to ensure we are on the right CPU.
 973          */
 974         blk_add_timer(rq);
 975 }
 976
 977 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 978                 bool async)
 979 {
 980         struct request_queue *q = rq->q;
 981         struct blk_mq_hw_ctx *hctx;
 982         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 983
 984         current_ctx = blk_mq_get_ctx(q);
 985         if (!cpu_online(ctx->cpu))
 986                 rq->mq_ctx = ctx = current_ctx;
 987
 988         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 989
 990         if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
 991             !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
 992                 blk_insert_flush(rq);
 993         } else {
 994                 spin_lock(&ctx->lock);
 995                 __blk_mq_insert_request(hctx, rq, at_head);
 996                 spin_unlock(&ctx->lock);
 997         }
 998
 999         if (run_queue)
1000                 blk_mq_run_hw_queue(hctx, async);
1001
1002         blk_mq_put_ctx(current_ctx);
1003 }
1004
1005 static void blk_mq_insert_requests(struct request_queue *q,
1006                                      struct blk_mq_ctx *ctx,
1007                                      struct list_head *list,
1008                                      int depth,
1009                                      bool from_schedule)
1010
1011 {
1012         struct blk_mq_hw_ctx *hctx;
1013         struct blk_mq_ctx *current_ctx;
1014
1015         trace_block_unplug(q, depth, !from_schedule);
1016
1017         current_ctx = blk_mq_get_ctx(q);
1018
1019         if (!cpu_online(ctx->cpu))
1020                 ctx = current_ctx;
1021         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1022
1023         /*
1024          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1025          * offline now
1026          */
1027         spin_lock(&ctx->lock);
1028         while (!list_empty(list)) {
1029                 struct request *rq;
1030
1031                 rq = list_first_entry(list, struct request, queuelist);
1032                 list_del_init(&rq->queuelist);
1033                 rq->mq_ctx = ctx;
1034                 __blk_mq_insert_request(hctx, rq, false);
1035         }
1036         spin_unlock(&ctx->lock);
1037
1038         blk_mq_run_hw_queue(hctx, from_schedule);
1039         blk_mq_put_ctx(current_ctx);
1040 }
1041
1042 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1043 {
1044         struct request *rqa = container_of(a, struct request, queuelist);
1045         struct request *rqb = container_of(b, struct request, queuelist);
1046
1047         return !(rqa->mq_ctx < rqb->mq_ctx ||
1048                  (rqa->mq_ctx == rqb->mq_ctx &&
1049                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1050 }
1051
1052 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1053 {
1054         struct blk_mq_ctx *this_ctx;
1055         struct request_queue *this_q;
1056         struct request *rq;
1057         LIST_HEAD(list);
1058         LIST_HEAD(ctx_list);
1059         unsigned int depth;
1060
1061         list_splice_init(&plug->mq_list, &list);
1062
1063         list_sort(NULL, &list, plug_ctx_cmp);
1064
1065         this_q = NULL;
1066         this_ctx = NULL;
1067         depth = 0;
1068
1069         while (!list_empty(&list)) {
1070                 rq = list_entry_rq(list.next);
1071                 list_del_init(&rq->queuelist);
1072                 BUG_ON(!rq->q);
1073                 if (rq->mq_ctx != this_ctx) {
1074                         if (this_ctx) {
1075                                 blk_mq_insert_requests(this_q, this_ctx,
1076                                                         &ctx_list, depth,
1077                                                         from_schedule);
1078                         }
1079
1080                         this_ctx = rq->mq_ctx;
1081                         this_q = rq->q;
1082                         depth = 0;
1083                 }
1084
1085                 depth++;
1086                 list_add_tail(&rq->queuelist, &ctx_list);
1087         }
1088
1089         /*
1090          * If 'this_ctx' is set, we know we have entries to complete
1091          * on 'ctx_list'. Do those.
