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
 305         blk_queue_enter_live(q);
 306         data->q = q;
 307         if (likely(!data->ctx))
 308                 data->ctx = blk_mq_get_ctx(q);
 309         if (likely(!data->hctx))
 310                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 311         if (op & REQ_NOWAIT)
 312                 data->flags |= BLK_MQ_REQ_NOWAIT;
 313
 314         if (e) {
 315                 data->flags |= BLK_MQ_REQ_INTERNAL;
 316
 317                 /*
 318                  * Flush requests are special and go directly to the
 319                  * dispatch list.
 320                  */
 321                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
 322                         e->type->ops.mq.limit_depth(op, data);
 323         }
 324
 325         tag = blk_mq_get_tag(data);
 326         if (tag == BLK_MQ_TAG_FAIL) {
 327                 blk_queue_exit(q);
 328                 return NULL;
 329         }
 330
 331         rq = blk_mq_rq_ctx_init(data, tag, op);
 332         if (!op_is_flush(op)) {
 333                 rq->elv.icq = NULL;
 334                 if (e && e->type->ops.mq.prepare_request) {
 335                         if (e->type->icq_cache && rq_ioc(bio))
 336                                 blk_mq_sched_assign_ioc(rq, bio);
 337
 338                         e->type->ops.mq.prepare_request(rq, bio);
 339                         rq->rq_flags |= RQF_ELVPRIV;
 340                 }
 341         }
 342         data->hctx->queued++;
 343         return rq;
 344 }
 345
 346 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 347                 unsigned int flags)
 348 {
 349         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 350         struct request *rq;
 351         int ret;
 352
 353         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 354         if (ret)
 355                 return ERR_PTR(ret);
 356
 357         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 358
 359         blk_mq_put_ctx(alloc_data.ctx);
 360         blk_queue_exit(q);
 361
 362         if (!rq)
 363                 return ERR_PTR(-EWOULDBLOCK);
 364
 365         rq->__data_len = 0;
 366         rq->__sector = (sector_t) -1;
 367         rq->bio = rq->biotail = NULL;
 368         return rq;
 369 }
 370 EXPORT_SYMBOL(blk_mq_alloc_request);
 371
 372 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 373                 unsigned int op, unsigned int flags, unsigned int hctx_idx)
 374 {
 375         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 376         struct request *rq;
 377         unsigned int cpu;
 378         int ret;
 379
 380         /*
 381          * If the tag allocator sleeps we could get an allocation for a
 382          * different hardware context.  No need to complicate the low level
 383          * allocator for this for the rare use case of a command tied to
 384          * a specific queue.
 385          */
 386         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 387                 return ERR_PTR(-EINVAL);
 388
 389         if (hctx_idx >= q->nr_hw_queues)
 390                 return ERR_PTR(-EIO);
 391
 392         ret = blk_queue_enter(q, true);
 393         if (ret)
 394                 return ERR_PTR(ret);
 395
 396         /*
 397          * Check if the hardware context is actually mapped to anything.
 398          * If not tell the caller that it should skip this queue.
 399          */
 400         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 401         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 402                 blk_queue_exit(q);
 403                 return ERR_PTR(-EXDEV);
 404         }
 405         cpu = cpumask_first(alloc_data.hctx->cpumask);
 406         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 407
 408         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 409
 410         blk_queue_exit(q);
 411
 412         if (!rq)
 413                 return ERR_PTR(-EWOULDBLOCK);
 414
 415         return rq;
 416 }
 417 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 418
 419 void blk_mq_free_request(struct request *rq)
 420 {
 421         struct request_queue *q = rq->q;
 422         struct elevator_queue *e = q->elevator;
 423         struct blk_mq_ctx *ctx = rq->mq_ctx;
 424         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 425         const int sched_tag = rq->internal_tag;
 426
 427         if (rq->rq_flags & RQF_ELVPRIV) {
 428                 if (e && e->type->ops.mq.finish_request)
 429                         e->type->ops.mq.finish_request(rq);
 430                 if (rq->elv.icq) {
 431                         put_io_context(rq->elv.icq->ioc);
 432                         rq->elv.icq = NULL;
 433                 }
 434         }
 435
 436         ctx->rq_completed[rq_is_sync(rq)]++;
 437         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 438                 atomic_dec(&hctx->nr_active);
 439
 440         wbt_done(q->rq_wb, &rq->issue_stat);
 441
 442         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 443         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 444         if (rq->tag != -1)
 445                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 446         if (sched_tag != -1)
 447                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 448         blk_mq_sched_restart(hctx);
 449         blk_queue_exit(q);
 450 }
 451 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 452
 453 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 454 {
 455         blk_account_io_done(rq);
 456
 457         if (rq->end_io) {
 458                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 459                 rq->end_io(rq, error);
 460         } else {
 461                 if (unlikely(blk_bidi_rq(rq)))
 462                         blk_mq_free_request(rq->next_rq);
 463                 blk_mq_free_request(rq);
 464         }
 465 }
 466 EXPORT_SYMBOL(__blk_mq_end_request);
 467
 468 void blk_mq_end_request(struct request *rq, blk_status_t error)
 469 {
 470         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 471                 BUG();
 472         __blk_mq_end_request(rq, error);
 473 }
 474 EXPORT_SYMBOL(blk_mq_end_request);
 475
 476 static void __blk_mq_complete_request_remote(void *data)
 477 {
 478         struct request *rq = data;
 479
 480         rq->q->softirq_done_fn(rq);
 481 }
 482
 483 static void __blk_mq_complete_request(struct request *rq)
 484 {
 485         struct blk_mq_ctx *ctx = rq->mq_ctx;
 486         bool shared = false;
 487         int cpu;
 488
 489         if (rq->internal_tag != -1)
 490                 blk_mq_sched_completed_request(rq);
 491         if (rq->rq_flags & RQF_STATS) {
 492                 blk_mq_poll_stats_start(rq->q);
 493                 blk_stat_add(rq);
 494         }
 495
 496         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 497                 rq->q->softirq_done_fn(rq);
 498                 return;
 499         }
 500
 501         cpu = get_cpu();
 502         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 503                 shared = cpus_share_cache(cpu, ctx->cpu);
 504
 505         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 506                 rq->csd.func = __blk_mq_complete_request_remote;
 507                 rq->csd.info = rq;
 508                 rq->csd.flags = 0;
 509                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 510         } else {
 511                 rq->q->softirq_done_fn(rq);
 512         }
 513         put_cpu();
 514 }
 515
 516 /**
 517  * blk_mq_complete_request - end I/O on a request
 518  * @rq:         the request being processed
 519  *
 520  * Description:
 521  *      Ends all I/O on a request. It does not handle partial completions.
 522  *      The actual completion happens out-of-order, through a IPI handler.
 523  **/
 524 void blk_mq_complete_request(struct request *rq)
 525 {
 526         struct request_queue *q = rq->q;
 527
 528         if (unlikely(blk_should_fake_timeout(q)))
 529                 return;
 530         if (!blk_mark_rq_complete(rq))
 531                 __blk_mq_complete_request(rq);
 532 }
 533 EXPORT_SYMBOL(blk_mq_complete_request);
 534
 535 int blk_mq_request_started(struct request *rq)
 536 {
 537         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 538 }
 539 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 540
 541 void blk_mq_start_request(struct request *rq)
 542 {
 543         struct request_queue *q = rq->q;
 544
 545         blk_mq_sched_started_request(rq);
 546
 547         trace_block_rq_issue(q, rq);
 548
 549         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 550                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
 551                 rq->rq_flags |= RQF_STATS;
 552                 wbt_issue(q->rq_wb, &rq->issue_stat);
 553         }
 554
 555         blk_add_timer(rq);
 556
 557         /*
 558          * Ensure that ->deadline is visible before set the started
 559          * flag and clear the completed flag.
 560          */
 561         smp_mb__before_atomic();
 562
 563         /*
 564          * Mark us as started and clear complete. Complete might have been
 565          * set if requeue raced with timeout, which then marked it as
 566          * complete. So be sure to clear complete again when we start
 567          * the request, otherwise we'll ignore the completion event.
 568          */
 569         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 570                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 571         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 572                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 573
 574         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 575                 /*
 576                  * Make sure space for the drain appears.  We know we can do
 577                  * this because max_hw_segments has been adjusted to be one
 578                  * fewer than the device can handle.
 579                  */
 580                 rq->nr_phys_segments++;
 581         }
 582 }
 583 EXPORT_SYMBOL(blk_mq_start_request);
 584
 585 /*
 586  * When we reach here because queue is busy, REQ_ATOM_COMPLETE
 587  * flag isn't set yet, so there may be race with timeout handler,
 588  * but given rq->deadline is just set in .queue_rq() under
 589  * this situation, the race won't be possible in reality because
 590  * rq->timeout should be set as big enough to cover the window
 591  * between blk_mq_start_request() called from .queue_rq() and
 592  * clearing REQ_ATOM_STARTED here.
