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 struct mq_inflight {
  87         struct hd_struct *part;
  88         unsigned int *inflight;
  89 };
  90
  91 static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  92                                   struct request *rq, void *priv,
  93                                   bool reserved)
  94 {
  95         struct mq_inflight *mi = priv;
  96
  97         if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
  98             !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
  99                 /*
 100                  * index[0] counts the specific partition that was asked
 101                  * for. index[1] counts the ones that are active on the
 102                  * whole device, so increment that if mi->part is indeed
 103                  * a partition, and not a whole device.
 104                  */
 105                 if (rq->part == mi->part)
 106                         mi->inflight[0]++;
 107                 if (mi->part->partno)
 108                         mi->inflight[1]++;
 109         }
 110 }
 111
 112 void blk_mq_in_flight(struct request_queue *q, struct hd_struct *part,
 113                       unsigned int inflight[2])
 114 {
 115         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 116
 117         inflight[0] = inflight[1] = 0;
 118         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 119 }
 120
 121 void blk_freeze_queue_start(struct request_queue *q)
 122 {
 123         int freeze_depth;
 124
 125         freeze_depth = atomic_inc_return(&q->mq_freeze_depth);
 126         if (freeze_depth == 1) {
 127                 percpu_ref_kill(&q->q_usage_counter);
 128                 blk_mq_run_hw_queues(q, false);
 129         }
 130 }
 131 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 132
 133 void blk_mq_freeze_queue_wait(struct request_queue *q)
 134 {
 135         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 136 }
 137 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 138
 139 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 140                                      unsigned long timeout)
 141 {
 142         return wait_event_timeout(q->mq_freeze_wq,
 143                                         percpu_ref_is_zero(&q->q_usage_counter),
 144                                         timeout);
 145 }
 146 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 147
 148 /*
 149  * Guarantee no request is in use, so we can change any data structure of
 150  * the queue afterward.
 151  */
 152 void blk_freeze_queue(struct request_queue *q)
 153 {
 154         /*
 155          * In the !blk_mq case we are only calling this to kill the
 156          * q_usage_counter, otherwise this increases the freeze depth
 157          * and waits for it to return to zero.  For this reason there is
 158          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 159          * exported to drivers as the only user for unfreeze is blk_mq.
 160          */
 161         blk_freeze_queue_start(q);
 162         if (!q->mq_ops)
 163                 blk_drain_queue(q);
 164         blk_mq_freeze_queue_wait(q);
 165 }
 166
 167 void blk_mq_freeze_queue(struct request_queue *q)
 168 {
 169         /*
 170          * ...just an alias to keep freeze and unfreeze actions balanced
 171          * in the blk_mq_* namespace
 172          */
 173         blk_freeze_queue(q);
 174 }
 175 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 176
 177 void blk_mq_unfreeze_queue(struct request_queue *q)
 178 {
 179         int freeze_depth;
 180
 181         freeze_depth = atomic_dec_return(&q->mq_freeze_depth);
 182         WARN_ON_ONCE(freeze_depth < 0);
 183         if (!freeze_depth) {
 184                 percpu_ref_reinit(&q->q_usage_counter);
 185                 wake_up_all(&q->mq_freeze_wq);
 186         }
 187 }
 188 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 189
 190 /*
 191  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 192  * mpt3sas driver such that this function can be removed.
 193  */
 194 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 195 {
 196         unsigned long flags;
 197
 198         spin_lock_irqsave(q->queue_lock, flags);
 199         queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 200         spin_unlock_irqrestore(q->queue_lock, flags);
 201 }
 202 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 203
 204 /**
 205  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 206  * @q: request queue.
 207  *
 208  * Note: this function does not prevent that the struct request end_io()
 209  * callback function is invoked. Once this function is returned, we make
 210  * sure no dispatch can happen until the queue is unquiesced via
 211  * blk_mq_unquiesce_queue().
 212  */
 213 void blk_mq_quiesce_queue(struct request_queue *q)
 214 {
 215         struct blk_mq_hw_ctx *hctx;
 216         unsigned int i;
 217         bool rcu = false;
 218
 219         blk_mq_quiesce_queue_nowait(q);
 220
 221         queue_for_each_hw_ctx(q, hctx, i) {
 222                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 223                         synchronize_srcu(hctx->queue_rq_srcu);
 224                 else
 225                         rcu = true;
 226         }
 227         if (rcu)
 228                 synchronize_rcu();
 229 }
 230 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 231
 232 /*
 233  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 234  * @q: request queue.
 235  *
 236  * This function recovers queue into the state before quiescing
 237  * which is done by blk_mq_quiesce_queue.
 238  */
 239 void blk_mq_unquiesce_queue(struct request_queue *q)
 240 {
 241         unsigned long flags;
 242
 243         spin_lock_irqsave(q->queue_lock, flags);
 244         queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 245         spin_unlock_irqrestore(q->queue_lock, flags);
 246
 247         /* dispatch requests which are inserted during quiescing */
 248         blk_mq_run_hw_queues(q, true);
 249 }
 250 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 251
 252 void blk_mq_wake_waiters(struct request_queue *q)
 253 {
 254         struct blk_mq_hw_ctx *hctx;
 255         unsigned int i;
 256
 257         queue_for_each_hw_ctx(q, hctx, i)
 258                 if (blk_mq_hw_queue_mapped(hctx))
 259                         blk_mq_tag_wakeup_all(hctx->tags, true);
 260
 261         /*
 262          * If we are called because the queue has now been marked as
 263          * dying, we need to ensure that processes currently waiting on
 264          * the queue are notified as well.
