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 = q->mq_ops->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         hctx = q->queue_hw_ctx[hctx_idx];
 270         ctx = __blk_mq_get_ctx(q, cpumask_first(hctx->cpumask));
 271
 272         blk_mq_set_alloc_data(&alloc_data, q, flags, ctx, hctx);
 273         rq = __blk_mq_alloc_request(&alloc_data, rw, 0);
 274         if (!rq) {
 275                 blk_queue_exit(q);
 276                 return ERR_PTR(-EWOULDBLOCK);
 277         }
 278
 279         return rq;
 280 }
 281 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 282
 283 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 284                                   struct blk_mq_ctx *ctx, struct request *rq)
 285 {
 286         const int tag = rq->tag;
 287         struct request_queue *q = rq->q;
 288
 289         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 290                 atomic_dec(&hctx->nr_active);
 291         rq->cmd_flags = 0;
 292
 293         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 294         blk_mq_put_tag(hctx, ctx, tag);
 295         blk_queue_exit(q);
 296 }
 297
 298 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 299 {
 300         struct blk_mq_ctx *ctx = rq->mq_ctx;
 301
 302         ctx->rq_completed[rq_is_sync(rq)]++;
 303         __blk_mq_free_request(hctx, ctx, rq);
 304
 305 }
 306 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 307
 308 void blk_mq_free_request(struct request *rq)
 309 {
 310         struct blk_mq_hw_ctx *hctx;
 311         struct request_queue *q = rq->q;
 312
 313         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 314         blk_mq_free_hctx_request(hctx, rq);
 315 }
 316 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 317
 318 inline void __blk_mq_end_request(struct request *rq, int error)
 319 {
 320         blk_account_io_done(rq);
 321
 322         if (rq->end_io) {
 323                 rq->end_io(rq, error);
 324         } else {
 325                 if (unlikely(blk_bidi_rq(rq)))
 326                         blk_mq_free_request(rq->next_rq);
 327                 blk_mq_free_request(rq);
 328         }
 329 }
 330 EXPORT_SYMBOL(__blk_mq_end_request);
 331
 332 void blk_mq_end_request(struct request *rq, int error)
 333 {
 334         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 335                 BUG();
 336         __blk_mq_end_request(rq, error);
 337 }
 338 EXPORT_SYMBOL(blk_mq_end_request);
 339
 340 static void __blk_mq_complete_request_remote(void *data)
 341 {
 342         struct request *rq = data;
 343
 344         rq->q->softirq_done_fn(rq);
 345 }
 346
 347 static void blk_mq_ipi_complete_request(struct request *rq)
 348 {
 349         struct blk_mq_ctx *ctx = rq->mq_ctx;
 350         bool shared = false;
 351         int cpu;
 352
 353         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 354                 rq->q->softirq_done_fn(rq);
 355                 return;
 356         }
 357
 358         cpu = get_cpu();
 359         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 360                 shared = cpus_share_cache(cpu, ctx->cpu);
 361
 362         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 363                 rq->csd.func = __blk_mq_complete_request_remote;
 364                 rq->csd.info = rq;
 365                 rq->csd.flags = 0;
 366                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 367         } else {
 368                 rq->q->softirq_done_fn(rq);
 369         }
 370         put_cpu();
 371 }
 372
 373 static void __blk_mq_complete_request(struct request *rq)
 374 {
 375         struct request_queue *q = rq->q;
 376
 377         if (!q->softirq_done_fn)
 378                 blk_mq_end_request(rq, rq->errors);
 379         else
 380                 blk_mq_ipi_complete_request(rq);
 381 }
 382
 383 /**
 384  * blk_mq_complete_request - end I/O on a request
 385  * @rq:         the request being processed
 386  *
 387  * Description:
 388  *      Ends all I/O on a request. It does not handle partial completions.
 389  *      The actual completion happens out-of-order, through a IPI handler.
 390  **/
 391 void blk_mq_complete_request(struct request *rq, int error)
 392 {
 393         struct request_queue *q = rq->q;
 394
 395         if (unlikely(blk_should_fake_timeout(q)))
 396                 return;
 397         if (!blk_mark_rq_complete(rq)) {
 398                 rq->errors = error;
 399                 __blk_mq_complete_request(rq);
 400         }
 401 }
 402 EXPORT_SYMBOL(blk_mq_complete_request);
 403
 404 int blk_mq_request_started(struct request *rq)
 405 {
 406         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 407 }
 408 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 409
 410 void blk_mq_start_request(struct request *rq)
 411 {
 412         struct request_queue *q = rq->q;
 413
 414         trace_block_rq_issue(q, rq);
 415
 416         rq->resid_len = blk_rq_bytes(rq);
 417         if (unlikely(blk_bidi_rq(rq)))
 418                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 419
 420         blk_add_timer(rq);
 421
 422         /*
 423          * Ensure that ->deadline is visible before set the started
 424          * flag and clear the completed flag.
 425          */
 426         smp_mb__before_atomic();
 427
 428         /*
 429          * Mark us as started and clear complete. Complete might have been
 430          * set if requeue raced with timeout, which then marked it as
 431          * complete. So be sure to clear complete again when we start
 432          * the request, otherwise we'll ignore the completion event.
 433          */
 434         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 435                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 436         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 437                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 438
 439         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 440                 /*
 441                  * Make sure space for the drain appears.  We know we can do
 442                  * this because max_hw_segments has been adjusted to be one
 443                  * fewer than the device can handle.
 444                  */
 445                 rq->nr_phys_segments++;
 446         }
 447 }
 448 EXPORT_SYMBOL(blk_mq_start_request);
 449
 450 static void __blk_mq_requeue_request(struct request *rq)
 451 {
 452         struct request_queue *q = rq->q;
 453
 454         trace_block_rq_requeue(q, rq);
 455
 456         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 457                 if (q->dma_drain_size && blk_rq_bytes(rq))
 458                         rq->nr_phys_segments--;
 459         }
 460 }
 461
 462 void blk_mq_requeue_request(struct request *rq)
 463 {
 464         __blk_mq_requeue_request(rq);
 465
 466         BUG_ON(blk_queued_rq(rq));
 467         blk_mq_add_to_requeue_list(rq, true);
 468 }
 469 EXPORT_SYMBOL(blk_mq_requeue_request);
 470
 471 static void blk_mq_requeue_work(struct work_struct *work)
 472 {
 473         struct request_queue *q =
 474                 container_of(work, struct request_queue, requeue_work.work);
 475         LIST_HEAD(rq_list);
 476         struct request *rq, *next;
 477         unsigned long flags;
 478
 479         spin_lock_irqsave(&q->requeue_lock, flags);
 480         list_splice_init(&q->requeue_list, &rq_list);
 481         spin_unlock_irqrestore(&q->requeue_lock, flags);
 482
 483         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 484                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 485                         continue;
 486
 487                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 488                 list_del_init(&rq->queuelist);
 489                 blk_mq_insert_request(rq, true, false, false);
 490         }
 491
 492         while (!list_empty(&rq_list)) {
 493                 rq = list_entry(rq_list.next, struct request, queuelist);
 494                 list_del_init(&rq->queuelist);
 495                 blk_mq_insert_request(rq, false, false, false);
 496         }
 497
 498         /*
 499          * Use the start variant of queue running here, so that running
 500          * the requeue work will kick stopped queues.
 501          */
 502         blk_mq_start_hw_queues(q);
 503 }
 504
 505 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 506 {
 507         struct request_queue *q = rq->q;
 508         unsigned long flags;
 509
 510         /*
 511          * We abuse this flag that is otherwise used by the I/O scheduler to
 512          * request head insertation from the workqueue.
