block/blk-mq.c

   1 #include <linux/kernel.h>
   2 #include <linux/module.h>
   3 #include <linux/backing-dev.h>
   4 #include <linux/bio.h>
   5 #include <linux/blkdev.h>
   6 #include <linux/mm.h>
   7 #include <linux/init.h>
   8 #include <linux/slab.h>
   9 #include <linux/workqueue.h>
  10 #include <linux/smp.h>
  11 #include <linux/llist.h>
  12 #include <linux/list_sort.h>
  13 #include <linux/cpu.h>
  14 #include <linux/cache.h>
  15 #include <linux/sched/sysctl.h>
  16 #include <linux/delay.h>
  17
  18 #include <trace/events/block.h>
  19
  20 #include <linux/blk-mq.h>
  21 #include "blk.h"
  22 #include "blk-mq.h"
  23 #include "blk-mq-tag.h"
  24
  25 static DEFINE_MUTEX(all_q_mutex);
  26 static LIST_HEAD(all_q_list);
  27
  28 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  29
  30 static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
  31                                            unsigned int cpu)
  32 {
  33         return per_cpu_ptr(q->queue_ctx, cpu);
  34 }
  35
  36 /*
  37  * This assumes per-cpu software queueing queues. They could be per-node
  38  * as well, for instance. For now this is hardcoded as-is. Note that we don't
  39  * care about preemption, since we know the ctx's are persistent. This does
  40  * mean that we can't rely on ctx always matching the currently running CPU.
  41  */
  42 static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
  43 {
  44         return __blk_mq_get_ctx(q, get_cpu());
  45 }
  46
  47 static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
  48 {
  49         put_cpu();
  50 }
  51
  52 /*
  53  * Check if any of the ctx's have pending work in this hardware queue
  54  */
  55 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  56 {
  57         unsigned int i;
  58
  59         for (i = 0; i < hctx->ctx_map.map_size; i++)
  60                 if (hctx->ctx_map.map[i].word)
  61                         return true;
  62
  63         return false;
  64 }
  65
  66 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  67                                               struct blk_mq_ctx *ctx)
  68 {
  69         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  70 }
  71
  72 #define CTX_TO_BIT(hctx, ctx)   \
  73         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  74
  75 /*
  76  * Mark this ctx as having pending work in this hardware queue
  77  */
  78 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  79                                      struct blk_mq_ctx *ctx)
  80 {
  81         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  82
  83         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  84                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  85 }
  86
  87 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  88                                       struct blk_mq_ctx *ctx)
  89 {
  90         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  91
  92         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  93 }
  94
  95 static int blk_mq_queue_enter(struct request_queue *q)
  96 {
  97         int ret;
  98
  99         __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
 100         smp_wmb();
 101         /* we have problems to freeze the queue if it's initializing */
 102         if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
 103                 return 0;
 104
 105         __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
 106
 107         spin_lock_irq(q->queue_lock);
 108         ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
 109                 !blk_queue_bypass(q) || blk_queue_dying(q),
 110                 *q->queue_lock);
 111         /* inc usage with lock hold to avoid freeze_queue runs here */
 112         if (!ret && !blk_queue_dying(q))
 113                 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
 114         else if (blk_queue_dying(q))
 115                 ret = -ENODEV;
 116         spin_unlock_irq(q->queue_lock);
 117
 118         return ret;
 119 }
 120
 121 static void blk_mq_queue_exit(struct request_queue *q)
 122 {
 123         __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
 124 }
 125
 126 static void __blk_mq_drain_queue(struct request_queue *q)
 127 {
 128         while (true) {
 129                 s64 count;
 130
 131                 spin_lock_irq(q->queue_lock);
 132                 count = percpu_counter_sum(&q->mq_usage_counter);
 133                 spin_unlock_irq(q->queue_lock);
 134
 135                 if (count == 0)
 136                         break;
 137                 blk_mq_run_queues(q, false);
 138                 msleep(10);
 139         }
 140 }
 141
 142 /*
 143  * Guarantee no request is in use, so we can change any data structure of
 144  * the queue afterward.
 145  */
 146 static void blk_mq_freeze_queue(struct request_queue *q)
 147 {
 148         bool drain;
 149
 150         spin_lock_irq(q->queue_lock);
 151         drain = !q->bypass_depth++;
 152         queue_flag_set(QUEUE_FLAG_BYPASS, q);
 153         spin_unlock_irq(q->queue_lock);
 154
 155         if (drain)
 156                 __blk_mq_drain_queue(q);
 157 }
 158
 159 void blk_mq_drain_queue(struct request_queue *q)
 160 {
 161         __blk_mq_drain_queue(q);
 162 }
 163
 164 static void blk_mq_unfreeze_queue(struct request_queue *q)
 165 {
 166         bool wake = false;
 167
 168         spin_lock_irq(q->queue_lock);
 169         if (!--q->bypass_depth) {
 170                 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
 171                 wake = true;
 172         }
 173         WARN_ON_ONCE(q->bypass_depth < 0);
 174         spin_unlock_irq(q->queue_lock);
 175         if (wake)
 176                 wake_up_all(&q->mq_freeze_wq);
 177 }
 178
 179 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 180 {
 181         return blk_mq_has_free_tags(hctx->tags);
 182 }
 183 EXPORT_SYMBOL(blk_mq_can_queue);
 184
 185 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 186                                struct request *rq, unsigned int rw_flags)
 187 {
 188         if (blk_queue_io_stat(q))
 189                 rw_flags |= REQ_IO_STAT;
 190
 191         INIT_LIST_HEAD(&rq->queuelist);
 192         /* csd/requeue_work/fifo_time is initialized before use */
 193         rq->q = q;
 194         rq->mq_ctx = ctx;
 195         rq->cmd_flags |= rw_flags;
 196         rq->cmd_type = 0;
 197         /* do not touch atomic flags, it needs atomic ops against the timer */
 198         rq->cpu = -1;
 199         rq->__data_len = 0;
 200         rq->__sector = (sector_t) -1;
 201         rq->bio = NULL;
 202         rq->biotail = NULL;
 203         INIT_HLIST_NODE(&rq->hash);
 204         RB_CLEAR_NODE(&rq->rb_node);
 205         memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv)));
 206         rq->rq_disk = NULL;
 207         rq->part = NULL;
 208         rq->start_time = jiffies;
 209 #ifdef CONFIG_BLK_CGROUP
 210         rq->rl = NULL;
 211         set_start_time_ns(rq);
 212         rq->io_start_time_ns = 0;
 213 #endif
 214         rq->nr_phys_segments = 0;
 215 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 216         rq->nr_integrity_segments = 0;
 217 #endif
 218         rq->ioprio = 0;
 219         rq->special = NULL;
 220         /* tag was already set */
 221         rq->errors = 0;
 222         memset(rq->__cmd, 0, sizeof(rq->__cmd));
 223         rq->cmd = rq->__cmd;
 224         rq->cmd_len = BLK_MAX_CDB;
 225
 226         rq->extra_len = 0;
 227         rq->sense_len = 0;
 228         rq->resid_len = 0;
 229         rq->sense = NULL;
 230
 231         rq->deadline = 0;
 232         INIT_LIST_HEAD(&rq->timeout_list);
 233         rq->timeout = 0;
 234         rq->retries = 0;
 235         rq->end_io = NULL;
 236         rq->end_io_data = NULL;
 237         rq->next_rq = NULL;
 238
 239         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 240 }
 241
 242 static struct request *
 243 __blk_mq_alloc_request(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 244                 struct blk_mq_ctx *ctx, int rw, gfp_t gfp, bool reserved)
 245 {
 246         struct request *rq;
 247         unsigned int tag;
 248
 249         tag = blk_mq_get_tag(hctx, &ctx->last_tag, gfp, reserved);
 250         if (tag != BLK_MQ_TAG_FAIL) {
 251                 rq = hctx->tags->rqs[tag];
 252
 253                 rq->cmd_flags = 0;
 254                 if (blk_mq_tag_busy(hctx)) {
 255                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 256                         atomic_inc(&hctx->nr_active);
 257                 }
 258
 259                 rq->tag = tag;
 260                 blk_mq_rq_ctx_init(q, ctx, rq, rw);
 261                 return rq;
 262         }
 263
 264         return NULL;
 265 }
 266
 267 static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
 268                                                    int rw, gfp_t gfp,
 269                                                    bool reserved)
 270 {
 271         bool gfp_mask = gfp & ~__GFP_WAIT;
 272         struct request *rq;
 273
 274         do {
