block/blk-mq.c

   1 /*
   2  * Block multiqueue core code
   3  *
   4  * Copyright (C) 2013-2014 Jens Axboe
   5  * Copyright (C) 2013-2014 Christoph Hellwig
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/mm.h>
  13 #include <linux/init.h>
  14 #include <linux/slab.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/smp.h>
  17 #include <linux/llist.h>
  18 #include <linux/list_sort.h>
  19 #include <linux/cpu.h>
  20 #include <linux/cache.h>
  21 #include <linux/sched/sysctl.h>
  22 #include <linux/delay.h>
  23 #include <linux/crash_dump.h>
  24
  25 #include <trace/events/block.h>
  26
  27 #include <linux/blk-mq.h>
  28 #include "blk.h"
  29 #include "blk-mq.h"
  30 #include "blk-mq-tag.h"
  31
  32 static DEFINE_MUTEX(all_q_mutex);
  33 static LIST_HEAD(all_q_list);
  34
  35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  36
  37 /*
  38  * Check if any of the ctx's have pending work in this hardware queue
  39  */
  40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  41 {
  42         unsigned int i;
  43
  44         for (i = 0; i < hctx->ctx_map.map_size; i++)
  45                 if (hctx->ctx_map.map[i].word)
  46                         return true;
  47
  48         return false;
  49 }
  50
  51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  52                                               struct blk_mq_ctx *ctx)
  53 {
  54         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  55 }
  56
  57 #define CTX_TO_BIT(hctx, ctx)   \
  58         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  59
  60 /*
  61  * Mark this ctx as having pending work in this hardware queue
  62  */
  63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  64                                      struct blk_mq_ctx *ctx)
  65 {
  66         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  67
  68         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  69                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  70 }
  71
  72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  73                                       struct blk_mq_ctx *ctx)
  74 {
  75         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  76
  77         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  78 }
  79
  80 static int blk_mq_queue_enter(struct request_queue *q)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 ret = wait_event_interruptible(q->mq_freeze_wq,
  89                                 !q->mq_freeze_depth || blk_queue_dying(q));
  90                 if (blk_queue_dying(q))
  91                         return -ENODEV;
  92                 if (ret)
  93                         return ret;
  94         }
  95 }
  96
  97 static void blk_mq_queue_exit(struct request_queue *q)
  98 {
  99         percpu_ref_put(&q->mq_usage_counter);
 100 }
 101
 102 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 103 {
 104         struct request_queue *q =
 105                 container_of(ref, struct request_queue, mq_usage_counter);
 106
 107         wake_up_all(&q->mq_freeze_wq);
 108 }
 109
 110 void blk_mq_freeze_queue_start(struct request_queue *q)
 111 {
 112         bool freeze;
 113
 114         spin_lock_irq(q->queue_lock);
 115         freeze = !q->mq_freeze_depth++;
 116         spin_unlock_irq(q->queue_lock);
 117
 118         if (freeze) {
 119                 percpu_ref_kill(&q->mq_usage_counter);
 120                 blk_mq_run_queues(q, false);
 121         }
 122 }
 123 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 124
 125 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 126 {
 127         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 128 }
 129
 130 /*
 131  * Guarantee no request is in use, so we can change any data structure of
 132  * the queue afterward.
 133  */
 134 void blk_mq_freeze_queue(struct request_queue *q)
 135 {
 136         blk_mq_freeze_queue_start(q);
 137         blk_mq_freeze_queue_wait(q);
 138 }
 139
 140 void blk_mq_unfreeze_queue(struct request_queue *q)
 141 {
 142         bool wake;
 143
 144         spin_lock_irq(q->queue_lock);
 145         wake = !--q->mq_freeze_depth;
 146         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 147         spin_unlock_irq(q->queue_lock);
 148         if (wake) {
 149                 percpu_ref_reinit(&q->mq_usage_counter);
 150                 wake_up_all(&q->mq_freeze_wq);
 151         }
 152 }
 153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 154
 155 void blk_mq_wake_waiters(struct request_queue *q)
 156 {
 157         struct blk_mq_hw_ctx *hctx;
 158         unsigned int i;
 159
 160         queue_for_each_hw_ctx(q, hctx, i)
 161                 if (blk_mq_hw_queue_mapped(hctx))
 162                         blk_mq_tag_wakeup_all(hctx->tags, true);
 163
 164         /*
 165          * If we are called because the queue has now been marked as
 166          * dying, we need to ensure that processes currently waiting on
 167          * the queue are notified as well.
 168          */
 169         wake_up_all(&q->mq_freeze_wq);
 170 }
 171
 172 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 173 {
 174         return blk_mq_has_free_tags(hctx->tags);
 175 }
 176 EXPORT_SYMBOL(blk_mq_can_queue);
 177
 178 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 179                                struct request *rq, unsigned int rw_flags)
 180 {
 181         if (blk_queue_io_stat(q))
 182                 rw_flags |= REQ_IO_STAT;
 183
 184         INIT_LIST_HEAD(&rq->queuelist);
 185         /* csd/requeue_work/fifo_time is initialized before use */
 186         rq->q = q;
 187         rq->mq_ctx = ctx;
 188         rq->cmd_flags |= rw_flags;
 189         /* do not touch atomic flags, it needs atomic ops against the timer */
 190         rq->cpu = -1;
 191         INIT_HLIST_NODE(&rq->hash);
 192         RB_CLEAR_NODE(&rq->rb_node);
 193         rq->rq_disk = NULL;
 194         rq->part = NULL;
 195         rq->start_time = jiffies;
 196 #ifdef CONFIG_BLK_CGROUP
 197         rq->rl = NULL;
 198         set_start_time_ns(rq);
 199         rq->io_start_time_ns = 0;
 200 #endif
 201         rq->nr_phys_segments = 0;
 202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 203         rq->nr_integrity_segments = 0;
 204 #endif
 205         rq->special = NULL;
 206         /* tag was already set */
 207         rq->errors = 0;
 208
 209         rq->cmd = rq->__cmd;
 210
 211         rq->extra_len = 0;
 212         rq->sense_len = 0;
 213         rq->resid_len = 0;
 214         rq->sense = NULL;
 215
 216         INIT_LIST_HEAD(&rq->timeout_list);
 217         rq->timeout = 0;
 218
 219         rq->end_io = NULL;
 220         rq->end_io_data = NULL;
 221         rq->next_rq = NULL;
 222
 223         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 224 }
 225
 226 static struct request *
 227 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 228 {
 229         struct request *rq;
 230         unsigned int tag;
 231
 232         tag = blk_mq_get_tag(data);
 233         if (tag != BLK_MQ_TAG_FAIL) {
 234                 rq = data->hctx->tags->rqs[tag];
 235
 236                 if (blk_mq_tag_busy(data->hctx)) {
 237                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 238                         atomic_inc(&data->hctx->nr_active);
 239                 }
 240
 241                 rq->tag = tag;
 242                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 243                 return rq;
 244         }
 245
 246         return NULL;
 247 }
 248
 249 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 250                 bool reserved)
 251 {
 252         struct blk_mq_ctx *ctx;
 253         struct blk_mq_hw_ctx *hctx;
 254         struct request *rq;
 255         struct blk_mq_alloc_data alloc_data;
 256         int ret;
 257
 258         ret = blk_mq_queue_enter(q);
 259         if (ret)
 260                 return ERR_PTR(ret);
 261
 262         ctx = blk_mq_get_ctx(q);
 263         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 264         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 265                         reserved, ctx, hctx);
 266
 267         rq = __blk_mq_alloc_request(&alloc_data, rw);
 268         if (!rq && (gfp & __GFP_WAIT)) {
 269                 __blk_mq_run_hw_queue(hctx);
 270                 blk_mq_put_ctx(ctx);
 271
 272                 ctx = blk_mq_get_ctx(q);
 273                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 274                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 275                                 hctx);
 276                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 277                 ctx = alloc_data.ctx;
 278         }
 279         blk_mq_put_ctx(ctx);
 280         if (!rq) {
 281                 blk_mq_queue_exit(q);
 282                 return ERR_PTR(-EWOULDBLOCK);
 283         }
 284         return rq;
 285 }
 286 EXPORT_SYMBOL(blk_mq_alloc_request);
 287
 288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 289                                   struct blk_mq_ctx *ctx, struct request *rq)
 290 {
 291         const int tag = rq->tag;
 292         struct request_queue *q = rq->q;
 293
 294         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 295                 atomic_dec(&hctx->nr_active);
 296         rq->cmd_flags = 0;
 297
 298         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 299         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 300         blk_mq_queue_exit(q);
 301 }
 302
 303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 304 {
 305         struct blk_mq_ctx *ctx = rq->mq_ctx;
 306
 307         ctx->rq_completed[rq_is_sync(rq)]++;
 308         __blk_mq_free_request(hctx, ctx, rq);
 309
 310 }
 311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 312
 313 void blk_mq_free_request(struct request *rq)
 314 {
 315         struct blk_mq_hw_ctx *hctx;
 316         struct request_queue *q = rq->q;
