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                 /*
 123                  * XXX: Temporary kludge to work around SCSI blk-mq stall.
 124                  * SCSI synchronously creates and destroys many queues
 125                  * back-to-back during probe leading to lengthy stalls.
 126                  * This will be fixed by keeping ->mq_usage_counter in
 127                  * atomic mode until genhd registration, but, for now,
 128                  * let's work around using expedited synchronization.
 129                  */
 130                 __percpu_ref_kill_expedited(&q->mq_usage_counter);
 131
 132                 blk_mq_run_queues(q, false);
 133         }
 134         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 135 }
 136
 137 static void blk_mq_unfreeze_queue(struct request_queue *q)
 138 {
 139         bool wake;
 140
 141         spin_lock_irq(q->queue_lock);
 142         wake = !--q->mq_freeze_depth;
 143         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 144         spin_unlock_irq(q->queue_lock);
 145         if (wake) {
 146                 percpu_ref_reinit(&q->mq_usage_counter);
 147                 wake_up_all(&q->mq_freeze_wq);
 148         }
 149 }
 150
 151 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 152 {
 153         return blk_mq_has_free_tags(hctx->tags);
 154 }
 155 EXPORT_SYMBOL(blk_mq_can_queue);
 156
 157 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 158                                struct request *rq, unsigned int rw_flags)
 159 {
 160         if (blk_queue_io_stat(q))
 161                 rw_flags |= REQ_IO_STAT;
 162
 163         INIT_LIST_HEAD(&rq->queuelist);
 164         /* csd/requeue_work/fifo_time is initialized before use */
 165         rq->q = q;
 166         rq->mq_ctx = ctx;
 167         rq->cmd_flags |= rw_flags;
 168         /* do not touch atomic flags, it needs atomic ops against the timer */
 169         rq->cpu = -1;
 170         INIT_HLIST_NODE(&rq->hash);
 171         RB_CLEAR_NODE(&rq->rb_node);
 172         rq->rq_disk = NULL;
 173         rq->part = NULL;
 174         rq->start_time = jiffies;
 175 #ifdef CONFIG_BLK_CGROUP
 176         rq->rl = NULL;
 177         set_start_time_ns(rq);
 178         rq->io_start_time_ns = 0;
 179 #endif
 180         rq->nr_phys_segments = 0;
 181 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 182         rq->nr_integrity_segments = 0;
 183 #endif
 184         rq->special = NULL;
 185         /* tag was already set */
 186         rq->errors = 0;
 187
 188         rq->cmd = rq->__cmd;
 189
 190         rq->extra_len = 0;
 191         rq->sense_len = 0;
 192         rq->resid_len = 0;
 193         rq->sense = NULL;
 194
 195         INIT_LIST_HEAD(&rq->timeout_list);
 196         rq->timeout = 0;
 197
 198         rq->end_io = NULL;
 199         rq->end_io_data = NULL;
 200         rq->next_rq = NULL;
 201
 202         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 203 }
 204
 205 static struct request *
 206 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 207 {
 208         struct request *rq;
 209         unsigned int tag;
 210
 211         tag = blk_mq_get_tag(data);
 212         if (tag != BLK_MQ_TAG_FAIL) {
 213                 rq = data->hctx->tags->rqs[tag];
 214
 215                 if (blk_mq_tag_busy(data->hctx)) {
 216                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 217                         atomic_inc(&data->hctx->nr_active);
 218                 }
 219
 220                 rq->tag = tag;
 221                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 222                 return rq;
 223         }
 224
 225         return NULL;
 226 }
 227
 228 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 229                 bool reserved)
 230 {
 231         struct blk_mq_ctx *ctx;
 232         struct blk_mq_hw_ctx *hctx;
 233         struct request *rq;
 234         struct blk_mq_alloc_data alloc_data;
 235
 236         if (blk_mq_queue_enter(q))
 237                 return NULL;
 238
 239         ctx = blk_mq_get_ctx(q);
 240         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 241         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 242                         reserved, ctx, hctx);
 243
 244         rq = __blk_mq_alloc_request(&alloc_data, rw);
 245         if (!rq && (gfp & __GFP_WAIT)) {
 246                 __blk_mq_run_hw_queue(hctx);
 247                 blk_mq_put_ctx(ctx);
 248
 249                 ctx = blk_mq_get_ctx(q);
 250                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 251                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 252                                 hctx);
 253                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 254                 ctx = alloc_data.ctx;
 255         }
 256         blk_mq_put_ctx(ctx);
 257         return rq;
 258 }
 259 EXPORT_SYMBOL(blk_mq_alloc_request);
 260
 261 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 262                                   struct blk_mq_ctx *ctx, struct request *rq)
 263 {
 264         const int tag = rq->tag;
 265         struct request_queue *q = rq->q;
 266
 267         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 268                 atomic_dec(&hctx->nr_active);
 269         rq->cmd_flags = 0;
 270
 271         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 272         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 273         blk_mq_queue_exit(q);
 274 }
 275
 276 void blk_mq_free_request(struct request *rq)
 277 {
 278         struct blk_mq_ctx *ctx = rq->mq_ctx;
 279         struct blk_mq_hw_ctx *hctx;
 280         struct request_queue *q = rq->q;
 281
 282         ctx->rq_completed[rq_is_sync(rq)]++;
 283
 284         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 285         __blk_mq_free_request(hctx, ctx, rq);
 286 }
 287
 288 /*
 289  * Clone all relevant state from a request that has been put on hold in
 290  * the flush state machine into the preallocated flush request that hangs
 291  * off the request queue.
 292  *
 293  * For a driver the flush request should be invisible, that's why we are
 294  * impersonating the original request here.
 295  */
 296 void blk_mq_clone_flush_request(struct request *flush_rq,
 297                 struct request *orig_rq)
 298 {
 299         struct blk_mq_hw_ctx *hctx =
 300                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 301
 302         flush_rq->mq_ctx = orig_rq->mq_ctx;
 303         flush_rq->tag = orig_rq->tag;
 304         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 305                 hctx->cmd_size);
 306 }
 307
 308 inline void __blk_mq_end_io(struct request *rq, int error)
 309 {
 310         blk_account_io_done(rq);
 311
 312         if (rq->end_io) {
 313                 rq->end_io(rq, error);
 314         } else {
 315                 if (unlikely(blk_bidi_rq(rq)))
 316                         blk_mq_free_request(rq->next_rq);
 317                 blk_mq_free_request(rq);
 318         }
 319 }
 320 EXPORT_SYMBOL(__blk_mq_end_io);
 321
 322 void blk_mq_end_io(struct request *rq, int error)
 323 {
 324         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 325                 BUG();
 326         __blk_mq_end_io(rq, error);
 327 }
 328 EXPORT_SYMBOL(blk_mq_end_io);
 329
 330 static void __blk_mq_complete_request_remote(void *data)
 331 {
 332         struct request *rq = data;
 333
 334         rq->q->softirq_done_fn(rq);
 335 }
 336
 337 static void blk_mq_ipi_complete_request(struct request *rq)
 338 {
 339         struct blk_mq_ctx *ctx = rq->mq_ctx;
 340         bool shared = false;
 341         int cpu;
 342
 343         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 344                 rq->q->softirq_done_fn(rq);
 345                 return;
 346         }
 347
 348         cpu = get_cpu();
 349         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 350                 shared = cpus_share_cache(cpu, ctx->cpu);
 351
 352         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 353                 rq->csd.func = __blk_mq_complete_request_remote;
 354                 rq->csd.info = rq;
 355                 rq->csd.flags = 0;
 356                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 357         } else {
 358                 rq->q->softirq_done_fn(rq);
 359         }
 360         put_cpu();
 361 }
 362
 363 void __blk_mq_complete_request(struct request *rq)
 364 {
 365         struct request_queue *q = rq->q;
 366
 367         if (!q->softirq_done_fn)
 368                 blk_mq_end_io(rq, rq->errors);
 369         else
 370                 blk_mq_ipi_complete_request(rq);
 371 }
 372
 373 /**
 374  * blk_mq_complete_request - end I/O on a request
 375  * @rq:         the request being processed
 376  *
 377  * Description:
 378  *      Ends all I/O on a request. It does not handle partial completions.
