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