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