 275                 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 276                 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
 277
 278                 rq = __blk_mq_alloc_request(q, hctx, ctx, rw, gfp_mask,
 279                                                 reserved);
 280                 if (rq)
 281                         break;
 282
 283                 if (!(gfp & __GFP_WAIT)) {
 284                         blk_mq_put_ctx(ctx);
 285                         break;
 286                 }
 287
 288                 __blk_mq_run_hw_queue(hctx);
 289                 blk_mq_put_ctx(ctx);
 290                 gfp_mask = gfp;
 291         } while (1);
 292
 293         return rq;
 294 }
 295
 296 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 297                 bool reserved)
 298 {
 299         struct request *rq;
 300
 301         if (blk_mq_queue_enter(q))
 302                 return NULL;
 303
 304         rq = blk_mq_alloc_request_pinned(q, rw, gfp, reserved);
 305         if (rq)
 306                 blk_mq_put_ctx(rq->mq_ctx);
 307         return rq;
 308 }
 309 EXPORT_SYMBOL(blk_mq_alloc_request);
 310
 311 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 312                                   struct blk_mq_ctx *ctx, struct request *rq)
 313 {
 314         const int tag = rq->tag;
 315         struct request_queue *q = rq->q;
 316
 317         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 318                 atomic_dec(&hctx->nr_active);
 319
 320         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 321         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 322         blk_mq_queue_exit(q);
 323 }
 324
 325 void blk_mq_free_request(struct request *rq)
 326 {
 327         struct blk_mq_ctx *ctx = rq->mq_ctx;
 328         struct blk_mq_hw_ctx *hctx;
 329         struct request_queue *q = rq->q;
 330
 331         ctx->rq_completed[rq_is_sync(rq)]++;
 332
 333         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 334         __blk_mq_free_request(hctx, ctx, rq);
 335 }
 336
 337 /*
 338  * Clone all relevant state from a request that has been put on hold in
 339  * the flush state machine into the preallocated flush request that hangs
 340  * off the request queue.
 341  *
 342  * For a driver the flush request should be invisible, that's why we are
 343  * impersonating the original request here.
 344  */
 345 void blk_mq_clone_flush_request(struct request *flush_rq,
 346                 struct request *orig_rq)
 347 {
 348         struct blk_mq_hw_ctx *hctx =
 349                 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
 350
 351         flush_rq->mq_ctx = orig_rq->mq_ctx;
 352         flush_rq->tag = orig_rq->tag;
 353         memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
 354                 hctx->cmd_size);
 355 }
 356
 357 inline void __blk_mq_end_io(struct request *rq, int error)
 358 {
 359         blk_account_io_done(rq);
 360
 361         if (rq->end_io) {
 362                 rq->end_io(rq, error);
 363         } else {
 364                 if (unlikely(blk_bidi_rq(rq)))
 365                         blk_mq_free_request(rq->next_rq);
 366                 blk_mq_free_request(rq);
 367         }
 368 }
 369 EXPORT_SYMBOL(__blk_mq_end_io);
 370
 371 void blk_mq_end_io(struct request *rq, int error)
 372 {
 373         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 374                 BUG();
 375         __blk_mq_end_io(rq, error);
 376 }
 377 EXPORT_SYMBOL(blk_mq_end_io);
 378
 379 static void __blk_mq_complete_request_remote(void *data)
 380 {
 381         struct request *rq = data;
 382
 383         rq->q->softirq_done_fn(rq);
 384 }
 385
 386 void __blk_mq_complete_request(struct request *rq)
 387 {
 388         struct blk_mq_ctx *ctx = rq->mq_ctx;
 389         bool shared = false;
 390         int cpu;
 391
 392         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 393                 rq->q->softirq_done_fn(rq);
 394                 return;
 395         }
 396
 397         cpu = get_cpu();
 398         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 399                 shared = cpus_share_cache(cpu, ctx->cpu);
 400
 401         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 402                 rq->csd.func = __blk_mq_complete_request_remote;
 403                 rq->csd.info = rq;
 404                 rq->csd.flags = 0;
 405                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 406         } else {
 407                 rq->q->softirq_done_fn(rq);
 408         }
 409         put_cpu();
 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                 if (q->softirq_done_fn)
 428                         __blk_mq_complete_request(rq);
 429                 else
 430                         blk_mq_end_io(rq, rq->errors);
 431         }
 432 }
 433 EXPORT_SYMBOL(blk_mq_complete_request);
 434
 435 static void blk_mq_start_request(struct request *rq, bool last)
 436 {
 437         struct request_queue *q = rq->q;
 438
 439         trace_block_rq_issue(q, rq);
 440
 441         rq->resid_len = blk_rq_bytes(rq);
 442         if (unlikely(blk_bidi_rq(rq)))
 443                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 444
 445         /*
 446          * Just mark start time and set the started bit. Due to memory
 447          * ordering, we know we'll see the correct deadline as long as
 448          * REQ_ATOMIC_STARTED is seen. Use the default queue timeout,
 449          * unless one has been set in the request.
 450          */
 451         if (!rq->timeout)
 452                 rq->deadline = jiffies + q->rq_timeout;
 453         else
 454                 rq->deadline = jiffies + rq->timeout;
 455
 456         /*
 457          * Mark us as started and clear complete. Complete might have been
 458          * set if requeue raced with timeout, which then marked it as
 459          * complete. So be sure to clear complete again when we start
 460          * the request, otherwise we'll ignore the completion event.
 461          */
 462         set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 463         clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 464
 465         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 466                 /*
 467                  * Make sure space for the drain appears.  We know we can do
 468                  * this because max_hw_segments has been adjusted to be one
 469                  * fewer than the device can handle.
 470                  */
 471                 rq->nr_phys_segments++;
 472         }
 473
 474         /*
 475          * Flag the last request in the series so that drivers know when IO
 476          * should be kicked off, if they don't do it on a per-request basis.
 477          *
 478          * Note: the flag isn't the only condition drivers should do kick off.
 479          * If drive is busy, the last request might not have the bit set.
 480          */
 481         if (last)
 482                 rq->cmd_flags |= REQ_END;
 483 }
 484
 485 static void __blk_mq_requeue_request(struct request *rq)
 486 {
 487         struct request_queue *q = rq->q;
 488
 489         trace_block_rq_requeue(q, rq);
 490         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 491
 492         rq->cmd_flags &= ~REQ_END;
 493
 494         if (q->dma_drain_size && blk_rq_bytes(rq))
 495                 rq->nr_phys_segments--;
 496 }
 497
 498 void blk_mq_requeue_request(struct request *rq)
 499 {
 500         __blk_mq_requeue_request(rq);
 501         blk_clear_rq_complete(rq);
 502
 503         BUG_ON(blk_queued_rq(rq));
 504         blk_mq_add_to_requeue_list(rq, true);
 505 }
 506 EXPORT_SYMBOL(blk_mq_requeue_request);
 507
 508 static void blk_mq_requeue_work(struct work_struct *work)
 509 {
 510         struct request_queue *q =
 511                 container_of(work, struct request_queue, requeue_work);
 512         LIST_HEAD(rq_list);
 513         struct request *rq, *next;
 514         unsigned long flags;
 515
 516         spin_lock_irqsave(&q->requeue_lock, flags);
 517         list_splice_init(&q->requeue_list, &rq_list);
 518         spin_unlock_irqrestore(&q->requeue_lock, flags);
 519
 520         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 521                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 522                         continue;
 523
 524                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 525                 list_del_init(&rq->queuelist);
 526                 blk_mq_insert_request(rq, true, false, false);
 527         }
 528
 529         while (!list_empty(&rq_list)) {
 530                 rq = list_entry(rq_list.next, struct request, queuelist);
 531                 list_del_init(&rq->queuelist);
 532                 blk_mq_insert_request(rq, false, false, false);
 533         }
 534
 535         blk_mq_run_queues(q, false);
 536 }
 537
 538 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 539 {
 540         struct request_queue *q = rq->q;
 541         unsigned long flags;
 542
 543         /*
 544          * We abuse this flag that is otherwise used by the I/O scheduler to
 545          * request head insertation from the workqueue.
