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/kmemleak.h>
  13 #include <linux/mm.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/smp.h>
  18 #include <linux/llist.h>
  19 #include <linux/list_sort.h>
  20 #include <linux/cpu.h>
  21 #include <linux/cache.h>
  22 #include <linux/sched/sysctl.h>
  23 #include <linux/delay.h>
  24 #include <linux/crash_dump.h>
  25 #include <linux/prefetch.h>
  26
  27 #include <trace/events/block.h>
  28
  29 #include <linux/blk-mq.h>
  30 #include "blk.h"
  31 #include "blk-mq.h"
  32 #include "blk-mq-tag.h"
  33
  34 static DEFINE_MUTEX(all_q_mutex);
  35 static LIST_HEAD(all_q_list);
  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         return sbitmap_any_bit_set(&hctx->ctx_map);
  43 }
  44
  45 /*
  46  * Mark this ctx as having pending work in this hardware queue
  47  */
  48 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  49                                      struct blk_mq_ctx *ctx)
  50 {
  51         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  52                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  53 }
  54
  55 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  56                                       struct blk_mq_ctx *ctx)
  57 {
  58         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  59 }
  60
  61 void blk_mq_freeze_queue_start(struct request_queue *q)
  62 {
  63         int freeze_depth;
  64
  65         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
  66         if (freeze_depth == 1) {
  67                 percpu_ref_kill(&q->q_usage_counter);
  68                 blk_mq_run_hw_queues(q, false);
  69         }
  70 }
  71 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
  72
  73 static void blk_mq_freeze_queue_wait(struct request_queue *q)
  74 {
  75         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
  76 }
  77
  78 /*
  79  * Guarantee no request is in use, so we can change any data structure of
  80  * the queue afterward.
  81  */
  82 void blk_freeze_queue(struct request_queue *q)
  83 {
  84         /*
  85          * In the !blk_mq case we are only calling this to kill the
  86          * q_usage_counter, otherwise this increases the freeze depth
  87          * and waits for it to return to zero.  For this reason there is
  88          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
  89          * exported to drivers as the only user for unfreeze is blk_mq.
  90          */
  91         blk_mq_freeze_queue_start(q);
  92         blk_mq_freeze_queue_wait(q);
  93 }
  94
  95 void blk_mq_freeze_queue(struct request_queue *q)
  96 {
  97         /*
  98          * ...just an alias to keep freeze and unfreeze actions balanced
  99          * in the blk_mq_* namespace
 100          */
 101         blk_freeze_queue(q);
 102 }
 103 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 104
 105 void blk_mq_unfreeze_queue(struct request_queue *q)
 106 {
 107         int freeze_depth;
 108
 109         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 110         WARN_ON_ONCE(freeze_depth < 0);
 111         if (!freeze_depth) {
 112                 percpu_ref_reinit(&q->q_usage_counter);
 113                 wake_up_all(&q->mq_freeze_wq);
 114         }
 115 }
 116 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 117
 118 void blk_mq_wake_waiters(struct request_queue *q)
 119 {
 120         struct blk_mq_hw_ctx *hctx;
 121         unsigned int i;
 122
 123         queue_for_each_hw_ctx(q, hctx, i)
 124                 if (blk_mq_hw_queue_mapped(hctx))
 125                         blk_mq_tag_wakeup_all(hctx->tags, true);
 126
 127         /*
 128          * If we are called because the queue has now been marked as
 129          * dying, we need to ensure that processes currently waiting on
 130          * the queue are notified as well.
 131          */
 132         wake_up_all(&q->mq_freeze_wq);
 133 }
 134
 135 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 136 {
 137         return blk_mq_has_free_tags(hctx->tags);
 138 }
 139 EXPORT_SYMBOL(blk_mq_can_queue);
 140
 141 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 142                                struct request *rq, int op,
 143                                unsigned int op_flags)
 144 {
 145         if (blk_queue_io_stat(q))
 146                 op_flags |= REQ_IO_STAT;
 147
 148         INIT_LIST_HEAD(&rq->queuelist);
 149         /* csd/requeue_work/fifo_time is initialized before use */
 150         rq->q = q;
 151         rq->mq_ctx = ctx;
 152         req_set_op_attrs(rq, op, op_flags);
 153         /* do not touch atomic flags, it needs atomic ops against the timer */
 154         rq->cpu = -1;
 155         INIT_HLIST_NODE(&rq->hash);
 156         RB_CLEAR_NODE(&rq->rb_node);
 157         rq->rq_disk = NULL;
 158         rq->part = NULL;
 159         rq->start_time = jiffies;
 160 #ifdef CONFIG_BLK_CGROUP
 161         rq->rl = NULL;
 162         set_start_time_ns(rq);
 163         rq->io_start_time_ns = 0;
 164 #endif
 165         rq->nr_phys_segments = 0;
 166 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 167         rq->nr_integrity_segments = 0;
 168 #endif
 169         rq->special = NULL;
 170         /* tag was already set */
 171         rq->errors = 0;
 172
 173         rq->cmd = rq->__cmd;
 174
 175         rq->extra_len = 0;
 176         rq->sense_len = 0;
 177         rq->resid_len = 0;
 178         rq->sense = NULL;
 179
 180         INIT_LIST_HEAD(&rq->timeout_list);
 181         rq->timeout = 0;
 182
 183         rq->end_io = NULL;
 184         rq->end_io_data = NULL;
 185         rq->next_rq = NULL;
 186
 187         ctx->rq_dispatched[rw_is_sync(op, op_flags)]++;
 188 }
 189
 190 static struct request *
 191 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int op, int op_flags)
 192 {
 193         struct request *rq;
 194         unsigned int tag;
 195
 196         tag = blk_mq_get_tag(data);
 197         if (tag != BLK_MQ_TAG_FAIL) {
 198                 rq = data->hctx->tags->rqs[tag];
 199
 200                 if (blk_mq_tag_busy(data->hctx)) {
 201                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 202                         atomic_inc(&data->hctx->nr_active);
 203                 }
 204
 205                 rq->tag = tag;
 206                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, op, op_flags);
 207                 return rq;
 208         }
 209
 210         return NULL;
 211 }
 212
 213 struct request *blk_mq_alloc_request(struct request_queue *q, int rw,
 214                 unsigned int flags)
 215 {
 216         struct blk_mq_ctx *ctx;
 217         struct blk_mq_hw_ctx *hctx;
 218         struct request *rq;
 219         struct blk_mq_alloc_data alloc_data;
 220         int ret;
 221
 222         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 223         if (ret)
 224                 return ERR_PTR(ret);
 225
 226         ctx = blk_mq_get_ctx(q);
 227         hctx = blk_mq_map_queue(q, ctx->cpu);
 228         blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
 229         rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
 230         blk_mq_put_ctx(ctx);
 231
 232         if (!rq) {
 233                 blk_queue_exit(q);
 234                 return ERR_PTR(-EWOULDBLOCK);
 235         }
 236
 237         rq->__data_len = 0;
 238         rq->__sector = (sector_t) -1;
 239         rq->bio = rq->biotail = NULL;
 240         return rq;
 241 }
 242 EXPORT_SYMBOL(blk_mq_alloc_request);
 243
 244 struct request *blk_mq_alloc_request_hctx(struct request_queue *q, int rw,
 245                 unsigned int flags, unsigned int hctx_idx)
 246 {
 247         struct blk_mq_hw_ctx *hctx;
 248         struct blk_mq_ctx *ctx;
 249         struct request *rq;
 250         struct blk_mq_alloc_data alloc_data;
 251         int ret;
 252
 253         /*
 254          * If the tag allocator sleeps we could get an allocation for a
 255          * different hardware context.  No need to complicate the low level
 256          * allocator for this for the rare use case of a command tied to
 257          * a specific queue.
 258          */
 259         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 260                 return ERR_PTR(-EINVAL);
 261
 262         if (hctx_idx >= q->nr_hw_queues)
 263                 return ERR_PTR(-EIO);
 264
 265         ret = blk_queue_enter(q, true);
 266         if (ret)
 267                 return ERR_PTR(ret);
 268
 269         /*
 270          * Check if the hardware context is actually mapped to anything.
 271          * If not tell the caller that it should skip this queue.
