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 static inline bool is_flush_request(struct request *rq,
 543                 struct blk_flush_queue *fq, unsigned int tag)
 544 {
 545         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 546                         fq->flush_rq->tag == tag);
 547 }
 548
 549 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 550 {
 551         struct request *rq = tags->rqs[tag];
 552         /* mq_ctx of flush rq is always cloned from the corresponding req */
 553         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 554
 555         if (!is_flush_request(rq, fq, tag))
 556                 return rq;
 557
 558         return fq->flush_rq;
 559 }
 560 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 561
 562 struct blk_mq_timeout_data {
 563         unsigned long next;
 564         unsigned int next_set;
 565 };
 566
 567 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 568 {
 569         struct blk_mq_ops *ops = req->q->mq_ops;
 570         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 571
 572         /*
 573          * We know that complete is set at this point. If STARTED isn't set
 574          * anymore, then the request isn't active and the "timeout" should
 575          * just be ignored. This can happen due to the bitflag ordering.
 576          * Timeout first checks if STARTED is set, and if it is, assumes
 577          * the request is active. But if we race with completion, then
 578          * we both flags will get cleared. So check here again, and ignore
 579          * a timeout event with a request that isn't active.
 580          */
 581         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 582                 return;
 583
 584         if (ops->timeout)
 585                 ret = ops->timeout(req, reserved);
 586
 587         switch (ret) {
 588         case BLK_EH_HANDLED:
 589                 __blk_mq_complete_request(req);
 590                 break;
 591         case BLK_EH_RESET_TIMER:
 592                 blk_add_timer(req);
 593                 blk_clear_rq_complete(req);
 594                 break;
 595         case BLK_EH_NOT_HANDLED:
 596                 break;
 597         default:
 598                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 599                 break;
 600         }
 601 }
 602
 603 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 604                 struct request *rq, void *priv, bool reserved)
 605 {
 606         struct blk_mq_timeout_data *data = priv;
 607
 608         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 609                 return;
 610
 611         if (time_after_eq(jiffies, rq->deadline)) {
 612                 if (!blk_mark_rq_complete(rq))
 613                         blk_mq_rq_timed_out(rq, reserved);
 614         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 615                 data->next = rq->deadline;
 616                 data->next_set = 1;
 617         }
 618 }
 619
 620 static void blk_mq_rq_timer(unsigned long priv)
 621 {
 622         struct request_queue *q = (struct request_queue *)priv;
 623         struct blk_mq_timeout_data data = {
 624                 .next           = 0,
 625                 .next_set       = 0,
 626         };
 627         struct blk_mq_hw_ctx *hctx;
 628         int i;
 629
 630         queue_for_each_hw_ctx(q, hctx, i) {
 631                 /*
 632                  * If not software queues are currently mapped to this
 633                  * hardware queue, there's nothing to check
 634                  */
 635                 if (!blk_mq_hw_queue_mapped(hctx))
 636                         continue;
 637
 638                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 639         }
 640
 641         if (data.next_set) {
 642                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 643                 mod_timer(&q->timeout, data.next);
 644         } else {
 645                 queue_for_each_hw_ctx(q, hctx, i)
 646                         blk_mq_tag_idle(hctx);
 647         }
 648 }
 649
 650 /*
 651  * Reverse check our software queue for entries that we could potentially
 652  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 653  * too much time checking for merges.
 654  */
 655 static bool blk_mq_attempt_merge(struct request_queue *q,
 656                                  struct blk_mq_ctx *ctx, struct bio *bio)
 657 {
 658         struct request *rq;
 659         int checked = 8;
 660
 661         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 662                 int el_ret;
 663
 664                 if (!checked--)
 665                         break;
 666
 667                 if (!blk_rq_merge_ok(rq, bio))
 668                         continue;
 669
 670                 el_ret = blk_try_merge(rq, bio);
 671                 if (el_ret == ELEVATOR_BACK_MERGE) {
 672                         if (bio_attempt_back_merge(q, rq, bio)) {
 673                                 ctx->rq_merged++;
 674                                 return true;
 675                         }
 676                         break;
 677                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 678                         if (bio_attempt_front_merge(q, rq, bio)) {
 679                                 ctx->rq_merged++;
 680                                 return true;
 681                         }
 682                         break;
 683                 }
 684         }
 685
 686         return false;
 687 }
 688
 689 /*
 690  * Process software queues that have been marked busy, splicing them
 691  * to the for-dispatch
 692  */
 693 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 694 {
 695         struct blk_mq_ctx *ctx;
 696         int i;
 697
 698         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 699                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 700                 unsigned int off, bit;
 701
 702                 if (!bm->word)
 703                         continue;
 704
 705                 bit = 0;
 706                 off = i * hctx->ctx_map.bits_per_word;
 707                 do {
 708                         bit = find_next_bit(&bm->word, bm->depth, bit);
 709                         if (bit >= bm->depth)
 710                                 break;
 711
 712                         ctx = hctx->ctxs[bit + off];
 713                         clear_bit(bit, &bm->word);
 714                         spin_lock(&ctx->lock);
 715                         list_splice_tail_init(&ctx->rq_list, list);
 716                         spin_unlock(&ctx->lock);
 717
 718                         bit++;
 719                 } while (1);
 720         }
 721 }
 722
 723 /*
 724  * Run this hardware queue, pulling any software queues mapped to it in.
 725  * Note that this function currently has various problems around ordering
 726  * of IO. In particular, we'd like FIFO behaviour on handling existing
 727  * items on the hctx->dispatch list. Ignore that for now.
 728  */
 729 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 730 {
 731         struct request_queue *q = hctx->queue;
 732         struct request *rq;
 733         LIST_HEAD(rq_list);
 734         LIST_HEAD(driver_list);
 735         struct list_head *dptr;
 736         int queued;
 737
 738         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 739
 740         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 741                 return;
 742
 743         hctx->run++;
 744
 745         /*
 746          * Touch any software queue that has pending entries.
 747          */
 748         flush_busy_ctxs(hctx, &rq_list);
 749
 750         /*
 751          * If we have previous entries on our dispatch list, grab them
 752          * and stuff them at the front for more fair dispatch.