1092          */
1093         if (this_ctx) {
1094                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1095                                        from_schedule);
1096         }
1097 }
1098
1099 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1100 {
1101         init_request_from_bio(rq, bio);
1102
1103         if (blk_do_io_stat(rq)) {
1104                 rq->start_time = jiffies;
1105                 blk_account_io_start(rq, 1);
1106         }
1107 }
1108
1109 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1110                                          struct blk_mq_ctx *ctx,
1111                                          struct request *rq, struct bio *bio)
1112 {
1113         struct request_queue *q = hctx->queue;
1114
1115         if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1116                 blk_mq_bio_to_request(rq, bio);
1117                 spin_lock(&ctx->lock);
1118 insert_rq:
1119                 __blk_mq_insert_request(hctx, rq, false);
1120                 spin_unlock(&ctx->lock);
1121                 return false;
1122         } else {
1123                 spin_lock(&ctx->lock);
1124                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1125                         blk_mq_bio_to_request(rq, bio);
1126                         goto insert_rq;
1127                 }
1128
1129                 spin_unlock(&ctx->lock);
1130                 __blk_mq_free_request(hctx, ctx, rq);
1131                 return true;
1132         }
1133 }
1134
1135 struct blk_map_ctx {
1136         struct blk_mq_hw_ctx *hctx;
1137         struct blk_mq_ctx *ctx;
1138 };
1139
1140 static struct request *blk_mq_map_request(struct request_queue *q,
1141                                           struct bio *bio,
1142                                           struct blk_map_ctx *data)
1143 {
1144         struct blk_mq_hw_ctx *hctx;
1145         struct blk_mq_ctx *ctx;
1146         struct request *rq;
1147         int rw = bio_data_dir(bio);
1148         struct blk_mq_alloc_data alloc_data;
1149
1150         if (unlikely(blk_mq_queue_enter(q))) {
1151                 bio_endio(bio, -EIO);
1152                 return NULL;
1153         }
1154
1155         ctx = blk_mq_get_ctx(q);
1156         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1157
1158         if (rw_is_sync(bio->bi_rw))
1159                 rw |= REQ_SYNC;
1160
1161         trace_block_getrq(q, bio, rw);
1162         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1163                         hctx);
1164         rq = __blk_mq_alloc_request(&alloc_data, rw);
1165         if (unlikely(!rq)) {
1166                 __blk_mq_run_hw_queue(hctx);
1167                 blk_mq_put_ctx(ctx);
1168                 trace_block_sleeprq(q, bio, rw);
1169
1170                 ctx = blk_mq_get_ctx(q);
1171                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1172                 blk_mq_set_alloc_data(&alloc_data, q,
1173                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1174                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1175                 ctx = alloc_data.ctx;
1176                 hctx = alloc_data.hctx;
1177         }
1178
1179         hctx->queued++;
1180         data->hctx = hctx;
1181         data->ctx = ctx;
1182         return rq;
1183 }
1184
1185 /*
1186  * Multiple hardware queue variant. This will not use per-process plugs,
1187  * but will attempt to bypass the hctx queueing if we can go straight to
1188  * hardware for SYNC IO.
1189  */
1190 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1191 {
1192         const int is_sync = rw_is_sync(bio->bi_rw);
1193         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1194         struct blk_map_ctx data;
1195         struct request *rq;
1196
1197         blk_queue_bounce(q, &bio);
1198
1199         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1200                 bio_endio(bio, -EIO);
1201                 return;
1202         }
1203
1204         rq = blk_mq_map_request(q, bio, &data);
1205         if (unlikely(!rq))
1206                 return;
1207
1208         if (unlikely(is_flush_fua)) {
1209                 blk_mq_bio_to_request(rq, bio);
1210                 blk_insert_flush(rq);
1211                 goto run_queue;
1212         }
1213
1214         if (is_sync) {
1215                 int ret;
1216
1217                 blk_mq_bio_to_request(rq, bio);
1218                 blk_mq_start_request(rq, true);
1219                 blk_add_timer(rq);
1220
1221                 /*
1222                  * For OK queue, we are done. For error, kill it. Any other
1223                  * error (busy), just add it to our list as we previously
1224                  * would have done
1225                  */
1226                 ret = q->mq_ops->queue_rq(data.hctx, rq);
1227                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1228                         goto done;
1229                 else {
1230                         __blk_mq_requeue_request(rq);
1231
1232                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1233                                 rq->errors = -EIO;
1234                                 blk_mq_end_io(rq, rq->errors);
1235                                 goto done;
1236                         }
1237                 }
1238         }
1239
1240         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1241                 /*
1242                  * For a SYNC request, send it to the hardware immediately. For
1243                  * an ASYNC request, just ensure that we run it later on. The
1244                  * latter allows for merging opportunities and more efficient
1245                  * dispatching.