 317
 318         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 319         blk_mq_free_hctx_request(hctx, rq);
 320 }
 321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 322
 323 inline void __blk_mq_end_request(struct request *rq, int error)
 324 {
 325         blk_account_io_done(rq);
 326
 327         if (rq->end_io) {
 328                 rq->end_io(rq, error);
 329         } else {
 330                 if (unlikely(blk_bidi_rq(rq)))
 331                         blk_mq_free_request(rq->next_rq);
 332                 blk_mq_free_request(rq);
 333         }
 334 }
 335 EXPORT_SYMBOL(__blk_mq_end_request);
 336
 337 void blk_mq_end_request(struct request *rq, int error)
 338 {
 339         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 340                 BUG();
 341         __blk_mq_end_request(rq, error);
 342 }
 343 EXPORT_SYMBOL(blk_mq_end_request);
 344
 345 static void __blk_mq_complete_request_remote(void *data)
 346 {
 347         struct request *rq = data;
 348
 349         rq->q->softirq_done_fn(rq);
 350 }
 351
 352 static void blk_mq_ipi_complete_request(struct request *rq)
 353 {
 354         struct blk_mq_ctx *ctx = rq->mq_ctx;
 355         bool shared = false;
 356         int cpu;
 357
 358         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 359                 rq->q->softirq_done_fn(rq);
 360                 return;
 361         }
 362
 363         cpu = get_cpu();
 364         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 365                 shared = cpus_share_cache(cpu, ctx->cpu);
 366
 367         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 368                 rq->csd.func = __blk_mq_complete_request_remote;
 369                 rq->csd.info = rq;
 370                 rq->csd.flags = 0;
 371                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 372         } else {
 373                 rq->q->softirq_done_fn(rq);
 374         }
 375         put_cpu();
 376 }
 377
 378 void __blk_mq_complete_request(struct request *rq)
 379 {
 380         struct request_queue *q = rq->q;
 381
 382         if (!q->softirq_done_fn)
 383                 blk_mq_end_request(rq, rq->errors);
 384         else
 385                 blk_mq_ipi_complete_request(rq);
 386 }
 387
 388 /**
 389  * blk_mq_complete_request - end I/O on a request
 390  * @rq:         the request being processed
 391  *
 392  * Description:
 393  *      Ends all I/O on a request. It does not handle partial completions.
 394  *      The actual completion happens out-of-order, through a IPI handler.
 395  **/
 396 void blk_mq_complete_request(struct request *rq)
 397 {
 398         struct request_queue *q = rq->q;
 399
 400         if (unlikely(blk_should_fake_timeout(q)))
 401                 return;
 402         if (!blk_mark_rq_complete(rq))
 403                 __blk_mq_complete_request(rq);
 404 }
 405 EXPORT_SYMBOL(blk_mq_complete_request);
 406
 407 int blk_mq_request_started(struct request *rq)
 408 {
 409         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 410 }
 411 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 412
 413 void blk_mq_start_request(struct request *rq)
 414 {
 415         struct request_queue *q = rq->q;
 416
 417         trace_block_rq_issue(q, rq);
 418
 419         rq->resid_len = blk_rq_bytes(rq);
 420         if (unlikely(blk_bidi_rq(rq)))
 421                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 422
 423         blk_add_timer(rq);
 424
 425         /*
 426          * Ensure that ->deadline is visible before set the started
 427          * flag and clear the completed flag.
 428          */
 429         smp_mb__before_atomic();
 430
 431         /*
 432          * Mark us as started and clear complete. Complete might have been
 433          * set if requeue raced with timeout, which then marked it as
 434          * complete. So be sure to clear complete again when we start
 435          * the request, otherwise we'll ignore the completion event.
 436          */
 437         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 438                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 439         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 440                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 441
 442         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 443                 /*
 444                  * Make sure space for the drain appears.  We know we can do
 445                  * this because max_hw_segments has been adjusted to be one
 446                  * fewer than the device can handle.
 447                  */
 448                 rq->nr_phys_segments++;
 449         }
 450 }
 451 EXPORT_SYMBOL(blk_mq_start_request);
 452
 453 static void __blk_mq_requeue_request(struct request *rq)
 454 {
 455         struct request_queue *q = rq->q;
 456
 457         trace_block_rq_requeue(q, rq);
 458
 459         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 460                 if (q->dma_drain_size && blk_rq_bytes(rq))
 461                         rq->nr_phys_segments--;
 462         }
 463 }
 464
 465 void blk_mq_requeue_request(struct request *rq)
 466 {
 467         __blk_mq_requeue_request(rq);
 468
 469         BUG_ON(blk_queued_rq(rq));
 470         blk_mq_add_to_requeue_list(rq, true);
 471 }
 472 EXPORT_SYMBOL(blk_mq_requeue_request);
 473
 474 static void blk_mq_requeue_work(struct work_struct *work)
 475 {
 476         struct request_queue *q =
 477                 container_of(work, struct request_queue, requeue_work);
 478         LIST_HEAD(rq_list);
 479         struct request *rq, *next;
 480         unsigned long flags;
 481
 482         spin_lock_irqsave(&q->requeue_lock, flags);
 483         list_splice_init(&q->requeue_list, &rq_list);
 484         spin_unlock_irqrestore(&q->requeue_lock, flags);
 485
 486         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 487                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 488                         continue;
 489
 490                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 491                 list_del_init(&rq->queuelist);
 492                 blk_mq_insert_request(rq, true, false, false);
 493         }
 494
 495         while (!list_empty(&rq_list)) {
 496                 rq = list_entry(rq_list.next, struct request, queuelist);
 497                 list_del_init(&rq->queuelist);
 498                 blk_mq_insert_request(rq, false, false, false);
 499         }
 500
 501         /*
 502          * Use the start variant of queue running here, so that running
 503          * the requeue work will kick stopped queues.
 504          */
 505         blk_mq_start_hw_queues(q);
 506 }
 507
 508 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 509 {
 510         struct request_queue *q = rq->q;
 511         unsigned long flags;
 512
 513         /*
 514          * We abuse this flag that is otherwise used by the I/O scheduler to
 515          * request head insertation from the workqueue.
 516          */
 517         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 518
 519         spin_lock_irqsave(&q->requeue_lock, flags);
 520         if (at_head) {
 521                 rq->cmd_flags |= REQ_SOFTBARRIER;
 522                 list_add(&rq->queuelist, &q->requeue_list);
 523         } else {
 524                 list_add_tail(&rq->queuelist, &q->requeue_list);
 525         }
 526         spin_unlock_irqrestore(&q->requeue_lock, flags);
 527 }
 528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 529
 530 void blk_mq_cancel_requeue_work(struct request_queue *q)
 531 {
 532         cancel_work_sync(&q->requeue_work);
 533 }
 534 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 535
 536 void blk_mq_kick_requeue_list(struct request_queue *q)
 537 {
 538         kblockd_schedule_work(&q->requeue_work);
 539 }
 540 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 541
 542 void blk_mq_abort_requeue_list(struct request_queue *q)
 543 {
 544         unsigned long flags;
 545         LIST_HEAD(rq_list);
 546
 547         spin_lock_irqsave(&q->requeue_lock, flags);
 548         list_splice_init(&q->requeue_list, &rq_list);
 549         spin_unlock_irqrestore(&q->requeue_lock, flags);
 550
 551         while (!list_empty(&rq_list)) {
 552                 struct request *rq;
 553
 554                 rq = list_first_entry(&rq_list, struct request, queuelist);
 555                 list_del_init(&rq->queuelist);
 556                 rq->errors = -EIO;
 557                 blk_mq_end_request(rq, rq->errors);
 558         }
 559 }
 560 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 561
 562 static inline bool is_flush_request(struct request *rq,
 563                 struct blk_flush_queue *fq, unsigned int tag)
 564 {
 565         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 566                         fq->flush_rq->tag == tag);
 567 }
 568
 569 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 570 {
 571         struct request *rq = tags->rqs[tag];
 572         /* mq_ctx of flush rq is always cloned from the corresponding req */
 573         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 574
 575         if (!is_flush_request(rq, fq, tag))
 576                 return rq;
 577
 578         return fq->flush_rq;
 579 }
 580 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 581
 582 struct blk_mq_timeout_data {
 583         unsigned long next;
 584         unsigned int next_set;
 585 };
 586
 587 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 588 {
 589         struct blk_mq_ops *ops = req->q->mq_ops;
 590         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 591
 592         /*
 593          * We know that complete is set at this point. If STARTED isn't set
 594          * anymore, then the request isn't active and the "timeout" should
 595          * just be ignored. This can happen due to the bitflag ordering.
