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                                 !atomic_read(&q->mq_freeze_depth) ||
  93                                 blk_queue_dying(q));
  94                 if (blk_queue_dying(q))
  95                         return -ENODEV;
  96                 if (ret)
  97                         return ret;
  98         }
  99 }
 100
 101 static void blk_mq_queue_exit(struct request_queue *q)
 102 {
 103         percpu_ref_put(&q->mq_usage_counter);
 104 }
 105
 106 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 107 {
 108         struct request_queue *q =
 109                 container_of(ref, struct request_queue, mq_usage_counter);
 110
 111         wake_up_all(&q->mq_freeze_wq);
 112 }
 113
 114 void blk_mq_freeze_queue_start(struct request_queue *q)
 115 {
 116         int freeze_depth;
 117
 118         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 119         if (freeze_depth == 1) {
 120                 percpu_ref_kill(&q->mq_usage_counter);
 121                 blk_mq_run_hw_queues(q, false);
 122         }
 123 }
 124 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 125
 126 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 127 {
 128         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 129 }
 130
 131 /*
 132  * Guarantee no request is in use, so we can change any data structure of
 133  * the queue afterward.
 134  */
 135 void blk_mq_freeze_queue(struct request_queue *q)
 136 {
 137         blk_mq_freeze_queue_start(q);
 138         blk_mq_freeze_queue_wait(q);
 139 }
 140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 141
 142 void blk_mq_unfreeze_queue(struct request_queue *q)
 143 {
 144         int freeze_depth;
 145
 146         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 147         WARN_ON_ONCE(freeze_depth < 0);
 148         if (!freeze_depth) {
 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, gfp);
 259         if (ret)
 260                 return ERR_PTR(ret);
 261
 262         ctx = blk_mq_get_ctx(q);
 263         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 264         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 265                         reserved, ctx, hctx);
 266
 267         rq = __blk_mq_alloc_request(&alloc_data, rw);
 268         if (!rq && (gfp & __GFP_WAIT)) {
 269                 __blk_mq_run_hw_queue(hctx);
 270                 blk_mq_put_ctx(ctx);
 271
 272                 ctx = blk_mq_get_ctx(q);
 273                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 274                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 275                                 hctx);
 276                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 277                 ctx = alloc_data.ctx;
 278         }
 279         blk_mq_put_ctx(ctx);
 280         if (!rq) {
 281                 blk_mq_queue_exit(q);
 282                 return ERR_PTR(-EWOULDBLOCK);
 283         }
 284         return rq;
 285 }
 286 EXPORT_SYMBOL(blk_mq_alloc_request);
 287
 288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 289                                   struct blk_mq_ctx *ctx, struct request *rq)
 290 {
 291         const int tag = rq->tag;
 292         struct request_queue *q = rq->q;
 293
 294         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 295                 atomic_dec(&hctx->nr_active);
 296         rq->cmd_flags = 0;
 297
 298         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 299         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 300         blk_mq_queue_exit(q);
 301 }
 302
 303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 304 {
 305         struct blk_mq_ctx *ctx = rq->mq_ctx;
 306
 307         ctx->rq_completed[rq_is_sync(rq)]++;
 308         __blk_mq_free_request(hctx, ctx, rq);
 309
 310 }
 311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 312
 313 void blk_mq_free_request(struct request *rq)
 314 {
 315         struct blk_mq_hw_ctx *hctx;
 316         struct request_queue *q = rq->q;
 317
 318         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 319         blk_mq_free_hctx_request(hctx, rq);
 320 }
 321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 322
 323 inline void __blk_mq_end_request(struct request *rq, int error)
 324 {
 325         blk_account_io_done(rq);
 326
 327         if (rq->end_io) {
 328                 rq->end_io(rq, error);
 329         } else {
 330                 if (unlikely(blk_bidi_rq(rq)))
 331                         blk_mq_free_request(rq->next_rq);
 332                 blk_mq_free_request(rq);
 333         }
 334 }
 335 EXPORT_SYMBOL(__blk_mq_end_request);
 336
 337 void blk_mq_end_request(struct request *rq, int error)
 338 {
 339         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 340                 BUG();
 341         __blk_mq_end_request(rq, error);
 342 }
 343 EXPORT_SYMBOL(blk_mq_end_request);
 344
 345 static void __blk_mq_complete_request_remote(void *data)
 346 {
 347         struct request *rq = data;
 348
 349         rq->q->softirq_done_fn(rq);
 350 }
 351
 352 static void blk_mq_ipi_complete_request(struct request *rq)
 353 {
 354         struct blk_mq_ctx *ctx = rq->mq_ctx;
 355         bool shared = false;
 356         int cpu;
 357
 358         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 359                 rq->q->softirq_done_fn(rq);
 360                 return;
 361         }
 362
 363         cpu = get_cpu();
 364         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 365                 shared = cpus_share_cache(cpu, ctx->cpu);
 366
 367         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 368                 rq->csd.func = __blk_mq_complete_request_remote;
 369                 rq->csd.info = rq;
 370                 rq->csd.flags = 0;
 371                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 372         } else {
 373                 rq->q->softirq_done_fn(rq);
 374         }
 375         put_cpu();
 376 }
 377
 378 void __blk_mq_complete_request(struct request *rq)
 379 {
 380         struct request_queue *q = rq->q;
 381
 382         if (!q->softirq_done_fn)
 383                 blk_mq_end_request(rq, rq->errors);
 384         else
 385                 blk_mq_ipi_complete_request(rq);
 386 }
 387
 388 /**
 389  * blk_mq_complete_request - end I/O on a request
 390  * @rq:         the request being processed
 391  *
 392  * Description:
 393  *      Ends all I/O on a request. It does not handle partial completions.
 394  *      The actual completion happens out-of-order, through a IPI handler.
 395  **/
 396 void blk_mq_complete_request(struct request *rq)
 397 {
 398         struct request_queue *q = rq->q;
 399
 400         if (unlikely(blk_should_fake_timeout(q)))
 401                 return;
 402         if (!blk_mark_rq_complete(rq))
 403                 __blk_mq_complete_request(rq);
 404 }
 405 EXPORT_SYMBOL(blk_mq_complete_request);
 406
 407 int blk_mq_request_started(struct request *rq)
 408 {
 409         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 410 }
 411 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 412
 413 void blk_mq_start_request(struct request *rq)
 414 {
 415         struct request_queue *q = rq->q;
 416
 417         trace_block_rq_issue(q, rq);
 418
 419         rq->resid_len = blk_rq_bytes(rq);
 420         if (unlikely(blk_bidi_rq(rq)))
 421                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 422
 423         blk_add_timer(rq);
 424
 425         /*
 426          * Ensure that ->deadline is visible before set the started
 427          * flag and clear the completed flag.
 428          */
 429         smp_mb__before_atomic();
 430
 431         /*
 432          * Mark us as started and clear complete. Complete might have been
 433          * set if requeue raced with timeout, which then marked it as
 434          * complete. So be sure to clear complete again when we start
 435          * the request, otherwise we'll ignore the completion event.
 436          */
 437         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 438                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 439         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 440                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 441
 442         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 443                 /*
 444                  * Make sure space for the drain appears.  We know we can do
 445                  * this because max_hw_segments has been adjusted to be one
 446                  * fewer than the device can handle.