 272          */
 273         hctx = q->queue_hw_ctx[hctx_idx];
 274         if (!blk_mq_hw_queue_mapped(hctx)) {
 275                 ret = -EXDEV;
 276                 goto out_queue_exit;
 277         }
 278         ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
 279
 280         blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
 281         rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
 282         if (!rq) {
 283                 ret = -EWOULDBLOCK;
 284                 goto out_queue_exit;
 285         }
 286
 287         return rq;
 288
 289 out_queue_exit:
 290         blk_queue_exit(q);
 291         return ERR_PTR(ret);
 292 }
 293 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 294
 295 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 296                                   struct blk_mq_ctx *ctx, struct request *rq)
 297 {
 298         const int tag = rq->tag;
 299         struct request_queue *q = rq->q;
 300
 301         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 302                 atomic_dec(&hctx->nr_active);
 303         rq->cmd_flags = 0;
 304
 305         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 306         blk_mq_put_tag(hctx, ctx, tag);
 307         blk_queue_exit(q);
 308 }
 309
 310 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 311 {
 312         struct blk_mq_ctx *ctx = rq->mq_ctx;
 313
 314         ctx->rq_completed[rq_is_sync(rq)]++;
 315         __blk_mq_free_request(hctx, ctx, rq);
 316
 317 }
 318 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 319
 320 void blk_mq_free_request(struct request *rq)
 321 {
 322         blk_mq_free_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
 323 }
 324 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 325
 326 inline void __blk_mq_end_request(struct request *rq, int error)
 327 {
 328         blk_account_io_done(rq);
 329
 330         if (rq->end_io) {
 331                 rq->end_io(rq, error);
 332         } else {
 333                 if (unlikely(blk_bidi_rq(rq)))
 334                         blk_mq_free_request(rq->next_rq);
 335                 blk_mq_free_request(rq);
 336         }
 337 }
 338 EXPORT_SYMBOL(__blk_mq_end_request);
 339
 340 void blk_mq_end_request(struct request *rq, int error)
 341 {
 342         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 343                 BUG();
 344         __blk_mq_end_request(rq, error);
 345 }
 346 EXPORT_SYMBOL(blk_mq_end_request);
 347
 348 static void __blk_mq_complete_request_remote(void *data)
 349 {
 350         struct request *rq = data;
 351
 352         rq->q->softirq_done_fn(rq);
 353 }
 354
 355 static void blk_mq_ipi_complete_request(struct request *rq)
 356 {
 357         struct blk_mq_ctx *ctx = rq->mq_ctx;
 358         bool shared = false;
 359         int cpu;
 360
 361         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 362                 rq->q->softirq_done_fn(rq);
 363                 return;
 364         }
 365
 366         cpu = get_cpu();
 367         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 368                 shared = cpus_share_cache(cpu, ctx->cpu);
 369
 370         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 371                 rq->csd.func = __blk_mq_complete_request_remote;
 372                 rq->csd.info = rq;
 373                 rq->csd.flags = 0;
 374                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 375         } else {
 376                 rq->q->softirq_done_fn(rq);
 377         }
 378         put_cpu();
 379 }
 380
 381 static void __blk_mq_complete_request(struct request *rq)
 382 {
 383         struct request_queue *q = rq->q;
 384
 385         if (!q->softirq_done_fn)
 386                 blk_mq_end_request(rq, rq->errors);
 387         else
 388                 blk_mq_ipi_complete_request(rq);
 389 }
 390
 391 /**
 392  * blk_mq_complete_request - end I/O on a request
 393  * @rq:         the request being processed
 394  *
 395  * Description:
 396  *      Ends all I/O on a request. It does not handle partial completions.
 397  *      The actual completion happens out-of-order, through a IPI handler.
 398  **/
 399 void blk_mq_complete_request(struct request *rq, int error)
 400 {
 401         struct request_queue *q = rq->q;
 402
 403         if (unlikely(blk_should_fake_timeout(q)))
 404                 return;
 405         if (!blk_mark_rq_complete(rq)) {
 406                 rq->errors = error;
 407                 __blk_mq_complete_request(rq);
 408         }
 409 }
 410 EXPORT_SYMBOL(blk_mq_complete_request);
 411
 412 int blk_mq_request_started(struct request *rq)
 413 {
 414         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 415 }
 416 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 417
 418 void blk_mq_start_request(struct request *rq)
 419 {
 420         struct request_queue *q = rq->q;
 421
 422         trace_block_rq_issue(q, rq);
 423
 424         rq->resid_len = blk_rq_bytes(rq);
 425         if (unlikely(blk_bidi_rq(rq)))
 426                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 427
 428         blk_add_timer(rq);
 429
 430         /*
 431          * Ensure that ->deadline is visible before set the started
 432          * flag and clear the completed flag.
 433          */
 434         smp_mb__before_atomic();
 435
 436         /*
 437          * Mark us as started and clear complete. Complete might have been
 438          * set if requeue raced with timeout, which then marked it as
 439          * complete. So be sure to clear complete again when we start
 440          * the request, otherwise we'll ignore the completion event.
 441          */
 442         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 443                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 444         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 445                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 446
 447         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 448                 /*
 449                  * Make sure space for the drain appears.  We know we can do
 450                  * this because max_hw_segments has been adjusted to be one
 451                  * fewer than the device can handle.
 452                  */
 453                 rq->nr_phys_segments++;
 454         }
 455 }
 456 EXPORT_SYMBOL(blk_mq_start_request);
 457
 458 static void __blk_mq_requeue_request(struct request *rq)
 459 {
 460         struct request_queue *q = rq->q;
 461
 462         trace_block_rq_requeue(q, rq);
 463
 464         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 465                 if (q->dma_drain_size && blk_rq_bytes(rq))
 466                         rq->nr_phys_segments--;
 467         }
 468 }
 469
 470 void blk_mq_requeue_request(struct request *rq)
 471 {
 472         __blk_mq_requeue_request(rq);
 473
 474         BUG_ON(blk_queued_rq(rq));
 475         blk_mq_add_to_requeue_list(rq, true);
 476 }
 477 EXPORT_SYMBOL(blk_mq_requeue_request);
 478
 479 static void blk_mq_requeue_work(struct work_struct *work)
 480 {
 481         struct request_queue *q =
 482                 container_of(work, struct request_queue, requeue_work.work);
 483         LIST_HEAD(rq_list);
 484         struct request *rq, *next;
 485         unsigned long flags;
 486
 487         spin_lock_irqsave(&q->requeue_lock, flags);
 488         list_splice_init(&q->requeue_list, &rq_list);
 489         spin_unlock_irqrestore(&q->requeue_lock, flags);
 490
 491         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 492                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 493                         continue;
 494
 495                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 496                 list_del_init(&rq->queuelist);
 497                 blk_mq_insert_request(rq, true, false, false);
 498         }
 499
 500         while (!list_empty(&rq_list)) {
 501                 rq = list_entry(rq_list.next, struct request, queuelist);
 502                 list_del_init(&rq->queuelist);
 503                 blk_mq_insert_request(rq, false, false, false);
 504         }
 505
 506         /*
 507          * Use the start variant of queue running here, so that running
 508          * the requeue work will kick stopped queues.
 509          */
 510         blk_mq_start_hw_queues(q);
 511 }
 512
 513 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 514 {
 515         struct request_queue *q = rq->q;
 516         unsigned long flags;
 517
 518         /*
 519          * We abuse this flag that is otherwise used by the I/O scheduler to
 520          * request head insertation from the workqueue.
 521          */
 522         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 523
 524         spin_lock_irqsave(&q->requeue_lock, flags);
 525         if (at_head) {
 526                 rq->cmd_flags |= REQ_SOFTBARRIER;
 527                 list_add(&rq->queuelist, &q->requeue_list);
 528         } else {
 529                 list_add_tail(&rq->queuelist, &q->requeue_list);
 530         }
 531         spin_unlock_irqrestore(&q->requeue_lock, flags);
 532 }
 533 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 534
 535 void blk_mq_cancel_requeue_work(struct request_queue *q)
 536 {
 537         cancel_delayed_work_sync(&q->requeue_work);
 538 }
 539 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 540
 541 void blk_mq_kick_requeue_list(struct request_queue *q)
 542 {
 543         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 544 }
 545 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 546
 547 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 548                                     unsigned long msecs)
 549 {
 550         kblockd_schedule_delayed_work(&q->requeue_work,
 551                                       msecs_to_jiffies(msecs));
 552 }
 553 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 554
 555 void blk_mq_abort_requeue_list(struct request_queue *q)
 556 {
 557         unsigned long flags;
 558         LIST_HEAD(rq_list);
 559
 560         spin_lock_irqsave(&q->requeue_lock, flags);
 561         list_splice_init(&q->requeue_list, &rq_list);
 562         spin_unlock_irqrestore(&q->requeue_lock, flags);
 563
 564         while (!list_empty(&rq_list)) {
 565                 struct request *rq;
 566
 567                 rq = list_first_entry(&rq_list, struct request, queuelist);
 568                 list_del_init(&rq->queuelist);
 569                 rq->errors = -EIO;
 570                 blk_mq_end_request(rq, rq->errors);
 571         }
 572 }
 573 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 574
 575 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 576 {
 577         if (tag < tags->nr_tags) {
 578                 prefetch(tags->rqs[tag]);
 579                 return tags->rqs[tag];
 580         }
 581
 582         return NULL;
 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_end_request(rq, rq->errors);
 640                 }
 641                 return;
 642         }
 643
 644         if (time_after_eq(jiffies, rq->deadline)) {
 645                 if (!blk_mark_rq_complete(rq))
 646                         blk_mq_rq_timed_out(rq, reserved);
 647         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 648                 data->next = rq->deadline;
 649                 data->next_set = 1;
 650         }
 651 }
 652
 653 static void blk_mq_timeout_work(struct work_struct *work)
 654 {
 655         struct request_queue *q =
 656                 container_of(work, struct request_queue, timeout_work);
 657         struct blk_mq_timeout_data data = {
 658                 .next           = 0,
 659                 .next_set       = 0,
 660         };
 661         int i;
 662
 663         /* A deadlock might occur if a request is stuck requiring a
 664          * timeout at the same time a queue freeze is waiting
 665          * completion, since the timeout code would not be able to
 666          * acquire the queue reference here.