 596          * Timeout first checks if STARTED is set, and if it is, assumes
 597          * the request is active. But if we race with completion, then
 598          * we both flags will get cleared. So check here again, and ignore
 599          * a timeout event with a request that isn't active.
 600          */
 601         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 602                 return;
 603
 604         if (ops->timeout)
 605                 ret = ops->timeout(req, reserved);
 606
 607         switch (ret) {
 608         case BLK_EH_HANDLED:
 609                 __blk_mq_complete_request(req);
 610                 break;
 611         case BLK_EH_RESET_TIMER:
 612                 blk_add_timer(req);
 613                 blk_clear_rq_complete(req);
 614                 break;
 615         case BLK_EH_NOT_HANDLED:
 616                 break;
 617         default:
 618                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 619                 break;
 620         }
 621 }
 622
 623 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 624                 struct request *rq, void *priv, bool reserved)
 625 {
 626         struct blk_mq_timeout_data *data = priv;
 627
 628         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 629                 return;
 630         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 631                 return;
 632
 633         if (time_after_eq(jiffies, rq->deadline)) {
 634                 if (!blk_mark_rq_complete(rq))
 635                         blk_mq_rq_timed_out(rq, reserved);
 636         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 637                 data->next = rq->deadline;
 638                 data->next_set = 1;
 639         }
 640 }
 641
 642 static void blk_mq_rq_timer(unsigned long priv)
 643 {
 644         struct request_queue *q = (struct request_queue *)priv;
 645         struct blk_mq_timeout_data data = {
 646                 .next           = 0,
 647                 .next_set       = 0,
 648         };
 649         struct blk_mq_hw_ctx *hctx;
 650         int i;
 651
 652         queue_for_each_hw_ctx(q, hctx, i) {
 653                 /*
 654                  * If not software queues are currently mapped to this
 655                  * hardware queue, there's nothing to check
 656                  */
 657                 if (!blk_mq_hw_queue_mapped(hctx))
 658                         continue;
 659
 660                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 661         }
 662
 663         if (data.next_set) {
 664                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 665                 mod_timer(&q->timeout, data.next);
 666         } else {
 667                 queue_for_each_hw_ctx(q, hctx, i)
 668                         blk_mq_tag_idle(hctx);
 669         }
 670 }
 671
 672 /*
 673  * Reverse check our software queue for entries that we could potentially
 674  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 675  * too much time checking for merges.
 676  */
 677 static bool blk_mq_attempt_merge(struct request_queue *q,
 678                                  struct blk_mq_ctx *ctx, struct bio *bio)
 679 {
 680         struct request *rq;
 681         int checked = 8;
 682
 683         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 684                 int el_ret;
 685
 686                 if (!checked--)
 687                         break;
 688
 689                 if (!blk_rq_merge_ok(rq, bio))
 690                         continue;
 691
 692                 el_ret = blk_try_merge(rq, bio);
 693                 if (el_ret == ELEVATOR_BACK_MERGE) {
 694                         if (bio_attempt_back_merge(q, rq, bio)) {
 695                                 ctx->rq_merged++;
 696                                 return true;
 697                         }
 698                         break;
 699                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 700                         if (bio_attempt_front_merge(q, rq, bio)) {
 701                                 ctx->rq_merged++;
 702                                 return true;
 703                         }
 704                         break;
 705                 }
 706         }
 707
 708         return false;
 709 }
 710
 711 /*
 712  * Process software queues that have been marked busy, splicing them
 713  * to the for-dispatch
 714  */
 715 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 716 {
 717         struct blk_mq_ctx *ctx;
 718         int i;
 719
 720         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 721                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 722                 unsigned int off, bit;
 723
 724                 if (!bm->word)
 725                         continue;
 726
 727                 bit = 0;
 728                 off = i * hctx->ctx_map.bits_per_word;
 729                 do {
 730                         bit = find_next_bit(&bm->word, bm->depth, bit);
 731                         if (bit >= bm->depth)
 732                                 break;
 733
 734                         ctx = hctx->ctxs[bit + off];
 735                         clear_bit(bit, &bm->word);
 736                         spin_lock(&ctx->lock);
 737                         list_splice_tail_init(&ctx->rq_list, list);
 738                         spin_unlock(&ctx->lock);
 739
 740                         bit++;
 741                 } while (1);
 742         }
 743 }
 744
 745 /*
 746  * Run this hardware queue, pulling any software queues mapped to it in.
 747  * Note that this function currently has various problems around ordering
 748  * of IO. In particular, we'd like FIFO behaviour on handling existing
 749  * items on the hctx->dispatch list. Ignore that for now.
 750  */
 751 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 752 {
 753         struct request_queue *q = hctx->queue;
 754         struct request *rq;
 755         LIST_HEAD(rq_list);
 756         LIST_HEAD(driver_list);
 757         struct list_head *dptr;
 758         int queued;
 759
 760         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 761
 762         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 763                 return;
 764
 765         hctx->run++;
 766
 767         /*
 768          * Touch any software queue that has pending entries.
 769          */
 770         flush_busy_ctxs(hctx, &rq_list);
 771
 772         /*
 773          * If we have previous entries on our dispatch list, grab them
 774          * and stuff them at the front for more fair dispatch.
 775          */
 776         if (!list_empty_careful(&hctx->dispatch)) {
 777                 spin_lock(&hctx->lock);
 778                 if (!list_empty(&hctx->dispatch))
 779                         list_splice_init(&hctx->dispatch, &rq_list);
 780                 spin_unlock(&hctx->lock);
 781         }
 782
 783         /*
 784          * Start off with dptr being NULL, so we start the first request
 785          * immediately, even if we have more pending.
 786          */
 787         dptr = NULL;
 788
 789         /*
 790          * Now process all the entries, sending them to the driver.
 791          */
 792         queued = 0;
 793         while (!list_empty(&rq_list)) {
 794                 struct blk_mq_queue_data bd;
 795                 int ret;
 796
 797                 rq = list_first_entry(&rq_list, struct request, queuelist);
 798                 list_del_init(&rq->queuelist);
 799
 800                 bd.rq = rq;
 801                 bd.list = dptr;
 802                 bd.last = list_empty(&rq_list);
 803
 804                 ret = q->mq_ops->queue_rq(hctx, &bd);
 805                 switch (ret) {
 806                 case BLK_MQ_RQ_QUEUE_OK:
 807                         queued++;
 808                         continue;
 809                 case BLK_MQ_RQ_QUEUE_BUSY:
 810                         list_add(&rq->queuelist, &rq_list);
 811                         __blk_mq_requeue_request(rq);
 812                         break;
 813                 default:
 814                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 815                 case BLK_MQ_RQ_QUEUE_ERROR:
 816                         rq->errors = -EIO;
 817                         blk_mq_end_request(rq, rq->errors);
 818                         break;
 819                 }
 820
 821                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 822                         break;
 823
 824                 /*
 825                  * We've done the first request. If we have more than 1
 826                  * left in the list, set dptr to defer issue.
 827                  */
 828                 if (!dptr && rq_list.next != rq_list.prev)
 829                         dptr = &driver_list;
 830         }
 831
 832         if (!queued)
 833                 hctx->dispatched[0]++;
 834         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 835                 hctx->dispatched[ilog2(queued) + 1]++;
 836
 837         /*
 838          * Any items that need requeuing? Stuff them into hctx->dispatch,
 839          * that is where we will continue on next queue run.
