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 #include <linux/crash_dump.h>
  24
  25 #include <trace/events/block.h>
  26
  27 #include <linux/blk-mq.h>
  28 #include "blk.h"
  29 #include "blk-mq.h"
  30 #include "blk-mq-tag.h"
  31
  32 static DEFINE_MUTEX(all_q_mutex);
  33 static LIST_HEAD(all_q_list);
  34
  35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  36
  37 /*
  38  * Check if any of the ctx's have pending work in this hardware queue
  39  */
  40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  41 {
  42         unsigned int i;
  43
  44         for (i = 0; i < hctx->ctx_map.size; i++)
  45                 if (hctx->ctx_map.map[i].word)
  46                         return true;
  47
  48         return false;
  49 }
  50
  51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  52                                               struct blk_mq_ctx *ctx)
  53 {
  54         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  55 }
  56
  57 #define CTX_TO_BIT(hctx, ctx)   \
  58         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  59
  60 /*
  61  * Mark this ctx as having pending work in this hardware queue
  62  */
  63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  64                                      struct blk_mq_ctx *ctx)
  65 {
  66         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  67
  68         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  69                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  70 }
  71
  72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  73                                       struct blk_mq_ctx *ctx)
  74 {
  75         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  76
  77         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  78 }
  79
  80 static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 if (!(gfp & __GFP_WAIT))
  89                         return -EBUSY;
  90
  91                 ret = wait_event_interruptible(q->mq_freeze_wq,
  92                                 !atomic_read(&q->mq_freeze_depth) ||
  93                                 blk_queue_dying(q));
  94                 if (blk_queue_dying(q))
  95                         return -ENODEV;
  96                 if (ret)
  97                         return ret;
  98         }
  99 }
 100
 101 static void blk_mq_queue_exit(struct request_queue *q)
 102 {
 103         percpu_ref_put(&q->mq_usage_counter);
 104 }
 105
 106 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 107 {
 108         struct request_queue *q =
 109                 container_of(ref, struct request_queue, mq_usage_counter);
 110
 111         wake_up_all(&q->mq_freeze_wq);
 112 }
 113
 114 void blk_mq_freeze_queue_start(struct request_queue *q)
 115 {
 116         int freeze_depth;
 117
 118         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 119         if (freeze_depth == 1) {
 120                 percpu_ref_kill(&q->mq_usage_counter);
 121                 blk_mq_run_hw_queues(q, false);
 122         }
 123 }
 124 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 125
 126 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 127 {
 128         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 129 }
 130
 131 /*
 132  * Guarantee no request is in use, so we can change any data structure of
 133  * the queue afterward.
 134  */
 135 void blk_mq_freeze_queue(struct request_queue *q)
 136 {
 137         blk_mq_freeze_queue_start(q);
 138         blk_mq_freeze_queue_wait(q);
 139 }
 140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 141
 142 void blk_mq_unfreeze_queue(struct request_queue *q)
 143 {
 144         int freeze_depth;
 145
 146         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 147         WARN_ON_ONCE(freeze_depth < 0);
 148         if (!freeze_depth) {
 149                 percpu_ref_reinit(&q->mq_usage_counter);
 150                 wake_up_all(&q->mq_freeze_wq);
 151         }
 152 }
 153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 154
 155 void blk_mq_wake_waiters(struct request_queue *q)
 156 {
 157         struct blk_mq_hw_ctx *hctx;
 158         unsigned int i;
 159
 160         queue_for_each_hw_ctx(q, hctx, i)
 161                 if (blk_mq_hw_queue_mapped(hctx))
 162                         blk_mq_tag_wakeup_all(hctx->tags, true);
 163
 164         /*
 165          * If we are called because the queue has now been marked as
 166          * dying, we need to ensure that processes currently waiting on
 167          * the queue are notified as well.
 168          */
 169         wake_up_all(&q->mq_freeze_wq);
 170 }
 171
 172 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 173 {
 174         return blk_mq_has_free_tags(hctx->tags);
 175 }
 176 EXPORT_SYMBOL(blk_mq_can_queue);
 177
 178 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 179                                struct request *rq, unsigned int rw_flags)
 180 {
 181         if (blk_queue_io_stat(q))
 182                 rw_flags |= REQ_IO_STAT;
 183
 184         INIT_LIST_HEAD(&rq->queuelist);
 185         /* csd/requeue_work/fifo_time is initialized before use */
 186         rq->q = q;
 187         rq->mq_ctx = ctx;
 188         rq->cmd_flags |= rw_flags;
 189         /* do not touch atomic flags, it needs atomic ops against the timer */
 190         rq->cpu = -1;
 191         INIT_HLIST_NODE(&rq->hash);
 192         RB_CLEAR_NODE(&rq->rb_node);
 193         rq->rq_disk = NULL;
 194         rq->part = NULL;
 195         rq->start_time = jiffies;
 196 #ifdef CONFIG_BLK_CGROUP
 197         rq->rl = NULL;
 198         set_start_time_ns(rq);
 199         rq->io_start_time_ns = 0;
 200 #endif
 201         rq->nr_phys_segments = 0;
 202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 203         rq->nr_integrity_segments = 0;
 204 #endif
 205         rq->special = NULL;
 206         /* tag was already set */
 207         rq->errors = 0;
 208
 209         rq->cmd = rq->__cmd;
 210
 211         rq->extra_len = 0;
 212         rq->sense_len = 0;
 213         rq->resid_len = 0;
 214         rq->sense = NULL;
 215
 216         INIT_LIST_HEAD(&rq->timeout_list);
 217         rq->timeout = 0;
 218
 219         rq->end_io = NULL;
 220         rq->end_io_data = NULL;
 221         rq->next_rq = NULL;
 222
 223         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 224 }
 225
 226 static struct request *
 227 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 228 {
 229         struct request *rq;
 230         unsigned int tag;
 231
 232         tag = blk_mq_get_tag(data);
 233         if (tag != BLK_MQ_TAG_FAIL) {
 234                 rq = data->hctx->tags->rqs[tag];
 235
 236                 if (blk_mq_tag_busy(data->hctx)) {
 237                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 238                         atomic_inc(&data->hctx->nr_active);
 239                 }
 240
 241                 rq->tag = tag;
 242                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 243                 return rq;
 244         }
 245
 246         return NULL;
 247 }
 248
 249 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 250                 bool reserved)
 251 {
 252         struct blk_mq_ctx *ctx;
 253         struct blk_mq_hw_ctx *hctx;
 254         struct request *rq;
 255         struct blk_mq_alloc_data alloc_data;
 256         int ret;
 257
 258         ret = blk_mq_queue_enter(q, gfp);
 259         if (ret)
 260                 return ERR_PTR(ret);
 261
 262         ctx = blk_mq_get_ctx(q);
 263         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 264         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 265                         reserved, ctx, hctx);
 266
 267         rq = __blk_mq_alloc_request(&alloc_data, rw);
 268         if (!rq && (gfp & __GFP_WAIT)) {
 269                 __blk_mq_run_hw_queue(hctx);
 270                 blk_mq_put_ctx(ctx);
 271
 272                 ctx = blk_mq_get_ctx(q);
 273                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 274                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 275                                 hctx);
 276                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 277                 ctx = alloc_data.ctx;
 278         }
 279         blk_mq_put_ctx(ctx);
 280         if (!rq) {
 281                 blk_mq_queue_exit(q);
 282                 return ERR_PTR(-EWOULDBLOCK);
 283         }
 284         return rq;
 285 }
 286 EXPORT_SYMBOL(blk_mq_alloc_request);
 287
 288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 289                                   struct blk_mq_ctx *ctx, struct request *rq)
 290 {
 291         const int tag = rq->tag;
 292         struct request_queue *q = rq->q;
 293
 294         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 295                 atomic_dec(&hctx->nr_active);
 296         rq->cmd_flags = 0;
 297
 298         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 299         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 300         blk_mq_queue_exit(q);
 301 }
 302
 303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 304 {
 305         struct blk_mq_ctx *ctx = rq->mq_ctx;
 306
 307         ctx->rq_completed[rq_is_sync(rq)]++;
 308         __blk_mq_free_request(hctx, ctx, rq);
 309
 310 }
 311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 312
 313 void blk_mq_free_request(struct request *rq)
 314 {
 315         struct blk_mq_hw_ctx *hctx;
 316         struct request_queue *q = rq->q;
 317
 318         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 319         blk_mq_free_hctx_request(hctx, rq);
 320 }
 321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 322
 323 inline void __blk_mq_end_request(struct request *rq, int error)
 324 {
 325         blk_account_io_done(rq);
 326
 327         if (rq->end_io) {
 328                 rq->end_io(rq, error);
 329         } else {
 330                 if (unlikely(blk_bidi_rq(rq)))
 331                         blk_mq_free_request(rq->next_rq);
 332                 blk_mq_free_request(rq);
 333         }
 334 }
 335 EXPORT_SYMBOL(__blk_mq_end_request);
 336
 337 void blk_mq_end_request(struct request *rq, int error)
 338 {
 339         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 340                 BUG();
 341         __blk_mq_end_request(rq, error);
 342 }
 343 EXPORT_SYMBOL(blk_mq_end_request);
 344
 345 static void __blk_mq_complete_request_remote(void *data)
 346 {
 347         struct request *rq = data;
 348
 349         rq->q->softirq_done_fn(rq);
 350 }
 351
 352 static void blk_mq_ipi_complete_request(struct request *rq)
 353 {
 354         struct blk_mq_ctx *ctx = rq->mq_ctx;
 355         bool shared = false;
 356         int cpu;
 357
 358         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 359                 rq->q->softirq_done_fn(rq);
 360                 return;
 361         }
 362
 363         cpu = get_cpu();
 364         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 365                 shared = cpus_share_cache(cpu, ctx->cpu);
 366
 367         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 368                 rq->csd.func = __blk_mq_complete_request_remote;
 369                 rq->csd.info = rq;
 370                 rq->csd.flags = 0;
 371                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 372         } else {
 373                 rq->q->softirq_done_fn(rq);
 374         }
 375         put_cpu();
 376 }
 377
 378 void __blk_mq_complete_request(struct request *rq)
 379 {
 380         struct request_queue *q = rq->q;
 381
 382         if (!q->softirq_done_fn)
 383                 blk_mq_end_request(rq, rq->errors);
 384         else
 385                 blk_mq_ipi_complete_request(rq);
 386 }
 387
 388 /**
 389  * blk_mq_complete_request - end I/O on a request
 390  * @rq:         the request being processed
 391  *
 392  * Description:
 393  *      Ends all I/O on a request. It does not handle partial completions.