 667          *
 668          * That's why we don't use blk_queue_enter here; instead, we use
 669          * percpu_ref_tryget directly, because we need to be able to
 670          * obtain a reference even in the short window between the queue
 671          * starting to freeze, by dropping the first reference in
 672          * blk_mq_freeze_queue_start, and the moment the last request is
 673          * consumed, marked by the instant q_usage_counter reaches
 674          * zero.
 675          */
 676         if (!percpu_ref_tryget(&q->q_usage_counter))
 677                 return;
 678
 679         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 680
 681         if (data.next_set) {
 682                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 683                 mod_timer(&q->timeout, data.next);
 684         } else {
 685                 struct blk_mq_hw_ctx *hctx;
 686
 687                 queue_for_each_hw_ctx(q, hctx, i) {
 688                         /* the hctx may be unmapped, so check it here */
 689                         if (blk_mq_hw_queue_mapped(hctx))
 690                                 blk_mq_tag_idle(hctx);
 691                 }
 692         }
 693         blk_queue_exit(q);
 694 }
 695
 696 /*
 697  * Reverse check our software queue for entries that we could potentially
 698  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 699  * too much time checking for merges.
 700  */
 701 static bool blk_mq_attempt_merge(struct request_queue *q,
 702                                  struct blk_mq_ctx *ctx, struct bio *bio)
 703 {
 704         struct request *rq;
 705         int checked = 8;
 706
 707         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 708                 int el_ret;
 709
 710                 if (!checked--)
 711                         break;
 712
 713                 if (!blk_rq_merge_ok(rq, bio))
 714                         continue;
 715
 716                 el_ret = blk_try_merge(rq, bio);
 717                 if (el_ret == ELEVATOR_BACK_MERGE) {
 718                         if (bio_attempt_back_merge(q, rq, bio)) {
 719                                 ctx->rq_merged++;
 720                                 return true;
 721                         }
 722                         break;
 723                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 724                         if (bio_attempt_front_merge(q, rq, bio)) {
 725                                 ctx->rq_merged++;
 726                                 return true;
 727                         }
 728                         break;
 729                 }
 730         }
 731
 732         return false;
 733 }
 734
 735 struct flush_busy_ctx_data {
 736         struct blk_mq_hw_ctx *hctx;
 737         struct list_head *list;
 738 };
 739
 740 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 741 {
 742         struct flush_busy_ctx_data *flush_data = data;
 743         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 744         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 745
 746         sbitmap_clear_bit(sb, bitnr);
 747         spin_lock(&ctx->lock);
 748         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 749         spin_unlock(&ctx->lock);
 750         return true;
 751 }
 752
 753 /*
 754  * Process software queues that have been marked busy, splicing them
 755  * to the for-dispatch
 756  */
 757 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 758 {
 759         struct flush_busy_ctx_data data = {
 760                 .hctx = hctx,
 761                 .list = list,
 762         };
 763
 764         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 765 }
 766
 767 static inline unsigned int queued_to_index(unsigned int queued)
 768 {
 769         if (!queued)
 770                 return 0;
 771
 772         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 773 }
 774
 775 /*
 776  * Run this hardware queue, pulling any software queues mapped to it in.
 777  * Note that this function currently has various problems around ordering
 778  * of IO. In particular, we'd like FIFO behaviour on handling existing
 779  * items on the hctx->dispatch list. Ignore that for now.
 780  */
 781 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 782 {
 783         struct request_queue *q = hctx->queue;
 784         struct request *rq;
 785         LIST_HEAD(rq_list);
 786         LIST_HEAD(driver_list);
 787         struct list_head *dptr;
 788         int queued;
 789
 790         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 791                 return;
 792
 793         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
 794                 cpu_online(hctx->next_cpu));
 795
 796         hctx->run++;
 797
 798         /*
 799          * Touch any software queue that has pending entries.
 800          */
 801         flush_busy_ctxs(hctx, &rq_list);
 802
 803         /*
 804          * If we have previous entries on our dispatch list, grab them
 805          * and stuff them at the front for more fair dispatch.
 806          */
 807         if (!list_empty_careful(&hctx->dispatch)) {
 808                 spin_lock(&hctx->lock);
 809                 if (!list_empty(&hctx->dispatch))
 810                         list_splice_init(&hctx->dispatch, &rq_list);
 811                 spin_unlock(&hctx->lock);
 812         }
 813
 814         /*
 815          * Start off with dptr being NULL, so we start the first request
 816          * immediately, even if we have more pending.
 817          */
 818         dptr = NULL;
 819
 820         /*
 821          * Now process all the entries, sending them to the driver.
 822          */
 823         queued = 0;
 824         while (!list_empty(&rq_list)) {
 825                 struct blk_mq_queue_data bd;
 826                 int ret;
 827
 828                 rq = list_first_entry(&rq_list, struct request, queuelist);
 829                 list_del_init(&rq->queuelist);
 830
 831                 bd.rq = rq;
 832                 bd.list = dptr;
 833                 bd.last = list_empty(&rq_list);
 834
 835                 ret = q->mq_ops->queue_rq(hctx, &bd);
 836                 switch (ret) {
 837                 case BLK_MQ_RQ_QUEUE_OK:
 838                         queued++;
 839                         break;
 840                 case BLK_MQ_RQ_QUEUE_BUSY:
 841                         list_add(&rq->queuelist, &rq_list);
 842                         __blk_mq_requeue_request(rq);
 843                         break;
 844                 default:
 845                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 846                 case BLK_MQ_RQ_QUEUE_ERROR:
 847                         rq->errors = -EIO;
 848                         blk_mq_end_request(rq, rq->errors);
 849                         break;
 850                 }
 851
 852                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 853                         break;
 854
 855                 /*
 856                  * We've done the first request. If we have more than 1
 857                  * left in the list, set dptr to defer issue.
 858                  */
 859                 if (!dptr && rq_list.next != rq_list.prev)
 860                         dptr = &driver_list;
 861         }
 862
 863         hctx->dispatched[queued_to_index(queued)]++;
 864
 865         /*
 866          * Any items that need requeuing? Stuff them into hctx->dispatch,
 867          * that is where we will continue on next queue run.
 868          */
 869         if (!list_empty(&rq_list)) {
 870                 spin_lock(&hctx->lock);
 871                 list_splice(&rq_list, &hctx->dispatch);
 872                 spin_unlock(&hctx->lock);
 873                 /*
 874                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 875                  * it's possible the queue is stopped and restarted again
 876                  * before this. Queue restart will dispatch requests. And since
 877                  * requests in rq_list aren't added into hctx->dispatch yet,
 878                  * the requests in rq_list might get lost.
 879                  *
 880                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 881                  **/
 882                 blk_mq_run_hw_queue(hctx, true);
 883         }
 884 }
 885
 886 /*
 887  * It'd be great if the workqueue API had a way to pass
 888  * in a mask and had some smarts for more clever placement.
 889  * For now we just round-robin here, switching for every
 890  * BLK_MQ_CPU_WORK_BATCH queued items.
