kernel/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 = swait_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         swait_wake_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         swait_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                 swait_wake_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         swait_wake_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 #ifdef CONFIG_PREEMPT_RT_FULL
 221         INIT_WORK(&rq->work, __blk_mq_complete_request_remote_work);
 222 #endif
 223         INIT_LIST_HEAD(&rq->timeout_list);
 224         rq->timeout = 0;
 225
 226         rq->end_io = NULL;
 227         rq->end_io_data = NULL;
 228         rq->next_rq = NULL;
 229
 230         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 231 }
 232
 233 static struct request *
 234 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 235 {
 236         struct request *rq;
 237         unsigned int tag;
 238
 239         tag = blk_mq_get_tag(data);
 240         if (tag != BLK_MQ_TAG_FAIL) {
 241                 rq = data->hctx->tags->rqs[tag];
 242
 243                 if (blk_mq_tag_busy(data->hctx)) {
 244                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 245                         atomic_inc(&data->hctx->nr_active);
 246                 }
 247
 248                 rq->tag = tag;
 249                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 250                 return rq;
 251         }
 252
 253         return NULL;
 254 }
 255
 256 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 257                 bool reserved)
 258 {
 259         struct blk_mq_ctx *ctx;
 260         struct blk_mq_hw_ctx *hctx;
 261         struct request *rq;
 262         struct blk_mq_alloc_data alloc_data;
 263         int ret;
 264
 265         ret = blk_mq_queue_enter(q, gfp);
 266         if (ret)
 267                 return ERR_PTR(ret);
 268
 269         ctx = blk_mq_get_ctx(q);
 270         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 271         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 272                         reserved, ctx, hctx);
 273
 274         rq = __blk_mq_alloc_request(&alloc_data, rw);
 275         if (!rq && (gfp & __GFP_WAIT)) {
 276                 __blk_mq_run_hw_queue(hctx);
 277                 blk_mq_put_ctx(ctx);
 278
 279                 ctx = blk_mq_get_ctx(q);
 280                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 281                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 282                                 hctx);
 283                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 284                 ctx = alloc_data.ctx;
 285         }
 286         blk_mq_put_ctx(ctx);
 287         if (!rq) {
 288                 blk_mq_queue_exit(q);
 289                 return ERR_PTR(-EWOULDBLOCK);
 290         }
 291         return rq;
 292 }
 293 EXPORT_SYMBOL(blk_mq_alloc_request);
 294
 295 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 296                                   struct blk_mq_ctx *ctx, struct request *rq)
 297 {
 298         const int tag = rq->tag;
 299         struct request_queue *q = rq->q;
 300
 301         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 302                 atomic_dec(&hctx->nr_active);
 303         rq->cmd_flags = 0;
 304
 305         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 306         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 307         blk_mq_queue_exit(q);
 308 }
 309
 310 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 311 {
 312         struct blk_mq_ctx *ctx = rq->mq_ctx;
 313
 314         ctx->rq_completed[rq_is_sync(rq)]++;
 315         __blk_mq_free_request(hctx, ctx, rq);
 316
 317 }
 318 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 319
 320 void blk_mq_free_request(struct request *rq)
 321 {
 322         struct blk_mq_hw_ctx *hctx;
 323         struct request_queue *q = rq->q;
 324
 325         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 326         blk_mq_free_hctx_request(hctx, rq);
 327 }
 328 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 329
 330 inline void __blk_mq_end_request(struct request *rq, int error)
 331 {
 332         blk_account_io_done(rq);
 333
 334         if (rq->end_io) {
 335                 rq->end_io(rq, error);
 336         } else {
 337                 if (unlikely(blk_bidi_rq(rq)))
 338                         blk_mq_free_request(rq->next_rq);
 339                 blk_mq_free_request(rq);
 340         }
 341 }
 342 EXPORT_SYMBOL(__blk_mq_end_request);
 343
 344 void blk_mq_end_request(struct request *rq, int error)
 345 {
 346         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 347                 BUG();
 348         __blk_mq_end_request(rq, error);
 349 }
 350 EXPORT_SYMBOL(blk_mq_end_request);
 351
 352 #ifdef CONFIG_PREEMPT_RT_FULL
 353
 354 void __blk_mq_complete_request_remote_work(struct work_struct *work)
 355 {
 356         struct request *rq = container_of(work, struct request, work);
 357
 358         rq->q->softirq_done_fn(rq);
 359 }
 360
 361 #else
 362
 363 static void __blk_mq_complete_request_remote(void *data)
 364 {
 365         struct request *rq = data;
 366
 367         rq->q->softirq_done_fn(rq);
 368 }
 369
 370 #endif
 371
 372 static void blk_mq_ipi_complete_request(struct request *rq)
 373 {
 374         struct blk_mq_ctx *ctx = rq->mq_ctx;
 375         bool shared = false;
 376         int cpu;
 377
 378         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 379                 rq->q->softirq_done_fn(rq);
 380                 return;
 381         }
 382
 383         cpu = get_cpu_light();
 384         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 385                 shared = cpus_share_cache(cpu, ctx->cpu);
 386
 387         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 388 #ifdef CONFIG_PREEMPT_RT_FULL
 389                 schedule_work_on(ctx->cpu, &rq->work);
 390 #else
 391                 rq->csd.func = __blk_mq_complete_request_remote;
 392                 rq->csd.info = rq;
 393                 rq->csd.flags = 0;
 394                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 395 #endif
 396         } else {
 397                 rq->q->softirq_done_fn(rq);
 398         }
 399         put_cpu_light();
 400 }
 401
 402 void __blk_mq_complete_request(struct request *rq)
 403 {
 404         struct request_queue *q = rq->q;
 405
 406         if (!q->softirq_done_fn)
 407                 blk_mq_end_request(rq, rq->errors);
 408         else
 409                 blk_mq_ipi_complete_request(rq);
 410 }
 411
 412 /**
 413  * blk_mq_complete_request - end I/O on a request
 414  * @rq:         the request being processed
 415  *
 416  * Description:
 417  *      Ends all I/O on a request. It does not handle partial completions.
 418  *      The actual completion happens out-of-order, through a IPI handler.
 419  **/
 420 void blk_mq_complete_request(struct request *rq)
 421 {
 422         struct request_queue *q = rq->q;
 423
 424         if (unlikely(blk_should_fake_timeout(q)))
 425                 return;
 426         if (!blk_mark_rq_complete(rq))
 427                 __blk_mq_complete_request(rq);
 428 }
 429 EXPORT_SYMBOL(blk_mq_complete_request);
 430
 431 int blk_mq_request_started(struct request *rq)
 432 {
 433         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 434 }
 435 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 436
 437 void blk_mq_start_request(struct request *rq)
 438 {
 439         struct request_queue *q = rq->q;
 440
 441         trace_block_rq_issue(q, rq);
 442
 443         rq->resid_len = blk_rq_bytes(rq);
 444         if (unlikely(blk_bidi_rq(rq)))
 445                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 446
 447         blk_add_timer(rq);
 448
 449         /*
 450          * Ensure that ->deadline is visible before set the started
 451          * flag and clear the completed flag.
 452          */
 453         smp_mb__before_atomic();
 454
 455         /*
 456          * Mark us as started and clear complete. Complete might have been
 457          * set if requeue raced with timeout, which then marked it as
 458          * complete. So be sure to clear complete again when we start
 459          * the request, otherwise we'll ignore the completion event.
 460          */
 461         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 462                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 463         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 464                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 465
 466         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 467                 /*
 468                  * Make sure space for the drain appears.  We know we can do
 469                  * this because max_hw_segments has been adjusted to be one
 470                  * fewer than the device can handle.