 265          */
 266         wake_up_all(&q->mq_freeze_wq);
 267 }
 268
 269 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 270 {
 271         return blk_mq_has_free_tags(hctx->tags);
 272 }
 273 EXPORT_SYMBOL(blk_mq_can_queue);
 274
 275 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 276                 unsigned int tag, unsigned int op)
 277 {
 278         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 279         struct request *rq = tags->static_rqs[tag];
 280
 281         rq->rq_flags = 0;
 282
 283         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 284                 rq->tag = -1;
 285                 rq->internal_tag = tag;
 286         } else {
 287                 if (blk_mq_tag_busy(data->hctx)) {
 288                         rq->rq_flags = RQF_MQ_INFLIGHT;
 289                         atomic_inc(&data->hctx->nr_active);
 290                 }
 291                 rq->tag = tag;
 292                 rq->internal_tag = -1;
 293                 data->hctx->tags->rqs[rq->tag] = rq;
 294         }
 295
 296         INIT_LIST_HEAD(&rq->queuelist);
 297         /* csd/requeue_work/fifo_time is initialized before use */
 298         rq->q = data->q;
 299         rq->mq_ctx = data->ctx;
 300         rq->cmd_flags = op;
 301         if (blk_queue_io_stat(data->q))
 302                 rq->rq_flags |= RQF_IO_STAT;
 303         /* do not touch atomic flags, it needs atomic ops against the timer */
 304         rq->cpu = -1;
 305         INIT_HLIST_NODE(&rq->hash);
 306         RB_CLEAR_NODE(&rq->rb_node);
 307         rq->rq_disk = NULL;
 308         rq->part = NULL;
 309         rq->start_time = jiffies;
 310 #ifdef CONFIG_BLK_CGROUP
 311         rq->rl = NULL;
 312         set_start_time_ns(rq);
 313         rq->io_start_time_ns = 0;
 314 #endif
 315         rq->nr_phys_segments = 0;
 316 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 317         rq->nr_integrity_segments = 0;
 318 #endif
 319         rq->special = NULL;
 320         /* tag was already set */
 321         rq->extra_len = 0;
 322
 323         INIT_LIST_HEAD(&rq->timeout_list);
 324         rq->timeout = 0;
 325
 326         rq->end_io = NULL;
 327         rq->end_io_data = NULL;
 328         rq->next_rq = NULL;
 329
 330         data->ctx->rq_dispatched[op_is_sync(op)]++;
 331         return rq;
 332 }
 333
 334 static struct request *blk_mq_get_request(struct request_queue *q,
 335                 struct bio *bio, unsigned int op,
 336                 struct blk_mq_alloc_data *data)
 337 {
 338         struct elevator_queue *e = q->elevator;
 339         struct request *rq;
 340         unsigned int tag;
 341         struct blk_mq_ctx *local_ctx = NULL;
 342
 343         blk_queue_enter_live(q);
 344         data->q = q;
 345         if (likely(!data->ctx))
 346                 data->ctx = local_ctx = blk_mq_get_ctx(q);
 347         if (likely(!data->hctx))
 348                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 349         if (op & REQ_NOWAIT)
 350                 data->flags |= BLK_MQ_REQ_NOWAIT;
 351
 352         if (e) {
 353                 data->flags |= BLK_MQ_REQ_INTERNAL;
 354
 355                 /*
 356                  * Flush requests are special and go directly to the
 357                  * dispatch list.
 358                  */
 359                 if (!op_is_flush(op) && e->type->ops.mq.limit_depth)
 360                         e->type->ops.mq.limit_depth(op, data);
 361         }
 362
 363         tag = blk_mq_get_tag(data);
 364         if (tag == BLK_MQ_TAG_FAIL) {
 365                 if (local_ctx) {
 366                         blk_mq_put_ctx(local_ctx);
 367                         data->ctx = NULL;
 368                 }
 369                 blk_queue_exit(q);
 370                 return NULL;
 371         }
 372
 373         rq = blk_mq_rq_ctx_init(data, tag, op);
 374         if (!op_is_flush(op)) {
 375                 rq->elv.icq = NULL;
 376                 if (e && e->type->ops.mq.prepare_request) {
 377                         if (e->type->icq_cache && rq_ioc(bio))
 378                                 blk_mq_sched_assign_ioc(rq, bio);
 379
 380                         e->type->ops.mq.prepare_request(rq, bio);
 381                         rq->rq_flags |= RQF_ELVPRIV;
 382                 }
 383         }
 384         data->hctx->queued++;
 385         return rq;
 386 }
 387
 388 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 389                 unsigned int flags)
 390 {
 391         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 392         struct request *rq;
 393         int ret;
 394
 395         ret = blk_queue_enter(q, flags & BLK_MQ_REQ_NOWAIT);
 396         if (ret)
 397                 return ERR_PTR(ret);
 398
 399         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 400         blk_queue_exit(q);
 401
 402         if (!rq)
 403                 return ERR_PTR(-EWOULDBLOCK);
 404
 405         blk_mq_put_ctx(alloc_data.ctx);
 406
 407         rq->__data_len = 0;
 408         rq->__sector = (sector_t) -1;
 409         rq->bio = rq->biotail = NULL;
 410         return rq;
 411 }
 412 EXPORT_SYMBOL(blk_mq_alloc_request);
 413
 414 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 415                 unsigned int op, unsigned int flags, unsigned int hctx_idx)
 416 {
 417         struct blk_mq_alloc_data alloc_data = { .flags = flags };
 418         struct request *rq;
 419         unsigned int cpu;
 420         int ret;
 421
 422         /*
 423          * If the tag allocator sleeps we could get an allocation for a
 424          * different hardware context.  No need to complicate the low level
 425          * allocator for this for the rare use case of a command tied to
 426          * a specific queue.
 427          */
 428         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 429                 return ERR_PTR(-EINVAL);
 430
 431         if (hctx_idx >= q->nr_hw_queues)
 432                 return ERR_PTR(-EIO);
 433
 434         ret = blk_queue_enter(q, true);
 435         if (ret)
 436                 return ERR_PTR(ret);
 437
 438         /*
 439          * Check if the hardware context is actually mapped to anything.
 440          * If not tell the caller that it should skip this queue.