 891  */
 892 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 893 {
 894         if (hctx->queue->nr_hw_queues == 1)
 895                 return WORK_CPU_UNBOUND;
 896
 897         if (--hctx->next_cpu_batch <= 0) {
 898                 int cpu = hctx->next_cpu, next_cpu;
 899
 900                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 901                 if (next_cpu >= nr_cpu_ids)
 902                         next_cpu = cpumask_first(hctx->cpumask);
 903
 904                 hctx->next_cpu = next_cpu;
 905                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 906
 907                 return cpu;
 908         }
 909
 910         return hctx->next_cpu;
 911 }
 912
 913 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 914 {
 915         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 916             !blk_mq_hw_queue_mapped(hctx)))
 917                 return;
 918
 919         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
 920                 int cpu = get_cpu();
 921                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 922                         __blk_mq_run_hw_queue(hctx);
 923                         put_cpu();
 924                         return;
 925                 }
 926
 927                 put_cpu();
 928         }
 929
 930         kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
 931 }
 932
 933 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 934 {
 935         struct blk_mq_hw_ctx *hctx;
 936         int i;
 937
 938         queue_for_each_hw_ctx(q, hctx, i) {
 939                 if ((!blk_mq_hctx_has_pending(hctx) &&
 940                     list_empty_careful(&hctx->dispatch)) ||
 941                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 942                         continue;
 943
 944                 blk_mq_run_hw_queue(hctx, async);
 945         }
 946 }
 947 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 948
 949 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 950 {
 951         cancel_work(&hctx->run_work);
 952         cancel_delayed_work(&hctx->delay_work);
 953         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 954 }
 955 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 956
 957 void blk_mq_stop_hw_queues(struct request_queue *q)
 958 {
 959         struct blk_mq_hw_ctx *hctx;
 960         int i;
 961
 962         queue_for_each_hw_ctx(q, hctx, i)
 963                 blk_mq_stop_hw_queue(hctx);
 964 }
 965 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 966
 967 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 968 {
 969         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 970
 971         blk_mq_run_hw_queue(hctx, false);
 972 }
 973 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 974
 975 void blk_mq_start_hw_queues(struct request_queue *q)
 976 {
 977         struct blk_mq_hw_ctx *hctx;
 978         int i;
 979
 980         queue_for_each_hw_ctx(q, hctx, i)
 981                 blk_mq_start_hw_queue(hctx);
 982 }
 983 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 984
 985 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 986 {
 987         struct blk_mq_hw_ctx *hctx;
 988         int i;
 989
 990         queue_for_each_hw_ctx(q, hctx, i) {
 991                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 992                         continue;
 993
 994                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 995                 blk_mq_run_hw_queue(hctx, async);
 996         }
 997 }
 998 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 999
1000 static void blk_mq_run_work_fn(struct work_struct *work)
1001 {
1002         struct blk_mq_hw_ctx *hctx;
1003
1004         hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
1005
1006         __blk_mq_run_hw_queue(hctx);
1007 }
1008
1009 static void blk_mq_delay_work_fn(struct work_struct *work)
1010 {
1011         struct blk_mq_hw_ctx *hctx;
1012
1013         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1014
1015         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1016                 __blk_mq_run_hw_queue(hctx);
1017 }
1018
1019 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1020 {
1021         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1022                 return;
1023
1024         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1025                         &hctx->delay_work, msecs_to_jiffies(msecs));
1026 }
1027 EXPORT_SYMBOL(blk_mq_delay_queue);
1028
1029 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1030                                             struct request *rq,
1031                                             bool at_head)
1032 {
1033         struct blk_mq_ctx *ctx = rq->mq_ctx;
1034
1035         trace_block_rq_insert(hctx->queue, rq);
1036
1037         if (at_head)
1038                 list_add(&rq->queuelist, &ctx->rq_list);
1039         else
1040                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1041 }
1042
1043 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1044                                     struct request *rq, bool at_head)
1045 {
1046         struct blk_mq_ctx *ctx = rq->mq_ctx;
1047
1048         __blk_mq_insert_req_list(hctx, rq, at_head);
1049         blk_mq_hctx_mark_pending(hctx, ctx);
1050 }
1051
1052 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1053                            bool async)
1054 {
1055         struct blk_mq_ctx *ctx = rq->mq_ctx;
1056         struct request_queue *q = rq->q;
1057         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1058
1059         spin_lock(&ctx->lock);
1060         __blk_mq_insert_request(hctx, rq, at_head);
1061         spin_unlock(&ctx->lock);
1062
1063         if (run_queue)
1064                 blk_mq_run_hw_queue(hctx, async);
1065 }
1066
1067 static void blk_mq_insert_requests(struct request_queue *q,
1068                                      struct blk_mq_ctx *ctx,
1069                                      struct list_head *list,
1070                                      int depth,
1071                                      bool from_schedule)
1072
1073 {
1074         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
1075
1076         trace_block_unplug(q, depth, !from_schedule);
1077
1078         /*
1079          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1080          * offline now
1081          */
1082         spin_lock(&ctx->lock);
1083         while (!list_empty(list)) {
1084                 struct request *rq;
1085
1086                 rq = list_first_entry(list, struct request, queuelist);
1087                 BUG_ON(rq->mq_ctx != ctx);
1088                 list_del_init(&rq->queuelist);
1089                 __blk_mq_insert_req_list(hctx, rq, false);
1090         }
1091         blk_mq_hctx_mark_pending(hctx, ctx);
1092         spin_unlock(&ctx->lock);
1093
1094         blk_mq_run_hw_queue(hctx, from_schedule);
1095 }
1096
1097 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1098 {
1099         struct request *rqa = container_of(a, struct request, queuelist);
1100         struct request *rqb = container_of(b, struct request, queuelist);
1101
1102         return !(rqa->mq_ctx < rqb->mq_ctx ||
1103                  (rqa->mq_ctx == rqb->mq_ctx &&
1104                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1105 }
1106
1107 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1108 {
1109         struct blk_mq_ctx *this_ctx;
1110         struct request_queue *this_q;
1111         struct request *rq;
1112         LIST_HEAD(list);
1113         LIST_HEAD(ctx_list);
1114         unsigned int depth;
1115
1116         list_splice_init(&plug->mq_list, &list);
1117
1118         list_sort(NULL, &list, plug_ctx_cmp);
1119
1120         this_q = NULL;
1121         this_ctx = NULL;
1122         depth = 0;
1123
1124         while (!list_empty(&list)) {
1125                 rq = list_entry_rq(list.next);
1126                 list_del_init(&rq->queuelist);
1127                 BUG_ON(!rq->q);
1128                 if (rq->mq_ctx != this_ctx) {
1129                         if (this_ctx) {
1130                                 blk_mq_insert_requests(this_q, this_ctx,
1131                                                         &ctx_list, depth,
1132                                                         from_schedule);
1133                         }
1134
1135                         this_ctx = rq->mq_ctx;
1136                         this_q = rq->q;
1137                         depth = 0;
1138                 }
1139
1140                 depth++;
1141                 list_add_tail(&rq->queuelist, &ctx_list);
1142         }
1143
1144         /*
1145          * If 'this_ctx' is set, we know we have entries to complete
1146          * on 'ctx_list'. Do those.
1147          */
1148         if (this_ctx) {
1149                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1150                                        from_schedule);
1151         }
1152 }
1153
1154 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1155 {
1156         init_request_from_bio(rq, bio);
1157
1158         blk_account_io_start(rq, 1);
1159 }
1160
1161 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1162 {
1163         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1164                 !blk_queue_nomerges(hctx->queue);
1165 }
1166
1167 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1168                                          struct blk_mq_ctx *ctx,
1169                                          struct request *rq, struct bio *bio)
1170 {
1171         if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1172                 blk_mq_bio_to_request(rq, bio);
1173                 spin_lock(&ctx->lock);
1174 insert_rq:
1175                 __blk_mq_insert_request(hctx, rq, false);
1176                 spin_unlock(&ctx->lock);
1177                 return false;
1178         } else {
1179                 struct request_queue *q = hctx->queue;
1180
1181                 spin_lock(&ctx->lock);
1182                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1183                         blk_mq_bio_to_request(rq, bio);
1184                         goto insert_rq;
1185                 }
1186
1187                 spin_unlock(&ctx->lock);
1188                 __blk_mq_free_request(hctx, ctx, rq);
1189                 return true;
1190         }
1191 }
1192
1193 struct blk_map_ctx {
1194         struct blk_mq_hw_ctx *hctx;
1195         struct blk_mq_ctx *ctx;
1196 };
1197
1198 static struct request *blk_mq_map_request(struct request_queue *q,
1199                                           struct bio *bio,
1200                                           struct blk_map_ctx *data)
1201 {
1202         struct blk_mq_hw_ctx *hctx;
1203         struct blk_mq_ctx *ctx;
1204         struct request *rq;
1205         int op = bio_data_dir(bio);
1206         int op_flags = 0;
1207         struct blk_mq_alloc_data alloc_data;
1208
1209         blk_queue_enter_live(q);
1210         ctx = blk_mq_get_ctx(q);
1211         hctx = blk_mq_map_queue(q, ctx->cpu);
1212
1213         if (rw_is_sync(bio_op(bio), bio->bi_opf))
1214                 op_flags |= REQ_SYNC;
1215
1216         trace_block_getrq(q, bio, op);
1217         blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1218         rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1219
1220         hctx->queued++;
1221         data->hctx = hctx;
1222         data->ctx = ctx;
1223         return rq;
1224 }
1225
1226 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1227 {
1228         int ret;
1229         struct request_queue *q = rq->q;
1230         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
1231         struct blk_mq_queue_data bd = {
1232                 .rq = rq,
1233                 .list = NULL,
1234                 .last = 1
1235         };
1236         blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1237
1238         /*
1239          * For OK queue, we are done. For error, kill it. Any other
1240          * error (busy), just add it to our list as we previously
1241          * would have done
1242          */
1243         ret = q->mq_ops->queue_rq(hctx, &bd);
1244         if (ret == BLK_MQ_RQ_QUEUE_OK) {
1245                 *cookie = new_cookie;
1246                 return 0;
1247         }
1248
1249         __blk_mq_requeue_request(rq);
1250
1251         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1252                 *cookie = BLK_QC_T_NONE;
1253                 rq->errors = -EIO;
1254                 blk_mq_end_request(rq, rq->errors);
1255                 return 0;
1256         }
1257
1258         return -1;
1259 }
1260
1261 /*
1262  * Multiple hardware queue variant. This will not use per-process plugs,
1263  * but will attempt to bypass the hctx queueing if we can go straight to
1264  * hardware for SYNC IO.