 471                  */
 472                 rq->nr_phys_segments++;
 473         }
 474 }
 475 EXPORT_SYMBOL(blk_mq_start_request);
 476
 477 static void __blk_mq_requeue_request(struct request *rq)
 478 {
 479         struct request_queue *q = rq->q;
 480
 481         trace_block_rq_requeue(q, rq);
 482
 483         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 484                 if (q->dma_drain_size && blk_rq_bytes(rq))
 485                         rq->nr_phys_segments--;
 486         }
 487 }
 488
 489 void blk_mq_requeue_request(struct request *rq)
 490 {
 491         __blk_mq_requeue_request(rq);
 492
 493         BUG_ON(blk_queued_rq(rq));
 494         blk_mq_add_to_requeue_list(rq, true);
 495 }
 496 EXPORT_SYMBOL(blk_mq_requeue_request);
 497
 498 static void blk_mq_requeue_work(struct work_struct *work)
 499 {
 500         struct request_queue *q =
 501                 container_of(work, struct request_queue, requeue_work);
 502         LIST_HEAD(rq_list);
 503         struct request *rq, *next;
 504         unsigned long flags;
 505
 506         spin_lock_irqsave(&q->requeue_lock, flags);
 507         list_splice_init(&q->requeue_list, &rq_list);
 508         spin_unlock_irqrestore(&q->requeue_lock, flags);
 509
 510         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 511                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 512                         continue;
 513
 514                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 515                 list_del_init(&rq->queuelist);
 516                 blk_mq_insert_request(rq, true, false, false);
 517         }
 518
 519         while (!list_empty(&rq_list)) {
 520                 rq = list_entry(rq_list.next, struct request, queuelist);
 521                 list_del_init(&rq->queuelist);
 522                 blk_mq_insert_request(rq, false, false, false);
 523         }
 524
 525         /*
 526          * Use the start variant of queue running here, so that running
 527          * the requeue work will kick stopped queues.
 528          */
 529         blk_mq_start_hw_queues(q);
 530 }
 531
 532 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 533 {
 534         struct request_queue *q = rq->q;
 535         unsigned long flags;
 536
 537         /*
 538          * We abuse this flag that is otherwise used by the I/O scheduler to
 539          * request head insertation from the workqueue.
 540          */
 541         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 542
 543         spin_lock_irqsave(&q->requeue_lock, flags);
 544         if (at_head) {
 545                 rq->cmd_flags |= REQ_SOFTBARRIER;
 546                 list_add(&rq->queuelist, &q->requeue_list);
 547         } else {
 548                 list_add_tail(&rq->queuelist, &q->requeue_list);
 549         }
 550         spin_unlock_irqrestore(&q->requeue_lock, flags);
 551 }
 552 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 553
 554 void blk_mq_cancel_requeue_work(struct request_queue *q)
 555 {
 556         cancel_work_sync(&q->requeue_work);
 557 }
 558 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 559
 560 void blk_mq_kick_requeue_list(struct request_queue *q)
 561 {
 562         kblockd_schedule_work(&q->requeue_work);
 563 }
 564 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 565
 566 void blk_mq_abort_requeue_list(struct request_queue *q)
 567 {
 568         unsigned long flags;
 569         LIST_HEAD(rq_list);
 570
 571         spin_lock_irqsave(&q->requeue_lock, flags);
 572         list_splice_init(&q->requeue_list, &rq_list);
 573         spin_unlock_irqrestore(&q->requeue_lock, flags);
 574
 575         while (!list_empty(&rq_list)) {
 576                 struct request *rq;
 577
 578                 rq = list_first_entry(&rq_list, struct request, queuelist);
 579                 list_del_init(&rq->queuelist);
 580                 rq->errors = -EIO;
 581                 blk_mq_end_request(rq, rq->errors);
 582         }
 583 }
 584 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 585
 586 static inline bool is_flush_request(struct request *rq,
 587                 struct blk_flush_queue *fq, unsigned int tag)
 588 {
 589         return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
 590                         fq->flush_rq->tag == tag);
 591 }
 592
 593 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 594 {
 595         struct request *rq = tags->rqs[tag];
 596         /* mq_ctx of flush rq is always cloned from the corresponding req */
 597         struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
 598
 599         if (!is_flush_request(rq, fq, tag))
 600                 return rq;
 601
 602         return fq->flush_rq;
 603 }
 604 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 605
 606 struct blk_mq_timeout_data {
 607         unsigned long next;
 608         unsigned int next_set;
 609 };
 610
 611 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 612 {
 613         struct blk_mq_ops *ops = req->q->mq_ops;
 614         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 615
 616         /*
 617          * We know that complete is set at this point. If STARTED isn't set
 618          * anymore, then the request isn't active and the "timeout" should
 619          * just be ignored. This can happen due to the bitflag ordering.
 620          * Timeout first checks if STARTED is set, and if it is, assumes
 621          * the request is active. But if we race with completion, then
 622          * we both flags will get cleared. So check here again, and ignore
 623          * a timeout event with a request that isn't active.
 624          */
 625         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 626                 return;
 627
 628         if (ops->timeout)
 629                 ret = ops->timeout(req, reserved);
 630
 631         switch (ret) {
 632         case BLK_EH_HANDLED:
 633                 __blk_mq_complete_request(req);
 634                 break;
 635         case BLK_EH_RESET_TIMER:
 636                 blk_add_timer(req);
 637                 blk_clear_rq_complete(req);
 638                 break;
 639         case BLK_EH_NOT_HANDLED:
 640                 break;
 641         default:
 642                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 643                 break;
 644         }
 645 }
 646
 647 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 648                 struct request *rq, void *priv, bool reserved)
 649 {
 650         struct blk_mq_timeout_data *data = priv;
 651
 652         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 653                 /*
 654                  * If a request wasn't started before the queue was
 655                  * marked dying, kill it here or it'll go unnoticed.
 656                  */
 657                 if (unlikely(blk_queue_dying(rq->q))) {
 658                         rq->errors = -EIO;
 659                         blk_mq_complete_request(rq);
 660                 }
 661                 return;
 662         }
 663         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 664                 return;
 665
 666         if (time_after_eq(jiffies, rq->deadline)) {
 667                 if (!blk_mark_rq_complete(rq))
 668                         blk_mq_rq_timed_out(rq, reserved);
 669         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 670                 data->next = rq->deadline;
 671                 data->next_set = 1;
 672         }
 673 }
 674
 675 static void blk_mq_rq_timer(unsigned long priv)
 676 {
 677         struct request_queue *q = (struct request_queue *)priv;
 678         struct blk_mq_timeout_data data = {
 679                 .next           = 0,
 680                 .next_set       = 0,
 681         };
 682         struct blk_mq_hw_ctx *hctx;
 683         int i;
 684
 685         queue_for_each_hw_ctx(q, hctx, i) {
 686                 /*
 687                  * If not software queues are currently mapped to this
 688                  * hardware queue, there's nothing to check
 689                  */
 690                 if (!blk_mq_hw_queue_mapped(hctx))
 691                         continue;
 692
 693                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 694         }
 695
 696         if (data.next_set) {
 697                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 698                 mod_timer(&q->timeout, data.next);
 699         } else {
 700                 queue_for_each_hw_ctx(q, hctx, i) {
 701                         /* the hctx may be unmapped, so check it here */
 702                         if (blk_mq_hw_queue_mapped(hctx))
 703                                 blk_mq_tag_idle(hctx);
 704                 }
 705         }
 706 }
 707
 708 /*
 709  * Reverse check our software queue for entries that we could potentially
 710  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 711  * too much time checking for merges.