 379  *      The actual completion happens out-of-order, through a IPI handler.
 380  **/
 381 void blk_mq_complete_request(struct request *rq)
 382 {
 383         struct request_queue *q = rq->q;
 384
 385         if (unlikely(blk_should_fake_timeout(q)))
 386                 return;
 387         if (!blk_mark_rq_complete(rq))
 388                 __blk_mq_complete_request(rq);
 389 }
 390 EXPORT_SYMBOL(blk_mq_complete_request);
 391
 392 static void blk_mq_start_request(struct request *rq, bool last)
 393 {
 394         struct request_queue *q = rq->q;
 395
 396         trace_block_rq_issue(q, rq);
 397
 398         rq->resid_len = blk_rq_bytes(rq);
 399         if (unlikely(blk_bidi_rq(rq)))
 400                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 401
 402         blk_add_timer(rq);
 403
 404         /*
 405          * Ensure that ->deadline is visible before set the started
 406          * flag and clear the completed flag.
 407          */
 408         smp_mb__before_atomic();
 409
 410         /*
 411          * Mark us as started and clear complete. Complete might have been
 412          * set if requeue raced with timeout, which then marked it as
 413          * complete. So be sure to clear complete again when we start
 414          * the request, otherwise we'll ignore the completion event.
 415          */
 416         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 417                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 418         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 419                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 420
 421         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 422                 /*
 423                  * Make sure space for the drain appears.  We know we can do
 424                  * this because max_hw_segments has been adjusted to be one
 425                  * fewer than the device can handle.
 426                  */
 427                 rq->nr_phys_segments++;
 428         }
 429
 430         /*
 431          * Flag the last request in the series so that drivers know when IO
 432          * should be kicked off, if they don't do it on a per-request basis.
 433          *
 434          * Note: the flag isn't the only condition drivers should do kick off.
 435          * If drive is busy, the last request might not have the bit set.
 436          */
 437         if (last)
 438                 rq->cmd_flags |= REQ_END;
 439 }
 440
 441 static void __blk_mq_requeue_request(struct request *rq)
 442 {
 443         struct request_queue *q = rq->q;
 444
 445         trace_block_rq_requeue(q, rq);
 446         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 447
 448         rq->cmd_flags &= ~REQ_END;
 449
 450         if (q->dma_drain_size && blk_rq_bytes(rq))
 451                 rq->nr_phys_segments--;
 452 }
 453
 454 void blk_mq_requeue_request(struct request *rq)
 455 {
 456         __blk_mq_requeue_request(rq);
 457         blk_clear_rq_complete(rq);
 458
 459         BUG_ON(blk_queued_rq(rq));
 460         blk_mq_add_to_requeue_list(rq, true);
 461 }
 462 EXPORT_SYMBOL(blk_mq_requeue_request);
 463
 464 static void blk_mq_requeue_work(struct work_struct *work)
 465 {
 466         struct request_queue *q =
 467                 container_of(work, struct request_queue, requeue_work);
 468         LIST_HEAD(rq_list);
 469         struct request *rq, *next;
 470         unsigned long flags;
 471
 472         spin_lock_irqsave(&q->requeue_lock, flags);
 473         list_splice_init(&q->requeue_list, &rq_list);
 474         spin_unlock_irqrestore(&q->requeue_lock, flags);
 475
 476         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 477                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 478                         continue;
 479
 480                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 481                 list_del_init(&rq->queuelist);
 482                 blk_mq_insert_request(rq, true, false, false);
 483         }
 484
 485         while (!list_empty(&rq_list)) {
 486                 rq = list_entry(rq_list.next, struct request, queuelist);
 487                 list_del_init(&rq->queuelist);
 488                 blk_mq_insert_request(rq, false, false, false);
 489         }
 490
 491         /*
 492          * Use the start variant of queue running here, so that running
 493          * the requeue work will kick stopped queues.
 494          */
 495         blk_mq_start_hw_queues(q);
 496 }
 497
 498 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 499 {
 500         struct request_queue *q = rq->q;
 501         unsigned long flags;
 502
 503         /*
 504          * We abuse this flag that is otherwise used by the I/O scheduler to
 505          * request head insertation from the workqueue.
 506          */
 507         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 508
 509         spin_lock_irqsave(&q->requeue_lock, flags);
 510         if (at_head) {
 511                 rq->cmd_flags |= REQ_SOFTBARRIER;
 512                 list_add(&rq->queuelist, &q->requeue_list);
 513         } else {
 514                 list_add_tail(&rq->queuelist, &q->requeue_list);
 515         }
 516         spin_unlock_irqrestore(&q->requeue_lock, flags);
 517 }
 518 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 519
 520 void blk_mq_kick_requeue_list(struct request_queue *q)
 521 {
 522         kblockd_schedule_work(&q->requeue_work);
 523 }
 524 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 525
 526 static inline bool is_flush_request(struct request *rq, unsigned int tag)
 527 {
 528         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 529                         rq->q->flush_rq->tag == tag);
 530 }
 531
 532 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 533 {
 534         struct request *rq = tags->rqs[tag];
 535
 536         if (!is_flush_request(rq, tag))
 537                 return rq;
 538
 539         return rq->q->flush_rq;
 540 }
 541 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 542
 543 struct blk_mq_timeout_data {
 544         struct blk_mq_hw_ctx *hctx;
 545         unsigned long *next;
 546         unsigned int *next_set;
 547 };
 548
 549 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
 550 {
 551         struct blk_mq_timeout_data *data = __data;
 552         struct blk_mq_hw_ctx *hctx = data->hctx;
 553         unsigned int tag;
 554
 555          /* It may not be in flight yet (this is where
 556          * the REQ_ATOMIC_STARTED flag comes in). The requests are
 557          * statically allocated, so we know it's always safe to access the
 558          * memory associated with a bit offset into ->rqs[].
 559          */
 560         tag = 0;
 561         do {
 562                 struct request *rq;
 563
 564                 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
 565                 if (tag >= hctx->tags->nr_tags)
 566                         break;
 567
 568                 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
 569                 if (rq->q != hctx->queue)
 570                         continue;
 571                 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 572                         continue;
 573
 574                 blk_rq_check_expired(rq, data->next, data->next_set);
 575         } while (1);
 576 }
 577
 578 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
 579                                         unsigned long *next,
 580                                         unsigned int *next_set)
 581 {
 582         struct blk_mq_timeout_data data = {
 583                 .hctx           = hctx,
 584                 .next           = next,
 585                 .next_set       = next_set,
 586         };
 587
 588         /*
 589          * Ask the tagging code to iterate busy requests, so we can
 590          * check them for timeout.
 591          */
 592         blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
 593 }
 594
 595 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
 596 {
 597         struct request_queue *q = rq->q;
 598
 599         /*
 600          * We know that complete is set at this point. If STARTED isn't set
 601          * anymore, then the request isn't active and the "timeout" should
 602          * just be ignored. This can happen due to the bitflag ordering.