 840          */
 841         if (!list_empty(&rq_list)) {
 842                 spin_lock(&hctx->lock);
 843                 list_splice(&rq_list, &hctx->dispatch);
 844                 spin_unlock(&hctx->lock);
 845         }
 846 }
 847
 848 /*
 849  * It'd be great if the workqueue API had a way to pass
 850  * in a mask and had some smarts for more clever placement.
 851  * For now we just round-robin here, switching for every
 852  * BLK_MQ_CPU_WORK_BATCH queued items.
 853  */
 854 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 855 {
 856         if (hctx->queue->nr_hw_queues == 1)
 857                 return WORK_CPU_UNBOUND;
 858
 859         if (--hctx->next_cpu_batch <= 0) {
 860                 int cpu = hctx->next_cpu, next_cpu;
 861
 862                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 863                 if (next_cpu >= nr_cpu_ids)
 864                         next_cpu = cpumask_first(hctx->cpumask);
 865
 866                 hctx->next_cpu = next_cpu;
 867                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 868
 869                 return cpu;
 870         }
 871
 872         return hctx->next_cpu;
 873 }
 874
 875 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 876 {
 877         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 878             !blk_mq_hw_queue_mapped(hctx)))
 879                 return;
 880
 881         if (!async) {
 882                 int cpu = get_cpu();
 883                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 884                         __blk_mq_run_hw_queue(hctx);
 885                         put_cpu();
 886                         return;
 887                 }
 888
 889                 put_cpu();
 890         }
 891
 892         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 893                         &hctx->run_work, 0);
 894 }
 895
 896 void blk_mq_run_queues(struct request_queue *q, bool async)
 897 {
 898         struct blk_mq_hw_ctx *hctx;
 899         int i;
 900
 901         queue_for_each_hw_ctx(q, hctx, i) {
 902                 if ((!blk_mq_hctx_has_pending(hctx) &&
 903                     list_empty_careful(&hctx->dispatch)) ||
 904                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 905                         continue;
 906
 907                 blk_mq_run_hw_queue(hctx, async);
 908         }
 909 }
 910 EXPORT_SYMBOL(blk_mq_run_queues);
 911
 912 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 913 {
 914         cancel_delayed_work(&hctx->run_work);
 915         cancel_delayed_work(&hctx->delay_work);
 916         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 917 }
 918 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 919
 920 void blk_mq_stop_hw_queues(struct request_queue *q)
 921 {
 922         struct blk_mq_hw_ctx *hctx;
 923         int i;
 924
 925         queue_for_each_hw_ctx(q, hctx, i)
 926                 blk_mq_stop_hw_queue(hctx);
 927 }
 928 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 929
 930 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 931 {
 932         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 933
 934         blk_mq_run_hw_queue(hctx, false);
 935 }
 936 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 937
 938 void blk_mq_start_hw_queues(struct request_queue *q)
 939 {
 940         struct blk_mq_hw_ctx *hctx;
 941         int i;
 942
 943         queue_for_each_hw_ctx(q, hctx, i)
 944                 blk_mq_start_hw_queue(hctx);
 945 }
 946 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 947
 948
 949 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 950 {
 951         struct blk_mq_hw_ctx *hctx;
 952         int i;
 953
 954         queue_for_each_hw_ctx(q, hctx, i) {
 955                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 956                         continue;
 957
 958                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 959                 blk_mq_run_hw_queue(hctx, async);
 960         }
 961 }
 962 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 963
 964 static void blk_mq_run_work_fn(struct work_struct *work)
 965 {
 966         struct blk_mq_hw_ctx *hctx;
 967
 968         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 969
 970         __blk_mq_run_hw_queue(hctx);
 971 }
 972
 973 static void blk_mq_delay_work_fn(struct work_struct *work)
 974 {
 975         struct blk_mq_hw_ctx *hctx;
 976
 977         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 978
 979         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 980                 __blk_mq_run_hw_queue(hctx);
 981 }
 982
 983 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 984 {
 985         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
 986                 return;
 987
 988         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 989                         &hctx->delay_work, msecs_to_jiffies(msecs));
 990 }
 991 EXPORT_SYMBOL(blk_mq_delay_queue);
 992
 993 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 994                                     struct request *rq, bool at_head)
 995 {
 996         struct blk_mq_ctx *ctx = rq->mq_ctx;
 997
 998         trace_block_rq_insert(hctx->queue, rq);
 999
1000         if (at_head)
1001                 list_add(&rq->queuelist, &ctx->rq_list);
1002         else
1003                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1004
1005         blk_mq_hctx_mark_pending(hctx, ctx);
1006 }
1007
1008 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1009                 bool async)
1010 {
1011         struct request_queue *q = rq->q;
1012         struct blk_mq_hw_ctx *hctx;
1013         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1014
1015         current_ctx = blk_mq_get_ctx(q);
1016         if (!cpu_online(ctx->cpu))
1017                 rq->mq_ctx = ctx = current_ctx;
1018
1019         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1020
1021         spin_lock(&ctx->lock);
1022         __blk_mq_insert_request(hctx, rq, at_head);
1023         spin_unlock(&ctx->lock);
1024
1025         if (run_queue)
1026                 blk_mq_run_hw_queue(hctx, async);
1027
1028         blk_mq_put_ctx(current_ctx);
1029 }
1030
1031 static void blk_mq_insert_requests(struct request_queue *q,
1032                                      struct blk_mq_ctx *ctx,
1033                                      struct list_head *list,
1034                                      int depth,
1035                                      bool from_schedule)
1036
1037 {
1038         struct blk_mq_hw_ctx *hctx;
1039         struct blk_mq_ctx *current_ctx;
1040
1041         trace_block_unplug(q, depth, !from_schedule);
1042
1043         current_ctx = blk_mq_get_ctx(q);
1044
1045         if (!cpu_online(ctx->cpu))
1046                 ctx = current_ctx;
1047         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1048
1049         /*
1050          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1051          * offline now
1052          */
1053         spin_lock(&ctx->lock);
1054         while (!list_empty(list)) {
1055                 struct request *rq;
1056
1057                 rq = list_first_entry(list, struct request, queuelist);
1058                 list_del_init(&rq->queuelist);
1059                 rq->mq_ctx = ctx;
1060                 __blk_mq_insert_request(hctx, rq, false);
1061         }
1062         spin_unlock(&ctx->lock);
1063
1064         blk_mq_run_hw_queue(hctx, from_schedule);
1065         blk_mq_put_ctx(current_ctx);
1066 }
1067
1068 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1069 {
1070         struct request *rqa = container_of(a, struct request, queuelist);
1071         struct request *rqb = container_of(b, struct request, queuelist);
1072
1073         return !(rqa->mq_ctx < rqb->mq_ctx ||
1074                  (rqa->mq_ctx == rqb->mq_ctx &&
1075                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1076 }
1077
1078 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1079 {
1080         struct blk_mq_ctx *this_ctx;
1081         struct request_queue *this_q;
1082         struct request *rq;
1083         LIST_HEAD(list);
1084         LIST_HEAD(ctx_list);
1085         unsigned int depth;
1086
1087         list_splice_init(&plug->mq_list, &list);
1088
1089         list_sort(NULL, &list, plug_ctx_cmp);
1090
1091         this_q = NULL;
1092         this_ctx = NULL;
1093         depth = 0;
1094
1095         while (!list_empty(&list)) {
1096                 rq = list_entry_rq(list.next);
1097                 list_del_init(&rq->queuelist);
1098                 BUG_ON(!rq->q);
1099                 if (rq->mq_ctx != this_ctx) {
1100                         if (this_ctx) {
1101                                 blk_mq_insert_requests(this_q, this_ctx,
1102                                                         &ctx_list, depth,
1103                                                         from_schedule);
1104                         }
1105
1106                         this_ctx = rq->mq_ctx;
1107                         this_q = rq->q;
1108                         depth = 0;
1109                 }
1110
1111                 depth++;
1112                 list_add_tail(&rq->queuelist, &ctx_list);
1113         }
1114
1115         /*
1116          * If 'this_ctx' is set, we know we have entries to complete
1117          * on 'ctx_list'. Do those.