 394  *      The actual completion happens out-of-order, through a IPI handler.
 395  **/
 396 void blk_mq_complete_request(struct request *rq)
 397 {
 398         struct request_queue *q = rq->q;
 399
 400         if (unlikely(blk_should_fake_timeout(q)))
 401                 return;
 402         if (!blk_mark_rq_complete(rq))
 403                 __blk_mq_complete_request(rq);
 404 }
 405 EXPORT_SYMBOL(blk_mq_complete_request);
 406
 407 int blk_mq_request_started(struct request *rq)
 408 {
 409         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 410 }
 411 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 412
 413 void blk_mq_start_request(struct request *rq)
 414 {
 415         struct request_queue *q = rq->q;
 416
 417         trace_block_rq_issue(q, rq);
 418
 419         rq->resid_len = blk_rq_bytes(rq);
 420         if (unlikely(blk_bidi_rq(rq)))
 421                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 422
 423         blk_add_timer(rq);
 424
 425         /*
 426          * Ensure that ->deadline is visible before set the started
 427          * flag and clear the completed flag.
 428          */
 429         smp_mb__before_atomic();
 430
 431         /*
 432          * Mark us as started and clear complete. Complete might have been
 433          * set if requeue raced with timeout, which then marked it as
 434          * complete. So be sure to clear complete again when we start
 435          * the request, otherwise we'll ignore the completion event.
 436          */
 437         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 438                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 439         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 440                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 441
 442         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 443                 /*
 444                  * Make sure space for the drain appears.  We know we can do
 445                  * this because max_hw_segments has been adjusted to be one
 446                  * fewer than the device can handle.
 447                  */
 448                 rq->nr_phys_segments++;
 449         }
 450 }
 451 EXPORT_SYMBOL(blk_mq_start_request);
 452
 453 static void __blk_mq_requeue_request(struct request *rq)
 454 {
 455         struct request_queue *q = rq->q;
 456
 457         trace_block_rq_requeue(q, rq);
 458
 459         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 460                 if (q->dma_drain_size && blk_rq_bytes(rq))
 461                         rq->nr_phys_segments--;
 462         }
 463 }
 464
 465 void blk_mq_requeue_request(struct request *rq)
 466 {
 467         __blk_mq_requeue_request(rq);
 468
 469         BUG_ON(blk_queued_rq(rq));
 470         blk_mq_add_to_requeue_list(rq, true);
 471 }
 472 EXPORT_SYMBOL(blk_mq_requeue_request);
 473
 474 static void blk_mq_requeue_work(struct work_struct *work)
 475 {
 476         struct request_queue *q =
 477                 container_of(work, struct request_queue, requeue_work);
 478         LIST_HEAD(rq_list);
 479         struct request *rq, *next;
 480         unsigned long flags;
 481
 482         spin_lock_irqsave(&q->requeue_lock, flags);
 483         list_splice_init(&q->requeue_list, &rq_list);
 484         spin_unlock_irqrestore(&q->requeue_lock, flags);
 485
 486         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 487                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 488                         continue;
 489
 490                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 491                 list_del_init(&rq->queuelist);
 492                 blk_mq_insert_request(rq, true, false, false);
 493         }
 494
 495         while (!list_empty(&rq_list)) {
 496                 rq = list_entry(rq_list.next, struct request, queuelist);
 497                 list_del_init(&rq->queuelist);
 498                 blk_mq_insert_request(rq, false, false, false);
 499         }
 500
 501         /*
 502          * Use the start variant of queue running here, so that running
 503          * the requeue work will kick stopped queues.
 504          */
 505         blk_mq_start_hw_queues(q);
 506 }
 507
 508 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 509 {
 510         struct request_queue *q = rq->q;
 511         unsigned long flags;
 512
 513         /*
 514          * We abuse this flag that is otherwise used by the I/O scheduler to
 515          * request head insertation from the workqueue.
 516          */
 517         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 518
 519         spin_lock_irqsave(&q->requeue_lock, flags);
 520         if (at_head) {
 521                 rq->cmd_flags |= REQ_SOFTBARRIER;
 522                 list_add(&rq->queuelist, &q->requeue_list);
 523         } else {
 524                 list_add_tail(&rq->queuelist, &q->requeue_list);
 525         }
 526         spin_unlock_irqrestore(&q->requeue_lock, flags);
 527 }
 528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 529
 530 void blk_mq_cancel_requeue_work(struct request_queue *q)
 531 {
 532         cancel_work_sync(&q->requeue_work);
 533 }
 534 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 535
 536 void blk_mq_kick_requeue_list(struct request_queue *q)
 537 {
 538         kblockd_schedule_work(&q->requeue_work);
 539 }
 540 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 541
 542 void blk_mq_abort_requeue_list(struct request_queue *q)
 543 {
 544         unsigned long flags;
 545         LIST_HEAD(rq_list);
 546
 547         spin_lock_irqsave(&q->requeue_lock, flags);
 548         list_splice_init(&q->requeue_list, &rq_list);
 549         spin_unlock_irqrestore(&q->requeue_lock, flags);
 550
 551         while (!list_empty(&rq_list)) {
 552                 struct request *rq;
 553
 554                 rq = list_first_entry(&rq_list, struct request, queuelist);
 555                 list_del_init(&rq->queuelist);
 556                 rq->errors = -EIO;
 557                 blk_mq_end_request(rq, rq->errors);
 558         }
 559 }
 560 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 561
 562 static inline bool is_flush_request(struct request *rq,
 563                 struct blk_flush_queue *fq, unsigned int tag)
 564 {
 565         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 566                         fq->flush_rq->tag == tag);
 567 }
 568
 569 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 570 {
 571         struct request *rq = tags->rqs[tag];
 572         /* mq_ctx of flush rq is always cloned from the corresponding req */
 573         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 574
 575         if (!is_flush_request(rq, fq, tag))
 576                 return rq;
 577
 578         return fq->flush_rq;
 579 }
 580 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 581
 582 struct blk_mq_timeout_data {
 583         unsigned long next;
 584         unsigned int next_set;
 585 };
 586
 587 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 588 {
 589         struct blk_mq_ops *ops = req->q->mq_ops;
 590         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 591
 592         /*
 593          * We know that complete is set at this point. If STARTED isn't set
 594          * anymore, then the request isn't active and the "timeout" should
 595          * just be ignored. This can happen due to the bitflag ordering.
 596          * Timeout first checks if STARTED is set, and if it is, assumes
 597          * the request is active. But if we race with completion, then
 598          * we both flags will get cleared. So check here again, and ignore
 599          * a timeout event with a request that isn't active.