 593  */
 594 static void __blk_mq_requeue_request(struct request *rq)
 595 {
 596         struct request_queue *q = rq->q;
 597
 598         trace_block_rq_requeue(q, rq);
 599         wbt_requeue(q->rq_wb, &rq->issue_stat);
 600         blk_mq_sched_requeue_request(rq);
 601
 602         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 603                 if (q->dma_drain_size && blk_rq_bytes(rq))
 604                         rq->nr_phys_segments--;
 605         }
 606 }
 607
 608 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 609 {
 610         __blk_mq_requeue_request(rq);
 611
 612         BUG_ON(blk_queued_rq(rq));
 613         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 614 }
 615 EXPORT_SYMBOL(blk_mq_requeue_request);
 616
 617 static void blk_mq_requeue_work(struct work_struct *work)
 618 {
 619         struct request_queue *q =
 620                 container_of(work, struct request_queue, requeue_work.work);
 621         LIST_HEAD(rq_list);
 622         struct request *rq, *next;
 623
 624         spin_lock_irq(&q->requeue_lock);
 625         list_splice_init(&q->requeue_list, &rq_list);
 626         spin_unlock_irq(&q->requeue_lock);
 627
 628         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 629                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 630                         continue;
 631
 632                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 633                 list_del_init(&rq->queuelist);
 634                 blk_mq_sched_insert_request(rq, true, false, false, true);
 635         }
 636
 637         while (!list_empty(&rq_list)) {
 638                 rq = list_entry(rq_list.next, struct request, queuelist);
 639                 list_del_init(&rq->queuelist);
 640                 blk_mq_sched_insert_request(rq, false, false, false, true);
 641         }
 642
 643         blk_mq_run_hw_queues(q, false);
 644 }
 645
 646 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 647                                 bool kick_requeue_list)
 648 {
 649         struct request_queue *q = rq->q;
 650         unsigned long flags;
 651
 652         /*
 653          * We abuse this flag that is otherwise used by the I/O scheduler to
 654          * request head insertation from the workqueue.
 655          */
 656         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 657
 658         spin_lock_irqsave(&q->requeue_lock, flags);
 659         if (at_head) {
 660                 rq->rq_flags |= RQF_SOFTBARRIER;
 661                 list_add(&rq->queuelist, &q->requeue_list);
 662         } else {
 663                 list_add_tail(&rq->queuelist, &q->requeue_list);
 664         }
 665         spin_unlock_irqrestore(&q->requeue_lock, flags);
 666
 667         if (kick_requeue_list)
 668                 blk_mq_kick_requeue_list(q);
 669 }
 670 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 671
 672 void blk_mq_kick_requeue_list(struct request_queue *q)
 673 {
 674         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 675 }
 676 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 677
 678 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 679                                     unsigned long msecs)
 680 {
 681         kblockd_schedule_delayed_work(&q->requeue_work,
 682                                       msecs_to_jiffies(msecs));
 683 }
 684 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 685
 686 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 687 {
 688         if (tag < tags->nr_tags) {
 689                 prefetch(tags->rqs[tag]);
 690                 return tags->rqs[tag];
 691         }
 692
 693         return NULL;
 694 }
 695 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 696
 697 struct blk_mq_timeout_data {
 698         unsigned long next;
 699         unsigned int next_set;
 700 };
 701
 702 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 703 {
 704         const struct blk_mq_ops *ops = req->q->mq_ops;
 705         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 706
 707         /*
 708          * We know that complete is set at this point. If STARTED isn't set
 709          * anymore, then the request isn't active and the "timeout" should
 710          * just be ignored. This can happen due to the bitflag ordering.
 711          * Timeout first checks if STARTED is set, and if it is, assumes
 712          * the request is active. But if we race with completion, then
 713          * both flags will get cleared. So check here again, and ignore
 714          * a timeout event with a request that isn't active.
 715          */
 716         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 717                 return;
 718
 719         if (ops->timeout)
 720                 ret = ops->timeout(req, reserved);
 721
 722         switch (ret) {
 723         case BLK_EH_HANDLED:
 724                 __blk_mq_complete_request(req);
 725                 break;
 726         case BLK_EH_RESET_TIMER:
 727                 blk_add_timer(req);
 728                 blk_clear_rq_complete(req);
 729                 break;
 730         case BLK_EH_NOT_HANDLED:
 731                 break;
 732         default:
 733                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 734                 break;
 735         }
 736 }
 737
 738 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 739                 struct request *rq, void *priv, bool reserved)
 740 {
 741         struct blk_mq_timeout_data *data = priv;
 742
 743         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 744                 return;
 745
 746         /*
 747          * The rq being checked may have been freed and reallocated
 748          * out already here, we avoid this race by checking rq->deadline
 749          * and REQ_ATOM_COMPLETE flag together:
 750          *
 751          * - if rq->deadline is observed as new value because of
 752          *   reusing, the rq won't be timed out because of timing.
 753          * - if rq->deadline is observed as previous value,
 754          *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
 755          *   because we put a barrier between setting rq->deadline
 756          *   and clearing the flag in blk_mq_start_request(), so
 757          *   this rq won't be timed out too.
 758          */
 759         if (time_after_eq(jiffies, rq->deadline)) {
 760                 if (!blk_mark_rq_complete(rq))
 761                         blk_mq_rq_timed_out(rq, reserved);
 762         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 763                 data->next = rq->deadline;
 764                 data->next_set = 1;
 765         }
 766 }
 767
 768 static void blk_mq_timeout_work(struct work_struct *work)
 769 {
 770         struct request_queue *q =
 771                 container_of(work, struct request_queue, timeout_work);
 772         struct blk_mq_timeout_data data = {
 773                 .next           = 0,
 774                 .next_set       = 0,
 775         };
 776         int i;
 777
 778         /* A deadlock might occur if a request is stuck requiring a
 779          * timeout at the same time a queue freeze is waiting
 780          * completion, since the timeout code would not be able to
 781          * acquire the queue reference here.
 782          *
 783          * That's why we don't use blk_queue_enter here; instead, we use
 784          * percpu_ref_tryget directly, because we need to be able to
 785          * obtain a reference even in the short window between the queue
 786          * starting to freeze, by dropping the first reference in
 787          * blk_freeze_queue_start, and the moment the last request is
 788          * consumed, marked by the instant q_usage_counter reaches
 789          * zero.
 790          */
 791         if (!percpu_ref_tryget(&q->q_usage_counter))
 792                 return;
 793
 794         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 795
 796         if (data.next_set) {
 797                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 798                 mod_timer(&q->timeout, data.next);
 799         } else {
 800                 struct blk_mq_hw_ctx *hctx;
 801
 802                 queue_for_each_hw_ctx(q, hctx, i) {
 803                         /* the hctx may be unmapped, so check it here */
 804                         if (blk_mq_hw_queue_mapped(hctx))
 805                                 blk_mq_tag_idle(hctx);
 806                 }
 807         }
 808         blk_queue_exit(q);
 809 }
 810
 811 struct flush_busy_ctx_data {
 812         struct blk_mq_hw_ctx *hctx;
 813         struct list_head *list;
 814 };
 815
 816 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 817 {
 818         struct flush_busy_ctx_data *flush_data = data;
 819         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 820         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 821
 822         sbitmap_clear_bit(sb, bitnr);
 823         spin_lock(&ctx->lock);
 824         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 825         spin_unlock(&ctx->lock);
 826         return true;
 827 }
 828
 829 /*
 830  * Process software queues that have been marked busy, splicing them
 831  * to the for-dispatch
 832  */
 833 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 834 {
 835         struct flush_busy_ctx_data data = {
 836                 .hctx = hctx,
 837                 .list = list,
 838         };
 839
 840         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 841 }
 842 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 843
 844 static inline unsigned int queued_to_index(unsigned int queued)
 845 {
 846         if (!queued)
 847                 return 0;
 848
 849         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 850 }
 851
 852 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 853                            bool wait)
 854 {
 855         struct blk_mq_alloc_data data = {
 856                 .q = rq->q,
 857                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 858                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 859         };
 860
 861         might_sleep_if(wait);
 862
 863         if (rq->tag != -1)
 864                 goto done;
 865
 866         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 867                 data.flags |= BLK_MQ_REQ_RESERVED;
 868
 869         rq->tag = blk_mq_get_tag(&data);
 870         if (rq->tag >= 0) {
 871                 if (blk_mq_tag_busy(data.hctx)) {
 872                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 873                         atomic_inc(&data.hctx->nr_active);
 874                 }
 875                 data.hctx->tags->rqs[rq->tag] = rq;
 876         }
 877
 878 done:
 879         if (hctx)
 880                 *hctx = data.hctx;
 881         return rq->tag != -1;
 882 }
 883
 884 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 885                                     struct request *rq)
 886 {
 887         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
 888         rq->tag = -1;
 889
 890         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 891                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
 892                 atomic_dec(&hctx->nr_active);
 893         }
 894 }
 895
 896 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
 897                                        struct request *rq)
 898 {
 899         if (rq->tag == -1 || rq->internal_tag == -1)
 900                 return;
 901
 902         __blk_mq_put_driver_tag(hctx, rq);
 903 }
 904
 905 static void blk_mq_put_driver_tag(struct request *rq)
 906 {
 907         struct blk_mq_hw_ctx *hctx;
 908
 909         if (rq->tag == -1 || rq->internal_tag == -1)
 910                 return;
 911
 912         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
 913         __blk_mq_put_driver_tag(hctx, rq);
 914 }
 915
 916 /*
 917  * If we fail getting a driver tag because all the driver tags are already
 918  * assigned and on the dispatch list, BUT the first entry does not have a
 919  * tag, then we could deadlock. For that case, move entries with assigned
 920  * driver tags to the front, leaving the set of tagged requests in the
 921  * same order, and the untagged set in the same order.