 712  */
 713 static bool blk_mq_attempt_merge(struct request_queue *q,
 714                                  struct blk_mq_ctx *ctx, struct bio *bio)
 715 {
 716         struct request *rq;
 717         int checked = 8;
 718
 719         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 720                 int el_ret;
 721
 722                 if (!checked--)
 723                         break;
 724
 725                 if (!blk_rq_merge_ok(rq, bio))
 726                         continue;
 727
 728                 el_ret = blk_try_merge(rq, bio);
 729                 if (el_ret == ELEVATOR_BACK_MERGE) {
 730                         if (bio_attempt_back_merge(q, rq, bio)) {
 731                                 ctx->rq_merged++;
 732                                 return true;
 733                         }
 734                         break;
 735                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 736                         if (bio_attempt_front_merge(q, rq, bio)) {
 737                                 ctx->rq_merged++;
 738                                 return true;
 739                         }
 740                         break;
 741                 }
 742         }
 743
 744         return false;
 745 }
 746
 747 /*
 748  * Process software queues that have been marked busy, splicing them
 749  * to the for-dispatch
 750  */
 751 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 752 {
 753         struct blk_mq_ctx *ctx;
 754         int i;
 755
 756         for (i = 0; i < hctx->ctx_map.size; i++) {
 757                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 758                 unsigned int off, bit;
 759
 760                 if (!bm->word)
 761                         continue;
 762
 763                 bit = 0;
 764                 off = i * hctx->ctx_map.bits_per_word;
 765                 do {
 766                         bit = find_next_bit(&bm->word, bm->depth, bit);
 767                         if (bit >= bm->depth)
 768                                 break;
 769
 770                         ctx = hctx->ctxs[bit + off];
 771                         clear_bit(bit, &bm->word);
 772                         spin_lock(&ctx->lock);
 773                         list_splice_tail_init(&ctx->rq_list, list);
 774                         spin_unlock(&ctx->lock);
 775
 776                         bit++;
 777                 } while (1);
 778         }
 779 }
 780
 781 /*
 782  * Run this hardware queue, pulling any software queues mapped to it in.
 783  * Note that this function currently has various problems around ordering
 784  * of IO. In particular, we'd like FIFO behaviour on handling existing
 785  * items on the hctx->dispatch list. Ignore that for now.
 786  */
 787 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 788 {
 789         struct request_queue *q = hctx->queue;
 790         struct request *rq;
 791         LIST_HEAD(rq_list);
 792         LIST_HEAD(driver_list);
 793         struct list_head *dptr;
 794         int queued;
 795
 796         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 797
 798         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 799                 return;
 800
 801         hctx->run++;
 802
 803         /*
 804          * Touch any software queue that has pending entries.
 805          */
 806         flush_busy_ctxs(hctx, &rq_list);
 807
 808         /*
 809          * If we have previous entries on our dispatch list, grab them
 810          * and stuff them at the front for more fair dispatch.
 811          */
 812         if (!list_empty_careful(&hctx->dispatch)) {
 813                 spin_lock(&hctx->lock);
 814                 if (!list_empty(&hctx->dispatch))
 815                         list_splice_init(&hctx->dispatch, &rq_list);
 816                 spin_unlock(&hctx->lock);
 817         }
 818
 819         /*
 820          * Start off with dptr being NULL, so we start the first request
 821          * immediately, even if we have more pending.
 822          */
 823         dptr = NULL;
 824
 825         /*
 826          * Now process all the entries, sending them to the driver.
 827          */
 828         queued = 0;
 829         while (!list_empty(&rq_list)) {
 830                 struct blk_mq_queue_data bd;
 831                 int ret;
 832
 833                 rq = list_first_entry(&rq_list, struct request, queuelist);
 834                 list_del_init(&rq->queuelist);
 835
 836                 bd.rq = rq;
 837                 bd.list = dptr;
 838                 bd.last = list_empty(&rq_list);
 839
 840                 ret = q->mq_ops->queue_rq(hctx, &bd);
 841                 switch (ret) {
 842                 case BLK_MQ_RQ_QUEUE_OK:
 843                         queued++;
 844                         continue;
 845                 case BLK_MQ_RQ_QUEUE_BUSY:
 846                         list_add(&rq->queuelist, &rq_list);
 847                         __blk_mq_requeue_request(rq);
 848                         break;
 849                 default:
 850                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 851                 case BLK_MQ_RQ_QUEUE_ERROR:
 852                         rq->errors = -EIO;
 853                         blk_mq_end_request(rq, rq->errors);
 854                         break;
 855                 }
 856
 857                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 858                         break;
 859
 860                 /*
 861                  * We've done the first request. If we have more than 1
 862                  * left in the list, set dptr to defer issue.
 863                  */
 864                 if (!dptr && rq_list.next != rq_list.prev)
 865                         dptr = &driver_list;
 866         }
 867
 868         if (!queued)
 869                 hctx->dispatched[0]++;
 870         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 871                 hctx->dispatched[ilog2(queued) + 1]++;
 872
 873         /*
 874          * Any items that need requeuing? Stuff them into hctx->dispatch,
 875          * that is where we will continue on next queue run.
 876          */
 877         if (!list_empty(&rq_list)) {
 878                 spin_lock(&hctx->lock);
 879                 list_splice(&rq_list, &hctx->dispatch);
 880                 spin_unlock(&hctx->lock);
 881                 /*
 882                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 883                  * it's possible the queue is stopped and restarted again
 884                  * before this. Queue restart will dispatch requests. And since
 885                  * requests in rq_list aren't added into hctx->dispatch yet,
 886                  * the requests in rq_list might get lost.
 887                  *
 888                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 889                  **/
 890                 blk_mq_run_hw_queue(hctx, true);
 891         }
 892 }
 893
 894 /*
 895  * It'd be great if the workqueue API had a way to pass
 896  * in a mask and had some smarts for more clever placement.
 897  * For now we just round-robin here, switching for every
 898  * BLK_MQ_CPU_WORK_BATCH queued items.