 546          */
 547         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 548
 549         spin_lock_irqsave(&q->requeue_lock, flags);
 550         if (at_head) {
 551                 rq->cmd_flags |= REQ_SOFTBARRIER;
 552                 list_add(&rq->queuelist, &q->requeue_list);
 553         } else {
 554                 list_add_tail(&rq->queuelist, &q->requeue_list);
 555         }
 556         spin_unlock_irqrestore(&q->requeue_lock, flags);
 557 }
 558 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 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 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 567 {
 568         return tags->rqs[tag];
 569 }
 570 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 571
 572 struct blk_mq_timeout_data {
 573         struct blk_mq_hw_ctx *hctx;
 574         unsigned long *next;
 575         unsigned int *next_set;
 576 };
 577
 578 static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
 579 {
 580         struct blk_mq_timeout_data *data = __data;
 581         struct blk_mq_hw_ctx *hctx = data->hctx;
 582         unsigned int tag;
 583
 584          /* It may not be in flight yet (this is where
 585          * the REQ_ATOMIC_STARTED flag comes in). The requests are
 586          * statically allocated, so we know it's always safe to access the
 587          * memory associated with a bit offset into ->rqs[].
 588          */
 589         tag = 0;
 590         do {
 591                 struct request *rq;
 592
 593                 tag = find_next_zero_bit(free_tags, hctx->tags->nr_tags, tag);
 594                 if (tag >= hctx->tags->nr_tags)
 595                         break;
 596
 597                 rq = blk_mq_tag_to_rq(hctx->tags, tag++);
 598                 if (rq->q != hctx->queue)
 599                         continue;
 600                 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 601                         continue;
 602
 603                 blk_rq_check_expired(rq, data->next, data->next_set);
 604         } while (1);
 605 }
 606
 607 static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
 608                                         unsigned long *next,
 609                                         unsigned int *next_set)
 610 {
 611         struct blk_mq_timeout_data data = {
 612                 .hctx           = hctx,
 613                 .next           = next,
 614                 .next_set       = next_set,
 615         };
 616
 617         /*
 618          * Ask the tagging code to iterate busy requests, so we can
 619          * check them for timeout.
 620          */
 621         blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
 622 }
 623
 624 static enum blk_eh_timer_return blk_mq_rq_timed_out(struct request *rq)
 625 {
 626         struct request_queue *q = rq->q;
 627
 628         /*
 629          * We know that complete is set at this point. If STARTED isn't set
 630          * anymore, then the request isn't active and the "timeout" should
 631          * just be ignored. This can happen due to the bitflag ordering.
 632          * Timeout first checks if STARTED is set, and if it is, assumes
 633          * the request is active. But if we race with completion, then
 634          * we both flags will get cleared. So check here again, and ignore
 635          * a timeout event with a request that isn't active.
 636          */
 637         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 638                 return BLK_EH_NOT_HANDLED;
 639
 640         if (!q->mq_ops->timeout)
 641                 return BLK_EH_RESET_TIMER;
 642
 643         return q->mq_ops->timeout(rq);
 644 }
 645
 646 static void blk_mq_rq_timer(unsigned long data)
 647 {
 648         struct request_queue *q = (struct request_queue *) data;
 649         struct blk_mq_hw_ctx *hctx;
 650         unsigned long next = 0;
 651         int i, next_set = 0;
 652
 653         queue_for_each_hw_ctx(q, hctx, i) {
 654                 /*
 655                  * If not software queues are currently mapped to this
 656                  * hardware queue, there's nothing to check
 657                  */
 658                 if (!hctx->nr_ctx || !hctx->tags)
 659                         continue;
 660
 661                 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
 662         }
 663
 664         if (next_set) {
 665                 next = blk_rq_timeout(round_jiffies_up(next));
 666                 mod_timer(&q->timeout, next);
 667         } else {
 668                 queue_for_each_hw_ctx(q, hctx, i)
 669                         blk_mq_tag_idle(hctx);
 670         }
 671 }
 672
 673 /*
 674  * Reverse check our software queue for entries that we could potentially
 675  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 676  * too much time checking for merges.
 677  */
 678 static bool blk_mq_attempt_merge(struct request_queue *q,
 679                                  struct blk_mq_ctx *ctx, struct bio *bio)
 680 {
 681         struct request *rq;
 682         int checked = 8;
 683
 684         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 685                 int el_ret;
 686
 687                 if (!checked--)
 688                         break;
 689
 690                 if (!blk_rq_merge_ok(rq, bio))
 691                         continue;
 692
 693                 el_ret = blk_try_merge(rq, bio);
 694                 if (el_ret == ELEVATOR_BACK_MERGE) {
 695                         if (bio_attempt_back_merge(q, rq, bio)) {
 696                                 ctx->rq_merged++;
 697                                 return true;
 698                         }
 699                         break;
 700                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 701                         if (bio_attempt_front_merge(q, rq, bio)) {
 702                                 ctx->rq_merged++;
 703                                 return true;
 704                         }
 705                         break;
 706                 }
 707         }
 708
 709         return false;
 710 }
 711
 712 /*
 713  * Process software queues that have been marked busy, splicing them
 714  * to the for-dispatch
 715  */
 716 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 717 {
 718         struct blk_mq_ctx *ctx;
 719         int i;
 720
 721         for (i = 0; i < hctx->ctx_map.map_size; i++) {
 722                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 723                 unsigned int off, bit;
 724
 725                 if (!bm->word)
 726                         continue;
 727
 728                 bit = 0;
 729                 off = i * hctx->ctx_map.bits_per_word;
 730                 do {
 731                         bit = find_next_bit(&bm->word, bm->depth, bit);
 732                         if (bit >= bm->depth)
 733                                 break;
 734
 735                         ctx = hctx->ctxs[bit + off];
 736                         clear_bit(bit, &bm->word);
 737                         spin_lock(&ctx->lock);
 738                         list_splice_tail_init(&ctx->rq_list, list);
 739                         spin_unlock(&ctx->lock);
 740
 741                         bit++;
 742                 } while (1);
 743         }
 744 }
 745
 746 /*
 747  * Run this hardware queue, pulling any software queues mapped to it in.
 748  * Note that this function currently has various problems around ordering
 749  * of IO. In particular, we'd like FIFO behaviour on handling existing
 750  * items on the hctx->dispatch list. Ignore that for now.
 751  */
 752 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 753 {
 754         struct request_queue *q = hctx->queue;
 755         struct request *rq;
 756         LIST_HEAD(rq_list);
 757         int queued;
 758
 759         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 760
 761         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 762                 return;
 763
 764         hctx->run++;
 765
 766         /*
 767          * Touch any software queue that has pending entries.
 768          */
 769         flush_busy_ctxs(hctx, &rq_list);
 770
 771         /*
 772          * If we have previous entries on our dispatch list, grab them
 773          * and stuff them at the front for more fair dispatch.
 774          */
 775         if (!list_empty_careful(&hctx->dispatch)) {
 776                 spin_lock(&hctx->lock);
 777                 if (!list_empty(&hctx->dispatch))
 778                         list_splice_init(&hctx->dispatch, &rq_list);
 779                 spin_unlock(&hctx->lock);
 780         }
 781
 782         /*
 783          * Now process all the entries, sending them to the driver.
