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         bool freeze;
 116
 117         spin_lock_irq(q->queue_lock);
 118         freeze = !q->mq_freeze_depth++;
 119         spin_unlock_irq(q->queue_lock);
 120
 121         if (freeze) {
 122                 percpu_ref_kill(&q->mq_usage_counter);
 123                 blk_mq_run_queues(q, false);
 124         }
 125         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 126 }
 127
 128 static void blk_mq_unfreeze_queue(struct request_queue *q)
 129 {
 130         bool wake;
 131
 132         spin_lock_irq(q->queue_lock);
 133         wake = !--q->mq_freeze_depth;
 134         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 135         spin_unlock_irq(q->queue_lock);
 136         if (wake) {
 137                 percpu_ref_reinit(&q->mq_usage_counter);
 138                 wake_up_all(&q->mq_freeze_wq);
 139         }
 140 }
 141
 142 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 143 {
 144         return blk_mq_has_free_tags(hctx->tags);
 145 }
 146 EXPORT_SYMBOL(blk_mq_can_queue);
 147
 148 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 149                                struct request *rq, unsigned int rw_flags)
 150 {
 151         if (blk_queue_io_stat(q))
 152                 rw_flags |= REQ_IO_STAT;
 153
 154         INIT_LIST_HEAD(&rq->queuelist);
 155         /* csd/requeue_work/fifo_time is initialized before use */
 156         rq->q = q;
 157         rq->mq_ctx = ctx;
 158         rq->cmd_flags |= rw_flags;
 159         /* do not touch atomic flags, it needs atomic ops against the timer */
 160         rq->cpu = -1;
 161         INIT_HLIST_NODE(&rq->hash);
 162         RB_CLEAR_NODE(&rq->rb_node);
 163         rq->rq_disk = NULL;
 164         rq->part = NULL;
 165         rq->start_time = jiffies;
 166 #ifdef CONFIG_BLK_CGROUP
 167         rq->rl = NULL;
 168         set_start_time_ns(rq);
 169         rq->io_start_time_ns = 0;
 170 #endif
 171         rq->nr_phys_segments = 0;
 172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 173         rq->nr_integrity_segments = 0;
 174 #endif
 175         rq->special = NULL;
 176         /* tag was already set */
 177         rq->errors = 0;
 178
 179         rq->cmd = rq->__cmd;
 180
 181         rq->extra_len = 0;
 182         rq->sense_len = 0;
 183         rq->resid_len = 0;
 184         rq->sense = NULL;
 185
 186         INIT_LIST_HEAD(&rq->timeout_list);
 187         rq->timeout = 0;
 188
 189         rq->end_io = NULL;
 190         rq->end_io_data = NULL;
 191         rq->next_rq = NULL;
 192
 193         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 194 }
 195
 196 static struct request *
 197 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 198 {
 199         struct request *rq;
 200         unsigned int tag;
 201
 202         tag = blk_mq_get_tag(data);
 203         if (tag != BLK_MQ_TAG_FAIL) {
 204                 rq = data->hctx->tags->rqs[tag];
 205
 206                 if (blk_mq_tag_busy(data->hctx)) {
 207                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 208                         atomic_inc(&data->hctx->nr_active);
 209                 }
 210
 211                 rq->tag = tag;
 212                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 213                 return rq;
 214         }
 215
 216         return NULL;
 217 }
 218
 219 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 220                 bool reserved)
 221 {
 222         struct blk_mq_ctx *ctx;
 223         struct blk_mq_hw_ctx *hctx;
 224         struct request *rq;
 225         struct blk_mq_alloc_data alloc_data;
 226
 227         if (blk_mq_queue_enter(q))
 228                 return NULL;
 229
 230         ctx = blk_mq_get_ctx(q);
 231         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 232         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 233                         reserved, ctx, hctx);
 234
 235         rq = __blk_mq_alloc_request(&alloc_data, rw);
 236         if (!rq && (gfp & __GFP_WAIT)) {
 237                 __blk_mq_run_hw_queue(hctx);
 238                 blk_mq_put_ctx(ctx);
 239
 240                 ctx = blk_mq_get_ctx(q);
 241                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 242                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 243                                 hctx);
 244                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 245                 ctx = alloc_data.ctx;
 246         }
 247         blk_mq_put_ctx(ctx);
 248         return rq;
 249 }
 250 EXPORT_SYMBOL(blk_mq_alloc_request);
 251
 252 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 253                                   struct blk_mq_ctx *ctx, struct request *rq)
 254 {
 255         const int tag = rq->tag;
 256         struct request_queue *q = rq->q;
 257
 258         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 259                 atomic_dec(&hctx->nr_active);
 260         rq->cmd_flags = 0;
 261
 262         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 263         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 264         blk_mq_queue_exit(q);
 265 }
 266
 267 void blk_mq_free_request(struct request *rq)
 268 {
 269         struct blk_mq_ctx *ctx = rq->mq_ctx;
 270         struct blk_mq_hw_ctx *hctx;
 271         struct request_queue *q = rq->q;
 272
 273         ctx->rq_completed[rq_is_sync(rq)]++;
 274
 275         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 276         __blk_mq_free_request(hctx, ctx, rq);
 277 }
 278
 279 /*
 280  * Clone all relevant state from a request that has been put on hold in
 281  * the flush state machine into the preallocated flush request that hangs
 282  * off the request queue.
 283  *
 284  * For a driver the flush request should be invisible, that's why we are
 285  * impersonating the original request here.
 286  */
 287 void blk_mq_clone_flush_request(struct request *flush_rq,
 288                 struct request *orig_rq)
 289 {
 290         struct blk_mq_hw_ctx *hctx =
 291                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 292
 293         flush_rq->mq_ctx = orig_rq->mq_ctx;
 294         flush_rq->tag = orig_rq->tag;
 295         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 296                 hctx->cmd_size);
 297 }
 298
 299 inline void __blk_mq_end_io(struct request *rq, int error)
 300 {
 301         blk_account_io_done(rq);
 302
 303         if (rq->end_io) {
 304                 rq->end_io(rq, error);
 305         } else {
 306                 if (unlikely(blk_bidi_rq(rq)))
 307                         blk_mq_free_request(rq->next_rq);
 308                 blk_mq_free_request(rq);
 309         }
 310 }
 311 EXPORT_SYMBOL(__blk_mq_end_io);
 312
 313 void blk_mq_end_io(struct request *rq, int error)
 314 {
 315         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 316                 BUG();
 317         __blk_mq_end_io(rq, error);
 318 }
 319 EXPORT_SYMBOL(blk_mq_end_io);
 320
 321 static void __blk_mq_complete_request_remote(void *data)
 322 {
 323         struct request *rq = data;
 324
 325         rq->q->softirq_done_fn(rq);
 326 }
 327
 328 static void blk_mq_ipi_complete_request(struct request *rq)
 329 {
 330         struct blk_mq_ctx *ctx = rq->mq_ctx;
 331         bool shared = false;
 332         int cpu;
 333
 334         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 335                 rq->q->softirq_done_fn(rq);
 336                 return;
 337         }
 338
 339         cpu = get_cpu();
 340         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 341                 shared = cpus_share_cache(cpu, ctx->cpu);
 342
 343         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 344                 rq->csd.func = __blk_mq_complete_request_remote;
 345                 rq->csd.info = rq;
 346                 rq->csd.flags = 0;
 347                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 348         } else {
 349                 rq->q->softirq_done_fn(rq);
 350         }
 351         put_cpu();
 352 }
 353
 354 void __blk_mq_complete_request(struct request *rq)
 355 {
 356         struct request_queue *q = rq->q;
 357
 358         if (!q->softirq_done_fn)
 359                 blk_mq_end_io(rq, rq->errors);
 360         else
 361                 blk_mq_ipi_complete_request(rq);
 362 }
 363
 364 /**
 365  * blk_mq_complete_request - end I/O on a request
 366  * @rq:         the request being processed
 367  *
 368  * Description:
 369  *      Ends all I/O on a request. It does not handle partial completions.