 513          */
 514         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 515
 516         spin_lock_irqsave(&q->requeue_lock, flags);
 517         if (at_head) {
 518                 rq->cmd_flags |= REQ_SOFTBARRIER;
 519                 list_add(&rq->queuelist, &q->requeue_list);
 520         } else {
 521                 list_add_tail(&rq->queuelist, &q->requeue_list);
 522         }
 523         spin_unlock_irqrestore(&q->requeue_lock, flags);
 524 }
 525 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 526
 527 void blk_mq_cancel_requeue_work(struct request_queue *q)
 528 {
 529         cancel_delayed_work_sync(&q->requeue_work);
 530 }
 531 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 532
 533 void blk_mq_kick_requeue_list(struct request_queue *q)
 534 {
 535         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 536 }
 537 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 538
 539 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 540                                     unsigned long msecs)
 541 {
 542         kblockd_schedule_delayed_work(&q->requeue_work,
 543                                       msecs_to_jiffies(msecs));
 544 }
 545 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 546
 547 void blk_mq_abort_requeue_list(struct request_queue *q)
 548 {
 549         unsigned long flags;
 550         LIST_HEAD(rq_list);
 551
 552         spin_lock_irqsave(&q->requeue_lock, flags);
 553         list_splice_init(&q->requeue_list, &rq_list);
 554         spin_unlock_irqrestore(&q->requeue_lock, flags);
 555
 556         while (!list_empty(&rq_list)) {
 557                 struct request *rq;
 558
 559                 rq = list_first_entry(&rq_list, struct request, queuelist);
 560                 list_del_init(&rq->queuelist);
 561                 rq->errors = -EIO;
 562                 blk_mq_end_request(rq, rq->errors);
 563         }
 564 }
 565 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 566
 567 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 568 {
 569         if (tag < tags->nr_tags) {
 570                 prefetch(tags->rqs[tag]);
 571                 return tags->rqs[tag];
 572         }
 573
 574         return NULL;
 575 }
 576 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 577
 578 struct blk_mq_timeout_data {
 579         unsigned long next;
 580         unsigned int next_set;
 581 };
 582
 583 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 584 {
 585         struct blk_mq_ops *ops = req->q->mq_ops;
 586         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 587
 588         /*
 589          * We know that complete is set at this point. If STARTED isn't set
 590          * anymore, then the request isn't active and the "timeout" should
 591          * just be ignored. This can happen due to the bitflag ordering.
 592          * Timeout first checks if STARTED is set, and if it is, assumes
 593          * the request is active. But if we race with completion, then
 594          * we both flags will get cleared. So check here again, and ignore
 595          * a timeout event with a request that isn't active.
 596          */
 597         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 598                 return;
 599
 600         if (ops->timeout)
 601                 ret = ops->timeout(req, reserved);
 602
 603         switch (ret) {
 604         case BLK_EH_HANDLED:
 605                 __blk_mq_complete_request(req);
 606                 break;
 607         case BLK_EH_RESET_TIMER:
 608                 blk_add_timer(req);
 609                 blk_clear_rq_complete(req);
 610                 break;
 611         case BLK_EH_NOT_HANDLED:
 612                 break;
 613         default:
 614                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 615                 break;
 616         }
 617 }
 618
 619 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 620                 struct request *rq, void *priv, bool reserved)
 621 {
 622         struct blk_mq_timeout_data *data = priv;
 623
 624         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 625                 /*
 626                  * If a request wasn't started before the queue was
 627                  * marked dying, kill it here or it'll go unnoticed.
 628                  */
 629                 if (unlikely(blk_queue_dying(rq->q))) {
 630                         rq->errors = -EIO;
 631                         blk_mq_end_request(rq, rq->errors);
 632                 }
 633                 return;
 634         }
 635
 636         if (time_after_eq(jiffies, rq->deadline)) {
 637                 if (!blk_mark_rq_complete(rq))
 638                         blk_mq_rq_timed_out(rq, reserved);
 639         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 640                 data->next = rq->deadline;
 641                 data->next_set = 1;
 642         }
 643 }
 644
 645 static void blk_mq_timeout_work(struct work_struct *work)
 646 {
 647         struct request_queue *q =
 648                 container_of(work, struct request_queue, timeout_work);
 649         struct blk_mq_timeout_data data = {
 650                 .next           = 0,
 651                 .next_set       = 0,
 652         };
 653         int i;
 654
 655         /* A deadlock might occur if a request is stuck requiring a
 656          * timeout at the same time a queue freeze is waiting
 657          * completion, since the timeout code would not be able to
 658          * acquire the queue reference here.
 659          *
 660          * That's why we don't use blk_queue_enter here; instead, we use
 661          * percpu_ref_tryget directly, because we need to be able to
 662          * obtain a reference even in the short window between the queue
 663          * starting to freeze, by dropping the first reference in
 664          * blk_mq_freeze_queue_start, and the moment the last request is
 665          * consumed, marked by the instant q_usage_counter reaches
 666          * zero.
 667          */
 668         if (!percpu_ref_tryget(&q->q_usage_counter))
 669                 return;
 670
 671         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 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                 struct blk_mq_hw_ctx *hctx;
 678
 679                 queue_for_each_hw_ctx(q, hctx, i) {
 680                         /* the hctx may be unmapped, so check it here */
 681                         if (blk_mq_hw_queue_mapped(hctx))
 682                                 blk_mq_tag_idle(hctx);
 683                 }
 684         }
 685         blk_queue_exit(q);
 686 }
 687
 688 /*
 689  * Reverse check our software queue for entries that we could potentially
 690  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 691  * too much time checking for merges.
 692  */
 693 static bool blk_mq_attempt_merge(struct request_queue *q,
 694                                  struct blk_mq_ctx *ctx, struct bio *bio)
 695 {
 696         struct request *rq;
 697         int checked = 8;
 698
 699         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 700                 int el_ret;
 701
 702                 if (!checked--)
 703                         break;
 704
 705                 if (!blk_rq_merge_ok(rq, bio))
 706                         continue;
 707
 708                 el_ret = blk_try_merge(rq, bio);
 709                 if (el_ret == ELEVATOR_BACK_MERGE) {
 710                         if (bio_attempt_back_merge(q, rq, bio)) {
 711                                 ctx->rq_merged++;
 712                                 return true;
 713                         }
 714                         break;
 715                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 716                         if (bio_attempt_front_merge(q, rq, bio)) {
 717                                 ctx->rq_merged++;
 718                                 return true;
 719                         }
 720                         break;
 721                 }
 722         }
 723
 724         return false;
 725 }
 726
 727 struct flush_busy_ctx_data {
 728         struct blk_mq_hw_ctx *hctx;
 729         struct list_head *list;
 730 };
 731
 732 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 733 {
 734         struct flush_busy_ctx_data *flush_data = data;
 735         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 736         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 737
 738         sbitmap_clear_bit(sb, bitnr);
 739         spin_lock(&ctx->lock);
 740         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 741         spin_unlock(&ctx->lock);
 742         return true;
 743 }
 744
 745 /*
 746  * Process software queues that have been marked busy, splicing them
 747  * to the for-dispatch
 748  */
 749 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 750 {
 751         struct flush_busy_ctx_data data = {
 752                 .hctx = hctx,
 753                 .list = list,
 754         };
 755
 756         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 757 }
 758
 759 static inline unsigned int queued_to_index(unsigned int queued)
 760 {
 761         if (!queued)
 762                 return 0;
 763
 764         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 765 }
 766
 767 /*
 768  * Run this hardware queue, pulling any software queues mapped to it in.
 769  * Note that this function currently has various problems around ordering
 770  * of IO. In particular, we'd like FIFO behaviour on handling existing
 771  * items on the hctx->dispatch list. Ignore that for now.
 772  */
 773 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 774 {
 775         struct request_queue *q = hctx->queue;
 776         struct request *rq;
 777         LIST_HEAD(rq_list);
 778         LIST_HEAD(driver_list);
 779         struct list_head *dptr;
 780         int queued;
 781
 782         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 783                 return;
 784
 785         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
 786                 cpu_online(hctx->next_cpu));
 787
 788         hctx->run++;
 789
 790         /*
 791          * Touch any software queue that has pending entries.
 792          */
 793         flush_busy_ctxs(hctx, &rq_list);
 794
 795         /*
 796          * If we have previous entries on our dispatch list, grab them
 797          * and stuff them at the front for more fair dispatch.
