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.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, gfp_t gfp)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 if (!(gfp & __GFP_WAIT))
  89                         return -EBUSY;
  90
  91                 ret = wait_event_interruptible(q->mq_freeze_wq,
  92                                 !q->mq_freeze_depth || blk_queue_dying(q));
  93                 if (blk_queue_dying(q))
  94                         return -ENODEV;
  95                 if (ret)
  96                         return ret;
  97         }
  98 }
  99
 100 static void blk_mq_queue_exit(struct request_queue *q)
 101 {
 102         percpu_ref_put(&q->mq_usage_counter);
 103 }
 104
 105 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 106 {
 107         struct request_queue *q =
 108                 container_of(ref, struct request_queue, mq_usage_counter);
 109
 110         wake_up_all(&q->mq_freeze_wq);
 111 }
 112
 113 void blk_mq_freeze_queue_start(struct request_queue *q)
 114 {
 115         bool freeze;
 116
 117         spin_lock_irq(q->queue_lock);
 118         freeze = !q->mq_freeze_depth++;
 119         spin_unlock_irq(q->queue_lock);
 120
 121         if (freeze) {
 122                 percpu_ref_kill(&q->mq_usage_counter);
 123                 blk_mq_run_hw_queues(q, false);
 124         }
 125 }
 126 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 127
 128 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 129 {
 130         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 131 }
 132
 133 /*
 134  * Guarantee no request is in use, so we can change any data structure of
 135  * the queue afterward.
 136  */
 137 void blk_mq_freeze_queue(struct request_queue *q)
 138 {
 139         blk_mq_freeze_queue_start(q);
 140         blk_mq_freeze_queue_wait(q);
 141 }
 142 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 143
 144 void blk_mq_unfreeze_queue(struct request_queue *q)
 145 {
 146         bool wake;
 147
 148         spin_lock_irq(q->queue_lock);
 149         wake = !--q->mq_freeze_depth;
 150         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 151         spin_unlock_irq(q->queue_lock);
 152         if (wake) {
 153                 percpu_ref_reinit(&q->mq_usage_counter);
 154                 wake_up_all(&q->mq_freeze_wq);
 155         }
 156 }
 157 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 158
 159 void blk_mq_wake_waiters(struct request_queue *q)
 160 {
 161         struct blk_mq_hw_ctx *hctx;
 162         unsigned int i;
 163
 164         queue_for_each_hw_ctx(q, hctx, i)
 165                 if (blk_mq_hw_queue_mapped(hctx))
 166                         blk_mq_tag_wakeup_all(hctx->tags, true);
 167
 168         /*
 169          * If we are called because the queue has now been marked as
 170          * dying, we need to ensure that processes currently waiting on
 171          * the queue are notified as well.
 172          */
 173         wake_up_all(&q->mq_freeze_wq);
 174 }
 175
 176 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 177 {
 178         return blk_mq_has_free_tags(hctx->tags);
 179 }
 180 EXPORT_SYMBOL(blk_mq_can_queue);
 181
 182 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 183                                struct request *rq, unsigned int rw_flags)
 184 {
 185         if (blk_queue_io_stat(q))
 186                 rw_flags |= REQ_IO_STAT;
 187
 188         INIT_LIST_HEAD(&rq->queuelist);
 189         /* csd/requeue_work/fifo_time is initialized before use */
 190         rq->q = q;
 191         rq->mq_ctx = ctx;
 192         rq->cmd_flags |= rw_flags;
 193         /* do not touch atomic flags, it needs atomic ops against the timer */
 194         rq->cpu = -1;
 195         INIT_HLIST_NODE(&rq->hash);
 196         RB_CLEAR_NODE(&rq->rb_node);
 197         rq->rq_disk = NULL;
 198         rq->part = NULL;
 199         rq->start_time = jiffies;
 200 #ifdef CONFIG_BLK_CGROUP
 201         rq->rl = NULL;
 202         set_start_time_ns(rq);
 203         rq->io_start_time_ns = 0;
 204 #endif
 205         rq->nr_phys_segments = 0;
 206 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 207         rq->nr_integrity_segments = 0;
 208 #endif
 209         rq->special = NULL;
 210         /* tag was already set */
 211         rq->errors = 0;
 212
 213         rq->cmd = rq->__cmd;
 214
 215         rq->extra_len = 0;
 216         rq->sense_len = 0;
 217         rq->resid_len = 0;
 218         rq->sense = NULL;
 219
 220         INIT_LIST_HEAD(&rq->timeout_list);
 221         rq->timeout = 0;
 222
 223         rq->end_io = NULL;
 224         rq->end_io_data = NULL;
 225         rq->next_rq = NULL;
 226
 227         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 228 }
 229
 230 static struct request *
 231 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 232 {
 233         struct request *rq;
 234         unsigned int tag;
 235
 236         tag = blk_mq_get_tag(data);
 237         if (tag != BLK_MQ_TAG_FAIL) {
 238                 rq = data->hctx->tags->rqs[tag];
 239
 240                 if (blk_mq_tag_busy(data->hctx)) {
 241                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 242                         atomic_inc(&data->hctx->nr_active);
 243                 }
 244
 245                 rq->tag = tag;
 246                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 247                 return rq;
 248         }
 249
 250         return NULL;
 251 }
 252
 253 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 254                 bool reserved)
 255 {
 256         struct blk_mq_ctx *ctx;
 257         struct blk_mq_hw_ctx *hctx;
 258         struct request *rq;
 259         struct blk_mq_alloc_data alloc_data;
 260         int ret;
 261
 262         ret = blk_mq_queue_enter(q, gfp);
 263         if (ret)
 264                 return ERR_PTR(ret);
 265
 266         ctx = blk_mq_get_ctx(q);
 267         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 268         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 269                         reserved, ctx, hctx);
 270
 271         rq = __blk_mq_alloc_request(&alloc_data, rw);
 272         if (!rq && (gfp & __GFP_WAIT)) {
 273                 __blk_mq_run_hw_queue(hctx);
 274                 blk_mq_put_ctx(ctx);
 275
 276                 ctx = blk_mq_get_ctx(q);
 277                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 278                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 279                                 hctx);
 280                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 281                 ctx = alloc_data.ctx;
 282         }
 283         blk_mq_put_ctx(ctx);
 284         if (!rq) {
 285                 blk_mq_queue_exit(q);
 286                 return ERR_PTR(-EWOULDBLOCK);
 287         }
 288         return rq;
 289 }
 290 EXPORT_SYMBOL(blk_mq_alloc_request);
 291
 292 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 293                                   struct blk_mq_ctx *ctx, struct request *rq)
 294 {
 295         const int tag = rq->tag;
 296         struct request_queue *q = rq->q;
 297
 298         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 299                 atomic_dec(&hctx->nr_active);
 300         rq->cmd_flags = 0;
 301
 302         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 303         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 304         blk_mq_queue_exit(q);
 305 }
 306
 307 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 308 {
 309         struct blk_mq_ctx *ctx = rq->mq_ctx;
 310
 311         ctx->rq_completed[rq_is_sync(rq)]++;
 312         __blk_mq_free_request(hctx, ctx, rq);
 313
 314 }
 315 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 316
 317 void blk_mq_free_request(struct request *rq)
 318 {
 319         struct blk_mq_hw_ctx *hctx;
 320         struct request_queue *q = rq->q;
 321
 322         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 323         blk_mq_free_hctx_request(hctx, rq);
 324 }
 325 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 326
 327 inline void __blk_mq_end_request(struct request *rq, int error)
 328 {
 329         blk_account_io_done(rq);
 330
 331         if (rq->end_io) {
 332                 rq->end_io(rq, error);
 333         } else {
 334                 if (unlikely(blk_bidi_rq(rq)))
 335                         blk_mq_free_request(rq->next_rq);
 336                 blk_mq_free_request(rq);
 337         }
 338 }
 339 EXPORT_SYMBOL(__blk_mq_end_request);
 340
 341 void blk_mq_end_request(struct request *rq, int error)
 342 {
 343         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 344                 BUG();
 345         __blk_mq_end_request(rq, error);
 346 }
 347 EXPORT_SYMBOL(blk_mq_end_request);
 348
 349 static void __blk_mq_complete_request_remote(void *data)
 350 {
 351         struct request *rq = data;
 352
 353         rq->q->softirq_done_fn(rq);
 354 }
 355
 356 static void blk_mq_ipi_complete_request(struct request *rq)
 357 {
 358         struct blk_mq_ctx *ctx = rq->mq_ctx;
 359         bool shared = false;
 360         int cpu;
 361
 362         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 363                 rq->q->softirq_done_fn(rq);
 364                 return;
 365         }
 366
 367         cpu = get_cpu();
 368         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 369                 shared = cpus_share_cache(cpu, ctx->cpu);
 370
 371         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 372                 rq->csd.func = __blk_mq_complete_request_remote;
 373                 rq->csd.info = rq;
 374                 rq->csd.flags = 0;
 375                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 376         } else {
 377                 rq->q->softirq_done_fn(rq);
 378         }
 379         put_cpu();
 380 }
 381
 382 void __blk_mq_complete_request(struct request *rq)
 383 {
 384         struct request_queue *q = rq->q;
 385
 386         if (!q->softirq_done_fn)
 387                 blk_mq_end_request(rq, rq->errors);
 388         else
 389                 blk_mq_ipi_complete_request(rq);
 390 }
 391
 392 /**
 393  * blk_mq_complete_request - end I/O on a request
 394  * @rq:         the request being processed
 395  *
 396  * Description:
 397  *      Ends all I/O on a request. It does not handle partial completions.