 447                  */
 448                 rq->nr_phys_segments++;
 449         }
 450 }
 451 EXPORT_SYMBOL(blk_mq_start_request);
 452
 453 static void __blk_mq_requeue_request(struct request *rq)
 454 {
 455         struct request_queue *q = rq->q;
 456
 457         trace_block_rq_requeue(q, rq);
 458
 459         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 460                 if (q->dma_drain_size && blk_rq_bytes(rq))
 461                         rq->nr_phys_segments--;
 462         }
 463 }
 464
 465 void blk_mq_requeue_request(struct request *rq)
 466 {
 467         __blk_mq_requeue_request(rq);
 468
 469         BUG_ON(blk_queued_rq(rq));
 470         blk_mq_add_to_requeue_list(rq, true);
 471 }
 472 EXPORT_SYMBOL(blk_mq_requeue_request);
 473
 474 static void blk_mq_requeue_work(struct work_struct *work)
 475 {
 476         struct request_queue *q =
 477                 container_of(work, struct request_queue, requeue_work);
 478         LIST_HEAD(rq_list);
 479         struct request *rq, *next;
 480         unsigned long flags;
 481
 482         spin_lock_irqsave(&q->requeue_lock, flags);
 483         list_splice_init(&q->requeue_list, &rq_list);
 484         spin_unlock_irqrestore(&q->requeue_lock, flags);
 485
 486         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 487                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 488                         continue;
 489
 490                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 491                 list_del_init(&rq->queuelist);
 492                 blk_mq_insert_request(rq, true, false, false);
 493         }
 494
 495         while (!list_empty(&rq_list)) {
 496                 rq = list_entry(rq_list.next, struct request, queuelist);
 497                 list_del_init(&rq->queuelist);
 498                 blk_mq_insert_request(rq, false, false, false);
 499         }
 500
 501         /*
 502          * Use the start variant of queue running here, so that running
 503          * the requeue work will kick stopped queues.
 504          */
 505         blk_mq_start_hw_queues(q);
 506 }
 507
 508 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 509 {
 510         struct request_queue *q = rq->q;
 511         unsigned long flags;
 512
 513         /*
 514          * We abuse this flag that is otherwise used by the I/O scheduler to
 515          * request head insertation from the workqueue.
 516          */
 517         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 518
 519         spin_lock_irqsave(&q->requeue_lock, flags);
 520         if (at_head) {
 521                 rq->cmd_flags |= REQ_SOFTBARRIER;
 522                 list_add(&rq->queuelist, &q->requeue_list);
 523         } else {
 524                 list_add_tail(&rq->queuelist, &q->requeue_list);
 525         }
 526         spin_unlock_irqrestore(&q->requeue_lock, flags);
 527 }
 528 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 529
 530 void blk_mq_cancel_requeue_work(struct request_queue *q)
 531 {
 532         cancel_work_sync(&q->requeue_work);
 533 }
 534 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 535
 536 void blk_mq_kick_requeue_list(struct request_queue *q)
 537 {
 538         kblockd_schedule_work(&q->requeue_work);
 539 }
 540 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 541
 542 void blk_mq_abort_requeue_list(struct request_queue *q)
 543 {
 544         unsigned long flags;
 545         LIST_HEAD(rq_list);
 546
 547         spin_lock_irqsave(&q->requeue_lock, flags);
 548         list_splice_init(&q->requeue_list, &rq_list);
 549         spin_unlock_irqrestore(&q->requeue_lock, flags);
 550
 551         while (!list_empty(&rq_list)) {
 552                 struct request *rq;
 553
 554                 rq = list_first_entry(&rq_list, struct request, queuelist);
 555                 list_del_init(&rq->queuelist);
 556                 rq->errors = -EIO;
 557                 blk_mq_end_request(rq, rq->errors);
 558         }
 559 }
 560 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 561
 562 static inline bool is_flush_request(struct request *rq,
 563                 struct blk_flush_queue *fq, unsigned int tag)
 564 {
 565         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 566                         fq->flush_rq->tag == tag);
 567 }
 568
 569 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 570 {
 571         struct request *rq = tags->rqs[tag];
 572         /* mq_ctx of flush rq is always cloned from the corresponding req */
 573         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 574
 575         if (!is_flush_request(rq, fq, tag))
 576                 return rq;
 577
 578         return fq->flush_rq;
 579 }
 580 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 581
 582 struct blk_mq_timeout_data {
 583         unsigned long next;
 584         unsigned int next_set;
 585 };
 586
 587 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 588 {
 589         struct blk_mq_ops *ops = req->q->mq_ops;
 590         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 591
 592         /*
 593          * We know that complete is set at this point. If STARTED isn't set
 594          * anymore, then the request isn't active and the "timeout" should
 595          * just be ignored. This can happen due to the bitflag ordering.
 596          * Timeout first checks if STARTED is set, and if it is, assumes
 597          * the request is active. But if we race with completion, then
 598          * we both flags will get cleared. So check here again, and ignore
 599          * a timeout event with a request that isn't active.
 600          */
 601         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 602                 return;
 603
 604         if (ops->timeout)
 605                 ret = ops->timeout(req, reserved);
 606
 607         switch (ret) {
 608         case BLK_EH_HANDLED:
 609                 __blk_mq_complete_request(req);
 610                 break;
 611         case BLK_EH_RESET_TIMER:
 612                 blk_add_timer(req);
 613                 blk_clear_rq_complete(req);
 614                 break;
 615         case BLK_EH_NOT_HANDLED:
 616                 break;
 617         default:
 618                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 619                 break;
 620         }
 621 }
 622
 623 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 624                 struct request *rq, void *priv, bool reserved)
 625 {
 626         struct blk_mq_timeout_data *data = priv;
 627
 628         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 629                 /*
 630                  * If a request wasn't started before the queue was
 631                  * marked dying, kill it here or it'll go unnoticed.
 632                  */
 633                 if (unlikely(blk_queue_dying(rq->q))) {
 634                         rq->errors = -EIO;
 635                         blk_mq_complete_request(rq);
 636                 }
 637                 return;
 638         }
 639         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 640                 return;
 641
 642         if (time_after_eq(jiffies, rq->deadline)) {
 643                 if (!blk_mark_rq_complete(rq))
 644                         blk_mq_rq_timed_out(rq, reserved);
 645         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 646                 data->next = rq->deadline;
 647                 data->next_set = 1;
 648         }
 649 }
 650
 651 static void blk_mq_rq_timer(unsigned long priv)
 652 {
 653         struct request_queue *q = (struct request_queue *)priv;
 654         struct blk_mq_timeout_data data = {
 655                 .next           = 0,
 656                 .next_set       = 0,
 657         };
 658         struct blk_mq_hw_ctx *hctx;
 659         int i;
 660
 661         queue_for_each_hw_ctx(q, hctx, i) {
 662                 /*
 663                  * If not software queues are currently mapped to this
 664                  * hardware queue, there's nothing to check
 665                  */
 666                 if (!blk_mq_hw_queue_mapped(hctx))
 667                         continue;
 668
 669                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 670         }
 671
 672         if (data.next_set) {
 673                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 674                 mod_timer(&q->timeout, data.next);
 675         } else {
 676                 queue_for_each_hw_ctx(q, hctx, i)
 677                         blk_mq_tag_idle(hctx);
 678         }
 679 }
 680
 681 /*
 682  * Reverse check our software queue for entries that we could potentially
 683  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 684  * too much time checking for merges.
 685  */
 686 static bool blk_mq_attempt_merge(struct request_queue *q,
 687                                  struct blk_mq_ctx *ctx, struct bio *bio)
 688 {
 689         struct request *rq;
 690         int checked = 8;
 691
 692         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 693                 int el_ret;
 694
 695                 if (!checked--)
 696                         break;
 697
 698                 if (!blk_rq_merge_ok(rq, bio))
 699                         continue;
 700
 701                 el_ret = blk_try_merge(rq, bio);
 702                 if (el_ret == ELEVATOR_BACK_MERGE) {
 703                         if (bio_attempt_back_merge(q, rq, bio)) {
 704                                 ctx->rq_merged++;
 705                                 return true;
 706                         }
 707                         break;
 708                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 709                         if (bio_attempt_front_merge(q, rq, bio)) {
 710                                 ctx->rq_merged++;
 711                                 return true;
 712                         }
 713                         break;
 714                 }
 715         }
 716
 717         return false;
 718 }
 719
 720 /*
 721  * Process software queues that have been marked busy, splicing them
 722  * to the for-dispatch
 723  */
 724 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 725 {
 726         struct blk_mq_ctx *ctx;
 727         int i;
 728
 729         for (i = 0; i < hctx->ctx_map.size; i++) {
 730                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 731                 unsigned int off, bit;
 732
 733                 if (!bm->word)
 734                         continue;
 735
 736                 bit = 0;
 737                 off = i * hctx->ctx_map.bits_per_word;
 738                 do {
 739                         bit = find_next_bit(&bm->word, bm->depth, bit);
 740                         if (bit >= bm->depth)
 741                                 break;
 742
 743                         ctx = hctx->ctxs[bit + off];
 744                         clear_bit(bit, &bm->word);
 745                         spin_lock(&ctx->lock);
 746                         list_splice_tail_init(&ctx->rq_list, list);
 747                         spin_unlock(&ctx->lock);
 748
 749                         bit++;
 750                 } while (1);
 751         }
 752 }
 753
 754 /*
 755  * Run this hardware queue, pulling any software queues mapped to it in.