1246                  */
1247 run_queue:
1248                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1249         }
1250 done:
1251         blk_mq_put_ctx(data.ctx);
1252 }
1253
1254 /*
1255  * Single hardware queue variant. This will attempt to use any per-process
1256  * plug for merging and IO deferral.
1257  */
1258 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1259 {
1260         const int is_sync = rw_is_sync(bio->bi_rw);
1261         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1262         unsigned int use_plug, request_count = 0;
1263         struct blk_map_ctx data;
1264         struct request *rq;
1265
1266         /*
1267          * If we have multiple hardware queues, just go directly to
1268          * one of those for sync IO.
1269          */
1270         use_plug = !is_flush_fua && !is_sync;
1271
1272         blk_queue_bounce(q, &bio);
1273
1274         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1275                 bio_endio(bio, -EIO);
1276                 return;
1277         }
1278
1279         if (use_plug && !blk_queue_nomerges(q) &&
1280             blk_attempt_plug_merge(q, bio, &request_count))
1281                 return;
1282
1283         rq = blk_mq_map_request(q, bio, &data);
1284         if (unlikely(!rq))
1285                 return;
1286
1287         if (unlikely(is_flush_fua)) {
1288                 blk_mq_bio_to_request(rq, bio);
1289                 blk_insert_flush(rq);
1290                 goto run_queue;
1291         }
1292
1293         /*
1294          * A task plug currently exists. Since this is completely lockless,
1295          * utilize that to temporarily store requests until the task is
1296          * either done or scheduled away.
1297          */
1298         if (use_plug) {
1299                 struct blk_plug *plug = current->plug;
1300
1301                 if (plug) {
1302                         blk_mq_bio_to_request(rq, bio);
1303                         if (list_empty(&plug->mq_list))
1304                                 trace_block_plug(q);
1305                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1306                                 blk_flush_plug_list(plug, false);
1307                                 trace_block_plug(q);
1308                         }
1309                         list_add_tail(&rq->queuelist, &plug->mq_list);
1310                         blk_mq_put_ctx(data.ctx);
1311                         return;
1312                 }
1313         }
1314
1315         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1316                 /*
1317                  * For a SYNC request, send it to the hardware immediately. For
1318                  * an ASYNC request, just ensure that we run it later on. The
1319                  * latter allows for merging opportunities and more efficient
1320                  * dispatching.
1321                  */
1322 run_queue:
1323                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1324         }
1325
1326         blk_mq_put_ctx(data.ctx);
1327 }
1328
1329 /*
1330  * Default mapping to a software queue, since we use one per CPU.