 798          */
 799         if (!list_empty_careful(&hctx->dispatch)) {
 800                 spin_lock(&hctx->lock);
 801                 if (!list_empty(&hctx->dispatch))
 802                         list_splice_init(&hctx->dispatch, &rq_list);
 803                 spin_unlock(&hctx->lock);
 804         }
 805
 806         /*
 807          * Start off with dptr being NULL, so we start the first request
 808          * immediately, even if we have more pending.
 809          */
 810         dptr = NULL;
 811
 812         /*
 813          * Now process all the entries, sending them to the driver.
 814          */
 815         queued = 0;
 816         while (!list_empty(&rq_list)) {
 817                 struct blk_mq_queue_data bd;
 818                 int ret;
 819
 820                 rq = list_first_entry(&rq_list, struct request, queuelist);
 821                 list_del_init(&rq->queuelist);
 822
 823                 bd.rq = rq;
 824                 bd.list = dptr;
 825                 bd.last = list_empty(&rq_list);
 826
 827                 ret = q->mq_ops->queue_rq(hctx, &bd);
 828                 switch (ret) {
 829                 case BLK_MQ_RQ_QUEUE_OK:
 830                         queued++;
 831                         break;
 832                 case BLK_MQ_RQ_QUEUE_BUSY:
 833                         list_add(&rq->queuelist, &rq_list);
 834                         __blk_mq_requeue_request(rq);
 835                         break;
 836                 default:
 837                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 838                 case BLK_MQ_RQ_QUEUE_ERROR:
 839                         rq->errors = -EIO;
 840                         blk_mq_end_request(rq, rq->errors);
 841                         break;
 842                 }
 843
 844                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 845                         break;
 846
 847                 /*
 848                  * We've done the first request. If we have more than 1
 849                  * left in the list, set dptr to defer issue.
 850                  */
 851                 if (!dptr && rq_list.next != rq_list.prev)
 852                         dptr = &driver_list;
 853         }
 854
 855         hctx->dispatched[queued_to_index(queued)]++;
 856
 857         /*
 858          * Any items that need requeuing? Stuff them into hctx->dispatch,
 859          * that is where we will continue on next queue run.
 860          */
 861         if (!list_empty(&rq_list)) {
 862                 spin_lock(&hctx->lock);
 863                 list_splice(&rq_list, &hctx->dispatch);
 864                 spin_unlock(&hctx->lock);
 865                 /*
 866                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 867                  * it's possible the queue is stopped and restarted again
 868                  * before this. Queue restart will dispatch requests. And since
 869                  * requests in rq_list aren't added into hctx->dispatch yet,
 870                  * the requests in rq_list might get lost.
 871                  *
 872                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 873                  **/
 874                 blk_mq_run_hw_queue(hctx, true);
 875         }
 876 }
 877
 878 /*
 879  * It'd be great if the workqueue API had a way to pass
 880  * in a mask and had some smarts for more clever placement.
 881  * For now we just round-robin here, switching for every
 882  * BLK_MQ_CPU_WORK_BATCH queued items.
 883  */
 884 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 885 {
 886         if (hctx->queue->nr_hw_queues == 1)
 887                 return WORK_CPU_UNBOUND;
 888
 889         if (--hctx->next_cpu_batch <= 0) {
 890                 int cpu = hctx->next_cpu, next_cpu;
 891
 892                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 893                 if (next_cpu >= nr_cpu_ids)
 894                         next_cpu = cpumask_first(hctx->cpumask);
 895
 896                 hctx->next_cpu = next_cpu;
 897                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 898
 899                 return cpu;
 900         }
 901
 902         return hctx->next_cpu;
 903 }
 904
 905 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 906 {
 907         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 908             !blk_mq_hw_queue_mapped(hctx)))
 909                 return;
 910
 911         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
 912                 int cpu = get_cpu();
 913                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 914                         __blk_mq_run_hw_queue(hctx);
 915                         put_cpu();
 916                         return;
 917                 }
 918
 919                 put_cpu();
 920         }
 921
 922         kblockd_schedule_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work);
 923 }
 924
 925 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 926 {
 927         struct blk_mq_hw_ctx *hctx;
 928         int i;
 929
 930         queue_for_each_hw_ctx(q, hctx, i) {
 931                 if ((!blk_mq_hctx_has_pending(hctx) &&
 932                     list_empty_careful(&hctx->dispatch)) ||
 933                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 934                         continue;
 935
 936                 blk_mq_run_hw_queue(hctx, async);
 937         }
 938 }
 939 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 940
 941 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 942 {
 943         cancel_work(&hctx->run_work);
 944         cancel_delayed_work(&hctx->delay_work);
 945         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 946 }
 947 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 948
 949 void blk_mq_stop_hw_queues(struct request_queue *q)
 950 {
 951         struct blk_mq_hw_ctx *hctx;
 952         int i;
 953
 954         queue_for_each_hw_ctx(q, hctx, i)
 955                 blk_mq_stop_hw_queue(hctx);
 956 }
 957 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 958
 959 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 960 {
 961         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 962
 963         blk_mq_run_hw_queue(hctx, false);
 964 }
 965 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 966
 967 void blk_mq_start_hw_queues(struct request_queue *q)
 968 {
 969         struct blk_mq_hw_ctx *hctx;
 970         int i;
 971
 972         queue_for_each_hw_ctx(q, hctx, i)
 973                 blk_mq_start_hw_queue(hctx);
 974 }
 975 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 976
 977 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 978 {
 979         struct blk_mq_hw_ctx *hctx;
 980         int i;
 981
 982         queue_for_each_hw_ctx(q, hctx, i) {
 983                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 984                         continue;
 985
 986                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 987                 blk_mq_run_hw_queue(hctx, async);
 988         }
 989 }
 990 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 991
 992 static void blk_mq_run_work_fn(struct work_struct *work)
 993 {
 994         struct blk_mq_hw_ctx *hctx;
 995
 996         hctx = container_of(work, struct blk_mq_hw_ctx, run_work);
 997
 998         __blk_mq_run_hw_queue(hctx);
 999 }
1000
1001 static void blk_mq_delay_work_fn(struct work_struct *work)
1002 {
1003         struct blk_mq_hw_ctx *hctx;
1004
1005         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1006
1007         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1008                 __blk_mq_run_hw_queue(hctx);
1009 }
1010
1011 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1012 {
1013         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1014                 return;
1015
1016         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1017                         &hctx->delay_work, msecs_to_jiffies(msecs));
1018 }
1019 EXPORT_SYMBOL(blk_mq_delay_queue);
1020
1021 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1022                                             struct request *rq,
1023                                             bool at_head)
1024 {
1025         struct blk_mq_ctx *ctx = rq->mq_ctx;
1026
1027         trace_block_rq_insert(hctx->queue, rq);
1028
1029         if (at_head)
1030                 list_add(&rq->queuelist, &ctx->rq_list);
1031         else
1032                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1033 }
1034
1035 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1036                                     struct request *rq, bool at_head)
1037 {
1038         struct blk_mq_ctx *ctx = rq->mq_ctx;
1039
1040         __blk_mq_insert_req_list(hctx, rq, at_head);
1041         blk_mq_hctx_mark_pending(hctx, ctx);
1042 }
1043
1044 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1045                            bool async)
1046 {
1047         struct blk_mq_ctx *ctx = rq->mq_ctx;
1048         struct request_queue *q = rq->q;
1049         struct blk_mq_hw_ctx *hctx;
1050
1051         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1052
1053         spin_lock(&ctx->lock);
1054         __blk_mq_insert_request(hctx, rq, at_head);
1055         spin_unlock(&ctx->lock);
1056
1057         if (run_queue)
1058                 blk_mq_run_hw_queue(hctx, async);
1059 }
1060
1061 static void blk_mq_insert_requests(struct request_queue *q,
1062                                      struct blk_mq_ctx *ctx,
1063                                      struct list_head *list,
1064                                      int depth,
1065                                      bool from_schedule)
1066
1067 {
1068         struct blk_mq_hw_ctx *hctx;
1069
1070         trace_block_unplug(q, depth, !from_schedule);
1071
1072         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1073
1074         /*
1075          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1076          * offline now
1077          */
1078         spin_lock(&ctx->lock);
1079         while (!list_empty(list)) {
1080                 struct request *rq;
1081
1082                 rq = list_first_entry(list, struct request, queuelist);
1083                 BUG_ON(rq->mq_ctx != ctx);
1084                 list_del_init(&rq->queuelist);
1085                 __blk_mq_insert_req_list(hctx, rq, false);
1086         }
1087         blk_mq_hctx_mark_pending(hctx, ctx);
1088         spin_unlock(&ctx->lock);
1089
1090         blk_mq_run_hw_queue(hctx, from_schedule);
1091 }
1092
1093 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1094 {
1095         struct request *rqa = container_of(a, struct request, queuelist);
1096         struct request *rqb = container_of(b, struct request, queuelist);
1097
1098         return !(rqa->mq_ctx < rqb->mq_ctx ||
1099                  (rqa->mq_ctx == rqb->mq_ctx &&
1100                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1101 }
1102
1103 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1104 {
1105         struct blk_mq_ctx *this_ctx;
1106         struct request_queue *this_q;
1107         struct request *rq;
1108         LIST_HEAD(list);
1109         LIST_HEAD(ctx_list);
1110         unsigned int depth;
1111
1112         list_splice_init(&plug->mq_list, &list);
1113
1114         list_sort(NULL, &list, plug_ctx_cmp);
1115
1116         this_q = NULL;
1117         this_ctx = NULL;
1118         depth = 0;
1119
1120         while (!list_empty(&list)) {
1121                 rq = list_entry_rq(list.next);
1122                 list_del_init(&rq->queuelist);
1123                 BUG_ON(!rq->q);
1124                 if (rq->mq_ctx != this_ctx) {
1125                         if (this_ctx) {
1126                                 blk_mq_insert_requests(this_q, this_ctx,
1127                                                         &ctx_list, depth,
1128                                                         from_schedule);
1129                         }
1130
1131                         this_ctx = rq->mq_ctx;
1132                         this_q = rq->q;
1133                         depth = 0;
1134                 }
1135
1136                 depth++;
1137                 list_add_tail(&rq->queuelist, &ctx_list);
1138         }
1139
1140         /*
1141          * If 'this_ctx' is set, we know we have entries to complete
1142          * on 'ctx_list'. Do those.