1118          */
1119         if (this_ctx) {
1120                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1121                                        from_schedule);
1122         }
1123 }
1124
1125 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1126 {
1127         init_request_from_bio(rq, bio);
1128
1129         if (blk_do_io_stat(rq))
1130                 blk_account_io_start(rq, 1);
1131 }
1132
1133 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1134 {
1135         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1136                 !blk_queue_nomerges(hctx->queue);
1137 }
1138
1139 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1140                                          struct blk_mq_ctx *ctx,
1141                                          struct request *rq, struct bio *bio)
1142 {
1143         if (!hctx_allow_merges(hctx)) {
1144                 blk_mq_bio_to_request(rq, bio);
1145                 spin_lock(&ctx->lock);
1146 insert_rq:
1147                 __blk_mq_insert_request(hctx, rq, false);
1148                 spin_unlock(&ctx->lock);
1149                 return false;
1150         } else {
1151                 struct request_queue *q = hctx->queue;
1152
1153                 spin_lock(&ctx->lock);
1154                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1155                         blk_mq_bio_to_request(rq, bio);
1156                         goto insert_rq;
1157                 }
1158
1159                 spin_unlock(&ctx->lock);
1160                 __blk_mq_free_request(hctx, ctx, rq);
1161                 return true;
1162         }
1163 }
1164
1165 struct blk_map_ctx {
1166         struct blk_mq_hw_ctx *hctx;
1167         struct blk_mq_ctx *ctx;
1168 };
1169
1170 static struct request *blk_mq_map_request(struct request_queue *q,
1171                                           struct bio *bio,
1172                                           struct blk_map_ctx *data)
1173 {
1174         struct blk_mq_hw_ctx *hctx;
1175         struct blk_mq_ctx *ctx;
1176         struct request *rq;
1177         int rw = bio_data_dir(bio);
1178         struct blk_mq_alloc_data alloc_data;
1179
1180         if (unlikely(blk_mq_queue_enter(q))) {
1181                 bio_endio(bio, -EIO);
1182                 return NULL;
1183         }
1184
1185         ctx = blk_mq_get_ctx(q);
1186         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1187
1188         if (rw_is_sync(bio->bi_rw))
1189                 rw |= REQ_SYNC;
1190
1191         trace_block_getrq(q, bio, rw);
1192         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1193                         hctx);
1194         rq = __blk_mq_alloc_request(&alloc_data, rw);
1195         if (unlikely(!rq)) {
1196                 __blk_mq_run_hw_queue(hctx);
1197                 blk_mq_put_ctx(ctx);
1198                 trace_block_sleeprq(q, bio, rw);
1199
1200                 ctx = blk_mq_get_ctx(q);
1201                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1202                 blk_mq_set_alloc_data(&alloc_data, q,
1203                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1204                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1205                 ctx = alloc_data.ctx;
1206                 hctx = alloc_data.hctx;
1207         }
1208
1209         hctx->queued++;
1210         data->hctx = hctx;
1211         data->ctx = ctx;
1212         return rq;
1213 }
1214
1215 /*
1216  * Multiple hardware queue variant. This will not use per-process plugs,
1217  * but will attempt to bypass the hctx queueing if we can go straight to
1218  * hardware for SYNC IO.
1219  */
1220 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1221 {
1222         const int is_sync = rw_is_sync(bio->bi_rw);
1223         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1224         struct blk_map_ctx data;
1225         struct request *rq;
1226
1227         blk_queue_bounce(q, &bio);
1228
1229         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1230                 bio_endio(bio, -EIO);
1231                 return;
1232         }
1233
1234         rq = blk_mq_map_request(q, bio, &data);
1235         if (unlikely(!rq))
1236                 return;
1237
1238         if (unlikely(is_flush_fua)) {
1239                 blk_mq_bio_to_request(rq, bio);
1240                 blk_insert_flush(rq);
1241                 goto run_queue;
1242         }
1243
1244         /*
1245          * If the driver supports defer issued based on 'last', then
1246          * queue it up like normal since we can potentially save some
1247          * CPU this way.
1248          */
1249         if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1250                 struct blk_mq_queue_data bd = {
1251                         .rq = rq,
1252                         .list = NULL,
1253                         .last = 1
1254                 };
1255                 int ret;
1256
1257                 blk_mq_bio_to_request(rq, bio);
1258
1259                 /*
1260                  * For OK queue, we are done. For error, kill it. Any other
1261                  * error (busy), just add it to our list as we previously
1262                  * would have done
1263                  */
1264                 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1265                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1266                         goto done;
1267                 else {
1268                         __blk_mq_requeue_request(rq);
1269
1270                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1271                                 rq->errors = -EIO;
1272                                 blk_mq_end_request(rq, rq->errors);
1273                                 goto done;
1274                         }
1275                 }
1276         }
1277
1278         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1279                 /*
1280                  * For a SYNC request, send it to the hardware immediately. For
1281                  * an ASYNC request, just ensure that we run it later on. The
1282                  * latter allows for merging opportunities and more efficient
1283                  * dispatching.
1284                  */
1285 run_queue:
1286                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1287         }
1288 done:
1289         blk_mq_put_ctx(data.ctx);
1290 }
1291
1292 /*
1293  * Single hardware queue variant. This will attempt to use any per-process
1294  * plug for merging and IO deferral.
1295  */
1296 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1297 {
1298         const int is_sync = rw_is_sync(bio->bi_rw);
1299         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1300         unsigned int use_plug, request_count = 0;
1301         struct blk_map_ctx data;
1302         struct request *rq;
1303
1304         /*
1305          * If we have multiple hardware queues, just go directly to
1306          * one of those for sync IO.
1307          */
1308         use_plug = !is_flush_fua && !is_sync;
1309
1310         blk_queue_bounce(q, &bio);
1311
1312         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1313                 bio_endio(bio, -EIO);
1314                 return;
1315         }
1316
1317         if (use_plug && !blk_queue_nomerges(q) &&
1318             blk_attempt_plug_merge(q, bio, &request_count))
1319                 return;
1320
1321         rq = blk_mq_map_request(q, bio, &data);
1322         if (unlikely(!rq))
1323                 return;
1324
1325         if (unlikely(is_flush_fua)) {
1326                 blk_mq_bio_to_request(rq, bio);
1327                 blk_insert_flush(rq);
1328                 goto run_queue;
1329         }
1330
1331         /*
1332          * A task plug currently exists. Since this is completely lockless,
1333          * utilize that to temporarily store requests until the task is
1334          * either done or scheduled away.
1335          */
1336         if (use_plug) {
1337                 struct blk_plug *plug = current->plug;
1338
1339                 if (plug) {
1340                         blk_mq_bio_to_request(rq, bio);
1341                         if (list_empty(&plug->mq_list))
1342                                 trace_block_plug(q);
1343                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1344                                 blk_flush_plug_list(plug, false);
1345                                 trace_block_plug(q);
1346                         }
1347                         list_add_tail(&rq->queuelist, &plug->mq_list);
1348                         blk_mq_put_ctx(data.ctx);
1349                         return;
1350                 }
1351         }
1352
1353         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1354                 /*
1355                  * For a SYNC request, send it to the hardware immediately. For
1356                  * an ASYNC request, just ensure that we run it later on. The
1357                  * latter allows for merging opportunities and more efficient
1358                  * dispatching.
1359                  */
1360 run_queue:
1361                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1362         }
1363
1364         blk_mq_put_ctx(data.ctx);
1365 }
1366
1367 /*
1368  * Default mapping to a software queue, since we use one per CPU.