 370  *      The actual completion happens out-of-order, through a IPI handler.
 371  **/
 372 void blk_mq_complete_request(struct request *rq)
 373 {
 374         struct request_queue *q = rq->q;
 375
 376         if (unlikely(blk_should_fake_timeout(q)))
 377                 return;
 378         if (!blk_mark_rq_complete(rq))
 379                 __blk_mq_complete_request(rq);
 380 }
 381 EXPORT_SYMBOL(blk_mq_complete_request);
 382
 383 static void blk_mq_start_request(struct request *rq, bool last)
 384 {
 385         struct request_queue *q = rq->q;
 386
 387         trace_block_rq_issue(q, rq);
 388
 389         rq->resid_len = blk_rq_bytes(rq);
 390         if (unlikely(blk_bidi_rq(rq)))
 391                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 392
 393         blk_add_timer(rq);
 394
 395         /*
 396          * Ensure that ->deadline is visible before set the started
 397          * flag and clear the completed flag.
 398          */
 399         smp_mb__before_atomic();
 400
 401         /*
 402          * Mark us as started and clear complete. Complete might have been
 403          * set if requeue raced with timeout, which then marked it as
 404          * complete. So be sure to clear complete again when we start
 405          * the request, otherwise we'll ignore the completion event.
 406          */
 407         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 408                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 409         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 410                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 411
 412         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 413                 /*
 414                  * Make sure space for the drain appears.  We know we can do
 415                  * this because max_hw_segments has been adjusted to be one
 416                  * fewer than the device can handle.
 417                  */
 418                 rq->nr_phys_segments++;
 419         }
 420
 421         /*
 422          * Flag the last request in the series so that drivers know when IO
 423          * should be kicked off, if they don't do it on a per-request basis.
 424          *
 425          * Note: the flag isn't the only condition drivers should do kick off.
 426          * If drive is busy, the last request might not have the bit set.
 427          */
 428         if (last)
 429                 rq->cmd_flags |= REQ_END;
 430 }
 431
 432 static void __blk_mq_requeue_request(struct request *rq)
 433 {
 434         struct request_queue *q = rq->q;
 435
 436         trace_block_rq_requeue(q, rq);
 437         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 438
 439         rq->cmd_flags &= ~REQ_END;
 440
 441         if (q->dma_drain_size && blk_rq_bytes(rq))
 442                 rq->nr_phys_segments--;
 443 }
 444
 445 void blk_mq_requeue_request(struct request *rq)
 446 {
 447         __blk_mq_requeue_request(rq);
 448         blk_clear_rq_complete(rq);
 449
 450         BUG_ON(blk_queued_rq(rq));
 451         blk_mq_add_to_requeue_list(rq, true);
 452 }
 453 EXPORT_SYMBOL(blk_mq_requeue_request);
 454
 455 static void blk_mq_requeue_work(struct work_struct *work)
 456 {
 457         struct request_queue *q =
 458                 container_of(work, struct request_queue, requeue_work);
 459         LIST_HEAD(rq_list);
 460         struct request *rq, *next;
 461         unsigned long flags;
 462
 463         spin_lock_irqsave(&q->requeue_lock, flags);
 464         list_splice_init(&q->requeue_list, &rq_list);
 465         spin_unlock_irqrestore(&q->requeue_lock, flags);
 466
 467         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 468                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 469                         continue;
 470
 471                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 472                 list_del_init(&rq->queuelist);
 473                 blk_mq_insert_request(rq, true, false, false);
 474         }
 475
 476         while (!list_empty(&rq_list)) {
 477                 rq = list_entry(rq_list.next, struct request, queuelist);
 478                 list_del_init(&rq->queuelist);
 479                 blk_mq_insert_request(rq, false, false, false);
 480         }
 481
 482         /*
 483          * Use the start variant of queue running here, so that running
 484          * the requeue work will kick stopped queues.
 485          */
 486         blk_mq_start_hw_queues(q);
 487 }
 488
 489 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 490 {
 491         struct request_queue *q = rq->q;
 492         unsigned long flags;
 493
 494         /*
 495          * We abuse this flag that is otherwise used by the I/O scheduler to
 496          * request head insertation from the workqueue.
 497          */
 498         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 499
 500         spin_lock_irqsave(&q->requeue_lock, flags);
 501         if (at_head) {
 502                 rq->cmd_flags |= REQ_SOFTBARRIER;
 503                 list_add(&rq->queuelist, &q->requeue_list);
 504         } else {
 505                 list_add_tail(&rq->queuelist, &q->requeue_list);
 506         }
 507         spin_unlock_irqrestore(&q->requeue_lock, flags);
 508 }
 509 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 510
 511 void blk_mq_kick_requeue_list(struct request_queue *q)
 512 {
 513         kblockd_schedule_work(&q->requeue_work);
 514 }
 515 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 516
 517 static inline bool is_flush_request(struct request *rq, unsigned int tag)
 518 {
 519         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 520                         rq->q->flush_rq->tag == tag);
 521 }
 522
 523 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 524 {
 525         struct request *rq = tags->rqs[tag];
 526
 527         if (!is_flush_request(rq, tag))
 528                 return rq;
 529
 530         return rq->q->flush_rq;
 531 }
 532 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 533
 534 struct blk_mq_timeout_data {
 535         struct blk_mq_hw_ctx *hctx;
 536         unsigned long *next;
 537         unsigned int *next_set;
 538 };
 539
 540 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
 541 {
 542         struct blk_mq_timeout_data *data = __data;
 543         struct blk_mq_hw_ctx *hctx = data->hctx;
 544         unsigned int tag;
 545
 546          /* It may not be in flight yet (this is where
 547          * the REQ_ATOMIC_STARTED flag comes in). The requests are
 548          * statically allocated, so we know it's always safe to access the
 549          * memory associated with a bit offset into ->rqs[].
 550          */
 551         tag = 0;
 552         do {
 553                 struct request *rq;
 554
 555                 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
 556                 if (tag >= hctx->tags->nr_tags)
 557                         break;
 558
 559                 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
 560                 if (rq->q != hctx->queue)
 561                         continue;
 562                 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 563                         continue;
 564
 565                 blk_rq_check_expired(rq, data->next, data->next_set);
 566         } while (1);
 567 }
 568
 569 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
 570                                         unsigned long *next,
 571                                         unsigned int *next_set)
 572 {
 573         struct blk_mq_timeout_data data = {
 574                 .hctx           = hctx,
 575                 .next           = next,
 576                 .next_set       = next_set,
 577         };
 578
 579         /*
 580          * Ask the tagging code to iterate busy requests, so we can
 581          * check them for timeout.
 582          */
 583         blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
 584 }
 585
 586 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
 587 {
 588         struct request_queue *q = rq->q;
 589
 590         /*
 591          * We know that complete is set at this point. If STARTED isn't set
 592          * anymore, then the request isn't active and the "timeout" should
 593          * just be ignored. This can happen due to the bitflag ordering.