 441          */
 442         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 443         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 444                 blk_queue_exit(q);
 445                 return ERR_PTR(-EXDEV);
 446         }
 447         cpu = cpumask_first(alloc_data.hctx->cpumask);
 448         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 449
 450         rq = blk_mq_get_request(q, NULL, op, &alloc_data);
 451         blk_queue_exit(q);
 452
 453         if (!rq)
 454                 return ERR_PTR(-EWOULDBLOCK);
 455
 456         return rq;
 457 }
 458 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 459
 460 void blk_mq_free_request(struct request *rq)
 461 {
 462         struct request_queue *q = rq->q;
 463         struct elevator_queue *e = q->elevator;
 464         struct blk_mq_ctx *ctx = rq->mq_ctx;
 465         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 466         const int sched_tag = rq->internal_tag;
 467
 468         if (rq->rq_flags & RQF_ELVPRIV) {
 469                 if (e && e->type->ops.mq.finish_request)
 470                         e->type->ops.mq.finish_request(rq);
 471                 if (rq->elv.icq) {
 472                         put_io_context(rq->elv.icq->ioc);
 473                         rq->elv.icq = NULL;
 474                 }
 475         }
 476
 477         ctx->rq_completed[rq_is_sync(rq)]++;
 478         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 479                 atomic_dec(&hctx->nr_active);
 480
 481         wbt_done(q->rq_wb, &rq->issue_stat);
 482
 483         clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 484         clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 485         if (rq->tag != -1)
 486                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 487         if (sched_tag != -1)
 488                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 489         blk_mq_sched_restart(hctx);
 490         blk_queue_exit(q);
 491 }
 492 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 493
 494 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 495 {
 496         blk_account_io_done(rq);
 497
 498         if (rq->end_io) {
 499                 wbt_done(rq->q->rq_wb, &rq->issue_stat);
 500                 rq->end_io(rq, error);
 501         } else {
 502                 if (unlikely(blk_bidi_rq(rq)))
 503                         blk_mq_free_request(rq->next_rq);
 504                 blk_mq_free_request(rq);
 505         }
 506 }
 507 EXPORT_SYMBOL(__blk_mq_end_request);
 508
 509 void blk_mq_end_request(struct request *rq, blk_status_t error)
 510 {
 511         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 512                 BUG();
 513         __blk_mq_end_request(rq, error);
 514 }
 515 EXPORT_SYMBOL(blk_mq_end_request);
 516
 517 static void __blk_mq_complete_request_remote(void *data)
 518 {
 519         struct request *rq = data;
 520
 521         rq->q->softirq_done_fn(rq);
 522 }
 523
 524 static void __blk_mq_complete_request(struct request *rq)
 525 {
 526         struct blk_mq_ctx *ctx = rq->mq_ctx;
 527         bool shared = false;
 528         int cpu;
 529
 530         if (rq->internal_tag != -1)
 531                 blk_mq_sched_completed_request(rq);
 532         if (rq->rq_flags & RQF_STATS) {
 533                 blk_mq_poll_stats_start(rq->q);
 534                 blk_stat_add(rq);
 535         }
 536
 537         if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
 538                 rq->q->softirq_done_fn(rq);
 539                 return;
 540         }
 541
 542         cpu = get_cpu();
 543         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
 544                 shared = cpus_share_cache(cpu, ctx->cpu);
 545
 546         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 547                 rq->csd.func = __blk_mq_complete_request_remote;
 548                 rq->csd.info = rq;
 549                 rq->csd.flags = 0;
 550                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 551         } else {
 552                 rq->q->softirq_done_fn(rq);
 553         }
 554         put_cpu();
 555 }
 556
 557 /**
 558  * blk_mq_complete_request - end I/O on a request
 559  * @rq:         the request being processed
 560  *
 561  * Description:
 562  *      Ends all I/O on a request. It does not handle partial completions.
 563  *      The actual completion happens out-of-order, through a IPI handler.
 564  **/
 565 void blk_mq_complete_request(struct request *rq)
 566 {
 567         struct request_queue *q = rq->q;
 568
 569         if (unlikely(blk_should_fake_timeout(q)))
 570                 return;
 571         if (!blk_mark_rq_complete(rq))
 572                 __blk_mq_complete_request(rq);
 573 }
 574 EXPORT_SYMBOL(blk_mq_complete_request);
 575
 576 int blk_mq_request_started(struct request *rq)
 577 {
 578         return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 579 }
 580 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 581
 582 void blk_mq_start_request(struct request *rq)
 583 {
 584         struct request_queue *q = rq->q;
 585
 586         blk_mq_sched_started_request(rq);
 587
 588         trace_block_rq_issue(q, rq);
 589
 590         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 591                 blk_stat_set_issue(&rq->issue_stat, blk_rq_sectors(rq));
 592                 rq->rq_flags |= RQF_STATS;
 593                 wbt_issue(q->rq_wb, &rq->issue_stat);
 594         }
 595
 596         blk_add_timer(rq);
 597
 598         /*
 599          * Ensure that ->deadline is visible before set the started
 600          * flag and clear the completed flag.
 601          */
 602         smp_mb__before_atomic();
 603
 604         /*
 605          * Mark us as started and clear complete. Complete might have been
 606          * set if requeue raced with timeout, which then marked it as
 607          * complete. So be sure to clear complete again when we start
 608          * the request, otherwise we'll ignore the completion event.
 609          */
 610         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 611                 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 612         if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 613                 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 614
 615         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 616                 /*
 617                  * Make sure space for the drain appears.  We know we can do
 618                  * this because max_hw_segments has been adjusted to be one
 619                  * fewer than the device can handle.
 620                  */
 621                 rq->nr_phys_segments++;
 622         }
 623 }
 624 EXPORT_SYMBOL(blk_mq_start_request);
 625
 626 /*
 627  * When we reach here because queue is busy, REQ_ATOM_COMPLETE
 628  * flag isn't set yet, so there may be race with timeout handler,
 629  * but given rq->deadline is just set in .queue_rq() under
 630  * this situation, the race won't be possible in reality because
 631  * rq->timeout should be set as big enough to cover the window
 632  * between blk_mq_start_request() called from .queue_rq() and
 633  * clearing REQ_ATOM_STARTED here.