 899  */
 900 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 901 {
 902         if (hctx->queue->nr_hw_queues == 1)
 903                 return WORK_CPU_UNBOUND;
 904
 905         if (--hctx->next_cpu_batch <= 0) {
 906                 int cpu = hctx->next_cpu, next_cpu;
 907
 908                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 909                 if (next_cpu >= nr_cpu_ids)
 910                         next_cpu = cpumask_first(hctx->cpumask);
 911
 912                 hctx->next_cpu = next_cpu;
 913                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 914
 915                 return cpu;
 916         }
 917
 918         return hctx->next_cpu;
 919 }
 920
 921 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 922 {
 923         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 924             !blk_mq_hw_queue_mapped(hctx)))
 925                 return;
 926
 927         if (!async) {
 928                 int cpu = get_cpu_light();
 929                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 930                         __blk_mq_run_hw_queue(hctx);
 931                         put_cpu_light();
 932                         return;
 933                 }
 934
 935                 put_cpu_light();
 936         }
 937
 938         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 939                         &hctx->run_work, 0);
 940 }
 941
 942 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 943 {
 944         struct blk_mq_hw_ctx *hctx;
 945         int i;
 946
 947         queue_for_each_hw_ctx(q, hctx, i) {
 948                 if ((!blk_mq_hctx_has_pending(hctx) &&
 949                     list_empty_careful(&hctx->dispatch)) ||
 950                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 951                         continue;
 952
 953                 blk_mq_run_hw_queue(hctx, async);
 954         }
 955 }
 956 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 957
 958 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 959 {
 960         cancel_delayed_work(&hctx->run_work);
 961         cancel_delayed_work(&hctx->delay_work);
 962         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 963 }
 964 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 965
 966 void blk_mq_stop_hw_queues(struct request_queue *q)
 967 {
 968         struct blk_mq_hw_ctx *hctx;
 969         int i;
 970
 971         queue_for_each_hw_ctx(q, hctx, i)
 972                 blk_mq_stop_hw_queue(hctx);
 973 }
 974 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 975
 976 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 977 {
 978         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 979
 980         blk_mq_run_hw_queue(hctx, false);
 981 }
 982 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 983
 984 void blk_mq_start_hw_queues(struct request_queue *q)
 985 {
 986         struct blk_mq_hw_ctx *hctx;
 987         int i;
 988
 989         queue_for_each_hw_ctx(q, hctx, i)
 990                 blk_mq_start_hw_queue(hctx);
 991 }
 992 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 993
 994 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 995 {
 996         struct blk_mq_hw_ctx *hctx;
 997         int i;
 998
 999         queue_for_each_hw_ctx(q, hctx, i) {
1000                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1001                         continue;
1002
1003                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1004                 blk_mq_run_hw_queue(hctx, async);
1005         }
1006 }
1007 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1008
1009 static void blk_mq_run_work_fn(struct work_struct *work)
1010 {
1011         struct blk_mq_hw_ctx *hctx;
1012
1013         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1014
1015         __blk_mq_run_hw_queue(hctx);
1016 }
1017
1018 static void blk_mq_delay_work_fn(struct work_struct *work)
1019 {
1020         struct blk_mq_hw_ctx *hctx;
1021
1022         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
1023
1024         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
1025                 __blk_mq_run_hw_queue(hctx);
1026 }
1027
1028 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1029 {
1030         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
1031                 return;
1032
1033         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1034                         &hctx->delay_work, msecs_to_jiffies(msecs));
1035 }
1036 EXPORT_SYMBOL(blk_mq_delay_queue);
1037
1038 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1039                                     struct request *rq, bool at_head)
1040 {
1041         struct blk_mq_ctx *ctx = rq->mq_ctx;
1042
1043         trace_block_rq_insert(hctx->queue, rq);
1044
1045         if (at_head)
1046                 list_add(&rq->queuelist, &ctx->rq_list);
1047         else
1048                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1049
1050         blk_mq_hctx_mark_pending(hctx, ctx);
1051 }
1052
1053 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1054                 bool async)
1055 {
1056         struct request_queue *q = rq->q;
1057         struct blk_mq_hw_ctx *hctx;
1058         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1059
1060         current_ctx = blk_mq_get_ctx(q);
1061         if (!cpu_online(ctx->cpu))
1062                 rq->mq_ctx = ctx = current_ctx;
1063
1064         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1065
1066         spin_lock(&ctx->lock);
1067         __blk_mq_insert_request(hctx, rq, at_head);
1068         spin_unlock(&ctx->lock);
1069
1070         if (run_queue)
1071                 blk_mq_run_hw_queue(hctx, async);
1072
1073         blk_mq_put_ctx(current_ctx);
1074 }
1075
1076 static void blk_mq_insert_requests(struct request_queue *q,
1077                                      struct blk_mq_ctx *ctx,
1078                                      struct list_head *list,
1079                                      int depth,
1080                                      bool from_schedule)
1081
1082 {
1083         struct blk_mq_hw_ctx *hctx;
1084         struct blk_mq_ctx *current_ctx;
1085
1086         trace_block_unplug(q, depth, !from_schedule);
1087
1088         current_ctx = blk_mq_get_ctx(q);
1089
1090         if (!cpu_online(ctx->cpu))
1091                 ctx = current_ctx;
1092         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1093
1094         /*
1095          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1096          * offline now
1097          */
1098         spin_lock(&ctx->lock);
1099         while (!list_empty(list)) {
1100                 struct request *rq;
1101
1102                 rq = list_first_entry(list, struct request, queuelist);
1103                 list_del_init(&rq->queuelist);
1104                 rq->mq_ctx = ctx;
1105                 __blk_mq_insert_request(hctx, rq, false);
1106         }
1107         spin_unlock(&ctx->lock);
1108
1109         blk_mq_run_hw_queue(hctx, from_schedule);
1110         blk_mq_put_ctx(current_ctx);
1111 }
1112
1113 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1114 {
1115         struct request *rqa = container_of(a, struct request, queuelist);
1116         struct request *rqb = container_of(b, struct request, queuelist);
1117
1118         return !(rqa->mq_ctx < rqb->mq_ctx ||
1119                  (rqa->mq_ctx == rqb->mq_ctx &&
1120                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1121 }
1122
1123 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1124 {
1125         struct blk_mq_ctx *this_ctx;
1126         struct request_queue *this_q;
1127         struct request *rq;
1128         LIST_HEAD(list);
1129         LIST_HEAD(ctx_list);
1130         unsigned int depth;
1131
1132         list_splice_init(&plug->mq_list, &list);
1133
1134         list_sort(NULL, &list, plug_ctx_cmp);
1135
1136         this_q = NULL;
1137         this_ctx = NULL;
1138         depth = 0;
1139
1140         while (!list_empty(&list)) {
1141                 rq = list_entry_rq(list.next);
1142                 list_del_init(&rq->queuelist);
1143                 BUG_ON(!rq->q);
1144                 if (rq->mq_ctx != this_ctx) {
1145                         if (this_ctx) {
1146                                 blk_mq_insert_requests(this_q, this_ctx,
1147                                                         &ctx_list, depth,
1148                                                         from_schedule);
1149                         }
1150
1151                         this_ctx = rq->mq_ctx;
1152                         this_q = rq->q;
1153                         depth = 0;
1154                 }
1155
1156                 depth++;
1157                 list_add_tail(&rq->queuelist, &ctx_list);
1158         }
1159
1160         /*
1161          * If 'this_ctx' is set, we know we have entries to complete
1162          * on 'ctx_list'. Do those.