 600          */
 601         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 602                 return;
 603
 604         if (ops->timeout)
 605                 ret = ops->timeout(req, reserved);
 606
 607         switch (ret) {
 608         case BLK_EH_HANDLED:
 609                 __blk_mq_complete_request(req);
 610                 break;
 611         case BLK_EH_RESET_TIMER:
 612                 blk_add_timer(req);
 613                 blk_clear_rq_complete(req);
 614                 break;
 615         case BLK_EH_NOT_HANDLED:
 616                 break;
 617         default:
 618                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 619                 break;
 620         }
 621 }
 622
 623 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 624                 struct request *rq, void *priv, bool reserved)
 625 {
 626         struct blk_mq_timeout_data *data = priv;
 627
 628         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 629                 /*
 630                  * If a request wasn't started before the queue was
 631                  * marked dying, kill it here or it'll go unnoticed.
 632                  */
 633                 if (unlikely(blk_queue_dying(rq->q))) {
 634                         rq->errors = -EIO;
 635                         blk_mq_complete_request(rq);
 636                 }
 637                 return;
 638         }
 639         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 640                 return;
 641
 642         if (time_after_eq(jiffies, rq->deadline)) {
 643                 if (!blk_mark_rq_complete(rq))
 644                         blk_mq_rq_timed_out(rq, reserved);
 645         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 646                 data->next = rq->deadline;
 647                 data->next_set = 1;
 648         }
 649 }
 650
 651 static void blk_mq_rq_timer(unsigned long priv)
 652 {
 653         struct request_queue *q = (struct request_queue *)priv;
 654         struct blk_mq_timeout_data data = {
 655                 .next           = 0,
 656                 .next_set       = 0,
 657         };
 658         struct blk_mq_hw_ctx *hctx;
 659         int i;
 660
 661         queue_for_each_hw_ctx(q, hctx, i) {
 662                 /*
 663                  * If not software queues are currently mapped to this
 664                  * hardware queue, there's nothing to check
 665                  */
 666                 if (!blk_mq_hw_queue_mapped(hctx))
 667                         continue;
 668
 669                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 670         }
 671
 672         if (data.next_set) {
 673                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 674                 mod_timer(&q->timeout, data.next);
 675         } else {
 676                 queue_for_each_hw_ctx(q, hctx, i) {
 677                         /* the hctx may be unmapped, so check it here */
 678                         if (blk_mq_hw_queue_mapped(hctx))
 679                                 blk_mq_tag_idle(hctx);
 680                 }
 681         }
 682 }
 683
 684 /*
 685  * Reverse check our software queue for entries that we could potentially
 686  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 687  * too much time checking for merges.
 688  */
 689 static bool blk_mq_attempt_merge(struct request_queue *q,
 690                                  struct blk_mq_ctx *ctx, struct bio *bio)
 691 {
 692         struct request *rq;
 693         int checked = 8;
 694
 695         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 696                 int el_ret;
 697
 698                 if (!checked--)
 699                         break;
 700
 701                 if (!blk_rq_merge_ok(rq, bio))
 702                         continue;
 703
 704                 el_ret = blk_try_merge(rq, bio);
 705                 if (el_ret == ELEVATOR_BACK_MERGE) {
 706                         if (bio_attempt_back_merge(q, rq, bio)) {
 707                                 ctx->rq_merged++;
 708                                 return true;
 709                         }
 710                         break;
 711                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 712                         if (bio_attempt_front_merge(q, rq, bio)) {
 713                                 ctx->rq_merged++;
 714                                 return true;
 715                         }
 716                         break;
 717                 }
 718         }
 719
 720         return false;
 721 }
 722
 723 /*
 724  * Process software queues that have been marked busy, splicing them
 725  * to the for-dispatch
 726  */
 727 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 728 {
 729         struct blk_mq_ctx *ctx;
 730         int i;
 731
 732         for (i = 0; i < hctx->ctx_map.size; i++) {
 733                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 734                 unsigned int off, bit;
 735
 736                 if (!bm->word)
 737                         continue;
 738
 739                 bit = 0;
 740                 off = i * hctx->ctx_map.bits_per_word;
 741                 do {
 742                         bit = find_next_bit(&bm->word, bm->depth, bit);
 743                         if (bit >= bm->depth)
 744                                 break;
 745
 746                         ctx = hctx->ctxs[bit + off];
 747                         clear_bit(bit, &bm->word);
 748                         spin_lock(&ctx->lock);
 749                         list_splice_tail_init(&ctx->rq_list, list);
 750                         spin_unlock(&ctx->lock);
 751
 752                         bit++;
 753                 } while (1);
 754         }
 755 }
 756
 757 /*
 758  * Run this hardware queue, pulling any software queues mapped to it in.
 759  * Note that this function currently has various problems around ordering
 760  * of IO. In particular, we'd like FIFO behaviour on handling existing
 761  * items on the hctx->dispatch list. Ignore that for now.
 762  */
 763 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 764 {
 765         struct request_queue *q = hctx->queue;
 766         struct request *rq;
 767         LIST_HEAD(rq_list);
 768         LIST_HEAD(driver_list);
 769         struct list_head *dptr;
 770         int queued;
 771
 772         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 773
 774         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 775                 return;
 776
 777         hctx->run++;
 778
 779         /*
 780          * Touch any software queue that has pending entries.
 781          */
 782         flush_busy_ctxs(hctx, &rq_list);
 783
 784         /*
 785          * If we have previous entries on our dispatch list, grab them
 786          * and stuff them at the front for more fair dispatch.
 787          */
 788         if (!list_empty_careful(&hctx->dispatch)) {
 789                 spin_lock(&hctx->lock);
 790                 if (!list_empty(&hctx->dispatch))
 791                         list_splice_init(&hctx->dispatch, &rq_list);
 792                 spin_unlock(&hctx->lock);
 793         }
 794
 795         /*
 796          * Start off with dptr being NULL, so we start the first request
 797          * immediately, even if we have more pending.
 798          */
 799         dptr = NULL;
 800
 801         /*
 802          * Now process all the entries, sending them to the driver.
 803          */
 804         queued = 0;
 805         while (!list_empty(&rq_list)) {
 806                 struct blk_mq_queue_data bd;
 807                 int ret;
 808
 809                 rq = list_first_entry(&rq_list, struct request, queuelist);
 810                 list_del_init(&rq->queuelist);
 811
 812                 bd.rq = rq;
 813                 bd.list = dptr;
 814                 bd.last = list_empty(&rq_list);
 815
 816                 ret = q->mq_ops->queue_rq(hctx, &bd);
 817                 switch (ret) {
 818                 case BLK_MQ_RQ_QUEUE_OK:
 819                         queued++;
 820                         continue;
 821                 case BLK_MQ_RQ_QUEUE_BUSY:
 822                         list_add(&rq->queuelist, &rq_list);
 823                         __blk_mq_requeue_request(rq);
 824                         break;
 825                 default:
 826                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 827                 case BLK_MQ_RQ_QUEUE_ERROR:
 828                         rq->errors = -EIO;
 829                         blk_mq_end_request(rq, rq->errors);
 830                         break;
 831                 }
 832
 833                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 834                         break;
 835
 836                 /*
 837                  * We've done the first request. If we have more than 1
 838                  * left in the list, set dptr to defer issue.
 839                  */
 840                 if (!dptr && rq_list.next != rq_list.prev)
 841                         dptr = &driver_list;
 842         }
 843
 844         if (!queued)
 845                 hctx->dispatched[0]++;
 846         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 847                 hctx->dispatched[ilog2(queued) + 1]++;
 848
 849         /*
 850          * Any items that need requeuing? Stuff them into hctx->dispatch,
 851          * that is where we will continue on next queue run.
 852          */
 853         if (!list_empty(&rq_list)) {
 854                 spin_lock(&hctx->lock);
 855                 list_splice(&rq_list, &hctx->dispatch);
 856                 spin_unlock(&hctx->lock);
 857                 /*
 858                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 859                  * it's possible the queue is stopped and restarted again
 860                  * before this. Queue restart will dispatch requests. And since
 861                  * requests in rq_list aren't added into hctx->dispatch yet,
 862                  * the requests in rq_list might get lost.
 863                  *
 864                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 865                  **/
 866                 blk_mq_run_hw_queue(hctx, true);
 867         }
 868 }
 869
 870 /*
 871  * It'd be great if the workqueue API had a way to pass
 872  * in a mask and had some smarts for more clever placement.
 873  * For now we just round-robin here, switching for every
 874  * BLK_MQ_CPU_WORK_BATCH queued items.