 603          * Timeout first checks if STARTED is set, and if it is, assumes
 604          * the request is active. But if we race with completion, then
 605          * we both flags will get cleared. So check here again, and ignore
 606          * a timeout event with a request that isn't active.
 607          */
 608         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 609                 return BLK_EH_NOT_HANDLED;
 610
 611         if (!q->mq_ops->timeout)
 612                 return BLK_EH_RESET_TIMER;
 613
 614         return q->mq_ops->timeout(rq);
 615 }
 616
 617 static void blk_mq_rq_timer(unsigned long data)
 618 {
 619         struct request_queue *q = (struct request_queue *) data;
 620         struct blk_mq_hw_ctx *hctx;
 621         unsigned long next = 0;
 622         int i, next_set = 0;
 623
 624         queue_for_each_hw_ctx(q, hctx, i) {
 625                 /*
 626                  * If not software queues are currently mapped to this
 627                  * hardware queue, there's nothing to check
 628                  */
 629                 if (!hctx->nr_ctx || !hctx->tags)
 630                         continue;
 631
 632                 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
 633         }
 634
 635         if (next_set) {
 636                 next = blk_rq_timeout(round_jiffies_up(next));
 637                 mod_timer(&q->timeout, next);
 638         } else {
 639                 queue_for_each_hw_ctx(q, hctx, i)
 640                         blk_mq_tag_idle(hctx);
 641         }
 642 }
 643
 644 /*
 645  * Reverse check our software queue for entries that we could potentially
 646  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 647  * too much time checking for merges.
 648  */
 649 static bool blk_mq_attempt_merge(struct request_queue *q,
 650                                  struct blk_mq_ctx *ctx, struct bio *bio)
 651 {
 652         struct request *rq;
 653         int checked = 8;
 654
 655         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 656                 int el_ret;
 657
 658                 if (!checked--)
 659                         break;
 660
 661                 if (!blk_rq_merge_ok(rq, bio))
 662                         continue;
 663
 664                 el_ret = blk_try_merge(rq, bio);
 665                 if (el_ret == ELEVATOR_BACK_MERGE) {
 666                         if (bio_attempt_back_merge(q, rq, bio)) {
 667                                 ctx->rq_merged++;
 668                                 return true;
 669                         }
 670                         break;
 671                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 672                         if (bio_attempt_front_merge(q, rq, bio)) {
 673                                 ctx->rq_merged++;
 674                                 return true;
 675                         }
 676                         break;
 677                 }
 678         }
 679
 680         return false;
 681 }
 682
 683 /*
 684  * Process software queues that have been marked busy, splicing them
 685  * to the for-dispatch
 686  */
 687 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 688 {
 689         struct blk_mq_ctx *ctx;
 690         int i;
 691
 692         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 693                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 694                 unsigned int off, bit;
 695
 696                 if (!bm->word)
 697                         continue;
 698
 699                 bit = 0;
 700                 off = i * hctx->ctx_map.bits_per_word;
 701                 do {
 702                         bit = find_next_bit(&bm->word, bm->depth, bit);
 703                         if (bit >= bm->depth)
 704                                 break;
 705
 706                         ctx = hctx->ctxs[bit + off];
 707                         clear_bit(bit, &bm->word);
 708                         spin_lock(&ctx->lock);
 709                         list_splice_tail_init(&ctx->rq_list, list);
 710                         spin_unlock(&ctx->lock);
 711
 712                         bit++;
 713                 } while (1);
 714         }
 715 }
 716
 717 /*
 718  * Run this hardware queue, pulling any software queues mapped to it in.
 719  * Note that this function currently has various problems around ordering
 720  * of IO. In particular, we'd like FIFO behaviour on handling existing
 721  * items on the hctx->dispatch list. Ignore that for now.
 722  */
 723 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 724 {
 725         struct request_queue *q = hctx->queue;
 726         struct request *rq;
 727         LIST_HEAD(rq_list);
 728         int queued;
 729
 730         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 731
 732         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 733                 return;
 734
 735         hctx->run++;
 736
 737         /*
 738          * Touch any software queue that has pending entries.
 739          */
 740         flush_busy_ctxs(hctx, &rq_list);
 741
 742         /*
 743          * If we have previous entries on our dispatch list, grab them
 744          * and stuff them at the front for more fair dispatch.
 745          */
 746         if (!list_empty_careful(&hctx->dispatch)) {
 747                 spin_lock(&hctx->lock);
 748                 if (!list_empty(&hctx->dispatch))
 749                         list_splice_init(&hctx->dispatch, &rq_list);
 750                 spin_unlock(&hctx->lock);
 751         }
 752
 753         /*
 754          * Now process all the entries, sending them to the driver.
 755          */
 756         queued = 0;
 757         while (!list_empty(&rq_list)) {
 758                 int ret;
 759
 760                 rq = list_first_entry(&rq_list, struct request, queuelist);
 761                 list_del_init(&rq->queuelist);
 762
 763                 blk_mq_start_request(rq, list_empty(&rq_list));
 764
 765                 ret = q->mq_ops->queue_rq(hctx, rq);
 766                 switch (ret) {
 767                 case BLK_MQ_RQ_QUEUE_OK:
 768                         queued++;
 769                         continue;
 770                 case BLK_MQ_RQ_QUEUE_BUSY:
 771                         list_add(&rq->queuelist, &rq_list);
 772                         __blk_mq_requeue_request(rq);
 773                         break;
 774                 default:
 775                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 776                 case BLK_MQ_RQ_QUEUE_ERROR:
 777                         rq->errors = -EIO;
 778                         blk_mq_end_io(rq, rq->errors);
 779                         break;
 780                 }
 781
 782                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 783                         break;
 784         }
 785
 786         if (!queued)
 787                 hctx->dispatched[0]++;
 788         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 789                 hctx->dispatched[ilog2(queued) + 1]++;
 790
 791         /*
 792          * Any items that need requeuing? Stuff them into hctx->dispatch,
 793          * that is where we will continue on next queue run.
 794          */
 795         if (!list_empty(&rq_list)) {
 796                 spin_lock(&hctx->lock);
 797                 list_splice(&rq_list, &hctx->dispatch);
 798                 spin_unlock(&hctx->lock);
 799         }
 800 }
 801
 802 /*
 803  * It'd be great if the workqueue API had a way to pass
 804  * in a mask and had some smarts for more clever placement.
 805  * For now we just round-robin here, switching for every
 806  * BLK_MQ_CPU_WORK_BATCH queued items.