1163          */
1164         if (this_ctx) {
1165                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1166                                        from_schedule);
1167         }
1168 }
1169
1170 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1171 {
1172         init_request_from_bio(rq, bio);
1173
1174         if (blk_do_io_stat(rq))
1175                 blk_account_io_start(rq, 1);
1176 }
1177
1178 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1179 {
1180         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1181                 !blk_queue_nomerges(hctx->queue);
1182 }
1183
1184 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1185                                          struct blk_mq_ctx *ctx,
1186                                          struct request *rq, struct bio *bio)
1187 {
1188         if (!hctx_allow_merges(hctx)) {
1189                 blk_mq_bio_to_request(rq, bio);
1190                 spin_lock(&ctx->lock);
1191 insert_rq:
1192                 __blk_mq_insert_request(hctx, rq, false);
1193                 spin_unlock(&ctx->lock);
1194                 return false;
1195         } else {
1196                 struct request_queue *q = hctx->queue;
1197
1198                 spin_lock(&ctx->lock);
1199                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1200                         blk_mq_bio_to_request(rq, bio);
1201                         goto insert_rq;
1202                 }
1203
1204                 spin_unlock(&ctx->lock);
1205                 __blk_mq_free_request(hctx, ctx, rq);
1206                 return true;
1207         }
1208 }
1209
1210 struct blk_map_ctx {
1211         struct blk_mq_hw_ctx *hctx;
1212         struct blk_mq_ctx *ctx;
1213 };
1214
1215 static struct request *blk_mq_map_request(struct request_queue *q,
1216                                           struct bio *bio,
1217                                           struct blk_map_ctx *data)
1218 {
1219         struct blk_mq_hw_ctx *hctx;
1220         struct blk_mq_ctx *ctx;
1221         struct request *rq;
1222         int rw = bio_data_dir(bio);
1223         struct blk_mq_alloc_data alloc_data;
1224
1225         if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1226                 bio_endio(bio, -EIO);
1227                 return NULL;
1228         }
1229
1230         ctx = blk_mq_get_ctx(q);
1231         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1232
1233         if (rw_is_sync(bio->bi_rw))
1234                 rw |= REQ_SYNC;
1235
1236         trace_block_getrq(q, bio, rw);
1237         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1238                         hctx);
1239         rq = __blk_mq_alloc_request(&alloc_data, rw);
1240         if (unlikely(!rq)) {
1241                 __blk_mq_run_hw_queue(hctx);
1242                 blk_mq_put_ctx(ctx);
1243                 trace_block_sleeprq(q, bio, rw);
1244
1245                 ctx = blk_mq_get_ctx(q);
1246                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1247                 blk_mq_set_alloc_data(&alloc_data, q,
1248                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1249                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1250                 ctx = alloc_data.ctx;
1251                 hctx = alloc_data.hctx;
1252         }
1253
1254         hctx->queued++;
1255         data->hctx = hctx;
1256         data->ctx = ctx;
1257         return rq;
1258 }
1259
1260 /*
1261  * Multiple hardware queue variant. This will not use per-process plugs,
1262  * but will attempt to bypass the hctx queueing if we can go straight to
1263  * hardware for SYNC IO.
1264  */
1265 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1266 {
1267         const int is_sync = rw_is_sync(bio->bi_rw);
1268         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1269         struct blk_map_ctx data;
1270         struct request *rq;
1271
1272         blk_queue_bounce(q, &bio);
1273
1274         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1275                 bio_endio(bio, -EIO);
1276                 return;
1277         }
1278
1279         rq = blk_mq_map_request(q, bio, &data);
1280         if (unlikely(!rq))
1281                 return;
1282
1283         if (unlikely(is_flush_fua)) {
1284                 blk_mq_bio_to_request(rq, bio);
1285                 blk_insert_flush(rq);
1286                 goto run_queue;
1287         }
1288
1289         /*
1290          * If the driver supports defer issued based on 'last', then
1291          * queue it up like normal since we can potentially save some
1292          * CPU this way.
1293          */
1294         if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1295                 struct blk_mq_queue_data bd = {
1296                         .rq = rq,
1297                         .list = NULL,
1298                         .last = 1
1299                 };
1300                 int ret;
1301
1302                 blk_mq_bio_to_request(rq, bio);
1303
1304                 /*
1305                  * For OK queue, we are done. For error, kill it. Any other
1306                  * error (busy), just add it to our list as we previously
1307                  * would have done
1308                  */
1309                 ret = q->mq_ops->queue_rq(data.hctx, &bd);
1310                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1311                         goto done;
1312                 else {
1313                         __blk_mq_requeue_request(rq);
1314
1315                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1316                                 rq->errors = -EIO;
1317                                 blk_mq_end_request(rq, rq->errors);
1318                                 goto done;
1319                         }
1320                 }
1321         }
1322
1323         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1324                 /*
1325                  * For a SYNC request, send it to the hardware immediately. For
1326                  * an ASYNC request, just ensure that we run it later on. The
1327                  * latter allows for merging opportunities and more efficient
1328                  * dispatching.
1329                  */
1330 run_queue:
1331                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1332         }
1333 done:
1334         blk_mq_put_ctx(data.ctx);
1335 }
1336
1337 /*
1338  * Single hardware queue variant. This will attempt to use any per-process
1339  * plug for merging and IO deferral.
1340  */
1341 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1342 {
1343         const int is_sync = rw_is_sync(bio->bi_rw);
1344         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1345         unsigned int use_plug, request_count = 0;
1346         struct blk_map_ctx data;
1347         struct request *rq;
1348
1349         /*
1350          * If we have multiple hardware queues, just go directly to
1351          * one of those for sync IO.
1352          */
1353         use_plug = !is_flush_fua && !is_sync;
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 (use_plug && !blk_queue_nomerges(q) &&
1363             blk_attempt_plug_merge(q, bio, &request_count))
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         if (use_plug) {
1382                 struct blk_plug *plug = current->plug;
1383
1384                 if (plug) {
1385                         blk_mq_bio_to_request(rq, bio);
1386                         if (list_empty(&plug->mq_list))
1387                                 trace_block_plug(q);
1388                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1389                                 blk_flush_plug_list(plug, false);
1390                                 trace_block_plug(q);
1391                         }
1392                         list_add_tail(&rq->queuelist, &plug->mq_list);
1393                         blk_mq_put_ctx(data.ctx);
1394                         return;
1395                 }
1396         }
1397
1398         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1399                 /*
1400                  * For a SYNC request, send it to the hardware immediately. For
1401                  * an ASYNC request, just ensure that we run it later on. The
1402                  * latter allows for merging opportunities and more efficient
1403                  * dispatching.
1404                  */
1405 run_queue:
1406                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1407         }
1408
1409         blk_mq_put_ctx(data.ctx);
1410 }
1411
1412 /*
1413  * Default mapping to a software queue, since we use one per CPU.