 875  */
 876 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 877 {
 878         if (hctx->queue->nr_hw_queues == 1)
 879                 return WORK_CPU_UNBOUND;
 880
 881         if (--hctx->next_cpu_batch <= 0) {
 882                 int cpu = hctx->next_cpu, next_cpu;
 883
 884                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 885                 if (next_cpu >= nr_cpu_ids)
 886                         next_cpu = cpumask_first(hctx->cpumask);
 887
 888                 hctx->next_cpu = next_cpu;
 889                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 890
 891                 return cpu;
 892         }
 893
 894         return hctx->next_cpu;
 895 }
 896
 897 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 898 {
 899         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 900             !blk_mq_hw_queue_mapped(hctx)))
 901                 return;
 902
 903         if (!async) {
 904                 int cpu = get_cpu();
 905                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 906                         __blk_mq_run_hw_queue(hctx);
 907                         put_cpu();
 908                         return;
 909                 }
 910
 911                 put_cpu();
 912         }
 913
 914         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 915                         &hctx->run_work, 0);
 916 }
 917
 918 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 919 {
 920         struct blk_mq_hw_ctx *hctx;
 921         int i;
 922
 923         queue_for_each_hw_ctx(q, hctx, i) {
 924                 if ((!blk_mq_hctx_has_pending(hctx) &&
 925                     list_empty_careful(&hctx->dispatch)) ||
 926                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 927                         continue;
 928
 929                 blk_mq_run_hw_queue(hctx, async);
 930         }
 931 }
 932 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 933
 934 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 935 {
 936         cancel_delayed_work(&hctx->run_work);
 937         cancel_delayed_work(&hctx->delay_work);
 938         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 939 }
 940 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 941
 942 void blk_mq_stop_hw_queues(struct request_queue *q)
 943 {
 944         struct blk_mq_hw_ctx *hctx;
 945         int i;
 946
 947         queue_for_each_hw_ctx(q, hctx, i)
 948                 blk_mq_stop_hw_queue(hctx);
 949 }
 950 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 951
 952 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 953 {
 954         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 955
 956         blk_mq_run_hw_queue(hctx, false);
 957 }
 958 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 959
 960 void blk_mq_start_hw_queues(struct request_queue *q)
 961 {
 962         struct blk_mq_hw_ctx *hctx;
 963         int i;
 964
 965         queue_for_each_hw_ctx(q, hctx, i)
 966                 blk_mq_start_hw_queue(hctx);
 967 }
 968 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 969
 970 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 971 {
 972         struct blk_mq_hw_ctx *hctx;
 973         int i;
 974
 975         queue_for_each_hw_ctx(q, hctx, i) {
 976                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 977                         continue;
 978
 979                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 980                 blk_mq_run_hw_queue(hctx, async);
 981         }
 982 }
 983 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 984
 985 static void blk_mq_run_work_fn(struct work_struct *work)
 986 {
 987         struct blk_mq_hw_ctx *hctx;
 988
 989         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 990
 991         __blk_mq_run_hw_queue(hctx);
 992 }
 993
 994 static void blk_mq_delay_work_fn(struct work_struct *work)
 995 {
 996         struct blk_mq_hw_ctx *hctx;
 997
 998         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 999
1000         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1001                 __blk_mq_run_hw_queue(hctx);
1002 }
1003
1004 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1005 {
1006         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1007                 return;
1008
1009         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1010                         &hctx->delay_work, msecs_to_jiffies(msecs));
1011 }
1012 EXPORT_SYMBOL(blk_mq_delay_queue);
1013
1014 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1015                                     struct request *rq, bool at_head)
1016 {
1017         struct blk_mq_ctx *ctx = rq->mq_ctx;
1018
1019         trace_block_rq_insert(hctx->queue, rq);
1020
1021         if (at_head)
1022                 list_add(&rq->queuelist, &ctx->rq_list);
1023         else
1024                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1025
1026         blk_mq_hctx_mark_pending(hctx, ctx);
1027 }
1028
1029 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1030                 bool async)
1031 {
1032         struct request_queue *q = rq->q;
1033         struct blk_mq_hw_ctx *hctx;
1034         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1035
1036         current_ctx = blk_mq_get_ctx(q);
1037         if (!cpu_online(ctx->cpu))
1038                 rq->mq_ctx = ctx = current_ctx;
1039
1040         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1041
1042         spin_lock(&ctx->lock);
1043         __blk_mq_insert_request(hctx, rq, at_head);
1044         spin_unlock(&ctx->lock);
1045
1046         if (run_queue)
1047                 blk_mq_run_hw_queue(hctx, async);
1048
1049         blk_mq_put_ctx(current_ctx);
1050 }
1051
1052 static void blk_mq_insert_requests(struct request_queue *q,
1053                                      struct blk_mq_ctx *ctx,
1054                                      struct list_head *list,
1055                                      int depth,
1056                                      bool from_schedule)
1057
1058 {
1059         struct blk_mq_hw_ctx *hctx;
1060         struct blk_mq_ctx *current_ctx;
1061
1062         trace_block_unplug(q, depth, !from_schedule);
1063
1064         current_ctx = blk_mq_get_ctx(q);
1065
1066         if (!cpu_online(ctx->cpu))
1067                 ctx = current_ctx;
1068         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1069
1070         /*
1071          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1072          * offline now
1073          */
1074         spin_lock(&ctx->lock);
1075         while (!list_empty(list)) {
1076                 struct request *rq;
1077
1078                 rq = list_first_entry(list, struct request, queuelist);
1079                 list_del_init(&rq->queuelist);
1080                 rq->mq_ctx = ctx;
1081                 __blk_mq_insert_request(hctx, rq, false);
1082         }
1083         spin_unlock(&ctx->lock);
1084
1085         blk_mq_run_hw_queue(hctx, from_schedule);
1086         blk_mq_put_ctx(current_ctx);
1087 }
1088
1089 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1090 {
1091         struct request *rqa = container_of(a, struct request, queuelist);
1092         struct request *rqb = container_of(b, struct request, queuelist);
1093
1094         return !(rqa->mq_ctx < rqb->mq_ctx ||
1095                  (rqa->mq_ctx == rqb->mq_ctx &&
1096                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1097 }
1098
1099 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1100 {
1101         struct blk_mq_ctx *this_ctx;
1102         struct request_queue *this_q;
1103         struct request *rq;
1104         LIST_HEAD(list);
1105         LIST_HEAD(ctx_list);
1106         unsigned int depth;
1107
1108         list_splice_init(&plug->mq_list, &list);
1109
1110         list_sort(NULL, &list, plug_ctx_cmp);
1111
1112         this_q = NULL;
1113         this_ctx = NULL;
1114         depth = 0;
1115
1116         while (!list_empty(&list)) {
1117                 rq = list_entry_rq(list.next);
1118                 list_del_init(&rq->queuelist);
1119                 BUG_ON(!rq->q);
1120                 if (rq->mq_ctx != this_ctx) {
1121                         if (this_ctx) {
1122                                 blk_mq_insert_requests(this_q, this_ctx,
1123                                                         &ctx_list, depth,
1124                                                         from_schedule);
1125                         }
1126
1127                         this_ctx = rq->mq_ctx;
1128                         this_q = rq->q;
1129                         depth = 0;
1130                 }
1131
1132                 depth++;
1133                 list_add_tail(&rq->queuelist, &ctx_list);
1134         }
1135
1136         /*
1137          * If 'this_ctx' is set, we know we have entries to complete
1138          * on 'ctx_list'. Do those.
