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/kmemleak.h>
  13 #include <linux/mm.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/smp.h>
  18 #include <linux/llist.h>
  19 #include <linux/list_sort.h>
  20 #include <linux/cpu.h>
  21 #include <linux/cache.h>
  22 #include <linux/sched/sysctl.h>
  23 #include <linux/sched/topology.h>
  24 #include <linux/sched/signal.h>
  25 #include <linux/delay.h>
  26 #include <linux/crash_dump.h>
  27 #include <linux/prefetch.h>
  28
  29 #include <trace/events/block.h>
  30
  31 #include <linux/blk-mq.h>
  32 #include "blk.h"
  33 #include "blk-mq.h"
  34 #include "blk-mq-debugfs.h"
  35 #include "blk-mq-tag.h"
  36 #include "blk-stat.h"
  37 #include "blk-wbt.h"
  38 #include "blk-mq-sched.h"
  39
  40 static void blk_mq_poll_stats_start(struct request_queue *q);
  41 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  42
  43 static int blk_mq_poll_stats_bkt(const struct request *rq)
  44 {
  45         int ddir, bytes, bucket;
  46
  47         ddir = rq_data_dir(rq);
  48         bytes = blk_rq_bytes(rq);
  49
  50         bucket = ddir + 2*(ilog2(bytes) - 9);
  51
  52         if (bucket < 0)
  53                 return -1;
  54         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  55                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  56
  57         return bucket;
  58 }
  59
  60 /*
  61  * Check if any of the ctx's have pending work in this hardware queue
  62  */
  63 bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  64 {
  65         return sbitmap_any_bit_set(&hctx->ctx_map) ||
  66                         !list_empty_careful(&hctx->dispatch) ||
  67                         blk_mq_sched_has_work(hctx);
  68 }
  69
  70 /*
  71  * Mark this ctx as having pending work in this hardware queue
  72  */
  73 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  74                                      struct blk_mq_ctx *ctx)
  75 {
  76         if (!sbitmap_test_bit(&hctx->ctx_map, ctx->index_hw))
  77                 sbitmap_set_bit(&hctx->ctx_map, ctx->index_hw);
  78 }
  79
  80 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  81                                       struct blk_mq_ctx *ctx)
  82 {
  83         sbitmap_clear_bit(&hctx->ctx_map, ctx->index_hw);
  84 }
  85
  86 void blk_freeze_queue_start(struct request_queue *q)
  87 {
  88         int freeze_depth;
  89
  90         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
  91         if (freeze_depth == 1) {
  92                 percpu_ref_kill(&q->q_usage_counter);
  93                 blk_mq_run_hw_queues(q, false);
  94         }
  95 }
  96 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
  97
  98 void blk_mq_freeze_queue_wait(struct request_queue *q)
  99 {
 100         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 101 }
 102 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 103
 104 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 105                                      unsigned long timeout)
 106 {
 107         return wait_event_timeout(q->mq_freeze_wq,
 108                                         percpu_ref_is_zero(&q->q_usage_counter),
 109                                         timeout);
 110 }
 111 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 112
 113 /*
 114  * Guarantee no request is in use, so we can change any data structure of
 115  * the queue afterward.
 116  */
 117 void blk_freeze_queue(struct request_queue *q)
 118 {
 119         /*
 120          * In the !blk_mq case we are only calling this to kill the
 121          * q_usage_counter, otherwise this increases the freeze depth
 122          * and waits for it to return to zero.  For this reason there is
 123          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 124          * exported to drivers as the only user for unfreeze is blk_mq.
 125          */
 126         blk_freeze_queue_start(q);
 127         blk_mq_freeze_queue_wait(q);
 128 }
 129
 130 void blk_mq_freeze_queue(struct request_queue *q)
 131 {
 132         /*
 133          * ...just an alias to keep freeze and unfreeze actions balanced
 134          * in the blk_mq_* namespace
 135          */
 136         blk_freeze_queue(q);
 137 }
 138 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 139
 140 void blk_mq_unfreeze_queue(struct request_queue *q)
 141 {
 142         int freeze_depth;
 143
 144         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 145         WARN_ON_ONCE(freeze_depth < 0);
 146         if (!freeze_depth) {
 147                 percpu_ref_reinit(&q->q_usage_counter);
 148                 wake_up_all(&q->mq_freeze_wq);
 149         }
 150 }
 151 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 152
 153 /*
 154  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 155  * mpt3sas driver such that this function can be removed.
 156  */
 157 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 158 {
 159         unsigned long flags;
 160
 161         spin_lock_irqsave(q->queue_lock, flags);
 162         queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 163         spin_unlock_irqrestore(q->queue_lock, flags);
 164 }
 165 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 166
 167 /**
 168  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 169  * @q: request queue.
 170  *
 171  * Note: this function does not prevent that the struct request end_io()
 172  * callback function is invoked. Once this function is returned, we make
 173  * sure no dispatch can happen until the queue is unquiesced via
 174  * blk_mq_unquiesce_queue().
 175  */
 176 void blk_mq_quiesce_queue(struct request_queue *q)
 177 {
 178         struct blk_mq_hw_ctx *hctx;
 179         unsigned int i;
 180         bool rcu = false;
 181
 182         blk_mq_quiesce_queue_nowait(q);
 183
 184         queue_for_each_hw_ctx(q, hctx, i) {
 185                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 186                         synchronize_srcu(hctx->queue_rq_srcu);
 187                 else
 188                         rcu = true;
 189         }
 190         if (rcu)
 191                 synchronize_rcu();
 192 }
 193 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 194
 195 /*
 196  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 197  * @q: request queue.
 198  *
 199  * This function recovers queue into the state before quiescing
 200  * which is done by blk_mq_quiesce_queue.
 201  */
 202 void blk_mq_unquiesce_queue(struct request_queue *q)
 203 {
 204         unsigned long flags;
 205
 206         spin_lock_irqsave(q->queue_lock, flags);
 207         queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 208         spin_unlock_irqrestore(q->queue_lock, flags);
 209
 210         /* dispatch requests which are inserted during quiescing */
 211         blk_mq_run_hw_queues(q, true);
 212 }
 213 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 214
 215 void blk_mq_wake_waiters(struct request_queue *q)
 216 {
 217         struct blk_mq_hw_ctx *hctx;
 218         unsigned int i;
 219
 220         queue_for_each_hw_ctx(q, hctx, i)
 221                 if (blk_mq_hw_queue_mapped(hctx))
 222                         blk_mq_tag_wakeup_all(hctx->tags, true);
 223
 224         /*
 225          * If we are called because the queue has now been marked as
 226          * dying, we need to ensure that processes currently waiting on
 227          * the queue are notified as well.
 228          */
 229         wake_up_all(&q->mq_freeze_wq);
 230 }
 231
 232 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 233 {
 234         return blk_mq_has_free_tags(hctx->tags);
 235 }
 236 EXPORT_SYMBOL(blk_mq_can_queue);
 237
 238 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 239                 unsigned int tag, unsigned int op)
 240 {
 241         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 242         struct request *rq = tags->static_rqs[tag];
 243
 244         rq->rq_flags = 0;
 245
 246         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 247                 rq->tag = -1;
 248                 rq->internal_tag = tag;
 249         } else {
 250                 if (blk_mq_tag_busy(data->hctx)) {
 251                         rq->rq_flags = RQF_MQ_INFLIGHT;
 252                         atomic_inc(&data->hctx->nr_active);
 253                 }
 254                 rq->tag = tag;
 255                 rq->internal_tag = -1;
 256                 data->hctx->tags->rqs[rq->tag] = rq;
 257         }
 258
 259         INIT_LIST_HEAD(&rq->queuelist);
 260         /* csd/requeue_work/fifo_time is initialized before use */
 261         rq->q = data->q;
 262         rq->mq_ctx = data->ctx;
 263         rq->cmd_flags = op;
 264         if (blk_queue_io_stat(data->q))
 265                 rq->rq_flags |= RQF_IO_STAT;
 266         /* do not touch atomic flags, it needs atomic ops against the timer */
 267         rq->cpu = -1;
 268         INIT_HLIST_NODE(&rq->hash);
 269         RB_CLEAR_NODE(&rq->rb_node);
 270         rq->rq_disk = NULL;
 271         rq->part = NULL;
 272         rq->start_time = jiffies;
 273 #ifdef CONFIG_BLK_CGROUP
 274         rq->rl = NULL;
 275         set_start_time_ns(rq);
 276         rq->io_start_time_ns = 0;
 277 #endif
 278         rq->nr_phys_segments = 0;
 279 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 280         rq->nr_integrity_segments = 0;
 281 #endif
 282         rq->special = NULL;
 283         /* tag was already set */
 284         rq->extra_len = 0;
 285
 286         INIT_LIST_HEAD(&rq->timeout_list);
 287         rq->timeout = 0;
 288
 289         rq->end_io = NULL;
 290         rq->end_io_data = NULL;
 291         rq->next_rq = NULL;
 292
 293         data->ctx->rq_dispatched[op_is_sync(op)]++;
 294         return rq;
 295 }
 296
 297 static struct request *blk_mq_get_request(struct request_queue *q,
 298                 struct bio *bio, unsigned int op,
 299                 struct blk_mq_alloc_data *data)
 300 {
 301         struct elevator_queue *e = q->elevator;
 302         struct request *rq;
 303         unsigned int tag;
 304         struct blk_mq_ctx *local_ctx = NULL;
 305
 306         blk_queue_enter_live(q);
 307         data->q = q;
 308         if (likely(!data->ctx))
 309                 data->ctx = local_ctx = blk_mq_get_ctx(q);
 310         if (likely(!data->hctx))
 311                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 312         if (op & REQ_NOWAIT)
 313                 data->flags |= BLK_MQ_REQ_NOWAIT;
 314
 315         if (e) {
 316                 data->flags |= BLK_MQ_REQ_INTERNAL;
 317
 318                 /*
 319                  * Flush requests are special and go directly to the
 320                  * dispatch list.