 398  *      The actual completion happens out-of-order, through a IPI handler.
 399  **/
 400 void blk_mq_complete_request(struct request *rq)
 401 {
 402         struct request_queue *q = rq->q;
 403
 404         if (unlikely(blk_should_fake_timeout(q)))
 405                 return;
 406         if (!blk_mark_rq_complete(rq))
 407                 __blk_mq_complete_request(rq);
 408 }
 409 EXPORT_SYMBOL(blk_mq_complete_request);
 410
 411 int blk_mq_request_started(struct request *rq)
 412 {
 413         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 414 }
 415 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 416
 417 void blk_mq_start_request(struct request *rq)
 418 {
 419         struct request_queue *q = rq->q;
 420
 421         trace_block_rq_issue(q, rq);
 422
 423         rq->resid_len = blk_rq_bytes(rq);
 424         if (unlikely(blk_bidi_rq(rq)))
 425                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 426
 427         blk_add_timer(rq);
 428
 429         /*
 430          * Ensure that ->deadline is visible before set the started
 431          * flag and clear the completed flag.
 432          */
 433         smp_mb__before_atomic();
 434
 435         /*
 436          * Mark us as started and clear complete. Complete might have been
 437          * set if requeue raced with timeout, which then marked it as
 438          * complete. So be sure to clear complete again when we start
 439          * the request, otherwise we'll ignore the completion event.
 440          */
 441         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 442                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 443         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 444                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 445
 446         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 447                 /*
 448                  * Make sure space for the drain appears.  We know we can do
 449                  * this because max_hw_segments has been adjusted to be one
 450                  * fewer than the device can handle.
 451                  */
 452                 rq->nr_phys_segments++;
 453         }
 454 }
 455 EXPORT_SYMBOL(blk_mq_start_request);
 456
 457 static void __blk_mq_requeue_request(struct request *rq)
 458 {
 459         struct request_queue *q = rq->q;
 460
 461         trace_block_rq_requeue(q, rq);
 462
 463         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 464                 if (q->dma_drain_size && blk_rq_bytes(rq))
 465                         rq->nr_phys_segments--;
 466         }
 467 }
 468
 469 void blk_mq_requeue_request(struct request *rq)
 470 {
 471         __blk_mq_requeue_request(rq);
 472
 473         BUG_ON(blk_queued_rq(rq));
 474         blk_mq_add_to_requeue_list(rq, true);
 475 }
 476 EXPORT_SYMBOL(blk_mq_requeue_request);
 477
 478 static void blk_mq_requeue_work(struct work_struct *work)
 479 {
 480         struct request_queue *q =
 481                 container_of(work, struct request_queue, requeue_work);
 482         LIST_HEAD(rq_list);
 483         struct request *rq, *next;
 484         unsigned long flags;
 485
 486         spin_lock_irqsave(&q->requeue_lock, flags);
 487         list_splice_init(&q->requeue_list, &rq_list);
 488         spin_unlock_irqrestore(&q->requeue_lock, flags);
 489
 490         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 491                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 492                         continue;
 493
 494                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 495                 list_del_init(&rq->queuelist);
 496                 blk_mq_insert_request(rq, true, false, false);
 497         }
 498
 499         while (!list_empty(&rq_list)) {
 500                 rq = list_entry(rq_list.next, struct request, queuelist);
 501                 list_del_init(&rq->queuelist);
 502                 blk_mq_insert_request(rq, false, false, false);
 503         }
 504
 505         /*
 506          * Use the start variant of queue running here, so that running
 507          * the requeue work will kick stopped queues.
 508          */
 509         blk_mq_start_hw_queues(q);
 510 }
 511
 512 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 513 {
 514         struct request_queue *q = rq->q;
 515         unsigned long flags;
 516
 517         /*
 518          * We abuse this flag that is otherwise used by the I/O scheduler to
 519          * request head insertation from the workqueue.
 520          */
 521         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 522
 523         spin_lock_irqsave(&q->requeue_lock, flags);
 524         if (at_head) {
 525                 rq->cmd_flags |= REQ_SOFTBARRIER;
 526                 list_add(&rq->queuelist, &q->requeue_list);
 527         } else {
 528                 list_add_tail(&rq->queuelist, &q->requeue_list);
 529         }
 530         spin_unlock_irqrestore(&q->requeue_lock, flags);
 531 }
 532 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 533
 534 void blk_mq_cancel_requeue_work(struct request_queue *q)
 535 {
 536         cancel_work_sync(&q->requeue_work);
 537 }
 538 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 539
 540 void blk_mq_kick_requeue_list(struct request_queue *q)
 541 {
 542         kblockd_schedule_work(&q->requeue_work);
 543 }
 544 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 545
 546 void blk_mq_abort_requeue_list(struct request_queue *q)
 547 {
 548         unsigned long flags;
 549         LIST_HEAD(rq_list);
 550
 551         spin_lock_irqsave(&q->requeue_lock, flags);
 552         list_splice_init(&q->requeue_list, &rq_list);
 553         spin_unlock_irqrestore(&q->requeue_lock, flags);
 554
 555         while (!list_empty(&rq_list)) {
 556                 struct request *rq;
 557
 558                 rq = list_first_entry(&rq_list, struct request, queuelist);
 559                 list_del_init(&rq->queuelist);
 560                 rq->errors = -EIO;
 561                 blk_mq_end_request(rq, rq->errors);
 562         }
 563 }
 564 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 565
 566 static inline bool is_flush_request(struct request *rq,
 567                 struct blk_flush_queue *fq, unsigned int tag)
 568 {
 569         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 570                         fq->flush_rq->tag == tag);
 571 }
 572
 573 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 574 {
 575         struct request *rq = tags->rqs[tag];
 576         /* mq_ctx of flush rq is always cloned from the corresponding req */
 577         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 578
 579         if (!is_flush_request(rq, fq, tag))
 580                 return rq;
 581
 582         return fq->flush_rq;
 583 }
 584 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 585
 586 struct blk_mq_timeout_data {
 587         unsigned long next;
 588         unsigned int next_set;
 589 };
 590
 591 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 592 {
 593         struct blk_mq_ops *ops = req->q->mq_ops;
 594         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 595
 596         /*
 597          * We know that complete is set at this point. If STARTED isn't set
 598          * anymore, then the request isn't active and the "timeout" should
 599          * just be ignored. This can happen due to the bitflag ordering.
 600          * Timeout first checks if STARTED is set, and if it is, assumes
 601          * the request is active. But if we race with completion, then
 602          * we both flags will get cleared. So check here again, and ignore
 603          * a timeout event with a request that isn't active.