1331  */
1332 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1333 {
1334         return q->queue_hw_ctx[q->mq_map[cpu]];
1335 }
1336 EXPORT_SYMBOL(blk_mq_map_queue);
1337
1338 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1339                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1340 {
1341         struct page *page;
1342
1343         if (tags->rqs && set->ops->exit_request) {
1344                 int i;
1345
1346                 for (i = 0; i < tags->nr_tags; i++) {
1347                         if (!tags->rqs[i])
1348                                 continue;
1349                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1350                                                 hctx_idx, i);
1351                 }
1352         }
1353
1354         while (!list_empty(&tags->page_list)) {
1355                 page = list_first_entry(&tags->page_list, struct page, lru);
1356                 list_del_init(&page->lru);
1357                 __free_pages(page, page->private);
1358         }
1359
1360         kfree(tags->rqs);
1361
1362         blk_mq_free_tags(tags);
1363 }
1364
1365 static size_t order_to_size(unsigned int order)
1366 {
1367         return (size_t)PAGE_SIZE << order;
1368 }
1369
1370 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1371                 unsigned int hctx_idx)
1372 {
1373         struct blk_mq_tags *tags;
1374         unsigned int i, j, entries_per_page, max_order = 4;
1375         size_t rq_size, left;
1376
1377         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1378                                 set->numa_node);
1379         if (!tags)
1380                 return NULL;
1381
1382         INIT_LIST_HEAD(&tags->page_list);
1383
1384         tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1385                                         GFP_KERNEL, set->numa_node);
1386         if (!tags->rqs) {
1387                 blk_mq_free_tags(tags);
1388                 return NULL;
1389         }
1390
1391         /*
1392          * rq_size is the size of the request plus driver payload, rounded
1393          * to the cacheline size
1394          */
1395         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1396                                 cache_line_size());
1397         left = rq_size * set->queue_depth;
1398
1399         for (i = 0; i < set->queue_depth; ) {
1400                 int this_order = max_order;
1401                 struct page *page;
1402                 int to_do;
1403                 void *p;
1404
1405                 while (left < order_to_size(this_order - 1) && this_order)
1406                         this_order--;
1407
1408                 do {
1409                         page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1410                                                 this_order);
1411                         if (page)
1412                                 break;
1413                         if (!this_order--)
1414                                 break;
1415                         if (order_to_size(this_order) < rq_size)
1416                                 break;
1417                 } while (1);
1418
1419                 if (!page)
1420                         goto fail;
1421
1422                 page->private = this_order;
1423                 list_add_tail(&page->lru, &tags->page_list);
1424
1425                 p = page_address(page);
1426                 entries_per_page = order_to_size(this_order) / rq_size;
1427                 to_do = min(entries_per_page, set->queue_depth - i);
1428                 left -= to_do * rq_size;
1429                 for (j = 0; j < to_do; j++) {
1430                         tags->rqs[i] = p;
1431                         if (set->ops->init_request) {
1432                                 if (set->ops->init_request(set->driver_data,
1433                                                 tags->rqs[i], hctx_idx, i,
1434                                                 set->numa_node))
1435                                         goto fail;
1436                         }
1437
1438                         p += rq_size;
1439                         i++;
1440                 }
1441         }
1442
1443         return tags;
1444
1445 fail:
1446         pr_warn("%s: failed to allocate requests\n", __func__);
1447         blk_mq_free_rq_map(set, tags, hctx_idx);
1448         return NULL;
1449 }
1450
1451 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1452 {
1453         kfree(bitmap->map);
1454 }
1455
1456 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1457 {
1458         unsigned int bpw = 8, total, num_maps, i;
1459
1460         bitmap->bits_per_word = bpw;
1461
1462         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1463         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1464                                         GFP_KERNEL, node);
1465         if (!bitmap->map)
1466                 return -ENOMEM;
1467
1468         bitmap->map_size = num_maps;
1469
1470         total = nr_cpu_ids;
1471         for (i = 0; i < num_maps; i++) {
1472                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1473                 total -= bitmap->map[i].depth;
1474         }
1475
1476         return 0;
1477 }
1478
1479 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1480 {
1481         struct request_queue *q = hctx->queue;
1482         struct blk_mq_ctx *ctx;
1483         LIST_HEAD(tmp);
1484
1485         /*
1486          * Move ctx entries to new CPU, if this one is going away.