1265  */
1266 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1267 {
1268         const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1269         const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1270         struct blk_map_ctx data;
1271         struct request *rq;
1272         unsigned int request_count = 0;
1273         struct blk_plug *plug;
1274         struct request *same_queue_rq = NULL;
1275         blk_qc_t cookie;
1276
1277         blk_queue_bounce(q, &bio);
1278
1279         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1280                 bio_io_error(bio);
1281                 return BLK_QC_T_NONE;
1282         }
1283
1284         blk_queue_split(q, &bio, q->bio_split);
1285
1286         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1287             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1288                 return BLK_QC_T_NONE;
1289
1290         rq = blk_mq_map_request(q, bio, &data);
1291         if (unlikely(!rq))
1292                 return BLK_QC_T_NONE;
1293
1294         cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1295
1296         if (unlikely(is_flush_fua)) {
1297                 blk_mq_bio_to_request(rq, bio);
1298                 blk_insert_flush(rq);
1299                 goto run_queue;
1300         }
1301
1302         plug = current->plug;
1303         /*
1304          * If the driver supports defer issued based on 'last', then
1305          * queue it up like normal since we can potentially save some
1306          * CPU this way.
1307          */
1308         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1309             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1310                 struct request *old_rq = NULL;
1311
1312                 blk_mq_bio_to_request(rq, bio);
1313
1314                 /*
1315                  * We do limited pluging. If the bio can be merged, do that.
1316                  * Otherwise the existing request in the plug list will be
1317                  * issued. So the plug list will have one request at most
1318                  */
1319                 if (plug) {
1320                         /*
1321                          * The plug list might get flushed before this. If that
1322                          * happens, same_queue_rq is invalid and plug list is
1323                          * empty
1324                          */
1325                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1326                                 old_rq = same_queue_rq;
1327                                 list_del_init(&old_rq->queuelist);
1328                         }
1329                         list_add_tail(&rq->queuelist, &plug->mq_list);
1330                 } else /* is_sync */
1331                         old_rq = rq;
1332                 blk_mq_put_ctx(data.ctx);
1333                 if (!old_rq)
1334                         goto done;
1335                 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1336                         goto done;
1337                 blk_mq_insert_request(old_rq, false, true, true);
1338                 goto done;
1339         }
1340
1341         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1342                 /*
1343                  * For a SYNC request, send it to the hardware immediately. For
1344                  * an ASYNC request, just ensure that we run it later on. The
1345                  * latter allows for merging opportunities and more efficient
1346                  * dispatching.
1347                  */
1348 run_queue:
1349                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1350         }
1351         blk_mq_put_ctx(data.ctx);
1352 done:
1353         return cookie;
1354 }
1355
1356 /*
1357  * Single hardware queue variant. This will attempt to use any per-process
1358  * plug for merging and IO deferral.
1359  */
1360 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1361 {
1362         const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1363         const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1364         struct blk_plug *plug;
1365         unsigned int request_count = 0;
1366         struct blk_map_ctx data;
1367         struct request *rq;
1368         blk_qc_t cookie;
1369
1370         blk_queue_bounce(q, &bio);
1371
1372         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1373                 bio_io_error(bio);
1374                 return BLK_QC_T_NONE;
1375         }
1376
1377         blk_queue_split(q, &bio, q->bio_split);
1378
1379         if (!is_flush_fua && !blk_queue_nomerges(q)) {
1380                 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1381                         return BLK_QC_T_NONE;
1382         } else
1383                 request_count = blk_plug_queued_count(q);
1384
1385         rq = blk_mq_map_request(q, bio, &data);
1386         if (unlikely(!rq))
1387                 return BLK_QC_T_NONE;
1388
1389         cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1390
1391         if (unlikely(is_flush_fua)) {
1392                 blk_mq_bio_to_request(rq, bio);
1393                 blk_insert_flush(rq);
1394                 goto run_queue;
1395         }
1396
1397         /*
1398          * A task plug currently exists. Since this is completely lockless,
1399          * utilize that to temporarily store requests until the task is
1400          * either done or scheduled away.
1401          */
1402         plug = current->plug;
1403         if (plug) {
1404                 blk_mq_bio_to_request(rq, bio);
1405                 if (!request_count)
1406                         trace_block_plug(q);
1407
1408                 blk_mq_put_ctx(data.ctx);
1409
1410                 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1411                         blk_flush_plug_list(plug, false);
1412                         trace_block_plug(q);
1413                 }
1414
1415                 list_add_tail(&rq->queuelist, &plug->mq_list);
1416                 return cookie;
1417         }
1418
1419         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1420                 /*
1421                  * For a SYNC request, send it to the hardware immediately. For
1422                  * an ASYNC request, just ensure that we run it later on. The
1423                  * latter allows for merging opportunities and more efficient
1424                  * dispatching.
1425                  */
1426 run_queue:
1427                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1428         }
1429
1430         blk_mq_put_ctx(data.ctx);
1431         return cookie;
1432 }
1433
1434 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1435                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1436 {
1437         struct page *page;
1438
1439         if (tags->rqs && set->ops->exit_request) {
1440                 int i;
1441
1442                 for (i = 0; i < tags->nr_tags; i++) {
1443                         if (!tags->rqs[i])
1444                                 continue;
1445                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1446                                                 hctx_idx, i);
1447                         tags->rqs[i] = NULL;
1448                 }
1449         }
1450
1451         while (!list_empty(&tags->page_list)) {
1452                 page = list_first_entry(&tags->page_list, struct page, lru);
1453                 list_del_init(&page->lru);
1454                 /*
1455                  * Remove kmemleak object previously allocated in
1456                  * blk_mq_init_rq_map().
1457                  */
1458                 kmemleak_free(page_address(page));
1459                 __free_pages(page, page->private);
1460         }
1461
1462         kfree(tags->rqs);
1463
1464         blk_mq_free_tags(tags);
1465 }
1466
1467 static size_t order_to_size(unsigned int order)
1468 {
1469         return (size_t)PAGE_SIZE << order;
1470 }
1471
1472 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1473                 unsigned int hctx_idx)
1474 {
1475         struct blk_mq_tags *tags;
1476         unsigned int i, j, entries_per_page, max_order = 4;
1477         size_t rq_size, left;
1478
1479         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1480                                 set->numa_node,
1481                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1482         if (!tags)
1483                 return NULL;
1484
1485         INIT_LIST_HEAD(&tags->page_list);
1486
1487         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1488                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1489                                  set->numa_node);
1490         if (!tags->rqs) {
1491                 blk_mq_free_tags(tags);
1492                 return NULL;
1493         }
1494
1495         /*
1496          * rq_size is the size of the request plus driver payload, rounded
1497          * to the cacheline size
1498          */
1499         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1500                                 cache_line_size());
1501         left = rq_size * set->queue_depth;
1502
1503         for (i = 0; i < set->queue_depth; ) {
1504                 int this_order = max_order;
1505                 struct page *page;
1506                 int to_do;
1507                 void *p;
1508
1509                 while (this_order && left < order_to_size(this_order - 1))
1510                         this_order--;
1511
1512                 do {
1513                         page = alloc_pages_node(set->numa_node,
1514                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1515                                 this_order);
1516                         if (page)
1517                                 break;
1518                         if (!this_order--)
1519                                 break;
1520                         if (order_to_size(this_order) < rq_size)
1521                                 break;
1522                 } while (1);
1523
1524                 if (!page)
1525                         goto fail;
1526
1527                 page->private = this_order;
1528                 list_add_tail(&page->lru, &tags->page_list);
1529
1530                 p = page_address(page);
1531                 /*
1532                  * Allow kmemleak to scan these pages as they contain pointers
1533                  * to additional allocations like via ops->init_request().