 321                  */
 322                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
 323                         e->type->ops.mq.limit_depth(op, data);
 324         }
 325
 326         tag = blk_mq_get_tag(data);
 327         if (tag == BLK_MQ_TAG_FAIL) {
 328                 if (local_ctx) {
 329                         blk_mq_put_ctx(local_ctx);
 330                         data->ctx = NULL;
 331                 }
 332                 blk_queue_exit(q);
 333                 return NULL;
 334         }
 335
 336         rq = blk_mq_rq_ctx_init(data, tag, op);
 337         if (!op_is_flush(op)) {
 338                 rq->elv.icq = NULL;
 339                 if (e && e->type->ops.mq.prepare_request) {
 340                         if (e->type->icq_cache && rq_ioc(bio))
 341                                 blk_mq_sched_assign_ioc(rq, bio);
 342
 343                         e->type->ops.mq.prepare_request(rq, bio);
 344                         rq->rq_flags |= RQF_ELVPRIV;
 345                 }
 346         }
 347         data->hctx->queued++;
 348         return rq;
 349 }
 350
 351 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 352                 unsigned int flags)
 353 {
 354         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 355         struct request *rq;
 356         int ret;
 357
 358         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 359         if (ret)
 360                 return ERR_PTR(ret);
 361
 362         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 363         blk_queue_exit(q);
 364
 365         if (!rq)
 366                 return ERR_PTR(-EWOULDBLOCK);
 367
 368         blk_mq_put_ctx(alloc_data.ctx);
 369
 370         rq->__data_len = 0;
 371         rq->__sector = (sector_t) -1;
 372         rq->bio = rq->biotail = NULL;
 373         return rq;
 374 }
 375 EXPORT_SYMBOL(blk_mq_alloc_request);
 376
 377 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 378                 unsigned int op, unsigned int flags, unsigned int hctx_idx)
 379 {
 380         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 381         struct request *rq;
 382         unsigned int cpu;
 383         int ret;
 384
 385         /*
 386          * If the tag allocator sleeps we could get an allocation for a
 387          * different hardware context.  No need to complicate the low level
 388          * allocator for this for the rare use case of a command tied to
 389          * a specific queue.
 390          */
 391         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 392                 return ERR_PTR(-EINVAL);
 393
 394         if (hctx_idx >= q->nr_hw_queues)
 395                 return ERR_PTR(-EIO);
 396
 397         ret = blk_queue_enter(q, true);
 398         if (ret)
 399                 return ERR_PTR(ret);
 400
 401         /*
 402          * Check if the hardware context is actually mapped to anything.
 403          * If not tell the caller that it should skip this queue.
 404          */
 405         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 406         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 407                 blk_queue_exit(q);
 408                 return ERR_PTR(-EXDEV);
 409         }
 410         cpu = cpumask_first(alloc_data.hctx->cpumask);
 411         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 412
 413         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 414         blk_queue_exit(q);
 415
 416         if (!rq)
 417                 return ERR_PTR(-EWOULDBLOCK);
 418
 419         return rq;
 420 }
 421 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 422
 423 void blk_mq_free_request(struct request *rq)
 424 {
 425         struct request_queue *q = rq->q;
 426         struct elevator_queue *e = q->elevator;
 427         struct blk_mq_ctx *ctx = rq->mq_ctx;
 428         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 429         const int sched_tag = rq->internal_tag;
 430
 431         if (rq->rq_flags & RQF_ELVPRIV) {
 432                 if (e && e->type->ops.mq.finish_request)
 433                         e->type->ops.mq.finish_request(rq);
 434                 if (rq->elv.icq) {
 435                         put_io_context(rq->elv.icq->ioc);
 436                         rq->elv.icq = NULL;
 437                 }
 438         }
 439
 440         ctx->rq_completed[rq_is_sync(rq)]++;
 441         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 442                 atomic_dec(&hctx->nr_active);
 443
 444         wbt_done(q->rq_wb, &rq->issue_stat);
 445
 446         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 447         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 448         if (rq->tag != -1)
 449                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 450         if (sched_tag != -1)
 451                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 452         blk_mq_sched_restart(hctx);
 453         blk_queue_exit(q);
 454 }
 455 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 456
 457 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 458 {
 459         blk_account_io_done(rq);
 460
 461         if (rq->end_io) {
 462                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 463                 rq->end_io(rq, error);
 464         } else {
 465                 if (unlikely(blk_bidi_rq(rq)))
 466                         blk_mq_free_request(rq->next_rq);
 467                 blk_mq_free_request(rq);
 468         }
 469 }
 470 EXPORT_SYMBOL(__blk_mq_end_request);
 471
 472 void blk_mq_end_request(struct request *rq, blk_status_t error)
 473 {
 474         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 475                 BUG();
 476         __blk_mq_end_request(rq, error);
 477 }
 478 EXPORT_SYMBOL(blk_mq_end_request);
 479
 480 static void __blk_mq_complete_request_remote(void *data)
 481 {
 482         struct request *rq = data;
 483
 484         rq->q->softirq_done_fn(rq);
 485 }
 486
 487 static void __blk_mq_complete_request(struct request *rq)
 488 {
 489         struct blk_mq_ctx *ctx = rq->mq_ctx;
 490         bool shared = false;
 491         int cpu;
 492
 493         if (rq->internal_tag != -1)
 494                 blk_mq_sched_completed_request(rq);
 495         if (rq->rq_flags & RQF_STATS) {
 496                 blk_mq_poll_stats_start(rq->q);
 497                 blk_stat_add(rq);
 498         }
 499
 500         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 501                 rq->q->softirq_done_fn(rq);
 502                 return;
 503         }
 504
 505         cpu = get_cpu();
 506         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 507                 shared = cpus_share_cache(cpu, ctx->cpu);
 508
 509         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 510                 rq->csd.func = __blk_mq_complete_request_remote;
 511                 rq->csd.info = rq;
 512                 rq->csd.flags = 0;
 513                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 514         } else {
 515                 rq->q->softirq_done_fn(rq);
 516         }
 517         put_cpu();
 518 }
 519
 520 /**
 521  * blk_mq_complete_request - end I/O on a request
 522  * @rq:         the request being processed
 523  *
 524  * Description:
 525  *      Ends all I/O on a request. It does not handle partial completions.
 526  *      The actual completion happens out-of-order, through a IPI handler.
 527  **/
 528 void blk_mq_complete_request(struct request *rq)
 529 {
 530         struct request_queue *q = rq->q;
 531
 532         if (unlikely(blk_should_fake_timeout(q)))
 533                 return;
 534         if (!blk_mark_rq_complete(rq))
 535                 __blk_mq_complete_request(rq);
 536 }
 537 EXPORT_SYMBOL(blk_mq_complete_request);
 538
 539 int blk_mq_request_started(struct request *rq)
 540 {
 541         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 542 }
 543 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 544
 545 void blk_mq_start_request(struct request *rq)
 546 {
 547         struct request_queue *q = rq->q;
 548
 549         blk_mq_sched_started_request(rq);
 550
 551         trace_block_rq_issue(q, rq);
 552
 553         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 554                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
 555                 rq->rq_flags |= RQF_STATS;
 556                 wbt_issue(q->rq_wb, &rq->issue_stat);
 557         }
 558
 559         blk_add_timer(rq);
 560
 561         /*
 562          * Ensure that ->deadline is visible before set the started
 563          * flag and clear the completed flag.
 564          */
 565         smp_mb__before_atomic();
 566
 567         /*
 568          * Mark us as started and clear complete. Complete might have been
 569          * set if requeue raced with timeout, which then marked it as
 570          * complete. So be sure to clear complete again when we start
 571          * the request, otherwise we'll ignore the completion event.
 572          */
 573         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 574                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 575         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 576                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 577
 578         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 579                 /*
 580                  * Make sure space for the drain appears.  We know we can do
 581                  * this because max_hw_segments has been adjusted to be one
 582                  * fewer than the device can handle.
 583                  */
 584                 rq->nr_phys_segments++;
 585         }
 586 }
 587 EXPORT_SYMBOL(blk_mq_start_request);
 588
 589 /*
 590  * When we reach here because queue is busy, REQ_ATOM_COMPLETE
 591  * flag isn't set yet, so there may be race with timeout handler,
 592  * but given rq->deadline is just set in .queue_rq() under
 593  * this situation, the race won't be possible in reality because
 594  * rq->timeout should be set as big enough to cover the window
 595  * between blk_mq_start_request() called from .queue_rq() and
 596  * clearing REQ_ATOM_STARTED here.
 597  */
 598 static void __blk_mq_requeue_request(struct request *rq)
 599 {
 600         struct request_queue *q = rq->q;
 601
 602         trace_block_rq_requeue(q, rq);
 603         wbt_requeue(q->rq_wb, &rq->issue_stat);
 604         blk_mq_sched_requeue_request(rq);
 605
 606         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 607                 if (q->dma_drain_size && blk_rq_bytes(rq))
 608                         rq->nr_phys_segments--;
 609         }
 610 }
 611
 612 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 613 {
 614         __blk_mq_requeue_request(rq);
 615
 616         BUG_ON(blk_queued_rq(rq));
 617         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 618 }
 619 EXPORT_SYMBOL(blk_mq_requeue_request);
 620
 621 static void blk_mq_requeue_work(struct work_struct *work)
 622 {
 623         struct request_queue *q =
 624                 container_of(work, struct request_queue, requeue_work.work);
 625         LIST_HEAD(rq_list);
 626         struct request *rq, *next;
 627         unsigned long flags;
 628
 629         spin_lock_irqsave(&q->requeue_lock, flags);
 630         list_splice_init(&q->requeue_list, &rq_list);
 631         spin_unlock_irqrestore(&q->requeue_lock, flags);
 632
 633         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 634                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 635                         continue;
 636
 637                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 638                 list_del_init(&rq->queuelist);
 639                 blk_mq_sched_insert_request(rq, true, false, false, true);
 640         }
 641
 642         while (!list_empty(&rq_list)) {
 643                 rq = list_entry(rq_list.next, struct request, queuelist);
 644                 list_del_init(&rq->queuelist);
 645                 blk_mq_sched_insert_request(rq, false, false, false, true);
 646         }
 647
 648         blk_mq_run_hw_queues(q, false);
 649 }
 650
 651 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 652                                 bool kick_requeue_list)
 653 {
 654         struct request_queue *q = rq->q;
 655         unsigned long flags;
 656
 657         /*
 658          * We abuse this flag that is otherwise used by the I/O scheduler to
 659          * request head insertation from the workqueue.