 784          */
 785         queued = 0;
 786         while (!list_empty(&rq_list)) {
 787                 int ret;
 788
 789                 rq = list_first_entry(&rq_list, struct request, queuelist);
 790                 list_del_init(&rq->queuelist);
 791
 792                 blk_mq_start_request(rq, list_empty(&rq_list));
 793
 794                 ret = q->mq_ops->queue_rq(hctx, rq);
 795                 switch (ret) {
 796                 case BLK_MQ_RQ_QUEUE_OK:
 797                         queued++;
 798                         continue;
 799                 case BLK_MQ_RQ_QUEUE_BUSY:
 800                         list_add(&rq->queuelist, &rq_list);
 801                         __blk_mq_requeue_request(rq);
 802                         break;
 803                 default:
 804                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 805                 case BLK_MQ_RQ_QUEUE_ERROR:
 806                         rq->errors = -EIO;
 807                         blk_mq_end_io(rq, rq->errors);
 808                         break;
 809                 }
 810
 811                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 812                         break;
 813         }
 814
 815         if (!queued)
 816                 hctx->dispatched[0]++;
 817         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 818                 hctx->dispatched[ilog2(queued) + 1]++;
 819
 820         /*
 821          * Any items that need requeuing? Stuff them into hctx->dispatch,
 822          * that is where we will continue on next queue run.
 823          */
 824         if (!list_empty(&rq_list)) {
 825                 spin_lock(&hctx->lock);
 826                 list_splice(&rq_list, &hctx->dispatch);
 827                 spin_unlock(&hctx->lock);
 828         }
 829 }
 830
 831 /*
 832  * It'd be great if the workqueue API had a way to pass
 833  * in a mask and had some smarts for more clever placement.
 834  * For now we just round-robin here, switching for every
 835  * BLK_MQ_CPU_WORK_BATCH queued items.
 836  */
 837 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 838 {
 839         int cpu = hctx->next_cpu;
 840
 841         if (--hctx->next_cpu_batch <= 0) {
 842                 int next_cpu;
 843
 844                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 845                 if (next_cpu >= nr_cpu_ids)
 846                         next_cpu = cpumask_first(hctx->cpumask);
 847
 848                 hctx->next_cpu = next_cpu;
 849                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 850         }
 851
 852         return cpu;
 853 }
 854
 855 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 856 {
 857         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 858                 return;
 859
 860         if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
 861                 __blk_mq_run_hw_queue(hctx);
 862         else if (hctx->queue->nr_hw_queues == 1)
 863                 kblockd_schedule_delayed_work(&hctx->run_work, 0);
 864         else {
 865                 unsigned int cpu;
 866
 867                 cpu = blk_mq_hctx_next_cpu(hctx);
 868                 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
 869         }
 870 }
 871
 872 void blk_mq_run_queues(struct request_queue *q, bool async)
 873 {
 874         struct blk_mq_hw_ctx *hctx;
 875         int i;
 876
 877         queue_for_each_hw_ctx(q, hctx, i) {
 878                 if ((!blk_mq_hctx_has_pending(hctx) &&
 879                     list_empty_careful(&hctx->dispatch)) ||
 880                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 881                         continue;
 882
 883                 preempt_disable();
 884                 blk_mq_run_hw_queue(hctx, async);
 885                 preempt_enable();
 886         }
 887 }
 888 EXPORT_SYMBOL(blk_mq_run_queues);
 889
 890 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 891 {
 892         cancel_delayed_work(&hctx->run_work);
 893         cancel_delayed_work(&hctx->delay_work);
 894         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 895 }
 896 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 897
 898 void blk_mq_stop_hw_queues(struct request_queue *q)
 899 {
 900         struct blk_mq_hw_ctx *hctx;
 901         int i;
 902
 903         queue_for_each_hw_ctx(q, hctx, i)
 904                 blk_mq_stop_hw_queue(hctx);
 905 }
 906 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 907
 908 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 909 {
 910         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 911
 912         preempt_disable();
 913         __blk_mq_run_hw_queue(hctx);
 914         preempt_enable();
 915 }
 916 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 917
 918 void blk_mq_start_hw_queues(struct request_queue *q)
 919 {
 920         struct blk_mq_hw_ctx *hctx;
 921         int i;
 922
 923         queue_for_each_hw_ctx(q, hctx, i)
 924                 blk_mq_start_hw_queue(hctx);
 925 }
 926 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 927
 928
 929 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 930 {
 931         struct blk_mq_hw_ctx *hctx;
 932         int i;
 933
 934         queue_for_each_hw_ctx(q, hctx, i) {
 935                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 936                         continue;
 937
 938                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 939                 preempt_disable();
 940                 blk_mq_run_hw_queue(hctx, async);
 941                 preempt_enable();
 942         }
 943 }
 944 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 945
 946 static void blk_mq_run_work_fn(struct work_struct *work)
 947 {
 948         struct blk_mq_hw_ctx *hctx;
 949
 950         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 951
 952         __blk_mq_run_hw_queue(hctx);
 953 }
 954
 955 static void blk_mq_delay_work_fn(struct work_struct *work)
 956 {
 957         struct blk_mq_hw_ctx *hctx;
 958
 959         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 960
 961         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 962                 __blk_mq_run_hw_queue(hctx);
 963 }
 964
 965 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 966 {
 967         unsigned long tmo = msecs_to_jiffies(msecs);
 968
 969         if (hctx->queue->nr_hw_queues == 1)
 970                 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
 971         else {
 972                 unsigned int cpu;
 973
 974                 cpu = blk_mq_hctx_next_cpu(hctx);
 975                 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
 976         }
 977 }
 978 EXPORT_SYMBOL(blk_mq_delay_queue);
 979
 980 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 981                                     struct request *rq, bool at_head)
 982 {
 983         struct blk_mq_ctx *ctx = rq->mq_ctx;
 984
 985         trace_block_rq_insert(hctx->queue, rq);
 986
 987         if (at_head)
 988                 list_add(&rq->queuelist, &ctx->rq_list);
 989         else
 990                 list_add_tail(&rq->queuelist, &ctx->rq_list);
 991
 992         blk_mq_hctx_mark_pending(hctx, ctx);
 993
 994         /*
 995          * We do this early, to ensure we are on the right CPU.
