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