1414  */
1415 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1416 {
1417         return q->queue_hw_ctx[q->mq_map[cpu]];
1418 }
1419 EXPORT_SYMBOL(blk_mq_map_queue);
1420
1421 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1422                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1423 {
1424         struct page *page;
1425
1426         if (tags->rqs && set->ops->exit_request) {
1427                 int i;
1428
1429                 for (i = 0; i < tags->nr_tags; i++) {
1430                         if (!tags->rqs[i])
1431                                 continue;
1432                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1433                                                 hctx_idx, i);
1434                         tags->rqs[i] = NULL;
1435                 }
1436         }
1437
1438         while (!list_empty(&tags->page_list)) {
1439                 page = list_first_entry(&tags->page_list, struct page, lru);
1440                 list_del_init(&page->lru);
1441                 __free_pages(page, page->private);
1442         }
1443
1444         kfree(tags->rqs);
1445
1446         blk_mq_free_tags(tags);
1447 }
1448
1449 static size_t order_to_size(unsigned int order)
1450 {
1451         return (size_t)PAGE_SIZE << order;
1452 }
1453
1454 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1455                 unsigned int hctx_idx)
1456 {
1457         struct blk_mq_tags *tags;
1458         unsigned int i, j, entries_per_page, max_order = 4;
1459         size_t rq_size, left;
1460
1461         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1462                                 set->numa_node,
1463                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1464         if (!tags)
1465                 return NULL;
1466
1467         INIT_LIST_HEAD(&tags->page_list);
1468
1469         tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1470                                  GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1471                                  set->numa_node);
1472         if (!tags->rqs) {
1473                 blk_mq_free_tags(tags);
1474                 return NULL;
1475         }
1476
1477         /*
1478          * rq_size is the size of the request plus driver payload, rounded
1479          * to the cacheline size
1480          */
1481         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1482                                 cache_line_size());
1483         left = rq_size * set->queue_depth;
1484
1485         for (i = 0; i < set->queue_depth; ) {
1486                 int this_order = max_order;
1487                 struct page *page;
1488                 int to_do;
1489                 void *p;
1490
1491                 while (left < order_to_size(this_order - 1) && this_order)
1492                         this_order--;
1493
1494                 do {
1495                         page = alloc_pages_node(set->numa_node,
1496                                 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1497                                 this_order);
1498                         if (page)
1499                                 break;
1500                         if (!this_order--)
1501                                 break;
1502                         if (order_to_size(this_order) < rq_size)
1503                                 break;
1504                 } while (1);
1505
1506                 if (!page)
1507                         goto fail;
1508
1509                 page->private = this_order;
1510                 list_add_tail(&page->lru, &tags->page_list);
1511
1512                 p = page_address(page);
1513                 entries_per_page = order_to_size(this_order) / rq_size;
1514                 to_do = min(entries_per_page, set->queue_depth - i);
1515                 left -= to_do * rq_size;
1516                 for (j = 0; j < to_do; j++) {
1517                         tags->rqs[i] = p;
1518                         if (set->ops->init_request) {
1519                                 if (set->ops->init_request(set->driver_data,
1520                                                 tags->rqs[i], hctx_idx, i,
1521                                                 set->numa_node)) {
1522                                         tags->rqs[i] = NULL;
1523                                         goto fail;
1524                                 }
1525                         }
1526
1527                         p += rq_size;
1528                         i++;
1529                 }
1530         }
1531
1532         return tags;
1533
1534 fail:
1535         blk_mq_free_rq_map(set, tags, hctx_idx);
1536         return NULL;
1537 }
1538
1539 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1540 {
1541         kfree(bitmap->map);
1542 }
1543
1544 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1545 {
1546         unsigned int bpw = 8, total, num_maps, i;
1547
1548         bitmap->bits_per_word = bpw;
1549
1550         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1551         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1552                                         GFP_KERNEL, node);
1553         if (!bitmap->map)
1554                 return -ENOMEM;
1555
1556         total = nr_cpu_ids;
1557         for (i = 0; i < num_maps; i++) {
1558                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1559                 total -= bitmap->map[i].depth;
1560         }
1561
1562         return 0;
1563 }
1564
1565 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1566 {
1567         struct request_queue *q = hctx->queue;
1568         struct blk_mq_ctx *ctx;
1569         LIST_HEAD(tmp);
1570
1571         /*
1572          * Move ctx entries to new CPU, if this one is going away.
1573          */
1574         ctx = __blk_mq_get_ctx(q, cpu);
1575
1576         spin_lock(&ctx->lock);
1577         if (!list_empty(&ctx->rq_list)) {
1578                 list_splice_init(&ctx->rq_list, &tmp);
1579                 blk_mq_hctx_clear_pending(hctx, ctx);
1580         }
1581         spin_unlock(&ctx->lock);
1582
1583         if (list_empty(&tmp))
1584                 return NOTIFY_OK;
1585
1586         ctx = blk_mq_get_ctx(q);
1587         spin_lock(&ctx->lock);
1588
1589         while (!list_empty(&tmp)) {
1590                 struct request *rq;
1591
1592                 rq = list_first_entry(&tmp, struct request, queuelist);
1593                 rq->mq_ctx = ctx;
1594                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1595         }
1596
1597         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1598         blk_mq_hctx_mark_pending(hctx, ctx);
1599
1600         spin_unlock(&ctx->lock);
1601
1602         blk_mq_run_hw_queue(hctx, true);
1603         blk_mq_put_ctx(ctx);
1604         return NOTIFY_OK;
1605 }
1606
1607 static int blk_mq_hctx_notify(void *data, unsigned long action,
1608                               unsigned int cpu)
1609 {
1610         struct blk_mq_hw_ctx *hctx = data;
1611
1612         if (action == CPU_POST_DEAD)
1613                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1614
1615         /*
1616          * In case of CPU online, tags may be reallocated
1617          * in blk_mq_map_swqueue() after mapping is updated.
1618          */
1619
1620         return NOTIFY_OK;
1621 }
1622
1623 /* hctx->ctxs will be freed in queue's release handler */
1624 static void blk_mq_exit_hctx(struct request_queue *q,
1625                 struct blk_mq_tag_set *set,
1626                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1627 {
1628         unsigned flush_start_tag = set->queue_depth;
1629
1630         blk_mq_tag_idle(hctx);
1631
1632         if (set->ops->exit_request)
1633                 set->ops->exit_request(set->driver_data,
1634                                        hctx->fq->flush_rq, hctx_idx,
1635                                        flush_start_tag + hctx_idx);
1636
1637         if (set->ops->exit_hctx)
1638                 set->ops->exit_hctx(hctx, hctx_idx);
1639
1640         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1641         blk_free_flush_queue(hctx->fq);
1642         blk_mq_free_bitmap(&hctx->ctx_map);
1643 }
1644
1645 static void blk_mq_exit_hw_queues(struct request_queue *q,
1646                 struct blk_mq_tag_set *set, int nr_queue)
1647 {
1648         struct blk_mq_hw_ctx *hctx;
1649         unsigned int i;
1650
1651         queue_for_each_hw_ctx(q, hctx, i) {
1652                 if (i == nr_queue)
1653                         break;
1654                 blk_mq_exit_hctx(q, set, hctx, i);
1655         }
1656 }
1657
1658 static void blk_mq_free_hw_queues(struct request_queue *q,
1659                 struct blk_mq_tag_set *set)
1660 {
1661         struct blk_mq_hw_ctx *hctx;
1662         unsigned int i;
1663
1664         queue_for_each_hw_ctx(q, hctx, i)
1665                 free_cpumask_var(hctx->cpumask);
1666 }
1667
1668 static int blk_mq_init_hctx(struct request_queue *q,
1669                 struct blk_mq_tag_set *set,
1670                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1671 {
1672         int node;
1673         unsigned flush_start_tag = set->queue_depth;
1674
1675         node = hctx->numa_node;
1676         if (node == NUMA_NO_NODE)
1677                 node = hctx->numa_node = set->numa_node;
1678
1679         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1680         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1681         spin_lock_init(&hctx->lock);
1682         INIT_LIST_HEAD(&hctx->dispatch);
1683         hctx->queue = q;
1684         hctx->queue_num = hctx_idx;
1685         hctx->flags = set->flags;
1686
1687         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1688                                         blk_mq_hctx_notify, hctx);
1689         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1690
1691         hctx->tags = set->tags[hctx_idx];
1692
1693         /*
1694          * Allocate space for all possible cpus to avoid allocation at
1695          * runtime