 660          */
 661         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 662
 663         spin_lock_irqsave(&q->requeue_lock, flags);
 664         if (at_head) {
 665                 rq->rq_flags |= RQF_SOFTBARRIER;
 666                 list_add(&rq->queuelist, &q->requeue_list);
 667         } else {
 668                 list_add_tail(&rq->queuelist, &q->requeue_list);
 669         }
 670         spin_unlock_irqrestore(&q->requeue_lock, flags);
 671
 672         if (kick_requeue_list)
 673                 blk_mq_kick_requeue_list(q);
 674 }
 675 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 676
 677 void blk_mq_kick_requeue_list(struct request_queue *q)
 678 {
 679         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 680 }
 681 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 682
 683 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 684                                     unsigned long msecs)
 685 {
 686         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 687                                     msecs_to_jiffies(msecs));
 688 }
 689 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 690
 691 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 692 {
 693         if (tag < tags->nr_tags) {
 694                 prefetch(tags->rqs[tag]);
 695                 return tags->rqs[tag];
 696         }
 697
 698         return NULL;
 699 }
 700 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 701
 702 struct blk_mq_timeout_data {
 703         unsigned long next;
 704         unsigned int next_set;
 705 };
 706
 707 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 708 {
 709         const struct blk_mq_ops *ops = req->q->mq_ops;
 710         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 711
 712         /*
 713          * We know that complete is set at this point. If STARTED isn't set
 714          * anymore, then the request isn't active and the "timeout" should
 715          * just be ignored. This can happen due to the bitflag ordering.
 716          * Timeout first checks if STARTED is set, and if it is, assumes
 717          * the request is active. But if we race with completion, then
 718          * both flags will get cleared. So check here again, and ignore
 719          * a timeout event with a request that isn't active.
 720          */
 721         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 722                 return;
 723
 724         if (ops->timeout)
 725                 ret = ops->timeout(req, reserved);
 726
 727         switch (ret) {
 728         case BLK_EH_HANDLED:
 729                 __blk_mq_complete_request(req);
 730                 break;
 731         case BLK_EH_RESET_TIMER:
 732                 blk_add_timer(req);
 733                 blk_clear_rq_complete(req);
 734                 break;
 735         case BLK_EH_NOT_HANDLED:
 736                 break;
 737         default:
 738                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 739                 break;
 740         }
 741 }
 742
 743 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 744                 struct request *rq, void *priv, bool reserved)
 745 {
 746         struct blk_mq_timeout_data *data = priv;
 747
 748         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 749                 return;
 750
 751         /*
 752          * The rq being checked may have been freed and reallocated
 753          * out already here, we avoid this race by checking rq->deadline
 754          * and REQ_ATOM_COMPLETE flag together:
 755          *
 756          * - if rq->deadline is observed as new value because of
 757          *   reusing, the rq won't be timed out because of timing.
 758          * - if rq->deadline is observed as previous value,
 759          *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
 760          *   because we put a barrier between setting rq->deadline
 761          *   and clearing the flag in blk_mq_start_request(), so
 762          *   this rq won't be timed out too.
 763          */
 764         if (time_after_eq(jiffies, rq->deadline)) {
 765                 if (!blk_mark_rq_complete(rq))
 766                         blk_mq_rq_timed_out(rq, reserved);
 767         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 768                 data->next = rq->deadline;
 769                 data->next_set = 1;
 770         }
 771 }
 772
 773 static void blk_mq_timeout_work(struct work_struct *work)
 774 {
 775         struct request_queue *q =
 776                 container_of(work, struct request_queue, timeout_work);
 777         struct blk_mq_timeout_data data = {
 778                 .next           = 0,
 779                 .next_set       = 0,
 780         };
 781         int i;
 782
 783         /* A deadlock might occur if a request is stuck requiring a
 784          * timeout at the same time a queue freeze is waiting
 785          * completion, since the timeout code would not be able to
 786          * acquire the queue reference here.
 787          *
 788          * That's why we don't use blk_queue_enter here; instead, we use
 789          * percpu_ref_tryget directly, because we need to be able to
 790          * obtain a reference even in the short window between the queue
 791          * starting to freeze, by dropping the first reference in
 792          * blk_freeze_queue_start, and the moment the last request is
 793          * consumed, marked by the instant q_usage_counter reaches
 794          * zero.
 795          */
 796         if (!percpu_ref_tryget(&q->q_usage_counter))
 797                 return;
 798
 799         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 800
 801         if (data.next_set) {
 802                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 803                 mod_timer(&q->timeout, data.next);
 804         } else {
 805                 struct blk_mq_hw_ctx *hctx;
 806
 807                 queue_for_each_hw_ctx(q, hctx, i) {
 808                         /* the hctx may be unmapped, so check it here */
 809                         if (blk_mq_hw_queue_mapped(hctx))
 810                                 blk_mq_tag_idle(hctx);
 811                 }
 812         }
 813         blk_queue_exit(q);
 814 }
 815
 816 struct flush_busy_ctx_data {
 817         struct blk_mq_hw_ctx *hctx;
 818         struct list_head *list;
 819 };
 820
 821 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 822 {
 823         struct flush_busy_ctx_data *flush_data = data;
 824         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 825         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 826
 827         sbitmap_clear_bit(sb, bitnr);
 828         spin_lock(&ctx->lock);
 829         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 830         spin_unlock(&ctx->lock);
 831         return true;
 832 }
 833
 834 /*
 835  * Process software queues that have been marked busy, splicing them
 836  * to the for-dispatch
 837  */
 838 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 839 {
 840         struct flush_busy_ctx_data data = {
 841                 .hctx = hctx,
 842                 .list = list,
 843         };
 844
 845         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 846 }
 847 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 848
 849 static inline unsigned int queued_to_index(unsigned int queued)
 850 {
 851         if (!queued)
 852                 return 0;
 853
 854         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 855 }
 856
 857 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 858                            bool wait)
 859 {
 860         struct blk_mq_alloc_data data = {
 861                 .q = rq->q,
 862                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 863                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 864         };
 865
 866         might_sleep_if(wait);
 867
 868         if (rq->tag != -1)
 869                 goto done;
 870
 871         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 872                 data.flags |= BLK_MQ_REQ_RESERVED;
 873
 874         rq->tag = blk_mq_get_tag(&data);
 875         if (rq->tag >= 0) {
 876                 if (blk_mq_tag_busy(data.hctx)) {
 877                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 878                         atomic_inc(&data.hctx->nr_active);
 879                 }
 880                 data.hctx->tags->rqs[rq->tag] = rq;
 881         }
 882
 883 done:
 884         if (hctx)
 885                 *hctx = data.hctx;
 886         return rq->tag != -1;
 887 }
 888
 889 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 890                                     struct request *rq)
 891 {
 892         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
 893         rq->tag = -1;
 894
 895         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 896                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
 897                 atomic_dec(&hctx->nr_active);
 898         }
 899 }
 900
 901 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
 902                                        struct request *rq)
 903 {
 904         if (rq->tag == -1 || rq->internal_tag == -1)
 905                 return;
 906
 907         __blk_mq_put_driver_tag(hctx, rq);
 908 }
 909
 910 static void blk_mq_put_driver_tag(struct request *rq)
 911 {
 912         struct blk_mq_hw_ctx *hctx;
 913
 914         if (rq->tag == -1 || rq->internal_tag == -1)
 915                 return;
 916
 917         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
 918         __blk_mq_put_driver_tag(hctx, rq);
 919 }
 920
 921 /*
 922  * If we fail getting a driver tag because all the driver tags are already
 923  * assigned and on the dispatch list, BUT the first entry does not have a
 924  * tag, then we could deadlock. For that case, move entries with assigned
 925  * driver tags to the front, leaving the set of tagged requests in the
 926  * same order, and the untagged set in the same order.
 927  */
 928 static bool reorder_tags_to_front(struct list_head *list)
 929 {
 930         struct request *rq, *tmp, *first = NULL;
 931
 932         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
 933                 if (rq == first)
 934                         break;
 935                 if (rq->tag != -1) {
 936                         list_move(&rq->queuelist, list);
 937                         if (!first)
 938                                 first = rq;
 939                 }
 940         }
 941
 942         return first != NULL;
 943 }
 944
 945 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
 946                                 void *key)
 947 {
 948         struct blk_mq_hw_ctx *hctx;
 949
 950         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
 951
 952         list_del(&wait->entry);
 953         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
 954         blk_mq_run_hw_queue(hctx, true);
 955         return 1;
 956 }
 957
 958 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
 959 {
 960         struct sbq_wait_state *ws;
 961
 962         /*
 963          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
 964          * The thread which wins the race to grab this bit adds the hardware
 965          * queue to the wait queue.
 966          */
 967         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
 968             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
 969                 return false;
 970
 971         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
 972         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
 973
 974         /*
 975          * As soon as this returns, it's no longer safe to fiddle with
 976          * hctx->dispatch_wait, since a completion can wake up the wait queue
 977          * and unlock the bit.
 978          */
 979         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
 980         return true;
 981 }
 982
 983 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
 984 {
 985         struct blk_mq_hw_ctx *hctx;
 986         struct request *rq;
 987         int errors, queued;
 988
 989         if (list_empty(list))
 990                 return false;
 991
 992         /*
 993          * Now process all the entries, sending them to the driver.
 994          */
 995         errors = queued = 0;
 996         do {
 997                 struct blk_mq_queue_data bd;
 998                 blk_status_t ret;
 999
1000                 rq = list_first_entry(list, struct request, queuelist);
1001                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1002                         if (!queued && reorder_tags_to_front(list))
1003                                 continue;
1004
1005                         /*
1006                          * The initial allocation attempt failed, so we need to
1007                          * rerun the hardware queue when a tag is freed.
1008                          */
1009                         if (!blk_mq_dispatch_wait_add(hctx))
1010                                 break;
1011
1012                         /*
1013                          * It's possible that a tag was freed in the window
1014                          * between the allocation failure and adding the
1015                          * hardware queue to the wait queue.
1016                          */
1017                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1018                                 break;
1019                 }
1020
1021                 list_del_init(&rq->queuelist);
1022
1023                 bd.rq = rq;
1024
1025                 /*
1026                  * Flag last if we have no more requests, or if we have more
1027                  * but can't assign a driver tag to it.
1028                  */
1029                 if (list_empty(list))
1030                         bd.last = true;
1031                 else {
1032                         struct request *nxt;
1033
1034                         nxt = list_first_entry(list, struct request, queuelist);
1035                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1036                 }
1037
1038                 ret = q->mq_ops->queue_rq(hctx, &bd);
1039                 if (ret == BLK_STS_RESOURCE) {
1040                         blk_mq_put_driver_tag_hctx(hctx, rq);
1041                         list_add(&rq->queuelist, list);
1042                         __blk_mq_requeue_request(rq);
1043                         break;
1044                 }
1045
1046                 if (unlikely(ret != BLK_STS_OK)) {
1047                         errors++;
1048                         blk_mq_end_request(rq, BLK_STS_IOERR);
1049                         continue;
1050                 }
1051
1052                 queued++;
1053         } while (!list_empty(list));
1054
1055         hctx->dispatched[queued_to_index(queued)]++;
1056
1057         /*
1058          * Any items that need requeuing? Stuff them into hctx->dispatch,
1059          * that is where we will continue on next queue run.