 922  */
 923 static bool reorder_tags_to_front(struct list_head *list)
 924 {
 925         struct request *rq, *tmp, *first = NULL;
 926
 927         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
 928                 if (rq == first)
 929                         break;
 930                 if (rq->tag != -1) {
 931                         list_move(&rq->queuelist, list);
 932                         if (!first)
 933                                 first = rq;
 934                 }
 935         }
 936
 937         return first != NULL;
 938 }
 939
 940 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
 941                                 void *key)
 942 {
 943         struct blk_mq_hw_ctx *hctx;
 944
 945         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
 946
 947         list_del(&wait->entry);
 948         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
 949         blk_mq_run_hw_queue(hctx, true);
 950         return 1;
 951 }
 952
 953 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
 954 {
 955         struct sbq_wait_state *ws;
 956
 957         /*
 958          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
 959          * The thread which wins the race to grab this bit adds the hardware
 960          * queue to the wait queue.
 961          */
 962         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
 963             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
 964                 return false;
 965
 966         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
 967         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
 968
 969         /*
 970          * As soon as this returns, it's no longer safe to fiddle with
 971          * hctx->dispatch_wait, since a completion can wake up the wait queue
 972          * and unlock the bit.
 973          */
 974         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
 975         return true;
 976 }
 977
 978 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
 979 {
 980         struct blk_mq_hw_ctx *hctx;
 981         struct request *rq;
 982         int errors, queued;
 983
 984         if (list_empty(list))
 985                 return false;
 986
 987         /*
 988          * Now process all the entries, sending them to the driver.
 989          */
 990         errors = queued = 0;
 991         do {
 992                 struct blk_mq_queue_data bd;
 993                 blk_status_t ret;
 994
 995                 rq = list_first_entry(list, struct request, queuelist);
 996                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
 997                         if (!queued && reorder_tags_to_front(list))
 998                                 continue;
 999
1000                         /*
1001                          * The initial allocation attempt failed, so we need to
1002                          * rerun the hardware queue when a tag is freed.
1003                          */
1004                         if (!blk_mq_dispatch_wait_add(hctx))
1005                                 break;
1006
1007                         /*
1008                          * It's possible that a tag was freed in the window
1009                          * between the allocation failure and adding the
1010                          * hardware queue to the wait queue.
1011                          */
1012                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1013                                 break;
1014                 }
1015
1016                 list_del_init(&rq->queuelist);
1017
1018                 bd.rq = rq;
1019
1020                 /*
1021                  * Flag last if we have no more requests, or if we have more
1022                  * but can't assign a driver tag to it.
1023                  */
1024                 if (list_empty(list))
1025                         bd.last = true;
1026                 else {
1027                         struct request *nxt;
1028
1029                         nxt = list_first_entry(list, struct request, queuelist);
1030                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1031                 }
1032
1033                 ret = q->mq_ops->queue_rq(hctx, &bd);
1034                 if (ret == BLK_STS_RESOURCE) {
1035                         blk_mq_put_driver_tag_hctx(hctx, rq);
1036                         list_add(&rq->queuelist, list);
1037                         __blk_mq_requeue_request(rq);
1038                         break;
1039                 }
1040
1041                 if (unlikely(ret != BLK_STS_OK)) {
1042                         errors++;
1043                         blk_mq_end_request(rq, BLK_STS_IOERR);
1044                         continue;
1045                 }
1046
1047                 queued++;
1048         } while (!list_empty(list));
1049
1050         hctx->dispatched[queued_to_index(queued)]++;
1051
1052         /*
1053          * Any items that need requeuing? Stuff them into hctx->dispatch,
1054          * that is where we will continue on next queue run.
1055          */
1056         if (!list_empty(list)) {
1057                 /*
1058                  * If an I/O scheduler has been configured and we got a driver
1059                  * tag for the next request already, free it again.
1060                  */
1061                 rq = list_first_entry(list, struct request, queuelist);
1062                 blk_mq_put_driver_tag(rq);
1063
1064                 spin_lock(&hctx->lock);
1065                 list_splice_init(list, &hctx->dispatch);
1066                 spin_unlock(&hctx->lock);
1067
1068                 /*
1069                  * If SCHED_RESTART was set by the caller of this function and
1070                  * it is no longer set that means that it was cleared by another
1071                  * thread and hence that a queue rerun is needed.
1072                  *
1073                  * If TAG_WAITING is set that means that an I/O scheduler has
1074                  * been configured and another thread is waiting for a driver
1075                  * tag. To guarantee fairness, do not rerun this hardware queue
1076                  * but let the other thread grab the driver tag.
1077                  *
1078                  * If no I/O scheduler has been configured it is possible that
1079                  * the hardware queue got stopped and restarted before requests
1080                  * were pushed back onto the dispatch list. Rerun the queue to
1081                  * avoid starvation. Notes:
1082                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1083                  *   been stopped before rerunning a queue.
1084                  * - Some but not all block drivers stop a queue before
1085                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1086                  *   and dm-rq.
1087                  */
1088                 if (!blk_mq_sched_needs_restart(hctx) &&
1089                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1090                         blk_mq_run_hw_queue(hctx, true);
1091         }
1092
1093         return (queued + errors) != 0;
1094 }
1095
1096 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1097 {
1098         int srcu_idx;
1099
1100         /*
1101          * We should be running this queue from one of the CPUs that
1102          * are mapped to it.
1103          */
1104         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1105                 cpu_online(hctx->next_cpu));
1106
1107         /*
1108          * We can't run the queue inline with ints disabled. Ensure that
1109          * we catch bad users of this early.
1110          */
1111         WARN_ON_ONCE(in_interrupt());
1112
1113         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1114                 rcu_read_lock();
1115                 blk_mq_sched_dispatch_requests(hctx);
1116                 rcu_read_unlock();
1117         } else {
1118                 might_sleep();
1119
1120                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1121                 blk_mq_sched_dispatch_requests(hctx);
1122                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1123         }
1124 }
1125
1126 /*
1127  * It'd be great if the workqueue API had a way to pass
1128  * in a mask and had some smarts for more clever placement.
1129  * For now we just round-robin here, switching for every
1130  * BLK_MQ_CPU_WORK_BATCH queued items.
1131  */
1132 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1133 {
1134         if (hctx->queue->nr_hw_queues == 1)
1135                 return WORK_CPU_UNBOUND;
1136
1137         if (--hctx->next_cpu_batch <= 0) {
1138                 int next_cpu;
1139
1140                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1141                 if (next_cpu >= nr_cpu_ids)
1142                         next_cpu = cpumask_first(hctx->cpumask);
1143
1144                 hctx->next_cpu = next_cpu;
1145                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1146         }
1147
1148         return hctx->next_cpu;
1149 }
1150
1151 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1152                                         unsigned long msecs)
1153 {
1154         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1155                 return;
1156
1157         if (unlikely(blk_mq_hctx_stopped(hctx)))
1158                 return;
1159
1160         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1161                 int cpu = get_cpu();
1162                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1163                         __blk_mq_run_hw_queue(hctx);
1164                         put_cpu();
1165                         return;
1166                 }
1167
1168                 put_cpu();
1169         }
1170
1171         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1172                                          &hctx->run_work,
1173                                          msecs_to_jiffies(msecs));
1174 }
1175
1176 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1177 {
1178         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1179 }
1180 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1181
1182 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1183 {
1184         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1185 }
1186 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1187
1188 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1189 {
1190         struct blk_mq_hw_ctx *hctx;
1191         int i;
1192
1193         queue_for_each_hw_ctx(q, hctx, i) {
1194                 if (!blk_mq_hctx_has_pending(hctx) ||
1195                     blk_mq_hctx_stopped(hctx))
1196                         continue;
1197
1198                 blk_mq_run_hw_queue(hctx, async);
1199         }
1200 }
1201 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1202
1203 /**
1204  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1205  * @q: request queue.
1206  *
1207  * The caller is responsible for serializing this function against
1208  * blk_mq_{start,stop}_hw_queue().
1209  */
1210 bool blk_mq_queue_stopped(struct request_queue *q)
1211 {
1212         struct blk_mq_hw_ctx *hctx;
1213         int i;
1214
1215         queue_for_each_hw_ctx(q, hctx, i)
1216                 if (blk_mq_hctx_stopped(hctx))
1217                         return true;
1218
1219         return false;
1220 }
1221 EXPORT_SYMBOL(blk_mq_queue_stopped);
1222
1223 /*
1224  * This function is often used for pausing .queue_rq() by driver when
1225  * there isn't enough resource or some conditions aren't satisfied, and
1226  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1227  *
1228  * We do not guarantee that dispatch can be drained or blocked
1229  * after blk_mq_stop_hw_queue() returns. Please use
1230  * blk_mq_quiesce_queue() for that requirement.
1231  */
1232 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1233 {
1234         cancel_delayed_work(&hctx->run_work);
1235
1236         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1237 }
1238 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1239
1240 /*
1241  * This function is often used for pausing .queue_rq() by driver when
1242  * there isn't enough resource or some conditions aren't satisfied, and
1243  * BLK_MQ_RQ_QUEUE_BUSY is usually returned.
1244  *
1245  * We do not guarantee that dispatch can be drained or blocked
1246  * after blk_mq_stop_hw_queues() returns. Please use
1247  * blk_mq_quiesce_queue() for that requirement.