 594          * Timeout first checks if STARTED is set, and if it is, assumes
 595          * the request is active. But if we race with completion, then
 596          * we both flags will get cleared. So check here again, and ignore
 597          * a timeout event with a request that isn't active.
 598          */
 599         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 600                 return BLK_EH_NOT_HANDLED;
 601
 602         if (!q->mq_ops->timeout)
 603                 return BLK_EH_RESET_TIMER;
 604
 605         return q->mq_ops->timeout(rq);
 606 }
 607
 608 static void blk_mq_rq_timer(unsigned long data)
 609 {
 610         struct request_queue *q = (struct request_queue *) data;
 611         struct blk_mq_hw_ctx *hctx;
 612         unsigned long next = 0;
 613         int i, next_set = 0;
 614
 615         queue_for_each_hw_ctx(q, hctx, i) {
 616                 /*
 617                  * If not software queues are currently mapped to this
 618                  * hardware queue, there's nothing to check
 619                  */
 620                 if (!hctx->nr_ctx || !hctx->tags)
 621                         continue;
 622
 623                 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
 624         }
 625
 626         if (next_set) {
 627                 next = blk_rq_timeout(round_jiffies_up(next));
 628                 mod_timer(&q->timeout, next);
 629         } else {
 630                 queue_for_each_hw_ctx(q, hctx, i)
 631                         blk_mq_tag_idle(hctx);
 632         }
 633 }
 634
 635 /*
 636  * Reverse check our software queue for entries that we could potentially
 637  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 638  * too much time checking for merges.
 639  */
 640 static bool blk_mq_attempt_merge(struct request_queue *q,
 641                                  struct blk_mq_ctx *ctx, struct bio *bio)
 642 {
 643         struct request *rq;
 644         int checked = 8;
 645
 646         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 647                 int el_ret;
 648
 649                 if (!checked--)
 650                         break;
 651
 652                 if (!blk_rq_merge_ok(rq, bio))
 653                         continue;
 654
 655                 el_ret = blk_try_merge(rq, bio);
 656                 if (el_ret == ELEVATOR_BACK_MERGE) {
 657                         if (bio_attempt_back_merge(q, rq, bio)) {
 658                                 ctx->rq_merged++;
 659                                 return true;
 660                         }
 661                         break;
 662                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 663                         if (bio_attempt_front_merge(q, rq, bio)) {
 664                                 ctx->rq_merged++;
 665                                 return true;
 666                         }
 667                         break;
 668                 }
 669         }
 670
 671         return false;
 672 }
 673
 674 /*
 675  * Process software queues that have been marked busy, splicing them
 676  * to the for-dispatch
 677  */
 678 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 679 {
 680         struct blk_mq_ctx *ctx;
 681         int i;
 682
 683         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 684                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 685                 unsigned int off, bit;
 686
 687                 if (!bm->word)
 688                         continue;
 689
 690                 bit = 0;
 691                 off = i * hctx->ctx_map.bits_per_word;
 692                 do {
 693                         bit = find_next_bit(&bm->word, bm->depth, bit);
 694                         if (bit >= bm->depth)
 695                                 break;
 696
 697                         ctx = hctx->ctxs[bit + off];
 698                         clear_bit(bit, &bm->word);
 699                         spin_lock(&ctx->lock);
 700                         list_splice_tail_init(&ctx->rq_list, list);
 701                         spin_unlock(&ctx->lock);
 702
 703                         bit++;
 704                 } while (1);
 705         }
 706 }
 707
 708 /*
 709  * Run this hardware queue, pulling any software queues mapped to it in.
 710  * Note that this function currently has various problems around ordering
 711  * of IO. In particular, we'd like FIFO behaviour on handling existing
 712  * items on the hctx->dispatch list. Ignore that for now.
 713  */
 714 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 715 {
 716         struct request_queue *q = hctx->queue;
 717         struct request *rq;
 718         LIST_HEAD(rq_list);
 719         int queued;
 720
 721         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 722
 723         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 724                 return;
 725
 726         hctx->run++;
 727
 728         /*
 729          * Touch any software queue that has pending entries.
 730          */
 731         flush_busy_ctxs(hctx, &rq_list);
 732
 733         /*
 734          * If we have previous entries on our dispatch list, grab them
 735          * and stuff them at the front for more fair dispatch.
 736          */
 737         if (!list_empty_careful(&hctx->dispatch)) {
 738                 spin_lock(&hctx->lock);
 739                 if (!list_empty(&hctx->dispatch))
 740                         list_splice_init(&hctx->dispatch, &rq_list);
 741                 spin_unlock(&hctx->lock);
 742         }
 743
 744         /*
 745          * Now process all the entries, sending them to the driver.
 746          */
 747         queued = 0;
 748         while (!list_empty(&rq_list)) {
 749                 int ret;
 750
 751                 rq = list_first_entry(&rq_list, struct request, queuelist);
 752                 list_del_init(&rq->queuelist);
 753
 754                 blk_mq_start_request(rq, list_empty(&rq_list));
 755
 756                 ret = q->mq_ops->queue_rq(hctx, rq);
 757                 switch (ret) {
 758                 case BLK_MQ_RQ_QUEUE_OK:
 759                         queued++;
 760                         continue;
 761                 case BLK_MQ_RQ_QUEUE_BUSY:
 762                         list_add(&rq->queuelist, &rq_list);
 763                         __blk_mq_requeue_request(rq);
 764                         break;
 765                 default:
 766                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 767                 case BLK_MQ_RQ_QUEUE_ERROR:
 768                         rq->errors = -EIO;
 769                         blk_mq_end_io(rq, rq->errors);
 770                         break;
 771                 }
 772
 773                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 774                         break;
 775         }
 776
 777         if (!queued)
 778                 hctx->dispatched[0]++;
 779         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 780                 hctx->dispatched[ilog2(queued) + 1]++;
 781
 782         /*
 783          * Any items that need requeuing? Stuff them into hctx->dispatch,
 784          * that is where we will continue on next queue run.
 785          */
 786         if (!list_empty(&rq_list)) {
 787                 spin_lock(&hctx->lock);
 788                 list_splice(&rq_list, &hctx->dispatch);
 789                 spin_unlock(&hctx->lock);
 790         }
 791 }
 792
 793 /*
 794  * It'd be great if the workqueue API had a way to pass
 795  * in a mask and had some smarts for more clever placement.
 796  * For now we just round-robin here, switching for every
 797  * BLK_MQ_CPU_WORK_BATCH queued items.