 996          */
 997         blk_add_timer(rq);
 998 }
 999
1000 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1001                 bool async)
1002 {
1003         struct request_queue *q = rq->q;
1004         struct blk_mq_hw_ctx *hctx;
1005         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1006
1007         current_ctx = blk_mq_get_ctx(q);
1008         if (!cpu_online(ctx->cpu))
1009                 rq->mq_ctx = ctx = current_ctx;
1010
1011         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1012
1013         if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
1014             !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
1015                 blk_insert_flush(rq);
1016         } else {
1017                 spin_lock(&ctx->lock);
1018                 __blk_mq_insert_request(hctx, rq, at_head);
1019                 spin_unlock(&ctx->lock);
1020         }
1021
1022         if (run_queue)
1023                 blk_mq_run_hw_queue(hctx, async);
1024
1025         blk_mq_put_ctx(current_ctx);
1026 }
1027
1028 static void blk_mq_insert_requests(struct request_queue *q,
1029                                      struct blk_mq_ctx *ctx,
1030                                      struct list_head *list,
1031                                      int depth,
1032                                      bool from_schedule)
1033
1034 {
1035         struct blk_mq_hw_ctx *hctx;
1036         struct blk_mq_ctx *current_ctx;
1037
1038         trace_block_unplug(q, depth, !from_schedule);
1039
1040         current_ctx = blk_mq_get_ctx(q);
1041
1042         if (!cpu_online(ctx->cpu))
1043                 ctx = current_ctx;
1044         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1045
1046         /*
1047          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1048          * offline now
1049          */
1050         spin_lock(&ctx->lock);
1051         while (!list_empty(list)) {
1052                 struct request *rq;
1053
1054                 rq = list_first_entry(list, struct request, queuelist);
1055                 list_del_init(&rq->queuelist);
1056                 rq->mq_ctx = ctx;
1057                 __blk_mq_insert_request(hctx, rq, false);
1058         }
1059         spin_unlock(&ctx->lock);
1060
1061         blk_mq_run_hw_queue(hctx, from_schedule);
1062         blk_mq_put_ctx(current_ctx);
1063 }
1064
1065 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1066 {
1067         struct request *rqa = container_of(a, struct request, queuelist);
1068         struct request *rqb = container_of(b, struct request, queuelist);
1069
1070         return !(rqa->mq_ctx < rqb->mq_ctx ||
1071                  (rqa->mq_ctx == rqb->mq_ctx &&
1072                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1073 }
1074
1075 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1076 {
1077         struct blk_mq_ctx *this_ctx;
1078         struct request_queue *this_q;
1079         struct request *rq;
1080         LIST_HEAD(list);
1081         LIST_HEAD(ctx_list);
1082         unsigned int depth;
1083
1084         list_splice_init(&plug->mq_list, &list);
1085
1086         list_sort(NULL, &list, plug_ctx_cmp);
1087
1088         this_q = NULL;
1089         this_ctx = NULL;
1090         depth = 0;
1091
1092         while (!list_empty(&list)) {
1093                 rq = list_entry_rq(list.next);
1094                 list_del_init(&rq->queuelist);
1095                 BUG_ON(!rq->q);
1096                 if (rq->mq_ctx != this_ctx) {
1097                         if (this_ctx) {
1098                                 blk_mq_insert_requests(this_q, this_ctx,
1099                                                         &ctx_list, depth,
1100                                                         from_schedule);
1101                         }
1102
1103                         this_ctx = rq->mq_ctx;
1104                         this_q = rq->q;
1105                         depth = 0;
1106                 }
1107
1108                 depth++;
1109                 list_add_tail(&rq->queuelist, &ctx_list);
1110         }
1111
1112         /*
1113          * If 'this_ctx' is set, we know we have entries to complete
1114          * on 'ctx_list'. Do those.
1115          */
1116         if (this_ctx) {
1117                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1118                                        from_schedule);
1119         }
1120 }
1121
1122 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1123 {
1124         init_request_from_bio(rq, bio);
1125         blk_account_io_start(rq, 1);
1126 }
1127
1128 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1129                                          struct blk_mq_ctx *ctx,
1130                                          struct request *rq, struct bio *bio)
1131 {
1132         struct request_queue *q = hctx->queue;
1133
1134         if (!(hctx->flags & BLK_MQ_F_SHOULD_MERGE)) {
1135                 blk_mq_bio_to_request(rq, bio);
1136                 spin_lock(&ctx->lock);
1137 insert_rq:
1138                 __blk_mq_insert_request(hctx, rq, false);
1139                 spin_unlock(&ctx->lock);
1140                 return false;
1141         } else {
1142                 spin_lock(&ctx->lock);
1143                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1144                         blk_mq_bio_to_request(rq, bio);
1145                         goto insert_rq;
1146                 }
1147
1148                 spin_unlock(&ctx->lock);
1149                 __blk_mq_free_request(hctx, ctx, rq);
1150                 return true;
1151         }
1152 }
1153
1154 struct blk_map_ctx {
1155         struct blk_mq_hw_ctx *hctx;
1156         struct blk_mq_ctx *ctx;
1157 };
1158
1159 static struct request *blk_mq_map_request(struct request_queue *q,
1160                                           struct bio *bio,
1161                                           struct blk_map_ctx *data)
1162 {
1163         struct blk_mq_hw_ctx *hctx;
1164         struct blk_mq_ctx *ctx;
1165         struct request *rq;
1166         int rw = bio_data_dir(bio);
1167
1168         if (unlikely(blk_mq_queue_enter(q))) {
1169                 bio_endio(bio, -EIO);
1170                 return NULL;
1171         }
1172
1173         ctx = blk_mq_get_ctx(q);
1174         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1175
1176         if (rw_is_sync(bio->bi_rw))
1177                 rw |= REQ_SYNC;
1178
1179         trace_block_getrq(q, bio, rw);
1180         rq = __blk_mq_alloc_request(q, hctx, ctx, rw, GFP_ATOMIC, false);
1181         if (unlikely(!rq)) {
1182                 __blk_mq_run_hw_queue(hctx);
1183                 blk_mq_put_ctx(ctx);
1184                 trace_block_sleeprq(q, bio, rw);
1185
1186                 ctx = blk_mq_get_ctx(q);
1187                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1188                 rq = __blk_mq_alloc_request(q, hctx, ctx, rw,
1189                                             __GFP_WAIT|GFP_ATOMIC, false);
1190         }
1191
1192         hctx->queued++;
1193         data->hctx = hctx;
1194         data->ctx = ctx;
1195         return rq;
1196 }
1197
1198 /*
1199  * Multiple hardware queue variant. This will not use per-process plugs,
1200  * but will attempt to bypass the hctx queueing if we can go straight to
1201  * hardware for SYNC IO.
1202  */
1203 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1204 {
1205         const int is_sync = rw_is_sync(bio->bi_rw);
1206         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1207         struct blk_map_ctx data;
1208         struct request *rq;
1209
1210         blk_queue_bounce(q, &bio);
1211
1212         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1213                 bio_endio(bio, -EIO);
1214                 return;
1215         }
1216
1217         rq = blk_mq_map_request(q, bio, &data);
1218         if (unlikely(!rq))
1219                 return;
1220
1221         if (unlikely(is_flush_fua)) {
1222                 blk_mq_bio_to_request(rq, bio);
1223                 blk_insert_flush(rq);
1224                 goto run_queue;
1225         }
1226
1227         if (is_sync) {
1228                 int ret;
1229
1230                 blk_mq_bio_to_request(rq, bio);
1231                 blk_mq_start_request(rq, true);
1232
1233                 /*
1234                  * For OK queue, we are done. For error, kill it. Any other
1235                  * error (busy), just add it to our list as we previously
1236                  * would have done
1237                  */
1238                 ret = q->mq_ops->queue_rq(data.hctx, rq);
1239                 if (ret == BLK_MQ_RQ_QUEUE_OK)
1240                         goto done;
1241                 else {
1242                         __blk_mq_requeue_request(rq);
1243
1244                         if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1245                                 rq->errors = -EIO;
1246                                 blk_mq_end_io(rq, rq->errors);
1247                                 goto done;
1248                         }
1249                 }
1250         }
1251
1252         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1253                 /*
1254                  * For a SYNC request, send it to the hardware immediately. For
1255                  * an ASYNC request, just ensure that we run it later on. The
1256                  * latter allows for merging opportunities and more efficient
1257                  * dispatching.
1258                  */
1259 run_queue:
1260                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1261         }
1262 done:
1263         blk_mq_put_ctx(data.ctx);
1264 }
1265
1266 /*
1267  * Single hardware queue variant. This will attempt to use any per-process
1268  * plug for merging and IO deferral.
1269  */
1270 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1271 {
1272         const int is_sync = rw_is_sync(bio->bi_rw);
1273         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1274         unsigned int use_plug, request_count = 0;
1275         struct blk_map_ctx data;
1276         struct request *rq;
1277
1278         /*
1279          * If we have multiple hardware queues, just go directly to
1280          * one of those for sync IO.
1281          */
1282         use_plug = !is_flush_fua && !is_sync;
1283
1284         blk_queue_bounce(q, &bio);
1285
1286         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1287                 bio_endio(bio, -EIO);
1288                 return;
1289         }
1290
1291         if (use_plug && !blk_queue_nomerges(q) &&
1292             blk_attempt_plug_merge(q, bio, &request_count))
1293                 return;
1294
1295         rq = blk_mq_map_request(q, bio, &data);
1296
1297         if (unlikely(is_flush_fua)) {
1298                 blk_mq_bio_to_request(rq, bio);
1299                 blk_insert_flush(rq);
1300                 goto run_queue;
1301         }
1302
1303         /*
1304          * A task plug currently exists. Since this is completely lockless,
1305          * utilize that to temporarily store requests until the task is
1306          * either done or scheduled away.