1143          */
1144         if (this_ctx) {
1145                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1146                                        from_schedule);
1147         }
1148 }
1149
1150 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1151 {
1152         init_request_from_bio(rq, bio);
1153
1154         blk_account_io_start(rq, 1);
1155 }
1156
1157 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1158 {
1159         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1160                 !blk_queue_nomerges(hctx->queue);
1161 }
1162
1163 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1164                                          struct blk_mq_ctx *ctx,
1165                                          struct request *rq, struct bio *bio)
1166 {
1167         if (!hctx_allow_merges(hctx) || !bio_mergeable(bio)) {
1168                 blk_mq_bio_to_request(rq, bio);
1169                 spin_lock(&ctx->lock);
1170 insert_rq:
1171                 __blk_mq_insert_request(hctx, rq, false);
1172                 spin_unlock(&ctx->lock);
1173                 return false;
1174         } else {
1175                 struct request_queue *q = hctx->queue;
1176
1177                 spin_lock(&ctx->lock);
1178                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1179                         blk_mq_bio_to_request(rq, bio);
1180                         goto insert_rq;
1181                 }
1182
1183                 spin_unlock(&ctx->lock);
1184                 __blk_mq_free_request(hctx, ctx, rq);
1185                 return true;
1186         }
1187 }
1188
1189 struct blk_map_ctx {
1190         struct blk_mq_hw_ctx *hctx;
1191         struct blk_mq_ctx *ctx;
1192 };
1193
1194 static struct request *blk_mq_map_request(struct request_queue *q,
1195                                           struct bio *bio,
1196                                           struct blk_map_ctx *data)
1197 {
1198         struct blk_mq_hw_ctx *hctx;
1199         struct blk_mq_ctx *ctx;
1200         struct request *rq;
1201         int op = bio_data_dir(bio);
1202         int op_flags = 0;
1203         struct blk_mq_alloc_data alloc_data;
1204
1205         blk_queue_enter_live(q);
1206         ctx = blk_mq_get_ctx(q);
1207         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1208
1209         if (rw_is_sync(bio_op(bio), bio->bi_opf))
1210                 op_flags |= REQ_SYNC;
1211
1212         trace_block_getrq(q, bio, op);
1213         blk_mq_set_alloc_data(&alloc_data, q, 0, ctx, hctx);
1214         rq = __blk_mq_alloc_request(&alloc_data, op, op_flags);
1215
1216         hctx->queued++;
1217         data->hctx = hctx;
1218         data->ctx = ctx;
1219         return rq;
1220 }
1221
1222 static int blk_mq_direct_issue_request(struct request *rq, blk_qc_t *cookie)
1223 {
1224         int ret;
1225         struct request_queue *q = rq->q;
1226         struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1227                         rq->mq_ctx->cpu);
1228         struct blk_mq_queue_data bd = {
1229                 .rq = rq,
1230                 .list = NULL,
1231                 .last = 1
1232         };
1233         blk_qc_t new_cookie = blk_tag_to_qc_t(rq->tag, hctx->queue_num);
1234
1235         /*
1236          * For OK queue, we are done. For error, kill it. Any other
1237          * error (busy), just add it to our list as we previously
1238          * would have done
1239          */
1240         ret = q->mq_ops->queue_rq(hctx, &bd);
1241         if (ret == BLK_MQ_RQ_QUEUE_OK) {
1242                 *cookie = new_cookie;
1243                 return 0;
1244         }
1245
1246         __blk_mq_requeue_request(rq);
1247
1248         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1249                 *cookie = BLK_QC_T_NONE;
1250                 rq->errors = -EIO;
1251                 blk_mq_end_request(rq, rq->errors);
1252                 return 0;
1253         }
1254
1255         return -1;
1256 }
1257
1258 /*
1259  * Multiple hardware queue variant. This will not use per-process plugs,
1260  * but will attempt to bypass the hctx queueing if we can go straight to
1261  * hardware for SYNC IO.
1262  */
1263 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1264 {
1265         const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1266         const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1267         struct blk_map_ctx data;
1268         struct request *rq;
1269         unsigned int request_count = 0;
1270         struct blk_plug *plug;
1271         struct request *same_queue_rq = NULL;
1272         blk_qc_t cookie;
1273
1274         blk_queue_bounce(q, &bio);
1275
1276         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1277                 bio_io_error(bio);
1278                 return BLK_QC_T_NONE;
1279         }
1280
1281         blk_queue_split(q, &bio, q->bio_split);
1282
1283         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1284             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1285                 return BLK_QC_T_NONE;
1286
1287         rq = blk_mq_map_request(q, bio, &data);
1288         if (unlikely(!rq))
1289                 return BLK_QC_T_NONE;
1290
1291         cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1292
1293         if (unlikely(is_flush_fua)) {
1294                 blk_mq_bio_to_request(rq, bio);
1295                 blk_insert_flush(rq);
1296                 goto run_queue;
1297         }
1298
1299         plug = current->plug;
1300         /*
1301          * If the driver supports defer issued based on 'last', then
1302          * queue it up like normal since we can potentially save some
1303          * CPU this way.
1304          */
1305         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1306             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1307                 struct request *old_rq = NULL;
1308
1309                 blk_mq_bio_to_request(rq, bio);
1310
1311                 /*
1312                  * We do limited pluging. If the bio can be merged, do that.
1313                  * Otherwise the existing request in the plug list will be
1314                  * issued. So the plug list will have one request at most
1315                  */
1316                 if (plug) {
1317                         /*
1318                          * The plug list might get flushed before this. If that
1319                          * happens, same_queue_rq is invalid and plug list is
1320                          * empty
1321                          */
1322                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1323                                 old_rq = same_queue_rq;
1324                                 list_del_init(&old_rq->queuelist);
1325                         }
1326                         list_add_tail(&rq->queuelist, &plug->mq_list);
1327                 } else /* is_sync */
1328                         old_rq = rq;
1329                 blk_mq_put_ctx(data.ctx);
1330                 if (!old_rq)
1331                         goto done;
1332                 if (!blk_mq_direct_issue_request(old_rq, &cookie))
1333                         goto done;
1334                 blk_mq_insert_request(old_rq, false, true, true);
1335                 goto done;
1336         }
1337
1338         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1339                 /*
1340                  * For a SYNC request, send it to the hardware immediately. For
1341                  * an ASYNC request, just ensure that we run it later on. The
1342                  * latter allows for merging opportunities and more efficient
1343                  * dispatching.