1487          */
1488         ctx = __blk_mq_get_ctx(q, cpu);
1489
1490         spin_lock(&ctx->lock);
1491         if (!list_empty(&ctx->rq_list)) {
1492                 list_splice_init(&ctx->rq_list, &tmp);
1493                 blk_mq_hctx_clear_pending(hctx, ctx);
1494         }
1495         spin_unlock(&ctx->lock);
1496
1497         if (list_empty(&tmp))
1498                 return NOTIFY_OK;
1499
1500         ctx = blk_mq_get_ctx(q);
1501         spin_lock(&ctx->lock);
1502
1503         while (!list_empty(&tmp)) {
1504                 struct request *rq;
1505
1506                 rq = list_first_entry(&tmp, struct request, queuelist);
1507                 rq->mq_ctx = ctx;
1508                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1509         }
1510
1511         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1512         blk_mq_hctx_mark_pending(hctx, ctx);
1513
1514         spin_unlock(&ctx->lock);
1515
1516         blk_mq_run_hw_queue(hctx, true);
1517         blk_mq_put_ctx(ctx);
1518         return NOTIFY_OK;
1519 }
1520
1521 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1522 {
1523         struct request_queue *q = hctx->queue;
1524         struct blk_mq_tag_set *set = q->tag_set;
1525
1526         if (set->tags[hctx->queue_num])
1527                 return NOTIFY_OK;
1528
1529         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1530         if (!set->tags[hctx->queue_num])
1531                 return NOTIFY_STOP;
1532
1533         hctx->tags = set->tags[hctx->queue_num];
1534         return NOTIFY_OK;
1535 }
1536
1537 static int blk_mq_hctx_notify(void *data, unsigned long action,
1538                               unsigned int cpu)
1539 {
1540         struct blk_mq_hw_ctx *hctx = data;
1541
1542         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1543                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1544         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1545                 return blk_mq_hctx_cpu_online(hctx, cpu);
1546
1547         return NOTIFY_OK;
1548 }
1549
1550 static void blk_mq_exit_hw_queues(struct request_queue *q,
1551                 struct blk_mq_tag_set *set, int nr_queue)
1552 {
1553         struct blk_mq_hw_ctx *hctx;
1554         unsigned int i;
1555
1556         queue_for_each_hw_ctx(q, hctx, i) {
1557                 if (i == nr_queue)
1558                         break;
1559
1560                 blk_mq_tag_idle(hctx);
1561
1562                 if (set->ops->exit_hctx)
1563                         set->ops->exit_hctx(hctx, i);
1564
1565                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1566                 kfree(hctx->ctxs);
1567                 blk_mq_free_bitmap(&hctx->ctx_map);
1568         }
1569
1570 }
1571
1572 static void blk_mq_free_hw_queues(struct request_queue *q,
1573                 struct blk_mq_tag_set *set)
1574 {
1575         struct blk_mq_hw_ctx *hctx;
1576         unsigned int i;
1577
1578         queue_for_each_hw_ctx(q, hctx, i) {
1579                 free_cpumask_var(hctx->cpumask);
1580                 kfree(hctx);
1581         }
1582 }
1583
1584 static int blk_mq_init_hw_queues(struct request_queue *q,
1585                 struct blk_mq_tag_set *set)
1586 {
1587         struct blk_mq_hw_ctx *hctx;
1588         unsigned int i;
1589
1590         /*
1591          * Initialize hardware queues
1592          */
1593         queue_for_each_hw_ctx(q, hctx, i) {
1594                 int node;
1595
1596                 node = hctx->numa_node;
1597                 if (node == NUMA_NO_NODE)
1598                         node = hctx->numa_node = set->numa_node;
1599
1600                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1601                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1602                 spin_lock_init(&hctx->lock);
1603                 INIT_LIST_HEAD(&hctx->dispatch);
1604                 hctx->queue = q;
1605                 hctx->queue_num = i;
1606                 hctx->flags = set->flags;
1607                 hctx->cmd_size = set->cmd_size;
1608
1609                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1610                                                 blk_mq_hctx_notify, hctx);
1611                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1612
1613                 hctx->tags = set->tags[i];
1614
1615                 /*
1616                  * Allocate space for all possible cpus to avoid allocation in
1617                  * runtime
1618                  */
1619                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1620                                                 GFP_KERNEL, node);
1621                 if (!hctx->ctxs)
1622                         break;
1623
1624                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1625                         break;
1626
1627                 hctx->nr_ctx = 0;
1628
1629                 if (set->ops->init_hctx &&
1630                     set->ops->init_hctx(hctx, set->driver_data, i))
1631                         break;
1632         }
1633
1634         if (i == q->nr_hw_queues)
1635                 return 0;