1369  */
1370 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1371 {
1372         return q->queue_hw_ctx[q->mq_map[cpu]];
1373 }
1374 EXPORT_SYMBOL(blk_mq_map_queue);
1375
1376 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1377                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1378 {
1379         struct page *page;
1380
1381         if (tags->rqs && set->ops->exit_request) {
1382                 int i;
1383
1384                 for (i = 0; i < tags->nr_tags; i++) {
1385                         if (!tags->rqs[i])
1386                                 continue;
1387                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1388                                                 hctx_idx, i);
1389                         tags->rqs[i] = NULL;
1390                 }
1391         }
1392
1393         while (!list_empty(&tags->page_list)) {
1394                 page = list_first_entry(&tags->page_list, struct page, lru);
1395                 list_del_init(&page->lru);
1396                 __free_pages(page, page->private);
1397         }
1398
1399         kfree(tags->rqs);
1400
1401         blk_mq_free_tags(tags);
1402 }
1403
1404 static size_t order_to_size(unsigned int order)
1405 {
1406         return (size_t)PAGE_SIZE << order;
1407 }
1408
1409 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1410                 unsigned int hctx_idx)
1411 {
1412         struct blk_mq_tags *tags;
1413         unsigned int i, j, entries_per_page, max_order = 4;
1414         size_t rq_size, left;
1415
1416         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1417                                 set->numa_node);
1418         if (!tags)
1419                 return NULL;
1420
1421         INIT_LIST_HEAD(&tags->page_list);
1422
1423         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1424                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1425                                  set->numa_node);
1426         if (!tags->rqs) {
1427                 blk_mq_free_tags(tags);
1428                 return NULL;
1429         }
1430
1431         /*
1432          * rq_size is the size of the request plus driver payload, rounded
1433          * to the cacheline size
1434          */
1435         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1436                                 cache_line_size());
1437         left = rq_size * set->queue_depth;
1438
1439         for (i = 0; i < set->queue_depth; ) {
1440                 int this_order = max_order;
1441                 struct page *page;
1442                 int to_do;
1443                 void *p;
1444
1445                 while (left < order_to_size(this_order - 1) && this_order)
1446                         this_order--;
1447
1448                 do {
1449                         page = alloc_pages_node(set->numa_node,
1450                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1451                                 this_order);
1452                         if (page)
1453                                 break;
1454                         if (!this_order--)
1455                                 break;
1456                         if (order_to_size(this_order) < rq_size)
1457                                 break;
1458                 } while (1);
1459
1460                 if (!page)
1461                         goto fail;
1462
1463                 page->private = this_order;
1464                 list_add_tail(&page->lru, &tags->page_list);
1465
1466                 p = page_address(page);
1467                 entries_per_page = order_to_size(this_order) / rq_size;
1468                 to_do = min(entries_per_page, set->queue_depth - i);
1469                 left -= to_do * rq_size;
1470                 for (j = 0; j < to_do; j++) {
1471                         tags->rqs[i] = p;
1472                         tags->rqs[i]->atomic_flags = 0;
1473                         tags->rqs[i]->cmd_flags = 0;
1474                         if (set->ops->init_request) {
1475                                 if (set->ops->init_request(set->driver_data,
1476                                                 tags->rqs[i], hctx_idx, i,
1477                                                 set->numa_node)) {
1478                                         tags->rqs[i] = NULL;
1479                                         goto fail;
1480                                 }
1481                         }
1482
1483                         p += rq_size;
1484                         i++;
1485                 }
1486         }
1487
1488         return tags;
1489
1490 fail:
1491         blk_mq_free_rq_map(set, tags, hctx_idx);
1492         return NULL;
1493 }
1494
1495 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1496 {
1497         kfree(bitmap->map);
1498 }
1499
1500 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1501 {
1502         unsigned int bpw = 8, total, num_maps, i;
1503
1504         bitmap->bits_per_word = bpw;
1505
1506         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1507         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1508                                         GFP_KERNEL, node);
1509         if (!bitmap->map)
1510                 return -ENOMEM;
1511
1512         bitmap->map_size = num_maps;
1513
1514         total = nr_cpu_ids;
1515         for (i = 0; i < num_maps; i++) {
1516                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1517                 total -= bitmap->map[i].depth;
1518         }
1519
1520         return 0;
1521 }
1522
1523 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1524 {
1525         struct request_queue *q = hctx->queue;
1526         struct blk_mq_ctx *ctx;
1527         LIST_HEAD(tmp);
1528
1529         /*
1530          * Move ctx entries to new CPU, if this one is going away.
1531          */
1532         ctx = __blk_mq_get_ctx(q, cpu);
1533
1534         spin_lock(&ctx->lock);
1535         if (!list_empty(&ctx->rq_list)) {
1536                 list_splice_init(&ctx->rq_list, &tmp);
1537                 blk_mq_hctx_clear_pending(hctx, ctx);
1538         }
1539         spin_unlock(&ctx->lock);
1540
1541         if (list_empty(&tmp))
1542                 return NOTIFY_OK;
1543
1544         ctx = blk_mq_get_ctx(q);
1545         spin_lock(&ctx->lock);
1546
1547         while (!list_empty(&tmp)) {
1548                 struct request *rq;
1549
1550                 rq = list_first_entry(&tmp, struct request, queuelist);
1551                 rq->mq_ctx = ctx;
1552                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1553         }
1554
1555         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1556         blk_mq_hctx_mark_pending(hctx, ctx);
1557
1558         spin_unlock(&ctx->lock);
1559
1560         blk_mq_run_hw_queue(hctx, true);
1561         blk_mq_put_ctx(ctx);
1562         return NOTIFY_OK;
1563 }
1564
1565 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1566 {
1567         struct request_queue *q = hctx->queue;
1568         struct blk_mq_tag_set *set = q->tag_set;
1569
1570         if (set->tags[hctx->queue_num])
1571                 return NOTIFY_OK;
1572
1573         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1574         if (!set->tags[hctx->queue_num])
1575                 return NOTIFY_STOP;
1576
1577         hctx->tags = set->tags[hctx->queue_num];
1578         return NOTIFY_OK;
1579 }
1580
1581 static int blk_mq_hctx_notify(void *data, unsigned long action,
1582                               unsigned int cpu)
1583 {
1584         struct blk_mq_hw_ctx *hctx = data;
1585
1586         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1587                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1588         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1589                 return blk_mq_hctx_cpu_online(hctx, cpu);
1590
1591         return NOTIFY_OK;
1592 }
1593
1594 static void blk_mq_exit_hctx(struct request_queue *q,
1595                 struct blk_mq_tag_set *set,
1596                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1597 {
1598         unsigned flush_start_tag = set->queue_depth;
1599
1600         blk_mq_tag_idle(hctx);
1601
1602         if (set->ops->exit_request)
1603                 set->ops->exit_request(set->driver_data,
1604                                        hctx->fq->flush_rq, hctx_idx,
1605                                        flush_start_tag + hctx_idx);
1606
1607         if (set->ops->exit_hctx)
1608                 set->ops->exit_hctx(hctx, hctx_idx);
1609
1610         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1611         blk_free_flush_queue(hctx->fq);
1612         kfree(hctx->ctxs);
1613         blk_mq_free_bitmap(&hctx->ctx_map);
1614 }
1615
1616 static void blk_mq_exit_hw_queues(struct request_queue *q,
1617                 struct blk_mq_tag_set *set, int nr_queue)
1618 {
1619         struct blk_mq_hw_ctx *hctx;
1620         unsigned int i;
1621
1622         queue_for_each_hw_ctx(q, hctx, i) {
1623                 if (i == nr_queue)
1624                         break;
1625                 blk_mq_exit_hctx(q, set, hctx, i);
1626         }
1627 }
1628
1629 static void blk_mq_free_hw_queues(struct request_queue *q,
1630                 struct blk_mq_tag_set *set)
1631 {
1632         struct blk_mq_hw_ctx *hctx;
1633         unsigned int i;
1634
1635         queue_for_each_hw_ctx(q, hctx, i) {
1636                 free_cpumask_var(hctx->cpumask);
1637                 kfree(hctx);
1638         }
1639 }
1640
1641 static int blk_mq_init_hctx(struct request_queue *q,
1642                 struct blk_mq_tag_set *set,
1643                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1644 {
1645         int node;
1646         unsigned flush_start_tag = set->queue_depth;
1647
1648         node = hctx->numa_node;
1649         if (node == NUMA_NO_NODE)
1650                 node = hctx->numa_node = set->numa_node;
1651
1652         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1653         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1654         spin_lock_init(&hctx->lock);
1655         INIT_LIST_HEAD(&hctx->dispatch);