 798  */
 799 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 800 {
 801         int cpu = hctx->next_cpu;
 802
 803         if (--hctx->next_cpu_batch <= 0) {
 804                 int next_cpu;
 805
 806                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 807                 if (next_cpu >= nr_cpu_ids)
 808                         next_cpu = cpumask_first(hctx->cpumask);
 809
 810                 hctx->next_cpu = next_cpu;
 811                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 812         }
 813
 814         return cpu;
 815 }
 816
 817 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 818 {
 819         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 820                 return;
 821
 822         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 823                 __blk_mq_run_hw_queue(hctx);
 824         else if (hctx->queue->nr_hw_queues == 1)
 825                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 826         else {
 827                 unsigned int cpu;
 828
 829                 cpu = blk_mq_hctx_next_cpu(hctx);
 830                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 831         }
 832 }
 833
 834 void blk_mq_run_queues(struct request_queue *q, bool async)
 835 {
 836         struct blk_mq_hw_ctx *hctx;
 837         int i;
 838
 839         queue_for_each_hw_ctx(q, hctx, i) {
 840                 if ((!blk_mq_hctx_has_pending(hctx) &&
 841                     list_empty_careful(&hctx->dispatch)) ||
 842                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 843                         continue;
 844
 845                 preempt_disable();
 846                 blk_mq_run_hw_queue(hctx, async);
 847                 preempt_enable();
 848         }
 849 }
 850 EXPORT_SYMBOL(blk_mq_run_queues);
 851
 852 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 853 {
 854         cancel_delayed_work(&hctx->run_work);
 855         cancel_delayed_work(&hctx->delay_work);
 856         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 857 }
 858 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 859
 860 void blk_mq_stop_hw_queues(struct request_queue *q)
 861 {
 862         struct blk_mq_hw_ctx *hctx;
 863         int i;
 864
 865         queue_for_each_hw_ctx(q, hctx, i)
 866                 blk_mq_stop_hw_queue(hctx);
 867 }
 868 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 869
 870 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 871 {
 872         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 873
 874         preempt_disable();
 875         blk_mq_run_hw_queue(hctx, false);
 876         preempt_enable();
 877 }
 878 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 879
 880 void blk_mq_start_hw_queues(struct request_queue *q)
 881 {
 882         struct blk_mq_hw_ctx *hctx;
 883         int i;
 884
 885         queue_for_each_hw_ctx(q, hctx, i)
 886                 blk_mq_start_hw_queue(hctx);
 887 }
 888 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 889
 890
 891 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 892 {
 893         struct blk_mq_hw_ctx *hctx;
 894         int i;
 895
 896         queue_for_each_hw_ctx(q, hctx, i) {
 897                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 898                         continue;
 899
 900                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 901                 preempt_disable();
 902                 blk_mq_run_hw_queue(hctx, async);
 903                 preempt_enable();
 904         }
 905 }
 906 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 907
 908 static void blk_mq_run_work_fn(struct work_struct *work)
 909 {
 910         struct blk_mq_hw_ctx *hctx;
 911
 912         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 913
 914         __blk_mq_run_hw_queue(hctx);
 915 }
 916
 917 static void blk_mq_delay_work_fn(struct work_struct *work)
 918 {
 919         struct blk_mq_hw_ctx *hctx;
 920
 921         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 922
 923         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 924                 __blk_mq_run_hw_queue(hctx);
 925 }
 926
 927 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 928 {
 929         unsigned long tmo = msecs_to_jiffies(msecs);
 930
 931         if (hctx->queue->nr_hw_queues == 1)
 932                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 933         else {
 934                 unsigned int cpu;
 935
 936                 cpu = blk_mq_hctx_next_cpu(hctx);
 937                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 938         }
 939 }
 940 EXPORT_SYMBOL(blk_mq_delay_queue);
 941
 942 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 943                                     struct request *rq, bool at_head)
 944 {
 945         struct blk_mq_ctx *ctx = rq->mq_ctx;
 946
 947         trace_block_rq_insert(hctx->queue, rq);
 948
 949         if (at_head)
 950                 list_add(&rq->queuelist, &ctx->rq_list);
 951         else
 952                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 953
 954         blk_mq_hctx_mark_pending(hctx, ctx);
 955 }
 956
 957 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 958                 bool async)
 959 {
 960         struct request_queue *q = rq->q;
 961         struct blk_mq_hw_ctx *hctx;
 962         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 963
 964         current_ctx = blk_mq_get_ctx(q);
 965         if (!cpu_online(ctx->cpu))
 966                 rq->mq_ctx = ctx = current_ctx;
 967
 968         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 969
 970         spin_lock(&ctx->lock);
 971         __blk_mq_insert_request(hctx, rq, at_head);
 972         spin_unlock(&ctx->lock);
 973
 974         if (run_queue)
 975                 blk_mq_run_hw_queue(hctx, async);
 976
 977         blk_mq_put_ctx(current_ctx);
 978 }
 979
 980 static void blk_mq_insert_requests(struct request_queue *q,
 981                                      struct blk_mq_ctx *ctx,
 982                                      struct list_head *list,
 983                                      int depth,
 984                                      bool from_schedule)
 985
 986 {
 987         struct blk_mq_hw_ctx *hctx;
 988         struct blk_mq_ctx *current_ctx;
 989
 990         trace_block_unplug(q, depth, !from_schedule);
 991
 992         current_ctx = blk_mq_get_ctx(q);
 993
 994         if (!cpu_online(ctx->cpu))
 995                 ctx = current_ctx;
 996         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 997
 998         /*
 999          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1000          * offline now
1001          */
1002         spin_lock(&ctx->lock);
1003         while (!list_empty(list)) {
1004                 struct request *rq;
1005
1006                 rq = list_first_entry(list, struct request, queuelist);
1007                 list_del_init(&rq->queuelist);
1008                 rq->mq_ctx = ctx;
1009                 __blk_mq_insert_request(hctx, rq, false);
1010         }
1011         spin_unlock(&ctx->lock);
1012
1013         blk_mq_run_hw_queue(hctx, from_schedule);
1014         blk_mq_put_ctx(current_ctx);
1015 }
1016
1017 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1018 {
1019         struct request *rqa = container_of(a, struct request, queuelist);
1020         struct request *rqb = container_of(b, struct request, queuelist);
1021
1022         return !(rqa->mq_ctx < rqb->mq_ctx ||
1023                  (rqa->mq_ctx == rqb->mq_ctx &&
1024                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1025 }
1026
1027 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1028 {
1029         struct blk_mq_ctx *this_ctx;
1030         struct request_queue *this_q;
1031         struct request *rq;
1032         LIST_HEAD(list);
1033         LIST_HEAD(ctx_list);
1034         unsigned int depth;
1035
1036         list_splice_init(&plug->mq_list, &list);
1037
1038         list_sort(NULL, &list, plug_ctx_cmp);
1039
1040         this_q = NULL;
1041         this_ctx = NULL;
1042         depth = 0;
1043
1044         while (!list_empty(&list)) {
1045                 rq = list_entry_rq(list.next);
1046                 list_del_init(&rq->queuelist);
1047                 BUG_ON(!rq->q);
1048                 if (rq->mq_ctx != this_ctx) {
1049                         if (this_ctx) {
1050                                 blk_mq_insert_requests(this_q, this_ctx,
1051                                                         &ctx_list, depth,
1052                                                         from_schedule);
1053                         }
1054
1055                         this_ctx = rq->mq_ctx;
1056                         this_q = rq->q;
1057                         depth = 0;
1058                 }
1059
1060                 depth++;
1061                 list_add_tail(&rq->queuelist, &ctx_list);
1062         }
1063
1064         /*
1065          * If 'this_ctx' is set, we know we have entries to complete
1066          * on 'ctx_list'. Do those.