1636
1637         /*
1638          * Init failed
1639          */
1640         blk_mq_exit_hw_queues(q, set, i);
1641
1642         return 1;
1643 }
1644
1645 static void blk_mq_init_cpu_queues(struct request_queue *q,
1646                                    unsigned int nr_hw_queues)
1647 {
1648         unsigned int i;
1649
1650         for_each_possible_cpu(i) {
1651                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1652                 struct blk_mq_hw_ctx *hctx;
1653
1654                 memset(__ctx, 0, sizeof(*__ctx));
1655                 __ctx->cpu = i;
1656                 spin_lock_init(&__ctx->lock);
1657                 INIT_LIST_HEAD(&__ctx->rq_list);
1658                 __ctx->queue = q;
1659
1660                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1661                 if (!cpu_online(i))
1662                         continue;
1663
1664                 hctx = q->mq_ops->map_queue(q, i);
1665                 cpumask_set_cpu(i, hctx->cpumask);
1666                 hctx->nr_ctx++;
1667
1668                 /*
1669                  * Set local node, IFF we have more than one hw queue. If
1670                  * not, we remain on the home node of the device
1671                  */
1672                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1673                         hctx->numa_node = cpu_to_node(i);
1674         }
1675 }
1676
1677 static void blk_mq_map_swqueue(struct request_queue *q)
1678 {
1679         unsigned int i;
1680         struct blk_mq_hw_ctx *hctx;
1681         struct blk_mq_ctx *ctx;
1682
1683         queue_for_each_hw_ctx(q, hctx, i) {
1684                 cpumask_clear(hctx->cpumask);
1685                 hctx->nr_ctx = 0;
1686         }
1687
1688         /*
1689          * Map software to hardware queues
1690          */
1691         queue_for_each_ctx(q, ctx, i) {
1692                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1693                 if (!cpu_online(i))
1694                         continue;
1695
1696                 hctx = q->mq_ops->map_queue(q, i);
1697                 cpumask_set_cpu(i, hctx->cpumask);
1698                 ctx->index_hw = hctx->nr_ctx;
1699                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1700         }
1701
1702         queue_for_each_hw_ctx(q, hctx, i) {
1703                 /*
1704                  * If not software queues are mapped to this hardware queue,
1705                  * disable it and free the request entries
1706                  */
1707                 if (!hctx->nr_ctx) {
1708                         struct blk_mq_tag_set *set = q->tag_set;
1709
1710                         if (set->tags[i]) {
1711                                 blk_mq_free_rq_map(set, set->tags[i], i);
1712                                 set->tags[i] = NULL;
1713                                 hctx->tags = NULL;
1714                         }
1715                         continue;
1716                 }
1717
1718                 /*
1719                  * Initialize batch roundrobin counts
1720                  */
1721                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1722                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1723         }
1724 }
1725
1726 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1727 {
1728         struct blk_mq_hw_ctx *hctx;
1729         struct request_queue *q;
1730         bool shared;
1731         int i;
1732
1733         if (set->tag_list.next == set->tag_list.prev)
1734                 shared = false;
1735         else
1736                 shared = true;
1737
1738         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1739                 blk_mq_freeze_queue(q);
1740
1741                 queue_for_each_hw_ctx(q, hctx, i) {
1742                         if (shared)
1743                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1744                         else
1745                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1746                 }
1747                 blk_mq_unfreeze_queue(q);
1748         }
1749 }
1750
1751 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1752 {
1753         struct blk_mq_tag_set *set = q->tag_set;
1754
1755         blk_mq_freeze_queue(q);
1756
1757         mutex_lock(&set->tag_list_lock);
1758         list_del_init(&q->tag_set_list);
1759         blk_mq_update_tag_set_depth(set);
1760         mutex_unlock(&set->tag_list_lock);
1761
1762         blk_mq_unfreeze_queue(q);
1763 }
1764
1765 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1766                                      struct request_queue *q)
1767 {
1768         q->tag_set = set;
1769
1770         mutex_lock(&set->tag_list_lock);
1771         list_add_tail(&q->tag_set_list, &set->tag_list);
1772         blk_mq_update_tag_set_depth(set);
1773         mutex_unlock(&set->tag_list_lock);
1774 }
1775
1776 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1777 {
1778         struct blk_mq_hw_ctx **hctxs;
1779         struct blk_mq_ctx __percpu *ctx;
1780         struct request_queue *q;
1781         unsigned int *map;
1782         int i;
1783
1784         ctx = alloc_percpu(struct blk_mq_ctx);