 604          */
 605         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 606                 return;
 607
 608         if (ops->timeout)
 609                 ret = ops->timeout(req, reserved);
 610
 611         switch (ret) {
 612         case BLK_EH_HANDLED:
 613                 __blk_mq_complete_request(req);
 614                 break;
 615         case BLK_EH_RESET_TIMER:
 616                 blk_add_timer(req);
 617                 blk_clear_rq_complete(req);
 618                 break;
 619         case BLK_EH_NOT_HANDLED:
 620                 break;
 621         default:
 622                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 623                 break;
 624         }
 625 }
 626
 627 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 628                 struct request *rq, void *priv, bool reserved)
 629 {
 630         struct blk_mq_timeout_data *data = priv;
 631
 632         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 633                 /*
 634                  * If a request wasn't started before the queue was
 635                  * marked dying, kill it here or it'll go unnoticed.
 636                  */
 637                 if (unlikely(blk_queue_dying(rq->q))) {
 638                         rq->errors = -EIO;
 639                         blk_mq_complete_request(rq);
 640                 }
 641                 return;
 642         }
 643         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 644                 return;
 645
 646         if (time_after_eq(jiffies, rq->deadline)) {
 647                 if (!blk_mark_rq_complete(rq))
 648                         blk_mq_rq_timed_out(rq, reserved);
 649         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 650                 data->next = rq->deadline;
 651                 data->next_set = 1;
 652         }
 653 }
 654
 655 static void blk_mq_rq_timer(unsigned long priv)
 656 {
 657         struct request_queue *q = (struct request_queue *)priv;
 658         struct blk_mq_timeout_data data = {
 659                 .next           = 0,
 660                 .next_set       = 0,
 661         };
 662         struct blk_mq_hw_ctx *hctx;
 663         int i;
 664
 665         queue_for_each_hw_ctx(q, hctx, i) {
 666                 /*
 667                  * If not software queues are currently mapped to this
 668                  * hardware queue, there's nothing to check
 669                  */
 670                 if (!blk_mq_hw_queue_mapped(hctx))
 671                         continue;
 672
 673                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 674         }
 675
 676         if (data.next_set) {
 677                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 678                 mod_timer(&q->timeout, data.next);
 679         } else {
 680                 queue_for_each_hw_ctx(q, hctx, i) {
 681                         /* the hctx may be unmapped, so check it here */
 682                         if (blk_mq_hw_queue_mapped(hctx))
 683                                 blk_mq_tag_idle(hctx);
 684                 }
 685         }
 686 }
 687
 688 /*
 689  * Reverse check our software queue for entries that we could potentially
 690  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 691  * too much time checking for merges.
 692  */
 693 static bool blk_mq_attempt_merge(struct request_queue *q,
 694                                  struct blk_mq_ctx *ctx, struct bio *bio)
 695 {
 696         struct request *rq;
 697         int checked = 8;
 698
 699         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 700                 int el_ret;
 701
 702                 if (!checked--)
 703                         break;
 704
 705                 if (!blk_rq_merge_ok(rq, bio))
 706                         continue;
 707
 708                 el_ret = blk_try_merge(rq, bio);
 709                 if (el_ret == ELEVATOR_BACK_MERGE) {
 710                         if (bio_attempt_back_merge(q, rq, bio)) {
 711                                 ctx->rq_merged++;
 712                                 return true;
 713                         }
 714                         break;
 715                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 716                         if (bio_attempt_front_merge(q, rq, bio)) {
 717                                 ctx->rq_merged++;
 718                                 return true;
 719                         }
 720                         break;
 721                 }
 722         }
 723
 724         return false;
 725 }
 726
 727 /*
 728  * Process software queues that have been marked busy, splicing them
 729  * to the for-dispatch
 730  */
 731 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 732 {
 733         struct blk_mq_ctx *ctx;
 734         int i;
 735
 736         for (i = 0; i < hctx->ctx_map.size; i++) {
 737                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 738                 unsigned int off, bit;
 739
 740                 if (!bm->word)
 741                         continue;
 742
 743                 bit = 0;
 744                 off = i * hctx->ctx_map.bits_per_word;
 745                 do {
 746                         bit = find_next_bit(&bm->word, bm->depth, bit);
 747                         if (bit >= bm->depth)
 748                                 break;
 749
 750                         ctx = hctx->ctxs[bit + off];
 751                         clear_bit(bit, &bm->word);
 752                         spin_lock(&ctx->lock);
 753                         list_splice_tail_init(&ctx->rq_list, list);
 754                         spin_unlock(&ctx->lock);
 755
 756                         bit++;
 757                 } while (1);
 758         }
 759 }
 760
 761 /*
 762  * Run this hardware queue, pulling any software queues mapped to it in.
 763  * Note that this function currently has various problems around ordering
 764  * of IO. In particular, we'd like FIFO behaviour on handling existing
 765  * items on the hctx->dispatch list. Ignore that for now.
 766  */
 767 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 768 {
 769         struct request_queue *q = hctx->queue;
 770         struct request *rq;
 771         LIST_HEAD(rq_list);
 772         LIST_HEAD(driver_list);
 773         struct list_head *dptr;
 774         int queued;
 775
 776         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 777
 778         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 779                 return;
 780
 781         hctx->run++;
 782
 783         /*
 784          * Touch any software queue that has pending entries.
 785          */
 786         flush_busy_ctxs(hctx, &rq_list);
 787
 788         /*
 789          * If we have previous entries on our dispatch list, grab them
 790          * and stuff them at the front for more fair dispatch.
 791          */
 792         if (!list_empty_careful(&hctx->dispatch)) {
 793                 spin_lock(&hctx->lock);
 794                 if (!list_empty(&hctx->dispatch))
 795                         list_splice_init(&hctx->dispatch, &rq_list);
 796                 spin_unlock(&hctx->lock);
 797         }
 798
 799         /*
 800          * Start off with dptr being NULL, so we start the first request
 801          * immediately, even if we have more pending.
 802          */
 803         dptr = NULL;
 804
 805         /*
 806          * Now process all the entries, sending them to the driver.
 807          */
 808         queued = 0;
 809         while (!list_empty(&rq_list)) {
 810                 struct blk_mq_queue_data bd;
 811                 int ret;
 812
 813                 rq = list_first_entry(&rq_list, struct request, queuelist);
 814                 list_del_init(&rq->queuelist);
 815
 816                 bd.rq = rq;
 817                 bd.list = dptr;
 818                 bd.last = list_empty(&rq_list);
 819
 820                 ret = q->mq_ops->queue_rq(hctx, &bd);
 821                 switch (ret) {
 822                 case BLK_MQ_RQ_QUEUE_OK:
 823                         queued++;
 824                         continue;
 825                 case BLK_MQ_RQ_QUEUE_BUSY:
 826                         list_add(&rq->queuelist, &rq_list);
 827                         __blk_mq_requeue_request(rq);
 828                         break;
 829                 default:
 830                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 831                 case BLK_MQ_RQ_QUEUE_ERROR:
 832                         rq->errors = -EIO;
 833                         blk_mq_end_request(rq, rq->errors);
 834                         break;
 835                 }
 836
 837                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 838                         break;
 839
 840                 /*
 841                  * We've done the first request. If we have more than 1
 842                  * left in the list, set dptr to defer issue.
 843                  */
 844                 if (!dptr && rq_list.next != rq_list.prev)
 845                         dptr = &driver_list;
 846         }
 847
 848         if (!queued)
 849                 hctx->dispatched[0]++;
 850         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 851                 hctx->dispatched[ilog2(queued) + 1]++;
 852
 853         /*
 854          * Any items that need requeuing? Stuff them into hctx->dispatch,
 855          * that is where we will continue on next queue run.
 856          */
 857         if (!list_empty(&rq_list)) {
 858                 spin_lock(&hctx->lock);
 859                 list_splice(&rq_list, &hctx->dispatch);
 860                 spin_unlock(&hctx->lock);
 861         }
 862 }
 863
 864 /*
 865  * It'd be great if the workqueue API had a way to pass
 866  * in a mask and had some smarts for more clever placement.
 867  * For now we just round-robin here, switching for every
 868  * BLK_MQ_CPU_WORK_BATCH queued items.
