block/blk-mq.c

   1 /*
   2  * Block multiqueue core code
   3  *
   4  * Copyright (C) 2013-2014 Jens Axboe
   5  * Copyright (C) 2013-2014 Christoph Hellwig
   6  */
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/mm.h>
  13 #include <linux/init.h>
  14 #include <linux/slab.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/smp.h>
  17 #include <linux/llist.h>
  18 #include <linux/list_sort.h>
  19 #include <linux/cpu.h>
  20 #include <linux/cache.h>
  21 #include <linux/sched/sysctl.h>
  22 #include <linux/delay.h>
  23 #include <linux/crash_dump.h>
  24
  25 #include <trace/events/block.h>
  26
  27 #include <linux/blk-mq.h>
  28 #include "blk.h"
  29 #include "blk-mq.h"
  30 #include "blk-mq-tag.h"
  31
  32 static DEFINE_MUTEX(all_q_mutex);
  33 static LIST_HEAD(all_q_list);
  34
  35 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
  36
  37 /*
  38  * Check if any of the ctx's have pending work in this hardware queue
  39  */
  40 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  41 {
  42         unsigned int i;
  43
  44         for (i = 0; i < hctx->ctx_map.size; i++)
  45                 if (hctx->ctx_map.map[i].word)
  46                         return true;
  47
  48         return false;
  49 }
  50
  51 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
  52                                               struct blk_mq_ctx *ctx)
  53 {
  54         return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
  55 }
  56
  57 #define CTX_TO_BIT(hctx, ctx)   \
  58         ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
  59
  60 /*
  61  * Mark this ctx as having pending work in this hardware queue
  62  */
  63 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  64                                      struct blk_mq_ctx *ctx)
  65 {
  66         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  67
  68         if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
  69                 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  70 }
  71
  72 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  73                                       struct blk_mq_ctx *ctx)
  74 {
  75         struct blk_align_bitmap *bm = get_bm(hctx, ctx);
  76
  77         clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
  78 }
  79
  80 static int blk_mq_queue_enter(struct request_queue *q, gfp_t gfp)
  81 {
  82         while (true) {
  83                 int ret;
  84
  85                 if (percpu_ref_tryget_live(&q->mq_usage_counter))
  86                         return 0;
  87
  88                 if (!(gfp & __GFP_WAIT))
  89                         return -EBUSY;
  90
  91                 ret = wait_event_interruptible(q->mq_freeze_wq,
  92                                 !atomic_read(&q->mq_freeze_depth) ||
  93                                 blk_queue_dying(q));
  94                 if (blk_queue_dying(q))
  95                         return -ENODEV;
  96                 if (ret)
  97                         return ret;
  98         }
  99 }
 100
 101 static void blk_mq_queue_exit(struct request_queue *q)
 102 {
 103         percpu_ref_put(&q->mq_usage_counter);
 104 }
 105
 106 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
 107 {
 108         struct request_queue *q =
 109                 container_of(ref, struct request_queue, mq_usage_counter);
 110
 111         wake_up_all(&q->mq_freeze_wq);
 112 }
 113
 114 void blk_mq_freeze_queue_start(struct request_queue *q)
 115 {
 116         int freeze_depth;
 117
 118         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 119         if (freeze_depth == 1) {
 120                 percpu_ref_kill(&q->mq_usage_counter);
 121                 blk_mq_run_hw_queues(q, false);
 122         }
 123 }
 124 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_start);
 125
 126 static void blk_mq_freeze_queue_wait(struct request_queue *q)
 127 {
 128         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
 129 }
 130
 131 /*
 132  * Guarantee no request is in use, so we can change any data structure of
 133  * the queue afterward.
 134  */
 135 void blk_mq_freeze_queue(struct request_queue *q)
 136 {
 137         blk_mq_freeze_queue_start(q);
 138         blk_mq_freeze_queue_wait(q);
 139 }
 140 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 141
 142 void blk_mq_unfreeze_queue(struct request_queue *q)
 143 {
 144         int freeze_depth;
 145
 146         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 147         WARN_ON_ONCE(freeze_depth < 0);
 148         if (!freeze_depth) {
 149                 percpu_ref_reinit(&q->mq_usage_counter);
 150                 wake_up_all(&q->mq_freeze_wq);
 151         }
 152 }
 153 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 154
 155 void blk_mq_wake_waiters(struct request_queue *q)
 156 {
 157         struct blk_mq_hw_ctx *hctx;
 158         unsigned int i;
 159
 160         queue_for_each_hw_ctx(q, hctx, i)
 161                 if (blk_mq_hw_queue_mapped(hctx))
 162                         blk_mq_tag_wakeup_all(hctx->tags, true);
 163
 164         /*
 165          * If we are called because the queue has now been marked as
 166          * dying, we need to ensure that processes currently waiting on
 167          * the queue are notified as well.
 168          */
 169         wake_up_all(&q->mq_freeze_wq);
 170 }
 171
 172 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 173 {
 174         return blk_mq_has_free_tags(hctx->tags);
 175 }
 176 EXPORT_SYMBOL(blk_mq_can_queue);
 177
 178 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 179                                struct request *rq, unsigned int rw_flags)
 180 {
 181         if (blk_queue_io_stat(q))
 182                 rw_flags |= REQ_IO_STAT;
 183
 184         INIT_LIST_HEAD(&rq->queuelist);
 185         /* csd/requeue_work/fifo_time is initialized before use */
 186         rq->q = q;
 187         rq->mq_ctx = ctx;
 188         rq->cmd_flags |= rw_flags;
 189         /* do not touch atomic flags, it needs atomic ops against the timer */
 190         rq->cpu = -1;
 191         INIT_HLIST_NODE(&rq->hash);
 192         RB_CLEAR_NODE(&rq->rb_node);
 193         rq->rq_disk = NULL;
 194         rq->part = NULL;
 195         rq->start_time = jiffies;
 196 #ifdef CONFIG_BLK_CGROUP
 197         rq->rl = NULL;
 198         set_start_time_ns(rq);
 199         rq->io_start_time_ns = 0;
 200 #endif
 201         rq->nr_phys_segments = 0;
 202 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 203         rq->nr_integrity_segments = 0;
 204 #endif
 205         rq->special = NULL;
 206         /* tag was already set */
 207         rq->errors = 0;
 208
 209         rq->cmd = rq->__cmd;
 210
 211         rq->extra_len = 0;
 212         rq->sense_len = 0;
 213         rq->resid_len = 0;
 214         rq->sense = NULL;
 215
 216         INIT_LIST_HEAD(&rq->timeout_list);
 217         rq->timeout = 0;
 218
 219         rq->end_io = NULL;
 220         rq->end_io_data = NULL;
 221         rq->next_rq = NULL;
 222
 223         ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
 224 }
 225
 226 static struct request *
 227 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
 228 {
 229         struct request *rq;
 230         unsigned int tag;
 231
 232         tag = blk_mq_get_tag(data);
 233         if (tag != BLK_MQ_TAG_FAIL) {
 234                 rq = data->hctx->tags->rqs[tag];
 235
 236                 if (blk_mq_tag_busy(data->hctx)) {
 237                         rq->cmd_flags = REQ_MQ_INFLIGHT;
 238                         atomic_inc(&data->hctx->nr_active);
 239                 }
 240
 241                 rq->tag = tag;
 242                 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
 243                 return rq;
 244         }
 245
 246         return NULL;
 247 }
 248
 249 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
 250                 bool reserved)
 251 {
 252         struct blk_mq_ctx *ctx;
 253         struct blk_mq_hw_ctx *hctx;
 254         struct request *rq;
 255         struct blk_mq_alloc_data alloc_data;
 256         int ret;
 257
 258         ret = blk_mq_queue_enter(q, gfp);
 259         if (ret)
 260                 return ERR_PTR(ret);
 261
 262         ctx = blk_mq_get_ctx(q);
 263         hctx = q->mq_ops->map_queue(q, ctx->cpu);
 264         blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
 265                         reserved, ctx, hctx);
 266
 267         rq = __blk_mq_alloc_request(&alloc_data, rw);
 268         if (!rq && (gfp & __GFP_WAIT)) {
 269                 __blk_mq_run_hw_queue(hctx);
 270                 blk_mq_put_ctx(ctx);
 271
 272                 ctx = blk_mq_get_ctx(q);
 273                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
 274                 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
 275                                 hctx);
 276                 rq =  __blk_mq_alloc_request(&alloc_data, rw);
 277                 ctx = alloc_data.ctx;
 278         }
 279         blk_mq_put_ctx(ctx);
 280         if (!rq) {
 281                 blk_mq_queue_exit(q);
 282                 return ERR_PTR(-EWOULDBLOCK);
 283         }
 284         return rq;
 285 }
 286 EXPORT_SYMBOL(blk_mq_alloc_request);
 287
 288 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 289                                   struct blk_mq_ctx *ctx, struct request *rq)
 290 {
 291         const int tag = rq->tag;
 292         struct request_queue *q = rq->q;
 293
 294         if (rq->cmd_flags & REQ_MQ_INFLIGHT)
 295                 atomic_dec(&hctx->nr_active);
 296         rq->cmd_flags = 0;
 297
 298         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 299         blk_mq_put_tag(hctx, tag, &ctx->last_tag);
 300         blk_mq_queue_exit(q);
 301 }
 302
 303 void blk_mq_free_hctx_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