1785         if (!ctx)
1786                 return ERR_PTR(-ENOMEM);
1787
1788         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1789                         set->numa_node);
1790
1791         if (!hctxs)
1792                 goto err_percpu;
1793
1794         map = blk_mq_make_queue_map(set);
1795         if (!map)
1796                 goto err_map;
1797
1798         for (i = 0; i < set->nr_hw_queues; i++) {
1799                 int node = blk_mq_hw_queue_to_node(map, i);
1800
1801                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1802                                         GFP_KERNEL, node);
1803                 if (!hctxs[i])
1804                         goto err_hctxs;
1805
1806                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1807                         goto err_hctxs;
1808
1809                 atomic_set(&hctxs[i]->nr_active, 0);
1810                 hctxs[i]->numa_node = node;
1811                 hctxs[i]->queue_num = i;
1812         }
1813
1814         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1815         if (!q)
1816                 goto err_hctxs;
1817
1818         if (percpu_counter_init(&q->mq_usage_counter, 0))
1819                 goto err_map;
1820
1821         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1822         blk_queue_rq_timeout(q, 30000);
1823
1824         q->nr_queues = nr_cpu_ids;
1825         q->nr_hw_queues = set->nr_hw_queues;
1826         q->mq_map = map;
1827
1828         q->queue_ctx = ctx;
1829         q->queue_hw_ctx = hctxs;
1830
1831         q->mq_ops = set->ops;
1832         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1833
1834         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1835                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1836
1837         q->sg_reserved_size = INT_MAX;
1838
1839         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1840         INIT_LIST_HEAD(&q->requeue_list);
1841         spin_lock_init(&q->requeue_lock);
1842
1843         if (q->nr_hw_queues > 1)
1844                 blk_queue_make_request(q, blk_mq_make_request);
1845         else
1846                 blk_queue_make_request(q, blk_sq_make_request);
1847
1848         blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1849         if (set->timeout)
1850                 blk_queue_rq_timeout(q, set->timeout);
1851
1852         /*
1853          * Do this after blk_queue_make_request() overrides it...
1854          */
1855         q->nr_requests = set->queue_depth;
1856
1857         if (set->ops->complete)
1858                 blk_queue_softirq_done(q, set->ops->complete);
1859
1860         blk_mq_init_flush(q);
1861         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1862
1863         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1864                                 set->cmd_size, cache_line_size()),
1865                                 GFP_KERNEL);
1866         if (!q->flush_rq)
1867                 goto err_hw;
1868
1869         if (blk_mq_init_hw_queues(q, set))
1870                 goto err_flush_rq;
1871
1872         mutex_lock(&all_q_mutex);
1873         list_add_tail(&q->all_q_node, &all_q_list);
1874         mutex_unlock(&all_q_mutex);
1875
1876         blk_mq_add_queue_tag_set(set, q);
1877
1878         blk_mq_map_swqueue(q);
1879
1880         return q;
1881
1882 err_flush_rq:
1883         kfree(q->flush_rq);
1884 err_hw:
1885         blk_cleanup_queue(q);
1886 err_hctxs:
1887         kfree(map);
1888         for (i = 0; i < set->nr_hw_queues; i++) {
1889                 if (!hctxs[i])
1890                         break;
1891                 free_cpumask_var(hctxs[i]->cpumask);
1892                 kfree(hctxs[i]);
1893         }
1894 err_map:
1895         kfree(hctxs);
1896 err_percpu:
1897         free_percpu(ctx);
1898         return ERR_PTR(-ENOMEM);
1899 }
1900 EXPORT_SYMBOL(blk_mq_init_queue);
1901
1902 void blk_mq_free_queue(struct request_queue *q)
1903 {
1904         struct blk_mq_tag_set   *set = q->tag_set;
1905
1906         blk_mq_del_queue_tag_set(q);
1907
1908         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1909         blk_mq_free_hw_queues(q, set);
1910
1911         percpu_counter_destroy(&q->mq_usage_counter);
1912
1913         free_percpu(q->queue_ctx);
1914         kfree(q->queue_hw_ctx);
1915         kfree(q->mq_map);
1916
1917         q->queue_ctx = NULL;
1918         q->queue_hw_ctx = NULL;
1919         q->mq_map = NULL;
1920
1921         mutex_lock(&all_q_mutex);
1922         list_del_init(&q->all_q_node);
1923         mutex_unlock(&all_q_mutex);
1924 }
1925
1926 /* Basically redo blk_mq_init_queue with queue frozen */
1927 static void blk_mq_queue_reinit(struct request_queue *q)
1928 {
1929         blk_mq_freeze_queue(q);
1930
1931         blk_mq_sysfs_unregister(q);
1932
1933         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1934
1935         /*
1936          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1937          * we should change hctx numa_node according to new topology (this
1938          * involves free and re-allocate memory, worthy doing?)