 869  */
 870 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 871 {
 872         if (hctx->queue->nr_hw_queues == 1)
 873                 return WORK_CPU_UNBOUND;
 874
 875         if (--hctx->next_cpu_batch <= 0) {
 876                 int cpu = hctx->next_cpu, next_cpu;
 877
 878                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 879                 if (next_cpu >= nr_cpu_ids)
 880                         next_cpu = cpumask_first(hctx->cpumask);
 881
 882                 hctx->next_cpu = next_cpu;
 883                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 884
 885                 return cpu;
 886         }
 887
 888         return hctx->next_cpu;
 889 }
 890
 891 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 892 {
 893         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 894             !blk_mq_hw_queue_mapped(hctx)))
 895                 return;
 896
 897         if (!async) {
 898                 int cpu = get_cpu();
 899                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 900                         __blk_mq_run_hw_queue(hctx);
 901                         put_cpu();
 902                         return;
 903                 }
 904
 905                 put_cpu();
 906         }
 907
 908         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 909                         &hctx->run_work, 0);
 910 }
 911
 912 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 913 {
 914         struct blk_mq_hw_ctx *hctx;
 915         int i;
 916
 917         queue_for_each_hw_ctx(q, hctx, i) {
 918                 if ((!blk_mq_hctx_has_pending(hctx) &&
 919                     list_empty_careful(&hctx->dispatch)) ||
 920                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 921                         continue;
 922
 923                 blk_mq_run_hw_queue(hctx, async);
 924         }
 925 }
 926 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 927
 928 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 929 {
 930         cancel_delayed_work(&hctx->run_work);
 931         cancel_delayed_work(&hctx->delay_work);
 932         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 933 }
 934 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 935
 936 void blk_mq_stop_hw_queues(struct request_queue *q)
 937 {
 938         struct blk_mq_hw_ctx *hctx;
 939         int i;
 940
 941         queue_for_each_hw_ctx(q, hctx, i)
 942                 blk_mq_stop_hw_queue(hctx);
 943 }
 944 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 945
 946 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 947 {
 948         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 949
 950         blk_mq_run_hw_queue(hctx, false);
 951 }
 952 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 953
 954 void blk_mq_start_hw_queues(struct request_queue *q)
 955 {
 956         struct blk_mq_hw_ctx *hctx;
 957         int i;
 958
 959         queue_for_each_hw_ctx(q, hctx, i)
 960                 blk_mq_start_hw_queue(hctx);
 961 }
 962 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 963
 964 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 965 {
 966         struct blk_mq_hw_ctx *hctx;
 967         int i;
 968
 969         queue_for_each_hw_ctx(q, hctx, i) {
 970                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 971                         continue;
 972
 973                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 974                 blk_mq_run_hw_queue(hctx, async);
 975         }
 976 }
 977 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 978
 979 static void blk_mq_run_work_fn(struct work_struct *work)
 980 {
 981         struct blk_mq_hw_ctx *hctx;
 982
 983         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 984
 985         __blk_mq_run_hw_queue(hctx);
 986 }
 987
 988 static void blk_mq_delay_work_fn(struct work_struct *work)
 989 {
 990         struct blk_mq_hw_ctx *hctx;
 991
 992         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 993
 994         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 995                 __blk_mq_run_hw_queue(hctx);
 996 }
 997
 998 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 999 {
1000         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1001                 return;
1002
1003         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1004                         &hctx->delay_work, msecs_to_jiffies(msecs));
1005 }
1006 EXPORT_SYMBOL(blk_mq_delay_queue);
1007
1008 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1009                                     struct request *rq, bool at_head)
1010 {
1011         struct blk_mq_ctx *ctx = rq->mq_ctx;
1012
1013         trace_block_rq_insert(hctx->queue, rq);
1014
1015         if (at_head)
1016                 list_add(&rq->queuelist, &ctx->rq_list);
1017         else
1018                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1019
1020         blk_mq_hctx_mark_pending(hctx, ctx);
1021 }
1022
1023 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1024                 bool async)
1025 {
1026         struct request_queue *q = rq->q;
1027         struct blk_mq_hw_ctx *hctx;
1028         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1029
1030         current_ctx = blk_mq_get_ctx(q);
1031         if (!cpu_online(ctx->cpu))
1032                 rq->mq_ctx = ctx = current_ctx;
1033
1034         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1035
1036         spin_lock(&ctx->lock);
1037         __blk_mq_insert_request(hctx, rq, at_head);
1038         spin_unlock(&ctx->lock);
1039
1040         if (run_queue)
1041                 blk_mq_run_hw_queue(hctx, async);
1042
1043         blk_mq_put_ctx(current_ctx);
1044 }
1045
1046 static void blk_mq_insert_requests(struct request_queue *q,
1047                                      struct blk_mq_ctx *ctx,
1048                                      struct list_head *list,
1049                                      int depth,
1050                                      bool from_schedule)
1051
1052 {
1053         struct blk_mq_hw_ctx *hctx;
1054         struct blk_mq_ctx *current_ctx;
1055
1056         trace_block_unplug(q, depth, !from_schedule);
1057
1058         current_ctx = blk_mq_get_ctx(q);
1059
1060         if (!cpu_online(ctx->cpu))
1061                 ctx = current_ctx;
1062         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1063
1064         /*
1065          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1066          * offline now
1067          */
1068         spin_lock(&ctx->lock);
1069         while (!list_empty(list)) {
1070                 struct request *rq;
1071
1072                 rq = list_first_entry(list, struct request, queuelist);
1073                 list_del_init(&rq->queuelist);
1074                 rq->mq_ctx = ctx;
1075                 __blk_mq_insert_request(hctx, rq, false);
1076         }
1077         spin_unlock(&ctx->lock);
1078
1079         blk_mq_run_hw_queue(hctx, from_schedule);
1080         blk_mq_put_ctx(current_ctx);
1081 }
1082
1083 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1084 {
1085         struct request *rqa = container_of(a, struct request, queuelist);
1086         struct request *rqb = container_of(b, struct request, queuelist);
1087
1088         return !(rqa->mq_ctx < rqb->mq_ctx ||
1089                  (rqa->mq_ctx == rqb->mq_ctx &&
1090                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1091 }
1092
1093 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1094 {
1095         struct blk_mq_ctx *this_ctx;
1096         struct request_queue *this_q;
1097         struct request *rq;
1098         LIST_HEAD(list);
1099         LIST_HEAD(ctx_list);
1100         unsigned int depth;
1101
1102         list_splice_init(&plug->mq_list, &list);
1103
1104         list_sort(NULL, &list, plug_ctx_cmp);
1105
1106         this_q = NULL;
1107         this_ctx = NULL;
1108         depth = 0;
1109
1110         while (!list_empty(&list)) {
1111                 rq = list_entry_rq(list.next);
1112                 list_del_init(&rq->queuelist);
1113                 BUG_ON(!rq->q);
1114                 if (rq->mq_ctx != this_ctx) {
1115                         if (this_ctx) {
1116                                 blk_mq_insert_requests(this_q, this_ctx,
1117                                                         &ctx_list, depth,
1118                                                         from_schedule);
1119                         }
1120
1121                         this_ctx = rq->mq_ctx;
1122                         this_q = rq->q;
1123                         depth = 0;
1124                 }
1125
1126                 depth++;
1127                 list_add_tail(&rq->queuelist, &ctx_list);
1128         }
1129
1130         /*
1131          * If 'this_ctx' is set, we know we have entries to complete
1132          * on 'ctx_list'. Do those.