 304 {
 305         struct blk_mq_ctx *ctx = rq->mq_ctx;
 306
 307         ctx->rq_completed[rq_is_sync(rq)]++;
 308         __blk_mq_free_request(hctx, ctx, rq);
 309
 310 }
 311 EXPORT_SYMBOL_GPL(blk_mq_free_hctx_request);
 312
 313 void blk_mq_free_request(struct request *rq)
 314 {
 315         struct blk_mq_hw_ctx *hctx;
 316         struct request_queue *q = rq->q;
 317
 318         hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
 319         blk_mq_free_hctx_request(hctx, rq);
 320 }
 321 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 322
 323 inline void __blk_mq_end_request(struct request *rq, int error)
 324 {
 325         blk_account_io_done(rq);
 326
 327         if (rq->end_io) {
 328                 rq->end_io(rq, error);
 329         } else {
 330                 if (unlikely(blk_bidi_rq(rq)))
 331                         blk_mq_free_request(rq->next_rq);
 332                 blk_mq_free_request(rq);
 333         }
 334 }
 335 EXPORT_SYMBOL(__blk_mq_end_request);
 336
 337 void blk_mq_end_request(struct request *rq, int error)
 338 {
 339         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 340                 BUG();
 341         __blk_mq_end_request(rq, error);
 342 }
 343 EXPORT_SYMBOL(blk_mq_end_request);
 344
 345 static void __blk_mq_complete_request_remote(void *data)
 346 {
 347         struct request *rq = data;
 348
 349         rq->q->softirq_done_fn(rq);
 350 }
 351
 352 static void blk_mq_ipi_complete_request(struct request *rq)
 353 {
 354         struct blk_mq_ctx *ctx = rq->mq_ctx;
 355         bool shared = false;
 356         int cpu;
 357
 358         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 359                 rq->q->softirq_done_fn(rq);
 360                 return;
 361         }
 362
 363         cpu = get_cpu();
 364         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 365                 shared = cpus_share_cache(cpu, ctx->cpu);
 366
 367         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 368                 rq->csd.func = __blk_mq_complete_request_remote;
 369                 rq->csd.info = rq;
 370                 rq->csd.flags = 0;
 371                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 372         } else {
 373                 rq->q->softirq_done_fn(rq);
 374         }
 375         put_cpu();
 376 }
 377
 378 void __blk_mq_complete_request(struct request *rq)
 379 {
 380         struct request_queue *q = rq->q;
 381
 382         if (!q->softirq_done_fn)
 383                 blk_mq_end_request(rq, rq->errors);
 384         else
 385                 blk_mq_ipi_complete_request(rq);
 386 }
 387
 388 /**
 389  * blk_mq_complete_request - end I/O on a request
 390  * @rq:         the request being processed
 391  *
 392  * Description:
 393  *      Ends all I/O on a request. It does not handle partial completions.
 394  *      The actual completion happens out-of-order, through a IPI handler.
 395  **/
 396 void blk_mq_complete_request(struct request *rq, int error)
 397 {
 398         struct request_queue *q = rq->q;
 399
 400         if (unlikely(blk_should_fake_timeout(q)))
 401                 return;
 402         if (!blk_mark_rq_complete(rq)) {
 403                 rq->errors = error;
 404                 __blk_mq_complete_request(rq);
 405         }
 406 }
 407 EXPORT_SYMBOL(blk_mq_complete_request);
 408
 409 int blk_mq_request_started(struct request *rq)
 410 {
 411         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 412 }
 413 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 414
 415 void blk_mq_start_request(struct request *rq)
 416 {
 417         struct request_queue *q = rq->q;
 418
 419         trace_block_rq_issue(q, rq);
 420
 421         rq->resid_len = blk_rq_bytes(rq);
 422         if (unlikely(blk_bidi_rq(rq)))
 423                 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
 424
 425         blk_add_timer(rq);
 426
 427         /*
 428          * Ensure that ->deadline is visible before set the started
 429          * flag and clear the completed flag.
 430          */
 431         smp_mb__before_atomic();
 432
 433         /*
 434          * Mark us as started and clear complete. Complete might have been
 435          * set if requeue raced with timeout, which then marked it as
 436          * complete. So be sure to clear complete again when we start
 437          * the request, otherwise we'll ignore the completion event.
 438          */
 439         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 440                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 441         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 442                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 443
 444         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 445                 /*
 446                  * Make sure space for the drain appears.  We know we can do
 447                  * this because max_hw_segments has been adjusted to be one
 448                  * fewer than the device can handle.
 449                  */
 450                 rq->nr_phys_segments++;
 451         }
 452 }
 453 EXPORT_SYMBOL(blk_mq_start_request);
 454
 455 static void __blk_mq_requeue_request(struct request *rq)
 456 {
 457         struct request_queue *q = rq->q;
 458
 459         trace_block_rq_requeue(q, rq);
 460
 461         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 462                 if (q->dma_drain_size && blk_rq_bytes(rq))
 463                         rq->nr_phys_segments--;
 464         }
 465 }
 466
 467 void blk_mq_requeue_request(struct request *rq)
 468 {
 469         __blk_mq_requeue_request(rq);
 470
 471         BUG_ON(blk_queued_rq(rq));
 472         blk_mq_add_to_requeue_list(rq, true);
 473 }
 474 EXPORT_SYMBOL(blk_mq_requeue_request);
 475
 476 static void blk_mq_requeue_work(struct work_struct *work)
 477 {
 478         struct request_queue *q =
 479                 container_of(work, struct request_queue, requeue_work);
 480         LIST_HEAD(rq_list);
 481         struct request *rq, *next;
 482         unsigned long flags;
 483
 484         spin_lock_irqsave(&q->requeue_lock, flags);
 485         list_splice_init(&q->requeue_list, &rq_list);
 486         spin_unlock_irqrestore(&q->requeue_lock, flags);
 487
 488         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 489                 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
 490                         continue;
 491
 492                 rq->cmd_flags &= ~REQ_SOFTBARRIER;
 493                 list_del_init(&rq->queuelist);
 494                 blk_mq_insert_request(rq, true, false, false);
 495         }
 496
 497         while (!list_empty(&rq_list)) {
 498                 rq = list_entry(rq_list.next, struct request, queuelist);
 499                 list_del_init(&rq->queuelist);
 500                 blk_mq_insert_request(rq, false, false, false);
 501         }
 502
 503         /*
 504          * Use the start variant of queue running here, so that running
 505          * the requeue work will kick stopped queues.
 506          */
 507         blk_mq_start_hw_queues(q);
 508 }
 509
 510 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
 511 {
 512         struct request_queue *q = rq->q;
 513         unsigned long flags;
 514
 515         /*
 516          * We abuse this flag that is otherwise used by the I/O scheduler to
 517          * request head insertation from the workqueue.
 518          */
 519         BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
 520
 521         spin_lock_irqsave(&q->requeue_lock, flags);
 522         if (at_head) {
 523                 rq->cmd_flags |= REQ_SOFTBARRIER;
 524                 list_add(&rq->queuelist, &q->requeue_list);
 525         } else {
 526                 list_add_tail(&rq->queuelist, &q->requeue_list);
 527         }
 528         spin_unlock_irqrestore(&q->requeue_lock, flags);
 529 }
 530 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 531
 532 void blk_mq_cancel_requeue_work(struct request_queue *q)
 533 {
 534         cancel_work_sync(&q->requeue_work);
 535 }
 536 EXPORT_SYMBOL_GPL(blk_mq_cancel_requeue_work);
 537
 538 void blk_mq_kick_requeue_list(struct request_queue *q)
 539 {
 540         kblockd_schedule_work(&q->requeue_work);
 541 }
 542 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 543
 544 void blk_mq_abort_requeue_list(struct request_queue *q)
 545 {
 546         unsigned long flags;
 547         LIST_HEAD(rq_list);
 548
 549         spin_lock_irqsave(&q->requeue_lock, flags);
 550         list_splice_init(&q->requeue_list, &rq_list);
 551         spin_unlock_irqrestore(&q->requeue_lock, flags);
 552
 553         while (!list_empty(&rq_list)) {
 554                 struct request *rq;
 555
 556                 rq = list_first_entry(&rq_list, struct request, queuelist);
 557                 list_del_init(&rq->queuelist);
 558                 rq->errors = -EIO;
 559                 blk_mq_end_request(rq, rq->errors);
 560         }
 561 }
 562 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 563
 564 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 565 {
 566         return tags->rqs[tag];
 567 }
 568 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 569
 570 struct blk_mq_timeout_data {
 571         unsigned long next;
 572         unsigned int next_set;
 573 };
 574
 575 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 576 {
 577         struct blk_mq_ops *ops = req->q->mq_ops;
 578         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 579
 580         /*
 581          * We know that complete is set at this point. If STARTED isn't set
 582          * anymore, then the request isn't active and the "timeout" should
 583          * just be ignored. This can happen due to the bitflag ordering.