1060          */
1061         if (!list_empty(list)) {
1062                 /*
1063                  * If an I/O scheduler has been configured and we got a driver
1064                  * tag for the next request already, free it again.
1065                  */
1066                 rq = list_first_entry(list, struct request, queuelist);
1067                 blk_mq_put_driver_tag(rq);
1068
1069                 spin_lock(&hctx->lock);
1070                 list_splice_init(list, &hctx->dispatch);
1071                 spin_unlock(&hctx->lock);
1072
1073                 /*
1074                  * If SCHED_RESTART was set by the caller of this function and
1075                  * it is no longer set that means that it was cleared by another
1076                  * thread and hence that a queue rerun is needed.
1077                  *
1078                  * If TAG_WAITING is set that means that an I/O scheduler has
1079                  * been configured and another thread is waiting for a driver
1080                  * tag. To guarantee fairness, do not rerun this hardware queue
1081                  * but let the other thread grab the driver tag.
1082                  *
1083                  * If no I/O scheduler has been configured it is possible that
1084                  * the hardware queue got stopped and restarted before requests
1085                  * were pushed back onto the dispatch list. Rerun the queue to
1086                  * avoid starvation. Notes:
1087                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1088                  *   been stopped before rerunning a queue.
1089                  * - Some but not all block drivers stop a queue before
1090                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1091                  *   and dm-rq.
1092                  */
1093                 if (!blk_mq_sched_needs_restart(hctx) &&
1094                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1095                         blk_mq_run_hw_queue(hctx, true);
1096         }
1097
1098         return (queued + errors) != 0;
1099 }
1100
1101 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1102 {
1103         int srcu_idx;
1104
1105         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1106                 cpu_online(hctx->next_cpu));
1107
1108         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1109                 rcu_read_lock();
1110                 blk_mq_sched_dispatch_requests(hctx);
1111                 rcu_read_unlock();
1112         } else {
1113                 might_sleep();
1114
1115                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1116                 blk_mq_sched_dispatch_requests(hctx);
1117                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1118         }
1119 }
1120
1121 /*
1122  * It'd be great if the workqueue API had a way to pass
1123  * in a mask and had some smarts for more clever placement.
1124  * For now we just round-robin here, switching for every
1125  * BLK_MQ_CPU_WORK_BATCH queued items.
1126  */
1127 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1128 {
1129         if (hctx->queue->nr_hw_queues == 1)
1130                 return WORK_CPU_UNBOUND;
1131
1132         if (--hctx->next_cpu_batch <= 0) {
1133                 int next_cpu;
1134
1135                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1136                 if (next_cpu >= nr_cpu_ids)
1137                         next_cpu = cpumask_first(hctx->cpumask);
1138
1139                 hctx->next_cpu = next_cpu;
1140                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1141         }
1142
1143         return hctx->next_cpu;
1144 }
1145
1146 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1147                                         unsigned long msecs)
1148 {
1149         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1150                 return;
1151
1152         if (unlikely(blk_mq_hctx_stopped(hctx)))
1153                 return;
1154
1155         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1156                 int cpu = get_cpu();
1157                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1158                         __blk_mq_run_hw_queue(hctx);
1159                         put_cpu();
1160                         return;
1161                 }
1162
1163                 put_cpu();
1164         }
1165
1166         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1167                                          &hctx->run_work,
1168                                          msecs_to_jiffies(msecs));
1169 }
1170
1171 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1172 {
1173         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1174 }
1175 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1176
1177 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1178 {
1179         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1180 }
1181 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1182
1183 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1184 {
1185         struct blk_mq_hw_ctx *hctx;
1186         int i;
1187
1188         queue_for_each_hw_ctx(q, hctx, i) {
1189                 if (!blk_mq_hctx_has_pending(hctx) ||
1190                     blk_mq_hctx_stopped(hctx))
1191                         continue;
1192
1193                 blk_mq_run_hw_queue(hctx, async);
1194         }
1195 }
1196 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1197
1198 /**
1199  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1200  * @q: request queue.
1201  *
1202  * The caller is responsible for serializing this function against
1203  * blk_mq_{start,stop}_hw_queue().
1204  */
1205 bool blk_mq_queue_stopped(struct request_queue *q)
1206 {
1207         struct blk_mq_hw_ctx *hctx;
1208         int i;
1209
1210         queue_for_each_hw_ctx(q, hctx, i)
1211                 if (blk_mq_hctx_stopped(hctx))
1212                         return true;
1213
1214         return false;
1215 }
1216 EXPORT_SYMBOL(blk_mq_queue_stopped);
1217
1218 /*
1219  * This function is often used for pausing .queue_rq() by driver when
1220  * there isn't enough resource or some conditions aren't satisfied, and
1221  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1222  *
1223  * We do not guarantee that dispatch can be drained or blocked
1224  * after blk_mq_stop_hw_queue() returns. Please use
1225  * blk_mq_quiesce_queue() for that requirement.
1226  */
1227 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1228 {
1229         cancel_delayed_work(&hctx->run_work);
1230
1231         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1232 }
1233 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1234
1235 /*
1236  * This function is often used for pausing .queue_rq() by driver when
1237  * there isn't enough resource or some conditions aren't satisfied, and
1238  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1239  *
1240  * We do not guarantee that dispatch can be drained or blocked
1241  * after blk_mq_stop_hw_queues() returns. Please use
1242  * blk_mq_quiesce_queue() for that requirement.
1243  */
1244 void blk_mq_stop_hw_queues(struct request_queue *q)
1245 {
1246         struct blk_mq_hw_ctx *hctx;
1247         int i;
1248
1249         queue_for_each_hw_ctx(q, hctx, i)
1250                 blk_mq_stop_hw_queue(hctx);
1251 }
1252 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1253
1254 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1255 {
1256         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1257
1258         blk_mq_run_hw_queue(hctx, false);
1259 }
1260 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1261
1262 void blk_mq_start_hw_queues(struct request_queue *q)
1263 {
1264         struct blk_mq_hw_ctx *hctx;
1265         int i;
1266
1267         queue_for_each_hw_ctx(q, hctx, i)
1268                 blk_mq_start_hw_queue(hctx);
1269 }
1270 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1271
1272 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1273 {
1274         if (!blk_mq_hctx_stopped(hctx))
1275                 return;
1276
1277         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1278         blk_mq_run_hw_queue(hctx, async);
1279 }
1280 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1281
1282 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1283 {
1284         struct blk_mq_hw_ctx *hctx;
1285         int i;
1286
1287         queue_for_each_hw_ctx(q, hctx, i)
1288                 blk_mq_start_stopped_hw_queue(hctx, async);
1289 }
1290 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1291
1292 static void blk_mq_run_work_fn(struct work_struct *work)
1293 {
1294         struct blk_mq_hw_ctx *hctx;
1295
1296         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1297
1298         /*
1299          * If we are stopped, don't run the queue. The exception is if
1300          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1301          * the STOPPED bit and run it.
1302          */
1303         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1304                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1305                         return;
1306
1307                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1308                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1309         }
1310
1311         __blk_mq_run_hw_queue(hctx);
1312 }
1313
1314
1315 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1316 {
1317         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1318                 return;
1319
1320         /*
1321          * Stop the hw queue, then modify currently delayed work.
1322          * This should prevent us from running the queue prematurely.
1323          * Mark the queue as auto-clearing STOPPED when it runs.
1324          */
1325         blk_mq_stop_hw_queue(hctx);
1326         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1327         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1328                                         &hctx->run_work,
1329                                         msecs_to_jiffies(msecs));
1330 }
1331 EXPORT_SYMBOL(blk_mq_delay_queue);
1332
1333 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1334                                             struct request *rq,
1335                                             bool at_head)
1336 {
1337         struct blk_mq_ctx *ctx = rq->mq_ctx;
1338
1339         lockdep_assert_held(&ctx->lock);
1340
1341         trace_block_rq_insert(hctx->queue, rq);
1342
1343         if (at_head)
1344                 list_add(&rq->queuelist, &ctx->rq_list);
1345         else
1346                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1347 }
1348
1349 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1350                              bool at_head)
1351 {
1352         struct blk_mq_ctx *ctx = rq->mq_ctx;
1353
1354         lockdep_assert_held(&ctx->lock);
1355
1356         __blk_mq_insert_req_list(hctx, rq, at_head);
1357         blk_mq_hctx_mark_pending(hctx, ctx);
1358 }
1359
1360 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1361                             struct list_head *list)
1362
1363 {
1364         /*
1365          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1366          * offline now
1367          */
1368         spin_lock(&ctx->lock);
1369         while (!list_empty(list)) {
1370                 struct request *rq;
1371
1372                 rq = list_first_entry(list, struct request, queuelist);
1373                 BUG_ON(rq->mq_ctx != ctx);
1374                 list_del_init(&rq->queuelist);
1375                 __blk_mq_insert_req_list(hctx, rq, false);
1376         }
1377         blk_mq_hctx_mark_pending(hctx, ctx);
1378         spin_unlock(&ctx->lock);
1379 }
1380
1381 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1382 {
1383         struct request *rqa = container_of(a, struct request, queuelist);
1384         struct request *rqb = container_of(b, struct request, queuelist);
1385
1386         return !(rqa->mq_ctx < rqb->mq_ctx ||
1387                  (rqa->mq_ctx == rqb->mq_ctx &&
1388                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1389 }
1390
1391 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1392 {
1393         struct blk_mq_ctx *this_ctx;
1394         struct request_queue *this_q;
1395         struct request *rq;
1396         LIST_HEAD(list);
1397         LIST_HEAD(ctx_list);
1398         unsigned int depth;
1399
1400         list_splice_init(&plug->mq_list, &list);
1401
1402         list_sort(NULL, &list, plug_ctx_cmp);
1403
1404         this_q = NULL;
1405         this_ctx = NULL;
1406         depth = 0;
1407
1408         while (!list_empty(&list)) {
1409                 rq = list_entry_rq(list.next);
1410                 list_del_init(&rq->queuelist);
1411                 BUG_ON(!rq->q);
1412                 if (rq->mq_ctx != this_ctx) {
1413                         if (this_ctx) {
1414                                 trace_block_unplug(this_q, depth, from_schedule);
1415                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1416                                                                 &ctx_list,
1417                                                                 from_schedule);
1418                         }
1419
1420                         this_ctx = rq->mq_ctx;
1421                         this_q = rq->q;
1422                         depth = 0;
1423                 }
1424
1425                 depth++;
1426                 list_add_tail(&rq->queuelist, &ctx_list);
1427         }
1428
1429         /*
1430          * If 'this_ctx' is set, we know we have entries to complete
1431          * on 'ctx_list'. Do those.