 807  */
 808 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 809 {
 810         int cpu = hctx->next_cpu;
 811
 812         if (--hctx->next_cpu_batch <= 0) {
 813                 int next_cpu;
 814
 815                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 816                 if (next_cpu >= nr_cpu_ids)
 817                         next_cpu = cpumask_first(hctx->cpumask);
 818
 819                 hctx->next_cpu = next_cpu;
 820                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 821         }
 822
 823         return cpu;
 824 }
 825
 826 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 827 {
 828         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 829                 return;
 830
 831         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 832                 __blk_mq_run_hw_queue(hctx);
 833         else if (hctx->queue->nr_hw_queues == 1)
 834                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 835         else {
 836                 unsigned int cpu;
 837
 838                 cpu = blk_mq_hctx_next_cpu(hctx);
 839                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 840         }
 841 }
 842
 843 void blk_mq_run_queues(struct request_queue *q, bool async)
 844 {
 845         struct blk_mq_hw_ctx *hctx;
 846         int i;
 847
 848         queue_for_each_hw_ctx(q, hctx, i) {
 849                 if ((!blk_mq_hctx_has_pending(hctx) &&
 850                     list_empty_careful(&hctx->dispatch)) ||
 851                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 852                         continue;
 853
 854                 preempt_disable();
 855                 blk_mq_run_hw_queue(hctx, async);
 856                 preempt_enable();
 857         }
 858 }
 859 EXPORT_SYMBOL(blk_mq_run_queues);
 860
 861 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 862 {
 863         cancel_delayed_work(&hctx->run_work);
 864         cancel_delayed_work(&hctx->delay_work);
 865         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 866 }
 867 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 868
 869 void blk_mq_stop_hw_queues(struct request_queue *q)
 870 {
 871         struct blk_mq_hw_ctx *hctx;
 872         int i;
 873
 874         queue_for_each_hw_ctx(q, hctx, i)
 875                 blk_mq_stop_hw_queue(hctx);
 876 }
 877 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 878
 879 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 880 {
 881         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 882
 883         preempt_disable();
 884         blk_mq_run_hw_queue(hctx, false);
 885         preempt_enable();
 886 }
 887 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 888
 889 void blk_mq_start_hw_queues(struct request_queue *q)
 890 {
 891         struct blk_mq_hw_ctx *hctx;
 892         int i;
 893
 894         queue_for_each_hw_ctx(q, hctx, i)
 895                 blk_mq_start_hw_queue(hctx);
 896 }
 897 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 898
 899
 900 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 901 {
 902         struct blk_mq_hw_ctx *hctx;
 903         int i;
 904
 905         queue_for_each_hw_ctx(q, hctx, i) {
 906                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 907                         continue;
 908
 909                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 910                 preempt_disable();
 911                 blk_mq_run_hw_queue(hctx, async);
 912                 preempt_enable();
 913         }
 914 }
 915 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 916
 917 static void blk_mq_run_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, run_work.work);
 922
 923         __blk_mq_run_hw_queue(hctx);
 924 }
 925
 926 static void blk_mq_delay_work_fn(struct work_struct *work)
 927 {
 928         struct blk_mq_hw_ctx *hctx;
 929
 930         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 931
 932         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 933                 __blk_mq_run_hw_queue(hctx);
 934 }
 935
 936 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 937 {
 938         unsigned long tmo = msecs_to_jiffies(msecs);
 939
 940         if (hctx->queue->nr_hw_queues == 1)
 941                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 942         else {
 943                 unsigned int cpu;
 944
 945                 cpu = blk_mq_hctx_next_cpu(hctx);
 946                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 947         }
 948 }
 949 EXPORT_SYMBOL(blk_mq_delay_queue);
 950
 951 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 952                                     struct request *rq, bool at_head)
 953 {
 954         struct blk_mq_ctx *ctx = rq->mq_ctx;
 955
 956         trace_block_rq_insert(hctx->queue, rq);
 957
 958         if (at_head)
 959                 list_add(&rq->queuelist, &ctx->rq_list);
 960         else
 961                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 962
 963         blk_mq_hctx_mark_pending(hctx, ctx);
 964 }
 965
 966 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 967                 bool async)
 968 {
 969         struct request_queue *q = rq->q;
 970         struct blk_mq_hw_ctx *hctx;
 971         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 972
 973         current_ctx = blk_mq_get_ctx(q);
 974         if (!cpu_online(ctx->cpu))
 975                 rq->mq_ctx = ctx = current_ctx;
 976
 977         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 978
 979         spin_lock(&ctx->lock);
 980         __blk_mq_insert_request(hctx, rq, at_head);
 981         spin_unlock(&ctx->lock);
 982
 983         if (run_queue)
 984                 blk_mq_run_hw_queue(hctx, async);
 985
 986         blk_mq_put_ctx(current_ctx);
 987 }
 988
 989 static void blk_mq_insert_requests(struct request_queue *q,
 990                                      struct blk_mq_ctx *ctx,
 991                                      struct list_head *list,
 992                                      int depth,
 993                                      bool from_schedule)
 994
 995 {
 996         struct blk_mq_hw_ctx *hctx;
 997         struct blk_mq_ctx *current_ctx;
 998
 999         trace_block_unplug(q, depth, !from_schedule);
1000
1001         current_ctx = blk_mq_get_ctx(q);
1002
1003         if (!cpu_online(ctx->cpu))
1004                 ctx = current_ctx;
1005         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1006
1007         /*
1008          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1009          * offline now
1010          */
1011         spin_lock(&ctx->lock);
1012         while (!list_empty(list)) {
1013                 struct request *rq;
1014
1015                 rq = list_first_entry(list, struct request, queuelist);
1016                 list_del_init(&rq->queuelist);
1017                 rq->mq_ctx = ctx;
1018                 __blk_mq_insert_request(hctx, rq, false);
1019         }
1020         spin_unlock(&ctx->lock);
1021
1022         blk_mq_run_hw_queue(hctx, from_schedule);
1023         blk_mq_put_ctx(current_ctx);
1024 }
1025
1026 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1027 {
1028         struct request *rqa = container_of(a, struct request, queuelist);
1029         struct request *rqb = container_of(b, struct request, queuelist);
1030
1031         return !(rqa->mq_ctx < rqb->mq_ctx ||
1032                  (rqa->mq_ctx == rqb->mq_ctx &&
1033                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1034 }
1035
1036 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1037 {
1038         struct blk_mq_ctx *this_ctx;
1039         struct request_queue *this_q;
1040         struct request *rq;
1041         LIST_HEAD(list);
1042         LIST_HEAD(ctx_list);
1043         unsigned int depth;
1044
1045         list_splice_init(&plug->mq_list, &list);
1046
1047         list_sort(NULL, &list, plug_ctx_cmp);
1048
1049         this_q = NULL;
1050         this_ctx = NULL;
1051         depth = 0;
1052
1053         while (!list_empty(&list)) {
1054                 rq = list_entry_rq(list.next);
1055                 list_del_init(&rq->queuelist);
1056                 BUG_ON(!rq->q);
1057                 if (rq->mq_ctx != this_ctx) {
1058                         if (this_ctx) {
1059                                 blk_mq_insert_requests(this_q, this_ctx,
1060                                                         &ctx_list, depth,
1061                                                         from_schedule);
1062                         }
1063
1064                         this_ctx = rq->mq_ctx;
1065                         this_q = rq->q;
1066                         depth = 0;
1067                 }
1068
1069                 depth++;
1070                 list_add_tail(&rq->queuelist, &ctx_list);
1071         }
1072
1073         /*
1074          * If 'this_ctx' is set, we know we have entries to complete
1075          * on 'ctx_list'. Do those.