 753          */
 754         if (!list_empty_careful(&hctx->dispatch)) {
 755                 spin_lock(&hctx->lock);
 756                 if (!list_empty(&hctx->dispatch))
 757                         list_splice_init(&hctx->dispatch, &rq_list);
 758                 spin_unlock(&hctx->lock);
 759         }
 760
 761         /*
 762          * Start off with dptr being NULL, so we start the first request
 763          * immediately, even if we have more pending.
 764          */
 765         dptr = NULL;
 766
 767         /*
 768          * Now process all the entries, sending them to the driver.
 769          */
 770         queued = 0;
 771         while (!list_empty(&rq_list)) {
 772                 struct blk_mq_queue_data bd;
 773                 int ret;
 774
 775                 rq = list_first_entry(&rq_list, struct request, queuelist);
 776                 list_del_init(&rq->queuelist);
 777
 778                 bd.rq = rq;
 779                 bd.list = dptr;
 780                 bd.last = list_empty(&rq_list);
 781
 782                 ret = q->mq_ops->queue_rq(hctx, &bd);
 783                 switch (ret) {
 784                 case BLK_MQ_RQ_QUEUE_OK:
 785                         queued++;
 786                         continue;
 787                 case BLK_MQ_RQ_QUEUE_BUSY:
 788                         list_add(&rq->queuelist, &rq_list);
 789                         __blk_mq_requeue_request(rq);
 790                         break;
 791                 default:
 792                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 793                 case BLK_MQ_RQ_QUEUE_ERROR:
 794                         rq->errors = -EIO;
 795                         blk_mq_end_request(rq, rq->errors);
 796                         break;
 797                 }
 798
 799                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 800                         break;
 801
 802                 /*
 803                  * We've done the first request. If we have more than 1
 804                  * left in the list, set dptr to defer issue.
 805                  */
 806                 if (!dptr && rq_list.next != rq_list.prev)
 807                         dptr = &driver_list;
 808         }
 809
 810         if (!queued)
 811                 hctx->dispatched[0]++;
 812         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 813                 hctx->dispatched[ilog2(queued) + 1]++;
 814
 815         /*
 816          * Any items that need requeuing? Stuff them into hctx->dispatch,
 817          * that is where we will continue on next queue run.
 818          */
 819         if (!list_empty(&rq_list)) {
 820                 spin_lock(&hctx->lock);
 821                 list_splice(&rq_list, &hctx->dispatch);
 822                 spin_unlock(&hctx->lock);
 823         }
 824 }
 825
 826 /*
 827  * It'd be great if the workqueue API had a way to pass
 828  * in a mask and had some smarts for more clever placement.
 829  * For now we just round-robin here, switching for every
 830  * BLK_MQ_CPU_WORK_BATCH queued items.
 831  */
 832 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 833 {
 834         if (hctx->queue->nr_hw_queues == 1)
 835                 return WORK_CPU_UNBOUND;
 836
 837         if (--hctx->next_cpu_batch <= 0) {
 838                 int cpu = hctx->next_cpu, next_cpu;
 839
 840                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 841                 if (next_cpu >= nr_cpu_ids)
 842                         next_cpu = cpumask_first(hctx->cpumask);
 843
 844                 hctx->next_cpu = next_cpu;
 845                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 846
 847                 return cpu;
 848         }
 849
 850         return hctx->next_cpu;
 851 }
 852
 853 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 854 {
 855         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 856             !blk_mq_hw_queue_mapped(hctx)))
 857                 return;
 858
 859         if (!async) {
 860                 int cpu = get_cpu();
 861                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 862                         __blk_mq_run_hw_queue(hctx);
 863                         put_cpu();
 864                         return;
 865                 }
 866
 867                 put_cpu();
 868         }
 869
 870         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 871                         &hctx->run_work, 0);
 872 }
 873
 874 void blk_mq_run_queues(struct request_queue *q, bool async)
 875 {
 876         struct blk_mq_hw_ctx *hctx;
 877         int i;
 878
 879         queue_for_each_hw_ctx(q, hctx, i) {
 880                 if ((!blk_mq_hctx_has_pending(hctx) &&
 881                     list_empty_careful(&hctx->dispatch)) ||
 882                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 883                         continue;
 884
 885                 blk_mq_run_hw_queue(hctx, async);
 886         }
 887 }
 888 EXPORT_SYMBOL(blk_mq_run_queues);
 889
 890 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 891 {
 892         cancel_delayed_work(&hctx->run_work);
 893         cancel_delayed_work(&hctx->delay_work);
 894         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 895 }
 896 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 897
 898 void blk_mq_stop_hw_queues(struct request_queue *q)
 899 {
 900         struct blk_mq_hw_ctx *hctx;
 901         int i;
 902
 903         queue_for_each_hw_ctx(q, hctx, i)
 904                 blk_mq_stop_hw_queue(hctx);
 905 }
 906 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 907
 908 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 909 {
 910         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 911
 912         blk_mq_run_hw_queue(hctx, false);
 913 }
 914 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 915
 916 void blk_mq_start_hw_queues(struct request_queue *q)
 917 {
 918         struct blk_mq_hw_ctx *hctx;
 919         int i;
 920
 921         queue_for_each_hw_ctx(q, hctx, i)
 922                 blk_mq_start_hw_queue(hctx);
 923 }
 924 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 925
 926
 927 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 928 {
 929         struct blk_mq_hw_ctx *hctx;
 930         int i;
 931
 932         queue_for_each_hw_ctx(q, hctx, i) {
 933                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 934                         continue;
 935
 936                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 937                 blk_mq_run_hw_queue(hctx, async);
 938         }
 939 }
 940 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 941
 942 static void blk_mq_run_work_fn(struct work_struct *work)
 943 {
 944         struct blk_mq_hw_ctx *hctx;
 945
 946         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 947
 948         __blk_mq_run_hw_queue(hctx);
 949 }
 950
 951 static void blk_mq_delay_work_fn(struct work_struct *work)
 952 {
 953         struct blk_mq_hw_ctx *hctx;
 954
 955         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 956
 957         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 958                 __blk_mq_run_hw_queue(hctx);
 959 }
 960
 961 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 962 {
 963         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
 964                 return;
 965
 966         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 967                         &hctx->delay_work, msecs_to_jiffies(msecs));
 968 }
 969 EXPORT_SYMBOL(blk_mq_delay_queue);
 970
 971 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 972                                     struct request *rq, bool at_head)
 973 {
 974         struct blk_mq_ctx *ctx = rq->mq_ctx;
 975
 976         trace_block_rq_insert(hctx->queue, rq);