1432          */
1433         if (this_ctx) {
1434                 trace_block_unplug(this_q, depth, from_schedule);
1435                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1436                                                 from_schedule);
1437         }
1438 }
1439
1440 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1441 {
1442         blk_init_request_from_bio(rq, bio);
1443
1444         blk_account_io_start(rq, true);
1445 }
1446
1447 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1448 {
1449         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1450                 !blk_queue_nomerges(hctx->queue);
1451 }
1452
1453 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1454                                    struct blk_mq_ctx *ctx,
1455                                    struct request *rq)
1456 {
1457         spin_lock(&ctx->lock);
1458         __blk_mq_insert_request(hctx, rq, false);
1459         spin_unlock(&ctx->lock);
1460 }
1461
1462 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1463 {
1464         if (rq->tag != -1)
1465                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1466
1467         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1468 }
1469
1470 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1471                                         struct request *rq,
1472                                         blk_qc_t *cookie, bool may_sleep)
1473 {
1474         struct request_queue *q = rq->q;
1475         struct blk_mq_queue_data bd = {
1476                 .rq = rq,
1477                 .last = true,
1478         };
1479         blk_qc_t new_cookie;
1480         blk_status_t ret;
1481         bool run_queue = true;
1482
1483         /* RCU or SRCU read lock is needed before checking quiesced flag */
1484         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1485                 run_queue = false;
1486                 goto insert;
1487         }
1488
1489         if (q->elevator)
1490                 goto insert;
1491
1492         if (!blk_mq_get_driver_tag(rq, NULL, false))
1493                 goto insert;
1494
1495         new_cookie = request_to_qc_t(hctx, rq);
1496
1497         /*
1498          * For OK queue, we are done. For error, kill it. Any other
1499          * error (busy), just add it to our list as we previously
1500          * would have done
1501          */
1502         ret = q->mq_ops->queue_rq(hctx, &bd);
1503         switch (ret) {
1504         case BLK_STS_OK:
1505                 *cookie = new_cookie;
1506                 return;
1507         case BLK_STS_RESOURCE:
1508                 __blk_mq_requeue_request(rq);
1509                 goto insert;
1510         default:
1511                 *cookie = BLK_QC_T_NONE;
1512                 blk_mq_end_request(rq, ret);
1513                 return;
1514         }
1515
1516 insert:
1517         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1518 }
1519
1520 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1521                 struct request *rq, blk_qc_t *cookie)
1522 {
1523         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1524                 rcu_read_lock();
1525                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1526                 rcu_read_unlock();
1527         } else {
1528                 unsigned int srcu_idx;
1529
1530                 might_sleep();
1531
1532                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1533                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1534                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1535         }
1536 }
1537
1538 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1539 {
1540         const int is_sync = op_is_sync(bio->bi_opf);
1541         const int is_flush_fua = op_is_flush(bio->bi_opf);
1542         struct blk_mq_alloc_data data = { .flags = 0 };
1543         struct request *rq;
1544         unsigned int request_count = 0;
1545         struct blk_plug *plug;
1546         struct request *same_queue_rq = NULL;
1547         blk_qc_t cookie;
1548         unsigned int wb_acct;
1549
1550         blk_queue_bounce(q, &bio);
1551
1552         blk_queue_split(q, &bio);
1553
1554         if (!bio_integrity_prep(bio))
1555                 return BLK_QC_T_NONE;
1556
1557         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1558             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1559                 return BLK_QC_T_NONE;
1560
1561         if (blk_mq_sched_bio_merge(q, bio))
1562                 return BLK_QC_T_NONE;
1563
1564         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1565
1566         trace_block_getrq(q, bio, bio->bi_opf);
1567
1568         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1569         if (unlikely(!rq)) {
1570                 __wbt_done(q->rq_wb, wb_acct);
1571                 if (bio->bi_opf & REQ_NOWAIT)
1572                         bio_wouldblock_error(bio);
1573                 return BLK_QC_T_NONE;
1574         }
1575
1576         wbt_track(&rq->issue_stat, wb_acct);
1577
1578         cookie = request_to_qc_t(data.hctx, rq);
1579
1580         plug = current->plug;
1581         if (unlikely(is_flush_fua)) {
1582                 blk_mq_put_ctx(data.ctx);
1583                 blk_mq_bio_to_request(rq, bio);
1584                 if (q->elevator) {
1585                         blk_mq_sched_insert_request(rq, false, true, true,
1586                                         true);
1587                 } else {
1588                         blk_insert_flush(rq);
1589                         blk_mq_run_hw_queue(data.hctx, true);
1590                 }
1591         } else if (plug && q->nr_hw_queues == 1) {
1592                 struct request *last = NULL;
1593
1594                 blk_mq_put_ctx(data.ctx);
1595                 blk_mq_bio_to_request(rq, bio);
1596
1597                 /*
1598                  * @request_count may become stale because of schedule
1599                  * out, so check the list again.
1600                  */
1601                 if (list_empty(&plug->mq_list))
1602                         request_count = 0;
1603                 else if (blk_queue_nomerges(q))
1604                         request_count = blk_plug_queued_count(q);
1605
1606                 if (!request_count)
1607                         trace_block_plug(q);
1608                 else
1609                         last = list_entry_rq(plug->mq_list.prev);
1610
1611                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1612                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1613                         blk_flush_plug_list(plug, false);
1614                         trace_block_plug(q);
1615                 }
1616
1617                 list_add_tail(&rq->queuelist, &plug->mq_list);
1618         } else if (plug && !blk_queue_nomerges(q)) {
1619                 blk_mq_bio_to_request(rq, bio);
1620
1621                 /*
1622                  * We do limited plugging. If the bio can be merged, do that.
1623                  * Otherwise the existing request in the plug list will be
1624                  * issued. So the plug list will have one request at most
1625                  * The plug list might get flushed before this. If that happens,
1626                  * the plug list is empty, and same_queue_rq is invalid.
1627                  */
1628                 if (list_empty(&plug->mq_list))
1629                         same_queue_rq = NULL;
1630                 if (same_queue_rq)
1631                         list_del_init(&same_queue_rq->queuelist);
1632                 list_add_tail(&rq->queuelist, &plug->mq_list);
1633
1634                 blk_mq_put_ctx(data.ctx);
1635
1636                 if (same_queue_rq) {
1637                         data.hctx = blk_mq_map_queue(q,
1638                                         same_queue_rq->mq_ctx->cpu);
1639                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1640                                         &cookie);
1641                 }
1642         } else if (q->nr_hw_queues > 1 && is_sync) {
1643                 blk_mq_put_ctx(data.ctx);
1644                 blk_mq_bio_to_request(rq, bio);
1645                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1646         } else if (q->elevator) {
1647                 blk_mq_put_ctx(data.ctx);
1648                 blk_mq_bio_to_request(rq, bio);
1649                 blk_mq_sched_insert_request(rq, false, true, true, true);
1650         } else {
1651                 blk_mq_put_ctx(data.ctx);
1652                 blk_mq_bio_to_request(rq, bio);
1653                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1654                 blk_mq_run_hw_queue(data.hctx, true);
1655         }
1656
1657         return cookie;
1658 }
1659
1660 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1661                      unsigned int hctx_idx)
1662 {
1663         struct page *page;
1664
1665         if (tags->rqs && set->ops->exit_request) {
1666                 int i;
1667
1668                 for (i = 0; i < tags->nr_tags; i++) {
1669                         struct request *rq = tags->static_rqs[i];
1670
1671                         if (!rq)
1672                                 continue;
1673                         set->ops->exit_request(set, rq, hctx_idx);
1674                         tags->static_rqs[i] = NULL;
1675                 }
1676         }
1677
1678         while (!list_empty(&tags->page_list)) {
1679                 page = list_first_entry(&tags->page_list, struct page, lru);
1680                 list_del_init(&page->lru);
1681                 /*
1682                  * Remove kmemleak object previously allocated in
1683                  * blk_mq_init_rq_map().
1684                  */
1685                 kmemleak_free(page_address(page));
1686                 __free_pages(page, page->private);
1687         }
1688 }
1689
1690 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1691 {
1692         kfree(tags->rqs);
1693         tags->rqs = NULL;
1694         kfree(tags->static_rqs);
1695         tags->static_rqs = NULL;
1696
1697         blk_mq_free_tags(tags);
1698 }
1699
1700 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1701                                         unsigned int hctx_idx,
1702                                         unsigned int nr_tags,
1703                                         unsigned int reserved_tags)
1704 {
1705         struct blk_mq_tags *tags;
1706         int node;
1707
1708         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1709         if (node == NUMA_NO_NODE)
1710                 node = set->numa_node;
1711
1712         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1713                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1714         if (!tags)
1715                 return NULL;
1716
1717         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1718                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1719                                  node);
1720         if (!tags->rqs) {
1721                 blk_mq_free_tags(tags);
1722                 return NULL;
1723         }
1724
1725         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1726                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1727                                  node);
1728         if (!tags->static_rqs) {
1729                 kfree(tags->rqs);
1730                 blk_mq_free_tags(tags);
1731                 return NULL;
1732         }
1733
1734         return tags;
1735 }
1736
1737 static size_t order_to_size(unsigned int order)
1738 {
1739         return (size_t)PAGE_SIZE << order;
1740 }
1741
1742 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1743                      unsigned int hctx_idx, unsigned int depth)
1744 {
1745         unsigned int i, j, entries_per_page, max_order = 4;
1746         size_t rq_size, left;
1747         int node;
1748
1749         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1750         if (node == NUMA_NO_NODE)
1751                 node = set->numa_node;
1752
1753         INIT_LIST_HEAD(&tags->page_list);
1754
1755         /*
1756          * rq_size is the size of the request plus driver payload, rounded
1757          * to the cacheline size
1758          */
1759         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1760                                 cache_line_size());
1761         left = rq_size * depth;
1762
1763         for (i = 0; i < depth; ) {
1764                 int this_order = max_order;
1765                 struct page *page;
1766                 int to_do;
1767                 void *p;
1768
1769                 while (this_order && left < order_to_size(this_order - 1))
1770                         this_order--;
1771
1772                 do {
1773                         page = alloc_pages_node(node,
1774                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1775                                 this_order);
1776                         if (page)
1777                                 break;
1778                         if (!this_order--)
1779                                 break;
1780                         if (order_to_size(this_order) < rq_size)
1781                                 break;
1782                 } while (1);
1783
1784                 if (!page)
1785                         goto fail;
1786
1787                 page->private = this_order;
1788                 list_add_tail(&page->lru, &tags->page_list);
1789
1790                 p = page_address(page);
1791                 /*
1792                  * Allow kmemleak to scan these pages as they contain pointers
1793                  * to additional allocations like via ops->init_request().