1307          */
1308         if (use_plug) {
1309                 struct blk_plug *plug = current->plug;
1310
1311                 if (plug) {
1312                         blk_mq_bio_to_request(rq, bio);
1313                         if (list_empty(&plug->mq_list))
1314                                 trace_block_plug(q);
1315                         else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1316                                 blk_flush_plug_list(plug, false);
1317                                 trace_block_plug(q);
1318                         }
1319                         list_add_tail(&rq->queuelist, &plug->mq_list);
1320                         blk_mq_put_ctx(data.ctx);
1321                         return;
1322                 }
1323         }
1324
1325         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1326                 /*
1327                  * For a SYNC request, send it to the hardware immediately. For
1328                  * an ASYNC request, just ensure that we run it later on. The
1329                  * latter allows for merging opportunities and more efficient
1330                  * dispatching.
1331                  */
1332 run_queue:
1333                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1334         }
1335
1336         blk_mq_put_ctx(data.ctx);
1337 }
1338
1339 /*
1340  * Default mapping to a software queue, since we use one per CPU.
1341  */
1342 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1343 {
1344         return q->queue_hw_ctx[q->mq_map[cpu]];
1345 }
1346 EXPORT_SYMBOL(blk_mq_map_queue);
1347
1348 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
1349                                                    unsigned int hctx_index,
1350                                                    int node)
1351 {
1352         return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL, node);
1353 }
1354 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
1355
1356 void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
1357                                  unsigned int hctx_index)
1358 {
1359         kfree(hctx);
1360 }
1361 EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
1362
1363 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1364                 struct blk_mq_tags *tags, unsigned int hctx_idx)
1365 {
1366         struct page *page;
1367
1368         if (tags->rqs && set->ops->exit_request) {
1369                 int i;
1370
1371                 for (i = 0; i < tags->nr_tags; i++) {
1372                         if (!tags->rqs[i])
1373                                 continue;
1374                         set->ops->exit_request(set->driver_data, tags->rqs[i],
1375                                                 hctx_idx, i);
1376                 }
1377         }
1378
1379         while (!list_empty(&tags->page_list)) {
1380                 page = list_first_entry(&tags->page_list, struct page, lru);
1381                 list_del_init(&page->lru);
1382                 __free_pages(page, page->private);
1383         }
1384
1385         kfree(tags->rqs);
1386
1387         blk_mq_free_tags(tags);
1388 }
1389
1390 static size_t order_to_size(unsigned int order)
1391 {
1392         return (size_t)PAGE_SIZE << order;
1393 }
1394
1395 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1396                 unsigned int hctx_idx)
1397 {
1398         struct blk_mq_tags *tags;
1399         unsigned int i, j, entries_per_page, max_order = 4;
1400         size_t rq_size, left;
1401
1402         tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1403                                 set->numa_node);
1404         if (!tags)
1405                 return NULL;
1406
1407         INIT_LIST_HEAD(&tags->page_list);
1408
1409         tags->rqs = kmalloc_node(set->queue_depth * sizeof(struct request *),
1410                                         GFP_KERNEL, set->numa_node);
1411         if (!tags->rqs) {
1412                 blk_mq_free_tags(tags);
1413                 return NULL;
1414         }
1415
1416         /*
1417          * rq_size is the size of the request plus driver payload, rounded
1418          * to the cacheline size
1419          */
1420         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1421                                 cache_line_size());
1422         left = rq_size * set->queue_depth;
1423
1424         for (i = 0; i < set->queue_depth; ) {
1425                 int this_order = max_order;
1426                 struct page *page;
1427                 int to_do;
1428                 void *p;
1429
1430                 while (left < order_to_size(this_order - 1) && this_order)
1431                         this_order--;
1432
1433                 do {
1434                         page = alloc_pages_node(set->numa_node, GFP_KERNEL,
1435                                                 this_order);
1436                         if (page)
1437                                 break;
1438                         if (!this_order--)
1439                                 break;
1440                         if (order_to_size(this_order) < rq_size)
1441                                 break;
1442                 } while (1);
1443
1444                 if (!page)
1445                         goto fail;
1446
1447                 page->private = this_order;
1448                 list_add_tail(&page->lru, &tags->page_list);
1449
1450                 p = page_address(page);
1451                 entries_per_page = order_to_size(this_order) / rq_size;
1452                 to_do = min(entries_per_page, set->queue_depth - i);
1453                 left -= to_do * rq_size;
1454                 for (j = 0; j < to_do; j++) {
1455                         tags->rqs[i] = p;
1456                         if (set->ops->init_request) {
1457                                 if (set->ops->init_request(set->driver_data,
1458                                                 tags->rqs[i], hctx_idx, i,
1459                                                 set->numa_node))
1460                                         goto fail;
1461                         }
1462
1463                         p += rq_size;
1464                         i++;
1465                 }
1466         }
1467
1468         return tags;
1469
1470 fail:
1471         pr_warn("%s: failed to allocate requests\n", __func__);
1472         blk_mq_free_rq_map(set, tags, hctx_idx);
1473         return NULL;
1474 }
1475
1476 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1477 {
1478         kfree(bitmap->map);
1479 }
1480
1481 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1482 {
1483         unsigned int bpw = 8, total, num_maps, i;
1484
1485         bitmap->bits_per_word = bpw;
1486
1487         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1488         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1489                                         GFP_KERNEL, node);
1490         if (!bitmap->map)
1491                 return -ENOMEM;
1492
1493         bitmap->map_size = num_maps;
1494
1495         total = nr_cpu_ids;
1496         for (i = 0; i < num_maps; i++) {
1497                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1498                 total -= bitmap->map[i].depth;
1499         }
1500
1501         return 0;
1502 }
1503
1504 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1505 {
1506         struct request_queue *q = hctx->queue;
1507         struct blk_mq_ctx *ctx;
1508         LIST_HEAD(tmp);
1509
1510         /*
1511          * Move ctx entries to new CPU, if this one is going away.