1248  */
1249 void blk_mq_stop_hw_queues(struct request_queue *q)
1250 {
1251         struct blk_mq_hw_ctx *hctx;
1252         int i;
1253
1254         queue_for_each_hw_ctx(q, hctx, i)
1255                 blk_mq_stop_hw_queue(hctx);
1256 }
1257 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1258
1259 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1260 {
1261         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1262
1263         blk_mq_run_hw_queue(hctx, false);
1264 }
1265 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1266
1267 void blk_mq_start_hw_queues(struct request_queue *q)
1268 {
1269         struct blk_mq_hw_ctx *hctx;
1270         int i;
1271
1272         queue_for_each_hw_ctx(q, hctx, i)
1273                 blk_mq_start_hw_queue(hctx);
1274 }
1275 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1276
1277 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1278 {
1279         if (!blk_mq_hctx_stopped(hctx))
1280                 return;
1281
1282         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1283         blk_mq_run_hw_queue(hctx, async);
1284 }
1285 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1286
1287 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1288 {
1289         struct blk_mq_hw_ctx *hctx;
1290         int i;
1291
1292         queue_for_each_hw_ctx(q, hctx, i)
1293                 blk_mq_start_stopped_hw_queue(hctx, async);
1294 }
1295 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1296
1297 static void blk_mq_run_work_fn(struct work_struct *work)
1298 {
1299         struct blk_mq_hw_ctx *hctx;
1300
1301         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1302
1303         /*
1304          * If we are stopped, don't run the queue. The exception is if
1305          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1306          * the STOPPED bit and run it.
1307          */
1308         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1309                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1310                         return;
1311
1312                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1313                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1314         }
1315
1316         __blk_mq_run_hw_queue(hctx);
1317 }
1318
1319
1320 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1321 {
1322         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1323                 return;
1324
1325         /*
1326          * Stop the hw queue, then modify currently delayed work.
1327          * This should prevent us from running the queue prematurely.
1328          * Mark the queue as auto-clearing STOPPED when it runs.
1329          */
1330         blk_mq_stop_hw_queue(hctx);
1331         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1332         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1333                                         &hctx->run_work,
1334                                         msecs_to_jiffies(msecs));
1335 }
1336 EXPORT_SYMBOL(blk_mq_delay_queue);
1337
1338 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1339                                             struct request *rq,
1340                                             bool at_head)
1341 {
1342         struct blk_mq_ctx *ctx = rq->mq_ctx;
1343
1344         lockdep_assert_held(&ctx->lock);
1345
1346         trace_block_rq_insert(hctx->queue, rq);
1347
1348         if (at_head)
1349                 list_add(&rq->queuelist, &ctx->rq_list);
1350         else
1351                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1352 }
1353
1354 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1355                              bool at_head)
1356 {
1357         struct blk_mq_ctx *ctx = rq->mq_ctx;
1358
1359         lockdep_assert_held(&ctx->lock);
1360
1361         __blk_mq_insert_req_list(hctx, rq, at_head);
1362         blk_mq_hctx_mark_pending(hctx, ctx);
1363 }
1364
1365 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1366                             struct list_head *list)
1367
1368 {
1369         /*
1370          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1371          * offline now
1372          */
1373         spin_lock(&ctx->lock);
1374         while (!list_empty(list)) {
1375                 struct request *rq;
1376
1377                 rq = list_first_entry(list, struct request, queuelist);
1378                 BUG_ON(rq->mq_ctx != ctx);
1379                 list_del_init(&rq->queuelist);
1380                 __blk_mq_insert_req_list(hctx, rq, false);
1381         }
1382         blk_mq_hctx_mark_pending(hctx, ctx);
1383         spin_unlock(&ctx->lock);
1384 }
1385
1386 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1387 {
1388         struct request *rqa = container_of(a, struct request, queuelist);
1389         struct request *rqb = container_of(b, struct request, queuelist);
1390
1391         return !(rqa->mq_ctx < rqb->mq_ctx ||
1392                  (rqa->mq_ctx == rqb->mq_ctx &&
1393                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1394 }
1395
1396 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1397 {
1398         struct blk_mq_ctx *this_ctx;
1399         struct request_queue *this_q;
1400         struct request *rq;
1401         LIST_HEAD(list);
1402         LIST_HEAD(ctx_list);
1403         unsigned int depth;
1404
1405         list_splice_init(&plug->mq_list, &list);
1406
1407         list_sort(NULL, &list, plug_ctx_cmp);
1408
1409         this_q = NULL;
1410         this_ctx = NULL;
1411         depth = 0;
1412
1413         while (!list_empty(&list)) {
1414                 rq = list_entry_rq(list.next);
1415                 list_del_init(&rq->queuelist);
1416                 BUG_ON(!rq->q);
1417                 if (rq->mq_ctx != this_ctx) {
1418                         if (this_ctx) {
1419                                 trace_block_unplug(this_q, depth, from_schedule);
1420                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1421                                                                 &ctx_list,
1422                                                                 from_schedule);
1423                         }
1424
1425                         this_ctx = rq->mq_ctx;
1426                         this_q = rq->q;
1427                         depth = 0;
1428                 }
1429
1430                 depth++;
1431                 list_add_tail(&rq->queuelist, &ctx_list);
1432         }
1433
1434         /*
1435          * If 'this_ctx' is set, we know we have entries to complete
1436          * on 'ctx_list'. Do those.
1437          */
1438         if (this_ctx) {
1439                 trace_block_unplug(this_q, depth, from_schedule);
1440                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1441                                                 from_schedule);
1442         }
1443 }
1444
1445 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1446 {
1447         blk_init_request_from_bio(rq, bio);
1448
1449         blk_account_io_start(rq, true);
1450 }
1451
1452 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1453 {
1454         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1455                 !blk_queue_nomerges(hctx->queue);
1456 }
1457
1458 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1459                                    struct blk_mq_ctx *ctx,
1460                                    struct request *rq)
1461 {
1462         spin_lock(&ctx->lock);
1463         __blk_mq_insert_request(hctx, rq, false);
1464         spin_unlock(&ctx->lock);
1465 }
1466
1467 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1468 {
1469         if (rq->tag != -1)
1470                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1471
1472         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1473 }
1474
1475 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1476                                         struct request *rq,
1477                                         blk_qc_t *cookie, bool may_sleep)
1478 {
1479         struct request_queue *q = rq->q;
1480         struct blk_mq_queue_data bd = {
1481                 .rq = rq,
1482                 .last = true,
1483         };
1484         blk_qc_t new_cookie;
1485         blk_status_t ret;
1486         bool run_queue = true;
1487
1488         /* RCU or SRCU read lock is needed before checking quiesced flag */
1489         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1490                 run_queue = false;
1491                 goto insert;
1492         }
1493
1494         if (q->elevator)
1495                 goto insert;
1496
1497         if (!blk_mq_get_driver_tag(rq, NULL, false))
1498                 goto insert;
1499
1500         new_cookie = request_to_qc_t(hctx, rq);
1501
1502         /*
1503          * For OK queue, we are done. For error, kill it. Any other
1504          * error (busy), just add it to our list as we previously
1505          * would have done
1506          */
1507         ret = q->mq_ops->queue_rq(hctx, &bd);
1508         switch (ret) {
1509         case BLK_STS_OK:
1510                 *cookie = new_cookie;
1511                 return;
1512         case BLK_STS_RESOURCE:
1513                 __blk_mq_requeue_request(rq);
1514                 goto insert;
1515         default:
1516                 *cookie = BLK_QC_T_NONE;
1517                 blk_mq_end_request(rq, ret);
1518                 return;
1519         }
1520
1521 insert:
1522         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1523 }
1524
1525 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1526                 struct request *rq, blk_qc_t *cookie)
1527 {
1528         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1529                 rcu_read_lock();
1530                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1531                 rcu_read_unlock();
1532         } else {
1533                 unsigned int srcu_idx;
1534
1535                 might_sleep();
1536
1537                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1538                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1539                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1540         }
1541 }
1542
1543 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1544 {
1545         const int is_sync = op_is_sync(bio->bi_opf);
1546         const int is_flush_fua = op_is_flush(bio->bi_opf);
1547         struct blk_mq_alloc_data data = { .flags = 0 };
1548         struct request *rq;
1549         unsigned int request_count = 0;
1550         struct blk_plug *plug;
1551         struct request *same_queue_rq = NULL;
1552         blk_qc_t cookie;
1553         unsigned int wb_acct;
1554
1555         blk_queue_bounce(q, &bio);
1556
1557         blk_queue_split(q, &bio);
1558
1559         if (!bio_integrity_prep(bio))
1560                 return BLK_QC_T_NONE;
1561
1562         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1563             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1564                 return BLK_QC_T_NONE;
1565
1566         if (blk_mq_sched_bio_merge(q, bio))
1567                 return BLK_QC_T_NONE;
1568
1569         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1570
1571         trace_block_getrq(q, bio, bio->bi_opf);
1572
1573         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1574         if (unlikely(!rq)) {
1575                 __wbt_done(q->rq_wb, wb_acct);
1576                 if (bio->bi_opf & REQ_NOWAIT)
1577                         bio_wouldblock_error(bio);
1578                 return BLK_QC_T_NONE;
1579         }
1580
1581         wbt_track(&rq->issue_stat, wb_acct);
1582
1583         cookie = request_to_qc_t(data.hctx, rq);
1584
1585         plug = current->plug;
1586         if (unlikely(is_flush_fua)) {
1587                 blk_mq_put_ctx(data.ctx);
1588                 blk_mq_bio_to_request(rq, bio);
1589                 if (q->elevator) {
1590                         blk_mq_sched_insert_request(rq, false, true, true,
1591                                         true);
1592                 } else {
1593                         blk_insert_flush(rq);
1594                         blk_mq_run_hw_queue(data.hctx, true);
1595                 }
1596         } else if (plug && q->nr_hw_queues == 1) {
1597                 struct request *last = NULL;
1598
1599                 blk_mq_put_ctx(data.ctx);
1600                 blk_mq_bio_to_request(rq, bio);
1601
1602                 /*
1603                  * @request_count may become stale because of schedule
1604                  * out, so check the list again.