 977
 978         if (at_head)
 979                 list_add(&rq->queuelist, &ctx->rq_list);
 980         else
 981                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 982
 983         blk_mq_hctx_mark_pending(hctx, ctx);
 984 }
 985
 986 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
 987                 bool async)
 988 {
 989         struct request_queue *q = rq->q;
 990         struct blk_mq_hw_ctx *hctx;
 991         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
 992
 993         current_ctx = blk_mq_get_ctx(q);
 994         if (!cpu_online(ctx->cpu))
 995                 rq->mq_ctx = ctx = current_ctx;
 996
 997         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 998
 999         spin_lock(&ctx->lock);
1000         __blk_mq_insert_request(hctx, rq, at_head);
1001         spin_unlock(&ctx->lock);
1002
1003         if (run_queue)
1004                 blk_mq_run_hw_queue(hctx, async);
1005
1006         blk_mq_put_ctx(current_ctx);
1007 }
1008
1009 static void blk_mq_insert_requests(struct request_queue *q,
1010                                      struct blk_mq_ctx *ctx,
1011                                      struct list_head *list,
1012                                      int depth,
1013                                      bool from_schedule)
1014
1015 {
1016         struct blk_mq_hw_ctx *hctx;
1017         struct blk_mq_ctx *current_ctx;
1018
1019         trace_block_unplug(q, depth, !from_schedule);
1020
1021         current_ctx = blk_mq_get_ctx(q);
1022
1023         if (!cpu_online(ctx->cpu))
1024                 ctx = current_ctx;
1025         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1026
1027         /*
1028          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1029          * offline now
1030          */
1031         spin_lock(&ctx->lock);
1032         while (!list_empty(list)) {
1033                 struct request *rq;
1034
1035                 rq = list_first_entry(list, struct request, queuelist);
1036                 list_del_init(&rq->queuelist);
1037                 rq->mq_ctx = ctx;
1038                 __blk_mq_insert_request(hctx, rq, false);
1039         }
1040         spin_unlock(&ctx->lock);
1041
1042         blk_mq_run_hw_queue(hctx, from_schedule);
1043         blk_mq_put_ctx(current_ctx);
1044 }
1045
1046 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1047 {
1048         struct request *rqa = container_of(a, struct request, queuelist);
1049         struct request *rqb = container_of(b, struct request, queuelist);
1050
1051         return !(rqa->mq_ctx < rqb->mq_ctx ||
1052                  (rqa->mq_ctx == rqb->mq_ctx &&
1053                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1054 }
1055
1056 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1057 {
1058         struct blk_mq_ctx *this_ctx;
1059         struct request_queue *this_q;
1060         struct request *rq;
1061         LIST_HEAD(list);
1062         LIST_HEAD(ctx_list);
1063         unsigned int depth;
1064
1065         list_splice_init(&plug->mq_list, &list);
1066
1067         list_sort(NULL, &list, plug_ctx_cmp);
1068
1069         this_q = NULL;
1070         this_ctx = NULL;
1071         depth = 0;
1072
1073         while (!list_empty(&list)) {
1074                 rq = list_entry_rq(list.next);
1075                 list_del_init(&rq->queuelist);
1076                 BUG_ON(!rq->q);
1077                 if (rq->mq_ctx != this_ctx) {
1078                         if (this_ctx) {
1079                                 blk_mq_insert_requests(this_q, this_ctx,
1080                                                         &ctx_list, depth,
1081                                                         from_schedule);
1082                         }
1083
1084                         this_ctx = rq->mq_ctx;
1085                         this_q = rq->q;
1086                         depth = 0;
1087                 }
1088
1089                 depth++;
1090                 list_add_tail(&rq->queuelist, &ctx_list);
1091         }
1092
1093         /*
1094          * If 'this_ctx' is set, we know we have entries to complete
1095          * on 'ctx_list'. Do those.
1096          */
1097         if (this_ctx) {
1098                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1099                                        from_schedule);
1100         }
1101 }
1102
1103 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1104 {
1105         init_request_from_bio(rq, bio);
1106
1107         if (blk_do_io_stat(rq))
1108                 blk_account_io_start(rq, 1);
1109 }
1110
1111 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1112 {
1113         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1114                 !blk_queue_nomerges(hctx->queue);
1115 }
1116
1117 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1118                                          struct blk_mq_ctx *ctx,
1119                                          struct request *rq, struct bio *bio)
1120 {
1121         if (!hctx_allow_merges(hctx)) {
1122                 blk_mq_bio_to_request(rq, bio);
1123                 spin_lock(&ctx->lock);
1124 insert_rq:
1125                 __blk_mq_insert_request(hctx, rq, false);
1126                 spin_unlock(&ctx->lock);
1127                 return false;
1128         } else {
1129                 struct request_queue *q = hctx->queue;
1130
1131                 spin_lock(&ctx->lock);
1132                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1133                         blk_mq_bio_to_request(rq, bio);
1134                         goto insert_rq;
1135                 }
1136
1137                 spin_unlock(&ctx->lock);
1138                 __blk_mq_free_request(hctx, ctx, rq);
1139                 return true;
1140         }
1141 }
1142
1143 struct blk_map_ctx {
1144         struct blk_mq_hw_ctx *hctx;
1145         struct blk_mq_ctx *ctx;
1146 };
1147
1148 static struct request *blk_mq_map_request(struct request_queue *q,
1149                                           struct bio *bio,
1150                                           struct blk_map_ctx *data)
1151 {
1152         struct blk_mq_hw_ctx *hctx;
1153         struct blk_mq_ctx *ctx;
1154         struct request *rq;
1155         int rw = bio_data_dir(bio);
1156         struct blk_mq_alloc_data alloc_data;
1157
1158         if (unlikely(blk_mq_queue_enter(q))) {
1159                 bio_endio(bio, -EIO);
1160                 return NULL;
1161         }
1162
1163         ctx = blk_mq_get_ctx(q);
1164         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1165
1166         if (rw_is_sync(bio->bi_rw))
1167                 rw |= REQ_SYNC;
1168
1169         trace_block_getrq(q, bio, rw);
1170         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1171                         hctx);
1172         rq = __blk_mq_alloc_request(&alloc_data, rw);
1173         if (unlikely(!rq)) {
1174                 __blk_mq_run_hw_queue(hctx);
1175                 blk_mq_put_ctx(ctx);
1176                 trace_block_sleeprq(q, bio, rw);
1177
1178                 ctx = blk_mq_get_ctx(q);
1179                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1180                 blk_mq_set_alloc_data(&alloc_data, q,
1181                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1182                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1183                 ctx = alloc_data.ctx;
1184                 hctx = alloc_data.hctx;
1185         }
1186
1187         hctx->queued++;
1188         data->hctx = hctx;
1189         data->ctx = ctx;
1190         return rq;
1191 }
1192
1193 /*
1194  * Multiple hardware queue variant. This will not use per-process plugs,
1195  * but will attempt to bypass the hctx queueing if we can go straight to
1196  * hardware for SYNC IO.