1067          */
1068         if (this_ctx) {
1069                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1070                                        from_schedule);
1071         }
1072 }
1073
1074 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1075 {
1076         init_request_from_bio(rq, bio);
1077
1078         if (blk_do_io_stat(rq))
1079                 blk_account_io_start(rq, 1);
1080 }
1081
1082 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1083 {
1084         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1085                 !blk_queue_nomerges(hctx->queue);
1086 }
1087
1088 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1089                                          struct blk_mq_ctx *ctx,
1090                                          struct request *rq, struct bio *bio)
1091 {
1092         if (!hctx_allow_merges(hctx)) {
1093                 blk_mq_bio_to_request(rq, bio);
1094                 spin_lock(&ctx->lock);
1095 insert_rq:
1096                 __blk_mq_insert_request(hctx, rq, false);
1097                 spin_unlock(&ctx->lock);
1098                 return false;
1099         } else {
1100                 struct request_queue *q = hctx->queue;
1101
1102                 spin_lock(&ctx->lock);
1103                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1104                         blk_mq_bio_to_request(rq, bio);
1105                         goto insert_rq;
1106                 }
1107
1108                 spin_unlock(&ctx->lock);
1109                 __blk_mq_free_request(hctx, ctx, rq);
1110                 return true;
1111         }
1112 }
1113
1114 struct blk_map_ctx {
1115         struct blk_mq_hw_ctx *hctx;
1116         struct blk_mq_ctx *ctx;
1117 };
1118
1119 static struct request *blk_mq_map_request(struct request_queue *q,
1120                                           struct bio *bio,
1121                                           struct blk_map_ctx *data)
1122 {
1123         struct blk_mq_hw_ctx *hctx;
1124         struct blk_mq_ctx *ctx;
1125         struct request *rq;
1126         int rw = bio_data_dir(bio);
1127         struct blk_mq_alloc_data alloc_data;
1128
1129         if (unlikely(blk_mq_queue_enter(q))) {
1130                 bio_endio(bio, -EIO);
1131                 return NULL;
1132         }
1133
1134         ctx = blk_mq_get_ctx(q);
1135         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1136
1137         if (rw_is_sync(bio->bi_rw))
1138                 rw |= REQ_SYNC;
1139
1140         trace_block_getrq(q, bio, rw);
1141         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1142                         hctx);
1143         rq = __blk_mq_alloc_request(&alloc_data, rw);
1144         if (unlikely(!rq)) {
1145                 __blk_mq_run_hw_queue(hctx);
1146                 blk_mq_put_ctx(ctx);
1147                 trace_block_sleeprq(q, bio, rw);
1148
1149                 ctx = blk_mq_get_ctx(q);
1150                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1151                 blk_mq_set_alloc_data(&alloc_data, q,
1152                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1153                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1154                 ctx = alloc_data.ctx;
1155                 hctx = alloc_data.hctx;
1156         }
1157
1158         hctx->queued++;
1159         data->hctx = hctx;
1160         data->ctx = ctx;
1161         return rq;
1162 }
1163
1164 /*
1165  * Multiple hardware queue variant. This will not use per-process plugs,
1166  * but will attempt to bypass the hctx queueing if we can go straight to
1167  * hardware for SYNC IO.
1168  */
1169 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1170 {
1171         const int is_sync = rw_is_sync(bio->bi_rw);
1172         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1173         struct blk_map_ctx data;
1174         struct request *rq;
1175
1176         blk_queue_bounce(q, &bio);
1177
1178         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1179                 bio_endio(bio, -EIO);
1180                 return;
1181         }
1182
1183         rq = blk_mq_map_request(q, bio, &data);
1184         if (unlikely(!rq))
1185                 return;
1186
1187         if (unlikely(is_flush_fua)) {
1188                 blk_mq_bio_to_request(rq, bio);
1189                 blk_insert_flush(rq);
1190                 goto run_queue;
1191         }
1192
1193         if (is_sync) {
1194                 int ret;
1195
1196                 blk_mq_bio_to_request(rq, bio);
1197                 blk_mq_start_request(rq, true);
1198
1199                 /*
1200                  * For OK queue, we are done. For error, kill it. Any other
1201                  * error (busy), just add it to our list as we previously
1202                  * would have done
1203                  */
1204                 ret = q->mq_ops->queue_rq(data.hctx, rq);
1205                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1206                         goto done;
1207                 else {
1208                         __blk_mq_requeue_request(rq);
1209
1210                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1211                                 rq->errors = -EIO;
1212                                 blk_mq_end_io(rq, rq->errors);
1213                                 goto done;
1214                         }
1215                 }
1216         }
1217
1218         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1219                 /*
1220                  * For a SYNC request, send it to the hardware immediately. For
1221                  * an ASYNC request, just ensure that we run it later on. The
1222                  * latter allows for merging opportunities and more efficient
1223                  * dispatching.
1224                  */
1225 run_queue:
1226                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1227         }
1228 done:
1229         blk_mq_put_ctx(data.ctx);
1230 }
1231
1232 /*
1233  * Single hardware queue variant. This will attempt to use any per-process
1234  * plug for merging and IO deferral.
1235  */
1236 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1237 {
1238         const int is_sync = rw_is_sync(bio->bi_rw);
1239         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1240         unsigned int use_plug, request_count = 0;
1241         struct blk_map_ctx data;
1242         struct request *rq;
1243
1244         /*
1245          * If we have multiple hardware queues, just go directly to
1246          * one of those for sync IO.
1247          */
1248         use_plug = !is_flush_fua && !is_sync;
1249
1250         blk_queue_bounce(q, &bio);
1251
1252         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1253                 bio_endio(bio, -EIO);
1254                 return;
1255         }
1256
1257         if (use_plug && !blk_queue_nomerges(q) &&
1258             blk_attempt_plug_merge(q, bio, &request_count))
1259                 return;
1260
1261         rq = blk_mq_map_request(q, bio, &data);
1262         if (unlikely(!rq))
1263                 return;
1264
1265         if (unlikely(is_flush_fua)) {
1266                 blk_mq_bio_to_request(rq, bio);
1267                 blk_insert_flush(rq);
1268                 goto run_queue;
1269         }
1270
1271         /*
1272          * A task plug currently exists. Since this is completely lockless,
1273          * utilize that to temporarily store requests until the task is
1274          * either done or scheduled away.
1275          */
1276         if (use_plug) {
1277                 struct blk_plug *plug = current->plug;
1278
1279                 if (plug) {
1280                         blk_mq_bio_to_request(rq, bio);
1281                         if (list_empty(&plug->mq_list))
1282                                 trace_block_plug(q);
1283                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1284                                 blk_flush_plug_list(plug, false);
1285                                 trace_block_plug(q);
1286                         }
1287                         list_add_tail(&rq->queuelist, &plug->mq_list);
1288                         blk_mq_put_ctx(data.ctx);
1289                         return;
1290                 }
1291         }
1292
1293         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1294                 /*
1295                  * For a SYNC request, send it to the hardware immediately. For
1296                  * an ASYNC request, just ensure that we run it later on. The
1297                  * latter allows for merging opportunities and more efficient
1298                  * dispatching.
1299                  */
1300 run_queue:
1301                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1302         }
1303
1304         blk_mq_put_ctx(data.ctx);
1305 }
1306
1307 /*
1308  * Default mapping to a software queue, since we use one per CPU.