1605                  */
1606                 if (list_empty(&plug->mq_list))
1607                         request_count = 0;
1608                 else if (blk_queue_nomerges(q))
1609                         request_count = blk_plug_queued_count(q);
1610
1611                 if (!request_count)
1612                         trace_block_plug(q);
1613                 else
1614                         last = list_entry_rq(plug->mq_list.prev);
1615
1616                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1617                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1618                         blk_flush_plug_list(plug, false);
1619                         trace_block_plug(q);
1620                 }
1621
1622                 list_add_tail(&rq->queuelist, &plug->mq_list);
1623         } else if (plug && !blk_queue_nomerges(q)) {
1624                 blk_mq_bio_to_request(rq, bio);
1625
1626                 /*
1627                  * We do limited plugging. If the bio can be merged, do that.
1628                  * Otherwise the existing request in the plug list will be
1629                  * issued. So the plug list will have one request at most
1630                  * The plug list might get flushed before this. If that happens,
1631                  * the plug list is empty, and same_queue_rq is invalid.
1632                  */
1633                 if (list_empty(&plug->mq_list))
1634                         same_queue_rq = NULL;
1635                 if (same_queue_rq)
1636                         list_del_init(&same_queue_rq->queuelist);
1637                 list_add_tail(&rq->queuelist, &plug->mq_list);
1638
1639                 blk_mq_put_ctx(data.ctx);
1640
1641                 if (same_queue_rq) {
1642                         data.hctx = blk_mq_map_queue(q,
1643                                         same_queue_rq->mq_ctx->cpu);
1644                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1645                                         &cookie);
1646                 }
1647         } else if (q->nr_hw_queues > 1 && is_sync) {
1648                 blk_mq_put_ctx(data.ctx);
1649                 blk_mq_bio_to_request(rq, bio);
1650                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1651         } else if (q->elevator) {
1652                 blk_mq_put_ctx(data.ctx);
1653                 blk_mq_bio_to_request(rq, bio);
1654                 blk_mq_sched_insert_request(rq, false, true, true, true);
1655         } else {
1656                 blk_mq_put_ctx(data.ctx);
1657                 blk_mq_bio_to_request(rq, bio);
1658                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1659                 blk_mq_run_hw_queue(data.hctx, true);
1660         }
1661
1662         return cookie;
1663 }
1664
1665 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1666                      unsigned int hctx_idx)
1667 {
1668         struct page *page;
1669
1670         if (tags->rqs && set->ops->exit_request) {
1671                 int i;
1672
1673                 for (i = 0; i < tags->nr_tags; i++) {
1674                         struct request *rq = tags->static_rqs[i];
1675
1676                         if (!rq)
1677                                 continue;
1678                         set->ops->exit_request(set, rq, hctx_idx);
1679                         tags->static_rqs[i] = NULL;
1680                 }
1681         }
1682
1683         while (!list_empty(&tags->page_list)) {
1684                 page = list_first_entry(&tags->page_list, struct page, lru);
1685                 list_del_init(&page->lru);
1686                 /*
1687                  * Remove kmemleak object previously allocated in
1688                  * blk_mq_init_rq_map().
1689                  */
1690                 kmemleak_free(page_address(page));
1691                 __free_pages(page, page->private);
1692         }
1693 }
1694
1695 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1696 {
1697         kfree(tags->rqs);
1698         tags->rqs = NULL;
1699         kfree(tags->static_rqs);
1700         tags->static_rqs = NULL;
1701
1702         blk_mq_free_tags(tags);
1703 }
1704
1705 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1706                                         unsigned int hctx_idx,
1707                                         unsigned int nr_tags,
1708                                         unsigned int reserved_tags)
1709 {
1710         struct blk_mq_tags *tags;
1711         int node;
1712
1713         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1714         if (node == NUMA_NO_NODE)
1715                 node = set->numa_node;
1716
1717         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1718                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1719         if (!tags)
1720                 return NULL;
1721
1722         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1723                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1724                                  node);
1725         if (!tags->rqs) {
1726                 blk_mq_free_tags(tags);
1727                 return NULL;
1728         }
1729
1730         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1731                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1732                                  node);
1733         if (!tags->static_rqs) {
1734                 kfree(tags->rqs);
1735                 blk_mq_free_tags(tags);
1736                 return NULL;
1737         }
1738
1739         return tags;
1740 }
1741
1742 static size_t order_to_size(unsigned int order)
1743 {
1744         return (size_t)PAGE_SIZE << order;
1745 }
1746
1747 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1748                      unsigned int hctx_idx, unsigned int depth)
1749 {
1750         unsigned int i, j, entries_per_page, max_order = 4;
1751         size_t rq_size, left;
1752         int node;
1753
1754         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1755         if (node == NUMA_NO_NODE)
1756                 node = set->numa_node;
1757
1758         INIT_LIST_HEAD(&tags->page_list);
1759
1760         /*
1761          * rq_size is the size of the request plus driver payload, rounded
1762          * to the cacheline size
1763          */
1764         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1765                                 cache_line_size());
1766         left = rq_size * depth;
1767
1768         for (i = 0; i < depth; ) {
1769                 int this_order = max_order;
1770                 struct page *page;
1771                 int to_do;
1772                 void *p;
1773
1774                 while (this_order && left < order_to_size(this_order - 1))
1775                         this_order--;
1776
1777                 do {
1778                         page = alloc_pages_node(node,
1779                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1780                                 this_order);
1781                         if (page)
1782                                 break;
1783                         if (!this_order--)
1784                                 break;
1785                         if (order_to_size(this_order) < rq_size)
1786                                 break;
1787                 } while (1);
1788
1789                 if (!page)
1790                         goto fail;
1791
1792                 page->private = this_order;
1793                 list_add_tail(&page->lru, &tags->page_list);
1794
1795                 p = page_address(page);
1796                 /*
1797                  * Allow kmemleak to scan these pages as they contain pointers
1798                  * to additional allocations like via ops->init_request().
1799                  */
1800                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1801                 entries_per_page = order_to_size(this_order) / rq_size;
1802                 to_do = min(entries_per_page, depth - i);
1803                 left -= to_do * rq_size;
1804                 for (j = 0; j < to_do; j++) {
1805                         struct request *rq = p;
1806
1807                         tags->static_rqs[i] = rq;
1808                         if (set->ops->init_request) {
1809                                 if (set->ops->init_request(set, rq, hctx_idx,
1810                                                 node)) {
1811                                         tags->static_rqs[i] = NULL;
1812                                         goto fail;
1813                                 }
1814                         }
1815
1816                         p += rq_size;
1817                         i++;
1818                 }
1819         }
1820         return 0;
1821
1822 fail:
1823         blk_mq_free_rqs(set, tags, hctx_idx);
1824         return -ENOMEM;
1825 }
1826
1827 /*
1828  * 'cpu' is going away. splice any existing rq_list entries from this
1829  * software queue to the hw queue dispatch list, and ensure that it
1830  * gets run.