1344                  */
1345 run_queue:
1346                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1347         }
1348         blk_mq_put_ctx(data.ctx);
1349 done:
1350         return cookie;
1351 }
1352
1353 /*
1354  * Single hardware queue variant. This will attempt to use any per-process
1355  * plug for merging and IO deferral.
1356  */
1357 static blk_qc_t blk_sq_make_request(struct request_queue *q, struct bio *bio)
1358 {
1359         const int is_sync = rw_is_sync(bio_op(bio), bio->bi_opf);
1360         const int is_flush_fua = bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1361         struct blk_plug *plug;
1362         unsigned int request_count = 0;
1363         struct blk_map_ctx data;
1364         struct request *rq;
1365         blk_qc_t cookie;
1366
1367         blk_queue_bounce(q, &bio);
1368
1369         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1370                 bio_io_error(bio);
1371                 return BLK_QC_T_NONE;
1372         }
1373
1374         blk_queue_split(q, &bio, q->bio_split);
1375
1376         if (!is_flush_fua && !blk_queue_nomerges(q)) {
1377                 if (blk_attempt_plug_merge(q, bio, &request_count, NULL))
1378                         return BLK_QC_T_NONE;
1379         } else
1380                 request_count = blk_plug_queued_count(q);
1381
1382         rq = blk_mq_map_request(q, bio, &data);
1383         if (unlikely(!rq))
1384                 return BLK_QC_T_NONE;
1385
1386         cookie = blk_tag_to_qc_t(rq->tag, data.hctx->queue_num);
1387
1388         if (unlikely(is_flush_fua)) {
1389                 blk_mq_bio_to_request(rq, bio);
1390                 blk_insert_flush(rq);
1391                 goto run_queue;
1392         }
1393
1394         /*
1395          * A task plug currently exists. Since this is completely lockless,
1396          * utilize that to temporarily store requests until the task is
1397          * either done or scheduled away.
1398          */
1399         plug = current->plug;
1400         if (plug) {
1401                 blk_mq_bio_to_request(rq, bio);
1402                 if (!request_count)
1403                         trace_block_plug(q);
1404
1405                 blk_mq_put_ctx(data.ctx);
1406
1407                 if (request_count >= BLK_MAX_REQUEST_COUNT) {
1408                         blk_flush_plug_list(plug, false);
1409                         trace_block_plug(q);
1410                 }
1411
1412                 list_add_tail(&rq->queuelist, &plug->mq_list);
1413                 return cookie;
1414         }
1415
1416         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1417                 /*
1418                  * For a SYNC request, send it to the hardware immediately. For
1419                  * an ASYNC request, just ensure that we run it later on. The
1420                  * latter allows for merging opportunities and more efficient
1421                  * dispatching.
1422                  */
1423 run_queue:
1424                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1425         }
1426
1427         blk_mq_put_ctx(data.ctx);
1428         return cookie;
1429 }
1430
1431 /*
1432  * Default mapping to a software queue, since we use one per CPU.
1433  */
1434 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1435 {
1436         return q->queue_hw_ctx[q->mq_map[cpu]];
1437 }
1438 EXPORT_SYMBOL(blk_mq_map_queue);
1439
1440 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1441                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1442 {
1443         struct page *page;
1444
1445         if (tags->rqs && set->ops->exit_request) {
1446                 int i;
1447
1448                 for (i = 0; i < tags->nr_tags; i++) {
1449                         if (!tags->rqs[i])
1450                                 continue;
1451                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1452                                                 hctx_idx, i);
1453                         tags->rqs[i] = NULL;
1454                 }
1455         }
1456
1457         while (!list_empty(&tags->page_list)) {
1458                 page = list_first_entry(&tags->page_list, struct page, lru);
1459                 list_del_init(&page->lru);
1460                 /*
1461                  * Remove kmemleak object previously allocated in
1462                  * blk_mq_init_rq_map().
1463                  */
1464                 kmemleak_free(page_address(page));
1465                 __free_pages(page, page->private);
1466         }
1467
1468         kfree(tags->rqs);
1469
1470         blk_mq_free_tags(tags);
1471 }
1472
1473 static size_t order_to_size(unsigned int order)
1474 {
1475         return (size_t)PAGE_SIZE << order;
1476 }
1477
1478 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1479                 unsigned int hctx_idx)
1480 {
1481         struct blk_mq_tags *tags;
1482         unsigned int i, j, entries_per_page, max_order = 4;
1483         size_t rq_size, left;
1484
1485         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1486                                 set->numa_node,
1487                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1488         if (!tags)
1489                 return NULL;
1490
1491         INIT_LIST_HEAD(&tags->page_list);
1492
1493         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1494                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1495                                  set->numa_node);
1496         if (!tags->rqs) {
1497                 blk_mq_free_tags(tags);
1498                 return NULL;
1499         }
1500
1501         /*
1502          * rq_size is the size of the request plus driver payload, rounded
1503          * to the cacheline size
1504          */
1505         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1506                                 cache_line_size());
1507         left = rq_size * set->queue_depth;
1508
1509         for (i = 0; i < set->queue_depth; ) {
1510                 int this_order = max_order;
1511                 struct page *page;
1512                 int to_do;
1513                 void *p;
1514
1515                 while (this_order && left < order_to_size(this_order - 1))
1516                         this_order--;
1517
1518                 do {
1519                         page = alloc_pages_node(set->numa_node,
1520                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1521                                 this_order);
1522                         if (page)
1523                                 break;
1524                         if (!this_order--)
1525                                 break;
1526                         if (order_to_size(this_order) < rq_size)
1527                                 break;
1528                 } while (1);
1529
1530                 if (!page)
1531                         goto fail;
1532
1533                 page->private = this_order;
1534                 list_add_tail(&page->lru, &tags->page_list);
1535
1536                 p = page_address(page);
1537                 /*
1538                  * Allow kmemleak to scan these pages as they contain pointers
1539                  * to additional allocations like via ops->init_request().
1540                  */
1541                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_KERNEL);
1542                 entries_per_page = order_to_size(this_order) / rq_size;
1543                 to_do = min(entries_per_page, set->queue_depth - i);
1544                 left -= to_do * rq_size;
1545                 for (j = 0; j < to_do; j++) {
1546                         tags->rqs[i] = p;
1547                         if (set->ops->init_request) {
1548                                 if (set->ops->init_request(set->driver_data,
1549                                                 tags->rqs[i], hctx_idx, i,
1550                                                 set->numa_node)) {
1551                                         tags->rqs[i] = NULL;
1552                                         goto fail;
1553                                 }
1554                         }
1555
1556                         p += rq_size;
1557                         i++;
1558                 }
1559         }
1560         return tags;
1561
1562 fail:
1563         blk_mq_free_rq_map(set, tags, hctx_idx);
1564         return NULL;
1565 }
1566
1567 /*
1568  * 'cpu' is going away. splice any existing rq_list entries from this
1569  * software queue to the hw queue dispatch list, and ensure that it
1570  * gets run.
1571  */
1572 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1573 {
1574         struct blk_mq_ctx *ctx;
1575         LIST_HEAD(tmp);
1576
1577         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1578
1579         spin_lock(&ctx->lock);
1580         if (!list_empty(&ctx->rq_list)) {
1581                 list_splice_init(&ctx->rq_list, &tmp);
1582                 blk_mq_hctx_clear_pending(hctx, ctx);
1583         }
1584         spin_unlock(&ctx->lock);
1585
1586         if (list_empty(&tmp))
1587                 return NOTIFY_OK;
1588
1589         spin_lock(&hctx->lock);
1590         list_splice_tail_init(&tmp, &hctx->dispatch);
1591         spin_unlock(&hctx->lock);
1592
1593         blk_mq_run_hw_queue(hctx, true);
1594         return NOTIFY_OK;
1595 }
1596
1597 static int blk_mq_hctx_notify(void *data, unsigned long action,
1598                               unsigned int cpu)
1599 {
1600         struct blk_mq_hw_ctx *hctx = data;
1601
1602         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1603                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1604
1605         /*
1606          * In case of CPU online, tags may be reallocated
1607          * in blk_mq_map_swqueue() after mapping is updated.