1309  */
1310 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1311 {
1312         return q->queue_hw_ctx[q->mq_map[cpu]];
1313 }
1314 EXPORT_SYMBOL(blk_mq_map_queue);
1315
1316 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1317                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1318 {
1319         struct page *page;
1320
1321         if (tags->rqs && set->ops->exit_request) {
1322                 int i;
1323
1324                 for (i = 0; i < tags->nr_tags; i++) {
1325                         if (!tags->rqs[i])
1326                                 continue;
1327                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1328                                                 hctx_idx, i);
1329                         tags->rqs[i] = NULL;
1330                 }
1331         }
1332
1333         while (!list_empty(&tags->page_list)) {
1334                 page = list_first_entry(&tags->page_list, struct page, lru);
1335                 list_del_init(&page->lru);
1336                 __free_pages(page, page->private);
1337         }
1338
1339         kfree(tags->rqs);
1340
1341         blk_mq_free_tags(tags);
1342 }
1343
1344 static size_t order_to_size(unsigned int order)
1345 {
1346         return (size_t)PAGE_SIZE << order;
1347 }
1348
1349 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1350                 unsigned int hctx_idx)
1351 {
1352         struct blk_mq_tags *tags;
1353         unsigned int i, j, entries_per_page, max_order = 4;
1354         size_t rq_size, left;
1355
1356         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1357                                 set->numa_node);
1358         if (!tags)
1359                 return NULL;
1360
1361         INIT_LIST_HEAD(&tags->page_list);
1362
1363         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1364                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1365                                  set->numa_node);
1366         if (!tags->rqs) {
1367                 blk_mq_free_tags(tags);
1368                 return NULL;
1369         }
1370
1371         /*
1372          * rq_size is the size of the request plus driver payload, rounded
1373          * to the cacheline size
1374          */
1375         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1376                                 cache_line_size());
1377         left = rq_size * set->queue_depth;
1378
1379         for (i = 0; i < set->queue_depth; ) {
1380                 int this_order = max_order;
1381                 struct page *page;
1382                 int to_do;
1383                 void *p;
1384
1385                 while (left < order_to_size(this_order - 1) && this_order)
1386                         this_order--;
1387
1388                 do {
1389                         page = alloc_pages_node(set->numa_node,
1390                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1391                                 this_order);
1392                         if (page)
1393                                 break;
1394                         if (!this_order--)
1395                                 break;
1396                         if (order_to_size(this_order) < rq_size)
1397                                 break;
1398                 } while (1);
1399
1400                 if (!page)
1401                         goto fail;
1402
1403                 page->private = this_order;
1404                 list_add_tail(&page->lru, &tags->page_list);
1405
1406                 p = page_address(page);
1407                 entries_per_page = order_to_size(this_order) / rq_size;
1408                 to_do = min(entries_per_page, set->queue_depth - i);
1409                 left -= to_do * rq_size;
1410                 for (j = 0; j < to_do; j++) {
1411                         tags->rqs[i] = p;
1412                         tags->rqs[i]->atomic_flags = 0;
1413                         tags->rqs[i]->cmd_flags = 0;
1414                         if (set->ops->init_request) {
1415                                 if (set->ops->init_request(set->driver_data,
1416                                                 tags->rqs[i], hctx_idx, i,
1417                                                 set->numa_node)) {
1418                                         tags->rqs[i] = NULL;
1419                                         goto fail;
1420                                 }
1421                         }
1422
1423                         p += rq_size;
1424                         i++;
1425                 }
1426         }
1427
1428         return tags;
1429
1430 fail:
1431         blk_mq_free_rq_map(set, tags, hctx_idx);
1432         return NULL;
1433 }
1434
1435 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1436 {
1437         kfree(bitmap->map);
1438 }
1439
1440 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1441 {
1442         unsigned int bpw = 8, total, num_maps, i;
1443
1444         bitmap->bits_per_word = bpw;
1445
1446         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1447         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1448                                         GFP_KERNEL, node);
1449         if (!bitmap->map)
1450                 return -ENOMEM;
1451
1452         bitmap->map_size = num_maps;
1453
1454         total = nr_cpu_ids;
1455         for (i = 0; i < num_maps; i++) {
1456                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1457                 total -= bitmap->map[i].depth;
1458         }
1459
1460         return 0;
1461 }
1462
1463 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1464 {
1465         struct request_queue *q = hctx->queue;
1466         struct blk_mq_ctx *ctx;
1467         LIST_HEAD(tmp);
1468
1469         /*
1470          * Move ctx entries to new CPU, if this one is going away.
1471          */
1472         ctx = __blk_mq_get_ctx(q, cpu);
1473
1474         spin_lock(&ctx->lock);
1475         if (!list_empty(&ctx->rq_list)) {
1476                 list_splice_init(&ctx->rq_list, &tmp);
1477                 blk_mq_hctx_clear_pending(hctx, ctx);
1478         }
1479         spin_unlock(&ctx->lock);
1480
1481         if (list_empty(&tmp))
1482                 return NOTIFY_OK;
1483
1484         ctx = blk_mq_get_ctx(q);
1485         spin_lock(&ctx->lock);
1486
1487         while (!list_empty(&tmp)) {
1488                 struct request *rq;
1489
1490                 rq = list_first_entry(&tmp, struct request, queuelist);
1491                 rq->mq_ctx = ctx;
1492                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1493         }
1494
1495         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1496         blk_mq_hctx_mark_pending(hctx, ctx);
1497
1498         spin_unlock(&ctx->lock);
1499
1500         blk_mq_run_hw_queue(hctx, true);
1501         blk_mq_put_ctx(ctx);
1502         return NOTIFY_OK;
1503 }
1504
1505 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1506 {
1507         struct request_queue *q = hctx->queue;
1508         struct blk_mq_tag_set *set = q->tag_set;
1509
1510         if (set->tags[hctx->queue_num])
1511                 return NOTIFY_OK;
1512
1513         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1514         if (!set->tags[hctx->queue_num])
1515                 return NOTIFY_STOP;
1516
1517         hctx->tags = set->tags[hctx->queue_num];
1518         return NOTIFY_OK;
1519 }
1520
1521 static int blk_mq_hctx_notify(void *data, unsigned long action,
1522                               unsigned int cpu)
1523 {
1524         struct blk_mq_hw_ctx *hctx = data;
1525
1526         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1527                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1528         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1529                 return blk_mq_hctx_cpu_online(hctx, cpu);
1530
1531         return NOTIFY_OK;
1532 }
1533
1534 static void blk_mq_exit_hw_queues(struct request_queue *q,
1535                 struct blk_mq_tag_set *set, int nr_queue)
1536 {
1537         struct blk_mq_hw_ctx *hctx;
1538         unsigned int i;
1539
1540         queue_for_each_hw_ctx(q, hctx, i) {
1541                 if (i == nr_queue)
1542                         break;
1543
1544                 blk_mq_tag_idle(hctx);
1545
1546                 if (set->ops->exit_hctx)
1547                         set->ops->exit_hctx(hctx, i);
1548
1549                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1550                 kfree(hctx->ctxs);
1551                 blk_mq_free_bitmap(&hctx->ctx_map);
1552         }
1553
1554 }
1555
1556 static void blk_mq_free_hw_queues(struct request_queue *q,
1557                 struct blk_mq_tag_set *set)
1558 {
1559         struct blk_mq_hw_ctx *hctx;
1560         unsigned int i;
1561
1562         queue_for_each_hw_ctx(q, hctx, i) {
1563                 free_cpumask_var(hctx->cpumask);
1564                 kfree(hctx);
1565         }
1566 }
1567
1568 static int blk_mq_init_hw_queues(struct request_queue *q,
1569                 struct blk_mq_tag_set *set)
1570 {
1571         struct blk_mq_hw_ctx *hctx;
1572         unsigned int i;
1573
1574         /*
1575          * Initialize hardware queues
1576          */
1577         queue_for_each_hw_ctx(q, hctx, i) {
1578                 int node;
1579
1580                 node = hctx->numa_node;
1581                 if (node == NUMA_NO_NODE)