1656         hctx->queue = q;
1657         hctx->queue_num = hctx_idx;
1658         hctx->flags = set->flags;
1659
1660         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1661                                         blk_mq_hctx_notify, hctx);
1662         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1663
1664         hctx->tags = set->tags[hctx_idx];
1665
1666         /*
1667          * Allocate space for all possible cpus to avoid allocation at
1668          * runtime
1669          */
1670         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1671                                         GFP_KERNEL, node);
1672         if (!hctx->ctxs)
1673                 goto unregister_cpu_notifier;
1674
1675         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1676                 goto free_ctxs;
1677
1678         hctx->nr_ctx = 0;
1679
1680         if (set->ops->init_hctx &&
1681             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1682                 goto free_bitmap;
1683
1684         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1685         if (!hctx->fq)
1686                 goto exit_hctx;
1687
1688         if (set->ops->init_request &&
1689             set->ops->init_request(set->driver_data,
1690                                    hctx->fq->flush_rq, hctx_idx,
1691                                    flush_start_tag + hctx_idx, node))
1692                 goto free_fq;
1693
1694         return 0;
1695
1696  free_fq:
1697         kfree(hctx->fq);
1698  exit_hctx:
1699         if (set->ops->exit_hctx)
1700                 set->ops->exit_hctx(hctx, hctx_idx);
1701  free_bitmap:
1702         blk_mq_free_bitmap(&hctx->ctx_map);
1703  free_ctxs:
1704         kfree(hctx->ctxs);
1705  unregister_cpu_notifier:
1706         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1707
1708         return -1;
1709 }
1710
1711 static int blk_mq_init_hw_queues(struct request_queue *q,
1712                 struct blk_mq_tag_set *set)
1713 {
1714         struct blk_mq_hw_ctx *hctx;
1715         unsigned int i;
1716
1717         /*
1718          * Initialize hardware queues
1719          */
1720         queue_for_each_hw_ctx(q, hctx, i) {
1721                 if (blk_mq_init_hctx(q, set, hctx, i))
1722                         break;
1723         }
1724
1725         if (i == q->nr_hw_queues)
1726                 return 0;
1727
1728         /*
1729          * Init failed
1730          */
1731         blk_mq_exit_hw_queues(q, set, i);
1732
1733         return 1;
1734 }
1735
1736 static void blk_mq_init_cpu_queues(struct request_queue *q,
1737                                    unsigned int nr_hw_queues)
1738 {
1739         unsigned int i;
1740
1741         for_each_possible_cpu(i) {
1742                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1743                 struct blk_mq_hw_ctx *hctx;
1744
1745                 memset(__ctx, 0, sizeof(*__ctx));
1746                 __ctx->cpu = i;
1747                 spin_lock_init(&__ctx->lock);
1748                 INIT_LIST_HEAD(&__ctx->rq_list);
1749                 __ctx->queue = q;
1750
1751                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1752                 if (!cpu_online(i))
1753                         continue;
1754
1755                 hctx = q->mq_ops->map_queue(q, i);
1756                 cpumask_set_cpu(i, hctx->cpumask);
1757                 hctx->nr_ctx++;
1758
1759                 /*
1760                  * Set local node, IFF we have more than one hw queue. If
1761                  * not, we remain on the home node of the device
1762                  */
1763                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1764                         hctx->numa_node = cpu_to_node(i);
1765         }
1766 }
1767
1768 static void blk_mq_map_swqueue(struct request_queue *q)
1769 {
1770         unsigned int i;
1771         struct blk_mq_hw_ctx *hctx;
1772         struct blk_mq_ctx *ctx;
1773
1774         queue_for_each_hw_ctx(q, hctx, i) {
1775                 cpumask_clear(hctx->cpumask);
1776                 hctx->nr_ctx = 0;
1777         }
1778
1779         /*
1780          * Map software to hardware queues
1781          */
1782         queue_for_each_ctx(q, ctx, i) {
1783                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1784                 if (!cpu_online(i))
1785                         continue;
1786
1787                 hctx = q->mq_ops->map_queue(q, i);
1788                 cpumask_set_cpu(i, hctx->cpumask);
1789                 ctx->index_hw = hctx->nr_ctx;
1790                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1791         }
1792
1793         queue_for_each_hw_ctx(q, hctx, i) {
1794                 /*
1795                  * If no software queues are mapped to this hardware queue,
1796                  * disable it and free the request entries.
1797                  */
1798                 if (!hctx->nr_ctx) {
1799                         struct blk_mq_tag_set *set = q->tag_set;
1800
1801                         if (set->tags[i]) {
1802                                 blk_mq_free_rq_map(set, set->tags[i], i);
1803                                 set->tags[i] = NULL;
1804                                 hctx->tags = NULL;
1805                         }
1806                         continue;
1807                 }
1808
1809                 /*
1810                  * Initialize batch roundrobin counts
1811                  */
1812                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1813                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1814         }
1815 }
1816
1817 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1818 {
1819         struct blk_mq_hw_ctx *hctx;
1820         struct request_queue *q;
1821         bool shared;
1822         int i;
1823
1824         if (set->tag_list.next == set->tag_list.prev)
1825                 shared = false;
1826         else
1827                 shared = true;
1828
1829         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1830                 blk_mq_freeze_queue(q);
1831
1832                 queue_for_each_hw_ctx(q, hctx, i) {
1833                         if (shared)
1834                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1835                         else
1836                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1837                 }
1838                 blk_mq_unfreeze_queue(q);
1839         }
1840 }
1841
1842 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1843 {
1844         struct blk_mq_tag_set *set = q->tag_set;
1845
1846         mutex_lock(&set->tag_list_lock);
1847         list_del_init(&q->tag_set_list);
1848         blk_mq_update_tag_set_depth(set);
1849         mutex_unlock(&set->tag_list_lock);
1850 }
1851
1852 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1853                                      struct request_queue *q)
1854 {
1855         q->tag_set = set;
1856
1857         mutex_lock(&set->tag_list_lock);
1858         list_add_tail(&q->tag_set_list, &set->tag_list);
1859         blk_mq_update_tag_set_depth(set);
1860         mutex_unlock(&set->tag_list_lock);
1861 }
1862
1863 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1864 {
1865         struct blk_mq_hw_ctx **hctxs;
1866         struct blk_mq_ctx __percpu *ctx;
1867         struct request_queue *q;
1868         unsigned int *map;
1869         int i;
1870
1871         ctx = alloc_percpu(struct blk_mq_ctx);
1872         if (!ctx)
1873                 return ERR_PTR(-ENOMEM);
1874
1875         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1876                         set->numa_node);
1877
1878         if (!hctxs)
1879                 goto err_percpu;
1880
1881         map = blk_mq_make_queue_map(set);
1882         if (!map)
1883                 goto err_map;
1884
1885         for (i = 0; i < set->nr_hw_queues; i++) {
1886                 int node = blk_mq_hw_queue_to_node(map, i);
1887
1888                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1889                                         GFP_KERNEL, node);
1890                 if (!hctxs[i])
1891                         goto err_hctxs;
1892
1893                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1894                                                 node))
1895                         goto err_hctxs;
1896
1897                 atomic_set(&hctxs[i]->nr_active, 0);
1898                 hctxs[i]->numa_node = node;
1899                 hctxs[i]->queue_num = i;
1900         }
1901
1902         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1903         if (!q)
1904                 goto err_hctxs;
1905
1906         /*
1907          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1908          * See blk_register_queue() for details.
1909          */
1910         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1911                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1912                 goto err_map;
1913
1914         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1915         blk_queue_rq_timeout(q, 30000);
1916
1917         q->nr_queues = nr_cpu_ids;
1918         q->nr_hw_queues = set->nr_hw_queues;
1919         q->mq_map = map;
1920
1921         q->queue_ctx = ctx;
1922         q->queue_hw_ctx = hctxs;
1923
1924         q->mq_ops = set->ops;
1925         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1926
1927         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1928                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1929
1930         q->sg_reserved_size = INT_MAX;
1931
1932         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1933         INIT_LIST_HEAD(&q->requeue_list);
1934         spin_lock_init(&q->requeue_lock);
1935
1936         if (q->nr_hw_queues > 1)
1937                 blk_queue_make_request(q, blk_mq_make_request);
1938         else
1939                 blk_queue_make_request(q, blk_sq_make_request);
1940
1941         if (set->timeout)
1942                 blk_queue_rq_timeout(q, set->timeout);
1943
1944         /*
1945          * Do this after blk_queue_make_request() overrides it...