1831  */
1832 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1833 {
1834         struct blk_mq_hw_ctx *hctx;
1835         struct blk_mq_ctx *ctx;
1836         LIST_HEAD(tmp);
1837
1838         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1839         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1840
1841         spin_lock(&ctx->lock);
1842         if (!list_empty(&ctx->rq_list)) {
1843                 list_splice_init(&ctx->rq_list, &tmp);
1844                 blk_mq_hctx_clear_pending(hctx, ctx);
1845         }
1846         spin_unlock(&ctx->lock);
1847
1848         if (list_empty(&tmp))
1849                 return 0;
1850
1851         spin_lock(&hctx->lock);
1852         list_splice_tail_init(&tmp, &hctx->dispatch);
1853         spin_unlock(&hctx->lock);
1854
1855         blk_mq_run_hw_queue(hctx, true);
1856         return 0;
1857 }
1858
1859 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1860 {
1861         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1862                                             &hctx->cpuhp_dead);
1863 }
1864
1865 /* hctx->ctxs will be freed in queue's release handler */
1866 static void blk_mq_exit_hctx(struct request_queue *q,
1867                 struct blk_mq_tag_set *set,
1868                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1869 {
1870         blk_mq_debugfs_unregister_hctx(hctx);
1871
1872         blk_mq_tag_idle(hctx);
1873
1874         if (set->ops->exit_request)
1875                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1876
1877         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1878
1879         if (set->ops->exit_hctx)
1880                 set->ops->exit_hctx(hctx, hctx_idx);
1881
1882         if (hctx->flags & BLK_MQ_F_BLOCKING)
1883                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1884
1885         blk_mq_remove_cpuhp(hctx);
1886         blk_free_flush_queue(hctx->fq);
1887         sbitmap_free(&hctx->ctx_map);
1888 }
1889
1890 static void blk_mq_exit_hw_queues(struct request_queue *q,
1891                 struct blk_mq_tag_set *set, int nr_queue)
1892 {
1893         struct blk_mq_hw_ctx *hctx;
1894         unsigned int i;
1895
1896         queue_for_each_hw_ctx(q, hctx, i) {
1897                 if (i == nr_queue)
1898                         break;
1899                 blk_mq_exit_hctx(q, set, hctx, i);
1900         }
1901 }
1902
1903 static int blk_mq_init_hctx(struct request_queue *q,
1904                 struct blk_mq_tag_set *set,
1905                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1906 {
1907         int node;
1908
1909         node = hctx->numa_node;
1910         if (node == NUMA_NO_NODE)
1911                 node = hctx->numa_node = set->numa_node;
1912
1913         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1914         spin_lock_init(&hctx->lock);
1915         INIT_LIST_HEAD(&hctx->dispatch);
1916         hctx->queue = q;
1917         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1918
1919         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1920
1921         hctx->tags = set->tags[hctx_idx];
1922
1923         /*
1924          * Allocate space for all possible cpus to avoid allocation at
1925          * runtime
1926          */
1927         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1928                                         GFP_KERNEL, node);
1929         if (!hctx->ctxs)
1930                 goto unregister_cpu_notifier;
1931
1932         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1933                               node))
1934                 goto free_ctxs;
1935
1936         hctx->nr_ctx = 0;
1937
1938         if (set->ops->init_hctx &&
1939             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
1940                 goto free_bitmap;
1941
1942         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
1943                 goto exit_hctx;
1944
1945         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
1946         if (!hctx->fq)
1947                 goto sched_exit_hctx;
1948
1949         if (set->ops->init_request &&
1950             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
1951                                    node))
1952                 goto free_fq;
1953
1954         if (hctx->flags & BLK_MQ_F_BLOCKING)
1955                 init_srcu_struct(hctx->queue_rq_srcu);
1956
1957         blk_mq_debugfs_register_hctx(q, hctx);
1958
1959         return 0;
1960
1961  free_fq:
1962         kfree(hctx->fq);
1963  sched_exit_hctx:
1964         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1965  exit_hctx:
1966         if (set->ops->exit_hctx)
1967                 set->ops->exit_hctx(hctx, hctx_idx);
1968  free_bitmap:
1969         sbitmap_free(&hctx->ctx_map);
1970  free_ctxs:
1971         kfree(hctx->ctxs);
1972  unregister_cpu_notifier:
1973         blk_mq_remove_cpuhp(hctx);
1974         return -1;
1975 }
1976
1977 static void blk_mq_init_cpu_queues(struct request_queue *q,
1978                                    unsigned int nr_hw_queues)
1979 {
1980         unsigned int i;
1981
1982         for_each_possible_cpu(i) {
1983                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1984                 struct blk_mq_hw_ctx *hctx;
1985
1986                 __ctx->cpu = i;
1987                 spin_lock_init(&__ctx->lock);
1988                 INIT_LIST_HEAD(&__ctx->rq_list);
1989                 __ctx->queue = q;
1990
1991                 /* If the cpu isn't present, the cpu is mapped to first hctx */
1992                 if (!cpu_present(i))
1993                         continue;
1994
1995                 hctx = blk_mq_map_queue(q, i);
1996
1997                 /*
1998                  * Set local node, IFF we have more than one hw queue. If
1999                  * not, we remain on the home node of the device
2000                  */
2001                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2002                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2003         }
2004 }
2005
2006 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2007 {
2008         int ret = 0;
2009
2010         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2011                                         set->queue_depth, set->reserved_tags);
2012         if (!set->tags[hctx_idx])
2013                 return false;
2014
2015         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2016                                 set->queue_depth);
2017         if (!ret)
2018                 return true;
2019
2020         blk_mq_free_rq_map(set->tags[hctx_idx]);
2021         set->tags[hctx_idx] = NULL;
2022         return false;
2023 }
2024
2025 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2026                                          unsigned int hctx_idx)
2027 {
2028         if (set->tags[hctx_idx]) {
2029                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2030                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2031                 set->tags[hctx_idx] = NULL;
2032         }
2033 }
2034
2035 static void blk_mq_map_swqueue(struct request_queue *q)
2036 {
2037         unsigned int i, hctx_idx;
2038         struct blk_mq_hw_ctx *hctx;
2039         struct blk_mq_ctx *ctx;
2040         struct blk_mq_tag_set *set = q->tag_set;
2041
2042         /*
2043          * Avoid others reading imcomplete hctx->cpumask through sysfs
2044          */
2045         mutex_lock(&q->sysfs_lock);
2046
2047         queue_for_each_hw_ctx(q, hctx, i) {
2048                 cpumask_clear(hctx->cpumask);
2049                 hctx->nr_ctx = 0;
2050         }
2051
2052         /*
2053          * Map software to hardware queues.
2054          *
2055          * If the cpu isn't present, the cpu is mapped to first hctx.
2056          */
2057         for_each_present_cpu(i) {
2058                 hctx_idx = q->mq_map[i];
2059                 /* unmapped hw queue can be remapped after CPU topo changed */
2060                 if (!set->tags[hctx_idx] &&
2061                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2062                         /*
2063                          * If tags initialization fail for some hctx,
2064                          * that hctx won't be brought online.  In this
2065                          * case, remap the current ctx to hctx[0] which
2066                          * is guaranteed to always have tags allocated
2067                          */
2068                         q->mq_map[i] = 0;
2069                 }
2070
2071                 ctx = per_cpu_ptr(q->queue_ctx, i);
2072                 hctx = blk_mq_map_queue(q, i);
2073
2074                 cpumask_set_cpu(i, hctx->cpumask);
2075                 ctx->index_hw = hctx->nr_ctx;
2076                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2077         }
2078
2079         mutex_unlock(&q->sysfs_lock);
2080
2081         queue_for_each_hw_ctx(q, hctx, i) {
2082                 /*
2083                  * If no software queues are mapped to this hardware queue,
2084                  * disable it and free the request entries.
2085                  */
2086                 if (!hctx->nr_ctx) {
2087                         /* Never unmap queue 0.  We need it as a
2088                          * fallback in case of a new remap fails
2089                          * allocation
2090                          */
2091                         if (i && set->tags[i])
2092                                 blk_mq_free_map_and_requests(set, i);
2093
2094                         hctx->tags = NULL;
2095                         continue;
2096                 }
2097
2098                 hctx->tags = set->tags[i];
2099                 WARN_ON(!hctx->tags);
2100
2101                 /*
2102                  * Set the map size to the number of mapped software queues.
2103                  * This is more accurate and more efficient than looping
2104                  * over all possibly mapped software queues.
2105                  */
2106                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2107
2108                 /*
2109                  * Initialize batch roundrobin counts
2110                  */
2111                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2112                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2113         }
2114 }
2115
2116 /*
2117  * Caller needs to ensure that we're either frozen/quiesced, or that
2118  * the queue isn't live yet.
2119  */
2120 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2121 {
2122         struct blk_mq_hw_ctx *hctx;
2123         int i;
2124
2125         queue_for_each_hw_ctx(q, hctx, i) {
2126                 if (shared) {
2127                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2128                                 atomic_inc(&q->shared_hctx_restart);
2129                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2130                 } else {
2131                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2132                                 atomic_dec(&q->shared_hctx_restart);
2133                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2134                 }
2135         }
2136 }
2137
2138 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2139                                         bool shared)
2140 {
2141         struct request_queue *q;
2142
2143         lockdep_assert_held(&set->tag_list_lock);
2144
2145         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2146                 blk_mq_freeze_queue(q);
2147                 queue_set_hctx_shared(q, shared);
2148                 blk_mq_unfreeze_queue(q);
2149         }
2150 }
2151
2152 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2153 {
2154         struct blk_mq_tag_set *set = q->tag_set;
2155
2156         mutex_lock(&set->tag_list_lock);
2157         list_del_rcu(&q->tag_set_list);
2158         INIT_LIST_HEAD(&q->tag_set_list);
2159         if (list_is_singular(&set->tag_list)) {
2160                 /* just transitioned to unshared */
2161                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2162                 /* update existing queue */
2163                 blk_mq_update_tag_set_depth(set, false);
2164         }
2165         mutex_unlock(&set->tag_list_lock);
2166
2167         synchronize_rcu();
2168 }
2169
2170 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2171                                      struct request_queue *q)
2172 {
2173         q->tag_set = set;
2174
2175         mutex_lock(&set->tag_list_lock);
2176
2177         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2178         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2179                 set->flags |= BLK_MQ_F_TAG_SHARED;
2180                 /* update existing queue */
2181                 blk_mq_update_tag_set_depth(set, true);
2182         }
2183         if (set->flags & BLK_MQ_F_TAG_SHARED)
2184                 queue_set_hctx_shared(q, true);
2185         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2186
2187         mutex_unlock(&set->tag_list_lock);
2188 }
2189
2190 /*
2191  * It is the actual release handler for mq, but we do it from
2192  * request queue's release handler for avoiding use-after-free
2193  * and headache because q->mq_kobj shouldn't have been introduced,
2194  * but we can't group ctx/kctx kobj without it.