 634  */
 635 static void __blk_mq_requeue_request(struct request *rq)
 636 {
 637         struct request_queue *q = rq->q;
 638
 639         trace_block_rq_requeue(q, rq);
 640         wbt_requeue(q->rq_wb, &rq->issue_stat);
 641
 642         if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
 643                 if (q->dma_drain_size && blk_rq_bytes(rq))
 644                         rq->nr_phys_segments--;
 645         }
 646 }
 647
 648 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 649 {
 650         __blk_mq_requeue_request(rq);
 651
 652         /* this request will be re-inserted to io scheduler queue */
 653         blk_mq_sched_requeue_request(rq);
 654
 655         BUG_ON(blk_queued_rq(rq));
 656         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 657 }
 658 EXPORT_SYMBOL(blk_mq_requeue_request);
 659
 660 static void blk_mq_requeue_work(struct work_struct *work)
 661 {
 662         struct request_queue *q =
 663                 container_of(work, struct request_queue, requeue_work.work);
 664         LIST_HEAD(rq_list);
 665         struct request *rq, *next;
 666
 667         spin_lock_irq(&q->requeue_lock);
 668         list_splice_init(&q->requeue_list, &rq_list);
 669         spin_unlock_irq(&q->requeue_lock);
 670
 671         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 672                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 673                         continue;
 674
 675                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 676                 list_del_init(&rq->queuelist);
 677                 blk_mq_sched_insert_request(rq, true, false, false, true);
 678         }
 679
 680         while (!list_empty(&rq_list)) {
 681                 rq = list_entry(rq_list.next, struct request, queuelist);
 682                 list_del_init(&rq->queuelist);
 683                 blk_mq_sched_insert_request(rq, false, false, false, true);
 684         }
 685
 686         blk_mq_run_hw_queues(q, false);
 687 }
 688
 689 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 690                                 bool kick_requeue_list)
 691 {
 692         struct request_queue *q = rq->q;
 693         unsigned long flags;
 694
 695         /*
 696          * We abuse this flag that is otherwise used by the I/O scheduler to
 697          * request head insertation from the workqueue.
 698          */
 699         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 700
 701         spin_lock_irqsave(&q->requeue_lock, flags);
 702         if (at_head) {
 703                 rq->rq_flags |= RQF_SOFTBARRIER;
 704                 list_add(&rq->queuelist, &q->requeue_list);
 705         } else {
 706                 list_add_tail(&rq->queuelist, &q->requeue_list);
 707         }
 708         spin_unlock_irqrestore(&q->requeue_lock, flags);
 709
 710         if (kick_requeue_list)
 711                 blk_mq_kick_requeue_list(q);
 712 }
 713 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 714
 715 void blk_mq_kick_requeue_list(struct request_queue *q)
 716 {
 717         kblockd_schedule_delayed_work(&q->requeue_work, 0);
 718 }
 719 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 720
 721 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 722                                     unsigned long msecs)
 723 {
 724         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 725                                     msecs_to_jiffies(msecs));
 726 }
 727 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 728
 729 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 730 {
 731         if (tag < tags->nr_tags) {
 732                 prefetch(tags->rqs[tag]);
 733                 return tags->rqs[tag];
 734         }
 735
 736         return NULL;
 737 }
 738 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 739
 740 struct blk_mq_timeout_data {
 741         unsigned long next;
 742         unsigned int next_set;
 743 };
 744
 745 void blk_mq_rq_timed_out(struct request *req, bool reserved)
 746 {
 747         const struct blk_mq_ops *ops = req->q->mq_ops;
 748         enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 749
 750         /*
 751          * We know that complete is set at this point. If STARTED isn't set
 752          * anymore, then the request isn't active and the "timeout" should
 753          * just be ignored. This can happen due to the bitflag ordering.
 754          * Timeout first checks if STARTED is set, and if it is, assumes
 755          * the request is active. But if we race with completion, then
 756          * both flags will get cleared. So check here again, and ignore
 757          * a timeout event with a request that isn't active.
 758          */
 759         if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 760                 return;
 761
 762         if (ops->timeout)
 763                 ret = ops->timeout(req, reserved);
 764
 765         switch (ret) {
 766         case BLK_EH_HANDLED:
 767                 __blk_mq_complete_request(req);
 768                 break;
 769         case BLK_EH_RESET_TIMER:
 770                 blk_add_timer(req);
 771                 blk_clear_rq_complete(req);
 772                 break;
 773         case BLK_EH_NOT_HANDLED:
 774                 break;
 775         default:
 776                 printk(KERN_ERR "block: bad eh return: %d\n", ret);
 777                 break;
 778         }
 779 }
 780
 781 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 782                 struct request *rq, void *priv, bool reserved)
 783 {
 784         struct blk_mq_timeout_data *data = priv;
 785
 786         if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 787                 return;
 788
 789         /*
 790          * The rq being checked may have been freed and reallocated
 791          * out already here, we avoid this race by checking rq->deadline
 792          * and REQ_ATOM_COMPLETE flag together:
 793          *
 794          * - if rq->deadline is observed as new value because of
 795          *   reusing, the rq won't be timed out because of timing.
 796          * - if rq->deadline is observed as previous value,
 797          *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
 798          *   because we put a barrier between setting rq->deadline
 799          *   and clearing the flag in blk_mq_start_request(), so
 800          *   this rq won't be timed out too.
 801          */
 802         if (time_after_eq(jiffies, rq->deadline)) {
 803                 if (!blk_mark_rq_complete(rq))
 804                         blk_mq_rq_timed_out(rq, reserved);
 805         } else if (!data->next_set || time_after(data->next, rq->deadline)) {
 806                 data->next = rq->deadline;
 807                 data->next_set = 1;
 808         }
 809 }
 810
 811 static void blk_mq_timeout_work(struct work_struct *work)
 812 {
 813         struct request_queue *q =
 814                 container_of(work, struct request_queue, timeout_work);
 815         struct blk_mq_timeout_data data = {
 816                 .next           = 0,
 817                 .next_set       = 0,
 818         };
 819         int i;
 820
 821         /* A deadlock might occur if a request is stuck requiring a
 822          * timeout at the same time a queue freeze is waiting
 823          * completion, since the timeout code would not be able to
 824          * acquire the queue reference here.