1076          */
1077         if (this_ctx) {
1078                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1079                                        from_schedule);
1080         }
1081 }
1082
1083 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1084 {
1085         init_request_from_bio(rq, bio);
1086
1087         if (blk_do_io_stat(rq))
1088                 blk_account_io_start(rq, 1);
1089 }
1090
1091 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1092 {
1093         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1094                 !blk_queue_nomerges(hctx->queue);
1095 }
1096
1097 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1098                                          struct blk_mq_ctx *ctx,
1099                                          struct request *rq, struct bio *bio)
1100 {
1101         if (!hctx_allow_merges(hctx)) {
1102                 blk_mq_bio_to_request(rq, bio);
1103                 spin_lock(&ctx->lock);
1104 insert_rq:
1105                 __blk_mq_insert_request(hctx, rq, false);
1106                 spin_unlock(&ctx->lock);
1107                 return false;
1108         } else {
1109                 struct request_queue *q = hctx->queue;
1110
1111                 spin_lock(&ctx->lock);
1112                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1113                         blk_mq_bio_to_request(rq, bio);
1114                         goto insert_rq;
1115                 }
1116
1117                 spin_unlock(&ctx->lock);
1118                 __blk_mq_free_request(hctx, ctx, rq);
1119                 return true;
1120         }
1121 }
1122
1123 struct blk_map_ctx {
1124         struct blk_mq_hw_ctx *hctx;
1125         struct blk_mq_ctx *ctx;
1126 };
1127
1128 static struct request *blk_mq_map_request(struct request_queue *q,
1129                                           struct bio *bio,
1130                                           struct blk_map_ctx *data)
1131 {
1132         struct blk_mq_hw_ctx *hctx;
1133         struct blk_mq_ctx *ctx;
1134         struct request *rq;
1135         int rw = bio_data_dir(bio);
1136         struct blk_mq_alloc_data alloc_data;
1137
1138         if (unlikely(blk_mq_queue_enter(q))) {
1139                 bio_endio(bio, -EIO);
1140                 return NULL;
1141         }
1142
1143         ctx = blk_mq_get_ctx(q);
1144         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1145
1146         if (rw_is_sync(bio->bi_rw))
1147                 rw |= REQ_SYNC;
1148
1149         trace_block_getrq(q, bio, rw);
1150         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1151                         hctx);
1152         rq = __blk_mq_alloc_request(&alloc_data, rw);
1153         if (unlikely(!rq)) {
1154                 __blk_mq_run_hw_queue(hctx);
1155                 blk_mq_put_ctx(ctx);
1156                 trace_block_sleeprq(q, bio, rw);
1157
1158                 ctx = blk_mq_get_ctx(q);
1159                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1160                 blk_mq_set_alloc_data(&alloc_data, q,
1161                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1162                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1163                 ctx = alloc_data.ctx;
1164                 hctx = alloc_data.hctx;
1165         }
1166
1167         hctx->queued++;
1168         data->hctx = hctx;
1169         data->ctx = ctx;
1170         return rq;
1171 }
1172
1173 /*
1174  * Multiple hardware queue variant. This will not use per-process plugs,
1175  * but will attempt to bypass the hctx queueing if we can go straight to
1176  * hardware for SYNC IO.
1177  */
1178 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1179 {
1180         const int is_sync = rw_is_sync(bio->bi_rw);
1181         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1182         struct blk_map_ctx data;
1183         struct request *rq;
1184
1185         blk_queue_bounce(q, &bio);
1186
1187         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1188                 bio_endio(bio, -EIO);
1189                 return;
1190         }
1191
1192         rq = blk_mq_map_request(q, bio, &data);
1193         if (unlikely(!rq))
1194                 return;
1195
1196         if (unlikely(is_flush_fua)) {
1197                 blk_mq_bio_to_request(rq, bio);
1198                 blk_insert_flush(rq);
1199                 goto run_queue;
1200         }
1201
1202         if (is_sync) {
1203                 int ret;
1204
1205                 blk_mq_bio_to_request(rq, bio);
1206                 blk_mq_start_request(rq, true);
1207
1208                 /*
1209                  * For OK queue, we are done. For error, kill it. Any other
1210                  * error (busy), just add it to our list as we previously
1211                  * would have done
1212                  */
1213                 ret = q->mq_ops->queue_rq(data.hctx, rq);
1214                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1215                         goto done;
1216                 else {
1217                         __blk_mq_requeue_request(rq);
1218
1219                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1220                                 rq->errors = -EIO;
1221                                 blk_mq_end_io(rq, rq->errors);
1222                                 goto done;
1223                         }
1224                 }
1225         }
1226
1227         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1228                 /*
1229                  * For a SYNC request, send it to the hardware immediately. For
1230                  * an ASYNC request, just ensure that we run it later on. The
1231                  * latter allows for merging opportunities and more efficient
1232                  * dispatching.
1233                  */
1234 run_queue:
1235                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1236         }
1237 done:
1238         blk_mq_put_ctx(data.ctx);
1239 }
1240
1241 /*
1242  * Single hardware queue variant. This will attempt to use any per-process
1243  * plug for merging and IO deferral.
1244  */
1245 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1246 {
1247         const int is_sync = rw_is_sync(bio->bi_rw);
1248         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1249         unsigned int use_plug, request_count = 0;
1250         struct blk_map_ctx data;
1251         struct request *rq;
1252
1253         /*
1254          * If we have multiple hardware queues, just go directly to
1255          * one of those for sync IO.
1256          */
1257         use_plug = !is_flush_fua && !is_sync;
1258
1259         blk_queue_bounce(q, &bio);
1260
1261         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1262                 bio_endio(bio, -EIO);
1263                 return;
1264         }
1265
1266         if (use_plug && !blk_queue_nomerges(q) &&
1267             blk_attempt_plug_merge(q, bio, &request_count))
1268                 return;
1269
1270         rq = blk_mq_map_request(q, bio, &data);
1271         if (unlikely(!rq))
1272                 return;
1273
1274         if (unlikely(is_flush_fua)) {
1275                 blk_mq_bio_to_request(rq, bio);
1276                 blk_insert_flush(rq);
1277                 goto run_queue;
1278         }
1279
1280         /*
1281          * A task plug currently exists. Since this is completely lockless,
1282          * utilize that to temporarily store requests until the task is
1283          * either done or scheduled away.
1284          */
1285         if (use_plug) {
1286                 struct blk_plug *plug = current->plug;
1287
1288                 if (plug) {
1289                         blk_mq_bio_to_request(rq, bio);
1290                         if (list_empty(&plug->mq_list))
1291                                 trace_block_plug(q);
1292                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1293                                 blk_flush_plug_list(plug, false);
1294                                 trace_block_plug(q);
1295                         }
1296                         list_add_tail(&rq->queuelist, &plug->mq_list);
1297                         blk_mq_put_ctx(data.ctx);
1298                         return;
1299                 }
1300         }
1301
1302         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1303                 /*
1304                  * For a SYNC request, send it to the hardware immediately. For
1305                  * an ASYNC request, just ensure that we run it later on. The
1306                  * latter allows for merging opportunities and more efficient
1307                  * dispatching.
1308                  */
1309 run_queue:
1310                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1311         }
1312
1313         blk_mq_put_ctx(data.ctx);
1314 }
1315
1316 /*
1317  * Default mapping to a software queue, since we use one per CPU.