1794                  */
1795                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1796                 entries_per_page = order_to_size(this_order) / rq_size;
1797                 to_do = min(entries_per_page, depth - i);
1798                 left -= to_do * rq_size;
1799                 for (j = 0; j < to_do; j++) {
1800                         struct request *rq = p;
1801
1802                         tags->static_rqs[i] = rq;
1803                         if (set->ops->init_request) {
1804                                 if (set->ops->init_request(set, rq, hctx_idx,
1805                                                 node)) {
1806                                         tags->static_rqs[i] = NULL;
1807                                         goto fail;
1808                                 }
1809                         }
1810
1811                         p += rq_size;
1812                         i++;
1813                 }
1814         }
1815         return 0;
1816
1817 fail:
1818         blk_mq_free_rqs(set, tags, hctx_idx);
1819         return -ENOMEM;
1820 }
1821
1822 /*
1823  * 'cpu' is going away. splice any existing rq_list entries from this
1824  * software queue to the hw queue dispatch list, and ensure that it
1825  * gets run.
1826  */
1827 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1828 {
1829         struct blk_mq_hw_ctx *hctx;
1830         struct blk_mq_ctx *ctx;
1831         LIST_HEAD(tmp);
1832
1833         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1834         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1835
1836         spin_lock(&ctx->lock);
1837         if (!list_empty(&ctx->rq_list)) {
1838                 list_splice_init(&ctx->rq_list, &tmp);
1839                 blk_mq_hctx_clear_pending(hctx, ctx);
1840         }
1841         spin_unlock(&ctx->lock);
1842
1843         if (list_empty(&tmp))
1844                 return 0;
1845
1846         spin_lock(&hctx->lock);
1847         list_splice_tail_init(&tmp, &hctx->dispatch);
1848         spin_unlock(&hctx->lock);
1849
1850         blk_mq_run_hw_queue(hctx, true);
1851         return 0;
1852 }
1853
1854 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1855 {
1856         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1857                                             &hctx->cpuhp_dead);
1858 }
1859
1860 /* hctx->ctxs will be freed in queue's release handler */
1861 static void blk_mq_exit_hctx(struct request_queue *q,
1862                 struct blk_mq_tag_set *set,
1863                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1864 {
1865         blk_mq_debugfs_unregister_hctx(hctx);
1866
1867         blk_mq_tag_idle(hctx);
1868
1869         if (set->ops->exit_request)
1870                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1871
1872         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1873
1874         if (set->ops->exit_hctx)
1875                 set->ops->exit_hctx(hctx, hctx_idx);
1876
1877         if (hctx->flags & BLK_MQ_F_BLOCKING)
1878                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1879
1880         blk_mq_remove_cpuhp(hctx);
1881         blk_free_flush_queue(hctx->fq);
1882         sbitmap_free(&hctx->ctx_map);
1883 }
1884
1885 static void blk_mq_exit_hw_queues(struct request_queue *q,
1886                 struct blk_mq_tag_set *set, int nr_queue)
1887 {
1888         struct blk_mq_hw_ctx *hctx;
1889         unsigned int i;
1890
1891         queue_for_each_hw_ctx(q, hctx, i) {
1892                 if (i == nr_queue)
1893                         break;
1894                 blk_mq_exit_hctx(q, set, hctx, i);
1895         }
1896 }
1897
1898 static int blk_mq_init_hctx(struct request_queue *q,
1899                 struct blk_mq_tag_set *set,
1900                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1901 {
1902         int node;
1903
1904         node = hctx->numa_node;
1905         if (node == NUMA_NO_NODE)
1906                 node = hctx->numa_node = set->numa_node;
1907
1908         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1909         spin_lock_init(&hctx->lock);
1910         INIT_LIST_HEAD(&hctx->dispatch);
1911         hctx->queue = q;
1912         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1913
1914         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1915
1916         hctx->tags = set->tags[hctx_idx];
1917
1918         /*
1919          * Allocate space for all possible cpus to avoid allocation at
1920          * runtime
1921          */
1922         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1923                                         GFP_KERNEL, node);
1924         if (!hctx->ctxs)
1925                 goto unregister_cpu_notifier;
1926
1927         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1928                               node))
1929                 goto free_ctxs;
1930
1931         hctx->nr_ctx = 0;
1932
1933         if (set->ops->init_hctx &&
1934             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1935                 goto free_bitmap;
1936
1937         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1938                 goto exit_hctx;
1939
1940         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1941         if (!hctx->fq)
1942                 goto sched_exit_hctx;
1943
1944         if (set->ops->init_request &&
1945             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1946                                    node))
1947                 goto free_fq;
1948
1949         if (hctx->flags & BLK_MQ_F_BLOCKING)
1950                 init_srcu_struct(hctx->queue_rq_srcu);
1951
1952         blk_mq_debugfs_register_hctx(q, hctx);
1953
1954         return 0;
1955
1956  free_fq:
1957         kfree(hctx->fq);
1958  sched_exit_hctx:
1959         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1960  exit_hctx:
1961         if (set->ops->exit_hctx)
1962                 set->ops->exit_hctx(hctx, hctx_idx);
1963  free_bitmap:
1964         sbitmap_free(&hctx->ctx_map);
1965  free_ctxs:
1966         kfree(hctx->ctxs);
1967  unregister_cpu_notifier:
1968         blk_mq_remove_cpuhp(hctx);
1969         return -1;
1970 }
1971
1972 static void blk_mq_init_cpu_queues(struct request_queue *q,
1973                                    unsigned int nr_hw_queues)
1974 {
1975         unsigned int i;
1976
1977         for_each_possible_cpu(i) {
1978                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1979                 struct blk_mq_hw_ctx *hctx;
1980
1981                 __ctx->cpu = i;
1982                 spin_lock_init(&__ctx->lock);
1983                 INIT_LIST_HEAD(&__ctx->rq_list);
1984                 __ctx->queue = q;
1985
1986                 /* If the cpu isn't present, the cpu is mapped to first hctx */
1987                 if (!cpu_present(i))
1988                         continue;
1989
1990                 hctx = blk_mq_map_queue(q, i);
1991
1992                 /*
1993                  * Set local node, IFF we have more than one hw queue. If
1994                  * not, we remain on the home node of the device
1995                  */
1996                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1997                         hctx->numa_node = local_memory_node(cpu_to_node(i));
1998         }
1999 }
2000
2001 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2002 {
2003         int ret = 0;
2004
2005         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2006                                         set->queue_depth, set->reserved_tags);
2007         if (!set->tags[hctx_idx])
2008                 return false;
2009
2010         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2011                                 set->queue_depth);
2012         if (!ret)
2013                 return true;
2014
2015         blk_mq_free_rq_map(set->tags[hctx_idx]);
2016         set->tags[hctx_idx] = NULL;
2017         return false;
2018 }
2019
2020 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2021                                          unsigned int hctx_idx)
2022 {
2023         if (set->tags[hctx_idx]) {
2024                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2025                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2026                 set->tags[hctx_idx] = NULL;
2027         }
2028 }
2029
2030 static void blk_mq_map_swqueue(struct request_queue *q)
2031 {
2032         unsigned int i, hctx_idx;
2033         struct blk_mq_hw_ctx *hctx;
2034         struct blk_mq_ctx *ctx;
2035         struct blk_mq_tag_set *set = q->tag_set;
2036
2037         /*
2038          * Avoid others reading imcomplete hctx->cpumask through sysfs
2039          */
2040         mutex_lock(&q->sysfs_lock);
2041
2042         queue_for_each_hw_ctx(q, hctx, i) {
2043                 cpumask_clear(hctx->cpumask);
2044                 hctx->nr_ctx = 0;
2045         }
2046
2047         /*
2048          * Map software to hardware queues.
2049          *
2050          * If the cpu isn't present, the cpu is mapped to first hctx.
2051          */
2052         for_each_present_cpu(i) {
2053                 hctx_idx = q->mq_map[i];
2054                 /* unmapped hw queue can be remapped after CPU topo changed */
2055                 if (!set->tags[hctx_idx] &&
2056                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2057                         /*
2058                          * If tags initialization fail for some hctx,
2059                          * that hctx won't be brought online.  In this
2060                          * case, remap the current ctx to hctx[0] which
2061                          * is guaranteed to always have tags allocated
2062                          */
2063                         q->mq_map[i] = 0;
2064                 }
2065
2066                 ctx = per_cpu_ptr(q->queue_ctx, i);
2067                 hctx = blk_mq_map_queue(q, i);
2068
2069                 cpumask_set_cpu(i, hctx->cpumask);
2070                 ctx->index_hw = hctx->nr_ctx;
2071                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2072         }
2073
2074         mutex_unlock(&q->sysfs_lock);
2075
2076         queue_for_each_hw_ctx(q, hctx, i) {
2077                 /*
2078                  * If no software queues are mapped to this hardware queue,
2079                  * disable it and free the request entries.
2080                  */
2081                 if (!hctx->nr_ctx) {
2082                         /* Never unmap queue 0.  We need it as a
2083                          * fallback in case of a new remap fails
2084                          * allocation
2085                          */
2086                         if (i && set->tags[i])
2087                                 blk_mq_free_map_and_requests(set, i);
2088
2089                         hctx->tags = NULL;
2090                         continue;
2091                 }
2092
2093                 hctx->tags = set->tags[i];
2094                 WARN_ON(!hctx->tags);
2095
2096                 /*
2097                  * Set the map size to the number of mapped software queues.
2098                  * This is more accurate and more efficient than looping
2099                  * over all possibly mapped software queues.