 756  * Note that this function currently has various problems around ordering
 757  * of IO. In particular, we'd like FIFO behaviour on handling existing
 758  * items on the hctx->dispatch list. Ignore that for now.
 759  */
 760 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 761 {
 762         struct request_queue *q = hctx->queue;
 763         struct request *rq;
 764         LIST_HEAD(rq_list);
 765         LIST_HEAD(driver_list);
 766         struct list_head *dptr;
 767         int queued;
 768
 769         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 770
 771         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 772                 return;
 773
 774         hctx->run++;
 775
 776         /*
 777          * Touch any software queue that has pending entries.
 778          */
 779         flush_busy_ctxs(hctx, &rq_list);
 780
 781         /*
 782          * If we have previous entries on our dispatch list, grab them
 783          * and stuff them at the front for more fair dispatch.
 784          */
 785         if (!list_empty_careful(&hctx->dispatch)) {
 786                 spin_lock(&hctx->lock);
 787                 if (!list_empty(&hctx->dispatch))
 788                         list_splice_init(&hctx->dispatch, &rq_list);
 789                 spin_unlock(&hctx->lock);
 790         }
 791
 792         /*
 793          * Start off with dptr being NULL, so we start the first request
 794          * immediately, even if we have more pending.
 795          */
 796         dptr = NULL;
 797
 798         /*
 799          * Now process all the entries, sending them to the driver.
 800          */
 801         queued = 0;
 802         while (!list_empty(&rq_list)) {
 803                 struct blk_mq_queue_data bd;
 804                 int ret;
 805
 806                 rq = list_first_entry(&rq_list, struct request, queuelist);
 807                 list_del_init(&rq->queuelist);
 808
 809                 bd.rq = rq;
 810                 bd.list = dptr;
 811                 bd.last = list_empty(&rq_list);
 812
 813                 ret = q->mq_ops->queue_rq(hctx, &bd);
 814                 switch (ret) {
 815                 case BLK_MQ_RQ_QUEUE_OK:
 816                         queued++;
 817                         continue;
 818                 case BLK_MQ_RQ_QUEUE_BUSY:
 819                         list_add(&rq->queuelist, &rq_list);
 820                         __blk_mq_requeue_request(rq);
 821                         break;
 822                 default:
 823                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 824                 case BLK_MQ_RQ_QUEUE_ERROR:
 825                         rq->errors = -EIO;
 826                         blk_mq_end_request(rq, rq->errors);
 827                         break;
 828                 }
 829
 830                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 831                         break;
 832
 833                 /*
 834                  * We've done the first request. If we have more than 1
 835                  * left in the list, set dptr to defer issue.
 836                  */
 837                 if (!dptr && rq_list.next != rq_list.prev)
 838                         dptr = &driver_list;
 839         }
 840
 841         if (!queued)
 842                 hctx->dispatched[0]++;
 843         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 844                 hctx->dispatched[ilog2(queued) + 1]++;
 845
 846         /*
 847          * Any items that need requeuing? Stuff them into hctx->dispatch,
 848          * that is where we will continue on next queue run.
 849          */
 850         if (!list_empty(&rq_list)) {
 851                 spin_lock(&hctx->lock);
 852                 list_splice(&rq_list, &hctx->dispatch);
 853                 spin_unlock(&hctx->lock);
 854         }
 855 }
 856
 857 /*
 858  * It'd be great if the workqueue API had a way to pass
 859  * in a mask and had some smarts for more clever placement.
 860  * For now we just round-robin here, switching for every
 861  * BLK_MQ_CPU_WORK_BATCH queued items.
 862  */
 863 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 864 {
 865         if (hctx->queue->nr_hw_queues == 1)
 866                 return WORK_CPU_UNBOUND;
 867
 868         if (--hctx->next_cpu_batch <= 0) {
 869                 int cpu = hctx->next_cpu, next_cpu;
 870
 871                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 872                 if (next_cpu >= nr_cpu_ids)
 873                         next_cpu = cpumask_first(hctx->cpumask);
 874
 875                 hctx->next_cpu = next_cpu;
 876                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 877
 878                 return cpu;
 879         }
 880
 881         return hctx->next_cpu;
 882 }
 883
 884 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 885 {
 886         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 887             !blk_mq_hw_queue_mapped(hctx)))
 888                 return;
 889
 890         if (!async) {
 891                 int cpu = get_cpu();
 892                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 893                         __blk_mq_run_hw_queue(hctx);
 894                         put_cpu();
 895                         return;
 896                 }
 897
 898                 put_cpu();
 899         }
 900
 901         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 902                         &hctx->run_work, 0);
 903 }
 904
 905 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 906 {
 907         struct blk_mq_hw_ctx *hctx;
 908         int i;
 909
 910         queue_for_each_hw_ctx(q, hctx, i) {
 911                 if ((!blk_mq_hctx_has_pending(hctx) &&
 912                     list_empty_careful(&hctx->dispatch)) ||
 913                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 914                         continue;
 915
 916                 blk_mq_run_hw_queue(hctx, async);
 917         }
 918 }
 919 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 920
 921 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 922 {
 923         cancel_delayed_work(&hctx->run_work);
 924         cancel_delayed_work(&hctx->delay_work);
 925         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 926 }
 927 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 928
 929 void blk_mq_stop_hw_queues(struct request_queue *q)
 930 {
 931         struct blk_mq_hw_ctx *hctx;
 932         int i;
 933
 934         queue_for_each_hw_ctx(q, hctx, i)
 935                 blk_mq_stop_hw_queue(hctx);
 936 }
 937 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 938
 939 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 940 {
 941         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 942
 943         blk_mq_run_hw_queue(hctx, false);
 944 }
 945 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 946
 947 void blk_mq_start_hw_queues(struct request_queue *q)
 948 {
 949         struct blk_mq_hw_ctx *hctx;
 950         int i;
 951
 952         queue_for_each_hw_ctx(q, hctx, i)
 953                 blk_mq_start_hw_queue(hctx);
 954 }
 955 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 956
 957 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 958 {
 959         struct blk_mq_hw_ctx *hctx;
 960         int i;
 961
 962         queue_for_each_hw_ctx(q, hctx, i) {
 963                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 964                         continue;
 965
 966                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 967                 blk_mq_run_hw_queue(hctx, async);
 968         }
 969 }
 970 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 971
 972 static void blk_mq_run_work_fn(struct work_struct *work)
 973 {
 974         struct blk_mq_hw_ctx *hctx;
 975
 976         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 977
 978         __blk_mq_run_hw_queue(hctx);
 979 }
 980
 981 static void blk_mq_delay_work_fn(struct work_struct *work)
 982 {
 983         struct blk_mq_hw_ctx *hctx;
 984
 985         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 986
 987         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 988                 __blk_mq_run_hw_queue(hctx);
 989 }
 990
 991 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 992 {
 993         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
 994                 return;
 995
 996         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 997                         &hctx->delay_work, msecs_to_jiffies(msecs));
 998 }
 999 EXPORT_SYMBOL(blk_mq_delay_queue);
1000
1001 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1002                                     struct request *rq, bool at_head)
1003 {
1004         struct blk_mq_ctx *ctx = rq->mq_ctx;
1005
1006         trace_block_rq_insert(hctx->queue, rq);
1007
1008         if (at_head)