 584          * Timeout first checks if STARTED is set, and if it is, assumes
 585          * the request is active. But if we race with completion, then
 586          * we both flags will get cleared. So check here again, and ignore
 587          * a timeout event with a request that isn't active.
 588          */
 589         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 590                 return;
 591
 592         if (ops->timeout)
 593                 ret = ops->timeout(req, reserved);
 594
 595         switch (ret) {
 596         case BLK_EH_HANDLED:
 597                 __blk_mq_complete_request(req);
 598                 break;
 599         case BLK_EH_RESET_TIMER:
 600                 blk_add_timer(req);
 601                 blk_clear_rq_complete(req);
 602                 break;
 603         case BLK_EH_NOT_HANDLED:
 604                 break;
 605         default:
 606                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 607                 break;
 608         }
 609 }
 610
 611 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 612                 struct request *rq, void *priv, bool reserved)
 613 {
 614         struct blk_mq_timeout_data *data = priv;
 615
 616         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 617                 /*
 618                  * If a request wasn't started before the queue was
 619                  * marked dying, kill it here or it'll go unnoticed.
 620                  */
 621                 if (unlikely(blk_queue_dying(rq->q)))
 622                         blk_mq_complete_request(rq, -EIO);
 623                 return;
 624         }
 625         if (rq->cmd_flags & REQ_NO_TIMEOUT)
 626                 return;
 627
 628         if (time_after_eq(jiffies, rq->deadline)) {
 629                 if (!blk_mark_rq_complete(rq))
 630                         blk_mq_rq_timed_out(rq, reserved);
 631         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 632                 data->next = rq->deadline;
 633                 data->next_set = 1;
 634         }
 635 }
 636
 637 static void blk_mq_rq_timer(unsigned long priv)
 638 {
 639         struct request_queue *q = (struct request_queue *)priv;
 640         struct blk_mq_timeout_data data = {
 641                 .next           = 0,
 642                 .next_set       = 0,
 643         };
 644         struct blk_mq_hw_ctx *hctx;
 645         int i;
 646
 647         queue_for_each_hw_ctx(q, hctx, i) {
 648                 /*
 649                  * If not software queues are currently mapped to this
 650                  * hardware queue, there's nothing to check
 651                  */
 652                 if (!blk_mq_hw_queue_mapped(hctx))
 653                         continue;
 654
 655                 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
 656         }
 657
 658         if (data.next_set) {
 659                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 660                 mod_timer(&q->timeout, data.next);
 661         } else {
 662                 queue_for_each_hw_ctx(q, hctx, i) {
 663                         /* the hctx may be unmapped, so check it here */
 664                         if (blk_mq_hw_queue_mapped(hctx))
 665                                 blk_mq_tag_idle(hctx);
 666                 }
 667         }
 668 }
 669
 670 /*
 671  * Reverse check our software queue for entries that we could potentially
 672  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 673  * too much time checking for merges.
 674  */
 675 static bool blk_mq_attempt_merge(struct request_queue *q,
 676                                  struct blk_mq_ctx *ctx, struct bio *bio)
 677 {
 678         struct request *rq;
 679         int checked = 8;
 680
 681         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 682                 int el_ret;
 683
 684                 if (!checked--)
 685                         break;
 686
 687                 if (!blk_rq_merge_ok(rq, bio))
 688                         continue;
 689
 690                 el_ret = blk_try_merge(rq, bio);
 691                 if (el_ret == ELEVATOR_BACK_MERGE) {
 692                         if (bio_attempt_back_merge(q, rq, bio)) {
 693                                 ctx->rq_merged++;
 694                                 return true;
 695                         }
 696                         break;
 697                 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
 698                         if (bio_attempt_front_merge(q, rq, bio)) {
 699                                 ctx->rq_merged++;
 700                                 return true;
 701                         }
 702                         break;
 703                 }
 704         }
 705
 706         return false;
 707 }
 708
 709 /*
 710  * Process software queues that have been marked busy, splicing them
 711  * to the for-dispatch
 712  */
 713 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 714 {
 715         struct blk_mq_ctx *ctx;
 716         int i;
 717
 718         for (i = 0; i < hctx->ctx_map.size; i++) {
 719                 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
 720                 unsigned int off, bit;
 721
 722                 if (!bm->word)
 723                         continue;
 724
 725                 bit = 0;
 726                 off = i * hctx->ctx_map.bits_per_word;
 727                 do {
 728                         bit = find_next_bit(&bm->word, bm->depth, bit);
 729                         if (bit >= bm->depth)
 730                                 break;
 731
 732                         ctx = hctx->ctxs[bit + off];
 733                         clear_bit(bit, &bm->word);
 734                         spin_lock(&ctx->lock);
 735                         list_splice_tail_init(&ctx->rq_list, list);
 736                         spin_unlock(&ctx->lock);
 737
 738                         bit++;
 739                 } while (1);
 740         }
 741 }
 742
 743 /*
 744  * Run this hardware queue, pulling any software queues mapped to it in.
 745  * Note that this function currently has various problems around ordering
 746  * of IO. In particular, we'd like FIFO behaviour on handling existing
 747  * items on the hctx->dispatch list. Ignore that for now.
 748  */
 749 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 750 {
 751         struct request_queue *q = hctx->queue;
 752         struct request *rq;
 753         LIST_HEAD(rq_list);
 754         LIST_HEAD(driver_list);
 755         struct list_head *dptr;
 756         int queued;
 757
 758         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
 759
 760         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 761                 return;
 762
 763         hctx->run++;
 764
 765         /*
 766          * Touch any software queue that has pending entries.
 767          */
 768         flush_busy_ctxs(hctx, &rq_list);
 769
 770         /*
 771          * If we have previous entries on our dispatch list, grab them
 772          * and stuff them at the front for more fair dispatch.
 773          */
 774         if (!list_empty_careful(&hctx->dispatch)) {
 775                 spin_lock(&hctx->lock);
 776                 if (!list_empty(&hctx->dispatch))
 777                         list_splice_init(&hctx->dispatch, &rq_list);
 778                 spin_unlock(&hctx->lock);
 779         }
 780
 781         /*
 782          * Start off with dptr being NULL, so we start the first request
 783          * immediately, even if we have more pending.
 784          */
 785         dptr = NULL;
 786
 787         /*
 788          * Now process all the entries, sending them to the driver.
 789          */
 790         queued = 0;
 791         while (!list_empty(&rq_list)) {
 792                 struct blk_mq_queue_data bd;
 793                 int ret;
 794
 795                 rq = list_first_entry(&rq_list, struct request, queuelist);
 796                 list_del_init(&rq->queuelist);
 797
 798                 bd.rq = rq;
 799                 bd.list = dptr;
 800                 bd.last = list_empty(&rq_list);
 801
 802                 ret = q->mq_ops->queue_rq(hctx, &bd);
 803                 switch (ret) {
 804                 case BLK_MQ_RQ_QUEUE_OK:
 805                         queued++;
 806                         continue;
 807                 case BLK_MQ_RQ_QUEUE_BUSY:
 808                         list_add(&rq->queuelist, &rq_list);
 809                         __blk_mq_requeue_request(rq);
 810                         break;
 811                 default:
 812                         pr_err("blk-mq: bad return on queue: %d\n", ret);
 813                 case BLK_MQ_RQ_QUEUE_ERROR:
 814                         rq->errors = -EIO;
 815                         blk_mq_end_request(rq, rq->errors);
 816                         break;
 817                 }
 818
 819                 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
 820                         break;
 821
 822                 /*
 823                  * We've done the first request. If we have more than 1
 824                  * left in the list, set dptr to defer issue.
 825                  */
 826                 if (!dptr && rq_list.next != rq_list.prev)
 827                         dptr = &driver_list;
 828         }
 829
 830         if (!queued)
 831                 hctx->dispatched[0]++;
 832         else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
 833                 hctx->dispatched[ilog2(queued) + 1]++;
 834
 835         /*
 836          * Any items that need requeuing? Stuff them into hctx->dispatch,
 837          * that is where we will continue on next queue run.