1133          */
1134         if (this_ctx) {
1135                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1136                                        from_schedule);
1137         }
1138 }
1139
1140 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1141 {
1142         init_request_from_bio(rq, bio);
1143
1144         if (blk_do_io_stat(rq))
1145                 blk_account_io_start(rq, 1);
1146 }
1147
1148 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1149 {
1150         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1151                 !blk_queue_nomerges(hctx->queue);
1152 }
1153
1154 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1155                                          struct blk_mq_ctx *ctx,
1156                                          struct request *rq, struct bio *bio)
1157 {
1158         if (!hctx_allow_merges(hctx)) {
1159                 blk_mq_bio_to_request(rq, bio);
1160                 spin_lock(&ctx->lock);
1161 insert_rq:
1162                 __blk_mq_insert_request(hctx, rq, false);
1163                 spin_unlock(&ctx->lock);
1164                 return false;
1165         } else {
1166                 struct request_queue *q = hctx->queue;
1167
1168                 spin_lock(&ctx->lock);
1169                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1170                         blk_mq_bio_to_request(rq, bio);
1171                         goto insert_rq;
1172                 }
1173
1174                 spin_unlock(&ctx->lock);
1175                 __blk_mq_free_request(hctx, ctx, rq);
1176                 return true;
1177         }
1178 }
1179
1180 struct blk_map_ctx {
1181         struct blk_mq_hw_ctx *hctx;
1182         struct blk_mq_ctx *ctx;
1183 };
1184
1185 static struct request *blk_mq_map_request(struct request_queue *q,
1186                                           struct bio *bio,
1187                                           struct blk_map_ctx *data)
1188 {
1189         struct blk_mq_hw_ctx *hctx;
1190         struct blk_mq_ctx *ctx;
1191         struct request *rq;
1192         int rw = bio_data_dir(bio);
1193         struct blk_mq_alloc_data alloc_data;
1194
1195         if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1196                 bio_endio(bio, -EIO);
1197                 return NULL;
1198         }
1199
1200         ctx = blk_mq_get_ctx(q);
1201         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1202
1203         if (rw_is_sync(bio->bi_rw))
1204                 rw |= REQ_SYNC;
1205
1206         trace_block_getrq(q, bio, rw);
1207         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1208                         hctx);
1209         rq = __blk_mq_alloc_request(&alloc_data, rw);
1210         if (unlikely(!rq)) {
1211                 __blk_mq_run_hw_queue(hctx);
1212                 blk_mq_put_ctx(ctx);
1213                 trace_block_sleeprq(q, bio, rw);
1214
1215                 ctx = blk_mq_get_ctx(q);
1216                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1217                 blk_mq_set_alloc_data(&alloc_data, q,
1218                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1219                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1220                 ctx = alloc_data.ctx;
1221                 hctx = alloc_data.hctx;
1222         }
1223
1224         hctx->queued++;
1225         data->hctx = hctx;
1226         data->ctx = ctx;
1227         return rq;
1228 }
1229
1230 /*
1231  * Multiple hardware queue variant. This will not use per-process plugs,
1232  * but will attempt to bypass the hctx queueing if we can go straight to
1233  * hardware for SYNC IO.
1234  */
1235 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1236 {
1237         const int is_sync = rw_is_sync(bio->bi_rw);
1238         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1239         struct blk_map_ctx data;
1240         struct request *rq;
1241
1242         blk_queue_bounce(q, &bio);
1243
1244         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1245                 bio_endio(bio, -EIO);
1246                 return;
1247         }
1248
1249         rq = blk_mq_map_request(q, bio, &data);
1250         if (unlikely(!rq))
1251                 return;
1252
1253         if (unlikely(is_flush_fua)) {
1254                 blk_mq_bio_to_request(rq, bio);
1255                 blk_insert_flush(rq);
1256                 goto run_queue;
1257         }
1258
1259         /*
1260          * If the driver supports defer issued based on 'last', then
1261          * queue it up like normal since we can potentially save some
1262          * CPU this way.
1263          */
1264         if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1265                 struct blk_mq_queue_data bd = {
1266                         .rq = rq,
1267                         .list = NULL,
1268                         .last = 1
1269                 };
1270                 int ret;
1271
1272                 blk_mq_bio_to_request(rq, bio);
1273
1274                 /*
1275                  * For OK queue, we are done. For error, kill it. Any other
1276                  * error (busy), just add it to our list as we previously
1277                  * would have done
1278                  */
1279                 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1280                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1281                         goto done;
1282                 else {
1283                         __blk_mq_requeue_request(rq);
1284
1285                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1286                                 rq->errors = -EIO;
1287                                 blk_mq_end_request(rq, rq->errors);
1288                                 goto done;
1289                         }
1290                 }
1291         }
1292
1293         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1294                 /*
1295                  * For a SYNC request, send it to the hardware immediately. For
1296                  * an ASYNC request, just ensure that we run it later on. The
1297                  * latter allows for merging opportunities and more efficient
1298                  * dispatching.
1299                  */
1300 run_queue:
1301                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1302         }
1303 done:
1304         blk_mq_put_ctx(data.ctx);
1305 }
1306
1307 /*
1308  * Single hardware queue variant. This will attempt to use any per-process
1309  * plug for merging and IO deferral.
1310  */
1311 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1312 {
1313         const int is_sync = rw_is_sync(bio->bi_rw);
1314         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1315         unsigned int use_plug, request_count = 0;
1316         struct blk_map_ctx data;
1317         struct request *rq;
1318
1319         /*
1320          * If we have multiple hardware queues, just go directly to
1321          * one of those for sync IO.
1322          */
1323         use_plug = !is_flush_fua && !is_sync;
1324
1325         blk_queue_bounce(q, &bio);
1326
1327         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1328                 bio_endio(bio, -EIO);
1329                 return;
1330         }
1331
1332         if (use_plug && !blk_queue_nomerges(q) &&
1333             blk_attempt_plug_merge(q, bio, &request_count))
1334                 return;
1335
1336         rq = blk_mq_map_request(q, bio, &data);
1337         if (unlikely(!rq))
1338                 return;
1339
1340         if (unlikely(is_flush_fua)) {
1341                 blk_mq_bio_to_request(rq, bio);
1342                 blk_insert_flush(rq);
1343                 goto run_queue;
1344         }
1345
1346         /*
1347          * A task plug currently exists. Since this is completely lockless,
1348          * utilize that to temporarily store requests until the task is
1349          * either done or scheduled away.
1350          */
1351         if (use_plug) {
1352                 struct blk_plug *plug = current->plug;
1353
1354                 if (plug) {
1355                         blk_mq_bio_to_request(rq, bio);
1356                         if (list_empty(&plug->mq_list))
1357                                 trace_block_plug(q);
1358                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1359                                 blk_flush_plug_list(plug, false);
1360                                 trace_block_plug(q);
1361                         }
1362                         list_add_tail(&rq->queuelist, &plug->mq_list);
1363                         blk_mq_put_ctx(data.ctx);
1364                         return;
1365                 }
1366         }
1367
1368         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1369                 /*
1370                  * For a SYNC request, send it to the hardware immediately. For
1371                  * an ASYNC request, just ensure that we run it later on. The
1372                  * latter allows for merging opportunities and more efficient
1373                  * dispatching.
1374                  */
1375 run_queue:
1376                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1377         }
1378
1379         blk_mq_put_ctx(data.ctx);
1380 }
1381
1382 /*
1383  * Default mapping to a software queue, since we use one per CPU.