 825          *
 826          * That's why we don't use blk_queue_enter here; instead, we use
 827          * percpu_ref_tryget directly, because we need to be able to
 828          * obtain a reference even in the short window between the queue
 829          * starting to freeze, by dropping the first reference in
 830          * blk_freeze_queue_start, and the moment the last request is
 831          * consumed, marked by the instant q_usage_counter reaches
 832          * zero.
 833          */
 834         if (!percpu_ref_tryget(&q->q_usage_counter))
 835                 return;
 836
 837         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 838
 839         if (data.next_set) {
 840                 data.next = blk_rq_timeout(round_jiffies_up(data.next));
 841                 mod_timer(&q->timeout, data.next);
 842         } else {
 843                 struct blk_mq_hw_ctx *hctx;
 844
 845                 queue_for_each_hw_ctx(q, hctx, i) {
 846                         /* the hctx may be unmapped, so check it here */
 847                         if (blk_mq_hw_queue_mapped(hctx))
 848                                 blk_mq_tag_idle(hctx);
 849                 }
 850         }
 851         blk_queue_exit(q);
 852 }
 853
 854 struct flush_busy_ctx_data {
 855         struct blk_mq_hw_ctx *hctx;
 856         struct list_head *list;
 857 };
 858
 859 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 860 {
 861         struct flush_busy_ctx_data *flush_data = data;
 862         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 863         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 864
 865         sbitmap_clear_bit(sb, bitnr);
 866         spin_lock(&ctx->lock);
 867         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 868         spin_unlock(&ctx->lock);
 869         return true;
 870 }
 871
 872 /*
 873  * Process software queues that have been marked busy, splicing them
 874  * to the for-dispatch
 875  */
 876 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 877 {
 878         struct flush_busy_ctx_data data = {
 879                 .hctx = hctx,
 880                 .list = list,
 881         };
 882
 883         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 884 }
 885 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 886
 887 static inline unsigned int queued_to_index(unsigned int queued)
 888 {
 889         if (!queued)
 890                 return 0;
 891
 892         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 893 }
 894
 895 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 896                            bool wait)
 897 {
 898         struct blk_mq_alloc_data data = {
 899                 .q = rq->q,
 900                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 901                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 902         };
 903
 904         might_sleep_if(wait);
 905
 906         if (rq->tag != -1)
 907                 goto done;
 908
 909         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 910                 data.flags |= BLK_MQ_REQ_RESERVED;
 911
 912         rq->tag = blk_mq_get_tag(&data);
 913         if (rq->tag >= 0) {
 914                 if (blk_mq_tag_busy(data.hctx)) {
 915                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 916                         atomic_inc(&data.hctx->nr_active);
 917                 }
 918                 data.hctx->tags->rqs[rq->tag] = rq;
 919         }
 920
 921 done:
 922         if (hctx)
 923                 *hctx = data.hctx;
 924         return rq->tag != -1;
 925 }
 926
 927 static void __blk_mq_put_driver_tag(struct blk_mq_hw_ctx *hctx,
 928                                     struct request *rq)
 929 {
 930         blk_mq_put_tag(hctx, hctx->tags, rq->mq_ctx, rq->tag);
 931         rq->tag = -1;
 932
 933         if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 934                 rq->rq_flags &= ~RQF_MQ_INFLIGHT;
 935                 atomic_dec(&hctx->nr_active);
 936         }
 937 }
 938
 939 static void blk_mq_put_driver_tag_hctx(struct blk_mq_hw_ctx *hctx,
 940                                        struct request *rq)
 941 {
 942         if (rq->tag == -1 || rq->internal_tag == -1)
 943                 return;
 944
 945         __blk_mq_put_driver_tag(hctx, rq);
 946 }
 947
 948 static void blk_mq_put_driver_tag(struct request *rq)
 949 {
 950         struct blk_mq_hw_ctx *hctx;
 951
 952         if (rq->tag == -1 || rq->internal_tag == -1)
 953                 return;
 954
 955         hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
 956         __blk_mq_put_driver_tag(hctx, rq);
 957 }
 958
 959 /*
 960  * If we fail getting a driver tag because all the driver tags are already
 961  * assigned and on the dispatch list, BUT the first entry does not have a
 962  * tag, then we could deadlock. For that case, move entries with assigned
 963  * driver tags to the front, leaving the set of tagged requests in the
 964  * same order, and the untagged set in the same order.
 965  */
 966 static bool reorder_tags_to_front(struct list_head *list)
 967 {
 968         struct request *rq, *tmp, *first = NULL;
 969
 970         list_for_each_entry_safe_reverse(rq, tmp, list, queuelist) {
 971                 if (rq == first)
 972                         break;
 973                 if (rq->tag != -1) {
 974                         list_move(&rq->queuelist, list);
 975                         if (!first)
 976                                 first = rq;
 977                 }
 978         }
 979
 980         return first != NULL;
 981 }
 982
 983 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
 984                                 void *key)
 985 {
 986         struct blk_mq_hw_ctx *hctx;
 987
 988         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
 989
 990         list_del(&wait->entry);
 991         clear_bit_unlock(BLK_MQ_S_TAG_WAITING, &hctx->state);
 992         blk_mq_run_hw_queue(hctx, true);
 993         return 1;
 994 }
 995
 996 static bool blk_mq_dispatch_wait_add(struct blk_mq_hw_ctx *hctx)
 997 {
 998         struct sbq_wait_state *ws;
 999
1000         /*
1001          * The TAG_WAITING bit serves as a lock protecting hctx->dispatch_wait.
1002          * The thread which wins the race to grab this bit adds the hardware
1003          * queue to the wait queue.
1004          */
1005         if (test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state) ||
1006             test_and_set_bit_lock(BLK_MQ_S_TAG_WAITING, &hctx->state))
1007                 return false;
1008
1009         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
1010         ws = bt_wait_ptr(&hctx->tags->bitmap_tags, hctx);
1011
1012         /*
1013          * As soon as this returns, it's no longer safe to fiddle with
1014          * hctx->dispatch_wait, since a completion can wake up the wait queue
1015          * and unlock the bit.