 838          */
 839         if (!list_empty(&rq_list)) {
 840                 spin_lock(&hctx->lock);
 841                 list_splice(&rq_list, &hctx->dispatch);
 842                 spin_unlock(&hctx->lock);
 843                 /*
 844                  * the queue is expected stopped with BLK_MQ_RQ_QUEUE_BUSY, but
 845                  * it's possible the queue is stopped and restarted again
 846                  * before this. Queue restart will dispatch requests. And since
 847                  * requests in rq_list aren't added into hctx->dispatch yet,
 848                  * the requests in rq_list might get lost.
 849                  *
 850                  * blk_mq_run_hw_queue() already checks the STOPPED bit
 851                  **/
 852                 blk_mq_run_hw_queue(hctx, true);
 853         }
 854 }
 855
 856 /*
 857  * It'd be great if the workqueue API had a way to pass
 858  * in a mask and had some smarts for more clever placement.
 859  * For now we just round-robin here, switching for every
 860  * BLK_MQ_CPU_WORK_BATCH queued items.
 861  */
 862 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
 863 {
 864         if (hctx->queue->nr_hw_queues == 1)
 865                 return WORK_CPU_UNBOUND;
 866
 867         if (--hctx->next_cpu_batch <= 0) {
 868                 int cpu = hctx->next_cpu, next_cpu;
 869
 870                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
 871                 if (next_cpu >= nr_cpu_ids)
 872                         next_cpu = cpumask_first(hctx->cpumask);
 873
 874                 hctx->next_cpu = next_cpu;
 875                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
 876
 877                 return cpu;
 878         }
 879
 880         return hctx->next_cpu;
 881 }
 882
 883 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 884 {
 885         if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state) ||
 886             !blk_mq_hw_queue_mapped(hctx)))
 887                 return;
 888
 889         if (!async) {
 890                 int cpu = get_cpu();
 891                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
 892                         __blk_mq_run_hw_queue(hctx);
 893                         put_cpu();
 894                         return;
 895                 }
 896
 897                 put_cpu();
 898         }
 899
 900         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 901                         &hctx->run_work, 0);
 902 }
 903
 904 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
 905 {
 906         struct blk_mq_hw_ctx *hctx;
 907         int i;
 908
 909         queue_for_each_hw_ctx(q, hctx, i) {
 910                 if ((!blk_mq_hctx_has_pending(hctx) &&
 911                     list_empty_careful(&hctx->dispatch)) ||
 912                     test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 913                         continue;
 914
 915                 blk_mq_run_hw_queue(hctx, async);
 916         }
 917 }
 918 EXPORT_SYMBOL(blk_mq_run_hw_queues);
 919
 920 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
 921 {
 922         cancel_delayed_work(&hctx->run_work);
 923         cancel_delayed_work(&hctx->delay_work);
 924         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
 925 }
 926 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
 927
 928 void blk_mq_stop_hw_queues(struct request_queue *q)
 929 {
 930         struct blk_mq_hw_ctx *hctx;
 931         int i;
 932
 933         queue_for_each_hw_ctx(q, hctx, i)
 934                 blk_mq_stop_hw_queue(hctx);
 935 }
 936 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
 937
 938 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
 939 {
 940         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 941
 942         blk_mq_run_hw_queue(hctx, false);
 943 }
 944 EXPORT_SYMBOL(blk_mq_start_hw_queue);
 945
 946 void blk_mq_start_hw_queues(struct request_queue *q)
 947 {
 948         struct blk_mq_hw_ctx *hctx;
 949         int i;
 950
 951         queue_for_each_hw_ctx(q, hctx, i)
 952                 blk_mq_start_hw_queue(hctx);
 953 }
 954 EXPORT_SYMBOL(blk_mq_start_hw_queues);
 955
 956 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
 957 {
 958         struct blk_mq_hw_ctx *hctx;
 959         int i;
 960
 961         queue_for_each_hw_ctx(q, hctx, i) {
 962                 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
 963                         continue;
 964
 965                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
 966                 blk_mq_run_hw_queue(hctx, async);
 967         }
 968 }
 969 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
 970
 971 static void blk_mq_run_work_fn(struct work_struct *work)
 972 {
 973         struct blk_mq_hw_ctx *hctx;
 974
 975         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
 976
 977         __blk_mq_run_hw_queue(hctx);
 978 }
 979
 980 static void blk_mq_delay_work_fn(struct work_struct *work)
 981 {
 982         struct blk_mq_hw_ctx *hctx;
 983
 984         hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
 985
 986         if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
 987                 __blk_mq_run_hw_queue(hctx);
 988 }
 989
 990 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 991 {
 992         if (unlikely(!blk_mq_hw_queue_mapped(hctx)))
 993                 return;
 994
 995         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
 996                         &hctx->delay_work, msecs_to_jiffies(msecs));
 997 }
 998 EXPORT_SYMBOL(blk_mq_delay_queue);
 999
1000 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
1001                                     struct request *rq, bool at_head)
1002 {
1003         struct blk_mq_ctx *ctx = rq->mq_ctx;
1004
1005         trace_block_rq_insert(hctx->queue, rq);
1006
1007         if (at_head)
1008                 list_add(&rq->queuelist, &ctx->rq_list);
1009         else
1010                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1011
1012         blk_mq_hctx_mark_pending(hctx, ctx);
1013 }
1014
1015 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
1016                 bool async)
1017 {
1018         struct request_queue *q = rq->q;
1019         struct blk_mq_hw_ctx *hctx;
1020         struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
1021
1022         current_ctx = blk_mq_get_ctx(q);
1023         if (!cpu_online(ctx->cpu))
1024                 rq->mq_ctx = ctx = current_ctx;
1025
1026         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1027
1028         spin_lock(&ctx->lock);
1029         __blk_mq_insert_request(hctx, rq, at_head);
1030         spin_unlock(&ctx->lock);
1031
1032         if (run_queue)
1033                 blk_mq_run_hw_queue(hctx, async);
1034
1035         blk_mq_put_ctx(current_ctx);
1036 }
1037
1038 static void blk_mq_insert_requests(struct request_queue *q,
1039                                      struct blk_mq_ctx *ctx,
1040                                      struct list_head *list,
1041                                      int depth,
1042                                      bool from_schedule)
1043
1044 {
1045         struct blk_mq_hw_ctx *hctx;
1046         struct blk_mq_ctx *current_ctx;
1047
1048         trace_block_unplug(q, depth, !from_schedule);
1049
1050         current_ctx = blk_mq_get_ctx(q);
1051
1052         if (!cpu_online(ctx->cpu))
1053                 ctx = current_ctx;
1054         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1055
1056         /*
1057          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1058          * offline now
1059          */
1060         spin_lock(&ctx->lock);
1061         while (!list_empty(list)) {
1062                 struct request *rq;
1063
1064                 rq = list_first_entry(list, struct request, queuelist);
1065                 list_del_init(&rq->queuelist);
1066                 rq->mq_ctx = ctx;
1067                 __blk_mq_insert_request(hctx, rq, false);
1068         }
1069         spin_unlock(&ctx->lock);
1070
1071         blk_mq_run_hw_queue(hctx, from_schedule);
1072         blk_mq_put_ctx(current_ctx);
1073 }
1074
1075 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1076 {
1077         struct request *rqa = container_of(a, struct request, queuelist);
1078         struct request *rqb = container_of(b, struct request, queuelist);
1079
1080         return !(rqa->mq_ctx < rqb->mq_ctx ||
1081                  (rqa->mq_ctx == rqb->mq_ctx &&
1082                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1083 }
1084
1085 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1086 {
1087         struct blk_mq_ctx *this_ctx;
1088         struct request_queue *this_q;
1089         struct request *rq;
1090         LIST_HEAD(list);
1091         LIST_HEAD(ctx_list);
1092         unsigned int depth;
1093
1094         list_splice_init(&plug->mq_list, &list);
1095
1096         list_sort(NULL, &list, plug_ctx_cmp);
1097
1098         this_q = NULL;
1099         this_ctx = NULL;
1100         depth = 0;
1101
1102         while (!list_empty(&list)) {
1103                 rq = list_entry_rq(list.next);
1104                 list_del_init(&rq->queuelist);
1105                 BUG_ON(!rq->q);
1106                 if (rq->mq_ctx != this_ctx) {
1107                         if (this_ctx) {
1108                                 blk_mq_insert_requests(this_q, this_ctx,
1109                                                         &ctx_list, depth,
1110                                                         from_schedule);
1111                         }
1112
1113                         this_ctx = rq->mq_ctx;
1114                         this_q = rq->q;
1115                         depth = 0;
1116                 }
1117
1118                 depth++;
1119                 list_add_tail(&rq->queuelist, &ctx_list);
1120         }
1121
1122         /*
1123          * If 'this_ctx' is set, we know we have entries to complete
1124          * on 'ctx_list'. Do those.