1384  */
1385 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1386 {
1387         return q->queue_hw_ctx[q->mq_map[cpu]];
1388 }
1389 EXPORT_SYMBOL(blk_mq_map_queue);
1390
1391 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1392                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1393 {
1394         struct page *page;
1395
1396         if (tags->rqs && set->ops->exit_request) {
1397                 int i;
1398
1399                 for (i = 0; i < tags->nr_tags; i++) {
1400                         if (!tags->rqs[i])
1401                                 continue;
1402                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1403                                                 hctx_idx, i);
1404                         tags->rqs[i] = NULL;
1405                 }
1406         }
1407
1408         while (!list_empty(&tags->page_list)) {
1409                 page = list_first_entry(&tags->page_list, struct page, lru);
1410                 list_del_init(&page->lru);
1411                 __free_pages(page, page->private);
1412         }
1413
1414         kfree(tags->rqs);
1415
1416         blk_mq_free_tags(tags);
1417 }
1418
1419 static size_t order_to_size(unsigned int order)
1420 {
1421         return (size_t)PAGE_SIZE << order;
1422 }
1423
1424 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1425                 unsigned int hctx_idx)
1426 {
1427         struct blk_mq_tags *tags;
1428         unsigned int i, j, entries_per_page, max_order = 4;
1429         size_t rq_size, left;
1430
1431         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1432                                 set->numa_node,
1433                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1434         if (!tags)
1435                 return NULL;
1436
1437         INIT_LIST_HEAD(&tags->page_list);
1438
1439         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1440                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1441                                  set->numa_node);
1442         if (!tags->rqs) {
1443                 blk_mq_free_tags(tags);
1444                 return NULL;
1445         }
1446
1447         /*
1448          * rq_size is the size of the request plus driver payload, rounded
1449          * to the cacheline size
1450          */
1451         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1452                                 cache_line_size());
1453         left = rq_size * set->queue_depth;
1454
1455         for (i = 0; i < set->queue_depth; ) {
1456                 int this_order = max_order;
1457                 struct page *page;
1458                 int to_do;
1459                 void *p;
1460
1461                 while (left < order_to_size(this_order - 1) && this_order)
1462                         this_order--;
1463
1464                 do {
1465                         page = alloc_pages_node(set->numa_node,
1466                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1467                                 this_order);
1468                         if (page)
1469                                 break;
1470                         if (!this_order--)
1471                                 break;
1472                         if (order_to_size(this_order) < rq_size)
1473                                 break;
1474                 } while (1);
1475
1476                 if (!page)
1477                         goto fail;
1478
1479                 page->private = this_order;
1480                 list_add_tail(&page->lru, &tags->page_list);
1481
1482                 p = page_address(page);
1483                 entries_per_page = order_to_size(this_order) / rq_size;
1484                 to_do = min(entries_per_page, set->queue_depth - i);
1485                 left -= to_do * rq_size;
1486                 for (j = 0; j < to_do; j++) {
1487                         tags->rqs[i] = p;
1488                         if (set->ops->init_request) {
1489                                 if (set->ops->init_request(set->driver_data,
1490                                                 tags->rqs[i], hctx_idx, i,
1491                                                 set->numa_node)) {
1492                                         tags->rqs[i] = NULL;
1493                                         goto fail;
1494                                 }
1495                         }
1496
1497                         p += rq_size;
1498                         i++;
1499                 }
1500         }
1501
1502         return tags;
1503
1504 fail:
1505         blk_mq_free_rq_map(set, tags, hctx_idx);
1506         return NULL;
1507 }
1508
1509 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1510 {
1511         kfree(bitmap->map);
1512 }
1513
1514 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1515 {
1516         unsigned int bpw = 8, total, num_maps, i;
1517
1518         bitmap->bits_per_word = bpw;
1519
1520         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1521         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1522                                         GFP_KERNEL, node);
1523         if (!bitmap->map)
1524                 return -ENOMEM;
1525
1526         total = nr_cpu_ids;
1527         for (i = 0; i < num_maps; i++) {
1528                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1529                 total -= bitmap->map[i].depth;
1530         }
1531
1532         return 0;
1533 }
1534
1535 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1536 {
1537         struct request_queue *q = hctx->queue;
1538         struct blk_mq_ctx *ctx;
1539         LIST_HEAD(tmp);
1540
1541         /*
1542          * Move ctx entries to new CPU, if this one is going away.
1543          */
1544         ctx = __blk_mq_get_ctx(q, cpu);
1545
1546         spin_lock(&ctx->lock);
1547         if (!list_empty(&ctx->rq_list)) {
1548                 list_splice_init(&ctx->rq_list, &tmp);
1549                 blk_mq_hctx_clear_pending(hctx, ctx);
1550         }
1551         spin_unlock(&ctx->lock);
1552
1553         if (list_empty(&tmp))
1554                 return NOTIFY_OK;
1555
1556         ctx = blk_mq_get_ctx(q);
1557         spin_lock(&ctx->lock);
1558
1559         while (!list_empty(&tmp)) {
1560                 struct request *rq;
1561
1562                 rq = list_first_entry(&tmp, struct request, queuelist);
1563                 rq->mq_ctx = ctx;
1564                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1565         }
1566
1567         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1568         blk_mq_hctx_mark_pending(hctx, ctx);
1569
1570         spin_unlock(&ctx->lock);
1571
1572         blk_mq_run_hw_queue(hctx, true);
1573         blk_mq_put_ctx(ctx);
1574         return NOTIFY_OK;
1575 }
1576
1577 static int blk_mq_hctx_notify(void *data, unsigned long action,
1578                               unsigned int cpu)
1579 {
1580         struct blk_mq_hw_ctx *hctx = data;
1581
1582         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1583                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1584
1585         /*
1586          * In case of CPU online, tags may be reallocated
1587          * in blk_mq_map_swqueue() after mapping is updated.
1588          */
1589
1590         return NOTIFY_OK;
1591 }
1592
1593 static void blk_mq_exit_hctx(struct request_queue *q,
1594                 struct blk_mq_tag_set *set,
1595                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1596 {
1597         unsigned flush_start_tag = set->queue_depth;
1598
1599         blk_mq_tag_idle(hctx);
1600
1601         if (set->ops->exit_request)
1602                 set->ops->exit_request(set->driver_data,
1603                                        hctx->fq->flush_rq, hctx_idx,
1604                                        flush_start_tag + hctx_idx);
1605
1606         if (set->ops->exit_hctx)
1607                 set->ops->exit_hctx(hctx, hctx_idx);
1608
1609         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1610         blk_free_flush_queue(hctx->fq);
1611         kfree(hctx->ctxs);
1612         blk_mq_free_bitmap(&hctx->ctx_map);
1613 }
1614
1615 static void blk_mq_exit_hw_queues(struct request_queue *q,
1616                 struct blk_mq_tag_set *set, int nr_queue)
1617 {
1618         struct blk_mq_hw_ctx *hctx;
1619         unsigned int i;
1620
1621         queue_for_each_hw_ctx(q, hctx, i) {
1622                 if (i == nr_queue)
1623                         break;
1624                 blk_mq_exit_hctx(q, set, hctx, i);
1625         }
1626 }
1627
1628 static void blk_mq_free_hw_queues(struct request_queue *q,
1629                 struct blk_mq_tag_set *set)
1630 {
1631         struct blk_mq_hw_ctx *hctx;
1632         unsigned int i;
1633
1634         queue_for_each_hw_ctx(q, hctx, i)
1635                 free_cpumask_var(hctx->cpumask);
1636 }
1637
1638 static int blk_mq_init_hctx(struct request_queue *q,
1639                 struct blk_mq_tag_set *set,
1640                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1641 {
1642         int node;
1643         unsigned flush_start_tag = set->queue_depth;
1644
1645         node = hctx->numa_node;
1646         if (node == NUMA_NO_NODE)
1647                 node = hctx->numa_node = set->numa_node;
1648
1649         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1650         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1651         spin_lock_init(&hctx->lock);
1652         INIT_LIST_HEAD(&hctx->dispatch);
1653         hctx->queue = q;
1654         hctx->queue_num = hctx_idx;
1655         hctx->flags = set->flags;
1656
1657         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1658                                         blk_mq_hctx_notify, hctx);
1659         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1660
1661         hctx->tags = set->tags[hctx_idx];
1662
1663         /*
1664          * Allocate space for all possible cpus to avoid allocation at
1665          * runtime
1666          */
1667         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1668                                         GFP_KERNEL, node);
1669         if (!hctx->ctxs)
1670                 goto unregister_cpu_notifier;
1671
1672         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1673                 goto free_ctxs;
1674
1675         hctx->nr_ctx = 0;
1676
1677         if (set->ops->init_hctx &&
1678             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1679                 goto free_bitmap;
1680
1681         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1682         if (!hctx->fq)
1683                 goto exit_hctx;
1684
1685         if (set->ops->init_request &&
1686             set->ops->init_request(set->driver_data,
1687                                    hctx->fq->flush_rq, hctx_idx,
1688                                    flush_start_tag + hctx_idx, node))
1689                 goto free_fq;
1690
1691         return 0;
1692
1693  free_fq:
1694         kfree(hctx->fq);
1695  exit_hctx:
1696         if (set->ops->exit_hctx)
1697                 set->ops->exit_hctx(hctx, hctx_idx);
1698  free_bitmap:
1699         blk_mq_free_bitmap(&hctx->ctx_map);
1700  free_ctxs:
1701         kfree(hctx->ctxs);
1702  unregister_cpu_notifier:
1703         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1704
1705         return -1;
1706 }
1707
1708 static int blk_mq_init_hw_queues(struct request_queue *q,
1709                 struct blk_mq_tag_set *set)
1710 {
1711         struct blk_mq_hw_ctx *hctx;
1712         unsigned int i;
1713
1714         /*
1715          * Initialize hardware queues
1716          */
1717         queue_for_each_hw_ctx(q, hctx, i) {
1718                 if (blk_mq_init_hctx(q, set, hctx, i))
1719                         break;
1720         }
1721
1722         if (i == q->nr_hw_queues)
1723                 return 0;
1724
1725         /*
1726          * Init failed
1727          */
1728         blk_mq_exit_hw_queues(q, set, i);
1729
1730         return 1;
1731 }
1732
1733 static void blk_mq_init_cpu_queues(struct request_queue *q,
1734                                    unsigned int nr_hw_queues)
1735 {
1736         unsigned int i;
1737
1738         for_each_possible_cpu(i) {
1739                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1740                 struct blk_mq_hw_ctx *hctx;
1741
1742                 memset(__ctx, 0, sizeof(*__ctx));
1743                 __ctx->cpu = i;
1744                 spin_lock_init(&__ctx->lock);
1745                 INIT_LIST_HEAD(&__ctx->rq_list);
1746                 __ctx->queue = q;
1747
1748                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1749                 if (!cpu_online(i))
1750                         continue;
1751
1752                 hctx = q->mq_ops->map_queue(q, i);
1753
1754                 /*
1755                  * Set local node, IFF we have more than one hw queue. If
1756                  * not, we remain on the home node of the device
1757                  */
1758                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1759                         hctx->numa_node = cpu_to_node(i);
1760         }
1761 }
1762
1763 static void blk_mq_map_swqueue(struct request_queue *q)
1764 {
1765         unsigned int i;
1766         struct blk_mq_hw_ctx *hctx;
1767         struct blk_mq_ctx *ctx;
1768         struct blk_mq_tag_set *set = q->tag_set;
1769
1770         queue_for_each_hw_ctx(q, hctx, i) {
1771                 cpumask_clear(hctx->cpumask);
1772                 hctx->nr_ctx = 0;
1773         }
1774
1775         /*
1776          * Map software to hardware queues
1777          */
1778         queue_for_each_ctx(q, ctx, i) {
1779                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1780                 if (!cpu_online(i))
1781                         continue;
1782
1783                 hctx = q->mq_ops->map_queue(q, i);
1784                 cpumask_set_cpu(i, hctx->cpumask);
1785                 ctx->index_hw = hctx->nr_ctx;
1786                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1787         }
1788
1789         queue_for_each_hw_ctx(q, hctx, i) {
1790                 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1791
1792                 /*
1793                  * If no software queues are mapped to this hardware queue,
1794                  * disable it and free the request entries.