1608          */
1609
1610         return NOTIFY_OK;
1611 }
1612
1613 /* hctx->ctxs will be freed in queue's release handler */
1614 static void blk_mq_exit_hctx(struct request_queue *q,
1615                 struct blk_mq_tag_set *set,
1616                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1617 {
1618         unsigned flush_start_tag = set->queue_depth;
1619
1620         blk_mq_tag_idle(hctx);
1621
1622         if (set->ops->exit_request)
1623                 set->ops->exit_request(set->driver_data,
1624                                        hctx->fq->flush_rq, hctx_idx,
1625                                        flush_start_tag + hctx_idx);
1626
1627         if (set->ops->exit_hctx)
1628                 set->ops->exit_hctx(hctx, hctx_idx);
1629
1630         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1631         blk_free_flush_queue(hctx->fq);
1632         sbitmap_free(&hctx->ctx_map);
1633 }
1634
1635 static void blk_mq_exit_hw_queues(struct request_queue *q,
1636                 struct blk_mq_tag_set *set, int nr_queue)
1637 {
1638         struct blk_mq_hw_ctx *hctx;
1639         unsigned int i;
1640
1641         queue_for_each_hw_ctx(q, hctx, i) {
1642                 if (i == nr_queue)
1643                         break;
1644                 blk_mq_exit_hctx(q, set, hctx, i);
1645         }
1646 }
1647
1648 static void blk_mq_free_hw_queues(struct request_queue *q,
1649                 struct blk_mq_tag_set *set)
1650 {
1651         struct blk_mq_hw_ctx *hctx;
1652         unsigned int i;
1653
1654         queue_for_each_hw_ctx(q, hctx, i)
1655                 free_cpumask_var(hctx->cpumask);
1656 }
1657
1658 static int blk_mq_init_hctx(struct request_queue *q,
1659                 struct blk_mq_tag_set *set,
1660                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1661 {
1662         int node;
1663         unsigned flush_start_tag = set->queue_depth;
1664
1665         node = hctx->numa_node;
1666         if (node == NUMA_NO_NODE)
1667                 node = hctx->numa_node = set->numa_node;
1668
1669         INIT_WORK(&hctx->run_work, blk_mq_run_work_fn);
1670         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1671         spin_lock_init(&hctx->lock);
1672         INIT_LIST_HEAD(&hctx->dispatch);
1673         hctx->queue = q;
1674         hctx->queue_num = hctx_idx;
1675         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1676
1677         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1678                                         blk_mq_hctx_notify, hctx);
1679         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1680
1681         hctx->tags = set->tags[hctx_idx];
1682
1683         /*
1684          * Allocate space for all possible cpus to avoid allocation at
1685          * runtime
1686          */
1687         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1688                                         GFP_KERNEL, node);
1689         if (!hctx->ctxs)
1690                 goto unregister_cpu_notifier;
1691
1692         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1693                               node))
1694                 goto free_ctxs;
1695
1696         hctx->nr_ctx = 0;
1697
1698         if (set->ops->init_hctx &&
1699             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1700                 goto free_bitmap;
1701
1702         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1703         if (!hctx->fq)
1704                 goto exit_hctx;
1705
1706         if (set->ops->init_request &&
1707             set->ops->init_request(set->driver_data,
1708                                    hctx->fq->flush_rq, hctx_idx,
1709                                    flush_start_tag + hctx_idx, node))
1710                 goto free_fq;
1711
1712         return 0;
1713
1714  free_fq:
1715         kfree(hctx->fq);
1716  exit_hctx:
1717         if (set->ops->exit_hctx)
1718                 set->ops->exit_hctx(hctx, hctx_idx);
1719  free_bitmap:
1720         sbitmap_free(&hctx->ctx_map);
1721  free_ctxs:
1722         kfree(hctx->ctxs);
1723  unregister_cpu_notifier:
1724         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1725
1726         return -1;
1727 }
1728
1729 static void blk_mq_init_cpu_queues(struct request_queue *q,
1730                                    unsigned int nr_hw_queues)
1731 {
1732         unsigned int i;
1733
1734         for_each_possible_cpu(i) {
1735                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1736                 struct blk_mq_hw_ctx *hctx;
1737
1738                 memset(__ctx, 0, sizeof(*__ctx));
1739                 __ctx->cpu = i;
1740                 spin_lock_init(&__ctx->lock);
1741                 INIT_LIST_HEAD(&__ctx->rq_list);
1742                 __ctx->queue = q;
1743
1744                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1745                 if (!cpu_online(i))
1746                         continue;
1747
1748                 hctx = q->mq_ops->map_queue(q, i);
1749
1750                 /*
1751                  * Set local node, IFF we have more than one hw queue. If
1752                  * not, we remain on the home node of the device
1753                  */
1754                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1755                         hctx->numa_node = local_memory_node(cpu_to_node(i));
1756         }
1757 }
1758
1759 static void blk_mq_map_swqueue(struct request_queue *q,
1760                                const struct cpumask *online_mask)
1761 {
1762         unsigned int i;
1763         struct blk_mq_hw_ctx *hctx;
1764         struct blk_mq_ctx *ctx;
1765         struct blk_mq_tag_set *set = q->tag_set;
1766
1767         /*
1768          * Avoid others reading imcomplete hctx->cpumask through sysfs
1769          */
1770         mutex_lock(&q->sysfs_lock);
1771
1772         queue_for_each_hw_ctx(q, hctx, i) {
1773                 cpumask_clear(hctx->cpumask);
1774                 hctx->nr_ctx = 0;
1775         }
1776
1777         /*
1778          * Map software to hardware queues
1779          */
1780         for_each_possible_cpu(i) {
1781                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1782                 if (!cpumask_test_cpu(i, online_mask))
1783                         continue;
1784
1785                 ctx = per_cpu_ptr(q->queue_ctx, i);
1786                 hctx = q->mq_ops->map_queue(q, i);
1787
1788                 cpumask_set_cpu(i, hctx->cpumask);
1789                 ctx->index_hw = hctx->nr_ctx;
1790                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1791         }
1792
1793         mutex_unlock(&q->sysfs_lock);
1794
1795         queue_for_each_hw_ctx(q, hctx, i) {
1796                 /*
1797                  * If no software queues are mapped to this hardware queue,
1798                  * disable it and free the request entries.
1799                  */
1800                 if (!hctx->nr_ctx) {
1801                         if (set->tags[i]) {
1802                                 blk_mq_free_rq_map(set, set->tags[i], i);
1803                                 set->tags[i] = NULL;
1804                         }
1805                         hctx->tags = NULL;
1806                         continue;
1807                 }
1808
1809                 /* unmapped hw queue can be remapped after CPU topo changed */
1810                 if (!set->tags[i])
1811                         set->tags[i] = blk_mq_init_rq_map(set, i);
1812                 hctx->tags = set->tags[i];
1813                 WARN_ON(!hctx->tags);
1814
1815                 cpumask_copy(hctx->tags->cpumask, hctx->cpumask);
1816                 /*
1817                  * Set the map size to the number of mapped software queues.
1818                  * This is more accurate and more efficient than looping
1819                  * over all possibly mapped software queues.