1139          */
1140         if (this_ctx) {
1141                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1142                                        from_schedule);
1143         }
1144 }
1145
1146 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1147 {
1148         init_request_from_bio(rq, bio);
1149
1150         if (blk_do_io_stat(rq))
1151                 blk_account_io_start(rq, 1);
1152 }
1153
1154 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1155 {
1156         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1157                 !blk_queue_nomerges(hctx->queue);
1158 }
1159
1160 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1161                                          struct blk_mq_ctx *ctx,
1162                                          struct request *rq, struct bio *bio)
1163 {
1164         if (!hctx_allow_merges(hctx)) {
1165                 blk_mq_bio_to_request(rq, bio);
1166                 spin_lock(&ctx->lock);
1167 insert_rq:
1168                 __blk_mq_insert_request(hctx, rq, false);
1169                 spin_unlock(&ctx->lock);
1170                 return false;
1171         } else {
1172                 struct request_queue *q = hctx->queue;
1173
1174                 spin_lock(&ctx->lock);
1175                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1176                         blk_mq_bio_to_request(rq, bio);
1177                         goto insert_rq;
1178                 }
1179
1180                 spin_unlock(&ctx->lock);
1181                 __blk_mq_free_request(hctx, ctx, rq);
1182                 return true;
1183         }
1184 }
1185
1186 struct blk_map_ctx {
1187         struct blk_mq_hw_ctx *hctx;
1188         struct blk_mq_ctx *ctx;
1189 };
1190
1191 static struct request *blk_mq_map_request(struct request_queue *q,
1192                                           struct bio *bio,
1193                                           struct blk_map_ctx *data)
1194 {
1195         struct blk_mq_hw_ctx *hctx;
1196         struct blk_mq_ctx *ctx;
1197         struct request *rq;
1198         int rw = bio_data_dir(bio);
1199         struct blk_mq_alloc_data alloc_data;
1200
1201         if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1202                 bio_endio(bio, -EIO);
1203                 return NULL;
1204         }
1205
1206         ctx = blk_mq_get_ctx(q);
1207         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1208
1209         if (rw_is_sync(bio->bi_rw))
1210                 rw |= REQ_SYNC;
1211
1212         trace_block_getrq(q, bio, rw);
1213         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1214                         hctx);
1215         rq = __blk_mq_alloc_request(&alloc_data, rw);
1216         if (unlikely(!rq)) {
1217                 __blk_mq_run_hw_queue(hctx);
1218                 blk_mq_put_ctx(ctx);
1219                 trace_block_sleeprq(q, bio, rw);
1220
1221                 ctx = blk_mq_get_ctx(q);
1222                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1223                 blk_mq_set_alloc_data(&alloc_data, q,
1224                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1225                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1226                 ctx = alloc_data.ctx;
1227                 hctx = alloc_data.hctx;
1228         }
1229
1230         hctx->queued++;
1231         data->hctx = hctx;
1232         data->ctx = ctx;
1233         return rq;
1234 }
1235
1236 static int blk_mq_direct_issue_request(struct request *rq)
1237 {
1238         int ret;
1239         struct request_queue *q = rq->q;
1240         struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1241                         rq->mq_ctx->cpu);
1242         struct blk_mq_queue_data bd = {
1243                 .rq = rq,
1244                 .list = NULL,
1245                 .last = 1
1246         };
1247
1248         /*
1249          * For OK queue, we are done. For error, kill it. Any other
1250          * error (busy), just add it to our list as we previously
1251          * would have done
1252          */
1253         ret = q->mq_ops->queue_rq(hctx, &bd);
1254         if (ret == BLK_MQ_RQ_QUEUE_OK)
1255                 return 0;
1256         else {
1257                 __blk_mq_requeue_request(rq);
1258
1259                 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1260                         rq->errors = -EIO;
1261                         blk_mq_end_request(rq, rq->errors);
1262                         return 0;
1263                 }
1264                 return -1;
1265         }
1266 }
1267
1268 /*
1269  * Multiple hardware queue variant. This will not use per-process plugs,
1270  * but will attempt to bypass the hctx queueing if we can go straight to
1271  * hardware for SYNC IO.
1272  */
1273 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1274 {
1275         const int is_sync = rw_is_sync(bio->bi_rw);
1276         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1277         struct blk_map_ctx data;
1278         struct request *rq;
1279         unsigned int request_count = 0;
1280         struct blk_plug *plug;
1281         struct request *same_queue_rq = NULL;
1282
1283         blk_queue_bounce(q, &bio);
1284
1285         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1286                 bio_endio(bio, -EIO);
1287                 return;
1288         }
1289
1290         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1291             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1292                 return;
1293
1294         rq = blk_mq_map_request(q, bio, &data);
1295         if (unlikely(!rq))
1296                 return;
1297
1298         if (unlikely(is_flush_fua)) {
1299                 blk_mq_bio_to_request(rq, bio);
1300                 blk_insert_flush(rq);
1301                 goto run_queue;
1302         }
1303
1304         plug = current->plug;
1305         /*
1306          * If the driver supports defer issued based on 'last', then
1307          * queue it up like normal since we can potentially save some
1308          * CPU this way.
1309          */
1310         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1311             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1312                 struct request *old_rq = NULL;
1313
1314                 blk_mq_bio_to_request(rq, bio);
1315
1316                 /*
1317                  * we do limited pluging. If bio can be merged, do merge.
1318                  * Otherwise the existing request in the plug list will be
1319                  * issued. So the plug list will have one request at most
1320                  */
1321                 if (plug) {
1322                         /*
1323                          * The plug list might get flushed before this. If that
1324                          * happens, same_queue_rq is invalid and plug list is empty
1325                          **/
1326                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1327                                 old_rq = same_queue_rq;
1328                                 list_del_init(&old_rq->queuelist);
1329                         }
1330                         list_add_tail(&rq->queuelist, &plug->mq_list);
1331                 } else /* is_sync */
1332                         old_rq = rq;
1333                 blk_mq_put_ctx(data.ctx);
1334                 if (!old_rq)
1335                         return;
1336                 if (!blk_mq_direct_issue_request(old_rq))
1337                         return;
1338                 blk_mq_insert_request(old_rq, false, true, true);
1339                 return;
1340         }
1341
1342         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1343                 /*
1344                  * For a SYNC request, send it to the hardware immediately. For
1345                  * an ASYNC request, just ensure that we run it later on. The
1346                  * latter allows for merging opportunities and more efficient
1347                  * dispatching.
1348                  */
1349 run_queue:
1350                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1351         }
1352         blk_mq_put_ctx(data.ctx);
1353 }
1354
1355 /*
1356  * Single hardware queue variant. This will attempt to use any per-process
1357  * plug for merging and IO deferral.
1358  */
1359 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1360 {
1361         const int is_sync = rw_is_sync(bio->bi_rw);
1362         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1363         struct blk_plug *plug;
1364         unsigned int request_count = 0;
1365         struct blk_map_ctx data;
1366         struct request *rq;
1367
1368         blk_queue_bounce(q, &bio);
1369
1370         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1371                 bio_endio(bio, -EIO);
1372                 return;
1373         }
1374
1375         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1376             blk_attempt_plug_merge(q, bio, &request_count, NULL))
1377                 return;
1378
1379         rq = blk_mq_map_request(q, bio, &data);
1380         if (unlikely(!rq))
1381                 return;
1382
1383         if (unlikely(is_flush_fua)) {
1384                 blk_mq_bio_to_request(rq, bio);
1385                 blk_insert_flush(rq);
1386                 goto run_queue;
1387         }
1388
1389         /*
1390          * A task plug currently exists. Since this is completely lockless,
1391          * utilize that to temporarily store requests until the task is
1392          * either done or scheduled away.
1393          */
1394         plug = current->plug;
1395         if (plug) {
1396                 blk_mq_bio_to_request(rq, bio);
1397                 if (list_empty(&plug->mq_list))
1398                         trace_block_plug(q);
1399                 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1400                         blk_flush_plug_list(plug, false);
1401                         trace_block_plug(q);
1402                 }
1403                 list_add_tail(&rq->queuelist, &plug->mq_list);
1404                 blk_mq_put_ctx(data.ctx);
1405                 return;
1406         }
1407
1408         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1409                 /*
1410                  * For a SYNC request, send it to the hardware immediately. For
1411                  * an ASYNC request, just ensure that we run it later on. The
1412                  * latter allows for merging opportunities and more efficient
1413                  * dispatching.
1414                  */
1415 run_queue:
1416                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1417         }
1418
1419         blk_mq_put_ctx(data.ctx);
1420 }
1421
1422 /*
1423  * Default mapping to a software queue, since we use one per CPU.