1946          */
1947         q->nr_requests = set->queue_depth;
1948
1949         if (set->ops->complete)
1950                 blk_queue_softirq_done(q, set->ops->complete);
1951
1952         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1953
1954         if (blk_mq_init_hw_queues(q, set))
1955                 goto err_hw;
1956
1957         mutex_lock(&all_q_mutex);
1958         list_add_tail(&q->all_q_node, &all_q_list);
1959         mutex_unlock(&all_q_mutex);
1960
1961         blk_mq_add_queue_tag_set(set, q);
1962
1963         blk_mq_map_swqueue(q);
1964
1965         return q;
1966
1967 err_hw:
1968         blk_cleanup_queue(q);
1969 err_hctxs:
1970         kfree(map);
1971         for (i = 0; i < set->nr_hw_queues; i++) {
1972                 if (!hctxs[i])
1973                         break;
1974                 free_cpumask_var(hctxs[i]->cpumask);
1975                 kfree(hctxs[i]);
1976         }
1977 err_map:
1978         kfree(hctxs);
1979 err_percpu:
1980         free_percpu(ctx);
1981         return ERR_PTR(-ENOMEM);
1982 }
1983 EXPORT_SYMBOL(blk_mq_init_queue);
1984
1985 void blk_mq_free_queue(struct request_queue *q)
1986 {
1987         struct blk_mq_tag_set   *set = q->tag_set;
1988
1989         blk_mq_del_queue_tag_set(q);
1990
1991         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1992         blk_mq_free_hw_queues(q, set);
1993
1994         percpu_ref_exit(&q->mq_usage_counter);
1995
1996         free_percpu(q->queue_ctx);
1997         kfree(q->queue_hw_ctx);
1998         kfree(q->mq_map);
1999
2000         q->queue_ctx = NULL;
2001         q->queue_hw_ctx = NULL;
2002         q->mq_map = NULL;
2003
2004         mutex_lock(&all_q_mutex);
2005         list_del_init(&q->all_q_node);
2006         mutex_unlock(&all_q_mutex);
2007 }
2008
2009 /* Basically redo blk_mq_init_queue with queue frozen */
2010 static void blk_mq_queue_reinit(struct request_queue *q)
2011 {
2012         WARN_ON_ONCE(!q->mq_freeze_depth);
2013
2014         blk_mq_sysfs_unregister(q);
2015
2016         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2017
2018         /*
2019          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2020          * we should change hctx numa_node according to new topology (this
2021          * involves free and re-allocate memory, worthy doing?)
2022          */
2023
2024         blk_mq_map_swqueue(q);
2025
2026         blk_mq_sysfs_register(q);
2027 }
2028
2029 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2030                                       unsigned long action, void *hcpu)
2031 {
2032         struct request_queue *q;
2033
2034         /*
2035          * Before new mappings are established, hotadded cpu might already
2036          * start handling requests. This doesn't break anything as we map
2037          * offline CPUs to first hardware queue. We will re-init the queue
2038          * below to get optimal settings.
2039          */
2040         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2041             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2042                 return NOTIFY_OK;
2043
2044         mutex_lock(&all_q_mutex);
2045
2046         /*
2047          * We need to freeze and reinit all existing queues.  Freezing
2048          * involves synchronous wait for an RCU grace period and doing it
2049          * one by one may take a long time.  Start freezing all queues in
2050          * one swoop and then wait for the completions so that freezing can
2051          * take place in parallel.
2052          */
2053         list_for_each_entry(q, &all_q_list, all_q_node)
2054                 blk_mq_freeze_queue_start(q);
2055         list_for_each_entry(q, &all_q_list, all_q_node)
2056                 blk_mq_freeze_queue_wait(q);
2057
2058         list_for_each_entry(q, &all_q_list, all_q_node)
2059                 blk_mq_queue_reinit(q);
2060
2061         list_for_each_entry(q, &all_q_list, all_q_node)
2062                 blk_mq_unfreeze_queue(q);
2063
2064         mutex_unlock(&all_q_mutex);
2065         return NOTIFY_OK;
2066 }
2067
2068 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2069 {
2070         int i;
2071
2072         for (i = 0; i < set->nr_hw_queues; i++) {
2073                 set->tags[i] = blk_mq_init_rq_map(set, i);
2074                 if (!set->tags[i])
2075                         goto out_unwind;
2076         }
2077
2078         return 0;
2079
2080 out_unwind:
2081         while (--i >= 0)
2082                 blk_mq_free_rq_map(set, set->tags[i], i);
2083
2084         return -ENOMEM;
2085 }
2086
2087 /*
2088  * Allocate the request maps associated with this tag_set. Note that this
2089  * may reduce the depth asked for, if memory is tight. set->queue_depth
2090  * will be updated to reflect the allocated depth.
2091  */
2092 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2093 {
2094         unsigned int depth;
2095         int err;
2096
2097         depth = set->queue_depth;
2098         do {
2099                 err = __blk_mq_alloc_rq_maps(set);
2100                 if (!err)
2101                         break;
2102
2103                 set->queue_depth >>= 1;
2104                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2105                         err = -ENOMEM;
2106                         break;
2107                 }
2108         } while (set->queue_depth);
2109
2110         if (!set->queue_depth || err) {
2111                 pr_err("blk-mq: failed to allocate request map\n");
2112                 return -ENOMEM;
2113         }
2114
2115         if (depth != set->queue_depth)
2116                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2117                                                 depth, set->queue_depth);
2118
2119         return 0;
2120 }
2121
2122 /*
2123  * Alloc a tag set to be associated with one or more request queues.
2124  * May fail with EINVAL for various error conditions. May adjust the
2125  * requested depth down, if if it too large. In that case, the set
2126  * value will be stored in set->queue_depth.
2127  */
2128 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2129 {
2130         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2131
2132         if (!set->nr_hw_queues)
2133                 return -EINVAL;
2134         if (!set->queue_depth)
2135                 return -EINVAL;
2136         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2137                 return -EINVAL;
2138
2139         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2140                 return -EINVAL;
2141
2142         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2143                 pr_info("blk-mq: reduced tag depth to %u\n",
2144                         BLK_MQ_MAX_DEPTH);
2145                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2146         }
2147
2148         /*
2149          * If a crashdump is active, then we are potentially in a very
2150          * memory constrained environment. Limit us to 1 queue and
2151          * 64 tags to prevent using too much memory.
2152          */
2153         if (is_kdump_kernel()) {
2154                 set->nr_hw_queues = 1;
2155                 set->queue_depth = min(64U, set->queue_depth);
2156         }
2157
2158         set->tags = kmalloc_node(set->nr_hw_queues *
2159                                  sizeof(struct blk_mq_tags *),
2160                                  GFP_KERNEL, set->numa_node);
2161         if (!set->tags)
2162                 return -ENOMEM;
2163
2164         if (blk_mq_alloc_rq_maps(set))
2165                 goto enomem;
2166
2167         mutex_init(&set->tag_list_lock);
2168         INIT_LIST_HEAD(&set->tag_list);
2169
2170         return 0;
2171 enomem:
2172         kfree(set->tags);
2173         set->tags = NULL;
2174         return -ENOMEM;
2175 }
2176 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2177
2178 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2179 {
2180         int i;
2181
2182         for (i = 0; i < set->nr_hw_queues; i++) {
2183                 if (set->tags[i])
2184                         blk_mq_free_rq_map(set, set->tags[i], i);
2185         }
2186
2187         kfree(set->tags);
2188         set->tags = NULL;
2189 }
2190 EXPORT_SYMBOL(blk_mq_free_tag_set);
2191
2192 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2193 {
2194         struct blk_mq_tag_set *set = q->tag_set;
2195         struct blk_mq_hw_ctx *hctx;
2196         int i, ret;
2197
2198         if (!set || nr > set->queue_depth)
2199                 return -EINVAL;
2200
2201         ret = 0;
2202         queue_for_each_hw_ctx(q, hctx, i) {
2203                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2204                 if (ret)
2205                         break;
2206         }
2207
2208         if (!ret)
2209                 q->nr_requests = nr;
2210
2211         return ret;
2212 }
2213
2214 void blk_mq_disable_hotplug(void)
2215 {
2216         mutex_lock(&all_q_mutex);
2217 }
2218
2219 void blk_mq_enable_hotplug(void)
2220 {
2221         mutex_unlock(&all_q_mutex);
2222 }
2223
2224 static int __init blk_mq_init(void)
2225 {
2226         blk_mq_cpu_init();
2227
2228         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2229
2230         return 0;
2231 }
2232 subsys_initcall(blk_mq_init);