1582                         node = hctx->numa_node = set->numa_node;
1583
1584                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1585                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1586                 spin_lock_init(&hctx->lock);
1587                 INIT_LIST_HEAD(&hctx->dispatch);
1588                 hctx->queue = q;
1589                 hctx->queue_num = i;
1590                 hctx->flags = set->flags;
1591                 hctx->cmd_size = set->cmd_size;
1592
1593                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1594                                                 blk_mq_hctx_notify, hctx);
1595                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1596
1597                 hctx->tags = set->tags[i];
1598
1599                 /*
1600                  * Allocate space for all possible cpus to avoid allocation at
1601                  * runtime
1602                  */
1603                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1604                                                 GFP_KERNEL, node);
1605                 if (!hctx->ctxs)
1606                         break;
1607
1608                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1609                         break;
1610
1611                 hctx->nr_ctx = 0;
1612
1613                 if (set->ops->init_hctx &&
1614                     set->ops->init_hctx(hctx, set->driver_data, i))
1615                         break;
1616         }
1617
1618         if (i == q->nr_hw_queues)
1619                 return 0;
1620
1621         /*
1622          * Init failed
1623          */
1624         blk_mq_exit_hw_queues(q, set, i);
1625
1626         return 1;
1627 }
1628
1629 static void blk_mq_init_cpu_queues(struct request_queue *q,
1630                                    unsigned int nr_hw_queues)
1631 {
1632         unsigned int i;
1633
1634         for_each_possible_cpu(i) {
1635                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1636                 struct blk_mq_hw_ctx *hctx;
1637
1638                 memset(__ctx, 0, sizeof(*__ctx));
1639                 __ctx->cpu = i;
1640                 spin_lock_init(&__ctx->lock);
1641                 INIT_LIST_HEAD(&__ctx->rq_list);
1642                 __ctx->queue = q;
1643
1644                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1645                 if (!cpu_online(i))
1646                         continue;
1647
1648                 hctx = q->mq_ops->map_queue(q, i);
1649                 cpumask_set_cpu(i, hctx->cpumask);
1650                 hctx->nr_ctx++;
1651
1652                 /*
1653                  * Set local node, IFF we have more than one hw queue. If
1654                  * not, we remain on the home node of the device
1655                  */
1656                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1657                         hctx->numa_node = cpu_to_node(i);
1658         }
1659 }
1660
1661 static void blk_mq_map_swqueue(struct request_queue *q)
1662 {
1663         unsigned int i;
1664         struct blk_mq_hw_ctx *hctx;
1665         struct blk_mq_ctx *ctx;
1666
1667         queue_for_each_hw_ctx(q, hctx, i) {
1668                 cpumask_clear(hctx->cpumask);
1669                 hctx->nr_ctx = 0;
1670         }
1671
1672         /*
1673          * Map software to hardware queues
1674          */
1675         queue_for_each_ctx(q, ctx, i) {
1676                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1677                 if (!cpu_online(i))
1678                         continue;
1679
1680                 hctx = q->mq_ops->map_queue(q, i);
1681                 cpumask_set_cpu(i, hctx->cpumask);
1682                 ctx->index_hw = hctx->nr_ctx;
1683                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1684         }
1685
1686         queue_for_each_hw_ctx(q, hctx, i) {
1687                 /*
1688                  * If no software queues are mapped to this hardware queue,
1689                  * disable it and free the request entries.
1690                  */
1691                 if (!hctx->nr_ctx) {
1692                         struct blk_mq_tag_set *set = q->tag_set;
1693
1694                         if (set->tags[i]) {
1695                                 blk_mq_free_rq_map(set, set->tags[i], i);
1696                                 set->tags[i] = NULL;
1697                                 hctx->tags = NULL;
1698                         }
1699                         continue;
1700                 }
1701
1702                 /*
1703                  * Initialize batch roundrobin counts
1704                  */
1705                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1706                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1707         }
1708 }
1709
1710 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1711 {
1712         struct blk_mq_hw_ctx *hctx;
1713         struct request_queue *q;
1714         bool shared;
1715         int i;
1716
1717         if (set->tag_list.next == set->tag_list.prev)
1718                 shared = false;
1719         else
1720                 shared = true;
1721
1722         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1723                 blk_mq_freeze_queue(q);
1724
1725                 queue_for_each_hw_ctx(q, hctx, i) {
1726                         if (shared)
1727                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1728                         else
1729                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1730                 }
1731                 blk_mq_unfreeze_queue(q);
1732         }
1733 }
1734
1735 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1736 {
1737         struct blk_mq_tag_set *set = q->tag_set;
1738
1739         mutex_lock(&set->tag_list_lock);
1740         list_del_init(&q->tag_set_list);
1741         blk_mq_update_tag_set_depth(set);
1742         mutex_unlock(&set->tag_list_lock);
1743 }
1744
1745 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1746                                      struct request_queue *q)
1747 {
1748         q->tag_set = set;
1749
1750         mutex_lock(&set->tag_list_lock);
1751         list_add_tail(&q->tag_set_list, &set->tag_list);
1752         blk_mq_update_tag_set_depth(set);
1753         mutex_unlock(&set->tag_list_lock);
1754 }
1755
1756 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1757 {
1758         struct blk_mq_hw_ctx **hctxs;
1759         struct blk_mq_ctx __percpu *ctx;
1760         struct request_queue *q;
1761         unsigned int *map;
1762         int i;
1763
1764         ctx = alloc_percpu(struct blk_mq_ctx);
1765         if (!ctx)
1766                 return ERR_PTR(-ENOMEM);
1767
1768         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1769                         set->numa_node);
1770
1771         if (!hctxs)
1772                 goto err_percpu;
1773
1774         map = blk_mq_make_queue_map(set);
1775         if (!map)
1776                 goto err_map;
1777
1778         for (i = 0; i < set->nr_hw_queues; i++) {
1779                 int node = blk_mq_hw_queue_to_node(map, i);
1780
1781                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1782                                         GFP_KERNEL, node);
1783                 if (!hctxs[i])
1784                         goto err_hctxs;
1785
1786                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1787                         goto err_hctxs;
1788
1789                 atomic_set(&hctxs[i]->nr_active, 0);
1790                 hctxs[i]->numa_node = node;
1791                 hctxs[i]->queue_num = i;
1792         }
1793
1794         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1795         if (!q)
1796                 goto err_hctxs;
1797
1798         /*
1799          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1800          * See blk_register_queue() for details.
1801          */
1802         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1803                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1804                 goto err_map;
1805
1806         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1807         blk_queue_rq_timeout(q, 30000);
1808
1809         q->nr_queues = nr_cpu_ids;
1810         q->nr_hw_queues = set->nr_hw_queues;
1811         q->mq_map = map;
1812
1813         q->queue_ctx = ctx;
1814         q->queue_hw_ctx = hctxs;
1815
1816         q->mq_ops = set->ops;
1817         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1818
1819         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1820                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1821
1822         q->sg_reserved_size = INT_MAX;
1823
1824         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1825         INIT_LIST_HEAD(&q->requeue_list);
1826         spin_lock_init(&q->requeue_lock);
1827
1828         if (q->nr_hw_queues > 1)
1829                 blk_queue_make_request(q, blk_mq_make_request);
1830         else
1831                 blk_queue_make_request(q, blk_sq_make_request);
1832
1833         blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1834         if (set->timeout)
1835                 blk_queue_rq_timeout(q, set->timeout);
1836
1837         /*
1838          * Do this after blk_queue_make_request() overrides it...