1125          */
1126         if (this_ctx) {
1127                 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1128                                        from_schedule);
1129         }
1130 }
1131
1132 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1133 {
1134         init_request_from_bio(rq, bio);
1135
1136         if (blk_do_io_stat(rq))
1137                 blk_account_io_start(rq, 1);
1138 }
1139
1140 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1141 {
1142         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1143                 !blk_queue_nomerges(hctx->queue);
1144 }
1145
1146 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1147                                          struct blk_mq_ctx *ctx,
1148                                          struct request *rq, struct bio *bio)
1149 {
1150         if (!hctx_allow_merges(hctx)) {
1151                 blk_mq_bio_to_request(rq, bio);
1152                 spin_lock(&ctx->lock);
1153 insert_rq:
1154                 __blk_mq_insert_request(hctx, rq, false);
1155                 spin_unlock(&ctx->lock);
1156                 return false;
1157         } else {
1158                 struct request_queue *q = hctx->queue;
1159
1160                 spin_lock(&ctx->lock);
1161                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1162                         blk_mq_bio_to_request(rq, bio);
1163                         goto insert_rq;
1164                 }
1165
1166                 spin_unlock(&ctx->lock);
1167                 __blk_mq_free_request(hctx, ctx, rq);
1168                 return true;
1169         }
1170 }
1171
1172 struct blk_map_ctx {
1173         struct blk_mq_hw_ctx *hctx;
1174         struct blk_mq_ctx *ctx;
1175 };
1176
1177 static struct request *blk_mq_map_request(struct request_queue *q,
1178                                           struct bio *bio,
1179                                           struct blk_map_ctx *data)
1180 {
1181         struct blk_mq_hw_ctx *hctx;
1182         struct blk_mq_ctx *ctx;
1183         struct request *rq;
1184         int rw = bio_data_dir(bio);
1185         struct blk_mq_alloc_data alloc_data;
1186
1187         if (unlikely(blk_mq_queue_enter(q, GFP_KERNEL))) {
1188                 bio_io_error(bio);
1189                 return NULL;
1190         }
1191
1192         ctx = blk_mq_get_ctx(q);
1193         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1194
1195         if (rw_is_sync(bio->bi_rw))
1196                 rw |= REQ_SYNC;
1197
1198         trace_block_getrq(q, bio, rw);
1199         blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1200                         hctx);
1201         rq = __blk_mq_alloc_request(&alloc_data, rw);
1202         if (unlikely(!rq)) {
1203                 __blk_mq_run_hw_queue(hctx);
1204                 blk_mq_put_ctx(ctx);
1205                 trace_block_sleeprq(q, bio, rw);
1206
1207                 ctx = blk_mq_get_ctx(q);
1208                 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1209                 blk_mq_set_alloc_data(&alloc_data, q,
1210                                 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1211                 rq = __blk_mq_alloc_request(&alloc_data, rw);
1212                 ctx = alloc_data.ctx;
1213                 hctx = alloc_data.hctx;
1214         }
1215
1216         hctx->queued++;
1217         data->hctx = hctx;
1218         data->ctx = ctx;
1219         return rq;
1220 }
1221
1222 static int blk_mq_direct_issue_request(struct request *rq)
1223 {
1224         int ret;
1225         struct request_queue *q = rq->q;
1226         struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
1227                         rq->mq_ctx->cpu);
1228         struct blk_mq_queue_data bd = {
1229                 .rq = rq,
1230                 .list = NULL,
1231                 .last = 1
1232         };
1233
1234         /*
1235          * For OK queue, we are done. For error, kill it. Any other
1236          * error (busy), just add it to our list as we previously
1237          * would have done
1238          */
1239         ret = q->mq_ops->queue_rq(hctx, &bd);
1240         if (ret == BLK_MQ_RQ_QUEUE_OK)
1241                 return 0;
1242         else {
1243                 __blk_mq_requeue_request(rq);
1244
1245                 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1246                         rq->errors = -EIO;
1247                         blk_mq_end_request(rq, rq->errors);
1248                         return 0;
1249                 }
1250                 return -1;
1251         }
1252 }
1253
1254 /*
1255  * Multiple hardware queue variant. This will not use per-process plugs,
1256  * but will attempt to bypass the hctx queueing if we can go straight to
1257  * hardware for SYNC IO.
1258  */
1259 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1260 {
1261         const int is_sync = rw_is_sync(bio->bi_rw);
1262         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1263         struct blk_map_ctx data;
1264         struct request *rq;
1265         unsigned int request_count = 0;
1266         struct blk_plug *plug;
1267         struct request *same_queue_rq = NULL;
1268
1269         blk_queue_bounce(q, &bio);
1270
1271         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1272                 bio_io_error(bio);
1273                 return;
1274         }
1275
1276         blk_queue_split(q, &bio, q->bio_split);
1277
1278         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1279             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1280                 return;
1281
1282         rq = blk_mq_map_request(q, bio, &data);
1283         if (unlikely(!rq))
1284                 return;
1285
1286         if (unlikely(is_flush_fua)) {
1287                 blk_mq_bio_to_request(rq, bio);
1288                 blk_insert_flush(rq);
1289                 goto run_queue;
1290         }
1291
1292         plug = current->plug;
1293         /*
1294          * If the driver supports defer issued based on 'last', then
1295          * queue it up like normal since we can potentially save some
1296          * CPU this way.
1297          */
1298         if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
1299             !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
1300                 struct request *old_rq = NULL;
1301
1302                 blk_mq_bio_to_request(rq, bio);
1303
1304                 /*
1305                  * we do limited pluging. If bio can be merged, do merge.
1306                  * Otherwise the existing request in the plug list will be
1307                  * issued. So the plug list will have one request at most
1308                  */
1309                 if (plug) {
1310                         /*
1311                          * The plug list might get flushed before this. If that
1312                          * happens, same_queue_rq is invalid and plug list is empty
1313                          **/
1314                         if (same_queue_rq && !list_empty(&plug->mq_list)) {
1315                                 old_rq = same_queue_rq;
1316                                 list_del_init(&old_rq->queuelist);
1317                         }
1318                         list_add_tail(&rq->queuelist, &plug->mq_list);
1319                 } else /* is_sync */
1320                         old_rq = rq;
1321                 blk_mq_put_ctx(data.ctx);
1322                 if (!old_rq)
1323                         return;
1324                 if (!blk_mq_direct_issue_request(old_rq))
1325                         return;
1326                 blk_mq_insert_request(old_rq, false, true, true);
1327                 return;
1328         }
1329
1330         if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1331                 /*
1332                  * For a SYNC request, send it to the hardware immediately. For
1333                  * an ASYNC request, just ensure that we run it later on. The
1334                  * latter allows for merging opportunities and more efficient
1335                  * dispatching.
1336                  */
1337 run_queue:
1338                 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1339         }
1340         blk_mq_put_ctx(data.ctx);
1341 }
1342
1343 /*
1344  * Single hardware queue variant. This will attempt to use any per-process
1345  * plug for merging and IO deferral.
1346  */
1347 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1348 {
1349         const int is_sync = rw_is_sync(bio->bi_rw);
1350         const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1351         struct blk_plug *plug;
1352         unsigned int request_count = 0;
1353         struct blk_map_ctx data;
1354         struct request *rq;
1355
1356         blk_queue_bounce(q, &bio);
1357
1358         if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1359                 bio_io_error(bio);
1360                 return;
1361         }
1362
1363         blk_queue_split(q, &bio, q->bio_split);
1364
1365         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1366             blk_attempt_plug_merge(q, bio, &request_count, NULL))
1367                 return;
1368
1369         rq = blk_mq_map_request(q, bio, &data);
1370         if (unlikely(!rq))
1371                 return;
1372
1373         if (unlikely(is_flush_fua)) {
1374                 blk_mq_bio_to_request(rq, bio);
1375                 blk_insert_flush(rq);
1376                 goto run_queue;
1377         }
1378
1379         /*
1380          * A task plug currently exists. Since this is completely lockless,
1381          * utilize that to temporarily store requests until the task is
1382          * either done or scheduled away.