1696          */
1697         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1698                                         GFP_KERNEL, node);
1699         if (!hctx->ctxs)
1700                 goto unregister_cpu_notifier;
1701
1702         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1703                 goto free_ctxs;
1704
1705         hctx->nr_ctx = 0;
1706
1707         if (set->ops->init_hctx &&
1708             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1709                 goto free_bitmap;
1710
1711         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1712         if (!hctx->fq)
1713                 goto exit_hctx;
1714
1715         if (set->ops->init_request &&
1716             set->ops->init_request(set->driver_data,
1717                                    hctx->fq->flush_rq, hctx_idx,
1718                                    flush_start_tag + hctx_idx, node))
1719                 goto free_fq;
1720
1721         return 0;
1722
1723  free_fq:
1724         kfree(hctx->fq);
1725  exit_hctx:
1726         if (set->ops->exit_hctx)
1727                 set->ops->exit_hctx(hctx, hctx_idx);
1728  free_bitmap:
1729         blk_mq_free_bitmap(&hctx->ctx_map);
1730  free_ctxs:
1731         kfree(hctx->ctxs);
1732  unregister_cpu_notifier:
1733         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1734
1735         return -1;
1736 }
1737
1738 static int blk_mq_init_hw_queues(struct request_queue *q,
1739                 struct blk_mq_tag_set *set)
1740 {
1741         struct blk_mq_hw_ctx *hctx;
1742         unsigned int i;
1743
1744         /*
1745          * Initialize hardware queues
1746          */
1747         queue_for_each_hw_ctx(q, hctx, i) {
1748                 if (blk_mq_init_hctx(q, set, hctx, i))
1749                         break;
1750         }
1751
1752         if (i == q->nr_hw_queues)
1753                 return 0;
1754
1755         /*
1756          * Init failed
1757          */
1758         blk_mq_exit_hw_queues(q, set, i);
1759
1760         return 1;
1761 }
1762
1763 static void blk_mq_init_cpu_queues(struct request_queue *q,
1764                                    unsigned int nr_hw_queues)
1765 {
1766         unsigned int i;
1767
1768         for_each_possible_cpu(i) {
1769                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1770                 struct blk_mq_hw_ctx *hctx;
1771
1772                 memset(__ctx, 0, sizeof(*__ctx));
1773                 __ctx->cpu = i;
1774                 spin_lock_init(&__ctx->lock);
1775                 INIT_LIST_HEAD(&__ctx->rq_list);
1776                 __ctx->queue = q;
1777
1778                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1779                 if (!cpu_online(i))
1780                         continue;
1781
1782                 hctx = q->mq_ops->map_queue(q, i);
1783
1784                 /*
1785                  * Set local node, IFF we have more than one hw queue. If
1786                  * not, we remain on the home node of the device
1787                  */
1788                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1789                         hctx->numa_node = cpu_to_node(i);
1790         }
1791 }
1792
1793 static void blk_mq_map_swqueue(struct request_queue *q)
1794 {
1795         unsigned int i;
1796         struct blk_mq_hw_ctx *hctx;
1797         struct blk_mq_ctx *ctx;
1798         struct blk_mq_tag_set *set = q->tag_set;
1799
1800         queue_for_each_hw_ctx(q, hctx, i) {
1801                 cpumask_clear(hctx->cpumask);
1802                 hctx->nr_ctx = 0;
1803         }
1804
1805         /*
1806          * Map software to hardware queues
1807          */
1808         queue_for_each_ctx(q, ctx, i) {
1809                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1810                 if (!cpu_online(i))
1811                         continue;
1812
1813                 hctx = q->mq_ops->map_queue(q, i);
1814                 cpumask_set_cpu(i, hctx->cpumask);
1815                 ctx->index_hw = hctx->nr_ctx;
1816                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1817         }
1818
1819         queue_for_each_hw_ctx(q, hctx, i) {
1820                 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1821
1822                 /*
1823                  * If no software queues are mapped to this hardware queue,
1824                  * disable it and free the request entries.
1825                  */
1826                 if (!hctx->nr_ctx) {
1827                         if (set->tags[i]) {
1828                                 blk_mq_free_rq_map(set, set->tags[i], i);
1829                                 set->tags[i] = NULL;
1830                         }
1831                         hctx->tags = NULL;
1832                         continue;
1833                 }
1834
1835                 /* unmapped hw queue can be remapped after CPU topo changed */
1836                 if (!set->tags[i])
1837                         set->tags[i] = blk_mq_init_rq_map(set, i);
1838                 hctx->tags = set->tags[i];
1839                 WARN_ON(!hctx->tags);
1840
1841                 /*
1842                  * Set the map size to the number of mapped software queues.
1843                  * This is more accurate and more efficient than looping
1844                  * over all possibly mapped software queues.
1845                  */
1846                 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1847
1848                 /*
1849                  * Initialize batch roundrobin counts
1850                  */
1851                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1852                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1853         }
1854 }
1855
1856 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1857 {
1858         struct blk_mq_hw_ctx *hctx;
1859         struct request_queue *q;
1860         bool shared;
1861         int i;
1862
1863         if (set->tag_list.next == set->tag_list.prev)
1864                 shared = false;
1865         else
1866                 shared = true;
1867
1868         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1869                 blk_mq_freeze_queue(q);
1870
1871                 queue_for_each_hw_ctx(q, hctx, i) {
1872                         if (shared)
1873                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1874                         else
1875                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1876                 }
1877                 blk_mq_unfreeze_queue(q);
1878         }
1879 }
1880
1881 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1882 {
1883         struct blk_mq_tag_set *set = q->tag_set;
1884
1885         mutex_lock(&set->tag_list_lock);
1886         list_del_init(&q->tag_set_list);
1887         blk_mq_update_tag_set_depth(set);
1888         mutex_unlock(&set->tag_list_lock);
1889 }
1890
1891 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1892                                      struct request_queue *q)
1893 {
1894         q->tag_set = set;
1895
1896         mutex_lock(&set->tag_list_lock);
1897         list_add_tail(&q->tag_set_list, &set->tag_list);
1898         blk_mq_update_tag_set_depth(set);
1899         mutex_unlock(&set->tag_list_lock);
1900 }
1901
1902 /*
1903  * It is the actual release handler for mq, but we do it from
1904  * request queue's release handler for avoiding use-after-free
1905  * and headache because q->mq_kobj shouldn't have been introduced,
1906  * but we can't group ctx/kctx kobj without it.
1907  */
1908 void blk_mq_release(struct request_queue *q)
1909 {
1910         struct blk_mq_hw_ctx *hctx;
1911         unsigned int i;
1912
1913         /* hctx kobj stays in hctx */
1914         queue_for_each_hw_ctx(q, hctx, i) {
1915                 if (!hctx)
1916                         continue;
1917                 kfree(hctx->ctxs);
1918                 kfree(hctx);
1919         }
1920
1921         kfree(q->queue_hw_ctx);
1922
1923         /* ctx kobj stays in queue_ctx */
1924         free_percpu(q->queue_ctx);
1925 }
1926
1927 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1928 {
1929         struct request_queue *uninit_q, *q;
1930
1931         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1932         if (!uninit_q)
1933                 return ERR_PTR(-ENOMEM);
1934
1935         q = blk_mq_init_allocated_queue(set, uninit_q);
1936         if (IS_ERR(q))
1937                 blk_cleanup_queue(uninit_q);
1938
1939         return q;
1940 }
1941 EXPORT_SYMBOL(blk_mq_init_queue);
1942
1943 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1944                                                   struct request_queue *q)
1945 {
1946         struct blk_mq_hw_ctx **hctxs;
1947         struct blk_mq_ctx __percpu *ctx;
1948         unsigned int *map;
1949         int i;
1950
1951         ctx = alloc_percpu(struct blk_mq_ctx);
1952         if (!ctx)
1953                 return ERR_PTR(-ENOMEM);
1954
1955         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1956                         set->numa_node);
1957
1958         if (!hctxs)
1959                 goto err_percpu;
1960
1961         map = blk_mq_make_queue_map(set);
1962         if (!map)
1963                 goto err_map;
1964
1965         for (i = 0; i < set->nr_hw_queues; i++) {
1966                 int node = blk_mq_hw_queue_to_node(map, i);
1967
1968                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1969                                         GFP_KERNEL, node);
1970                 if (!hctxs[i])
1971                         goto err_hctxs;
1972
1973                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1974                                                 node))
1975                         goto err_hctxs;
1976
1977                 atomic_set(&hctxs[i]->nr_active, 0);
1978                 hctxs[i]->numa_node = node;
1979                 hctxs[i]->queue_num = i;
1980         }
1981
1982         /*
1983          * Init percpu_ref in atomic mode so that it's faster to shutdown.