1534                  */
1535                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1536                 entries_per_page = order_to_size(this_order) / rq_size;
1537                 to_do = min(entries_per_page, set->queue_depth - i);
1538                 left -= to_do * rq_size;
1539                 for (j = 0; j < to_do; j++) {
1540                         tags->rqs[i] = p;
1541                         if (set->ops->init_request) {
1542                                 if (set->ops->init_request(set->driver_data,
1543                                                 tags->rqs[i], hctx_idx, i,
1544                                                 set->numa_node)) {
1545                                         tags->rqs[i] = NULL;
1546                                         goto fail;
1547                                 }
1548                         }
1549
1550                         p += rq_size;
1551                         i++;
1552                 }
1553         }
1554         return tags;
1555
1556 fail:
1557         blk_mq_free_rq_map(set, tags, hctx_idx);
1558         return NULL;
1559 }
1560
1561 /*
1562  * 'cpu' is going away. splice any existing rq_list entries from this
1563  * software queue to the hw queue dispatch list, and ensure that it
1564  * gets run.
1565  */
1566 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1567 {
1568         struct blk_mq_ctx *ctx;
1569         LIST_HEAD(tmp);
1570
1571         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1572
1573         spin_lock(&ctx->lock);
1574         if (!list_empty(&ctx->rq_list)) {
1575                 list_splice_init(&ctx->rq_list, &tmp);
1576                 blk_mq_hctx_clear_pending(hctx, ctx);
1577         }
1578         spin_unlock(&ctx->lock);
1579
1580         if (list_empty(&tmp))
1581                 return NOTIFY_OK;
1582
1583         spin_lock(&hctx->lock);
1584         list_splice_tail_init(&tmp, &hctx->dispatch);
1585         spin_unlock(&hctx->lock);
1586
1587         blk_mq_run_hw_queue(hctx, true);
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
1599         /*
1600          * In case of CPU online, tags may be reallocated
1601          * in blk_mq_map_swqueue() after mapping is updated.
1602          */
1603
1604         return NOTIFY_OK;
1605 }
1606
1607 /* hctx->ctxs will be freed in queue's release handler */
1608 static void blk_mq_exit_hctx(struct request_queue *q,
1609                 struct blk_mq_tag_set *set,
1610                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1611 {
1612         unsigned flush_start_tag = set->queue_depth;
1613
1614         blk_mq_tag_idle(hctx);
1615
1616         if (set->ops->exit_request)
1617                 set->ops->exit_request(set->driver_data,
1618                                        hctx->fq->flush_rq, hctx_idx,
1619                                        flush_start_tag + hctx_idx);
1620
1621         if (set->ops->exit_hctx)
1622                 set->ops->exit_hctx(hctx, hctx_idx);
1623
1624         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1625         blk_free_flush_queue(hctx->fq);
1626         sbitmap_free(&hctx->ctx_map);
1627 }
1628
1629 static void blk_mq_exit_hw_queues(struct request_queue *q,
1630                 struct blk_mq_tag_set *set, int nr_queue)
1631 {
1632         struct blk_mq_hw_ctx *hctx;
1633         unsigned int i;
1634
1635         queue_for_each_hw_ctx(q, hctx, i) {
1636                 if (i == nr_queue)
1637                         break;
1638                 blk_mq_exit_hctx(q, set, hctx, i);
1639         }
1640 }
1641
1642 static void blk_mq_free_hw_queues(struct request_queue *q,
1643                 struct blk_mq_tag_set *set)
1644 {
1645         struct blk_mq_hw_ctx *hctx;
1646         unsigned int i;
1647
1648         queue_for_each_hw_ctx(q, hctx, i)
1649                 free_cpumask_var(hctx->cpumask);
1650 }
1651
1652 static int blk_mq_init_hctx(struct request_queue *q,
1653                 struct blk_mq_tag_set *set,
1654                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1655 {
1656         int node;
1657         unsigned flush_start_tag = set->queue_depth;
1658
1659         node = hctx->numa_node;
1660         if (node == NUMA_NO_NODE)
1661                 node = hctx->numa_node = set->numa_node;
1662
1663         INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1664         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1665         spin_lock_init(&hctx->lock);
1666         INIT_LIST_HEAD(&hctx->dispatch);
1667         hctx->queue = q;
1668         hctx->queue_num = hctx_idx;
1669         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1670
1671         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1672                                         blk_mq_hctx_notify, hctx);
1673         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1674
1675         hctx->tags = set->tags[hctx_idx];
1676
1677         /*
1678          * Allocate space for all possible cpus to avoid allocation at
1679          * runtime
1680          */
1681         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1682                                         GFP_KERNEL, node);
1683         if (!hctx->ctxs)
1684                 goto unregister_cpu_notifier;
1685
1686         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1687                               node))
1688                 goto free_ctxs;
1689
1690         hctx->nr_ctx = 0;
1691
1692         if (set->ops->init_hctx &&
1693             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1694                 goto free_bitmap;
1695
1696         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1697         if (!hctx->fq)
1698                 goto exit_hctx;
1699
1700         if (set->ops->init_request &&
1701             set->ops->init_request(set->driver_data,
1702                                    hctx->fq->flush_rq, hctx_idx,
1703                                    flush_start_tag + hctx_idx, node))
1704                 goto free_fq;
1705
1706         return 0;
1707
1708  free_fq:
1709         kfree(hctx->fq);
1710  exit_hctx:
1711         if (set->ops->exit_hctx)
1712                 set->ops->exit_hctx(hctx, hctx_idx);
1713  free_bitmap:
1714         sbitmap_free(&hctx->ctx_map);
1715  free_ctxs:
1716         kfree(hctx->ctxs);
1717  unregister_cpu_notifier:
1718         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1719
1720         return -1;
1721 }
1722
1723 static void blk_mq_init_cpu_queues(struct request_queue *q,
1724                                    unsigned int nr_hw_queues)
1725 {
1726         unsigned int i;
1727
1728         for_each_possible_cpu(i) {
1729                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1730                 struct blk_mq_hw_ctx *hctx;
1731
1732                 memset(__ctx, 0, sizeof(*__ctx));
1733                 __ctx->cpu = i;
1734                 spin_lock_init(&__ctx->lock);
1735                 INIT_LIST_HEAD(&__ctx->rq_list);
1736                 __ctx->queue = q;
1737
1738                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1739                 if (!cpu_online(i))
1740                         continue;
1741
1742                 hctx = blk_mq_map_queue(q, i);
1743
1744                 /*
1745                  * Set local node, IFF we have more than one hw queue. If
1746                  * not, we remain on the home node of the device
1747                  */
1748                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1749                         hctx->numa_node = local_memory_node(cpu_to_node(i));
1750         }
1751 }
1752
1753 static void blk_mq_map_swqueue(struct request_queue *q,
1754                                const struct cpumask *online_mask)
1755 {
1756         unsigned int i;
1757         struct blk_mq_hw_ctx *hctx;
1758         struct blk_mq_ctx *ctx;
1759         struct blk_mq_tag_set *set = q->tag_set;
1760
1761         /*
1762          * Avoid others reading imcomplete hctx->cpumask through sysfs
1763          */
1764         mutex_lock(&q->sysfs_lock);
1765
1766         queue_for_each_hw_ctx(q, hctx, i) {
1767                 cpumask_clear(hctx->cpumask);
1768                 hctx->nr_ctx = 0;
1769         }
1770
1771         /*
1772          * Map software to hardware queues
1773          */
1774         for_each_possible_cpu(i) {
1775                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1776                 if (!cpumask_test_cpu(i, online_mask))
1777                         continue;
1778
1779                 ctx = per_cpu_ptr(q->queue_ctx, i);
1780                 hctx = blk_mq_map_queue(q, i);
1781
1782                 cpumask_set_cpu(i, hctx->cpumask);
1783                 ctx->index_hw = hctx->nr_ctx;
1784                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1785         }
1786
1787         mutex_unlock(&q->sysfs_lock);
1788
1789         queue_for_each_hw_ctx(q, hctx, i) {
1790                 /*
1791                  * If no software queues are mapped to this hardware queue,
1792                  * disable it and free the request entries.
1793                  */
1794                 if (!hctx->nr_ctx) {
1795                         if (set->tags[i]) {
1796                                 blk_mq_free_rq_map(set, set->tags[i], i);
1797                                 set->tags[i] = NULL;
1798                         }
1799                         hctx->tags = NULL;
1800                         continue;
1801                 }
1802
1803                 /* unmapped hw queue can be remapped after CPU topo changed */
1804                 if (!set->tags[i])
1805                         set->tags[i] = blk_mq_init_rq_map(set, i);
1806                 hctx->tags = set->tags[i];
1807                 WARN_ON(!hctx->tags);
1808
1809                 /*
1810                  * Set the map size to the number of mapped software queues.