1820                  */
1821                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
1822
1823                 /*
1824                  * Initialize batch roundrobin counts
1825                  */
1826                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1827                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1828         }
1829 }
1830
1831 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
1832 {
1833         struct blk_mq_hw_ctx *hctx;
1834         int i;
1835
1836         queue_for_each_hw_ctx(q, hctx, i) {
1837                 if (shared)
1838                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
1839                 else
1840                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1841         }
1842 }
1843
1844 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
1845 {
1846         struct request_queue *q;
1847
1848         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1849                 blk_mq_freeze_queue(q);
1850                 queue_set_hctx_shared(q, shared);
1851                 blk_mq_unfreeze_queue(q);
1852         }
1853 }
1854
1855 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1856 {
1857         struct blk_mq_tag_set *set = q->tag_set;
1858
1859         mutex_lock(&set->tag_list_lock);
1860         list_del_init(&q->tag_set_list);
1861         if (list_is_singular(&set->tag_list)) {
1862                 /* just transitioned to unshared */
1863                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
1864                 /* update existing queue */
1865                 blk_mq_update_tag_set_depth(set, false);
1866         }
1867         mutex_unlock(&set->tag_list_lock);
1868 }
1869
1870 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1871                                      struct request_queue *q)
1872 {
1873         q->tag_set = set;
1874
1875         mutex_lock(&set->tag_list_lock);
1876
1877         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
1878         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
1879                 set->flags |= BLK_MQ_F_TAG_SHARED;
1880                 /* update existing queue */
1881                 blk_mq_update_tag_set_depth(set, true);
1882         }
1883         if (set->flags & BLK_MQ_F_TAG_SHARED)
1884                 queue_set_hctx_shared(q, true);
1885         list_add_tail(&q->tag_set_list, &set->tag_list);
1886
1887         mutex_unlock(&set->tag_list_lock);
1888 }
1889
1890 /*
1891  * It is the actual release handler for mq, but we do it from
1892  * request queue's release handler for avoiding use-after-free
1893  * and headache because q->mq_kobj shouldn't have been introduced,
1894  * but we can't group ctx/kctx kobj without it.
1895  */
1896 void blk_mq_release(struct request_queue *q)
1897 {
1898         struct blk_mq_hw_ctx *hctx;
1899         unsigned int i;
1900
1901         /* hctx kobj stays in hctx */
1902         queue_for_each_hw_ctx(q, hctx, i) {
1903                 if (!hctx)
1904                         continue;
1905                 kfree(hctx->ctxs);
1906                 kfree(hctx);
1907         }
1908
1909         kfree(q->mq_map);
1910         q->mq_map = NULL;
1911
1912         kfree(q->queue_hw_ctx);
1913
1914         /* ctx kobj stays in queue_ctx */
1915         free_percpu(q->queue_ctx);
1916 }
1917
1918 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1919 {
1920         struct request_queue *uninit_q, *q;
1921
1922         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1923         if (!uninit_q)
1924                 return ERR_PTR(-ENOMEM);
1925
1926         q = blk_mq_init_allocated_queue(set, uninit_q);
1927         if (IS_ERR(q))
1928                 blk_cleanup_queue(uninit_q);
1929
1930         return q;
1931 }
1932 EXPORT_SYMBOL(blk_mq_init_queue);
1933
1934 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
1935                                                 struct request_queue *q)
1936 {
1937         int i, j;
1938         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
1939
1940         blk_mq_sysfs_unregister(q);
1941         for (i = 0; i < set->nr_hw_queues; i++) {
1942                 int node;
1943
1944                 if (hctxs[i])
1945                         continue;
1946
1947                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
1948                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1949                                         GFP_KERNEL, node);
1950                 if (!hctxs[i])
1951                         break;
1952
1953                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1954                                                 node)) {
1955                         kfree(hctxs[i]);
1956                         hctxs[i] = NULL;
1957                         break;
1958                 }
1959
1960                 atomic_set(&hctxs[i]->nr_active, 0);
1961                 hctxs[i]->numa_node = node;
1962                 hctxs[i]->queue_num = i;
1963
1964                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
1965                         free_cpumask_var(hctxs[i]->cpumask);
1966                         kfree(hctxs[i]);
1967                         hctxs[i] = NULL;
1968                         break;
1969                 }
1970                 blk_mq_hctx_kobj_init(hctxs[i]);
1971         }
1972         for (j = i; j < q->nr_hw_queues; j++) {
1973                 struct blk_mq_hw_ctx *hctx = hctxs[j];
1974
1975                 if (hctx) {
1976                         if (hctx->tags) {
1977                                 blk_mq_free_rq_map(set, hctx->tags, j);
1978                                 set->tags[j] = NULL;
1979                         }
1980                         blk_mq_exit_hctx(q, set, hctx, j);
1981                         free_cpumask_var(hctx->cpumask);
1982                         kobject_put(&hctx->kobj);
1983                         kfree(hctx->ctxs);
1984                         kfree(hctx);
1985                         hctxs[j] = NULL;
1986
1987                 }
1988         }
1989         q->nr_hw_queues = i;
1990         blk_mq_sysfs_register(q);
1991 }
1992
1993 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1994                                                   struct request_queue *q)
1995 {
1996         /* mark the queue as mq asap */
1997         q->mq_ops = set->ops;
1998
1999         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2000         if (!q->queue_ctx)
2001                 goto err_exit;
2002
2003         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2004                                                 GFP_KERNEL, set->numa_node);
2005         if (!q->queue_hw_ctx)
2006                 goto err_percpu;
2007
2008         q->mq_map = blk_mq_make_queue_map(set);
2009         if (!q->mq_map)
2010                 goto err_map;
2011
2012         blk_mq_realloc_hw_ctxs(set, q);
2013         if (!q->nr_hw_queues)
2014                 goto err_hctxs;
2015
2016         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2017         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2018
2019         q->nr_queues = nr_cpu_ids;
2020
2021         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2022
2023         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2024                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2025
2026         q->sg_reserved_size = INT_MAX;
2027
2028         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2029         INIT_LIST_HEAD(&q->requeue_list);
2030         spin_lock_init(&q->requeue_lock);
2031
2032         if (q->nr_hw_queues > 1)
2033                 blk_queue_make_request(q, blk_mq_make_request);
2034         else
2035                 blk_queue_make_request(q, blk_sq_make_request);
2036
2037         /*
2038          * Do this after blk_queue_make_request() overrides it...
2039          */
2040         q->nr_requests = set->queue_depth;
2041
2042         if (set->ops->complete)
2043                 blk_queue_softirq_done(q, set->ops->complete);
2044
2045         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2046
2047         get_online_cpus();
2048         mutex_lock(&all_q_mutex);
2049
2050         list_add_tail(&q->all_q_node, &all_q_list);
2051         blk_mq_add_queue_tag_set(set, q);
2052         blk_mq_map_swqueue(q, cpu_online_mask);
2053
2054         mutex_unlock(&all_q_mutex);
2055         put_online_cpus();
2056
2057         return q;
2058
2059 err_hctxs:
2060         kfree(q->mq_map);
2061 err_map:
2062         kfree(q->queue_hw_ctx);
2063 err_percpu:
2064         free_percpu(q->queue_ctx);
2065 err_exit:
2066         q->mq_ops = NULL;
2067         return ERR_PTR(-ENOMEM);
2068 }
2069 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2070
2071 void blk_mq_free_queue(struct request_queue *q)
2072 {
2073         struct blk_mq_tag_set   *set = q->tag_set;
2074
2075         mutex_lock(&all_q_mutex);
2076         list_del_init(&q->all_q_node);
2077         mutex_unlock(&all_q_mutex);
2078
2079         blk_mq_del_queue_tag_set(q);
2080
2081         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2082         blk_mq_free_hw_queues(q, set);
2083 }
2084
2085 /* Basically redo blk_mq_init_queue with queue frozen */
2086 static void blk_mq_queue_reinit(struct request_queue *q,
2087                                 const struct cpumask *online_mask)
2088 {
2089         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2090
2091         blk_mq_sysfs_unregister(q);
2092
2093         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2094
2095         /*
2096          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2097          * we should change hctx numa_node according to new topology (this
2098          * involves free and re-allocate memory, worthy doing?)
2099          */
2100
2101         blk_mq_map_swqueue(q, online_mask);
2102
2103         blk_mq_sysfs_register(q);
2104 }
2105
2106 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2107                                       unsigned long action, void *hcpu)
2108 {
2109         struct request_queue *q;
2110         int cpu = (unsigned long)hcpu;
2111         /*
2112          * New online cpumask which is going to be set in this hotplug event.