1795                  */
1796                 if (!hctx->nr_ctx) {
1797                         if (set->tags[i]) {
1798                                 blk_mq_free_rq_map(set, set->tags[i], i);
1799                                 set->tags[i] = NULL;
1800                         }
1801                         hctx->tags = NULL;
1802                         continue;
1803                 }
1804
1805                 /* unmapped hw queue can be remapped after CPU topo changed */
1806                 if (!set->tags[i])
1807                         set->tags[i] = blk_mq_init_rq_map(set, i);
1808                 hctx->tags = set->tags[i];
1809                 WARN_ON(!hctx->tags);
1810
1811                 /*
1812                  * Set the map size to the number of mapped software queues.
1813                  * This is more accurate and more efficient than looping
1814                  * over all possibly mapped software queues.
1815                  */
1816                 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1817
1818                 /*
1819                  * Initialize batch roundrobin counts
1820                  */
1821                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1822                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1823         }
1824 }
1825
1826 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1827 {
1828         struct blk_mq_hw_ctx *hctx;
1829         struct request_queue *q;
1830         bool shared;
1831         int i;
1832
1833         if (set->tag_list.next == set->tag_list.prev)
1834                 shared = false;
1835         else
1836                 shared = true;
1837
1838         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1839                 blk_mq_freeze_queue(q);
1840
1841                 queue_for_each_hw_ctx(q, hctx, i) {
1842                         if (shared)
1843                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1844                         else
1845                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1846                 }
1847                 blk_mq_unfreeze_queue(q);
1848         }
1849 }
1850
1851 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1852 {
1853         struct blk_mq_tag_set *set = q->tag_set;
1854
1855         mutex_lock(&set->tag_list_lock);
1856         list_del_init(&q->tag_set_list);
1857         blk_mq_update_tag_set_depth(set);
1858         mutex_unlock(&set->tag_list_lock);
1859 }
1860
1861 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1862                                      struct request_queue *q)
1863 {
1864         q->tag_set = set;
1865
1866         mutex_lock(&set->tag_list_lock);
1867         list_add_tail(&q->tag_set_list, &set->tag_list);
1868         blk_mq_update_tag_set_depth(set);
1869         mutex_unlock(&set->tag_list_lock);
1870 }
1871
1872 /*
1873  * It is the actual release handler for mq, but we do it from
1874  * request queue's release handler for avoiding use-after-free
1875  * and headache because q->mq_kobj shouldn't have been introduced,
1876  * but we can't group ctx/kctx kobj without it.
1877  */
1878 void blk_mq_release(struct request_queue *q)
1879 {
1880         struct blk_mq_hw_ctx *hctx;
1881         unsigned int i;
1882
1883         /* hctx kobj stays in hctx */
1884         queue_for_each_hw_ctx(q, hctx, i)
1885                 kfree(hctx);
1886
1887         kfree(q->queue_hw_ctx);
1888
1889         /* ctx kobj stays in queue_ctx */
1890         free_percpu(q->queue_ctx);
1891 }
1892
1893 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1894 {
1895         struct request_queue *uninit_q, *q;
1896
1897         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1898         if (!uninit_q)
1899                 return ERR_PTR(-ENOMEM);
1900
1901         q = blk_mq_init_allocated_queue(set, uninit_q);
1902         if (IS_ERR(q))
1903                 blk_cleanup_queue(uninit_q);
1904
1905         return q;
1906 }
1907 EXPORT_SYMBOL(blk_mq_init_queue);
1908
1909 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1910                                                   struct request_queue *q)
1911 {
1912         struct blk_mq_hw_ctx **hctxs;
1913         struct blk_mq_ctx __percpu *ctx;
1914         unsigned int *map;
1915         int i;
1916
1917         ctx = alloc_percpu(struct blk_mq_ctx);
1918         if (!ctx)
1919                 return ERR_PTR(-ENOMEM);
1920
1921         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1922                         set->numa_node);
1923
1924         if (!hctxs)
1925                 goto err_percpu;
1926
1927         map = blk_mq_make_queue_map(set);
1928         if (!map)
1929                 goto err_map;
1930
1931         for (i = 0; i < set->nr_hw_queues; i++) {
1932                 int node = blk_mq_hw_queue_to_node(map, i);
1933
1934                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1935                                         GFP_KERNEL, node);
1936                 if (!hctxs[i])
1937                         goto err_hctxs;
1938
1939                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1940                                                 node))
1941                         goto err_hctxs;
1942
1943                 atomic_set(&hctxs[i]->nr_active, 0);
1944                 hctxs[i]->numa_node = node;
1945                 hctxs[i]->queue_num = i;
1946         }
1947
1948         /*
1949          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1950          * See blk_register_queue() for details.
1951          */
1952         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1953                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1954                 goto err_hctxs;
1955
1956         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1957         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
1958
1959         q->nr_queues = nr_cpu_ids;
1960         q->nr_hw_queues = set->nr_hw_queues;
1961         q->mq_map = map;
1962
1963         q->queue_ctx = ctx;
1964         q->queue_hw_ctx = hctxs;
1965
1966         q->mq_ops = set->ops;
1967         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1968
1969         if (!(set->flags & BLK_MQ_F_SG_MERGE))
1970                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1971
1972         q->sg_reserved_size = INT_MAX;
1973
1974         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1975         INIT_LIST_HEAD(&q->requeue_list);
1976         spin_lock_init(&q->requeue_lock);
1977
1978         if (q->nr_hw_queues > 1)
1979                 blk_queue_make_request(q, blk_mq_make_request);
1980         else
1981                 blk_queue_make_request(q, blk_sq_make_request);
1982
1983         /*
1984          * Do this after blk_queue_make_request() overrides it...