1424  */
1425 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1426 {
1427         return q->queue_hw_ctx[q->mq_map[cpu]];
1428 }
1429 EXPORT_SYMBOL(blk_mq_map_queue);
1430
1431 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1432                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1433 {
1434         struct page *page;
1435
1436         if (tags->rqs && set->ops->exit_request) {
1437                 int i;
1438
1439                 for (i = 0; i < tags->nr_tags; i++) {
1440                         if (!tags->rqs[i])
1441                                 continue;
1442                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1443                                                 hctx_idx, i);
1444                         tags->rqs[i] = NULL;
1445                 }
1446         }
1447
1448         while (!list_empty(&tags->page_list)) {
1449                 page = list_first_entry(&tags->page_list, struct page, lru);
1450                 list_del_init(&page->lru);
1451                 __free_pages(page, page->private);
1452         }
1453
1454         kfree(tags->rqs);
1455
1456         blk_mq_free_tags(tags);
1457 }
1458
1459 static size_t order_to_size(unsigned int order)
1460 {
1461         return (size_t)PAGE_SIZE << order;
1462 }
1463
1464 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1465                 unsigned int hctx_idx)
1466 {
1467         struct blk_mq_tags *tags;
1468         unsigned int i, j, entries_per_page, max_order = 4;
1469         size_t rq_size, left;
1470
1471         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1472                                 set->numa_node,
1473                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1474         if (!tags)
1475                 return NULL;
1476
1477         INIT_LIST_HEAD(&tags->page_list);
1478
1479         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1480                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1481                                  set->numa_node);
1482         if (!tags->rqs) {
1483                 blk_mq_free_tags(tags);
1484                 return NULL;
1485         }
1486
1487         /*
1488          * rq_size is the size of the request plus driver payload, rounded
1489          * to the cacheline size
1490          */
1491         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1492                                 cache_line_size());
1493         left = rq_size * set->queue_depth;
1494
1495         for (i = 0; i < set->queue_depth; ) {
1496                 int this_order = max_order;
1497                 struct page *page;
1498                 int to_do;
1499                 void *p;
1500
1501                 while (left < order_to_size(this_order - 1) && this_order)
1502                         this_order--;
1503
1504                 do {
1505                         page = alloc_pages_node(set->numa_node,
1506                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1507                                 this_order);
1508                         if (page)
1509                                 break;
1510                         if (!this_order--)
1511                                 break;
1512                         if (order_to_size(this_order) < rq_size)
1513                                 break;
1514                 } while (1);
1515
1516                 if (!page)
1517                         goto fail;
1518
1519                 page->private = this_order;
1520                 list_add_tail(&page->lru, &tags->page_list);
1521
1522                 p = page_address(page);
1523                 entries_per_page = order_to_size(this_order) / rq_size;
1524                 to_do = min(entries_per_page, set->queue_depth - i);
1525                 left -= to_do * rq_size;
1526                 for (j = 0; j < to_do; j++) {
1527                         tags->rqs[i] = p;
1528                         if (set->ops->init_request) {
1529                                 if (set->ops->init_request(set->driver_data,
1530                                                 tags->rqs[i], hctx_idx, i,
1531                                                 set->numa_node)) {
1532                                         tags->rqs[i] = NULL;
1533                                         goto fail;
1534                                 }
1535                         }
1536
1537                         p += rq_size;
1538                         i++;
1539                 }
1540         }
1541         return tags;
1542
1543 fail:
1544         blk_mq_free_rq_map(set, tags, hctx_idx);
1545         return NULL;
1546 }
1547
1548 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1549 {
1550         kfree(bitmap->map);
1551 }
1552
1553 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1554 {
1555         unsigned int bpw = 8, total, num_maps, i;
1556
1557         bitmap->bits_per_word = bpw;
1558
1559         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1560         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1561                                         GFP_KERNEL, node);
1562         if (!bitmap->map)
1563                 return -ENOMEM;
1564
1565         total = nr_cpu_ids;
1566         for (i = 0; i < num_maps; i++) {
1567                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1568                 total -= bitmap->map[i].depth;
1569         }
1570
1571         return 0;
1572 }
1573
1574 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1575 {
1576         struct request_queue *q = hctx->queue;
1577         struct blk_mq_ctx *ctx;
1578         LIST_HEAD(tmp);
1579
1580         /*
1581          * Move ctx entries to new CPU, if this one is going away.
1582          */
1583         ctx = __blk_mq_get_ctx(q, cpu);
1584
1585         spin_lock(&ctx->lock);
1586         if (!list_empty(&ctx->rq_list)) {
1587                 list_splice_init(&ctx->rq_list, &tmp);
1588                 blk_mq_hctx_clear_pending(hctx, ctx);
1589         }
1590         spin_unlock(&ctx->lock);
1591
1592         if (list_empty(&tmp))
1593                 return NOTIFY_OK;
1594
1595         ctx = blk_mq_get_ctx(q);
1596         spin_lock(&ctx->lock);
1597
1598         while (!list_empty(&tmp)) {
1599                 struct request *rq;
1600
1601                 rq = list_first_entry(&tmp, struct request, queuelist);
1602                 rq->mq_ctx = ctx;
1603                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1604         }
1605
1606         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1607         blk_mq_hctx_mark_pending(hctx, ctx);
1608
1609         spin_unlock(&ctx->lock);
1610
1611         blk_mq_run_hw_queue(hctx, true);
1612         blk_mq_put_ctx(ctx);
1613         return NOTIFY_OK;
1614 }
1615
1616 static int blk_mq_hctx_notify(void *data, unsigned long action,
1617                               unsigned int cpu)
1618 {
1619         struct blk_mq_hw_ctx *hctx = data;
1620
1621         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1622                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1623
1624         /*
1625          * In case of CPU online, tags may be reallocated
1626          * in blk_mq_map_swqueue() after mapping is updated.
1627          */
1628
1629         return NOTIFY_OK;
1630 }
1631
1632 /* hctx->ctxs will be freed in queue's release handler */
1633 static void blk_mq_exit_hctx(struct request_queue *q,
1634                 struct blk_mq_tag_set *set,
1635                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1636 {
1637         unsigned flush_start_tag = set->queue_depth;
1638
1639         blk_mq_tag_idle(hctx);
1640
1641         if (set->ops->exit_request)
1642                 set->ops->exit_request(set->driver_data,
1643                                        hctx->fq->flush_rq, hctx_idx,
1644                                        flush_start_tag + hctx_idx);
1645
1646         if (set->ops->exit_hctx)
1647                 set->ops->exit_hctx(hctx, hctx_idx);
1648
1649         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1650         blk_free_flush_queue(hctx->fq);
1651         blk_mq_free_bitmap(&hctx->ctx_map);
1652 }
1653
1654 static void blk_mq_exit_hw_queues(struct request_queue *q,
1655                 struct blk_mq_tag_set *set, int nr_queue)
1656 {
1657         struct blk_mq_hw_ctx *hctx;
1658         unsigned int i;
1659
1660         queue_for_each_hw_ctx(q, hctx, i) {
1661                 if (i == nr_queue)
1662                         break;
1663                 blk_mq_exit_hctx(q, set, hctx, i);
1664         }
1665 }
1666
1667 static void blk_mq_free_hw_queues(struct request_queue *q,
1668                 struct blk_mq_tag_set *set)
1669 {
1670         struct blk_mq_hw_ctx *hctx;
1671         unsigned int i;
1672
1673         queue_for_each_hw_ctx(q, hctx, i)
1674                 free_cpumask_var(hctx->cpumask);
1675 }
1676
1677 static int blk_mq_init_hctx(struct request_queue *q,
1678                 struct blk_mq_tag_set *set,
1679                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1680 {
1681         int node;
1682         unsigned flush_start_tag = set->queue_depth;
1683
1684         node = hctx->numa_node;
1685         if (node == NUMA_NO_NODE)
1686                 node = hctx->numa_node = set->numa_node;
1687
1688         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1689         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1690         spin_lock_init(&hctx->lock);
1691         INIT_LIST_HEAD(&hctx->dispatch);
1692         hctx->queue = q;
1693         hctx->queue_num = hctx_idx;
1694         hctx->flags = set->flags;
1695
1696         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1697                                         blk_mq_hctx_notify, hctx);
1698         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1699
1700         hctx->tags = set->tags[hctx_idx];
1701
1702         /*
1703          * Allocate space for all possible cpus to avoid allocation at
1704          * runtime
1705          */
1706         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1707                                         GFP_KERNEL, node);
1708         if (!hctx->ctxs)
1709                 goto unregister_cpu_notifier;
1710
1711         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1712                 goto free_ctxs;
1713
1714         hctx->nr_ctx = 0;
1715
1716         if (set->ops->init_hctx &&
1717             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1718                 goto free_bitmap;
1719
1720         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1721         if (!hctx->fq)
1722                 goto exit_hctx;
1723
1724         if (set->ops->init_request &&