1512          */
1513         ctx = __blk_mq_get_ctx(q, cpu);
1514
1515         spin_lock(&ctx->lock);
1516         if (!list_empty(&ctx->rq_list)) {
1517                 list_splice_init(&ctx->rq_list, &tmp);
1518                 blk_mq_hctx_clear_pending(hctx, ctx);
1519         }
1520         spin_unlock(&ctx->lock);
1521
1522         if (list_empty(&tmp))
1523                 return NOTIFY_OK;
1524
1525         ctx = blk_mq_get_ctx(q);
1526         spin_lock(&ctx->lock);
1527
1528         while (!list_empty(&tmp)) {
1529                 struct request *rq;
1530
1531                 rq = list_first_entry(&tmp, struct request, queuelist);
1532                 rq->mq_ctx = ctx;
1533                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1534         }
1535
1536         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1537         blk_mq_hctx_mark_pending(hctx, ctx);
1538
1539         spin_unlock(&ctx->lock);
1540
1541         blk_mq_run_hw_queue(hctx, true);
1542         blk_mq_put_ctx(ctx);
1543         return NOTIFY_OK;
1544 }
1545
1546 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1547 {
1548         struct request_queue *q = hctx->queue;
1549         struct blk_mq_tag_set *set = q->tag_set;
1550
1551         if (set->tags[hctx->queue_num])
1552                 return NOTIFY_OK;
1553
1554         set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1555         if (!set->tags[hctx->queue_num])
1556                 return NOTIFY_STOP;
1557
1558         hctx->tags = set->tags[hctx->queue_num];
1559         return NOTIFY_OK;
1560 }
1561
1562 static int blk_mq_hctx_notify(void *data, unsigned long action,
1563                               unsigned int cpu)
1564 {
1565         struct blk_mq_hw_ctx *hctx = data;
1566
1567         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1568                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1569         else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1570                 return blk_mq_hctx_cpu_online(hctx, cpu);
1571
1572         return NOTIFY_OK;
1573 }
1574
1575 static void blk_mq_exit_hw_queues(struct request_queue *q,
1576                 struct blk_mq_tag_set *set, int nr_queue)
1577 {
1578         struct blk_mq_hw_ctx *hctx;
1579         unsigned int i;
1580
1581         queue_for_each_hw_ctx(q, hctx, i) {
1582                 if (i == nr_queue)
1583                         break;
1584
1585                 if (set->ops->exit_hctx)
1586                         set->ops->exit_hctx(hctx, i);
1587
1588                 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1589                 kfree(hctx->ctxs);
1590                 blk_mq_free_bitmap(&hctx->ctx_map);
1591         }
1592
1593 }
1594
1595 static void blk_mq_free_hw_queues(struct request_queue *q,
1596                 struct blk_mq_tag_set *set)
1597 {
1598         struct blk_mq_hw_ctx *hctx;
1599         unsigned int i;
1600
1601         queue_for_each_hw_ctx(q, hctx, i) {
1602                 free_cpumask_var(hctx->cpumask);
1603                 set->ops->free_hctx(hctx, i);
1604         }
1605 }
1606
1607 static int blk_mq_init_hw_queues(struct request_queue *q,
1608                 struct blk_mq_tag_set *set)
1609 {
1610         struct blk_mq_hw_ctx *hctx;
1611         unsigned int i;
1612
1613         /*
1614          * Initialize hardware queues
1615          */
1616         queue_for_each_hw_ctx(q, hctx, i) {
1617                 int node;
1618
1619                 node = hctx->numa_node;
1620                 if (node == NUMA_NO_NODE)
1621                         node = hctx->numa_node = set->numa_node;
1622
1623                 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1624                 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1625                 spin_lock_init(&hctx->lock);
1626                 INIT_LIST_HEAD(&hctx->dispatch);
1627                 hctx->queue = q;
1628                 hctx->queue_num = i;
1629                 hctx->flags = set->flags;
1630                 hctx->cmd_size = set->cmd_size;
1631
1632                 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1633                                                 blk_mq_hctx_notify, hctx);
1634                 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1635
1636                 hctx->tags = set->tags[i];
1637
1638                 /*
1639                  * Allocate space for all possible cpus to avoid allocation in
1640                  * runtime
1641                  */
1642                 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1643                                                 GFP_KERNEL, node);
1644                 if (!hctx->ctxs)
1645                         break;
1646
1647                 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1648                         break;
1649
1650                 hctx->nr_ctx = 0;
1651
1652                 if (set->ops->init_hctx &&
1653                     set->ops->init_hctx(hctx, set->driver_data, i))
1654                         break;
1655         }
1656
1657         if (i == q->nr_hw_queues)
1658                 return 0;
1659
1660         /*
1661          * Init failed
1662          */
1663         blk_mq_exit_hw_queues(q, set, i);
1664
1665         return 1;
1666 }
1667
1668 static void blk_mq_init_cpu_queues(struct request_queue *q,
1669                                    unsigned int nr_hw_queues)
1670 {
1671         unsigned int i;
1672
1673         for_each_possible_cpu(i) {
1674                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1675                 struct blk_mq_hw_ctx *hctx;
1676
1677                 memset(__ctx, 0, sizeof(*__ctx));
1678                 __ctx->cpu = i;
1679                 spin_lock_init(&__ctx->lock);
1680                 INIT_LIST_HEAD(&__ctx->rq_list);
1681                 __ctx->queue = q;
1682
1683                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1684                 if (!cpu_online(i))
1685                         continue;
1686
1687                 hctx = q->mq_ops->map_queue(q, i);
1688                 cpumask_set_cpu(i, hctx->cpumask);
1689                 hctx->nr_ctx++;
1690
1691                 /*
1692                  * Set local node, IFF we have more than one hw queue. If
1693                  * not, we remain on the home node of the device
1694                  */
1695                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1696                         hctx->numa_node = cpu_to_node(i);
1697         }
1698 }
1699
1700 static void blk_mq_map_swqueue(struct request_queue *q)
1701 {
1702         unsigned int i;
1703         struct blk_mq_hw_ctx *hctx;
1704         struct blk_mq_ctx *ctx;
1705
1706         queue_for_each_hw_ctx(q, hctx, i) {
1707                 cpumask_clear(hctx->cpumask);
1708                 hctx->nr_ctx = 0;
1709         }
1710
1711         /*
1712          * Map software to hardware queues
1713          */
1714         queue_for_each_ctx(q, ctx, i) {
1715                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1716                 if (!cpu_online(i))
1717                         continue;
1718
1719                 hctx = q->mq_ops->map_queue(q, i);
1720                 cpumask_set_cpu(i, hctx->cpumask);
1721                 ctx->index_hw = hctx->nr_ctx;
1722                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1723         }
1724
1725         queue_for_each_hw_ctx(q, hctx, i) {
1726                 /*
1727                  * If not software queues are mapped to this hardware queue,
1728                  * disable it and free the request entries
1729                  */
1730                 if (!hctx->nr_ctx) {
1731                         struct blk_mq_tag_set *set = q->tag_set;
1732
1733                         if (set->tags[i]) {
1734                                 blk_mq_free_rq_map(set, set->tags[i], i);
1735                                 set->tags[i] = NULL;
1736                                 hctx->tags = NULL;
1737                         }
1738                         continue;
1739                 }
1740
1741                 /*
1742                  * Initialize batch roundrobin counts
1743                  */
1744                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1745                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1746         }
1747 }
1748
1749 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1750 {
1751         struct blk_mq_hw_ctx *hctx;
1752         struct request_queue *q;
1753         bool shared;
1754         int i;
1755
1756         if (set->tag_list.next == set->tag_list.prev)
1757                 shared = false;
1758         else
1759                 shared = true;
1760
1761         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1762                 blk_mq_freeze_queue(q);
1763
1764                 queue_for_each_hw_ctx(q, hctx, i) {
1765                         if (shared)
1766                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1767                         else
1768                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1769                 }
1770                 blk_mq_unfreeze_queue(q);
1771         }
1772 }
1773
1774 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1775 {
1776         struct blk_mq_tag_set *set = q->tag_set;
1777
1778         blk_mq_freeze_queue(q);
1779
1780         mutex_lock(&set->tag_list_lock);
1781         list_del_init(&q->tag_set_list);
1782         blk_mq_update_tag_set_depth(set);
1783         mutex_unlock(&set->tag_list_lock);
1784
1785         blk_mq_unfreeze_queue(q);
1786 }
1787
1788 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1789                                      struct request_queue *q)
1790 {
1791         q->tag_set = set;
1792
1793         mutex_lock(&set->tag_list_lock);
1794         list_add_tail(&q->tag_set_list, &set->tag_list);
1795         blk_mq_update_tag_set_depth(set);
1796         mutex_unlock(&set->tag_list_lock);
1797 }
1798
1799 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1800 {
1801         struct blk_mq_hw_ctx **hctxs;
1802         struct blk_mq_ctx *ctx;
1803         struct request_queue *q;
1804         unsigned int *map;
1805         int i;
1806
1807         ctx = alloc_percpu(struct blk_mq_ctx);
1808         if (!ctx)
1809                 return ERR_PTR(-ENOMEM);
1810
1811         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1812                         set->numa_node);
1813
1814         if (!hctxs)
1815                 goto err_percpu;
1816
1817         map = blk_mq_make_queue_map(set);
1818         if (!map)
1819                 goto err_map;
1820
1821         for (i = 0; i < set->nr_hw_queues; i++) {
1822                 int node = blk_mq_hw_queue_to_node(map, i);
1823
1824                 hctxs[i] = set->ops->alloc_hctx(set, i, node);
1825                 if (!hctxs[i])
1826                         goto err_hctxs;
1827
1828                 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1829                         goto err_hctxs;
1830
1831                 atomic_set(&hctxs[i]->nr_active, 0);
1832                 hctxs[i]->numa_node = node;
1833                 hctxs[i]->queue_num = i;
1834         }
1835
1836         q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1837         if (!q)
1838                 goto err_hctxs;
1839
1840         if (percpu_counter_init(&q->mq_usage_counter, 0))
1841                 goto err_map;
1842
1843         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1844         blk_queue_rq_timeout(q, 30000);
1845
1846         q->nr_queues = nr_cpu_ids;
1847         q->nr_hw_queues = set->nr_hw_queues;
1848         q->mq_map = map;
1849
1850         q->queue_ctx = ctx;
1851         q->queue_hw_ctx = hctxs;
1852
1853         q->mq_ops = set->ops;
1854         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1855
1856         q->sg_reserved_size = INT_MAX;
1857
1858         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1859         INIT_LIST_HEAD(&q->requeue_list);
1860         spin_lock_init(&q->requeue_lock);
1861
1862         if (q->nr_hw_queues > 1)
1863                 blk_queue_make_request(q, blk_mq_make_request);
1864         else
1865                 blk_queue_make_request(q, blk_sq_make_request);
1866
1867         blk_queue_rq_timed_out(q, blk_mq_rq_timed_out);
1868         if (set->timeout)
1869                 blk_queue_rq_timeout(q, set->timeout);
1870
1871         /*
1872          * Do this after blk_queue_make_request() overrides it...