1009                 list_add(&rq->queuelist, &ctx->rq_list);
1010         else
1011                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1012
1013         blk_mq_hctx_mark_pending(hctx, ctx);
1014 }
1015
1016 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1017                 bool async)
1018 {
1019         struct request_queue *q = rq->q;
1020         struct blk_mq_hw_ctx *hctx;
1021         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1022
1023         current_ctx = blk_mq_get_ctx(q);
1024         if (!cpu_online(ctx->cpu))
1025                 rq->mq_ctx = ctx = current_ctx;
1026
1027         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1028
1029         spin_lock(&ctx->lock);
1030         __blk_mq_insert_request(hctx, rq, at_head);
1031         spin_unlock(&ctx->lock);
1032
1033         if (run_queue)
1034                 blk_mq_run_hw_queue(hctx, async);
1035
1036         blk_mq_put_ctx(current_ctx);
1037 }
1038
1039 static void blk_mq_insert_requests(struct request_queue *q,
1040                                      struct blk_mq_ctx *ctx,
1041                                      struct list_head *list,
1042                                      int depth,
1043                                      bool from_schedule)
1044
1045 {
1046         struct blk_mq_hw_ctx *hctx;
1047         struct blk_mq_ctx *current_ctx;
1048
1049         trace_block_unplug(q, depth, !from_schedule);
1050
1051         current_ctx = blk_mq_get_ctx(q);
1052
1053         if (!cpu_online(ctx->cpu))
1054                 ctx = current_ctx;
1055         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1056
1057         /*
1058          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1059          * offline now
1060          */
1061         spin_lock(&ctx->lock);
1062         while (!list_empty(list)) {
1063                 struct request *rq;
1064
1065                 rq = list_first_entry(list, struct request, queuelist);
1066                 list_del_init(&rq->queuelist);
1067                 rq->mq_ctx = ctx;
1068                 __blk_mq_insert_request(hctx, rq, false);
1069         }
1070         spin_unlock(&ctx->lock);
1071
1072         blk_mq_run_hw_queue(hctx, from_schedule);
1073         blk_mq_put_ctx(current_ctx);
1074 }
1075
1076 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1077 {
1078         struct request *rqa = container_of(a, struct request, queuelist);
1079         struct request *rqb = container_of(b, struct request, queuelist);
1080
1081         return !(rqa->mq_ctx < rqb->mq_ctx ||
1082                  (rqa->mq_ctx == rqb->mq_ctx &&
1083                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1084 }
1085
1086 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1087 {
1088         struct blk_mq_ctx *this_ctx;
1089         struct request_queue *this_q;
1090         struct request *rq;
1091         LIST_HEAD(list);
1092         LIST_HEAD(ctx_list);
1093         unsigned int depth;
1094
1095         list_splice_init(&plug->mq_list, &list);
1096
1097         list_sort(NULL, &list, plug_ctx_cmp);
1098
1099         this_q = NULL;
1100         this_ctx = NULL;
1101         depth = 0;
1102
1103         while (!list_empty(&list)) {
1104                 rq = list_entry_rq(list.next);
1105                 list_del_init(&rq->queuelist);
1106                 BUG_ON(!rq->q);
1107                 if (rq->mq_ctx != this_ctx) {
1108                         if (this_ctx) {
1109                                 blk_mq_insert_requests(this_q, this_ctx,
1110                                                         &ctx_list, depth,
1111                                                         from_schedule);
1112                         }
1113
1114                         this_ctx = rq->mq_ctx;
1115                         this_q = rq->q;
1116                         depth = 0;
1117                 }
1118
1119                 depth++;
1120                 list_add_tail(&rq->queuelist, &ctx_list);
1121         }
1122
1123         /*
1124          * If 'this_ctx' is set, we know we have entries to complete
1125          * on 'ctx_list'. Do those.
1126          */
1127         if (this_ctx) {
1128                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1129                                        from_schedule);
1130         }
1131 }
1132
1133 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1134 {
1135         init_request_from_bio(rq, bio);
1136
1137         if (blk_do_io_stat(rq))
1138                 blk_account_io_start(rq, 1);
1139 }
1140
1141 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1142 {
1143         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1144                 !blk_queue_nomerges(hctx->queue);
1145 }
1146
1147 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1148                                          struct blk_mq_ctx *ctx,
1149                                          struct request *rq, struct bio *bio)
1150 {
1151         if (!hctx_allow_merges(hctx)) {
1152                 blk_mq_bio_to_request(rq, bio);
1153                 spin_lock(&ctx->lock);
1154 insert_rq:
1155                 __blk_mq_insert_request(hctx, rq, false);
1156                 spin_unlock(&ctx->lock);
1157                 return false;
1158         } else {
1159                 struct request_queue *q = hctx->queue;
1160
1161                 spin_lock(&ctx->lock);
1162                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1163                         blk_mq_bio_to_request(rq, bio);
1164                         goto insert_rq;
1165                 }
1166
1167                 spin_unlock(&ctx->lock);
1168                 __blk_mq_free_request(hctx, ctx, rq);
1169                 return true;
1170         }
1171 }
1172
1173 struct blk_map_ctx {
1174         struct blk_mq_hw_ctx *hctx;
1175         struct blk_mq_ctx *ctx;
1176 };
1177
1178 static struct request *blk_mq_map_request(struct request_queue *q,
1179                                           struct bio *bio,
1180                                           struct blk_map_ctx *data)
1181 {
1182         struct blk_mq_hw_ctx *hctx;
1183         struct blk_mq_ctx *ctx;
1184         struct request *rq;
1185         int rw = bio_data_dir(bio);
1186         struct blk_mq_alloc_data alloc_data;
1187
1188         if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1189                 bio_endio(bio, -EIO);
1190                 return NULL;
1191         }
1192
1193         ctx = blk_mq_get_ctx(q);
1194         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1195
1196         if (rw_is_sync(bio->bi_rw))
1197                 rw |= REQ_SYNC;
1198
1199         trace_block_getrq(q, bio, rw);
1200         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1201                         hctx);
1202         rq = __blk_mq_alloc_request(&alloc_data, rw);
1203         if (unlikely(!rq)) {
1204                 __blk_mq_run_hw_queue(hctx);
1205                 blk_mq_put_ctx(ctx);
1206                 trace_block_sleeprq(q, bio, rw);
1207
1208                 ctx = blk_mq_get_ctx(q);
1209                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1210                 blk_mq_set_alloc_data(&alloc_data, q,
1211                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1212                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1213                 ctx = alloc_data.ctx;
1214                 hctx = alloc_data.hctx;
1215         }
1216
1217         hctx->queued++;
1218         data->hctx = hctx;
1219         data->ctx = ctx;
1220         return rq;
1221 }
1222
1223 static int blk_mq_direct_issue_request(struct request *rq)
1224 {
1225         int ret;
1226         struct request_queue *q = rq->q;
1227         struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1228                         rq->mq_ctx->cpu);
1229         struct blk_mq_queue_data bd = {
1230                 .rq = rq,
1231                 .list = NULL,
1232                 .last = 1
1233         };
1234
1235         /*
1236          * For OK queue, we are done. For error, kill it. Any other
1237          * error (busy), just add it to our list as we previously
1238          * would have done
1239          */
1240         ret = q->mq_ops->queue_rq(hctx, &bd);
1241         if (ret == BLK_MQ_RQ_QUEUE_OK)
1242                 return 0;
1243         else {
1244                 __blk_mq_requeue_request(rq);
1245
1246                 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1247                         rq->errors = -EIO;
1248                         blk_mq_end_request(rq, rq->errors);
1249                         return 0;
1250                 }
1251                 return -1;
1252         }
1253 }
1254
1255 /*
1256  * Multiple hardware queue variant. This will not use per-process plugs,
1257  * but will attempt to bypass the hctx queueing if we can go straight to
1258  * hardware for SYNC IO.