2195  */
2196 void blk_mq_release(struct request_queue *q)
2197 {
2198         struct blk_mq_hw_ctx *hctx;
2199         unsigned int i;
2200
2201         /* hctx kobj stays in hctx */
2202         queue_for_each_hw_ctx(q, hctx, i) {
2203                 if (!hctx)
2204                         continue;
2205                 kobject_put(&hctx->kobj);
2206         }
2207
2208         q->mq_map = NULL;
2209
2210         kfree(q->queue_hw_ctx);
2211
2212         /*
2213          * release .mq_kobj and sw queue's kobject now because
2214          * both share lifetime with request queue.
2215          */
2216         blk_mq_sysfs_deinit(q);
2217
2218         free_percpu(q->queue_ctx);
2219 }
2220
2221 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2222 {
2223         struct request_queue *uninit_q, *q;
2224
2225         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2226         if (!uninit_q)
2227                 return ERR_PTR(-ENOMEM);
2228
2229         q = blk_mq_init_allocated_queue(set, uninit_q);
2230         if (IS_ERR(q))
2231                 blk_cleanup_queue(uninit_q);
2232
2233         return q;
2234 }
2235 EXPORT_SYMBOL(blk_mq_init_queue);
2236
2237 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2238 {
2239         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2240
2241         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2242                            __alignof__(struct blk_mq_hw_ctx)) !=
2243                      sizeof(struct blk_mq_hw_ctx));
2244
2245         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2246                 hw_ctx_size += sizeof(struct srcu_struct);
2247
2248         return hw_ctx_size;
2249 }
2250
2251 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2252                                                 struct request_queue *q)
2253 {
2254         int i, j;
2255         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2256
2257         blk_mq_sysfs_unregister(q);
2258         for (i = 0; i < set->nr_hw_queues; i++) {
2259                 int node;
2260
2261                 if (hctxs[i])
2262                         continue;
2263
2264                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2265                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2266                                         GFP_KERNEL, node);
2267                 if (!hctxs[i])
2268                         break;
2269
2270                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2271                                                 node)) {
2272                         kfree(hctxs[i]);
2273                         hctxs[i] = NULL;
2274                         break;
2275                 }
2276
2277                 atomic_set(&hctxs[i]->nr_active, 0);
2278                 hctxs[i]->numa_node = node;
2279                 hctxs[i]->queue_num = i;
2280
2281                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2282                         free_cpumask_var(hctxs[i]->cpumask);
2283                         kfree(hctxs[i]);
2284                         hctxs[i] = NULL;
2285                         break;
2286                 }
2287                 blk_mq_hctx_kobj_init(hctxs[i]);
2288         }
2289         for (j = i; j < q->nr_hw_queues; j++) {
2290                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2291
2292                 if (hctx) {
2293                         if (hctx->tags)
2294                                 blk_mq_free_map_and_requests(set, j);
2295                         blk_mq_exit_hctx(q, set, hctx, j);
2296                         kobject_put(&hctx->kobj);
2297                         hctxs[j] = NULL;
2298
2299                 }
2300         }
2301         q->nr_hw_queues = i;
2302         blk_mq_sysfs_register(q);
2303 }
2304
2305 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2306                                                   struct request_queue *q)
2307 {
2308         /* mark the queue as mq asap */
2309         q->mq_ops = set->ops;
2310
2311         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2312                                              blk_mq_poll_stats_bkt,
2313                                              BLK_MQ_POLL_STATS_BKTS, q);
2314         if (!q->poll_cb)
2315                 goto err_exit;
2316
2317         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2318         if (!q->queue_ctx)
2319                 goto err_exit;
2320
2321         /* init q->mq_kobj and sw queues' kobjects */
2322         blk_mq_sysfs_init(q);
2323
2324         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2325                                                 GFP_KERNEL, set->numa_node);
2326         if (!q->queue_hw_ctx)
2327                 goto err_percpu;
2328
2329         q->mq_map = set->mq_map;
2330
2331         blk_mq_realloc_hw_ctxs(set, q);
2332         if (!q->nr_hw_queues)
2333                 goto err_hctxs;
2334
2335         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2336         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2337
2338         q->nr_queues = nr_cpu_ids;
2339
2340         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2341
2342         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2343                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2344
2345         q->sg_reserved_size = INT_MAX;
2346
2347         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2348         INIT_LIST_HEAD(&q->requeue_list);
2349         spin_lock_init(&q->requeue_lock);
2350
2351         blk_queue_make_request(q, blk_mq_make_request);
2352
2353         /*
2354          * Do this after blk_queue_make_request() overrides it...
2355          */
2356         q->nr_requests = set->queue_depth;
2357
2358         /*
2359          * Default to classic polling
2360          */
2361         q->poll_nsec = -1;
2362
2363         if (set->ops->complete)
2364                 blk_queue_softirq_done(q, set->ops->complete);
2365
2366         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2367         blk_mq_add_queue_tag_set(set, q);
2368         blk_mq_map_swqueue(q);
2369
2370         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2371                 int ret;
2372
2373                 ret = blk_mq_sched_init(q);
2374                 if (ret)
2375                         return ERR_PTR(ret);
2376         }
2377
2378         return q;
2379
2380 err_hctxs:
2381         kfree(q->queue_hw_ctx);
2382 err_percpu:
2383         free_percpu(q->queue_ctx);
2384 err_exit:
2385         q->mq_ops = NULL;
2386         return ERR_PTR(-ENOMEM);
2387 }
2388 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2389
2390 void blk_mq_free_queue(struct request_queue *q)
2391 {
2392         struct blk_mq_tag_set   *set = q->tag_set;
2393
2394         blk_mq_del_queue_tag_set(q);
2395         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2396 }
2397
2398 /* Basically redo blk_mq_init_queue with queue frozen */
2399 static void blk_mq_queue_reinit(struct request_queue *q)
2400 {
2401         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2402
2403         blk_mq_debugfs_unregister_hctxs(q);
2404         blk_mq_sysfs_unregister(q);
2405
2406         /*
2407          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2408          * we should change hctx numa_node according to new topology (this
2409          * involves free and re-allocate memory, worthy doing?)
2410          */
2411
2412         blk_mq_map_swqueue(q);
2413
2414         blk_mq_sysfs_register(q);
2415         blk_mq_debugfs_register_hctxs(q);
2416 }
2417
2418 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2419 {
2420         int i;
2421
2422         for (i = 0; i < set->nr_hw_queues; i++)
2423                 if (!__blk_mq_alloc_rq_map(set, i))
2424                         goto out_unwind;
2425
2426         return 0;
2427
2428 out_unwind:
2429         while (--i >= 0)
2430                 blk_mq_free_rq_map(set->tags[i]);
2431
2432         return -ENOMEM;
2433 }
2434
2435 /*
2436  * Allocate the request maps associated with this tag_set. Note that this
2437  * may reduce the depth asked for, if memory is tight. set->queue_depth
2438  * will be updated to reflect the allocated depth.
2439  */
2440 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2441 {
2442         unsigned int depth;
2443         int err;
2444
2445         depth = set->queue_depth;
2446         do {
2447                 err = __blk_mq_alloc_rq_maps(set);
2448                 if (!err)
2449                         break;
2450
2451                 set->queue_depth >>= 1;
2452                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2453                         err = -ENOMEM;
2454                         break;
2455                 }
2456         } while (set->queue_depth);
2457
2458         if (!set->queue_depth || err) {
2459                 pr_err("blk-mq: failed to allocate request map\n");
2460                 return -ENOMEM;
2461         }
2462
2463         if (depth != set->queue_depth)
2464                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2465                                                 depth, set->queue_depth);
2466
2467         return 0;
2468 }
2469
2470 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2471 {
2472         if (set->ops->map_queues)
2473                 return set->ops->map_queues(set);
2474         else
2475                 return blk_mq_map_queues(set);
2476 }
2477
2478 /*
2479  * Alloc a tag set to be associated with one or more request queues.
2480  * May fail with EINVAL for various error conditions. May adjust the
2481  * requested depth down, if if it too large. In that case, the set
2482  * value will be stored in set->queue_depth.
2483  */
2484 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2485 {
2486         int ret;
2487
2488         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2489
2490         if (!set->nr_hw_queues)
2491                 return -EINVAL;
2492         if (!set->queue_depth)
2493                 return -EINVAL;
2494         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2495                 return -EINVAL;
2496
2497         if (!set->ops->queue_rq)
2498                 return -EINVAL;
2499
2500         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2501                 pr_info("blk-mq: reduced tag depth to %u\n",
2502                         BLK_MQ_MAX_DEPTH);
2503                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2504         }
2505
2506         /*
2507          * If a crashdump is active, then we are potentially in a very
2508          * memory constrained environment. Limit us to 1 queue and
2509          * 64 tags to prevent using too much memory.
2510          */
2511         if (is_kdump_kernel()) {
2512                 set->nr_hw_queues = 1;
2513                 set->queue_depth = min(64U, set->queue_depth);
2514         }
2515         /*
2516          * There is no use for more h/w queues than cpus.
