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