1259  */
1260 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1261 {
1262         const int is_sync = rw_is_sync(bio->bi_rw);
1263         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1264         struct blk_map_ctx data;
1265         struct request *rq;
1266         unsigned int request_count = 0;
1267         struct blk_plug *plug;
1268         struct request *same_queue_rq = NULL;
1269
1270         blk_queue_bounce(q, &bio);
1271
1272         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1273                 bio_endio(bio, -EIO);
1274                 return;
1275         }
1276
1277         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1278             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1279                 return;
1280
1281         rq = blk_mq_map_request(q, bio, &data);
1282         if (unlikely(!rq))
1283                 return;
1284
1285         if (unlikely(is_flush_fua)) {
1286                 blk_mq_bio_to_request(rq, bio);
1287                 blk_insert_flush(rq);
1288                 goto run_queue;
1289         }
1290
1291         plug = current->plug;
1292         /*
1293          * If the driver supports defer issued based on 'last', then
1294          * queue it up like normal since we can potentially save some
1295          * CPU this way.
1296          */
1297         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1298             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1299                 struct request *old_rq = NULL;
1300
1301                 blk_mq_bio_to_request(rq, bio);
1302
1303                 /*
1304                  * we do limited pluging. If bio can be merged, do merge.
1305                  * Otherwise the existing request in the plug list will be
1306                  * issued. So the plug list will have one request at most
1307                  */
1308                 if (plug) {
1309                         /*
1310                          * The plug list might get flushed before this. If that
1311                          * happens, same_queue_rq is invalid and plug list is empty
1312                          **/
1313                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1314                                 old_rq = same_queue_rq;
1315                                 list_del_init(&old_rq->queuelist);
1316                         }
1317                         list_add_tail(&rq->queuelist, &plug->mq_list);
1318                 } else /* is_sync */
1319                         old_rq = rq;
1320                 blk_mq_put_ctx(data.ctx);
1321                 if (!old_rq)
1322                         return;
1323                 if (!blk_mq_direct_issue_request(old_rq))
1324                         return;
1325                 blk_mq_insert_request(old_rq, false, true, true);
1326                 return;
1327         }
1328
1329         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1330                 /*
1331                  * For a SYNC request, send it to the hardware immediately. For
1332                  * an ASYNC request, just ensure that we run it later on. The
1333                  * latter allows for merging opportunities and more efficient
1334                  * dispatching.
1335                  */
1336 run_queue:
1337                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1338         }
1339         blk_mq_put_ctx(data.ctx);
1340 }
1341
1342 /*
1343  * Single hardware queue variant. This will attempt to use any per-process
1344  * plug for merging and IO deferral.
1345  */
1346 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1347 {
1348         const int is_sync = rw_is_sync(bio->bi_rw);
1349         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1350         struct blk_plug *plug;
1351         unsigned int request_count = 0;
1352         struct blk_map_ctx data;
1353         struct request *rq;
1354
1355         blk_queue_bounce(q, &bio);
1356
1357         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1358                 bio_endio(bio, -EIO);
1359                 return;
1360         }
1361
1362         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1363             blk_attempt_plug_merge(q, bio, &request_count, NULL))
1364                 return;
1365
1366         rq = blk_mq_map_request(q, bio, &data);
1367         if (unlikely(!rq))
1368                 return;
1369
1370         if (unlikely(is_flush_fua)) {
1371                 blk_mq_bio_to_request(rq, bio);
1372                 blk_insert_flush(rq);
1373                 goto run_queue;
1374         }
1375
1376         /*
1377          * A task plug currently exists. Since this is completely lockless,
1378          * utilize that to temporarily store requests until the task is
1379          * either done or scheduled away.
1380          */
1381         plug = current->plug;
1382         if (plug) {
1383                 blk_mq_bio_to_request(rq, bio);
1384                 if (list_empty(&plug->mq_list))
1385                         trace_block_plug(q);
1386                 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1387                         blk_flush_plug_list(plug, false);
1388                         trace_block_plug(q);
1389                 }
1390                 list_add_tail(&rq->queuelist, &plug->mq_list);
1391                 blk_mq_put_ctx(data.ctx);
1392                 return;
1393         }
1394
1395         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1396                 /*
1397                  * For a SYNC request, send it to the hardware immediately. For
1398                  * an ASYNC request, just ensure that we run it later on. The
1399                  * latter allows for merging opportunities and more efficient
1400                  * dispatching.
1401                  */
1402 run_queue:
1403                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1404         }
1405
1406         blk_mq_put_ctx(data.ctx);
1407 }
1408
1409 /*
1410  * Default mapping to a software queue, since we use one per CPU.
1411  */
1412 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1413 {
1414         return q->queue_hw_ctx[q->mq_map[cpu]];
1415 }
1416 EXPORT_SYMBOL(blk_mq_map_queue);
1417
1418 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1419                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1420 {
1421         struct page *page;
1422
1423         if (tags->rqs && set->ops->exit_request) {
1424                 int i;
1425
1426                 for (i = 0; i < tags->nr_tags; i++) {
1427                         if (!tags->rqs[i])
1428                                 continue;
1429                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1430                                                 hctx_idx, i);
1431                         tags->rqs[i] = NULL;
1432                 }
1433         }
1434
1435         while (!list_empty(&tags->page_list)) {
1436                 page = list_first_entry(&tags->page_list, struct page, lru);
1437                 list_del_init(&page->lru);
1438                 __free_pages(page, page->private);
1439         }
1440
1441         kfree(tags->rqs);
1442
1443         blk_mq_free_tags(tags);
1444 }
1445
1446 static size_t order_to_size(unsigned int order)
1447 {
1448         return (size_t)PAGE_SIZE << order;
1449 }
1450
1451 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1452                 unsigned int hctx_idx)
1453 {
1454         struct blk_mq_tags *tags;
1455         unsigned int i, j, entries_per_page, max_order = 4;
1456         size_t rq_size, left;
1457
1458         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1459                                 set->numa_node,
1460                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1461         if (!tags)
1462                 return NULL;
1463
1464         INIT_LIST_HEAD(&tags->page_list);
1465
1466         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1467                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1468                                  set->numa_node);
1469         if (!tags->rqs) {
1470                 blk_mq_free_tags(tags);
1471                 return NULL;
1472         }
1473
1474         /*
1475          * rq_size is the size of the request plus driver payload, rounded
1476          * to the cacheline size
1477          */
1478         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1479                                 cache_line_size());
1480         left = rq_size * set->queue_depth;
1481
1482         for (i = 0; i < set->queue_depth; ) {
1483                 int this_order = max_order;
1484                 struct page *page;
1485                 int to_do;
1486                 void *p;
1487
1488                 while (left < order_to_size(this_order - 1) && this_order)
1489                         this_order--;
1490
1491                 do {
1492                         page = alloc_pages_node(set->numa_node,
1493                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1494                                 this_order);
1495                         if (page)
1496                                 break;
1497                         if (!this_order--)
1498                                 break;
1499                         if (order_to_size(this_order) < rq_size)
1500                                 break;
1501                 } while (1);
1502
1503                 if (!page)
1504                         goto fail;
1505
1506                 page->private = this_order;
1507                 list_add_tail(&page->lru, &tags->page_list);
1508
1509                 p = page_address(page);
1510                 entries_per_page = order_to_size(this_order) / rq_size;
1511                 to_do = min(entries_per_page, set->queue_depth - i);
1512                 left -= to_do * rq_size;
1513                 for (j = 0; j < to_do; j++) {
1514                         tags->rqs[i] = p;
1515                         if (set->ops->init_request) {
1516                                 if (set->ops->init_request(set->driver_data,
1517                                                 tags->rqs[i], hctx_idx, i,
1518                                                 set->numa_node)) {
1519                                         tags->rqs[i] = NULL;
1520                                         goto fail;
1521                                 }
1522                         }
1523
1524                         p += rq_size;
1525                         i++;
1526                 }
1527         }
1528         return tags;
1529
1530 fail:
1531         blk_mq_free_rq_map(set, tags, hctx_idx);
1532         return NULL;
1533 }
1534
1535 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1536 {
1537         kfree(bitmap->map);
1538 }
1539
1540 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1541 {
1542         unsigned int bpw = 8, total, num_maps, i;
1543
1544         bitmap->bits_per_word = bpw;
1545
1546         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1547         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1548                                         GFP_KERNEL, node);
1549         if (!bitmap->map)
1550                 return -ENOMEM;
1551
1552         total = nr_cpu_ids;
1553         for (i = 0; i < num_maps; i++) {
1554                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1555                 total -= bitmap->map[i].depth;
1556         }
1557
1558         return 0;
1559 }
1560
1561 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1562 {
1563         struct request_queue *q = hctx->queue;
1564         struct blk_mq_ctx *ctx;
1565         LIST_HEAD(tmp);
1566
1567         /*
1568          * Move ctx entries to new CPU, if this one is going away.