2517          */
2518         if (set->nr_hw_queues > nr_cpu_ids)
2519                 set->nr_hw_queues = nr_cpu_ids;
2520
2521         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2522                                  GFP_KERNEL, set->numa_node);
2523         if (!set->tags)
2524                 return -ENOMEM;
2525
2526         ret = -ENOMEM;
2527         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2528                         GFP_KERNEL, set->numa_node);
2529         if (!set->mq_map)
2530                 goto out_free_tags;
2531
2532         ret = blk_mq_update_queue_map(set);
2533         if (ret)
2534                 goto out_free_mq_map;
2535
2536         ret = blk_mq_alloc_rq_maps(set);
2537         if (ret)
2538                 goto out_free_mq_map;
2539
2540         mutex_init(&set->tag_list_lock);
2541         INIT_LIST_HEAD(&set->tag_list);
2542
2543         return 0;
2544
2545 out_free_mq_map:
2546         kfree(set->mq_map);
2547         set->mq_map = NULL;
2548 out_free_tags:
2549         kfree(set->tags);
2550         set->tags = NULL;
2551         return ret;
2552 }
2553 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2554
2555 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2556 {
2557         int i;
2558
2559         for (i = 0; i < nr_cpu_ids; i++)
2560                 blk_mq_free_map_and_requests(set, i);
2561
2562         kfree(set->mq_map);
2563         set->mq_map = NULL;
2564
2565         kfree(set->tags);
2566         set->tags = NULL;
2567 }
2568 EXPORT_SYMBOL(blk_mq_free_tag_set);
2569
2570 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2571 {
2572         struct blk_mq_tag_set *set = q->tag_set;
2573         struct blk_mq_hw_ctx *hctx;
2574         int i, ret;
2575
2576         if (!set)
2577                 return -EINVAL;
2578
2579         blk_mq_freeze_queue(q);
2580
2581         ret = 0;
2582         queue_for_each_hw_ctx(q, hctx, i) {
2583                 if (!hctx->tags)
2584                         continue;
2585                 /*
2586                  * If we're using an MQ scheduler, just update the scheduler
2587                  * queue depth. This is similar to what the old code would do.
2588                  */
2589                 if (!hctx->sched_tags) {
2590                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2591                                                         min(nr, set->queue_depth),
2592                                                         false);
2593                 } else {
2594                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2595                                                         nr, true);
2596                 }
2597                 if (ret)
2598                         break;
2599         }
2600
2601         if (!ret)
2602                 q->nr_requests = nr;
2603
2604         blk_mq_unfreeze_queue(q);
2605
2606         return ret;
2607 }
2608
2609 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2610                                                         int nr_hw_queues)
2611 {
2612         struct request_queue *q;
2613
2614         lockdep_assert_held(&set->tag_list_lock);
2615
2616         if (nr_hw_queues > nr_cpu_ids)
2617                 nr_hw_queues = nr_cpu_ids;
2618         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2619                 return;
2620
2621         list_for_each_entry(q, &set->tag_list, tag_set_list)
2622                 blk_mq_freeze_queue(q);
2623
2624         set->nr_hw_queues = nr_hw_queues;
2625         blk_mq_update_queue_map(set);
2626         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2627                 blk_mq_realloc_hw_ctxs(set, q);
2628                 blk_mq_queue_reinit(q);
2629         }
2630
2631         list_for_each_entry(q, &set->tag_list, tag_set_list)
2632                 blk_mq_unfreeze_queue(q);
2633 }
2634
2635 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2636 {
2637         mutex_lock(&set->tag_list_lock);
2638         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2639         mutex_unlock(&set->tag_list_lock);
2640 }
2641 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2642
2643 /* Enable polling stats and return whether they were already enabled. */
2644 static bool blk_poll_stats_enable(struct request_queue *q)
2645 {
2646         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2647             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2648                 return true;
2649         blk_stat_add_callback(q, q->poll_cb);
2650         return false;
2651 }
2652
2653 static void blk_mq_poll_stats_start(struct request_queue *q)
2654 {
2655         /*
2656          * We don't arm the callback if polling stats are not enabled or the
2657          * callback is already active.
2658          */
2659         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2660             blk_stat_is_active(q->poll_cb))
2661                 return;
2662
2663         blk_stat_activate_msecs(q->poll_cb, 100);
2664 }
2665
2666 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2667 {
2668         struct request_queue *q = cb->data;
2669         int bucket;
2670
2671         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2672                 if (cb->stat[bucket].nr_samples)
2673                         q->poll_stat[bucket] = cb->stat[bucket];
2674         }
2675 }
2676
2677 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2678                                        struct blk_mq_hw_ctx *hctx,
2679                                        struct request *rq)
2680 {
2681         unsigned long ret = 0;
2682         int bucket;
2683
2684         /*
2685          * If stats collection isn't on, don't sleep but turn it on for
2686          * future users
2687          */
2688         if (!blk_poll_stats_enable(q))
2689                 return 0;
2690
2691         /*
2692          * As an optimistic guess, use half of the mean service time
2693          * for this type of request. We can (and should) make this smarter.
2694          * For instance, if the completion latencies are tight, we can
2695          * get closer than just half the mean. This is especially
2696          * important on devices where the completion latencies are longer
2697          * than ~10 usec. We do use the stats for the relevant IO size
2698          * if available which does lead to better estimates.
2699          */
2700         bucket = blk_mq_poll_stats_bkt(rq);
2701         if (bucket < 0)
2702                 return ret;
2703
2704         if (q->poll_stat[bucket].nr_samples)
2705                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2706
2707         return ret;
2708 }
2709
2710 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2711                                      struct blk_mq_hw_ctx *hctx,
2712                                      struct request *rq)
2713 {
2714         struct hrtimer_sleeper hs;
2715         enum hrtimer_mode mode;
2716         unsigned int nsecs;
2717         ktime_t kt;
2718
2719         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2720                 return false;
2721
2722         /*
2723          * poll_nsec can be:
2724          *
2725          * -1:  don't ever hybrid sleep
2726          *  0:  use half of prev avg
2727          * >0:  use this specific value
2728          */
2729         if (q->poll_nsec == -1)
2730                 return false;
2731         else if (q->poll_nsec > 0)
2732                 nsecs = q->poll_nsec;
2733         else
2734                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2735
2736         if (!nsecs)
2737                 return false;
2738
2739         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2740
2741         /*
2742          * This will be replaced with the stats tracking code, using
2743          * 'avg_completion_time / 2' as the pre-sleep target.
2744          */
2745         kt = nsecs;
2746
2747         mode = HRTIMER_MODE_REL;
2748         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2749         hrtimer_set_expires(&hs.timer, kt);
2750
2751         hrtimer_init_sleeper(&hs, current);
2752         do {
2753                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2754                         break;
2755                 set_current_state(TASK_UNINTERRUPTIBLE);
2756                 hrtimer_start_expires(&hs.timer, mode);
2757                 if (hs.task)
2758                         io_schedule();
2759                 hrtimer_cancel(&hs.timer);
2760                 mode = HRTIMER_MODE_ABS;
2761         } while (hs.task && !signal_pending(current));
2762
2763         __set_current_state(TASK_RUNNING);
2764         destroy_hrtimer_on_stack(&hs.timer);
2765         return true;
2766 }
2767
2768 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2769 {
2770         struct request_queue *q = hctx->queue;
2771         long state;
2772
2773         /*
2774          * If we sleep, have the caller restart the poll loop to reset
2775          * the state. Like for the other success return cases, the
2776          * caller is responsible for checking if the IO completed. If
2777          * the IO isn't complete, we'll get called again and will go
2778          * straight to the busy poll loop.
2779          */
2780         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2781                 return true;
2782
2783         hctx->poll_considered++;
2784
2785         state = current->state;
2786         while (!need_resched()) {
2787                 int ret;
2788
2789                 hctx->poll_invoked++;
2790
2791                 ret = q->mq_ops->poll(hctx, rq->tag);
2792                 if (ret > 0) {
2793                         hctx->poll_success++;
2794                         set_current_state(TASK_RUNNING);
2795                         return true;
2796                 }
2797
2798                 if (signal_pending_state(state, current))
2799                         set_current_state(TASK_RUNNING);
2800
2801                 if (current->state == TASK_RUNNING)
2802                         return true;
2803                 if (ret < 0)
2804                         break;
2805                 cpu_relax();
2806         }
2807
2808         return false;
2809 }
2810
2811 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2812 {
2813         struct blk_mq_hw_ctx *hctx;
2814         struct blk_plug *plug;
2815         struct request *rq;
2816
2817         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2818             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2819                 return false;
2820
2821         plug = current->plug;
2822         if (plug)
2823                 blk_flush_plug_list(plug, false);
2824
2825         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2826         if (!blk_qc_t_is_internal(cookie))
2827                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2828         else {
2829                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2830                 /*
2831                  * With scheduling, if the request has completed, we'll
2832                  * get a NULL return here, as we clear the sched tag when
2833                  * that happens. The request still remains valid, like always,
2834                  * so we should be safe with just the NULL check.
2835                  */
2836                 if (!rq)
2837                         return false;
2838         }
2839
2840         return __blk_mq_poll(hctx, rq);
2841 }
2842 EXPORT_SYMBOL_GPL(blk_mq_poll);
2843
2844 static int __init blk_mq_init(void)
2845 {
2846         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2847                                 blk_mq_hctx_notify_dead);
2848         return 0;
2849 }
2850 subsys_initcall(blk_mq_init);