2100                  */
2101                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2102
2103                 /*
2104                  * Initialize batch roundrobin counts
2105                  */
2106                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2107                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2108         }
2109 }
2110
2111 /*
2112  * Caller needs to ensure that we're either frozen/quiesced, or that
2113  * the queue isn't live yet.
2114  */
2115 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2116 {
2117         struct blk_mq_hw_ctx *hctx;
2118         int i;
2119
2120         queue_for_each_hw_ctx(q, hctx, i) {
2121                 if (shared) {
2122                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2123                                 atomic_inc(&q->shared_hctx_restart);
2124                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2125                 } else {
2126                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2127                                 atomic_dec(&q->shared_hctx_restart);
2128                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2129                 }
2130         }
2131 }
2132
2133 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2134                                         bool shared)
2135 {
2136         struct request_queue *q;
2137
2138         lockdep_assert_held(&set->tag_list_lock);
2139
2140         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2141                 blk_mq_freeze_queue(q);
2142                 queue_set_hctx_shared(q, shared);
2143                 blk_mq_unfreeze_queue(q);
2144         }
2145 }
2146
2147 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2148 {
2149         struct blk_mq_tag_set *set = q->tag_set;
2150
2151         mutex_lock(&set->tag_list_lock);
2152         list_del_rcu(&q->tag_set_list);
2153         INIT_LIST_HEAD(&q->tag_set_list);
2154         if (list_is_singular(&set->tag_list)) {
2155                 /* just transitioned to unshared */
2156                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2157                 /* update existing queue */
2158                 blk_mq_update_tag_set_depth(set, false);
2159         }
2160         mutex_unlock(&set->tag_list_lock);
2161
2162         synchronize_rcu();
2163 }
2164
2165 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2166                                      struct request_queue *q)
2167 {
2168         q->tag_set = set;
2169
2170         mutex_lock(&set->tag_list_lock);
2171
2172         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2173         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2174                 set->flags |= BLK_MQ_F_TAG_SHARED;
2175                 /* update existing queue */
2176                 blk_mq_update_tag_set_depth(set, true);
2177         }
2178         if (set->flags & BLK_MQ_F_TAG_SHARED)
2179                 queue_set_hctx_shared(q, true);
2180         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2181
2182         mutex_unlock(&set->tag_list_lock);
2183 }
2184
2185 /*
2186  * It is the actual release handler for mq, but we do it from
2187  * request queue's release handler for avoiding use-after-free
2188  * and headache because q->mq_kobj shouldn't have been introduced,
2189  * but we can't group ctx/kctx kobj without it.
2190  */
2191 void blk_mq_release(struct request_queue *q)
2192 {
2193         struct blk_mq_hw_ctx *hctx;
2194         unsigned int i;
2195
2196         /* hctx kobj stays in hctx */
2197         queue_for_each_hw_ctx(q, hctx, i) {
2198                 if (!hctx)
2199                         continue;
2200                 kobject_put(&hctx->kobj);
2201         }
2202
2203         q->mq_map = NULL;
2204
2205         kfree(q->queue_hw_ctx);
2206
2207         /*
2208          * release .mq_kobj and sw queue's kobject now because
2209          * both share lifetime with request queue.
2210          */
2211         blk_mq_sysfs_deinit(q);
2212
2213         free_percpu(q->queue_ctx);
2214 }
2215
2216 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2217 {
2218         struct request_queue *uninit_q, *q;
2219
2220         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2221         if (!uninit_q)
2222                 return ERR_PTR(-ENOMEM);
2223
2224         q = blk_mq_init_allocated_queue(set, uninit_q);
2225         if (IS_ERR(q))
2226                 blk_cleanup_queue(uninit_q);
2227
2228         return q;
2229 }
2230 EXPORT_SYMBOL(blk_mq_init_queue);
2231
2232 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2233 {
2234         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2235
2236         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2237                            __alignof__(struct blk_mq_hw_ctx)) !=
2238                      sizeof(struct blk_mq_hw_ctx));
2239
2240         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2241                 hw_ctx_size += sizeof(struct srcu_struct);
2242
2243         return hw_ctx_size;
2244 }
2245
2246 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2247                                                 struct request_queue *q)
2248 {
2249         int i, j;
2250         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2251
2252         blk_mq_sysfs_unregister(q);
2253         for (i = 0; i < set->nr_hw_queues; i++) {
2254                 int node;
2255
2256                 if (hctxs[i])
2257                         continue;
2258
2259                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2260                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2261                                         GFP_KERNEL, node);
2262                 if (!hctxs[i])
2263                         break;
2264
2265                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2266                                                 node)) {
2267                         kfree(hctxs[i]);
2268                         hctxs[i] = NULL;
2269                         break;
2270                 }
2271
2272                 atomic_set(&hctxs[i]->nr_active, 0);
2273                 hctxs[i]->numa_node = node;
2274                 hctxs[i]->queue_num = i;
2275
2276                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2277                         free_cpumask_var(hctxs[i]->cpumask);
2278                         kfree(hctxs[i]);
2279                         hctxs[i] = NULL;
2280                         break;
2281                 }
2282                 blk_mq_hctx_kobj_init(hctxs[i]);
2283         }
2284         for (j = i; j < q->nr_hw_queues; j++) {
2285                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2286
2287                 if (hctx) {
2288                         if (hctx->tags)
2289                                 blk_mq_free_map_and_requests(set, j);
2290                         blk_mq_exit_hctx(q, set, hctx, j);
2291                         kobject_put(&hctx->kobj);
2292                         hctxs[j] = NULL;
2293
2294                 }
2295         }
2296         q->nr_hw_queues = i;
2297         blk_mq_sysfs_register(q);
2298 }
2299
2300 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2301                                                   struct request_queue *q)
2302 {
2303         /* mark the queue as mq asap */
2304         q->mq_ops = set->ops;
2305
2306         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2307                                              blk_mq_poll_stats_bkt,
2308                                              BLK_MQ_POLL_STATS_BKTS, q);
2309         if (!q->poll_cb)
2310                 goto err_exit;
2311
2312         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2313         if (!q->queue_ctx)
2314                 goto err_exit;
2315
2316         /* init q->mq_kobj and sw queues' kobjects */
2317         blk_mq_sysfs_init(q);
2318
2319         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2320                                                 GFP_KERNEL, set->numa_node);
2321         if (!q->queue_hw_ctx)
2322                 goto err_percpu;
2323
2324         q->mq_map = set->mq_map;
2325
2326         blk_mq_realloc_hw_ctxs(set, q);
2327         if (!q->nr_hw_queues)
2328                 goto err_hctxs;
2329
2330         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2331         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2332
2333         q->nr_queues = nr_cpu_ids;
2334
2335         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2336
2337         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2338                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2339
2340         q->sg_reserved_size = INT_MAX;
2341
2342         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2343         INIT_LIST_HEAD(&q->requeue_list);
2344         spin_lock_init(&q->requeue_lock);
2345
2346         blk_queue_make_request(q, blk_mq_make_request);
2347
2348         /*
2349          * Do this after blk_queue_make_request() overrides it...
2350          */
2351         q->nr_requests = set->queue_depth;
2352
2353         /*
2354          * Default to classic polling
2355          */
2356         q->poll_nsec = -1;
2357
2358         if (set->ops->complete)
2359                 blk_queue_softirq_done(q, set->ops->complete);
2360
2361         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2362         blk_mq_add_queue_tag_set(set, q);
2363         blk_mq_map_swqueue(q);
2364
2365         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2366                 int ret;
2367
2368                 ret = blk_mq_sched_init(q);
2369                 if (ret)
2370                         return ERR_PTR(ret);
2371         }
2372
2373         return q;
2374
2375 err_hctxs:
2376         kfree(q->queue_hw_ctx);
2377 err_percpu:
2378         free_percpu(q->queue_ctx);
2379 err_exit:
2380         q->mq_ops = NULL;
2381         return ERR_PTR(-ENOMEM);
2382 }
2383 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2384
2385 void blk_mq_free_queue(struct request_queue *q)
2386 {
2387         struct blk_mq_tag_set   *set = q->tag_set;
2388
2389         blk_mq_del_queue_tag_set(q);
2390         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2391 }
2392
2393 /* Basically redo blk_mq_init_queue with queue frozen */
2394 static void blk_mq_queue_reinit(struct request_queue *q)
2395 {
2396         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2397
2398         blk_mq_debugfs_unregister_hctxs(q);
2399         blk_mq_sysfs_unregister(q);
2400
2401         /*
2402          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2403          * we should change hctx numa_node according to new topology (this
2404          * involves free and re-allocate memory, worthy doing?)
2405          */
2406
2407         blk_mq_map_swqueue(q);
2408
2409         blk_mq_sysfs_register(q);
2410         blk_mq_debugfs_register_hctxs(q);
2411 }
2412
2413 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2414 {
2415         int i;
2416
2417         for (i = 0; i < set->nr_hw_queues; i++)
2418                 if (!__blk_mq_alloc_rq_map(set, i))
2419                         goto out_unwind;
2420
2421         return 0;
2422
2423 out_unwind:
2424         while (--i >= 0)
2425                 blk_mq_free_rq_map(set->tags[i]);
2426
2427         return -ENOMEM;
2428 }
2429
2430 /*
2431  * Allocate the request maps associated with this tag_set. Note that this
2432  * may reduce the depth asked for, if memory is tight. set->queue_depth
2433  * will be updated to reflect the allocated depth.
2434  */
2435 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2436 {
2437         unsigned int depth;
2438         int err;
2439
2440         depth = set->queue_depth;
2441         do {
2442                 err = __blk_mq_alloc_rq_maps(set);
2443                 if (!err)
2444                         break;
2445
2446                 set->queue_depth >>= 1;
2447                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2448                         err = -ENOMEM;
2449                         break;
2450                 }
2451         } while (set->queue_depth);
2452
2453         if (!set->queue_depth || err) {
2454                 pr_err("blk-mq: failed to allocate request map\n");
2455                 return -ENOMEM;
2456         }
2457
2458         if (depth != set->queue_depth)
2459                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2460                                                 depth, set->queue_depth);
2461
2462         return 0;
2463 }
2464
2465 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2466 {
2467         if (set->ops->map_queues)
2468                 return set->ops->map_queues(set);
2469         else
2470                 return blk_mq_map_queues(set);
2471 }
2472
2473 /*
2474  * Alloc a tag set to be associated with one or more request queues.
2475  * May fail with EINVAL for various error conditions. May adjust the
2476  * requested depth down, if if it too large. In that case, the set
2477  * value will be stored in set->queue_depth.