1569          */
1570         ctx = __blk_mq_get_ctx(q, cpu);
1571
1572         spin_lock(&ctx->lock);
1573         if (!list_empty(&ctx->rq_list)) {
1574                 list_splice_init(&ctx->rq_list, &tmp);
1575                 blk_mq_hctx_clear_pending(hctx, ctx);
1576         }
1577         spin_unlock(&ctx->lock);
1578
1579         if (list_empty(&tmp))
1580                 return NOTIFY_OK;
1581
1582         ctx = blk_mq_get_ctx(q);
1583         spin_lock(&ctx->lock);
1584
1585         while (!list_empty(&tmp)) {
1586                 struct request *rq;
1587
1588                 rq = list_first_entry(&tmp, struct request, queuelist);
1589                 rq->mq_ctx = ctx;
1590                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1591         }
1592
1593         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1594         blk_mq_hctx_mark_pending(hctx, ctx);
1595
1596         spin_unlock(&ctx->lock);
1597
1598         blk_mq_run_hw_queue(hctx, true);
1599         blk_mq_put_ctx(ctx);
1600         return NOTIFY_OK;
1601 }
1602
1603 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1604 {
1605         struct request_queue *q = hctx->queue;
1606         struct blk_mq_tag_set *set = q->tag_set;
1607
1608         if (set->tags[hctx->queue_num])
1609                 return NOTIFY_OK;
1610
1611         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1612         if (!set->tags[hctx->queue_num])
1613                 return NOTIFY_STOP;
1614
1615         hctx->tags = set->tags[hctx->queue_num];
1616         return NOTIFY_OK;
1617 }
1618
1619 static int blk_mq_hctx_notify(void *data, unsigned long action,
1620                               unsigned int cpu)
1621 {
1622         struct blk_mq_hw_ctx *hctx = data;
1623
1624         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1625                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1626         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1627                 return blk_mq_hctx_cpu_online(hctx, cpu);
1628
1629         return NOTIFY_OK;
1630 }
1631
1632 static void blk_mq_exit_hctx(struct request_queue *q,
1633                 struct blk_mq_tag_set *set,
1634                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1635 {
1636         unsigned flush_start_tag = set->queue_depth;
1637
1638         blk_mq_tag_idle(hctx);
1639
1640         if (set->ops->exit_request)
1641                 set->ops->exit_request(set->driver_data,
1642                                        hctx->fq->flush_rq, hctx_idx,
1643                                        flush_start_tag + hctx_idx);
1644
1645         if (set->ops->exit_hctx)
1646                 set->ops->exit_hctx(hctx, hctx_idx);
1647
1648         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1649         blk_free_flush_queue(hctx->fq);
1650         kfree(hctx->ctxs);
1651         blk_mq_free_bitmap(&hctx->ctx_map);
1652 }
1653
1654 static void blk_mq_exit_hw_queues(struct request_queue *q,
1655                 struct blk_mq_tag_set *set, int nr_queue)
1656 {
1657         struct blk_mq_hw_ctx *hctx;
1658         unsigned int i;
1659
1660         queue_for_each_hw_ctx(q, hctx, i) {
1661                 if (i == nr_queue)
1662                         break;
1663                 blk_mq_exit_hctx(q, set, hctx, i);
1664         }
1665 }
1666
1667 static void blk_mq_free_hw_queues(struct request_queue *q,
1668                 struct blk_mq_tag_set *set)
1669 {
1670         struct blk_mq_hw_ctx *hctx;
1671         unsigned int i;
1672
1673         queue_for_each_hw_ctx(q, hctx, i)
1674                 free_cpumask_var(hctx->cpumask);
1675 }
1676
1677 static int blk_mq_init_hctx(struct request_queue *q,
1678                 struct blk_mq_tag_set *set,
1679                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1680 {
1681         int node;
1682         unsigned flush_start_tag = set->queue_depth;
1683
1684         node = hctx->numa_node;
1685         if (node == NUMA_NO_NODE)
1686                 node = hctx->numa_node = set->numa_node;
1687
1688         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1689         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1690         spin_lock_init(&hctx->lock);
1691         INIT_LIST_HEAD(&hctx->dispatch);
1692         hctx->queue = q;
1693         hctx->queue_num = hctx_idx;
1694         hctx->flags = set->flags;
1695
1696         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1697                                         blk_mq_hctx_notify, hctx);
1698         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1699
1700         hctx->tags = set->tags[hctx_idx];
1701
1702         /*
1703          * Allocate space for all possible cpus to avoid allocation at
1704          * runtime
1705          */
1706         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1707                                         GFP_KERNEL, node);
1708         if (!hctx->ctxs)
1709                 goto unregister_cpu_notifier;
1710
1711         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1712                 goto free_ctxs;
1713
1714         hctx->nr_ctx = 0;
1715
1716         if (set->ops->init_hctx &&
1717             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1718                 goto free_bitmap;
1719
1720         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1721         if (!hctx->fq)
1722                 goto exit_hctx;
1723
1724         if (set->ops->init_request &&
1725             set->ops->init_request(set->driver_data,
1726                                    hctx->fq->flush_rq, hctx_idx,
1727                                    flush_start_tag + hctx_idx, node))
1728                 goto free_fq;
1729
1730         return 0;
1731
1732  free_fq:
1733         kfree(hctx->fq);
1734  exit_hctx:
1735         if (set->ops->exit_hctx)
1736                 set->ops->exit_hctx(hctx, hctx_idx);
1737  free_bitmap:
1738         blk_mq_free_bitmap(&hctx->ctx_map);
1739  free_ctxs:
1740         kfree(hctx->ctxs);
1741  unregister_cpu_notifier:
1742         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1743
1744         return -1;
1745 }
1746
1747 static int blk_mq_init_hw_queues(struct request_queue *q,
1748                 struct blk_mq_tag_set *set)
1749 {
1750         struct blk_mq_hw_ctx *hctx;
1751         unsigned int i;
1752
1753         /*
1754          * Initialize hardware queues
1755          */
1756         queue_for_each_hw_ctx(q, hctx, i) {
1757                 if (blk_mq_init_hctx(q, set, hctx, i))
1758                         break;
1759         }
1760
1761         if (i == q->nr_hw_queues)
1762                 return 0;
1763
1764         /*
1765          * Init failed
1766          */
1767         blk_mq_exit_hw_queues(q, set, i);
1768
1769         return 1;
1770 }
1771
1772 static void blk_mq_init_cpu_queues(struct request_queue *q,
1773                                    unsigned int nr_hw_queues)
1774 {
1775         unsigned int i;
1776
1777         for_each_possible_cpu(i) {
1778                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1779                 struct blk_mq_hw_ctx *hctx;
1780
1781                 memset(__ctx, 0, sizeof(*__ctx));
1782                 __ctx->cpu = i;
1783                 spin_lock_init(&__ctx->lock);
1784                 INIT_LIST_HEAD(&__ctx->rq_list);
1785                 __ctx->queue = q;
1786
1787                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1788                 if (!cpu_online(i))
1789                         continue;
1790
1791                 hctx = q->mq_ops->map_queue(q, i);
1792
1793                 /*
1794                  * Set local node, IFF we have more than one hw queue. If
1795                  * not, we remain on the home node of the device
1796                  */
1797                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1798                         hctx->numa_node = cpu_to_node(i);
1799         }
1800 }
1801
1802 static void blk_mq_map_swqueue(struct request_queue *q)
1803 {
1804         unsigned int i;
1805         struct blk_mq_hw_ctx *hctx;
1806         struct blk_mq_ctx *ctx;
1807
1808         queue_for_each_hw_ctx(q, hctx, i) {
1809                 cpumask_clear(hctx->cpumask);
1810                 hctx->nr_ctx = 0;
1811         }
1812
1813         /*
1814          * Map software to hardware queues
1815          */
1816         queue_for_each_ctx(q, ctx, i) {
1817                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1818                 if (!cpu_online(i))
1819                         continue;
1820
1821                 hctx = q->mq_ops->map_queue(q, i);
1822                 cpumask_set_cpu(i, hctx->cpumask);
1823                 cpumask_set_cpu(i, hctx->tags->cpumask);
1824                 ctx->index_hw = hctx->nr_ctx;
1825                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1826         }
1827
1828         queue_for_each_hw_ctx(q, hctx, i) {
1829                 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1830
1831                 /*
1832                  * If no software queues are mapped to this hardware queue,
1833                  * disable it and free the request entries.