1197  */
1198 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1199 {
1200         const int is_sync = rw_is_sync(bio->bi_rw);
1201         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1202         struct blk_map_ctx data;
1203         struct request *rq;
1204
1205         blk_queue_bounce(q, &bio);
1206
1207         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1208                 bio_endio(bio, -EIO);
1209                 return;
1210         }
1211
1212         rq = blk_mq_map_request(q, bio, &data);
1213         if (unlikely(!rq))
1214                 return;
1215
1216         if (unlikely(is_flush_fua)) {
1217                 blk_mq_bio_to_request(rq, bio);
1218                 blk_insert_flush(rq);
1219                 goto run_queue;
1220         }
1221
1222         /*
1223          * If the driver supports defer issued based on 'last', then
1224          * queue it up like normal since we can potentially save some
1225          * CPU this way.
1226          */
1227         if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1228                 struct blk_mq_queue_data bd = {
1229                         .rq = rq,
1230                         .list = NULL,
1231                         .last = 1
1232                 };
1233                 int ret;
1234
1235                 blk_mq_bio_to_request(rq, bio);
1236
1237                 /*
1238                  * For OK queue, we are done. For error, kill it. Any other
1239                  * error (busy), just add it to our list as we previously
1240                  * would have done
1241                  */
1242                 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1243                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1244                         goto done;
1245                 else {
1246                         __blk_mq_requeue_request(rq);
1247
1248                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1249                                 rq->errors = -EIO;
1250                                 blk_mq_end_request(rq, rq->errors);
1251                                 goto done;
1252                         }
1253                 }
1254         }
1255
1256         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1257                 /*
1258                  * For a SYNC request, send it to the hardware immediately. For
1259                  * an ASYNC request, just ensure that we run it later on. The
1260                  * latter allows for merging opportunities and more efficient
1261                  * dispatching.
1262                  */
1263 run_queue:
1264                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1265         }
1266 done:
1267         blk_mq_put_ctx(data.ctx);
1268 }
1269
1270 /*
1271  * Single hardware queue variant. This will attempt to use any per-process
1272  * plug for merging and IO deferral.
1273  */
1274 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1275 {
1276         const int is_sync = rw_is_sync(bio->bi_rw);
1277         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1278         unsigned int use_plug, request_count = 0;
1279         struct blk_map_ctx data;
1280         struct request *rq;
1281
1282         /*
1283          * If we have multiple hardware queues, just go directly to
1284          * one of those for sync IO.
1285          */
1286         use_plug = !is_flush_fua && !is_sync;
1287
1288         blk_queue_bounce(q, &bio);
1289
1290         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1291                 bio_endio(bio, -EIO);
1292                 return;
1293         }
1294
1295         if (use_plug && !blk_queue_nomerges(q) &&
1296             blk_attempt_plug_merge(q, bio, &request_count))
1297                 return;
1298
1299         rq = blk_mq_map_request(q, bio, &data);
1300         if (unlikely(!rq))
1301                 return;
1302
1303         if (unlikely(is_flush_fua)) {
1304                 blk_mq_bio_to_request(rq, bio);
1305                 blk_insert_flush(rq);
1306                 goto run_queue;
1307         }
1308
1309         /*
1310          * A task plug currently exists. Since this is completely lockless,
1311          * utilize that to temporarily store requests until the task is
1312          * either done or scheduled away.
1313          */
1314         if (use_plug) {
1315                 struct blk_plug *plug = current->plug;
1316
1317                 if (plug) {
1318                         blk_mq_bio_to_request(rq, bio);
1319                         if (list_empty(&plug->mq_list))
1320                                 trace_block_plug(q);
1321                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1322                                 blk_flush_plug_list(plug, false);
1323                                 trace_block_plug(q);
1324                         }
1325                         list_add_tail(&rq->queuelist, &plug->mq_list);
1326                         blk_mq_put_ctx(data.ctx);
1327                         return;
1328                 }
1329         }
1330
1331         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1332                 /*
1333                  * For a SYNC request, send it to the hardware immediately. For
1334                  * an ASYNC request, just ensure that we run it later on. The
1335                  * latter allows for merging opportunities and more efficient
1336                  * dispatching.
1337                  */
1338 run_queue:
1339                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1340         }
1341
1342         blk_mq_put_ctx(data.ctx);
1343 }
1344
1345 /*
1346  * Default mapping to a software queue, since we use one per CPU.