1985          */
1986         q->nr_requests = set->queue_depth;
1987
1988         if (set->ops->complete)
1989                 blk_queue_softirq_done(q, set->ops->complete);
1990
1991         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1992
1993         if (blk_mq_init_hw_queues(q, set))
1994                 goto err_hctxs;
1995
1996         mutex_lock(&all_q_mutex);
1997         list_add_tail(&q->all_q_node, &all_q_list);
1998         mutex_unlock(&all_q_mutex);
1999
2000         blk_mq_add_queue_tag_set(set, q);
2001
2002         blk_mq_map_swqueue(q);
2003
2004         return q;
2005
2006 err_hctxs:
2007         kfree(map);
2008         for (i = 0; i < set->nr_hw_queues; i++) {
2009                 if (!hctxs[i])
2010                         break;
2011                 free_cpumask_var(hctxs[i]->cpumask);
2012                 kfree(hctxs[i]);
2013         }
2014 err_map:
2015         kfree(hctxs);
2016 err_percpu:
2017         free_percpu(ctx);
2018         return ERR_PTR(-ENOMEM);
2019 }
2020 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2021
2022 void blk_mq_free_queue(struct request_queue *q)
2023 {
2024         struct blk_mq_tag_set   *set = q->tag_set;
2025
2026         blk_mq_del_queue_tag_set(q);
2027
2028         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2029         blk_mq_free_hw_queues(q, set);
2030
2031         percpu_ref_exit(&q->mq_usage_counter);
2032
2033         kfree(q->mq_map);
2034
2035         q->mq_map = NULL;
2036
2037         mutex_lock(&all_q_mutex);
2038         list_del_init(&q->all_q_node);
2039         mutex_unlock(&all_q_mutex);
2040 }
2041
2042 /* Basically redo blk_mq_init_queue with queue frozen */
2043 static void blk_mq_queue_reinit(struct request_queue *q)
2044 {
2045         WARN_ON_ONCE(!q->mq_freeze_depth);
2046
2047         blk_mq_sysfs_unregister(q);
2048
2049         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2050
2051         /*
2052          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2053          * we should change hctx numa_node according to new topology (this
2054          * involves free and re-allocate memory, worthy doing?)
2055          */
2056
2057         blk_mq_map_swqueue(q);
2058
2059         blk_mq_sysfs_register(q);
2060 }
2061
2062 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2063                                       unsigned long action, void *hcpu)
2064 {
2065         struct request_queue *q;
2066
2067         /*
2068          * Before new mappings are established, hotadded cpu might already
2069          * start handling requests. This doesn't break anything as we map
2070          * offline CPUs to first hardware queue. We will re-init the queue
2071          * below to get optimal settings.
2072          */
2073         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2074             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2075                 return NOTIFY_OK;
2076
2077         mutex_lock(&all_q_mutex);
2078
2079         /*
2080          * We need to freeze and reinit all existing queues.  Freezing
2081          * involves synchronous wait for an RCU grace period and doing it
2082          * one by one may take a long time.  Start freezing all queues in
2083          * one swoop and then wait for the completions so that freezing can
2084          * take place in parallel.
2085          */
2086         list_for_each_entry(q, &all_q_list, all_q_node)
2087                 blk_mq_freeze_queue_start(q);
2088         list_for_each_entry(q, &all_q_list, all_q_node) {
2089                 blk_mq_freeze_queue_wait(q);
2090
2091                 /*
2092                  * timeout handler can't touch hw queue during the
2093                  * reinitialization
2094                  */
2095                 del_timer_sync(&q->timeout);
2096         }
2097
2098         list_for_each_entry(q, &all_q_list, all_q_node)
2099                 blk_mq_queue_reinit(q);
2100
2101         list_for_each_entry(q, &all_q_list, all_q_node)
2102                 blk_mq_unfreeze_queue(q);
2103
2104         mutex_unlock(&all_q_mutex);
2105         return NOTIFY_OK;
2106 }
2107
2108 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2109 {
2110         int i;
2111
2112         for (i = 0; i < set->nr_hw_queues; i++) {
2113                 set->tags[i] = blk_mq_init_rq_map(set, i);
2114                 if (!set->tags[i])
2115                         goto out_unwind;
2116         }
2117
2118         return 0;
2119
2120 out_unwind:
2121         while (--i >= 0)
2122                 blk_mq_free_rq_map(set, set->tags[i], i);
2123
2124         return -ENOMEM;
2125 }
2126
2127 /*
2128  * Allocate the request maps associated with this tag_set. Note that this
2129  * may reduce the depth asked for, if memory is tight. set->queue_depth
2130  * will be updated to reflect the allocated depth.
2131  */
2132 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2133 {
2134         unsigned int depth;
2135         int err;
2136
2137         depth = set->queue_depth;
2138         do {
2139                 err = __blk_mq_alloc_rq_maps(set);
2140                 if (!err)
2141                         break;
2142
2143                 set->queue_depth >>= 1;
2144                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2145                         err = -ENOMEM;
2146                         break;
2147                 }
2148         } while (set->queue_depth);
2149
2150         if (!set->queue_depth || err) {
2151                 pr_err("blk-mq: failed to allocate request map\n");
2152                 return -ENOMEM;
2153         }
2154
2155         if (depth != set->queue_depth)
2156                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2157                                                 depth, set->queue_depth);
2158
2159         return 0;
2160 }
2161
2162 /*
2163  * Alloc a tag set to be associated with one or more request queues.
2164  * May fail with EINVAL for various error conditions. May adjust the
2165  * requested depth down, if if it too large. In that case, the set
2166  * value will be stored in set->queue_depth.
2167  */
2168 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2169 {
2170         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2171
2172         if (!set->nr_hw_queues)
2173                 return -EINVAL;
2174         if (!set->queue_depth)
2175                 return -EINVAL;
2176         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2177                 return -EINVAL;
2178
2179         if (!set->ops->queue_rq || !set->ops->map_queue)
2180                 return -EINVAL;
2181
2182         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2183                 pr_info("blk-mq: reduced tag depth to %u\n",
2184                         BLK_MQ_MAX_DEPTH);
2185                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2186         }
2187
2188         /*
2189          * If a crashdump is active, then we are potentially in a very
2190          * memory constrained environment. Limit us to 1 queue and
2191          * 64 tags to prevent using too much memory.
2192          */
2193         if (is_kdump_kernel()) {
2194                 set->nr_hw_queues = 1;
2195                 set->queue_depth = min(64U, set->queue_depth);
2196         }
2197
2198         set->tags = kmalloc_node(set->nr_hw_queues *
2199                                  sizeof(struct blk_mq_tags *),
2200                                  GFP_KERNEL, set->numa_node);
2201         if (!set->tags)
2202                 return -ENOMEM;
2203
2204         if (blk_mq_alloc_rq_maps(set))
2205                 goto enomem;
2206
2207         mutex_init(&set->tag_list_lock);
2208         INIT_LIST_HEAD(&set->tag_list);
2209
2210         return 0;
2211 enomem:
2212         kfree(set->tags);
2213         set->tags = NULL;
2214         return -ENOMEM;
2215 }
2216 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2217
2218 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2219 {
2220         int i;
2221
2222         for (i = 0; i < set->nr_hw_queues; i++) {
2223                 if (set->tags[i])
2224                         blk_mq_free_rq_map(set, set->tags[i], i);
2225         }
2226
2227         kfree(set->tags);
2228         set->tags = NULL;
2229 }
2230 EXPORT_SYMBOL(blk_mq_free_tag_set);
2231
2232 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2233 {
2234         struct blk_mq_tag_set *set = q->tag_set;
2235         struct blk_mq_hw_ctx *hctx;
2236         int i, ret;
2237
2238         if (!set || nr > set->queue_depth)
2239                 return -EINVAL;
2240
2241         ret = 0;
2242         queue_for_each_hw_ctx(q, hctx, i) {
2243                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2244                 if (ret)
2245                         break;
2246         }
2247
2248         if (!ret)
2249                 q->nr_requests = nr;
2250
2251         return ret;
2252 }
2253
2254 void blk_mq_disable_hotplug(void)
2255 {
2256         mutex_lock(&all_q_mutex);
2257 }
2258
2259 void blk_mq_enable_hotplug(void)
2260 {
2261         mutex_unlock(&all_q_mutex);
2262 }
2263
2264 static int __init blk_mq_init(void)
2265 {
2266         blk_mq_cpu_init();
2267
2268         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2269
2270         return 0;
2271 }
2272 subsys_initcall(blk_mq_init);