1318  */
1319 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1320 {
1321         return q->queue_hw_ctx[q->mq_map[cpu]];
1322 }
1323 EXPORT_SYMBOL(blk_mq_map_queue);
1324
1325 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1326                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1327 {
1328         struct page *page;
1329
1330         if (tags->rqs && set->ops->exit_request) {
1331                 int i;
1332
1333                 for (i = 0; i < tags->nr_tags; i++) {
1334                         if (!tags->rqs[i])
1335                                 continue;
1336                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1337                                                 hctx_idx, i);
1338                         tags->rqs[i] = NULL;
1339                 }
1340         }
1341
1342         while (!list_empty(&tags->page_list)) {
1343                 page = list_first_entry(&tags->page_list, struct page, lru);
1344                 list_del_init(&page->lru);
1345                 __free_pages(page, page->private);
1346         }
1347
1348         kfree(tags->rqs);
1349
1350         blk_mq_free_tags(tags);
1351 }
1352
1353 static size_t order_to_size(unsigned int order)
1354 {
1355         return (size_t)PAGE_SIZE << order;
1356 }
1357
1358 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1359                 unsigned int hctx_idx)
1360 {
1361         struct blk_mq_tags *tags;
1362         unsigned int i, j, entries_per_page, max_order = 4;
1363         size_t rq_size, left;
1364
1365         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1366                                 set->numa_node);
1367         if (!tags)
1368                 return NULL;
1369
1370         INIT_LIST_HEAD(&tags->page_list);
1371
1372         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1373                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1374                                  set->numa_node);
1375         if (!tags->rqs) {
1376                 blk_mq_free_tags(tags);
1377                 return NULL;
1378         }
1379
1380         /*
1381          * rq_size is the size of the request plus driver payload, rounded
1382          * to the cacheline size
1383          */
1384         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1385                                 cache_line_size());
1386         left = rq_size * set->queue_depth;
1387
1388         for (i = 0; i < set->queue_depth; ) {
1389                 int this_order = max_order;
1390                 struct page *page;
1391                 int to_do;
1392                 void *p;
1393
1394                 while (left < order_to_size(this_order - 1) && this_order)
1395                         this_order--;
1396
1397                 do {
1398                         page = alloc_pages_node(set->numa_node,
1399                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1400                                 this_order);
1401                         if (page)
1402                                 break;
1403                         if (!this_order--)
1404                                 break;
1405                         if (order_to_size(this_order) < rq_size)
1406                                 break;
1407                 } while (1);
1408
1409                 if (!page)
1410                         goto fail;
1411
1412                 page->private = this_order;
1413                 list_add_tail(&page->lru, &tags->page_list);
1414
1415                 p = page_address(page);
1416                 entries_per_page = order_to_size(this_order) / rq_size;
1417                 to_do = min(entries_per_page, set->queue_depth - i);
1418                 left -= to_do * rq_size;
1419                 for (j = 0; j < to_do; j++) {
1420                         tags->rqs[i] = p;
1421                         tags->rqs[i]->atomic_flags = 0;
1422                         tags->rqs[i]->cmd_flags = 0;
1423                         if (set->ops->init_request) {
1424                                 if (set->ops->init_request(set->driver_data,
1425                                                 tags->rqs[i], hctx_idx, i,
1426                                                 set->numa_node)) {
1427                                         tags->rqs[i] = NULL;
1428                                         goto fail;
1429                                 }
1430                         }
1431
1432                         p += rq_size;
1433                         i++;
1434                 }
1435         }
1436
1437         return tags;
1438
1439 fail:
1440         blk_mq_free_rq_map(set, tags, hctx_idx);
1441         return NULL;
1442 }
1443
1444 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1445 {
1446         kfree(bitmap->map);
1447 }
1448
1449 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1450 {
1451         unsigned int bpw = 8, total, num_maps, i;
1452
1453         bitmap->bits_per_word = bpw;
1454
1455         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1456         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1457                                         GFP_KERNEL, node);
1458         if (!bitmap->map)
1459                 return -ENOMEM;
1460
1461         bitmap->map_size = num_maps;
1462
1463         total = nr_cpu_ids;
1464         for (i = 0; i < num_maps; i++) {
1465                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1466                 total -= bitmap->map[i].depth;
1467         }
1468
1469         return 0;
1470 }
1471
1472 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1473 {
1474         struct request_queue *q = hctx->queue;
1475         struct blk_mq_ctx *ctx;
1476         LIST_HEAD(tmp);
1477
1478         /*
1479          * Move ctx entries to new CPU, if this one is going away.
1480          */
1481         ctx = __blk_mq_get_ctx(q, cpu);
1482
1483         spin_lock(&ctx->lock);
1484         if (!list_empty(&ctx->rq_list)) {
1485                 list_splice_init(&ctx->rq_list, &tmp);
1486                 blk_mq_hctx_clear_pending(hctx, ctx);
1487         }
1488         spin_unlock(&ctx->lock);
1489
1490         if (list_empty(&tmp))
1491                 return NOTIFY_OK;
1492
1493         ctx = blk_mq_get_ctx(q);
1494         spin_lock(&ctx->lock);
1495
1496         while (!list_empty(&tmp)) {
1497                 struct request *rq;
1498
1499                 rq = list_first_entry(&tmp, struct request, queuelist);
1500                 rq->mq_ctx = ctx;
1501                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1502         }
1503
1504         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1505         blk_mq_hctx_mark_pending(hctx, ctx);
1506
1507         spin_unlock(&ctx->lock);
1508
1509         blk_mq_run_hw_queue(hctx, true);
1510         blk_mq_put_ctx(ctx);
1511         return NOTIFY_OK;
1512 }
1513
1514 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1515 {
1516         struct request_queue *q = hctx->queue;
1517         struct blk_mq_tag_set *set = q->tag_set;
1518
1519         if (set->tags[hctx->queue_num])
1520                 return NOTIFY_OK;
1521
1522         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1523         if (!set->tags[hctx->queue_num])
1524                 return NOTIFY_STOP;
1525
1526         hctx->tags = set->tags[hctx->queue_num];
1527         return NOTIFY_OK;
1528 }
1529
1530 static int blk_mq_hctx_notify(void *data, unsigned long action,
1531                               unsigned int cpu)
1532 {
1533         struct blk_mq_hw_ctx *hctx = data;
1534
1535         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1536                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1537         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1538                 return blk_mq_hctx_cpu_online(hctx, cpu);
1539
1540         return NOTIFY_OK;
1541 }
1542
1543 static void blk_mq_exit_hw_queues(struct request_queue *q,
1544                 struct blk_mq_tag_set *set, int nr_queue)
1545 {
1546         struct blk_mq_hw_ctx *hctx;
1547         unsigned int i;
1548
1549         queue_for_each_hw_ctx(q, hctx, i) {
1550                 if (i == nr_queue)
1551                         break;
1552
1553                 blk_mq_tag_idle(hctx);
1554
1555                 if (set->ops->exit_hctx)
1556                         set->ops->exit_hctx(hctx, i);
1557
1558                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1559                 kfree(hctx->ctxs);
1560                 blk_mq_free_bitmap(&hctx->ctx_map);
1561         }
1562
1563 }
1564
1565 static void blk_mq_free_hw_queues(struct request_queue *q,
1566                 struct blk_mq_tag_set *set)
1567 {
1568         struct blk_mq_hw_ctx *hctx;
1569         unsigned int i;
1570
1571         queue_for_each_hw_ctx(q, hctx, i) {
1572                 free_cpumask_var(hctx->cpumask);
1573                 kfree(hctx);
1574         }
1575 }
1576
1577 static int blk_mq_init_hw_queues(struct request_queue *q,
1578                 struct blk_mq_tag_set *set)
1579 {
1580         struct blk_mq_hw_ctx *hctx;
1581         unsigned int i;
1582
1583         /*
1584          * Initialize hardware queues
1585          */
1586         queue_for_each_hw_ctx(q, hctx, i) {