1347  */
1348 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1349 {
1350         return q->queue_hw_ctx[q->mq_map[cpu]];
1351 }
1352 EXPORT_SYMBOL(blk_mq_map_queue);
1353
1354 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1355                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1356 {
1357         struct page *page;
1358
1359         if (tags->rqs && set->ops->exit_request) {
1360                 int i;
1361
1362                 for (i = 0; i < tags->nr_tags; i++) {
1363                         if (!tags->rqs[i])
1364                                 continue;
1365                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1366                                                 hctx_idx, i);
1367                         tags->rqs[i] = NULL;
1368                 }
1369         }
1370
1371         while (!list_empty(&tags->page_list)) {
1372                 page = list_first_entry(&tags->page_list, struct page, lru);
1373                 list_del_init(&page->lru);
1374                 __free_pages(page, page->private);
1375         }
1376
1377         kfree(tags->rqs);
1378
1379         blk_mq_free_tags(tags);
1380 }
1381
1382 static size_t order_to_size(unsigned int order)
1383 {
1384         return (size_t)PAGE_SIZE << order;
1385 }
1386
1387 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1388                 unsigned int hctx_idx)
1389 {
1390         struct blk_mq_tags *tags;
1391         unsigned int i, j, entries_per_page, max_order = 4;
1392         size_t rq_size, left;
1393
1394         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1395                                 set->numa_node);
1396         if (!tags)
1397                 return NULL;
1398
1399         INIT_LIST_HEAD(&tags->page_list);
1400
1401         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1402                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1403                                  set->numa_node);
1404         if (!tags->rqs) {
1405                 blk_mq_free_tags(tags);
1406                 return NULL;
1407         }
1408
1409         /*
1410          * rq_size is the size of the request plus driver payload, rounded
1411          * to the cacheline size
1412          */
1413         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1414                                 cache_line_size());
1415         left = rq_size * set->queue_depth;
1416
1417         for (i = 0; i < set->queue_depth; ) {
1418                 int this_order = max_order;
1419                 struct page *page;
1420                 int to_do;
1421                 void *p;
1422
1423                 while (left < order_to_size(this_order - 1) && this_order)
1424                         this_order--;
1425
1426                 do {
1427                         page = alloc_pages_node(set->numa_node,
1428                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1429                                 this_order);
1430                         if (page)
1431                                 break;
1432                         if (!this_order--)
1433                                 break;
1434                         if (order_to_size(this_order) < rq_size)
1435                                 break;
1436                 } while (1);
1437
1438                 if (!page)
1439                         goto fail;
1440
1441                 page->private = this_order;
1442                 list_add_tail(&page->lru, &tags->page_list);
1443
1444                 p = page_address(page);
1445                 entries_per_page = order_to_size(this_order) / rq_size;
1446                 to_do = min(entries_per_page, set->queue_depth - i);
1447                 left -= to_do * rq_size;
1448                 for (j = 0; j < to_do; j++) {
1449                         tags->rqs[i] = p;
1450                         tags->rqs[i]->atomic_flags = 0;
1451                         tags->rqs[i]->cmd_flags = 0;
1452                         if (set->ops->init_request) {
1453                                 if (set->ops->init_request(set->driver_data,
1454                                                 tags->rqs[i], hctx_idx, i,
1455                                                 set->numa_node)) {
1456                                         tags->rqs[i] = NULL;
1457                                         goto fail;
1458                                 }
1459                         }
1460
1461                         p += rq_size;
1462                         i++;
1463                 }
1464         }
1465
1466         return tags;
1467
1468 fail:
1469         blk_mq_free_rq_map(set, tags, hctx_idx);
1470         return NULL;
1471 }
1472
1473 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1474 {
1475         kfree(bitmap->map);
1476 }
1477
1478 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1479 {
1480         unsigned int bpw = 8, total, num_maps, i;
1481
1482         bitmap->bits_per_word = bpw;
1483
1484         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1485         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1486                                         GFP_KERNEL, node);
1487         if (!bitmap->map)
1488                 return -ENOMEM;
1489
1490         bitmap->map_size = num_maps;
1491
1492         total = nr_cpu_ids;
1493         for (i = 0; i < num_maps; i++) {
1494                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1495                 total -= bitmap->map[i].depth;
1496         }
1497
1498         return 0;
1499 }
1500
1501 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1502 {
1503         struct request_queue *q = hctx->queue;
1504         struct blk_mq_ctx *ctx;
1505         LIST_HEAD(tmp);
1506
1507         /*
1508          * Move ctx entries to new CPU, if this one is going away.
1509          */
1510         ctx = __blk_mq_get_ctx(q, cpu);
1511
1512         spin_lock(&ctx->lock);
1513         if (!list_empty(&ctx->rq_list)) {
1514                 list_splice_init(&ctx->rq_list, &tmp);
1515                 blk_mq_hctx_clear_pending(hctx, ctx);
1516         }
1517         spin_unlock(&ctx->lock);
1518
1519         if (list_empty(&tmp))
1520                 return NOTIFY_OK;
1521
1522         ctx = blk_mq_get_ctx(q);
1523         spin_lock(&ctx->lock);
1524
1525         while (!list_empty(&tmp)) {
1526                 struct request *rq;
1527
1528                 rq = list_first_entry(&tmp, struct request, queuelist);
1529                 rq->mq_ctx = ctx;
1530                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1531         }
1532
1533         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1534         blk_mq_hctx_mark_pending(hctx, ctx);
1535
1536         spin_unlock(&ctx->lock);
1537
1538         blk_mq_run_hw_queue(hctx, true);
1539         blk_mq_put_ctx(ctx);
1540         return NOTIFY_OK;
1541 }
1542
1543 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1544 {
1545         struct request_queue *q = hctx->queue;
1546         struct blk_mq_tag_set *set = q->tag_set;
1547
1548         if (set->tags[hctx->queue_num])
1549                 return NOTIFY_OK;
1550
1551         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1552         if (!set->tags[hctx->queue_num])
1553                 return NOTIFY_STOP;
1554
1555         hctx->tags = set->tags[hctx->queue_num];
1556         return NOTIFY_OK;
1557 }
1558
1559 static int blk_mq_hctx_notify(void *data, unsigned long action,
1560                               unsigned int cpu)