2113          * Declare this cpumasks as global as cpu-hotplug operation is invoked
2114          * one-by-one and dynamically allocating this could result in a failure.
2115          */
2116         static struct cpumask online_new;
2117
2118         /*
2119          * Before hotadded cpu starts handling requests, new mappings must
2120          * be established.  Otherwise, these requests in hw queue might
2121          * never be dispatched.
2122          *
2123          * For example, there is a single hw queue (hctx) and two CPU queues
2124          * (ctx0 for CPU0, and ctx1 for CPU1).
2125          *
2126          * Now CPU1 is just onlined and a request is inserted into
2127          * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2128          * still zero.
2129          *
2130          * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2131          * set in pending bitmap and tries to retrieve requests in
2132          * hctx->ctxs[0]->rq_list.  But htx->ctxs[0] is a pointer to ctx0,
2133          * so the request in ctx1->rq_list is ignored.
2134          */
2135         switch (action & ~CPU_TASKS_FROZEN) {
2136         case CPU_DEAD:
2137         case CPU_UP_CANCELED:
2138                 cpumask_copy(&online_new, cpu_online_mask);
2139                 break;
2140         case CPU_UP_PREPARE:
2141                 cpumask_copy(&online_new, cpu_online_mask);
2142                 cpumask_set_cpu(cpu, &online_new);
2143                 break;
2144         default:
2145                 return NOTIFY_OK;
2146         }
2147
2148         mutex_lock(&all_q_mutex);
2149
2150         /*
2151          * We need to freeze and reinit all existing queues.  Freezing
2152          * involves synchronous wait for an RCU grace period and doing it
2153          * one by one may take a long time.  Start freezing all queues in
2154          * one swoop and then wait for the completions so that freezing can
2155          * take place in parallel.
2156          */
2157         list_for_each_entry(q, &all_q_list, all_q_node)
2158                 blk_mq_freeze_queue_start(q);
2159         list_for_each_entry(q, &all_q_list, all_q_node) {
2160                 blk_mq_freeze_queue_wait(q);
2161
2162                 /*
2163                  * timeout handler can't touch hw queue during the
2164                  * reinitialization
2165                  */
2166                 del_timer_sync(&q->timeout);
2167         }
2168
2169         list_for_each_entry(q, &all_q_list, all_q_node)
2170                 blk_mq_queue_reinit(q, &online_new);
2171
2172         list_for_each_entry(q, &all_q_list, all_q_node)
2173                 blk_mq_unfreeze_queue(q);
2174
2175         mutex_unlock(&all_q_mutex);
2176         return NOTIFY_OK;
2177 }
2178
2179 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2180 {
2181         int i;
2182
2183         for (i = 0; i < set->nr_hw_queues; i++) {
2184                 set->tags[i] = blk_mq_init_rq_map(set, i);
2185                 if (!set->tags[i])
2186                         goto out_unwind;
2187         }
2188
2189         return 0;
2190
2191 out_unwind:
2192         while (--i >= 0)
2193                 blk_mq_free_rq_map(set, set->tags[i], i);
2194
2195         return -ENOMEM;
2196 }
2197
2198 /*
2199  * Allocate the request maps associated with this tag_set. Note that this
2200  * may reduce the depth asked for, if memory is tight. set->queue_depth
2201  * will be updated to reflect the allocated depth.
2202  */
2203 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2204 {
2205         unsigned int depth;
2206         int err;
2207
2208         depth = set->queue_depth;
2209         do {
2210                 err = __blk_mq_alloc_rq_maps(set);
2211                 if (!err)
2212                         break;
2213
2214                 set->queue_depth >>= 1;
2215                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2216                         err = -ENOMEM;
2217                         break;
2218                 }
2219         } while (set->queue_depth);
2220
2221         if (!set->queue_depth || err) {
2222                 pr_err("blk-mq: failed to allocate request map\n");
2223                 return -ENOMEM;
2224         }
2225
2226         if (depth != set->queue_depth)
2227                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2228                                                 depth, set->queue_depth);
2229
2230         return 0;
2231 }
2232
2233 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2234 {
2235         return tags->cpumask;
2236 }
2237 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2238
2239 /*
2240  * Alloc a tag set to be associated with one or more request queues.
2241  * May fail with EINVAL for various error conditions. May adjust the
2242  * requested depth down, if if it too large. In that case, the set
2243  * value will be stored in set->queue_depth.
2244  */
2245 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2246 {
2247         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2248
2249         if (!set->nr_hw_queues)
2250                 return -EINVAL;
2251         if (!set->queue_depth)
2252                 return -EINVAL;
2253         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2254                 return -EINVAL;
2255
2256         if (!set->ops->queue_rq || !set->ops->map_queue)
2257                 return -EINVAL;
2258
2259         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2260                 pr_info("blk-mq: reduced tag depth to %u\n",
2261                         BLK_MQ_MAX_DEPTH);
2262                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2263         }
2264
2265         /*
2266          * If a crashdump is active, then we are potentially in a very
2267          * memory constrained environment. Limit us to 1 queue and
2268          * 64 tags to prevent using too much memory.
2269          */
2270         if (is_kdump_kernel()) {
2271                 set->nr_hw_queues = 1;
2272                 set->queue_depth = min(64U, set->queue_depth);
2273         }
2274         /*
2275          * There is no use for more h/w queues than cpus.
2276          */
2277         if (set->nr_hw_queues > nr_cpu_ids)
2278                 set->nr_hw_queues = nr_cpu_ids;
2279
2280         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2281                                  GFP_KERNEL, set->numa_node);
2282         if (!set->tags)
2283                 return -ENOMEM;
2284
2285         if (blk_mq_alloc_rq_maps(set))
2286                 goto enomem;
2287
2288         mutex_init(&set->tag_list_lock);
2289         INIT_LIST_HEAD(&set->tag_list);
2290
2291         return 0;
2292 enomem:
2293         kfree(set->tags);
2294         set->tags = NULL;
2295         return -ENOMEM;
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->tags);
2309         set->tags = NULL;
2310 }
2311 EXPORT_SYMBOL(blk_mq_free_tag_set);
2312
2313 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2314 {
2315         struct blk_mq_tag_set *set = q->tag_set;
2316         struct blk_mq_hw_ctx *hctx;
2317         int i, ret;
2318
2319         if (!set || nr > set->queue_depth)
2320                 return -EINVAL;
2321
2322         ret = 0;
2323         queue_for_each_hw_ctx(q, hctx, i) {
2324                 if (!hctx->tags)
2325                         continue;
2326                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2327                 if (ret)
2328                         break;
2329         }
2330
2331         if (!ret)
2332                 q->nr_requests = nr;
2333
2334         return ret;
2335 }
2336
2337 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2338 {
2339         struct request_queue *q;
2340
2341         if (nr_hw_queues > nr_cpu_ids)
2342                 nr_hw_queues = nr_cpu_ids;
2343         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2344                 return;
2345
2346         list_for_each_entry(q, &set->tag_list, tag_set_list)
2347                 blk_mq_freeze_queue(q);
2348
2349         set->nr_hw_queues = nr_hw_queues;
2350         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2351                 blk_mq_realloc_hw_ctxs(set, q);
2352
2353                 if (q->nr_hw_queues > 1)
2354                         blk_queue_make_request(q, blk_mq_make_request);
2355                 else
2356                         blk_queue_make_request(q, blk_sq_make_request);
2357
2358                 blk_mq_queue_reinit(q, cpu_online_mask);
2359         }
2360
2361         list_for_each_entry(q, &set->tag_list, tag_set_list)
2362                 blk_mq_unfreeze_queue(q);
2363 }
2364 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2365
2366 void blk_mq_disable_hotplug(void)
2367 {
2368         mutex_lock(&all_q_mutex);
2369 }
2370
2371 void blk_mq_enable_hotplug(void)
2372 {
2373         mutex_unlock(&all_q_mutex);
2374 }
2375
2376 static int __init blk_mq_init(void)
2377 {
2378         blk_mq_cpu_init();
2379
2380         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2381
2382         return 0;
2383 }
2384 subsys_initcall(blk_mq_init);