1839          */
1840         q->nr_requests = set->queue_depth;
1841
1842         if (set->ops->complete)
1843                 blk_queue_softirq_done(q, set->ops->complete);
1844
1845         blk_mq_init_flush(q);
1846         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1847
1848         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1849                                 set->cmd_size, cache_line_size()),
1850                                 GFP_KERNEL);
1851         if (!q->flush_rq)
1852                 goto err_hw;
1853
1854         if (blk_mq_init_hw_queues(q, set))
1855                 goto err_flush_rq;
1856
1857         mutex_lock(&all_q_mutex);
1858         list_add_tail(&q->all_q_node, &all_q_list);
1859         mutex_unlock(&all_q_mutex);
1860
1861         blk_mq_add_queue_tag_set(set, q);
1862
1863         blk_mq_map_swqueue(q);
1864
1865         return q;
1866
1867 err_flush_rq:
1868         kfree(q->flush_rq);
1869 err_hw:
1870         blk_cleanup_queue(q);
1871 err_hctxs:
1872         kfree(map);
1873         for (i = 0; i < set->nr_hw_queues; i++) {
1874                 if (!hctxs[i])
1875                         break;
1876                 free_cpumask_var(hctxs[i]->cpumask);
1877                 kfree(hctxs[i]);
1878         }
1879 err_map:
1880         kfree(hctxs);
1881 err_percpu:
1882         free_percpu(ctx);
1883         return ERR_PTR(-ENOMEM);
1884 }
1885 EXPORT_SYMBOL(blk_mq_init_queue);
1886
1887 void blk_mq_free_queue(struct request_queue *q)
1888 {
1889         struct blk_mq_tag_set   *set = q->tag_set;
1890
1891         blk_mq_del_queue_tag_set(q);
1892
1893         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1894         blk_mq_free_hw_queues(q, set);
1895
1896         percpu_ref_exit(&q->mq_usage_counter);
1897
1898         free_percpu(q->queue_ctx);
1899         kfree(q->queue_hw_ctx);
1900         kfree(q->mq_map);
1901
1902         q->queue_ctx = NULL;
1903         q->queue_hw_ctx = NULL;
1904         q->mq_map = NULL;
1905
1906         mutex_lock(&all_q_mutex);
1907         list_del_init(&q->all_q_node);
1908         mutex_unlock(&all_q_mutex);
1909 }
1910
1911 /* Basically redo blk_mq_init_queue with queue frozen */
1912 static void blk_mq_queue_reinit(struct request_queue *q)
1913 {
1914         blk_mq_freeze_queue(q);
1915
1916         blk_mq_sysfs_unregister(q);
1917
1918         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1919
1920         /*
1921          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1922          * we should change hctx numa_node according to new topology (this
1923          * involves free and re-allocate memory, worthy doing?)
1924          */
1925
1926         blk_mq_map_swqueue(q);
1927
1928         blk_mq_sysfs_register(q);
1929
1930         blk_mq_unfreeze_queue(q);
1931 }
1932
1933 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1934                                       unsigned long action, void *hcpu)
1935 {
1936         struct request_queue *q;
1937
1938         /*
1939          * Before new mappings are established, hotadded cpu might already
1940          * start handling requests. This doesn't break anything as we map
1941          * offline CPUs to first hardware queue. We will re-init the queue
1942          * below to get optimal settings.
1943          */
1944         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1945             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1946                 return NOTIFY_OK;
1947
1948         mutex_lock(&all_q_mutex);
1949         list_for_each_entry(q, &all_q_list, all_q_node)
1950                 blk_mq_queue_reinit(q);
1951         mutex_unlock(&all_q_mutex);
1952         return NOTIFY_OK;
1953 }
1954
1955 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1956 {
1957         int i;
1958
1959         for (i = 0; i < set->nr_hw_queues; i++) {
1960                 set->tags[i] = blk_mq_init_rq_map(set, i);
1961                 if (!set->tags[i])
1962                         goto out_unwind;
1963         }
1964
1965         return 0;
1966
1967 out_unwind:
1968         while (--i >= 0)
1969                 blk_mq_free_rq_map(set, set->tags[i], i);
1970
1971         return -ENOMEM;
1972 }
1973
1974 /*
1975  * Allocate the request maps associated with this tag_set. Note that this
1976  * may reduce the depth asked for, if memory is tight. set->queue_depth
1977  * will be updated to reflect the allocated depth.
1978  */
1979 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1980 {
1981         unsigned int depth;
1982         int err;
1983
1984         depth = set->queue_depth;
1985         do {
1986                 err = __blk_mq_alloc_rq_maps(set);
1987                 if (!err)
1988                         break;
1989
1990                 set->queue_depth >>= 1;
1991                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1992                         err = -ENOMEM;
1993                         break;
1994                 }
1995         } while (set->queue_depth);
1996
1997         if (!set->queue_depth || err) {
1998                 pr_err("blk-mq: failed to allocate request map\n");
1999                 return -ENOMEM;
2000         }
2001
2002         if (depth != set->queue_depth)
2003                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2004                                                 depth, set->queue_depth);
2005
2006         return 0;
2007 }
2008
2009 /*
2010  * Alloc a tag set to be associated with one or more request queues.
2011  * May fail with EINVAL for various error conditions. May adjust the
2012  * requested depth down, if if it too large. In that case, the set
2013  * value will be stored in set->queue_depth.
2014  */
2015 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2016 {
2017         if (!set->nr_hw_queues)
2018                 return -EINVAL;
2019         if (!set->queue_depth)
2020                 return -EINVAL;
2021         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2022                 return -EINVAL;
2023
2024         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2025                 return -EINVAL;
2026
2027         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2028                 pr_info("blk-mq: reduced tag depth to %u\n",
2029                         BLK_MQ_MAX_DEPTH);
2030                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2031         }
2032
2033         set->tags = kmalloc_node(set->nr_hw_queues *
2034                                  sizeof(struct blk_mq_tags *),
2035                                  GFP_KERNEL, set->numa_node);
2036         if (!set->tags)
2037                 return -ENOMEM;
2038
2039         if (blk_mq_alloc_rq_maps(set))
2040                 goto enomem;
2041
2042         mutex_init(&set->tag_list_lock);
2043         INIT_LIST_HEAD(&set->tag_list);
2044
2045         return 0;
2046 enomem:
2047         kfree(set->tags);
2048         set->tags = NULL;
2049         return -ENOMEM;
2050 }
2051 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2052
2053 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2054 {
2055         int i;
2056
2057         for (i = 0; i < set->nr_hw_queues; i++) {
2058                 if (set->tags[i])
2059                         blk_mq_free_rq_map(set, set->tags[i], i);
2060         }
2061
2062         kfree(set->tags);
2063         set->tags = NULL;
2064 }
2065 EXPORT_SYMBOL(blk_mq_free_tag_set);
2066
2067 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2068 {
2069         struct blk_mq_tag_set *set = q->tag_set;
2070         struct blk_mq_hw_ctx *hctx;
2071         int i, ret;
2072
2073         if (!set || nr > set->queue_depth)
2074                 return -EINVAL;
2075
2076         ret = 0;
2077         queue_for_each_hw_ctx(q, hctx, i) {
2078                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2079                 if (ret)
2080                         break;
2081         }
2082
2083         if (!ret)
2084                 q->nr_requests = nr;
2085
2086         return ret;
2087 }
2088
2089 void blk_mq_disable_hotplug(void)
2090 {
2091         mutex_lock(&all_q_mutex);
2092 }
2093
2094 void blk_mq_enable_hotplug(void)
2095 {
2096         mutex_unlock(&all_q_mutex);
2097 }
2098
2099 static int __init blk_mq_init(void)
2100 {
2101         blk_mq_cpu_init();
2102
2103         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2104
2105         return 0;
2106 }
2107 subsys_initcall(blk_mq_init);