1834                  */
1835                 if (!hctx->nr_ctx) {
1836                         struct blk_mq_tag_set *set = q->tag_set;
1837
1838                         if (set->tags[i]) {
1839                                 blk_mq_free_rq_map(set, set->tags[i], i);
1840                                 set->tags[i] = NULL;
1841                                 hctx->tags = NULL;
1842                         }
1843                         continue;
1844                 }
1845
1846                 /*
1847                  * Set the map size to the number of mapped software queues.
1848                  * This is more accurate and more efficient than looping
1849                  * over all possibly mapped software queues.
1850                  */
1851                 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1852
1853                 /*
1854                  * Initialize batch roundrobin counts
1855                  */
1856                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1857                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1858         }
1859 }
1860
1861 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1862 {
1863         struct blk_mq_hw_ctx *hctx;
1864         struct request_queue *q;
1865         bool shared;
1866         int i;
1867
1868         if (set->tag_list.next == set->tag_list.prev)
1869                 shared = false;
1870         else
1871                 shared = true;
1872
1873         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1874                 blk_mq_freeze_queue(q);
1875
1876                 queue_for_each_hw_ctx(q, hctx, i) {
1877                         if (shared)
1878                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1879                         else
1880                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1881                 }
1882                 blk_mq_unfreeze_queue(q);
1883         }
1884 }
1885
1886 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1887 {
1888         struct blk_mq_tag_set *set = q->tag_set;
1889
1890         mutex_lock(&set->tag_list_lock);
1891         list_del_init(&q->tag_set_list);
1892         blk_mq_update_tag_set_depth(set);
1893         mutex_unlock(&set->tag_list_lock);
1894 }
1895
1896 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1897                                      struct request_queue *q)
1898 {
1899         q->tag_set = set;
1900
1901         mutex_lock(&set->tag_list_lock);
1902         list_add_tail(&q->tag_set_list, &set->tag_list);
1903         blk_mq_update_tag_set_depth(set);
1904         mutex_unlock(&set->tag_list_lock);
1905 }
1906
1907 /*
1908  * It is the actual release handler for mq, but we do it from
1909  * request queue's release handler for avoiding use-after-free
1910  * and headache because q->mq_kobj shouldn't have been introduced,
1911  * but we can't group ctx/kctx kobj without it.
1912  */
1913 void blk_mq_release(struct request_queue *q)
1914 {
1915         struct blk_mq_hw_ctx *hctx;
1916         unsigned int i;
1917
1918         /* hctx kobj stays in hctx */
1919         queue_for_each_hw_ctx(q, hctx, i)
1920                 kfree(hctx);
1921
1922         kfree(q->queue_hw_ctx);
1923
1924         /* ctx kobj stays in queue_ctx */
1925         free_percpu(q->queue_ctx);
1926 }
1927
1928 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1929 {
1930         struct request_queue *uninit_q, *q;
1931
1932         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1933         if (!uninit_q)
1934                 return ERR_PTR(-ENOMEM);
1935
1936         q = blk_mq_init_allocated_queue(set, uninit_q);
1937         if (IS_ERR(q))
1938                 blk_cleanup_queue(uninit_q);
1939
1940         return q;
1941 }
1942 EXPORT_SYMBOL(blk_mq_init_queue);
1943
1944 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1945                                                   struct request_queue *q)
1946 {
1947         struct blk_mq_hw_ctx **hctxs;
1948         struct blk_mq_ctx __percpu *ctx;
1949         unsigned int *map;
1950         int i;
1951
1952         ctx = alloc_percpu(struct blk_mq_ctx);
1953         if (!ctx)
1954                 return ERR_PTR(-ENOMEM);
1955
1956         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1957                         set->numa_node);
1958
1959         if (!hctxs)
1960                 goto err_percpu;
1961
1962         map = blk_mq_make_queue_map(set);
1963         if (!map)
1964                 goto err_map;
1965
1966         for (i = 0; i < set->nr_hw_queues; i++) {
1967                 int node = blk_mq_hw_queue_to_node(map, i);
1968
1969                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1970                                         GFP_KERNEL, node);
1971                 if (!hctxs[i])
1972                         goto err_hctxs;
1973
1974                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1975                                                 node))
1976                         goto err_hctxs;
1977
1978                 atomic_set(&hctxs[i]->nr_active, 0);
1979                 hctxs[i]->numa_node = node;
1980                 hctxs[i]->queue_num = i;
1981         }
1982
1983         /*
1984          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1985          * See blk_register_queue() for details.
1986          */
1987         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1988                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1989                 goto err_hctxs;
1990
1991         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1992         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30000);
1993
1994         q->nr_queues = nr_cpu_ids;
1995         q->nr_hw_queues = set->nr_hw_queues;
1996         q->mq_map = map;
1997
1998         q->queue_ctx = ctx;
1999         q->queue_hw_ctx = hctxs;
2000
2001         q->mq_ops = set->ops;
2002         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2003
2004         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2005                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2006
2007         q->sg_reserved_size = INT_MAX;
2008
2009         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2010         INIT_LIST_HEAD(&q->requeue_list);
2011         spin_lock_init(&q->requeue_lock);
2012
2013         if (q->nr_hw_queues > 1)
2014                 blk_queue_make_request(q, blk_mq_make_request);
2015         else
2016                 blk_queue_make_request(q, blk_sq_make_request);
2017
2018         /*
2019          * Do this after blk_queue_make_request() overrides it...