1811                  * This is more accurate and more efficient than looping
1812                  * over all possibly mapped software queues.
1813                  */
1814                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
1815
1816                 /*
1817                  * Initialize batch roundrobin counts
1818                  */
1819                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1820                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1821         }
1822 }
1823
1824 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1825 {
1826         struct blk_mq_hw_ctx *hctx;
1827         int i;
1828
1829         queue_for_each_hw_ctx(q, hctx, i) {
1830                 if (shared)
1831                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
1832                 else
1833                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1834         }
1835 }
1836
1837 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1838 {
1839         struct request_queue *q;
1840
1841         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1842                 blk_mq_freeze_queue(q);
1843                 queue_set_hctx_shared(q, shared);
1844                 blk_mq_unfreeze_queue(q);
1845         }
1846 }
1847
1848 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1849 {
1850         struct blk_mq_tag_set *set = q->tag_set;
1851
1852         mutex_lock(&set->tag_list_lock);
1853         list_del_init(&q->tag_set_list);
1854         if (list_is_singular(&set->tag_list)) {
1855                 /* just transitioned to unshared */
1856                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1857                 /* update existing queue */
1858                 blk_mq_update_tag_set_depth(set, false);
1859         }
1860         mutex_unlock(&set->tag_list_lock);
1861 }
1862
1863 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1864                                      struct request_queue *q)
1865 {
1866         q->tag_set = set;
1867
1868         mutex_lock(&set->tag_list_lock);
1869
1870         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1871         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1872                 set->flags |= BLK_MQ_F_TAG_SHARED;
1873                 /* update existing queue */
1874                 blk_mq_update_tag_set_depth(set, true);
1875         }
1876         if (set->flags & BLK_MQ_F_TAG_SHARED)
1877                 queue_set_hctx_shared(q, true);
1878         list_add_tail(&q->tag_set_list, &set->tag_list);
1879
1880         mutex_unlock(&set->tag_list_lock);
1881 }
1882
1883 /*
1884  * It is the actual release handler for mq, but we do it from
1885  * request queue's release handler for avoiding use-after-free
1886  * and headache because q->mq_kobj shouldn't have been introduced,
1887  * but we can't group ctx/kctx kobj without it.
1888  */
1889 void blk_mq_release(struct request_queue *q)
1890 {
1891         struct blk_mq_hw_ctx *hctx;
1892         unsigned int i;
1893
1894         /* hctx kobj stays in hctx */
1895         queue_for_each_hw_ctx(q, hctx, i) {
1896                 if (!hctx)
1897                         continue;
1898                 kfree(hctx->ctxs);
1899                 kfree(hctx);
1900         }
1901
1902         q->mq_map = NULL;
1903
1904         kfree(q->queue_hw_ctx);
1905
1906         /* ctx kobj stays in queue_ctx */
1907         free_percpu(q->queue_ctx);
1908 }
1909
1910 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1911 {
1912         struct request_queue *uninit_q, *q;
1913
1914         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1915         if (!uninit_q)
1916                 return ERR_PTR(-ENOMEM);
1917
1918         q = blk_mq_init_allocated_queue(set, uninit_q);
1919         if (IS_ERR(q))
1920                 blk_cleanup_queue(uninit_q);
1921
1922         return q;
1923 }
1924 EXPORT_SYMBOL(blk_mq_init_queue);
1925
1926 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1927                                                 struct request_queue *q)
1928 {
1929         int i, j;
1930         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1931
1932         blk_mq_sysfs_unregister(q);
1933         for (i = 0; i < set->nr_hw_queues; i++) {
1934                 int node;
1935
1936                 if (hctxs[i])
1937                         continue;
1938
1939                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1940                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1941                                         GFP_KERNEL, node);
1942                 if (!hctxs[i])
1943                         break;
1944
1945                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1946                                                 node)) {
1947                         kfree(hctxs[i]);
1948                         hctxs[i] = NULL;
1949                         break;
1950                 }
1951
1952                 atomic_set(&hctxs[i]->nr_active, 0);
1953                 hctxs[i]->numa_node = node;
1954                 hctxs[i]->queue_num = i;
1955
1956                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1957                         free_cpumask_var(hctxs[i]->cpumask);
1958                         kfree(hctxs[i]);
1959                         hctxs[i] = NULL;
1960                         break;
1961                 }
1962                 blk_mq_hctx_kobj_init(hctxs[i]);
1963         }
1964         for (j = i; j < q->nr_hw_queues; j++) {
1965                 struct blk_mq_hw_ctx *hctx = hctxs[j];
1966
1967                 if (hctx) {
1968                         if (hctx->tags) {
1969                                 blk_mq_free_rq_map(set, hctx->tags, j);
1970                                 set->tags[j] = NULL;
1971                         }
1972                         blk_mq_exit_hctx(q, set, hctx, j);
1973                         free_cpumask_var(hctx->cpumask);
1974                         kobject_put(&hctx->kobj);
1975                         kfree(hctx->ctxs);
1976                         kfree(hctx);
1977                         hctxs[j] = NULL;
1978
1979                 }
1980         }
1981         q->nr_hw_queues = i;
1982         blk_mq_sysfs_register(q);
1983 }
1984
1985 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1986                                                   struct request_queue *q)
1987 {
1988         /* mark the queue as mq asap */
1989         q->mq_ops = set->ops;
1990
1991         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
1992         if (!q->queue_ctx)
1993                 goto err_exit;
1994
1995         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
1996                                                 GFP_KERNEL, set->numa_node);
1997         if (!q->queue_hw_ctx)
1998                 goto err_percpu;
1999
2000         q->mq_map = set->mq_map;
2001
2002         blk_mq_realloc_hw_ctxs(set, q);
2003         if (!q->nr_hw_queues)
2004                 goto err_hctxs;
2005
2006         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2007         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2008
2009         q->nr_queues = nr_cpu_ids;
2010
2011         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2012
2013         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2014                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2015
2016         q->sg_reserved_size = INT_MAX;
2017
2018         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2019         INIT_LIST_HEAD(&q->requeue_list);
2020         spin_lock_init(&q->requeue_lock);
2021
2022         if (q->nr_hw_queues > 1)
2023                 blk_queue_make_request(q, blk_mq_make_request);
2024         else
2025                 blk_queue_make_request(q, blk_sq_make_request);
2026
2027         /*
2028          * Do this after blk_queue_make_request() overrides it...
2029          */
2030         q->nr_requests = set->queue_depth;
2031
2032         if (set->ops->complete)
2033                 blk_queue_softirq_done(q, set->ops->complete);
2034
2035         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2036
2037         get_online_cpus();
2038         mutex_lock(&all_q_mutex);
2039
2040         list_add_tail(&q->all_q_node, &all_q_list);
2041         blk_mq_add_queue_tag_set(set, q);
2042         blk_mq_map_swqueue(q, cpu_online_mask);
2043
2044         mutex_unlock(&all_q_mutex);
2045         put_online_cpus();
2046
2047         return q;
2048
2049 err_hctxs:
2050         kfree(q->queue_hw_ctx);
2051 err_percpu:
2052         free_percpu(q->queue_ctx);
2053 err_exit:
2054         q->mq_ops = NULL;
2055         return ERR_PTR(-ENOMEM);
2056 }
2057 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2058
2059 void blk_mq_free_queue(struct request_queue *q)
2060 {
2061         struct blk_mq_tag_set   *set = q->tag_set;
2062
2063         mutex_lock(&all_q_mutex);
2064         list_del_init(&q->all_q_node);
2065         mutex_unlock(&all_q_mutex);
2066
2067         blk_mq_del_queue_tag_set(q);
2068
2069         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2070         blk_mq_free_hw_queues(q, set);
2071 }
2072
2073 /* Basically redo blk_mq_init_queue with queue frozen */
2074 static void blk_mq_queue_reinit(struct request_queue *q,
2075                                 const struct cpumask *online_mask)
2076 {
2077         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2078
2079         blk_mq_sysfs_unregister(q);
2080
2081         /*
2082          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2083          * we should change hctx numa_node according to new topology (this
2084          * involves free and re-allocate memory, worthy doing?)
2085          */
2086
2087         blk_mq_map_swqueue(q, online_mask);
2088
2089         blk_mq_sysfs_register(q);
2090 }
2091
2092 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2093                                       unsigned long action, void *hcpu)
2094 {
2095         struct request_queue *q;
2096         int cpu = (unsigned long)hcpu;
2097         /*
2098          * New online cpumask which is going to be set in this hotplug event.
2099          * Declare this cpumasks as global as cpu-hotplug operation is invoked
2100          * one-by-one and dynamically allocating this could result in a failure.