1016          */
1017         add_wait_queue(&ws->wait, &hctx->dispatch_wait);
1018         return true;
1019 }
1020
1021 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list)
1022 {
1023         struct blk_mq_hw_ctx *hctx;
1024         struct request *rq;
1025         int errors, queued;
1026
1027         if (list_empty(list))
1028                 return false;
1029
1030         /*
1031          * Now process all the entries, sending them to the driver.
1032          */
1033         errors = queued = 0;
1034         do {
1035                 struct blk_mq_queue_data bd;
1036                 blk_status_t ret;
1037
1038                 rq = list_first_entry(list, struct request, queuelist);
1039                 if (!blk_mq_get_driver_tag(rq, &hctx, false)) {
1040                         if (!queued && reorder_tags_to_front(list))
1041                                 continue;
1042
1043                         /*
1044                          * The initial allocation attempt failed, so we need to
1045                          * rerun the hardware queue when a tag is freed.
1046                          */
1047                         if (!blk_mq_dispatch_wait_add(hctx))
1048                                 break;
1049
1050                         /*
1051                          * It's possible that a tag was freed in the window
1052                          * between the allocation failure and adding the
1053                          * hardware queue to the wait queue.
1054                          */
1055                         if (!blk_mq_get_driver_tag(rq, &hctx, false))
1056                                 break;
1057                 }
1058
1059                 list_del_init(&rq->queuelist);
1060
1061                 bd.rq = rq;
1062
1063                 /*
1064                  * Flag last if we have no more requests, or if we have more
1065                  * but can't assign a driver tag to it.
1066                  */
1067                 if (list_empty(list))
1068                         bd.last = true;
1069                 else {
1070                         struct request *nxt;
1071
1072                         nxt = list_first_entry(list, struct request, queuelist);
1073                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1074                 }
1075
1076                 ret = q->mq_ops->queue_rq(hctx, &bd);
1077                 if (ret == BLK_STS_RESOURCE) {
1078                         blk_mq_put_driver_tag_hctx(hctx, rq);
1079                         list_add(&rq->queuelist, list);
1080                         __blk_mq_requeue_request(rq);
1081                         break;
1082                 }
1083
1084                 if (unlikely(ret != BLK_STS_OK)) {
1085                         errors++;
1086                         blk_mq_end_request(rq, BLK_STS_IOERR);
1087                         continue;
1088                 }
1089
1090                 queued++;
1091         } while (!list_empty(list));
1092
1093         hctx->dispatched[queued_to_index(queued)]++;
1094
1095         /*
1096          * Any items that need requeuing? Stuff them into hctx->dispatch,
1097          * that is where we will continue on next queue run.
1098          */
1099         if (!list_empty(list)) {
1100                 /*
1101                  * If an I/O scheduler has been configured and we got a driver
1102                  * tag for the next request already, free it again.
1103                  */
1104                 rq = list_first_entry(list, struct request, queuelist);
1105                 blk_mq_put_driver_tag(rq);
1106
1107                 spin_lock(&hctx->lock);
1108                 list_splice_init(list, &hctx->dispatch);
1109                 spin_unlock(&hctx->lock);
1110
1111                 /*
1112                  * If SCHED_RESTART was set by the caller of this function and
1113                  * it is no longer set that means that it was cleared by another
1114                  * thread and hence that a queue rerun is needed.
1115                  *
1116                  * If TAG_WAITING is set that means that an I/O scheduler has
1117                  * been configured and another thread is waiting for a driver
1118                  * tag. To guarantee fairness, do not rerun this hardware queue
1119                  * but let the other thread grab the driver tag.
1120                  *
1121                  * If no I/O scheduler has been configured it is possible that
1122                  * the hardware queue got stopped and restarted before requests
1123                  * were pushed back onto the dispatch list. Rerun the queue to
1124                  * avoid starvation. Notes:
1125                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1126                  *   been stopped before rerunning a queue.
1127                  * - Some but not all block drivers stop a queue before
1128                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1129                  *   and dm-rq.
1130                  */
1131                 if (!blk_mq_sched_needs_restart(hctx) &&
1132                     !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state))
1133                         blk_mq_run_hw_queue(hctx, true);
1134         }
1135
1136         return (queued + errors) != 0;
1137 }
1138
1139 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1140 {
1141         int srcu_idx;
1142
1143         /*
1144          * We should be running this queue from one of the CPUs that
1145          * are mapped to it.
1146          */
1147         WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1148                 cpu_online(hctx->next_cpu));
1149
1150         /*
1151          * We can't run the queue inline with ints disabled. Ensure that
1152          * we catch bad users of this early.
1153          */
1154         WARN_ON_ONCE(in_interrupt());
1155
1156         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1157                 rcu_read_lock();
1158                 blk_mq_sched_dispatch_requests(hctx);
1159                 rcu_read_unlock();
1160         } else {
1161                 might_sleep();
1162
1163                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1164                 blk_mq_sched_dispatch_requests(hctx);
1165                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1166         }
1167 }
1168
1169 /*
1170  * It'd be great if the workqueue API had a way to pass
1171  * in a mask and had some smarts for more clever placement.
1172  * For now we just round-robin here, switching for every
1173  * BLK_MQ_CPU_WORK_BATCH queued items.
1174  */
1175 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1176 {
1177         if (hctx->queue->nr_hw_queues == 1)
1178                 return WORK_CPU_UNBOUND;
1179
1180         if (--hctx->next_cpu_batch <= 0) {
1181                 int next_cpu;
1182
1183                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1184                 if (next_cpu >= nr_cpu_ids)
1185                         next_cpu = cpumask_first(hctx->cpumask);
1186
1187                 hctx->next_cpu = next_cpu;
1188                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1189         }
1190
1191         return hctx->next_cpu;
1192 }
1193
1194 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1195                                         unsigned long msecs)
1196 {
1197         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1198                 return;
1199
1200         if (unlikely(blk_mq_hctx_stopped(hctx)))
1201                 return;
1202
1203         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1204                 int cpu = get_cpu();
1205                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1206                         __blk_mq_run_hw_queue(hctx);
1207                         put_cpu();
1208                         return;
1209                 }
1210
1211                 put_cpu();
1212         }
1213
1214         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1215                                          &hctx->run_work,
1216                                          msecs_to_jiffies(msecs));
1217 }
1218
1219 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1220 {
1221         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1222 }
1223 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1224
1225 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1226 {
1227         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1228 }
1229 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1230
1231 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1232 {
1233         struct blk_mq_hw_ctx *hctx;
1234         int i;
1235
1236         queue_for_each_hw_ctx(q, hctx, i) {
1237                 if (!blk_mq_hctx_has_pending(hctx) ||
1238                     blk_mq_hctx_stopped(hctx))
1239                         continue;
1240
1241                 blk_mq_run_hw_queue(hctx, async);
1242         }
1243 }
1244 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1245
1246 /**
1247  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1248  * @q: request queue.