2020          */
2021         q->nr_requests = set->queue_depth;
2022
2023         if (set->ops->complete)
2024                 blk_queue_softirq_done(q, set->ops->complete);
2025
2026         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2027
2028         if (blk_mq_init_hw_queues(q, set))
2029                 goto err_hctxs;
2030
2031         mutex_lock(&all_q_mutex);
2032         list_add_tail(&q->all_q_node, &all_q_list);
2033         mutex_unlock(&all_q_mutex);
2034
2035         blk_mq_add_queue_tag_set(set, q);
2036
2037         blk_mq_map_swqueue(q);
2038
2039         return q;
2040
2041 err_hctxs:
2042         kfree(map);
2043         for (i = 0; i < set->nr_hw_queues; i++) {
2044                 if (!hctxs[i])
2045                         break;
2046                 free_cpumask_var(hctxs[i]->cpumask);
2047                 kfree(hctxs[i]);
2048         }
2049 err_map:
2050         kfree(hctxs);
2051 err_percpu:
2052         free_percpu(ctx);
2053         return ERR_PTR(-ENOMEM);
2054 }
2055 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2056
2057 void blk_mq_free_queue(struct request_queue *q)
2058 {
2059         struct blk_mq_tag_set   *set = q->tag_set;
2060
2061         blk_mq_del_queue_tag_set(q);
2062
2063         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2064         blk_mq_free_hw_queues(q, set);
2065
2066         percpu_ref_exit(&q->mq_usage_counter);
2067
2068         kfree(q->mq_map);
2069
2070         q->mq_map = NULL;
2071
2072         mutex_lock(&all_q_mutex);
2073         list_del_init(&q->all_q_node);
2074         mutex_unlock(&all_q_mutex);
2075 }
2076
2077 /* Basically redo blk_mq_init_queue with queue frozen */
2078 static void blk_mq_queue_reinit(struct request_queue *q)
2079 {
2080         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2081
2082         blk_mq_sysfs_unregister(q);
2083
2084         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2085
2086         /*
2087          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2088          * we should change hctx numa_node according to new topology (this
2089          * involves free and re-allocate memory, worthy doing?)
2090          */
2091
2092         blk_mq_map_swqueue(q);
2093
2094         blk_mq_sysfs_register(q);
2095 }
2096
2097 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2098                                       unsigned long action, void *hcpu)
2099 {
2100         struct request_queue *q;
2101
2102         /*
2103          * Before new mappings are established, hotadded cpu might already
2104          * start handling requests. This doesn't break anything as we map
2105          * offline CPUs to first hardware queue. We will re-init the queue
2106          * below to get optimal settings.
2107          */
2108         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2109             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2110                 return NOTIFY_OK;
2111
2112         mutex_lock(&all_q_mutex);
2113
2114         /*
2115          * We need to freeze and reinit all existing queues.  Freezing
2116          * involves synchronous wait for an RCU grace period and doing it
2117          * one by one may take a long time.  Start freezing all queues in
2118          * one swoop and then wait for the completions so that freezing can
2119          * take place in parallel.
2120          */
2121         list_for_each_entry(q, &all_q_list, all_q_node)
2122                 blk_mq_freeze_queue_start(q);
2123         list_for_each_entry(q, &all_q_list, all_q_node)
2124                 blk_mq_freeze_queue_wait(q);
2125
2126         list_for_each_entry(q, &all_q_list, all_q_node)
2127                 blk_mq_queue_reinit(q);
2128
2129         list_for_each_entry(q, &all_q_list, all_q_node)
2130                 blk_mq_unfreeze_queue(q);
2131
2132         mutex_unlock(&all_q_mutex);
2133         return NOTIFY_OK;
2134 }
2135
2136 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2137 {
2138         int i;
2139
2140         for (i = 0; i < set->nr_hw_queues; i++) {
2141                 set->tags[i] = blk_mq_init_rq_map(set, i);
2142                 if (!set->tags[i])
2143                         goto out_unwind;
2144         }
2145
2146         return 0;
2147
2148 out_unwind:
2149         while (--i >= 0)
2150                 blk_mq_free_rq_map(set, set->tags[i], i);
2151
2152         return -ENOMEM;
2153 }
2154
2155 /*
2156  * Allocate the request maps associated with this tag_set. Note that this
2157  * may reduce the depth asked for, if memory is tight. set->queue_depth
2158  * will be updated to reflect the allocated depth.
2159  */
2160 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2161 {
2162         unsigned int depth;
2163         int err;
2164
2165         depth = set->queue_depth;
2166         do {
2167                 err = __blk_mq_alloc_rq_maps(set);
2168                 if (!err)
2169                         break;
2170
2171                 set->queue_depth >>= 1;
2172                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2173                         err = -ENOMEM;
2174                         break;
2175                 }
2176         } while (set->queue_depth);
2177
2178         if (!set->queue_depth || err) {
2179                 pr_err("blk-mq: failed to allocate request map\n");
2180                 return -ENOMEM;
2181         }
2182
2183         if (depth != set->queue_depth)
2184                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2185                                                 depth, set->queue_depth);
2186
2187         return 0;
2188 }
2189
2190 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2191 {
2192         return tags->cpumask;
2193 }
2194 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2195
2196 /*
2197  * Alloc a tag set to be associated with one or more request queues.
2198  * May fail with EINVAL for various error conditions. May adjust the
2199  * requested depth down, if if it too large. In that case, the set
2200  * value will be stored in set->queue_depth.
2201  */
2202 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2203 {
2204         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2205
2206         if (!set->nr_hw_queues)
2207                 return -EINVAL;
2208         if (!set->queue_depth)
2209                 return -EINVAL;
2210         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2211                 return -EINVAL;
2212
2213         if (!set->ops->queue_rq || !set->ops->map_queue)
2214                 return -EINVAL;
2215
2216         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2217                 pr_info("blk-mq: reduced tag depth to %u\n",
2218                         BLK_MQ_MAX_DEPTH);
2219                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2220         }
2221
2222         /*
2223          * If a crashdump is active, then we are potentially in a very
2224          * memory constrained environment. Limit us to 1 queue and
2225          * 64 tags to prevent using too much memory.
2226          */
2227         if (is_kdump_kernel()) {
2228                 set->nr_hw_queues = 1;
2229                 set->queue_depth = min(64U, set->queue_depth);
2230         }
2231
2232         set->tags = kmalloc_node(set->nr_hw_queues *
2233                                  sizeof(struct blk_mq_tags *),
2234                                  GFP_KERNEL, set->numa_node);
2235         if (!set->tags)
2236                 return -ENOMEM;
2237
2238         if (blk_mq_alloc_rq_maps(set))
2239                 goto enomem;
2240
2241         mutex_init(&set->tag_list_lock);
2242         INIT_LIST_HEAD(&set->tag_list);
2243
2244         return 0;
2245 enomem:
2246         kfree(set->tags);
2247         set->tags = NULL;
2248         return -ENOMEM;
2249 }
2250 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2251
2252 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2253 {
2254         int i;
2255
2256         for (i = 0; i < set->nr_hw_queues; i++) {
2257                 if (set->tags[i]) {
2258                         blk_mq_free_rq_map(set, set->tags[i], i);
2259                         free_cpumask_var(set->tags[i]->cpumask);
2260                 }
2261         }
2262
2263         kfree(set->tags);
2264         set->tags = NULL;
2265 }
2266 EXPORT_SYMBOL(blk_mq_free_tag_set);
2267
2268 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2269 {
2270         struct blk_mq_tag_set *set = q->tag_set;
2271         struct blk_mq_hw_ctx *hctx;
2272         int i, ret;
2273
2274         if (!set || nr > set->queue_depth)
2275                 return -EINVAL;
2276
2277         ret = 0;
2278         queue_for_each_hw_ctx(q, hctx, i) {
2279                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2280                 if (ret)
2281                         break;
2282         }
2283
2284         if (!ret)
2285                 q->nr_requests = nr;
2286
2287         return ret;
2288 }
2289
2290 void blk_mq_disable_hotplug(void)
2291 {
2292         mutex_lock(&all_q_mutex);
2293 }
2294
2295 void blk_mq_enable_hotplug(void)
2296 {
2297         mutex_unlock(&all_q_mutex);
2298 }
2299
2300 static int __init blk_mq_init(void)
2301 {
2302         blk_mq_cpu_init();
2303
2304         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2305
2306         return 0;
2307 }
2308 subsys_initcall(blk_mq_init);