1939          */
1940
1941         blk_mq_map_swqueue(q);
1942
1943         blk_mq_sysfs_register(q);
1944
1945         blk_mq_unfreeze_queue(q);
1946 }
1947
1948 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1949                                       unsigned long action, void *hcpu)
1950 {
1951         struct request_queue *q;
1952
1953         /*
1954          * Before new mappings are established, hotadded cpu might already
1955          * start handling requests. This doesn't break anything as we map
1956          * offline CPUs to first hardware queue. We will re-init the queue
1957          * below to get optimal settings.
1958          */
1959         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1960             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1961                 return NOTIFY_OK;
1962
1963         mutex_lock(&all_q_mutex);
1964         list_for_each_entry(q, &all_q_list, all_q_node)
1965                 blk_mq_queue_reinit(q);
1966         mutex_unlock(&all_q_mutex);
1967         return NOTIFY_OK;
1968 }
1969
1970 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1971 {
1972         int i;
1973
1974         if (!set->nr_hw_queues)
1975                 return -EINVAL;
1976         if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1977                 return -EINVAL;
1978         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1979                 return -EINVAL;
1980
1981         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1982                 return -EINVAL;
1983
1984
1985         set->tags = kmalloc_node(set->nr_hw_queues *
1986                                  sizeof(struct blk_mq_tags *),
1987                                  GFP_KERNEL, set->numa_node);
1988         if (!set->tags)
1989                 goto out;
1990
1991         for (i = 0; i < set->nr_hw_queues; i++) {
1992                 set->tags[i] = blk_mq_init_rq_map(set, i);
1993                 if (!set->tags[i])
1994                         goto out_unwind;
1995         }
1996
1997         mutex_init(&set->tag_list_lock);
1998         INIT_LIST_HEAD(&set->tag_list);
1999
2000         return 0;
2001
2002 out_unwind:
2003         while (--i >= 0)
2004                 blk_mq_free_rq_map(set, set->tags[i], i);
2005 out:
2006         return -ENOMEM;
2007 }
2008 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2009
2010 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2011 {
2012         int i;
2013
2014         for (i = 0; i < set->nr_hw_queues; i++) {
2015                 if (set->tags[i])
2016                         blk_mq_free_rq_map(set, set->tags[i], i);
2017         }
2018
2019         kfree(set->tags);
2020 }
2021 EXPORT_SYMBOL(blk_mq_free_tag_set);
2022
2023 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2024 {
2025         struct blk_mq_tag_set *set = q->tag_set;
2026         struct blk_mq_hw_ctx *hctx;
2027         int i, ret;
2028
2029         if (!set || nr > set->queue_depth)
2030                 return -EINVAL;
2031
2032         ret = 0;
2033         queue_for_each_hw_ctx(q, hctx, i) {
2034                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2035                 if (ret)
2036                         break;
2037         }
2038
2039         if (!ret)
2040                 q->nr_requests = nr;
2041
2042         return ret;
2043 }
2044
2045 void blk_mq_disable_hotplug(void)
2046 {
2047         mutex_lock(&all_q_mutex);
2048 }
2049
2050 void blk_mq_enable_hotplug(void)
2051 {
2052         mutex_unlock(&all_q_mutex);
2053 }
2054
2055 static int __init blk_mq_init(void)
2056 {
2057         blk_mq_cpu_init();
2058
2059         /* Must be called after percpu_counter_hotcpu_callback() */
2060         hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2061
2062         return 0;
2063 }
2064 subsys_initcall(blk_mq_init);