1561 {
1562         struct blk_mq_hw_ctx *hctx = data;
1563
1564         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1565                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1566         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1567                 return blk_mq_hctx_cpu_online(hctx, cpu);
1568
1569         return NOTIFY_OK;
1570 }
1571
1572 static void blk_mq_exit_hctx(struct request_queue *q,
1573                 struct blk_mq_tag_set *set,
1574                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1575 {
1576         unsigned flush_start_tag = set->queue_depth;
1577
1578         blk_mq_tag_idle(hctx);
1579
1580         if (set->ops->exit_request)
1581                 set->ops->exit_request(set->driver_data,
1582                                        hctx->fq->flush_rq, hctx_idx,
1583                                        flush_start_tag + hctx_idx);
1584
1585         if (set->ops->exit_hctx)
1586                 set->ops->exit_hctx(hctx, hctx_idx);
1587
1588         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1589         blk_free_flush_queue(hctx->fq);
1590         kfree(hctx->ctxs);
1591         blk_mq_free_bitmap(&hctx->ctx_map);
1592 }
1593
1594 static void blk_mq_exit_hw_queues(struct request_queue *q,
1595                 struct blk_mq_tag_set *set, int nr_queue)
1596 {
1597         struct blk_mq_hw_ctx *hctx;
1598         unsigned int i;
1599
1600         queue_for_each_hw_ctx(q, hctx, i) {
1601                 if (i == nr_queue)
1602                         break;
1603                 blk_mq_exit_hctx(q, set, hctx, i);
1604         }
1605 }
1606
1607 static void blk_mq_free_hw_queues(struct request_queue *q,
1608                 struct blk_mq_tag_set *set)
1609 {
1610         struct blk_mq_hw_ctx *hctx;
1611         unsigned int i;
1612
1613         queue_for_each_hw_ctx(q, hctx, i) {
1614                 free_cpumask_var(hctx->cpumask);
1615                 kfree(hctx);
1616         }
1617 }
1618
1619 static int blk_mq_init_hctx(struct request_queue *q,
1620                 struct blk_mq_tag_set *set,
1621                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1622 {
1623         int node;
1624         unsigned flush_start_tag = set->queue_depth;
1625
1626         node = hctx->numa_node;
1627         if (node == NUMA_NO_NODE)
1628                 node = hctx->numa_node = set->numa_node;
1629
1630         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1631         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1632         spin_lock_init(&hctx->lock);
1633         INIT_LIST_HEAD(&hctx->dispatch);
1634         hctx->queue = q;
1635         hctx->queue_num = hctx_idx;
1636         hctx->flags = set->flags;
1637
1638         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1639                                         blk_mq_hctx_notify, hctx);
1640         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1641
1642         hctx->tags = set->tags[hctx_idx];
1643
1644         /*
1645          * Allocate space for all possible cpus to avoid allocation at
1646          * runtime
1647          */
1648         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1649                                         GFP_KERNEL, node);
1650         if (!hctx->ctxs)
1651                 goto unregister_cpu_notifier;
1652
1653         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1654                 goto free_ctxs;
1655
1656         hctx->nr_ctx = 0;
1657
1658         if (set->ops->init_hctx &&
1659             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1660                 goto free_bitmap;
1661
1662         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1663         if (!hctx->fq)
1664                 goto exit_hctx;
1665
1666         if (set->ops->init_request &&
1667             set->ops->init_request(set->driver_data,
1668                                    hctx->fq->flush_rq, hctx_idx,
1669                                    flush_start_tag + hctx_idx, node))
1670                 goto free_fq;
1671
1672         return 0;
1673
1674  free_fq:
1675         kfree(hctx->fq);
1676  exit_hctx:
1677         if (set->ops->exit_hctx)
1678                 set->ops->exit_hctx(hctx, hctx_idx);
1679  free_bitmap:
1680         blk_mq_free_bitmap(&hctx->ctx_map);
1681  free_ctxs:
1682         kfree(hctx->ctxs);
1683  unregister_cpu_notifier:
1684         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1685
1686         return -1;
1687 }
1688
1689 static int blk_mq_init_hw_queues(struct request_queue *q,
1690                 struct blk_mq_tag_set *set)
1691 {
1692         struct blk_mq_hw_ctx *hctx;
1693         unsigned int i;
1694
1695         /*
1696          * Initialize hardware queues
1697          */
1698         queue_for_each_hw_ctx(q, hctx, i) {
1699                 if (blk_mq_init_hctx(q, set, hctx, i))
1700                         break;
1701         }
1702
1703         if (i == q->nr_hw_queues)
1704                 return 0;
1705
1706         /*
1707          * Init failed
1708          */
1709         blk_mq_exit_hw_queues(q, set, i);
1710
1711         return 1;
1712 }
1713
1714 static void blk_mq_init_cpu_queues(struct request_queue *q,
1715                                    unsigned int nr_hw_queues)
1716 {
1717         unsigned int i;
1718
1719         for_each_possible_cpu(i) {
1720                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1721                 struct blk_mq_hw_ctx *hctx;
1722
1723                 memset(__ctx, 0, sizeof(*__ctx));
1724                 __ctx->cpu = i;
1725                 spin_lock_init(&__ctx->lock);
1726                 INIT_LIST_HEAD(&__ctx->rq_list);
1727                 __ctx->queue = q;
1728
1729                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1730                 if (!cpu_online(i))
1731                         continue;
1732
1733                 hctx = q->mq_ops->map_queue(q, i);
1734                 cpumask_set_cpu(i, hctx->cpumask);
1735                 hctx->nr_ctx++;
1736
1737                 /*
1738                  * Set local node, IFF we have more than one hw queue. If
1739                  * not, we remain on the home node of the device
1740                  */
1741                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1742                         hctx->numa_node = cpu_to_node(i);
1743         }
1744 }
1745
1746 static void blk_mq_map_swqueue(struct request_queue *q)
1747 {
1748         unsigned int i;
1749         struct blk_mq_hw_ctx *hctx;
1750         struct blk_mq_ctx *ctx;
1751
1752         queue_for_each_hw_ctx(q, hctx, i) {
1753                 cpumask_clear(hctx->cpumask);
1754                 hctx->nr_ctx = 0;
1755         }
1756
1757         /*
1758          * Map software to hardware queues
1759          */
1760         queue_for_each_ctx(q, ctx, i) {
1761                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1762                 if (!cpu_online(i))
1763                         continue;
1764
1765                 hctx = q->mq_ops->map_queue(q, i);
1766                 cpumask_set_cpu(i, hctx->cpumask);
1767                 ctx->index_hw = hctx->nr_ctx;
1768                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1769         }
1770
1771         queue_for_each_hw_ctx(q, hctx, i) {
1772                 /*
1773                  * If no software queues are mapped to this hardware queue,
1774                  * disable it and free the request entries.