1249  *
1250  * The caller is responsible for serializing this function against
1251  * blk_mq_{start,stop}_hw_queue().
1252  */
1253 bool blk_mq_queue_stopped(struct request_queue *q)
1254 {
1255         struct blk_mq_hw_ctx *hctx;
1256         int i;
1257
1258         queue_for_each_hw_ctx(q, hctx, i)
1259                 if (blk_mq_hctx_stopped(hctx))
1260                         return true;
1261
1262         return false;
1263 }
1264 EXPORT_SYMBOL(blk_mq_queue_stopped);
1265
1266 /*
1267  * This function is often used for pausing .queue_rq() by driver when
1268  * there isn't enough resource or some conditions aren't satisfied, and
1269  * BLK_STS_RESOURCE is usually returned.
1270  *
1271  * We do not guarantee that dispatch can be drained or blocked
1272  * after blk_mq_stop_hw_queue() returns. Please use
1273  * blk_mq_quiesce_queue() for that requirement.
1274  */
1275 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1276 {
1277         cancel_delayed_work(&hctx->run_work);
1278
1279         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1280 }
1281 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1282
1283 /*
1284  * This function is often used for pausing .queue_rq() by driver when
1285  * there isn't enough resource or some conditions aren't satisfied, and
1286  * BLK_STS_RESOURCE is usually returned.
1287  *
1288  * We do not guarantee that dispatch can be drained or blocked
1289  * after blk_mq_stop_hw_queues() returns. Please use
1290  * blk_mq_quiesce_queue() for that requirement.
1291  */
1292 void blk_mq_stop_hw_queues(struct request_queue *q)
1293 {
1294         struct blk_mq_hw_ctx *hctx;
1295         int i;
1296
1297         queue_for_each_hw_ctx(q, hctx, i)
1298                 blk_mq_stop_hw_queue(hctx);
1299 }
1300 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1301
1302 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1303 {
1304         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1305
1306         blk_mq_run_hw_queue(hctx, false);
1307 }
1308 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1309
1310 void blk_mq_start_hw_queues(struct request_queue *q)
1311 {
1312         struct blk_mq_hw_ctx *hctx;
1313         int i;
1314
1315         queue_for_each_hw_ctx(q, hctx, i)
1316                 blk_mq_start_hw_queue(hctx);
1317 }
1318 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1319
1320 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1321 {
1322         if (!blk_mq_hctx_stopped(hctx))
1323                 return;
1324
1325         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1326         blk_mq_run_hw_queue(hctx, async);
1327 }
1328 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1329
1330 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1331 {
1332         struct blk_mq_hw_ctx *hctx;
1333         int i;
1334
1335         queue_for_each_hw_ctx(q, hctx, i)
1336                 blk_mq_start_stopped_hw_queue(hctx, async);
1337 }
1338 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1339
1340 static void blk_mq_run_work_fn(struct work_struct *work)
1341 {
1342         struct blk_mq_hw_ctx *hctx;
1343
1344         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1345
1346         /*
1347          * If we are stopped, don't run the queue. The exception is if
1348          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1349          * the STOPPED bit and run it.
1350          */
1351         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1352                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1353                         return;
1354
1355                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1356                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1357         }
1358
1359         __blk_mq_run_hw_queue(hctx);
1360 }
1361
1362
1363 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1364 {
1365         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1366                 return;
1367
1368         /*
1369          * Stop the hw queue, then modify currently delayed work.
1370          * This should prevent us from running the queue prematurely.
1371          * Mark the queue as auto-clearing STOPPED when it runs.
1372          */
1373         blk_mq_stop_hw_queue(hctx);
1374         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1375         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1376                                         &hctx->run_work,
1377                                         msecs_to_jiffies(msecs));
1378 }
1379 EXPORT_SYMBOL(blk_mq_delay_queue);
1380
1381 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1382                                             struct request *rq,
1383                                             bool at_head)
1384 {
1385         struct blk_mq_ctx *ctx = rq->mq_ctx;
1386
1387         lockdep_assert_held(&ctx->lock);
1388
1389         trace_block_rq_insert(hctx->queue, rq);
1390
1391         if (at_head)
1392                 list_add(&rq->queuelist, &ctx->rq_list);
1393         else
1394                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1395 }
1396
1397 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1398                              bool at_head)
1399 {
1400         struct blk_mq_ctx *ctx = rq->mq_ctx;
1401
1402         lockdep_assert_held(&ctx->lock);
1403
1404         __blk_mq_insert_req_list(hctx, rq, at_head);
1405         blk_mq_hctx_mark_pending(hctx, ctx);
1406 }
1407
1408 /*
1409  * Should only be used carefully, when the caller knows we want to
1410  * bypass a potential IO scheduler on the target device.