1873          */
1874         q->nr_requests = set->queue_depth;
1875
1876         if (set->ops->complete)
1877                 blk_queue_softirq_done(q, set->ops->complete);
1878
1879         blk_mq_init_flush(q);
1880         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1881
1882         q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1883                                 set->cmd_size, cache_line_size()),
1884                                 GFP_KERNEL);
1885         if (!q->flush_rq)
1886                 goto err_hw;
1887
1888         if (blk_mq_init_hw_queues(q, set))
1889                 goto err_flush_rq;
1890
1891         mutex_lock(&all_q_mutex);
1892         list_add_tail(&q->all_q_node, &all_q_list);
1893         mutex_unlock(&all_q_mutex);
1894
1895         blk_mq_add_queue_tag_set(set, q);
1896
1897         blk_mq_map_swqueue(q);
1898
1899         return q;
1900
1901 err_flush_rq:
1902         kfree(q->flush_rq);
1903 err_hw:
1904         blk_cleanup_queue(q);
1905 err_hctxs:
1906         kfree(map);
1907         for (i = 0; i < set->nr_hw_queues; i++) {
1908                 if (!hctxs[i])
1909                         break;
1910                 free_cpumask_var(hctxs[i]->cpumask);
1911                 set->ops->free_hctx(hctxs[i], i);
1912         }
1913 err_map:
1914         kfree(hctxs);
1915 err_percpu:
1916         free_percpu(ctx);
1917         return ERR_PTR(-ENOMEM);
1918 }
1919 EXPORT_SYMBOL(blk_mq_init_queue);
1920
1921 void blk_mq_free_queue(struct request_queue *q)
1922 {
1923         struct blk_mq_tag_set   *set = q->tag_set;
1924
1925         blk_mq_del_queue_tag_set(q);
1926
1927         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1928         blk_mq_free_hw_queues(q, set);
1929
1930         percpu_counter_destroy(&q->mq_usage_counter);
1931
1932         free_percpu(q->queue_ctx);
1933         kfree(q->queue_hw_ctx);
1934         kfree(q->mq_map);
1935
1936         q->queue_ctx = NULL;
1937         q->queue_hw_ctx = NULL;
1938         q->mq_map = NULL;
1939
1940         mutex_lock(&all_q_mutex);
1941         list_del_init(&q->all_q_node);
1942         mutex_unlock(&all_q_mutex);
1943 }
1944
1945 /* Basically redo blk_mq_init_queue with queue frozen */
1946 static void blk_mq_queue_reinit(struct request_queue *q)
1947 {
1948         blk_mq_freeze_queue(q);
1949
1950         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1951
1952         /*
1953          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1954          * we should change hctx numa_node according to new topology (this
1955          * involves free and re-allocate memory, worthy doing?)
1956          */
1957
1958         blk_mq_map_swqueue(q);
1959
1960         blk_mq_unfreeze_queue(q);
1961 }
1962
1963 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1964                                       unsigned long action, void *hcpu)
1965 {
1966         struct request_queue *q;
1967
1968         /*
1969          * Before new mappings are established, hotadded cpu might already
1970          * start handling requests. This doesn't break anything as we map
1971          * offline CPUs to first hardware queue. We will re-init the queue
1972          * below to get optimal settings.
1973          */
1974         if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1975             action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1976                 return NOTIFY_OK;
1977
1978         mutex_lock(&all_q_mutex);
1979         list_for_each_entry(q, &all_q_list, all_q_node)
1980                 blk_mq_queue_reinit(q);
1981         mutex_unlock(&all_q_mutex);
1982         return NOTIFY_OK;
1983 }
1984
1985 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1986 {
1987         int i;
1988
1989         if (!set->nr_hw_queues)
1990                 return -EINVAL;
1991         if (!set->queue_depth || set->queue_depth > BLK_MQ_MAX_DEPTH)
1992                 return -EINVAL;
1993         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1994                 return -EINVAL;
1995
1996         if (!set->nr_hw_queues ||
1997             !set->ops->queue_rq || !set->ops->map_queue ||
1998             !set->ops->alloc_hctx || !set->ops->free_hctx)
1999                 return -EINVAL;
2000
2001
2002         set->tags = kmalloc_node(set->nr_hw_queues *
2003                                  sizeof(struct blk_mq_tags *),
2004                                  GFP_KERNEL, set->numa_node);
2005         if (!set->tags)
2006                 goto out;
2007
2008         for (i = 0; i < set->nr_hw_queues; i++) {
2009                 set->tags[i] = blk_mq_init_rq_map(set, i);
2010                 if (!set->tags[i])
2011                         goto out_unwind;
2012         }
2013
2014         mutex_init(&set->tag_list_lock);
2015         INIT_LIST_HEAD(&set->tag_list);
2016
2017         return 0;
2018
2019 out_unwind:
2020         while (--i >= 0)
2021                 blk_mq_free_rq_map(set, set->tags[i], i);
2022 out:
2023         return -ENOMEM;
2024 }
2025 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2026
2027 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2028 {
2029         int i;
2030
2031         for (i = 0; i < set->nr_hw_queues; i++) {
2032                 if (set->tags[i])
2033                         blk_mq_free_rq_map(set, set->tags[i], i);
2034         }
2035
2036         kfree(set->tags);
2037 }
2038 EXPORT_SYMBOL(blk_mq_free_tag_set);
2039
2040 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2041 {
2042         struct blk_mq_tag_set *set = q->tag_set;
2043         struct blk_mq_hw_ctx *hctx;
2044         int i, ret;
2045
2046         if (!set || nr > set->queue_depth)
2047                 return -EINVAL;
2048
2049         ret = 0;
2050         queue_for_each_hw_ctx(q, hctx, i) {
2051                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2052                 if (ret)
2053                         break;
2054         }
2055
2056         if (!ret)
2057                 q->nr_requests = nr;
2058
2059         return ret;
2060 }
2061
2062 void blk_mq_disable_hotplug(void)
2063 {
2064         mutex_lock(&all_q_mutex);
2065 }
2066
2067 void blk_mq_enable_hotplug(void)
2068 {
2069         mutex_unlock(&all_q_mutex);
2070 }
2071
2072 static int __init blk_mq_init(void)
2073 {
2074         blk_mq_cpu_init();
2075
2076         /* Must be called after percpu_counter_hotcpu_callback() */
2077         hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
2078
2079         return 0;
2080 }
2081 subsys_initcall(blk_mq_init);