1383          */
1384         plug = current->plug;
1385         if (plug) {
1386                 blk_mq_bio_to_request(rq, bio);
1387                 if (list_empty(&plug->mq_list))
1388                         trace_block_plug(q);
1389                 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1390                         blk_flush_plug_list(plug, false);
1391                         trace_block_plug(q);
1392                 }
1393                 list_add_tail(&rq->queuelist, &plug->mq_list);
1394                 blk_mq_put_ctx(data.ctx);
1395                 return;
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         return tags;
1532
1533 fail:
1534         blk_mq_free_rq_map(set, tags, hctx_idx);
1535         return NULL;
1536 }
1537
1538 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1539 {
1540         kfree(bitmap->map);
1541 }
1542
1543 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1544 {
1545         unsigned int bpw = 8, total, num_maps, i;
1546
1547         bitmap->bits_per_word = bpw;
1548
1549         num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1550         bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1551                                         GFP_KERNEL, node);
1552         if (!bitmap->map)
1553                 return -ENOMEM;
1554
1555         total = nr_cpu_ids;
1556         for (i = 0; i < num_maps; i++) {
1557                 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1558                 total -= bitmap->map[i].depth;
1559         }
1560
1561         return 0;
1562 }
1563
1564 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1565 {
1566         struct request_queue *q = hctx->queue;
1567         struct blk_mq_ctx *ctx;
1568         LIST_HEAD(tmp);
1569
1570         /*
1571          * Move ctx entries to new CPU, if this one is going away.
1572          */
1573         ctx = __blk_mq_get_ctx(q, cpu);
1574
1575         spin_lock(&ctx->lock);
1576         if (!list_empty(&ctx->rq_list)) {
1577                 list_splice_init(&ctx->rq_list, &tmp);
1578                 blk_mq_hctx_clear_pending(hctx, ctx);
1579         }
1580         spin_unlock(&ctx->lock);
1581
1582         if (list_empty(&tmp))
1583                 return NOTIFY_OK;
1584
1585         ctx = blk_mq_get_ctx(q);
1586         spin_lock(&ctx->lock);
1587
1588         while (!list_empty(&tmp)) {
1589                 struct request *rq;
1590
1591                 rq = list_first_entry(&tmp, struct request, queuelist);
1592                 rq->mq_ctx = ctx;
1593                 list_move_tail(&rq->queuelist, &ctx->rq_list);
1594         }
1595
1596         hctx = q->mq_ops->map_queue(q, ctx->cpu);
1597         blk_mq_hctx_mark_pending(hctx, ctx);
1598
1599         spin_unlock(&ctx->lock);
1600
1601         blk_mq_run_hw_queue(hctx, true);
1602         blk_mq_put_ctx(ctx);
1603         return NOTIFY_OK;
1604 }
1605
1606 static int blk_mq_hctx_notify(void *data, unsigned long action,
1607                               unsigned int cpu)
1608 {
1609         struct blk_mq_hw_ctx *hctx = data;
1610
1611         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1612                 return blk_mq_hctx_cpu_offline(hctx, cpu);
1613
1614         /*
1615          * In case of CPU online, tags may be reallocated
1616          * in blk_mq_map_swqueue() after mapping is updated.
1617          */
1618
1619         return NOTIFY_OK;
1620 }
1621
1622 /* hctx->ctxs will be freed in queue's release handler */
1623 static void blk_mq_exit_hctx(struct request_queue *q,
1624                 struct blk_mq_tag_set *set,
1625                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1626 {
1627         unsigned flush_start_tag = set->queue_depth;
1628
1629         blk_mq_tag_idle(hctx);
1630
1631         if (set->ops->exit_request)
1632                 set->ops->exit_request(set->driver_data,
1633                                        hctx->fq->flush_rq, hctx_idx,
1634                                        flush_start_tag + hctx_idx);
1635
1636         if (set->ops->exit_hctx)
1637                 set->ops->exit_hctx(hctx, hctx_idx);
1638
1639         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1640         blk_free_flush_queue(hctx->fq);
1641         blk_mq_free_bitmap(&hctx->ctx_map);
1642 }
1643
1644 static void blk_mq_exit_hw_queues(struct request_queue *q,
1645                 struct blk_mq_tag_set *set, int nr_queue)
1646 {
1647         struct blk_mq_hw_ctx *hctx;
1648         unsigned int i;
1649
1650         queue_for_each_hw_ctx(q, hctx, i) {
1651                 if (i == nr_queue)
1652                         break;
1653                 blk_mq_exit_hctx(q, set, hctx, i);
1654         }
1655 }
1656
1657 static void blk_mq_free_hw_queues(struct request_queue *q,
1658                 struct blk_mq_tag_set *set)
1659 {
1660         struct blk_mq_hw_ctx *hctx;
1661         unsigned int i;
1662
1663         queue_for_each_hw_ctx(q, hctx, i)
1664                 free_cpumask_var(hctx->cpumask);
1665 }
1666
1667 static int blk_mq_init_hctx(struct request_queue *q,
1668                 struct blk_mq_tag_set *set,
1669                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1670 {
1671         int node;
1672         unsigned flush_start_tag = set->queue_depth;
1673
1674         node = hctx->numa_node;
1675         if (node == NUMA_NO_NODE)
1676                 node = hctx->numa_node = set->numa_node;
1677
1678         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1679         INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1680         spin_lock_init(&hctx->lock);
1681         INIT_LIST_HEAD(&hctx->dispatch);
1682         hctx->queue = q;
1683         hctx->queue_num = hctx_idx;
1684         hctx->flags = set->flags;
1685
1686         blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1687                                         blk_mq_hctx_notify, hctx);
1688         blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1689
1690         hctx->tags = set->tags[hctx_idx];
1691
1692         /*
1693          * Allocate space for all possible cpus to avoid allocation at
1694          * runtime
1695          */
1696         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1697                                         GFP_KERNEL, node);
1698         if (!hctx->ctxs)
1699                 goto unregister_cpu_notifier;
1700
1701         if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1702                 goto free_ctxs;
1703
1704         hctx->nr_ctx = 0;
1705
1706         if (set->ops->init_hctx &&
1707             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1708                 goto free_bitmap;
1709
1710         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1711         if (!hctx->fq)
1712                 goto exit_hctx;
1713
1714         if (set->ops->init_request &&
1715             set->ops->init_request(set->driver_data,
1716                                    hctx->fq->flush_rq, hctx_idx,
1717                                    flush_start_tag + hctx_idx, node))
1718                 goto free_fq;
1719
1720         return 0;
1721
1722  free_fq:
1723         kfree(hctx->fq);
1724  exit_hctx:
1725         if (set->ops->exit_hctx)
1726                 set->ops->exit_hctx(hctx, hctx_idx);
1727  free_bitmap:
1728         blk_mq_free_bitmap(&hctx->ctx_map);
1729  free_ctxs:
1730         kfree(hctx->ctxs);
1731  unregister_cpu_notifier:
1732         blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1733
1734         return -1;
1735 }
1736
1737 static int blk_mq_init_hw_queues(struct request_queue *q,
1738                 struct blk_mq_tag_set *set)
1739 {
1740         struct blk_mq_hw_ctx *hctx;
1741         unsigned int i;
1742
1743         /*
1744          * Initialize hardware queues
1745          */
1746         queue_for_each_hw_ctx(q, hctx, i) {
1747                 if (blk_mq_init_hctx(q, set, hctx, i))
1748                         break;
1749         }
1750
1751         if (i == q->nr_hw_queues)
1752                 return 0;
1753
1754         /*
1755          * Init failed
1756          */
1757         blk_mq_exit_hw_queues(q, set, i);
1758
1759         return 1;
1760 }
1761
1762 static void blk_mq_init_cpu_queues(struct request_queue *q,
1763                                    unsigned int nr_hw_queues)
1764 {
1765         unsigned int i;
1766
1767         for_each_possible_cpu(i) {
1768                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1769                 struct blk_mq_hw_ctx *hctx;
1770
1771                 memset(__ctx, 0, sizeof(*__ctx));
1772                 __ctx->cpu = i;
1773                 spin_lock_init(&__ctx->lock);
1774                 INIT_LIST_HEAD(&__ctx->rq_list);
1775                 __ctx->queue = q;
1776
1777                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1778                 if (!cpu_online(i))
1779                         continue;
1780
1781                 hctx = q->mq_ops->map_queue(q, i);
1782
1783                 /*
1784                  * Set local node, IFF we have more than one hw queue. If
1785                  * not, we remain on the home node of the device
1786                  */
1787                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1788                         hctx->numa_node = cpu_to_node(i);
1789         }
1790 }
1791
1792 static void blk_mq_map_swqueue(struct request_queue *q,
1793                                const struct cpumask *online_mask)
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         /*
1801          * Avoid others reading imcomplete hctx->cpumask through sysfs
1802          */
1803         mutex_lock(&q->sysfs_lock);
1804
1805         queue_for_each_hw_ctx(q, hctx, i) {
1806                 cpumask_clear(hctx->cpumask);
1807                 hctx->nr_ctx = 0;
1808         }
1809
1810         /*
1811          * Map software to hardware queues
1812          */
1813         queue_for_each_ctx(q, ctx, i) {
1814                 /* If the cpu isn't online, the cpu is mapped to first hctx */
1815                 if (!cpumask_test_cpu(i, online_mask))
1816                         continue;
1817
1818                 hctx = q->mq_ops->map_queue(q, i);
1819                 cpumask_set_cpu(i, hctx->cpumask);
1820                 ctx->index_hw = hctx->nr_ctx;
1821                 hctx->ctxs[hctx->nr_ctx++] = ctx;
1822         }
1823
1824         mutex_unlock(&q->sysfs_lock);
1825
1826         queue_for_each_hw_ctx(q, hctx, i) {
1827                 struct blk_mq_ctxmap *map = &hctx->ctx_map;
1828
1829                 /*
1830                  * If no software queues are mapped to this hardware queue,
1831                  * disable it and free the request entries.