1984          * See blk_register_queue() for details.
1985          */
1986         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
1987                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
1988                 goto err_hctxs;
1989
1990         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1991         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
1992
1993         q->nr_queues = nr_cpu_ids;
1994         q->nr_hw_queues = set->nr_hw_queues;
1995         q->mq_map = map;
1996
1997         q->queue_ctx = ctx;
1998         q->queue_hw_ctx = hctxs;
1999
2000         q->mq_ops = set->ops;
2001         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2002
2003         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2004                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2005
2006         q->sg_reserved_size = INT_MAX;
2007
2008         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2009         INIT_LIST_HEAD(&q->requeue_list);
2010         spin_lock_init(&q->requeue_lock);
2011
2012         if (q->nr_hw_queues > 1)
2013                 blk_queue_make_request(q, blk_mq_make_request);
2014         else
2015                 blk_queue_make_request(q, blk_sq_make_request);
2016
2017         /*
2018          * Do this after blk_queue_make_request() overrides it...
2019          */
2020         q->nr_requests = set->queue_depth;
2021
2022         if (set->ops->complete)
2023                 blk_queue_softirq_done(q, set->ops->complete);
2024
2025         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2026
2027         if (blk_mq_init_hw_queues(q, set))
2028                 goto err_hctxs;
2029
2030         mutex_lock(&all_q_mutex);
2031         list_add_tail(&q->all_q_node, &all_q_list);
2032         mutex_unlock(&all_q_mutex);
2033
2034         blk_mq_add_queue_tag_set(set, q);
2035
2036         blk_mq_map_swqueue(q);
2037
2038         return q;
2039
2040 err_hctxs:
2041         kfree(map);
2042         for (i = 0; i < set->nr_hw_queues; i++) {
2043                 if (!hctxs[i])
2044                         break;
2045                 free_cpumask_var(hctxs[i]->cpumask);
2046                 kfree(hctxs[i]);
2047         }
2048 err_map:
2049         kfree(hctxs);
2050 err_percpu:
2051         free_percpu(ctx);
2052         return ERR_PTR(-ENOMEM);
2053 }
2054 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2055
2056 void blk_mq_free_queue(struct request_queue *q)
2057 {
2058         struct blk_mq_tag_set   *set = q->tag_set;
2059
2060         blk_mq_del_queue_tag_set(q);
2061
2062         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2063         blk_mq_free_hw_queues(q, set);
2064
2065         percpu_ref_exit(&q->mq_usage_counter);
2066
2067         kfree(q->mq_map);
2068
2069         q->mq_map = NULL;
2070
2071         mutex_lock(&all_q_mutex);
2072         list_del_init(&q->all_q_node);
2073         mutex_unlock(&all_q_mutex);
2074 }
2075
2076 /* Basically redo blk_mq_init_queue with queue frozen */
2077 static void blk_mq_queue_reinit(struct request_queue *q)
2078 {
2079         WARN_ON_ONCE(!q->mq_freeze_depth);
2080
2081         blk_mq_sysfs_unregister(q);
2082
2083         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
2084
2085         /*
2086          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2087          * we should change hctx numa_node according to new topology (this
2088          * involves free and re-allocate memory, worthy doing?)
2089          */
2090
2091         blk_mq_map_swqueue(q);
2092
2093         blk_mq_sysfs_register(q);
2094 }
2095
2096 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2097                                       unsigned long action, void *hcpu)
2098 {
2099         struct request_queue *q;
2100
2101         /*
2102          * Before new mappings are established, hotadded cpu might already
2103          * start handling requests. This doesn't break anything as we map
2104          * offline CPUs to first hardware queue. We will re-init the queue
2105          * below to get optimal settings.
2106          */
2107         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
2108             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
2109                 return NOTIFY_OK;
2110
2111         mutex_lock(&all_q_mutex);
2112
2113         /*
2114          * We need to freeze and reinit all existing queues.  Freezing
2115          * involves synchronous wait for an RCU grace period and doing it
2116          * one by one may take a long time.  Start freezing all queues in
2117          * one swoop and then wait for the completions so that freezing can
2118          * take place in parallel.
2119          */
2120         list_for_each_entry(q, &all_q_list, all_q_node)
2121                 blk_mq_freeze_queue_start(q);
2122         list_for_each_entry(q, &all_q_list, all_q_node) {
2123                 blk_mq_freeze_queue_wait(q);
2124
2125                 /*
2126                  * timeout handler can't touch hw queue during the
2127                  * reinitialization
2128                  */
2129                 del_timer_sync(&q->timeout);
2130         }
2131
2132         list_for_each_entry(q, &all_q_list, all_q_node)
2133                 blk_mq_queue_reinit(q);
2134
2135         list_for_each_entry(q, &all_q_list, all_q_node)
2136                 blk_mq_unfreeze_queue(q);
2137
2138         mutex_unlock(&all_q_mutex);
2139         return NOTIFY_OK;
2140 }
2141
2142 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2143 {
2144         int i;
2145
2146         for (i = 0; i < set->nr_hw_queues; i++) {
2147                 set->tags[i] = blk_mq_init_rq_map(set, i);
2148                 if (!set->tags[i])
2149                         goto out_unwind;
2150         }
2151
2152         return 0;
2153
2154 out_unwind:
2155         while (--i >= 0)
2156                 blk_mq_free_rq_map(set, set->tags[i], i);
2157
2158         return -ENOMEM;
2159 }
2160
2161 /*
2162  * Allocate the request maps associated with this tag_set. Note that this
2163  * may reduce the depth asked for, if memory is tight. set->queue_depth
2164  * will be updated to reflect the allocated depth.
2165  */
2166 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2167 {
2168         unsigned int depth;
2169         int err;
2170
2171         depth = set->queue_depth;
2172         do {
2173                 err = __blk_mq_alloc_rq_maps(set);
2174                 if (!err)
2175                         break;
2176
2177                 set->queue_depth >>= 1;
2178                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2179                         err = -ENOMEM;
2180                         break;
2181                 }
2182         } while (set->queue_depth);
2183
2184         if (!set->queue_depth || err) {
2185                 pr_err("blk-mq: failed to allocate request map\n");
2186                 return -ENOMEM;
2187         }
2188
2189         if (depth != set->queue_depth)
2190                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2191                                                 depth, set->queue_depth);
2192
2193         return 0;
2194 }
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         }
2260
2261         kfree(set->tags);
2262         set->tags = NULL;
2263 }
2264 EXPORT_SYMBOL(blk_mq_free_tag_set);
2265
2266 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2267 {
2268         struct blk_mq_tag_set *set = q->tag_set;
2269         struct blk_mq_hw_ctx *hctx;
2270         int i, ret;
2271
2272         if (!set || nr > set->queue_depth)
2273                 return -EINVAL;
2274
2275         ret = 0;
2276         queue_for_each_hw_ctx(q, hctx, i) {
2277                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2278                 if (ret)
2279                         break;
2280         }
2281
2282         if (!ret)
2283                 q->nr_requests = nr;
2284
2285         return ret;
2286 }
2287
2288 void blk_mq_disable_hotplug(void)
2289 {
2290         mutex_lock(&all_q_mutex);
2291 }
2292
2293 void blk_mq_enable_hotplug(void)
2294 {
2295         mutex_unlock(&all_q_mutex);
2296 }
2297
2298 static int __init blk_mq_init(void)
2299 {
2300         blk_mq_cpu_init();
2301
2302         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2303
2304         return 0;
2305 }
2306 subsys_initcall(blk_mq_init);