1587                 int node;
1588
1589                 node = hctx->numa_node;
1590                 if (node == NUMA_NO_NODE)
1591                         node = hctx->numa_node = set->numa_node;
1592
1593                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1594                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1595                 spin_lock_init(&hctx->lock);
1596                 INIT_LIST_HEAD(&hctx->dispatch);
1597                 hctx->queue = q;
1598                 hctx->queue_num = i;
1599                 hctx->flags = set->flags;
1600                 hctx->cmd_size = set->cmd_size;
1601
1602                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1603                                                 blk_mq_hctx_notify, hctx);
1604                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1605
1606                 hctx->tags = set->tags[i];
1607
1608                 /*
1609                  * Allocate space for all possible cpus to avoid allocation at
1610                  * runtime
1611                  */
1612                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1613                                                 GFP_KERNEL, node);
1614                 if (!hctx->ctxs)
1615                         break;
1616
1617                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1618                         break;
1619
1620                 hctx->nr_ctx = 0;
1621
1622                 if (set->ops->init_hctx &&
1623                     set->ops->init_hctx(hctx, set->driver_data, i))
1624                         break;
1625         }
1626
1627         if (i == q->nr_hw_queues)
1628                 return 0;
1629
1630         /*
1631          * Init failed
1632          */
1633         blk_mq_exit_hw_queues(q, set, i);
1634
1635         return 1;
1636 }
1637
1638 static void blk_mq_init_cpu_queues(struct request_queue *q,
1639                                    unsigned int nr_hw_queues)
1640 {
1641         unsigned int i;
1642
1643         for_each_possible_cpu(i) {
1644                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1645                 struct blk_mq_hw_ctx *hctx;
1646
1647                 memset(__ctx, 0, sizeof(*__ctx));
1648                 __ctx->cpu = i;
1649                 spin_lock_init(&__ctx->lock);
1650                 INIT_LIST_HEAD(&__ctx->rq_list);
1651                 __ctx->queue = q;
1652
1653                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1654                 if (!cpu_online(i))
1655                         continue;
1656
1657                 hctx = q->mq_ops->map_queue(q, i);
1658                 cpumask_set_cpu(i, hctx->cpumask);
1659                 hctx->nr_ctx++;
1660
1661                 /*
1662                  * Set local node, IFF we have more than one hw queue. If
1663                  * not, we remain on the home node of the device
1664                  */
1665                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1666                         hctx->numa_node = cpu_to_node(i);
1667         }
1668 }
1669
1670 static void blk_mq_map_swqueue(struct request_queue *q)
1671 {
1672         unsigned int i;
1673         struct blk_mq_hw_ctx *hctx;
1674         struct blk_mq_ctx *ctx;
1675
1676         queue_for_each_hw_ctx(q, hctx, i) {
1677                 cpumask_clear(hctx->cpumask);
1678                 hctx->nr_ctx = 0;
1679         }
1680
1681         /*
1682          * Map software to hardware queues
1683          */
1684         queue_for_each_ctx(q, ctx, i) {
1685                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1686                 if (!cpu_online(i))
1687                         continue;
1688
1689                 hctx = q->mq_ops->map_queue(q, i);
1690                 cpumask_set_cpu(i, hctx->cpumask);
1691                 ctx->index_hw = hctx->nr_ctx;
1692                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1693         }
1694
1695         queue_for_each_hw_ctx(q, hctx, i) {
1696                 /*
1697                  * If no software queues are mapped to this hardware queue,
1698                  * disable it and free the request entries.
1699                  */
1700                 if (!hctx->nr_ctx) {
1701                         struct blk_mq_tag_set *set = q->tag_set;
1702
1703                         if (set->tags[i]) {
1704                                 blk_mq_free_rq_map(set, set->tags[i], i);
1705                                 set->tags[i] = NULL;
1706                                 hctx->tags = NULL;
1707                         }
1708                         continue;
1709                 }
1710
1711                 /*
1712                  * Initialize batch roundrobin counts
1713                  */
1714                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1715                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1716         }
1717 }
1718
1719 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1720 {
1721         struct blk_mq_hw_ctx *hctx;
1722         struct request_queue *q;
1723         bool shared;
1724         int i;
1725
1726         if (set->tag_list.next == set->tag_list.prev)
1727                 shared = false;
1728         else
1729                 shared = true;
1730
1731         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1732                 blk_mq_freeze_queue(q);
1733
1734                 queue_for_each_hw_ctx(q, hctx, i) {
1735                         if (shared)
1736                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1737                         else
1738                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1739                 }
1740                 blk_mq_unfreeze_queue(q);
1741         }
1742 }
1743
1744 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1745 {
1746         struct blk_mq_tag_set *set = q->tag_set;
1747
1748         mutex_lock(&set->tag_list_lock);
1749         list_del_init(&q->tag_set_list);
1750         blk_mq_update_tag_set_depth(set);
1751         mutex_unlock(&set->tag_list_lock);
1752 }
1753
1754 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1755                                      struct request_queue *q)
1756 {
1757         q->tag_set = set;
1758
1759         mutex_lock(&set->tag_list_lock);
1760         list_add_tail(&q->tag_set_list, &set->tag_list);
1761         blk_mq_update_tag_set_depth(set);
1762         mutex_unlock(&set->tag_list_lock);
1763 }
1764
1765 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1766 {
1767         struct blk_mq_hw_ctx **hctxs;
1768         struct blk_mq_ctx __percpu *ctx;
1769         struct request_queue *q;
1770         unsigned int *map;
1771         int i;
1772
1773         ctx = alloc_percpu(struct blk_mq_ctx);
1774         if (!ctx)
1775                 return ERR_PTR(-ENOMEM);
1776
1777         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1778                         set->numa_node);
1779
1780         if (!hctxs)
1781                 goto err_percpu;
1782
1783         map = blk_mq_make_queue_map(set);
1784         if (!map)
1785                 goto err_map;
1786
1787         for (i = 0; i < set->nr_hw_queues; i++) {
1788                 int node = blk_mq_hw_queue_to_node(map, i);
1789
1790                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1791                                         GFP_KERNEL, node);
1792                 if (!hctxs[i])
1793                         goto err_hctxs;
1794
1795                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1796                         goto err_hctxs;
1797
1798                 atomic_set(&hctxs[i]->nr_active, 0);
1799                 hctxs[i]->numa_node = node;
1800                 hctxs[i]->queue_num = i;
1801         }
1802
1803         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1804         if (!q)
1805                 goto err_hctxs;
1806
1807         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1808                 goto err_map;
1809
1810         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1811         blk_queue_rq_timeout(q, 30000);
1812
1813         q->nr_queues = nr_cpu_ids;
1814         q->nr_hw_queues = set->nr_hw_queues;
1815         q->mq_map = map;
1816
1817         q->queue_ctx = ctx;
1818         q->queue_hw_ctx = hctxs;
1819
1820         q->mq_ops = set->ops;
1821         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1822
1823         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1824                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1825
1826         q->sg_reserved_size = INT_MAX;
1827
1828         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1829         INIT_LIST_HEAD(&q->requeue_list);
1830         spin_lock_init(&q->requeue_lock);
1831
1832         if (q->nr_hw_queues > 1)
1833                 blk_queue_make_request(q, blk_mq_make_request);
1834         else
1835                 blk_queue_make_request(q, blk_sq_make_request);
1836
1837         blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1838         if (set->timeout)
1839                 blk_queue_rq_timeout(q, set->timeout);
1840
1841         /*
1842          * Do this after blk_queue_make_request() overrides it...