1832                  */
1833                 if (!hctx->nr_ctx) {
1834                         if (set->tags[i]) {
1835                                 blk_mq_free_rq_map(set, set->tags[i], i);
1836                                 set->tags[i] = NULL;
1837                         }
1838                         hctx->tags = NULL;
1839                         continue;
1840                 }
1841
1842                 /* unmapped hw queue can be remapped after CPU topo changed */
1843                 if (!set->tags[i])
1844                         set->tags[i] = blk_mq_init_rq_map(set, i);
1845                 hctx->tags = set->tags[i];
1846                 WARN_ON(!hctx->tags);
1847
1848                 /*
1849                  * Set the map size to the number of mapped software queues.
1850                  * This is more accurate and more efficient than looping
1851                  * over all possibly mapped software queues.
1852                  */
1853                 map->size = DIV_ROUND_UP(hctx->nr_ctx, map->bits_per_word);
1854
1855                 /*
1856                  * Initialize batch roundrobin counts
1857                  */
1858                 hctx->next_cpu = cpumask_first(hctx->cpumask);
1859                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1860         }
1861
1862         queue_for_each_ctx(q, ctx, i) {
1863                 if (!cpumask_test_cpu(i, online_mask))
1864                         continue;
1865
1866                 hctx = q->mq_ops->map_queue(q, i);
1867                 cpumask_set_cpu(i, hctx->tags->cpumask);
1868         }
1869 }
1870
1871 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1872 {
1873         struct blk_mq_hw_ctx *hctx;
1874         struct request_queue *q;
1875         bool shared;
1876         int i;
1877
1878         if (set->tag_list.next == set->tag_list.prev)
1879                 shared = false;
1880         else
1881                 shared = true;
1882
1883         list_for_each_entry(q, &set->tag_list, tag_set_list) {
1884                 blk_mq_freeze_queue(q);
1885
1886                 queue_for_each_hw_ctx(q, hctx, i) {
1887                         if (shared)
1888                                 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1889                         else
1890                                 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1891                 }
1892                 blk_mq_unfreeze_queue(q);
1893         }
1894 }
1895
1896 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1897 {
1898         struct blk_mq_tag_set *set = q->tag_set;
1899
1900         mutex_lock(&set->tag_list_lock);
1901         list_del_init(&q->tag_set_list);
1902         blk_mq_update_tag_set_depth(set);
1903         mutex_unlock(&set->tag_list_lock);
1904 }
1905
1906 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1907                                      struct request_queue *q)
1908 {
1909         q->tag_set = set;
1910
1911         mutex_lock(&set->tag_list_lock);
1912         list_add_tail(&q->tag_set_list, &set->tag_list);
1913         blk_mq_update_tag_set_depth(set);
1914         mutex_unlock(&set->tag_list_lock);
1915 }
1916
1917 /*
1918  * It is the actual release handler for mq, but we do it from
1919  * request queue's release handler for avoiding use-after-free
1920  * and headache because q->mq_kobj shouldn't have been introduced,
1921  * but we can't group ctx/kctx kobj without it.
1922  */
1923 void blk_mq_release(struct request_queue *q)
1924 {
1925         struct blk_mq_hw_ctx *hctx;
1926         unsigned int i;
1927
1928         /* hctx kobj stays in hctx */
1929         queue_for_each_hw_ctx(q, hctx, i) {
1930                 if (!hctx)
1931                         continue;
1932                 kfree(hctx->ctxs);
1933                 kfree(hctx);
1934         }
1935
1936         kfree(q->mq_map);
1937         q->mq_map = NULL;
1938
1939         kfree(q->queue_hw_ctx);
1940
1941         /* ctx kobj stays in queue_ctx */
1942         free_percpu(q->queue_ctx);
1943 }
1944
1945 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1946 {
1947         struct request_queue *uninit_q, *q;
1948
1949         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1950         if (!uninit_q)
1951                 return ERR_PTR(-ENOMEM);
1952
1953         q = blk_mq_init_allocated_queue(set, uninit_q);
1954         if (IS_ERR(q))
1955                 blk_cleanup_queue(uninit_q);
1956
1957         return q;
1958 }
1959 EXPORT_SYMBOL(blk_mq_init_queue);
1960
1961 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
1962                                                   struct request_queue *q)
1963 {
1964         struct blk_mq_hw_ctx **hctxs;
1965         struct blk_mq_ctx __percpu *ctx;
1966         unsigned int *map;
1967         int i;
1968
1969         ctx = alloc_percpu(struct blk_mq_ctx);
1970         if (!ctx)
1971                 return ERR_PTR(-ENOMEM);
1972
1973         hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1974                         set->numa_node);
1975
1976         if (!hctxs)
1977                 goto err_percpu;
1978
1979         map = blk_mq_make_queue_map(set);
1980         if (!map)
1981                 goto err_map;
1982
1983         for (i = 0; i < set->nr_hw_queues; i++) {
1984                 int node = blk_mq_hw_queue_to_node(map, i);
1985
1986                 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1987                                         GFP_KERNEL, node);
1988                 if (!hctxs[i])
1989                         goto err_hctxs;
1990
1991                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
1992                                                 node))
1993                         goto err_hctxs;
1994
1995                 atomic_set(&hctxs[i]->nr_active, 0);
1996                 hctxs[i]->numa_node = node;
1997                 hctxs[i]->queue_num = i;
1998         }
1999
2000         /*
2001          * Init percpu_ref in atomic mode so that it's faster to shutdown.
2002          * See blk_register_queue() for details.
2003          */
2004         if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release,
2005                             PERCPU_REF_INIT_ATOMIC, GFP_KERNEL))
2006                 goto err_hctxs;
2007
2008         setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
2009         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2010
2011         q->nr_queues = nr_cpu_ids;
2012         q->nr_hw_queues = set->nr_hw_queues;
2013         q->mq_map = map;
2014
2015         q->queue_ctx = ctx;
2016         q->queue_hw_ctx = hctxs;
2017
2018         q->mq_ops = set->ops;
2019         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2020
2021         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2022                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2023
2024         q->sg_reserved_size = INT_MAX;
2025
2026         INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
2027         INIT_LIST_HEAD(&q->requeue_list);
2028         spin_lock_init(&q->requeue_lock);
2029
2030         if (q->nr_hw_queues > 1)
2031                 blk_queue_make_request(q, blk_mq_make_request);
2032         else
2033                 blk_queue_make_request(q, blk_sq_make_request);
2034
2035         /*
2036          * Do this after blk_queue_make_request() overrides it...
2037          */
2038         q->nr_requests = set->queue_depth;
2039
2040         if (set->ops->complete)
2041                 blk_queue_softirq_done(q, set->ops->complete);
2042
2043         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2044
2045         if (blk_mq_init_hw_queues(q, set))
2046                 goto err_hctxs;
2047
2048         get_online_cpus();
2049         mutex_lock(&all_q_mutex);
2050
2051         list_add_tail(&q->all_q_node, &all_q_list);
2052         blk_mq_add_queue_tag_set(set, q);
2053         blk_mq_map_swqueue(q, cpu_online_mask);
2054
2055         mutex_unlock(&all_q_mutex);
2056         put_online_cpus();
2057
2058         return q;
2059
2060 err_hctxs:
2061         kfree(map);
2062         for (i = 0; i < set->nr_hw_queues; i++) {
2063                 if (!hctxs[i])
2064                         break;
2065                 free_cpumask_var(hctxs[i]->cpumask);
2066                 kfree(hctxs[i]);
2067         }
2068 err_map:
2069         kfree(hctxs);
2070 err_percpu:
2071         free_percpu(ctx);
2072         return ERR_PTR(-ENOMEM);
2073 }
2074 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2075
2076 void blk_mq_free_queue(struct request_queue *q)
2077 {
2078         struct blk_mq_tag_set   *set = q->tag_set;
2079
2080         mutex_lock(&all_q_mutex);
2081         list_del_init(&q->all_q_node);
2082         mutex_unlock(&all_q_mutex);
2083
2084         blk_mq_del_queue_tag_set(q);
2085
2086         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2087         blk_mq_free_hw_queues(q, set);
2088
2089         percpu_ref_exit(&q->mq_usage_counter);
2090 }
2091
2092 /* Basically redo blk_mq_init_queue with queue frozen */
2093 static void blk_mq_queue_reinit(struct request_queue *q,
2094                                 const struct cpumask *online_mask)
2095 {
2096         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2097
2098         blk_mq_sysfs_unregister(q);
2099
2100         blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues, online_mask);
2101
2102         /*
2103          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2104          * we should change hctx numa_node according to new topology (this
2105          * involves free and re-allocate memory, worthy doing?)