1775                  */
1776                 if (!hctx->nr_ctx) {
1777                         struct blk_mq_tag_set *set = q->tag_set;
1778
1779                         if (set->tags[i]) {
1780                                 blk_mq_free_rq_map(set, set->tags[i], i);
1781                                 set->tags[i] = NULL;
1782                                 hctx->tags = NULL;
1783                         }
1784                         continue;
1785                 }
1786
1787                 /*
1788                  * Initialize batch roundrobin counts
1789                  */
1790                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1791                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1792         }
1793 }
1794
1795 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1796 {
1797         struct blk_mq_hw_ctx *hctx;
1798         struct request_queue *q;
1799         bool shared;
1800         int i;
1801
1802         if (set->tag_list.next == set->tag_list.prev)
1803                 shared = false;
1804         else
1805                 shared = true;
1806
1807         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1808                 blk_mq_freeze_queue(q);
1809
1810                 queue_for_each_hw_ctx(q, hctx, i) {
1811                         if (shared)
1812                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1813                         else
1814                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1815                 }
1816                 blk_mq_unfreeze_queue(q);
1817         }
1818 }
1819
1820 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1821 {
1822         struct blk_mq_tag_set *set = q->tag_set;
1823
1824         mutex_lock(&set->tag_list_lock);
1825         list_del_init(&q->tag_set_list);
1826         blk_mq_update_tag_set_depth(set);
1827         mutex_unlock(&set->tag_list_lock);
1828 }
1829
1830 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1831                                      struct request_queue *q)
1832 {
1833         q->tag_set = set;
1834
1835         mutex_lock(&set->tag_list_lock);
1836         list_add_tail(&q->tag_set_list, &set->tag_list);
1837         blk_mq_update_tag_set_depth(set);
1838         mutex_unlock(&set->tag_list_lock);
1839 }
1840
1841 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1842 {
1843         struct blk_mq_hw_ctx **hctxs;
1844         struct blk_mq_ctx __percpu *ctx;
1845         struct request_queue *q;
1846         unsigned int *map;
1847         int i;
1848
1849         ctx = alloc_percpu(struct blk_mq_ctx);
1850         if (!ctx)
1851                 return ERR_PTR(-ENOMEM);
1852
1853         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1854                         set->numa_node);
1855
1856         if (!hctxs)
1857                 goto err_percpu;
1858
1859         map = blk_mq_make_queue_map(set);
1860         if (!map)
1861                 goto err_map;
1862
1863         for (i = 0; i < set->nr_hw_queues; i++) {
1864                 int node = blk_mq_hw_queue_to_node(map, i);
1865
1866                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1867                                         GFP_KERNEL, node);
1868                 if (!hctxs[i])
1869                         goto err_hctxs;
1870
1871                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1872                                                 node))
1873                         goto err_hctxs;
1874
1875                 atomic_set(&hctxs[i]->nr_active, 0);
1876                 hctxs[i]->numa_node = node;
1877                 hctxs[i]->queue_num = i;
1878         }
1879
1880         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1881         if (!q)
1882                 goto err_hctxs;
1883
1884         /*
1885          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1886          * See blk_register_queue() for details.
1887          */
1888         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1889                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1890                 goto err_map;
1891
1892         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1893         blk_queue_rq_timeout(q, 30000);
1894
1895         q->nr_queues = nr_cpu_ids;
1896         q->nr_hw_queues = set->nr_hw_queues;
1897         q->mq_map = map;
1898
1899         q->queue_ctx = ctx;
1900         q->queue_hw_ctx = hctxs;
1901
1902         q->mq_ops = set->ops;
1903         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1904
1905         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1906                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1907
1908         q->sg_reserved_size = INT_MAX;
1909
1910         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1911         INIT_LIST_HEAD(&q->requeue_list);
1912         spin_lock_init(&q->requeue_lock);
1913
1914         if (q->nr_hw_queues > 1)
1915                 blk_queue_make_request(q, blk_mq_make_request);
1916         else
1917                 blk_queue_make_request(q, blk_sq_make_request);
1918
1919         if (set->timeout)
1920                 blk_queue_rq_timeout(q, set->timeout);
1921
1922         /*
1923          * Do this after blk_queue_make_request() overrides it...
1924          */
1925         q->nr_requests = set->queue_depth;
1926
1927         if (set->ops->complete)
1928                 blk_queue_softirq_done(q, set->ops->complete);
1929
1930         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1931
1932         if (blk_mq_init_hw_queues(q, set))
1933                 goto err_hw;
1934
1935         mutex_lock(&all_q_mutex);
1936         list_add_tail(&q->all_q_node, &all_q_list);
1937         mutex_unlock(&all_q_mutex);
1938
1939         blk_mq_add_queue_tag_set(set, q);
1940
1941         blk_mq_map_swqueue(q);
1942
1943         return q;
1944
1945 err_hw:
1946         blk_cleanup_queue(q);
1947 err_hctxs:
1948         kfree(map);
1949         for (i = 0; i < set->nr_hw_queues; i++) {
1950                 if (!hctxs[i])
1951                         break;
1952                 free_cpumask_var(hctxs[i]->cpumask);
1953                 kfree(hctxs[i]);
1954         }
1955 err_map:
1956         kfree(hctxs);
1957 err_percpu:
1958         free_percpu(ctx);
1959         return ERR_PTR(-ENOMEM);
1960 }
1961 EXPORT_SYMBOL(blk_mq_init_queue);
1962
1963 void blk_mq_free_queue(struct request_queue *q)
1964 {
1965         struct blk_mq_tag_set   *set = q->tag_set;
1966
1967         blk_mq_del_queue_tag_set(q);
1968
1969         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1970         blk_mq_free_hw_queues(q, set);
1971
1972         percpu_ref_exit(&q->mq_usage_counter);
1973
1974         free_percpu(q->queue_ctx);
1975         kfree(q->queue_hw_ctx);
1976         kfree(q->mq_map);
1977
1978         q->queue_ctx = NULL;
1979         q->queue_hw_ctx = NULL;
1980         q->mq_map = NULL;
1981
1982         mutex_lock(&all_q_mutex);
1983         list_del_init(&q->all_q_node);
1984         mutex_unlock(&all_q_mutex);
1985 }
1986
1987 /* Basically redo blk_mq_init_queue with queue frozen */
1988 static void blk_mq_queue_reinit(struct request_queue *q)
1989 {
1990         WARN_ON_ONCE(!q->mq_freeze_depth);
1991
1992         blk_mq_sysfs_unregister(q);
1993
1994         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1995
1996         /*
1997          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1998          * we should change hctx numa_node according to new topology (this
1999          * involves free and re-allocate memory, worthy doing?)