2478  */
2479 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2480 {
2481         int ret;
2482
2483         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2484
2485         if (!set->nr_hw_queues)
2486                 return -EINVAL;
2487         if (!set->queue_depth)
2488                 return -EINVAL;
2489         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2490                 return -EINVAL;
2491
2492         if (!set->ops->queue_rq)
2493                 return -EINVAL;
2494
2495         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2496                 pr_info("blk-mq: reduced tag depth to %u\n",
2497                         BLK_MQ_MAX_DEPTH);
2498                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2499         }
2500
2501         /*
2502          * If a crashdump is active, then we are potentially in a very
2503          * memory constrained environment. Limit us to 1 queue and
2504          * 64 tags to prevent using too much memory.
2505          */
2506         if (is_kdump_kernel()) {
2507                 set->nr_hw_queues = 1;
2508                 set->queue_depth = min(64U, set->queue_depth);
2509         }
2510         /*
2511          * There is no use for more h/w queues than cpus.
2512          */
2513         if (set->nr_hw_queues > nr_cpu_ids)
2514                 set->nr_hw_queues = nr_cpu_ids;
2515
2516         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2517                                  GFP_KERNEL, set->numa_node);
2518         if (!set->tags)
2519                 return -ENOMEM;
2520
2521         ret = -ENOMEM;
2522         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2523                         GFP_KERNEL, set->numa_node);
2524         if (!set->mq_map)
2525                 goto out_free_tags;
2526
2527         ret = blk_mq_update_queue_map(set);
2528         if (ret)
2529                 goto out_free_mq_map;
2530
2531         ret = blk_mq_alloc_rq_maps(set);
2532         if (ret)
2533                 goto out_free_mq_map;
2534
2535         mutex_init(&set->tag_list_lock);
2536         INIT_LIST_HEAD(&set->tag_list);
2537
2538         return 0;
2539
2540 out_free_mq_map:
2541         kfree(set->mq_map);
2542         set->mq_map = NULL;
2543 out_free_tags:
2544         kfree(set->tags);
2545         set->tags = NULL;
2546         return ret;
2547 }
2548 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2549
2550 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2551 {
2552         int i;
2553
2554         for (i = 0; i < nr_cpu_ids; i++)
2555                 blk_mq_free_map_and_requests(set, i);
2556
2557         kfree(set->mq_map);
2558         set->mq_map = NULL;
2559
2560         kfree(set->tags);
2561         set->tags = NULL;
2562 }
2563 EXPORT_SYMBOL(blk_mq_free_tag_set);
2564
2565 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2566 {
2567         struct blk_mq_tag_set *set = q->tag_set;
2568         struct blk_mq_hw_ctx *hctx;
2569         int i, ret;
2570
2571         if (!set)
2572                 return -EINVAL;
2573
2574         blk_mq_freeze_queue(q);
2575
2576         ret = 0;
2577         queue_for_each_hw_ctx(q, hctx, i) {
2578                 if (!hctx->tags)
2579                         continue;
2580                 /*
2581                  * If we're using an MQ scheduler, just update the scheduler
2582                  * queue depth. This is similar to what the old code would do.
2583                  */
2584                 if (!hctx->sched_tags) {
2585                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2586                                                         min(nr, set->queue_depth),
2587                                                         false);
2588                 } else {
2589                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2590                                                         nr, true);
2591                 }
2592                 if (ret)
2593                         break;
2594         }
2595
2596         if (!ret)
2597                 q->nr_requests = nr;
2598
2599         blk_mq_unfreeze_queue(q);
2600
2601         return ret;
2602 }
2603
2604 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2605                                                         int nr_hw_queues)
2606 {
2607         struct request_queue *q;
2608
2609         lockdep_assert_held(&set->tag_list_lock);
2610
2611         if (nr_hw_queues > nr_cpu_ids)
2612                 nr_hw_queues = nr_cpu_ids;
2613         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2614                 return;
2615
2616         list_for_each_entry(q, &set->tag_list, tag_set_list)
2617                 blk_mq_freeze_queue(q);
2618
2619         set->nr_hw_queues = nr_hw_queues;
2620         blk_mq_update_queue_map(set);
2621         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2622                 blk_mq_realloc_hw_ctxs(set, q);
2623                 blk_mq_queue_reinit(q);
2624         }
2625
2626         list_for_each_entry(q, &set->tag_list, tag_set_list)
2627                 blk_mq_unfreeze_queue(q);
2628 }
2629
2630 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2631 {
2632         mutex_lock(&set->tag_list_lock);
2633         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2634         mutex_unlock(&set->tag_list_lock);
2635 }
2636 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2637
2638 /* Enable polling stats and return whether they were already enabled. */
2639 static bool blk_poll_stats_enable(struct request_queue *q)
2640 {
2641         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2642             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2643                 return true;
2644         blk_stat_add_callback(q, q->poll_cb);
2645         return false;
2646 }
2647
2648 static void blk_mq_poll_stats_start(struct request_queue *q)
2649 {
2650         /*
2651          * We don't arm the callback if polling stats are not enabled or the
2652          * callback is already active.
2653          */
2654         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2655             blk_stat_is_active(q->poll_cb))
2656                 return;
2657
2658         blk_stat_activate_msecs(q->poll_cb, 100);
2659 }
2660
2661 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2662 {
2663         struct request_queue *q = cb->data;
2664         int bucket;
2665
2666         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2667                 if (cb->stat[bucket].nr_samples)
2668                         q->poll_stat[bucket] = cb->stat[bucket];
2669         }
2670 }
2671
2672 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2673                                        struct blk_mq_hw_ctx *hctx,
2674                                        struct request *rq)
2675 {
2676         unsigned long ret = 0;
2677         int bucket;
2678
2679         /*
2680          * If stats collection isn't on, don't sleep but turn it on for
2681          * future users
2682          */
2683         if (!blk_poll_stats_enable(q))
2684                 return 0;
2685
2686         /*
2687          * As an optimistic guess, use half of the mean service time
2688          * for this type of request. We can (and should) make this smarter.
2689          * For instance, if the completion latencies are tight, we can
2690          * get closer than just half the mean. This is especially
2691          * important on devices where the completion latencies are longer
2692          * than ~10 usec. We do use the stats for the relevant IO size
2693          * if available which does lead to better estimates.
2694          */
2695         bucket = blk_mq_poll_stats_bkt(rq);
2696         if (bucket < 0)
2697                 return ret;
2698
2699         if (q->poll_stat[bucket].nr_samples)
2700                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2701
2702         return ret;
2703 }
2704
2705 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2706                                      struct blk_mq_hw_ctx *hctx,
2707                                      struct request *rq)
2708 {
2709         struct hrtimer_sleeper hs;
2710         enum hrtimer_mode mode;
2711         unsigned int nsecs;
2712         ktime_t kt;
2713
2714         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2715                 return false;
2716
2717         /*
2718          * poll_nsec can be:
2719          *
2720          * -1:  don't ever hybrid sleep
2721          *  0:  use half of prev avg
2722          * >0:  use this specific value
2723          */
2724         if (q->poll_nsec == -1)
2725                 return false;
2726         else if (q->poll_nsec > 0)
2727                 nsecs = q->poll_nsec;
2728         else
2729                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2730
2731         if (!nsecs)
2732                 return false;
2733
2734         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2735
2736         /*
2737          * This will be replaced with the stats tracking code, using
2738          * 'avg_completion_time / 2' as the pre-sleep target.
2739          */
2740         kt = nsecs;
2741
2742         mode = HRTIMER_MODE_REL;
2743         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2744         hrtimer_set_expires(&hs.timer, kt);
2745
2746         hrtimer_init_sleeper(&hs, current);
2747         do {
2748                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2749                         break;
2750                 set_current_state(TASK_UNINTERRUPTIBLE);
2751                 hrtimer_start_expires(&hs.timer, mode);
2752                 if (hs.task)
2753                         io_schedule();
2754                 hrtimer_cancel(&hs.timer);
2755                 mode = HRTIMER_MODE_ABS;
2756         } while (hs.task && !signal_pending(current));
2757
2758         __set_current_state(TASK_RUNNING);
2759         destroy_hrtimer_on_stack(&hs.timer);
2760         return true;
2761 }
2762
2763 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2764 {
2765         struct request_queue *q = hctx->queue;
2766         long state;
2767
2768         /*
2769          * If we sleep, have the caller restart the poll loop to reset
2770          * the state. Like for the other success return cases, the
2771          * caller is responsible for checking if the IO completed. If
2772          * the IO isn't complete, we'll get called again and will go
2773          * straight to the busy poll loop.
2774          */
2775         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2776                 return true;
2777
2778         hctx->poll_considered++;
2779
2780         state = current->state;
2781         while (!need_resched()) {
2782                 int ret;
2783
2784                 hctx->poll_invoked++;
2785
2786                 ret = q->mq_ops->poll(hctx, rq->tag);
2787                 if (ret > 0) {
2788                         hctx->poll_success++;
2789                         set_current_state(TASK_RUNNING);
2790                         return true;
2791                 }
2792
2793                 if (signal_pending_state(state, current))
2794                         set_current_state(TASK_RUNNING);
2795
2796                 if (current->state == TASK_RUNNING)
2797                         return true;
2798                 if (ret < 0)
2799                         break;
2800                 cpu_relax();
2801         }
2802
2803         return false;
2804 }
2805
2806 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2807 {
2808         struct blk_mq_hw_ctx *hctx;
2809         struct blk_plug *plug;
2810         struct request *rq;
2811
2812         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2813             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2814                 return false;
2815
2816         plug = current->plug;
2817         if (plug)
2818                 blk_flush_plug_list(plug, false);
2819
2820         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2821         if (!blk_qc_t_is_internal(cookie))
2822                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2823         else {
2824                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2825                 /*
2826                  * With scheduling, if the request has completed, we'll
2827                  * get a NULL return here, as we clear the sched tag when
2828                  * that happens. The request still remains valid, like always,
2829                  * so we should be safe with just the NULL check.
2830                  */
2831                 if (!rq)
2832                         return false;
2833         }
2834
2835         return __blk_mq_poll(hctx, rq);
2836 }
2837 EXPORT_SYMBOL_GPL(blk_mq_poll);
2838
2839 static int __init blk_mq_init(void)
2840 {
2841         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2842                                 blk_mq_hctx_notify_dead);
2843         return 0;
2844 }
2845 subsys_initcall(blk_mq_init);