1725             set->ops->init_request(set->driver_data,
1726                                    hctx->fq->flush_rq, hctx_idx,
1727                                    flush_start_tag + hctx_idx, node))
1728                 goto free_fq;
1729
1730         return 0;
1731
1732  free_fq:
1733         kfree(hctx->fq);
1734  exit_hctx:
1735         if (set->ops->exit_hctx)
1736                 set->ops->exit_hctx(hctx, hctx_idx);
1737  free_bitmap:
1738         blk_mq_free_bitmap(&hctx->ctx_map);
1739  free_ctxs:
1740         kfree(hctx->ctxs);
1741  unregister_cpu_notifier:
1742         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1743
1744         return -1;
1745 }
1746
1747 static int blk_mq_init_hw_queues(struct request_queue *q,
1748                 struct blk_mq_tag_set *set)
1749 {
1750         struct blk_mq_hw_ctx *hctx;
1751         unsigned int i;
1752
1753         /*
1754          * Initialize hardware queues
1755          */
1756         queue_for_each_hw_ctx(q, hctx, i) {
1757                 if (blk_mq_init_hctx(q, set, hctx, i))
1758                         break;
1759         }
1760
1761         if (i == q->nr_hw_queues)
1762                 return 0;
1763
1764         /*
1765          * Init failed
1766          */
1767         blk_mq_exit_hw_queues(q, set, i);
1768
1769         return 1;
1770 }
1771
1772 static void blk_mq_init_cpu_queues(struct request_queue *q,
1773                                    unsigned int nr_hw_queues)
1774 {
1775         unsigned int i;
1776
1777         for_each_possible_cpu(i) {
1778                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1779                 struct blk_mq_hw_ctx *hctx;
1780
1781                 memset(__ctx, 0, sizeof(*__ctx));
1782                 __ctx->cpu = i;
1783                 spin_lock_init(&__ctx->lock);
1784                 INIT_LIST_HEAD(&__ctx->rq_list);
1785                 __ctx->queue = q;
1786
1787                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1788                 if (!cpu_online(i))
1789                         continue;
1790
1791                 hctx = q->mq_ops->map_queue(q, i);
1792
1793                 /*
1794                  * Set local node, IFF we have more than one hw queue. If
1795                  * not, we remain on the home node of the device
1796                  */
1797                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1798                         hctx->numa_node = cpu_to_node(i);
1799         }
1800 }
1801
1802 static void blk_mq_map_swqueue(struct request_queue *q)
1803 {
1804         unsigned int i;
1805         struct blk_mq_hw_ctx *hctx;
1806         struct blk_mq_ctx *ctx;
1807         struct blk_mq_tag_set *set = q->tag_set;
1808
1809         queue_for_each_hw_ctx(q, hctx, i) {
1810                 cpumask_clear(hctx->cpumask);
1811                 hctx->nr_ctx = 0;
1812         }
1813
1814         /*
1815          * Map software to hardware queues
1816          */
1817         queue_for_each_ctx(q, ctx, i) {
1818                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1819                 if (!cpu_online(i))
1820                         continue;
1821
1822                 hctx = q->mq_ops->map_queue(q, i);
1823                 cpumask_set_cpu(i, hctx->cpumask);
1824                 cpumask_set_cpu(i, hctx->tags->cpumask);
1825                 ctx->index_hw = hctx->nr_ctx;
1826                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1827         }
1828
1829         queue_for_each_hw_ctx(q, hctx, i) {
1830                 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1831
1832                 /*
1833                  * If no software queues are mapped to this hardware queue,
1834                  * disable it and free the request entries.
1835                  */
1836                 if (!hctx->nr_ctx) {
1837                         if (set->tags[i]) {
1838                                 blk_mq_free_rq_map(set, set->tags[i], i);
1839                                 set->tags[i] = NULL;
1840                         }
1841                         hctx->tags = NULL;
1842                         continue;
1843                 }
1844
1845                 /* unmapped hw queue can be remapped after CPU topo changed */
1846                 if (!set->tags[i])
1847                         set->tags[i] = blk_mq_init_rq_map(set, i);
1848                 hctx->tags = set->tags[i];
1849                 WARN_ON(!hctx->tags);
1850
1851                 /*
1852                  * Set the map size to the number of mapped software queues.
1853                  * This is more accurate and more efficient than looping
1854                  * over all possibly mapped software queues.
1855                  */
1856                 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1857
1858                 /*
1859                  * Initialize batch roundrobin counts
1860                  */
1861                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1862                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1863         }
1864 }
1865
1866 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1867 {
1868         struct blk_mq_hw_ctx *hctx;
1869         struct request_queue *q;
1870         bool shared;
1871         int i;
1872
1873         if (set->tag_list.next == set->tag_list.prev)
1874                 shared = false;
1875         else
1876                 shared = true;
1877
1878         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1879                 blk_mq_freeze_queue(q);
1880
1881                 queue_for_each_hw_ctx(q, hctx, i) {
1882                         if (shared)
1883                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1884                         else
1885                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1886                 }
1887                 blk_mq_unfreeze_queue(q);
1888         }
1889 }
1890
1891 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1892 {
1893         struct blk_mq_tag_set *set = q->tag_set;
1894
1895         mutex_lock(&set->tag_list_lock);
1896         list_del_init(&q->tag_set_list);
1897         blk_mq_update_tag_set_depth(set);
1898         mutex_unlock(&set->tag_list_lock);
1899 }
1900
1901 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1902                                      struct request_queue *q)
1903 {
1904         q->tag_set = set;
1905
1906         mutex_lock(&set->tag_list_lock);
1907         list_add_tail(&q->tag_set_list, &set->tag_list);
1908         blk_mq_update_tag_set_depth(set);
1909         mutex_unlock(&set->tag_list_lock);
1910 }
1911
1912 /*
1913  * It is the actual release handler for mq, but we do it from
1914  * request queue's release handler for avoiding use-after-free
1915  * and headache because q->mq_kobj shouldn't have been introduced,
1916  * but we can't group ctx/kctx kobj without it.
1917  */
1918 void blk_mq_release(struct request_queue *q)
1919 {
1920         struct blk_mq_hw_ctx *hctx;
1921         unsigned int i;
1922
1923         /* hctx kobj stays in hctx */
1924         queue_for_each_hw_ctx(q, hctx, i) {
1925                 if (!hctx)
1926                         continue;
1927                 kfree(hctx->ctxs);
1928                 kfree(hctx);
1929         }
1930
1931         kfree(q->queue_hw_ctx);
1932
1933         /* ctx kobj stays in queue_ctx */
1934         free_percpu(q->queue_ctx);
1935 }
1936
1937 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1938 {
1939         struct request_queue *uninit_q, *q;
1940
1941         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1942         if (!uninit_q)
1943                 return ERR_PTR(-ENOMEM);
1944
1945         q = blk_mq_init_allocated_queue(set, uninit_q);
1946         if (IS_ERR(q))
1947                 blk_cleanup_queue(uninit_q);
1948
1949         return q;
1950 }
1951 EXPORT_SYMBOL(blk_mq_init_queue);
1952
1953 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1954                                                   struct request_queue *q)
1955 {
1956         struct blk_mq_hw_ctx **hctxs;
1957         struct blk_mq_ctx __percpu *ctx;
1958         unsigned int *map;
1959         int i;
1960
1961         ctx = alloc_percpu(struct blk_mq_ctx);
1962         if (!ctx)
1963                 return ERR_PTR(-ENOMEM);
1964
1965         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1966                         set->numa_node);
1967
1968         if (!hctxs)
1969                 goto err_percpu;
1970
1971         map = blk_mq_make_queue_map(set);
1972         if (!map)
1973                 goto err_map;
1974
1975         for (i = 0; i < set->nr_hw_queues; i++) {
1976                 int node = blk_mq_hw_queue_to_node(map, i);
1977
1978                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1979                                         GFP_KERNEL, node);
1980                 if (!hctxs[i])
1981                         goto err_hctxs;
1982
1983                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1984                                                 node))
1985                         goto err_hctxs;
1986
1987                 atomic_set(&hctxs[i]->nr_active, 0);
1988                 hctxs[i]->numa_node = node;
1989                 hctxs[i]->queue_num = i;
1990         }
1991
1992         /*
1993          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1994          * See blk_register_queue() for details.
1995          */
1996         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1997                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1998                 goto err_hctxs;
1999
2000         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2001         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
2002
2003         q->nr_queues = nr_cpu_ids;
2004         q->nr_hw_queues = set->nr_hw_queues;
2005         q->mq_map = map;
2006
2007         q->queue_ctx = ctx;
2008         q->queue_hw_ctx = hctxs;
2009
2010         q->mq_ops = set->ops;
2011         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2012
2013         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2014                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2015
2016         q->sg_reserved_size = INT_MAX;
2017
2018         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2019         INIT_LIST_HEAD(&q->requeue_list);
2020         spin_lock_init(&q->requeue_lock);
2021
2022         if (q->nr_hw_queues > 1)
2023                 blk_queue_make_request(q, blk_mq_make_request);
2024         else
2025                 blk_queue_make_request(q, blk_sq_make_request);
2026
2027         /*
2028          * Do this after blk_queue_make_request() overrides it...