2106          */
2107
2108         blk_mq_map_swqueue(q, online_mask);
2109
2110         blk_mq_sysfs_register(q);
2111 }
2112
2113 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
2114                                       unsigned long action, void *hcpu)
2115 {
2116         struct request_queue *q;
2117         int cpu = (unsigned long)hcpu;
2118         /*
2119          * New online cpumask which is going to be set in this hotplug event.
2120          * Declare this cpumasks as global as cpu-hotplug operation is invoked
2121          * one-by-one and dynamically allocating this could result in a failure.
2122          */
2123         static struct cpumask online_new;
2124
2125         /*
2126          * Before hotadded cpu starts handling requests, new mappings must
2127          * be established.  Otherwise, these requests in hw queue might
2128          * never be dispatched.
2129          *
2130          * For example, there is a single hw queue (hctx) and two CPU queues
2131          * (ctx0 for CPU0, and ctx1 for CPU1).
2132          *
2133          * Now CPU1 is just onlined and a request is inserted into
2134          * ctx1->rq_list and set bit0 in pending bitmap as ctx1->index_hw is
2135          * still zero.
2136          *
2137          * And then while running hw queue, flush_busy_ctxs() finds bit0 is
2138          * set in pending bitmap and tries to retrieve requests in
2139          * hctx->ctxs[0]->rq_list.  But htx->ctxs[0] is a pointer to ctx0,
2140          * so the request in ctx1->rq_list is ignored.
2141          */
2142         switch (action & ~CPU_TASKS_FROZEN) {
2143         case CPU_DEAD:
2144         case CPU_UP_CANCELED:
2145                 cpumask_copy(&online_new, cpu_online_mask);
2146                 break;
2147         case CPU_UP_PREPARE:
2148                 cpumask_copy(&online_new, cpu_online_mask);
2149                 cpumask_set_cpu(cpu, &online_new);
2150                 break;
2151         default:
2152                 return NOTIFY_OK;
2153         }
2154
2155         mutex_lock(&all_q_mutex);
2156
2157         /*
2158          * We need to freeze and reinit all existing queues.  Freezing
2159          * involves synchronous wait for an RCU grace period and doing it
2160          * one by one may take a long time.  Start freezing all queues in
2161          * one swoop and then wait for the completions so that freezing can
2162          * take place in parallel.
2163          */
2164         list_for_each_entry(q, &all_q_list, all_q_node)
2165                 blk_mq_freeze_queue_start(q);
2166         list_for_each_entry(q, &all_q_list, all_q_node) {
2167                 blk_mq_freeze_queue_wait(q);
2168
2169                 /*
2170                  * timeout handler can't touch hw queue during the
2171                  * reinitialization
2172                  */
2173                 del_timer_sync(&q->timeout);
2174         }
2175
2176         list_for_each_entry(q, &all_q_list, all_q_node)
2177                 blk_mq_queue_reinit(q, &online_new);
2178
2179         list_for_each_entry(q, &all_q_list, all_q_node)
2180                 blk_mq_unfreeze_queue(q);
2181
2182         mutex_unlock(&all_q_mutex);
2183         return NOTIFY_OK;
2184 }
2185
2186 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2187 {
2188         int i;
2189
2190         for (i = 0; i < set->nr_hw_queues; i++) {
2191                 set->tags[i] = blk_mq_init_rq_map(set, i);
2192                 if (!set->tags[i])
2193                         goto out_unwind;
2194         }
2195
2196         return 0;
2197
2198 out_unwind:
2199         while (--i >= 0)
2200                 blk_mq_free_rq_map(set, set->tags[i], i);
2201
2202         return -ENOMEM;
2203 }
2204
2205 /*
2206  * Allocate the request maps associated with this tag_set. Note that this
2207  * may reduce the depth asked for, if memory is tight. set->queue_depth
2208  * will be updated to reflect the allocated depth.
2209  */
2210 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2211 {
2212         unsigned int depth;
2213         int err;
2214
2215         depth = set->queue_depth;
2216         do {
2217                 err = __blk_mq_alloc_rq_maps(set);
2218                 if (!err)
2219                         break;
2220
2221                 set->queue_depth >>= 1;
2222                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2223                         err = -ENOMEM;
2224                         break;
2225                 }
2226         } while (set->queue_depth);
2227
2228         if (!set->queue_depth || err) {
2229                 pr_err("blk-mq: failed to allocate request map\n");
2230                 return -ENOMEM;
2231         }
2232
2233         if (depth != set->queue_depth)
2234                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2235                                                 depth, set->queue_depth);
2236
2237         return 0;
2238 }
2239
2240 struct cpumask *blk_mq_tags_cpumask(struct blk_mq_tags *tags)
2241 {
2242         return tags->cpumask;
2243 }
2244 EXPORT_SYMBOL_GPL(blk_mq_tags_cpumask);
2245
2246 /*
2247  * Alloc a tag set to be associated with one or more request queues.
2248  * May fail with EINVAL for various error conditions. May adjust the
2249  * requested depth down, if if it too large. In that case, the set
2250  * value will be stored in set->queue_depth.
2251  */
2252 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2253 {
2254         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2255
2256         if (!set->nr_hw_queues)
2257                 return -EINVAL;
2258         if (!set->queue_depth)
2259                 return -EINVAL;
2260         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2261                 return -EINVAL;
2262
2263         if (!set->ops->queue_rq || !set->ops->map_queue)
2264                 return -EINVAL;
2265
2266         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2267                 pr_info("blk-mq: reduced tag depth to %u\n",
2268                         BLK_MQ_MAX_DEPTH);
2269                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2270         }
2271
2272         /*
2273          * If a crashdump is active, then we are potentially in a very
2274          * memory constrained environment. Limit us to 1 queue and
2275          * 64 tags to prevent using too much memory.
2276          */
2277         if (is_kdump_kernel()) {
2278                 set->nr_hw_queues = 1;
2279                 set->queue_depth = min(64U, set->queue_depth);
2280         }
2281
2282         set->tags = kmalloc_node(set->nr_hw_queues *
2283                                  sizeof(struct blk_mq_tags *),
2284                                  GFP_KERNEL, set->numa_node);
2285         if (!set->tags)
2286                 return -ENOMEM;
2287
2288         if (blk_mq_alloc_rq_maps(set))
2289                 goto enomem;
2290
2291         mutex_init(&set->tag_list_lock);
2292         INIT_LIST_HEAD(&set->tag_list);
2293
2294         return 0;
2295 enomem:
2296         kfree(set->tags);
2297         set->tags = NULL;
2298         return -ENOMEM;
2299 }
2300 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2301
2302 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2303 {
2304         int i;
2305
2306         for (i = 0; i < set->nr_hw_queues; i++) {
2307                 if (set->tags[i]) {
2308                         blk_mq_free_rq_map(set, set->tags[i], i);
2309                         free_cpumask_var(set->tags[i]->cpumask);
2310                 }
2311         }
2312
2313         kfree(set->tags);
2314         set->tags = NULL;
2315 }
2316 EXPORT_SYMBOL(blk_mq_free_tag_set);
2317
2318 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2319 {
2320         struct blk_mq_tag_set *set = q->tag_set;
2321         struct blk_mq_hw_ctx *hctx;
2322         int i, ret;
2323
2324         if (!set || nr > set->queue_depth)
2325                 return -EINVAL;
2326
2327         ret = 0;
2328         queue_for_each_hw_ctx(q, hctx, i) {
2329                 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2330                 if (ret)
2331                         break;
2332         }
2333
2334         if (!ret)
2335                 q->nr_requests = nr;
2336
2337         return ret;
2338 }
2339
2340 void blk_mq_disable_hotplug(void)
2341 {
2342         mutex_lock(&all_q_mutex);
2343 }
2344
2345 void blk_mq_enable_hotplug(void)
2346 {
2347         mutex_unlock(&all_q_mutex);
2348 }
2349
2350 static int __init blk_mq_init(void)
2351 {
2352         blk_mq_cpu_init();
2353
2354         hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2355
2356         return 0;
2357 }
2358 subsys_initcall(blk_mq_init);