2000          */
2001
2002         blk_mq_map_swqueue(q);
2003
2004         blk_mq_sysfs_register(q);
2005 }
2006
2007 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2008                                       unsigned long action, void *hcpu)
2009 {
2010         struct request_queue *q;
2011
2012         /*
2013          * Before new mappings are established, hotadded cpu might already
2014          * start handling requests. This doesn't break anything as we map
2015          * offline CPUs to first hardware queue. We will re-init the queue
2016          * below to get optimal settings.
2017          */
2018         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2019             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2020                 return NOTIFY_OK;
2021
2022         mutex_lock(&all_q_mutex);
2023
2024         /*
2025          * We need to freeze and reinit all existing queues.  Freezing
2026          * involves synchronous wait for an RCU grace period and doing it
2027          * one by one may take a long time.  Start freezing all queues in
2028          * one swoop and then wait for the completions so that freezing can
2029          * take place in parallel.
2030          */
2031         list_for_each_entry(q, &all_q_list, all_q_node)
2032                 blk_mq_freeze_queue_start(q);
2033         list_for_each_entry(q, &all_q_list, all_q_node)
2034                 blk_mq_freeze_queue_wait(q);
2035
2036         list_for_each_entry(q, &all_q_list, all_q_node)
2037                 blk_mq_queue_reinit(q);
2038
2039         list_for_each_entry(q, &all_q_list, all_q_node)
2040                 blk_mq_unfreeze_queue(q);
2041
2042         mutex_unlock(&all_q_mutex);
2043         return NOTIFY_OK;
2044 }
2045
2046 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2047 {
2048         int i;
2049
2050         for (i = 0; i < set->nr_hw_queues; i++) {
2051                 set->tags[i] = blk_mq_init_rq_map(set, i);
2052                 if (!set->tags[i])
2053                         goto out_unwind;
2054         }
2055
2056         return 0;
2057
2058 out_unwind:
2059         while (--i >= 0)
2060                 blk_mq_free_rq_map(set, set->tags[i], i);
2061
2062         return -ENOMEM;
2063 }
2064
2065 /*
2066  * Allocate the request maps associated with this tag_set. Note that this
2067  * may reduce the depth asked for, if memory is tight. set->queue_depth
2068  * will be updated to reflect the allocated depth.
2069  */
2070 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2071 {
2072         unsigned int depth;
2073         int err;
2074
2075         depth = set->queue_depth;
2076         do {
2077                 err = __blk_mq_alloc_rq_maps(set);
2078                 if (!err)
2079                         break;
2080
2081                 set->queue_depth >>= 1;
2082                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2083                         err = -ENOMEM;
2084                         break;
2085                 }
2086         } while (set->queue_depth);
2087
2088         if (!set->queue_depth || err) {
2089                 pr_err("blk-mq: failed to allocate request map\n");
2090                 return -ENOMEM;
2091         }
2092
2093         if (depth != set->queue_depth)
2094                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2095                                                 depth, set->queue_depth);
2096
2097         return 0;
2098 }
2099
2100 /*
2101  * Alloc a tag set to be associated with one or more request queues.
2102  * May fail with EINVAL for various error conditions. May adjust the
2103  * requested depth down, if if it too large. In that case, the set
2104  * value will be stored in set->queue_depth.
2105  */
2106 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2107 {
2108         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2109
2110         if (!set->nr_hw_queues)
2111                 return -EINVAL;
2112         if (!set->queue_depth)
2113                 return -EINVAL;
2114         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2115                 return -EINVAL;
2116
2117         if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
2118                 return -EINVAL;
2119
2120         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2121                 pr_info("blk-mq: reduced tag depth to %u\n",
2122                         BLK_MQ_MAX_DEPTH);
2123                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2124         }
2125
2126         /*
2127          * If a crashdump is active, then we are potentially in a very
2128          * memory constrained environment. Limit us to 1 queue and
2129          * 64 tags to prevent using too much memory.
2130          */
2131         if (is_kdump_kernel()) {
2132                 set->nr_hw_queues = 1;
2133                 set->queue_depth = min(64U, set->queue_depth);
2134         }
2135
2136         set->tags = kmalloc_node(set->nr_hw_queues *
2137                                  sizeof(struct blk_mq_tags *),
2138                                  GFP_KERNEL, set->numa_node);
2139         if (!set->tags)
2140                 return -ENOMEM;
2141
2142         if (blk_mq_alloc_rq_maps(set))
2143                 goto enomem;
2144
2145         mutex_init(&set->tag_list_lock);
2146         INIT_LIST_HEAD(&set->tag_list);
2147
2148         return 0;
2149 enomem:
2150         kfree(set->tags);
2151         set->tags = NULL;
2152         return -ENOMEM;
2153 }
2154 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2155
2156 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2157 {
2158         int i;
2159
2160         for (i = 0; i < set->nr_hw_queues; i++) {
2161                 if (set->tags[i])
2162                         blk_mq_free_rq_map(set, set->tags[i], i);
2163         }
2164
2165         kfree(set->tags);
2166         set->tags = NULL;
2167 }
2168 EXPORT_SYMBOL(blk_mq_free_tag_set);
2169
2170 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2171 {
2172         struct blk_mq_tag_set *set = q->tag_set;
2173         struct blk_mq_hw_ctx *hctx;
2174         int i, ret;
2175
2176         if (!set || nr > set->queue_depth)
2177                 return -EINVAL;
2178
2179         ret = 0;
2180         queue_for_each_hw_ctx(q, hctx, i) {
2181                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2182                 if (ret)
2183                         break;
2184         }
2185
2186         if (!ret)
2187                 q->nr_requests = nr;
2188
2189         return ret;
2190 }
2191
2192 void blk_mq_disable_hotplug(void)
2193 {
2194         mutex_lock(&all_q_mutex);
2195 }
2196
2197 void blk_mq_enable_hotplug(void)
2198 {
2199         mutex_unlock(&all_q_mutex);
2200 }
2201
2202 static int __init blk_mq_init(void)
2203 {
2204         blk_mq_cpu_init();
2205
2206         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2207
2208         return 0;
2209 }
2210 subsys_initcall(blk_mq_init);