1843          */
1844         q->nr_requests = set->queue_depth;
1845
1846         if (set->ops->complete)
1847                 blk_queue_softirq_done(q, set->ops->complete);
1848
1849         blk_mq_init_flush(q);
1850         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1851
1852         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1853                                 set->cmd_size, cache_line_size()),
1854                                 GFP_KERNEL);
1855         if (!q->flush_rq)
1856                 goto err_hw;
1857
1858         if (blk_mq_init_hw_queues(q, set))
1859                 goto err_flush_rq;
1860
1861         mutex_lock(&all_q_mutex);
1862         list_add_tail(&q->all_q_node, &all_q_list);
1863         mutex_unlock(&all_q_mutex);
1864
1865         blk_mq_add_queue_tag_set(set, q);
1866
1867         blk_mq_map_swqueue(q);
1868
1869         return q;
1870
1871 err_flush_rq:
1872         kfree(q->flush_rq);
1873 err_hw:
1874         blk_cleanup_queue(q);
1875 err_hctxs:
1876         kfree(map);
1877         for (i = 0; i < set->nr_hw_queues; i++) {
1878                 if (!hctxs[i])
1879                         break;
1880                 free_cpumask_var(hctxs[i]->cpumask);
1881                 kfree(hctxs[i]);
1882         }
1883 err_map:
1884         kfree(hctxs);
1885 err_percpu:
1886         free_percpu(ctx);
1887         return ERR_PTR(-ENOMEM);
1888 }
1889 EXPORT_SYMBOL(blk_mq_init_queue);
1890
1891 void blk_mq_free_queue(struct request_queue *q)
1892 {
1893         struct blk_mq_tag_set   *set = q->tag_set;
1894
1895         blk_mq_del_queue_tag_set(q);
1896
1897         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1898         blk_mq_free_hw_queues(q, set);
1899
1900         percpu_ref_exit(&q->mq_usage_counter);
1901
1902         free_percpu(q->queue_ctx);
1903         kfree(q->queue_hw_ctx);
1904         kfree(q->mq_map);
1905
1906         q->queue_ctx = NULL;
1907         q->queue_hw_ctx = NULL;
1908         q->mq_map = NULL;
1909
1910         mutex_lock(&all_q_mutex);
1911         list_del_init(&q->all_q_node);
1912         mutex_unlock(&all_q_mutex);
1913 }
1914
1915 /* Basically redo blk_mq_init_queue with queue frozen */
1916 static void blk_mq_queue_reinit(struct request_queue *q)
1917 {
1918         blk_mq_freeze_queue(q);
1919
1920         blk_mq_sysfs_unregister(q);
1921
1922         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1923
1924         /*
1925          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1926          * we should change hctx numa_node according to new topology (this
1927          * involves free and re-allocate memory, worthy doing?)
1928          */
1929
1930         blk_mq_map_swqueue(q);
1931
1932         blk_mq_sysfs_register(q);
1933
1934         blk_mq_unfreeze_queue(q);
1935 }
1936
1937 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1938                                       unsigned long action, void *hcpu)
1939 {
1940         struct request_queue *q;
1941
1942         /*
1943          * Before new mappings are established, hotadded cpu might already
1944          * start handling requests. This doesn't break anything as we map
1945          * offline CPUs to first hardware queue. We will re-init the queue
1946          * below to get optimal settings.
1947          */
1948         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1949             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1950                 return NOTIFY_OK;
1951
1952         mutex_lock(&all_q_mutex);
1953         list_for_each_entry(q, &all_q_list, all_q_node)
1954                 blk_mq_queue_reinit(q);
1955         mutex_unlock(&all_q_mutex);
1956         return NOTIFY_OK;
1957 }
1958
1959 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1960 {
1961         int i;
1962
1963         for (i = 0; i < set->nr_hw_queues; i++) {
1964                 set->tags[i] = blk_mq_init_rq_map(set, i);
1965                 if (!set->tags[i])
1966                         goto out_unwind;
1967         }
1968
1969         return 0;
1970
1971 out_unwind:
1972         while (--i >= 0)
1973                 blk_mq_free_rq_map(set, set->tags[i], i);
1974
1975         return -ENOMEM;
1976 }
1977
1978 /*
1979  * Allocate the request maps associated with this tag_set. Note that this
1980  * may reduce the depth asked for, if memory is tight. set->queue_depth
1981  * will be updated to reflect the allocated depth.
1982  */
1983 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1984 {
1985         unsigned int depth;
1986         int err;
1987
1988         depth = set->queue_depth;
1989         do {
1990                 err = __blk_mq_alloc_rq_maps(set);
1991                 if (!err)
1992                         break;
1993
1994                 set->queue_depth >>= 1;
1995                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1996                         err = -ENOMEM;
1997                         break;
1998                 }
1999         } while (set->queue_depth);
2000
2001         if (!set->queue_depth || err) {
2002                 pr_err("blk-mq: failed to allocate request map\n");
2003                 return -ENOMEM;
2004         }
2005
2006         if (depth != set->queue_depth)
2007                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2008                                                 depth, set->queue_depth);
2009
2010         return 0;
2011 }
2012
2013 /*
2014  * Alloc a tag set to be associated with one or more request queues.
2015  * May fail with EINVAL for various error conditions. May adjust the
2016  * requested depth down, if if it too large. In that case, the set
2017  * value will be stored in set->queue_depth.
2018  */
2019 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2020 {
2021         if (!set->nr_hw_queues)
2022                 return -EINVAL;
2023         if (!set->queue_depth)
2024                 return -EINVAL;
2025         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2026                 return -EINVAL;
2027
2028         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2029                 return -EINVAL;
2030
2031         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2032                 pr_info("blk-mq: reduced tag depth to %u\n",
2033                         BLK_MQ_MAX_DEPTH);
2034                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2035         }
2036
2037         set->tags = kmalloc_node(set->nr_hw_queues *
2038                                  sizeof(struct blk_mq_tags *),
2039                                  GFP_KERNEL, set->numa_node);
2040         if (!set->tags)
2041                 return -ENOMEM;
2042
2043         if (blk_mq_alloc_rq_maps(set))
2044                 goto enomem;
2045
2046         mutex_init(&set->tag_list_lock);
2047         INIT_LIST_HEAD(&set->tag_list);
2048
2049         return 0;
2050 enomem:
2051         kfree(set->tags);
2052         set->tags = NULL;
2053         return -ENOMEM;
2054 }
2055 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2056
2057 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2058 {
2059         int i;
2060
2061         for (i = 0; i < set->nr_hw_queues; i++) {
2062                 if (set->tags[i])
2063                         blk_mq_free_rq_map(set, set->tags[i], i);
2064         }
2065
2066         kfree(set->tags);
2067         set->tags = NULL;
2068 }
2069 EXPORT_SYMBOL(blk_mq_free_tag_set);
2070
2071 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2072 {
2073         struct blk_mq_tag_set *set = q->tag_set;
2074         struct blk_mq_hw_ctx *hctx;
2075         int i, ret;
2076
2077         if (!set || nr > set->queue_depth)
2078                 return -EINVAL;
2079
2080         ret = 0;
2081         queue_for_each_hw_ctx(q, hctx, i) {
2082                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2083                 if (ret)
2084                         break;
2085         }
2086
2087         if (!ret)
2088                 q->nr_requests = nr;
2089
2090         return ret;
2091 }
2092
2093 void blk_mq_disable_hotplug(void)
2094 {
2095         mutex_lock(&all_q_mutex);
2096 }
2097
2098 void blk_mq_enable_hotplug(void)
2099 {
2100         mutex_unlock(&all_q_mutex);
2101 }
2102
2103 static int __init blk_mq_init(void)
2104 {
2105         blk_mq_cpu_init();
2106
2107         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2108
2109         return 0;
2110 }
2111 subsys_initcall(blk_mq_init);