2101          */
2102         static struct cpumask online_new;
2103
2104         /*
2105          * Before hotadded cpu starts handling requests, new mappings must
2106          * be established.  Otherwise, these requests in hw queue might
2107          * never be dispatched.
2108          *
2109          * For example, there is a single hw queue (hctx) and two CPU queues
2110          * (ctx0 for CPU0, and ctx1 for CPU1).
2111          *
2112          * Now CPU1 is just onlined and a request is inserted into
2113          * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2114          * still zero.
2115          *
2116          * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2117          * set in pending bitmap and tries to retrieve requests in
2118          * hctx->ctxs[0]->rq_list.  But htx->ctxs[0] is a pointer to ctx0,
2119          * so the request in ctx1->rq_list is ignored.
2120          */
2121         switch (action & ~CPU_TASKS_FROZEN) {
2122         case CPU_DEAD:
2123         case CPU_UP_CANCELED:
2124                 cpumask_copy(&online_new, cpu_online_mask);
2125                 break;
2126         case CPU_UP_PREPARE:
2127                 cpumask_copy(&online_new, cpu_online_mask);
2128                 cpumask_set_cpu(cpu, &online_new);
2129                 break;
2130         default:
2131                 return NOTIFY_OK;
2132         }
2133
2134         mutex_lock(&all_q_mutex);
2135
2136         /*
2137          * We need to freeze and reinit all existing queues.  Freezing
2138          * involves synchronous wait for an RCU grace period and doing it
2139          * one by one may take a long time.  Start freezing all queues in
2140          * one swoop and then wait for the completions so that freezing can
2141          * take place in parallel.
2142          */
2143         list_for_each_entry(q, &all_q_list, all_q_node)
2144                 blk_mq_freeze_queue_start(q);
2145         list_for_each_entry(q, &all_q_list, all_q_node) {
2146                 blk_mq_freeze_queue_wait(q);
2147
2148                 /*
2149                  * timeout handler can't touch hw queue during the
2150                  * reinitialization
2151                  */
2152                 del_timer_sync(&q->timeout);
2153         }
2154
2155         list_for_each_entry(q, &all_q_list, all_q_node)
2156                 blk_mq_queue_reinit(q, &online_new);
2157
2158         list_for_each_entry(q, &all_q_list, all_q_node)
2159                 blk_mq_unfreeze_queue(q);
2160
2161         mutex_unlock(&all_q_mutex);
2162         return NOTIFY_OK;
2163 }
2164
2165 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2166 {
2167         int i;
2168
2169         for (i = 0; i < set->nr_hw_queues; i++) {
2170                 set->tags[i] = blk_mq_init_rq_map(set, i);
2171                 if (!set->tags[i])
2172                         goto out_unwind;
2173         }
2174
2175         return 0;
2176
2177 out_unwind:
2178         while (--i >= 0)
2179                 blk_mq_free_rq_map(set, set->tags[i], i);
2180
2181         return -ENOMEM;
2182 }
2183
2184 /*
2185  * Allocate the request maps associated with this tag_set. Note that this
2186  * may reduce the depth asked for, if memory is tight. set->queue_depth
2187  * will be updated to reflect the allocated depth.
2188  */
2189 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2190 {
2191         unsigned int depth;
2192         int err;
2193
2194         depth = set->queue_depth;
2195         do {
2196                 err = __blk_mq_alloc_rq_maps(set);
2197                 if (!err)
2198                         break;
2199
2200                 set->queue_depth >>= 1;
2201                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2202                         err = -ENOMEM;
2203                         break;
2204                 }
2205         } while (set->queue_depth);
2206
2207         if (!set->queue_depth || err) {
2208                 pr_err("blk-mq: failed to allocate request map\n");
2209                 return -ENOMEM;
2210         }
2211
2212         if (depth != set->queue_depth)
2213                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2214                                                 depth, set->queue_depth);
2215
2216         return 0;
2217 }
2218
2219 /*
2220  * Alloc a tag set to be associated with one or more request queues.
2221  * May fail with EINVAL for various error conditions. May adjust the
2222  * requested depth down, if if it too large. In that case, the set
2223  * value will be stored in set->queue_depth.
2224  */
2225 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2226 {
2227         int ret;
2228
2229         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2230
2231         if (!set->nr_hw_queues)
2232                 return -EINVAL;
2233         if (!set->queue_depth)
2234                 return -EINVAL;
2235         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2236                 return -EINVAL;
2237
2238         if (!set->ops->queue_rq)
2239                 return -EINVAL;
2240
2241         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2242                 pr_info("blk-mq: reduced tag depth to %u\n",
2243                         BLK_MQ_MAX_DEPTH);
2244                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2245         }
2246
2247         /*
2248          * If a crashdump is active, then we are potentially in a very
2249          * memory constrained environment. Limit us to 1 queue and
2250          * 64 tags to prevent using too much memory.
2251          */
2252         if (is_kdump_kernel()) {
2253                 set->nr_hw_queues = 1;
2254                 set->queue_depth = min(64U, set->queue_depth);
2255         }
2256         /*
2257          * There is no use for more h/w queues than cpus.
2258          */
2259         if (set->nr_hw_queues > nr_cpu_ids)
2260                 set->nr_hw_queues = nr_cpu_ids;
2261
2262         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2263                                  GFP_KERNEL, set->numa_node);
2264         if (!set->tags)
2265                 return -ENOMEM;
2266
2267         ret = -ENOMEM;
2268         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2269                         GFP_KERNEL, set->numa_node);
2270         if (!set->mq_map)
2271                 goto out_free_tags;
2272
2273         if (set->ops->map_queues)
2274                 ret = set->ops->map_queues(set);
2275         else
2276                 ret = blk_mq_map_queues(set);
2277         if (ret)
2278                 goto out_free_mq_map;
2279
2280         ret = blk_mq_alloc_rq_maps(set);
2281         if (ret)
2282                 goto out_free_mq_map;
2283
2284         mutex_init(&set->tag_list_lock);
2285         INIT_LIST_HEAD(&set->tag_list);
2286
2287         return 0;
2288
2289 out_free_mq_map:
2290         kfree(set->mq_map);
2291         set->mq_map = NULL;
2292 out_free_tags:
2293         kfree(set->tags);
2294         set->tags = NULL;
2295         return ret;
2296 }
2297 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2298
2299 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2300 {
2301         int i;
2302
2303         for (i = 0; i < nr_cpu_ids; i++) {
2304                 if (set->tags[i])
2305                         blk_mq_free_rq_map(set, set->tags[i], i);
2306         }
2307
2308         kfree(set->mq_map);
2309         set->mq_map = NULL;
2310
2311         kfree(set->tags);
2312         set->tags = NULL;
2313 }
2314 EXPORT_SYMBOL(blk_mq_free_tag_set);
2315
2316 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2317 {
2318         struct blk_mq_tag_set *set = q->tag_set;
2319         struct blk_mq_hw_ctx *hctx;
2320         int i, ret;
2321
2322         if (!set || nr > set->queue_depth)
2323                 return -EINVAL;
2324
2325         ret = 0;
2326         queue_for_each_hw_ctx(q, hctx, i) {
2327                 if (!hctx->tags)
2328                         continue;
2329                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2330                 if (ret)
2331                         break;
2332         }
2333
2334         if (!ret)
2335                 q->nr_requests = nr;
2336
2337         return ret;
2338 }
2339
2340 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2341 {
2342         struct request_queue *q;
2343
2344         if (nr_hw_queues > nr_cpu_ids)
2345                 nr_hw_queues = nr_cpu_ids;
2346         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2347                 return;
2348
2349         list_for_each_entry(q, &set->tag_list, tag_set_list)
2350                 blk_mq_freeze_queue(q);
2351
2352         set->nr_hw_queues = nr_hw_queues;
2353         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2354                 blk_mq_realloc_hw_ctxs(set, q);
2355
2356                 if (q->nr_hw_queues > 1)
2357                         blk_queue_make_request(q, blk_mq_make_request);
2358                 else
2359                         blk_queue_make_request(q, blk_sq_make_request);
2360
2361                 blk_mq_queue_reinit(q, cpu_online_mask);
2362         }
2363
2364         list_for_each_entry(q, &set->tag_list, tag_set_list)
2365                 blk_mq_unfreeze_queue(q);
2366 }
2367 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2368
2369 void blk_mq_disable_hotplug(void)
2370 {
2371         mutex_lock(&all_q_mutex);
2372 }
2373
2374 void blk_mq_enable_hotplug(void)
2375 {
2376         mutex_unlock(&all_q_mutex);
2377 }
2378
2379 static int __init blk_mq_init(void)
2380 {
2381         blk_mq_cpu_init();
2382
2383         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2384
2385         return 0;
2386 }
2387 subsys_initcall(blk_mq_init);