2029          */
2030         q->nr_requests = set->queue_depth;
2031
2032         if (set->ops->complete)
2033                 blk_queue_softirq_done(q, set->ops->complete);
2034
2035         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2036
2037         if (blk_mq_init_hw_queues(q, set))
2038                 goto err_hctxs;
2039
2040         mutex_lock(&all_q_mutex);
2041         list_add_tail(&q->all_q_node, &all_q_list);
2042         mutex_unlock(&all_q_mutex);
2043
2044         blk_mq_add_queue_tag_set(set, q);
2045
2046         blk_mq_map_swqueue(q);
2047
2048         return q;
2049
2050 err_hctxs:
2051         kfree(map);
2052         for (i = 0; i < set->nr_hw_queues; i++) {
2053                 if (!hctxs[i])
2054                         break;
2055                 free_cpumask_var(hctxs[i]->cpumask);
2056                 kfree(hctxs[i]);
2057         }
2058 err_map:
2059         kfree(hctxs);
2060 err_percpu:
2061         free_percpu(ctx);
2062         return ERR_PTR(-ENOMEM);
2063 }
2064 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2065
2066 void blk_mq_free_queue(struct request_queue *q)
2067 {
2068         struct blk_mq_tag_set   *set = q->tag_set;
2069
2070         blk_mq_del_queue_tag_set(q);
2071
2072         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2073         blk_mq_free_hw_queues(q, set);
2074
2075         percpu_ref_exit(&q->mq_usage_counter);
2076
2077         kfree(q->mq_map);
2078
2079         q->mq_map = NULL;
2080
2081         mutex_lock(&all_q_mutex);
2082         list_del_init(&q->all_q_node);
2083         mutex_unlock(&all_q_mutex);
2084 }
2085
2086 /* Basically redo blk_mq_init_queue with queue frozen */
2087 static void blk_mq_queue_reinit(struct request_queue *q)
2088 {
2089         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2090
2091         blk_mq_sysfs_unregister(q);
2092
2093         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2094
2095         /*
2096          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2097          * we should change hctx numa_node according to new topology (this
2098          * involves free and re-allocate memory, worthy doing?)
2099          */
2100
2101         blk_mq_map_swqueue(q);
2102
2103         blk_mq_sysfs_register(q);
2104 }
2105
2106 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2107                                       unsigned long action, void *hcpu)
2108 {
2109         struct request_queue *q;
2110
2111         /*
2112          * Before new mappings are established, hotadded cpu might already
2113          * start handling requests. This doesn't break anything as we map
2114          * offline CPUs to first hardware queue. We will re-init the queue
2115          * below to get optimal settings.
2116          */
2117         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2118             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2119                 return NOTIFY_OK;
2120
2121         mutex_lock(&all_q_mutex);
2122
2123         /*
2124          * We need to freeze and reinit all existing queues.  Freezing
2125          * involves synchronous wait for an RCU grace period and doing it
2126          * one by one may take a long time.  Start freezing all queues in
2127          * one swoop and then wait for the completions so that freezing can
2128          * take place in parallel.
2129          */
2130         list_for_each_entry(q, &all_q_list, all_q_node)
2131                 blk_mq_freeze_queue_start(q);
2132         list_for_each_entry(q, &all_q_list, all_q_node) {
2133                 blk_mq_freeze_queue_wait(q);
2134
2135                 /*
2136                  * timeout handler can't touch hw queue during the
2137                  * reinitialization
2138                  */
2139                 del_timer_sync(&q->timeout);
2140         }
2141
2142         list_for_each_entry(q, &all_q_list, all_q_node)
2143                 blk_mq_queue_reinit(q);
2144
2145         list_for_each_entry(q, &all_q_list, all_q_node)
2146                 blk_mq_unfreeze_queue(q);
2147
2148         mutex_unlock(&all_q_mutex);
2149         return NOTIFY_OK;
2150 }
2151
2152 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2153 {
2154         int i;
2155
2156         for (i = 0; i < set->nr_hw_queues; i++) {
2157                 set->tags[i] = blk_mq_init_rq_map(set, i);
2158                 if (!set->tags[i])
2159                         goto out_unwind;
2160         }
2161
2162         return 0;
2163
2164 out_unwind:
2165         while (--i >= 0)
2166                 blk_mq_free_rq_map(set, set->tags[i], i);
2167
2168         return -ENOMEM;
2169 }
2170
2171 /*
2172  * Allocate the request maps associated with this tag_set. Note that this
2173  * may reduce the depth asked for, if memory is tight. set->queue_depth
2174  * will be updated to reflect the allocated depth.
2175  */
2176 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2177 {
2178         unsigned int depth;
2179         int err;
2180
2181         depth = set->queue_depth;
2182         do {
2183                 err = __blk_mq_alloc_rq_maps(set);
2184                 if (!err)
2185                         break;
2186
2187                 set->queue_depth >>= 1;
2188                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2189                         err = -ENOMEM;
2190                         break;
2191                 }
2192         } while (set->queue_depth);
2193
2194         if (!set->queue_depth || err) {
2195                 pr_err("blk-mq: failed to allocate request map\n");
2196                 return -ENOMEM;
2197         }
2198
2199         if (depth != set->queue_depth)
2200                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2201                                                 depth, set->queue_depth);
2202
2203         return 0;
2204 }
2205
2206 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2207 {
2208         return tags->cpumask;
2209 }
2210 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2211
2212 /*
2213  * Alloc a tag set to be associated with one or more request queues.
2214  * May fail with EINVAL for various error conditions. May adjust the
2215  * requested depth down, if if it too large. In that case, the set
2216  * value will be stored in set->queue_depth.
2217  */
2218 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2219 {
2220         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2221
2222         if (!set->nr_hw_queues)
2223                 return -EINVAL;
2224         if (!set->queue_depth)
2225                 return -EINVAL;
2226         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2227                 return -EINVAL;
2228
2229         if (!set->ops->queue_rq || !set->ops->map_queue)
2230                 return -EINVAL;
2231
2232         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2233                 pr_info("blk-mq: reduced tag depth to %u\n",
2234                         BLK_MQ_MAX_DEPTH);
2235                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2236         }
2237
2238         /*
2239          * If a crashdump is active, then we are potentially in a very
2240          * memory constrained environment. Limit us to 1 queue and
2241          * 64 tags to prevent using too much memory.
2242          */
2243         if (is_kdump_kernel()) {
2244                 set->nr_hw_queues = 1;
2245                 set->queue_depth = min(64U, set->queue_depth);
2246         }
2247
2248         set->tags = kmalloc_node(set->nr_hw_queues *
2249                                  sizeof(struct blk_mq_tags *),
2250                                  GFP_KERNEL, set->numa_node);
2251         if (!set->tags)
2252                 return -ENOMEM;
2253
2254         if (blk_mq_alloc_rq_maps(set))
2255                 goto enomem;
2256
2257         mutex_init(&set->tag_list_lock);
2258         INIT_LIST_HEAD(&set->tag_list);
2259
2260         return 0;
2261 enomem:
2262         kfree(set->tags);
2263         set->tags = NULL;
2264         return -ENOMEM;
2265 }
2266 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2267
2268 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2269 {
2270         int i;
2271
2272         for (i = 0; i < set->nr_hw_queues; i++) {
2273                 if (set->tags[i]) {
2274                         blk_mq_free_rq_map(set, set->tags[i], i);
2275                         free_cpumask_var(set->tags[i]->cpumask);
2276                 }
2277         }
2278
2279         kfree(set->tags);
2280         set->tags = NULL;
2281 }
2282 EXPORT_SYMBOL(blk_mq_free_tag_set);
2283
2284 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2285 {
2286         struct blk_mq_tag_set *set = q->tag_set;
2287         struct blk_mq_hw_ctx *hctx;
2288         int i, ret;
2289
2290         if (!set || nr > set->queue_depth)
2291                 return -EINVAL;
2292
2293         ret = 0;
2294         queue_for_each_hw_ctx(q, hctx, i) {
2295                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2296                 if (ret)
2297                         break;
2298         }
2299
2300         if (!ret)
2301                 q->nr_requests = nr;
2302
2303         return ret;
2304 }
2305
2306 void blk_mq_disable_hotplug(void)
2307 {
2308         mutex_lock(&all_q_mutex);
2309 }
2310
2311 void blk_mq_enable_hotplug(void)
2312 {
2313         mutex_unlock(&all_q_mutex);
2314 }
2315
2316 static int __init blk_mq_init(void)
2317 {
2318         blk_mq_cpu_init();
2319
2320         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2321
2322         return 0;
2323 }
2324 subsys_initcall(blk_mq_init);