1411  */
1412 void blk_mq_request_bypass_insert(struct request *rq)
1413 {
1414         struct blk_mq_ctx *ctx = rq->mq_ctx;
1415         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1416
1417         spin_lock(&hctx->lock);
1418         list_add_tail(&rq->queuelist, &hctx->dispatch);
1419         spin_unlock(&hctx->lock);
1420
1421         blk_mq_run_hw_queue(hctx, false);
1422 }
1423
1424 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1425                             struct list_head *list)
1426
1427 {
1428         /*
1429          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1430          * offline now
1431          */
1432         spin_lock(&ctx->lock);
1433         while (!list_empty(list)) {
1434                 struct request *rq;
1435
1436                 rq = list_first_entry(list, struct request, queuelist);
1437                 BUG_ON(rq->mq_ctx != ctx);
1438                 list_del_init(&rq->queuelist);
1439                 __blk_mq_insert_req_list(hctx, rq, false);
1440         }
1441         blk_mq_hctx_mark_pending(hctx, ctx);
1442         spin_unlock(&ctx->lock);
1443 }
1444
1445 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1446 {
1447         struct request *rqa = container_of(a, struct request, queuelist);
1448         struct request *rqb = container_of(b, struct request, queuelist);
1449
1450         return !(rqa->mq_ctx < rqb->mq_ctx ||
1451                  (rqa->mq_ctx == rqb->mq_ctx &&
1452                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1453 }
1454
1455 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1456 {
1457         struct blk_mq_ctx *this_ctx;
1458         struct request_queue *this_q;
1459         struct request *rq;
1460         LIST_HEAD(list);
1461         LIST_HEAD(ctx_list);
1462         unsigned int depth;
1463
1464         list_splice_init(&plug->mq_list, &list);
1465
1466         list_sort(NULL, &list, plug_ctx_cmp);
1467
1468         this_q = NULL;
1469         this_ctx = NULL;
1470         depth = 0;
1471
1472         while (!list_empty(&list)) {
1473                 rq = list_entry_rq(list.next);
1474                 list_del_init(&rq->queuelist);
1475                 BUG_ON(!rq->q);
1476                 if (rq->mq_ctx != this_ctx) {
1477                         if (this_ctx) {
1478                                 trace_block_unplug(this_q, depth, from_schedule);
1479                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1480                                                                 &ctx_list,
1481                                                                 from_schedule);
1482                         }
1483
1484                         this_ctx = rq->mq_ctx;
1485                         this_q = rq->q;
1486                         depth = 0;
1487                 }
1488
1489                 depth++;
1490                 list_add_tail(&rq->queuelist, &ctx_list);
1491         }
1492
1493         /*
1494          * If 'this_ctx' is set, we know we have entries to complete
1495          * on 'ctx_list'. Do those.
1496          */
1497         if (this_ctx) {
1498                 trace_block_unplug(this_q, depth, from_schedule);
1499                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1500                                                 from_schedule);
1501         }
1502 }
1503
1504 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1505 {
1506         blk_init_request_from_bio(rq, bio);
1507
1508         blk_account_io_start(rq, true);
1509 }
1510
1511 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1512 {
1513         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1514                 !blk_queue_nomerges(hctx->queue);
1515 }
1516
1517 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1518                                    struct blk_mq_ctx *ctx,
1519                                    struct request *rq)
1520 {
1521         spin_lock(&ctx->lock);
1522         __blk_mq_insert_request(hctx, rq, false);
1523         spin_unlock(&ctx->lock);
1524 }
1525
1526 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1527 {
1528         if (rq->tag != -1)
1529                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1530
1531         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1532 }
1533
1534 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1535                                         struct request *rq,
1536                                         blk_qc_t *cookie, bool may_sleep)
1537 {
1538         struct request_queue *q = rq->q;
1539         struct blk_mq_queue_data bd = {
1540                 .rq = rq,
1541                 .last = true,
1542         };
1543         blk_qc_t new_cookie;
1544         blk_status_t ret;
1545         bool run_queue = true;
1546
1547         /* RCU or SRCU read lock is needed before checking quiesced flag */
1548         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1549                 run_queue = false;
1550                 goto insert;
1551         }
1552
1553         if (q->elevator)
1554                 goto insert;
1555
1556         if (!blk_mq_get_driver_tag(rq, NULL, false))
1557                 goto insert;
1558
1559         new_cookie = request_to_qc_t(hctx, rq);
1560
1561         /*
1562          * For OK queue, we are done. For error, kill it. Any other
1563          * error (busy), just add it to our list as we previously
1564          * would have done
1565          */
1566         ret = q->mq_ops->queue_rq(hctx, &bd);
1567         switch (ret) {
1568         case BLK_STS_OK:
1569                 *cookie = new_cookie;
1570                 return;
1571         case BLK_STS_RESOURCE:
1572                 __blk_mq_requeue_request(rq);
1573                 goto insert;
1574         default:
1575                 *cookie = BLK_QC_T_NONE;
1576                 blk_mq_end_request(rq, ret);
1577                 return;
1578         }
1579
1580 insert:
1581         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1582 }
1583
1584 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1585                 struct request *rq, blk_qc_t *cookie)
1586 {
1587         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1588                 rcu_read_lock();
1589                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1590                 rcu_read_unlock();
1591         } else {
1592                 unsigned int srcu_idx;
1593
1594                 might_sleep();
1595
1596                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1597                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1598                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1599         }
1600 }
1601
1602 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1603 {
1604         const int is_sync = op_is_sync(bio->bi_opf);
1605         const int is_flush_fua = op_is_flush(bio->bi_opf);
1606         struct blk_mq_alloc_data data = { .flags = 0 };
1607         struct request *rq;
1608         unsigned int request_count = 0;
1609         struct blk_plug *plug;
1610         struct request *same_queue_rq = NULL;
1611         blk_qc_t cookie;
1612         unsigned int wb_acct;
1613
1614         blk_queue_bounce(q, &bio);
1615
1616         blk_queue_split(q, &bio);
1617
1618         if (!bio_integrity_prep(bio))
1619                 return BLK_QC_T_NONE;
1620
1621         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1622             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1623                 return BLK_QC_T_NONE;
1624
1625         if (blk_mq_sched_bio_merge(q, bio))
1626                 return BLK_QC_T_NONE;
1627
1628         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1629
1630         trace_block_getrq(q, bio, bio->bi_opf);
1631
1632         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1633         if (unlikely(!rq)) {
1634                 __wbt_done(q->rq_wb, wb_acct);
1635                 if (bio->bi_opf & REQ_NOWAIT)
1636                         bio_wouldblock_error(bio);
1637                 return BLK_QC_T_NONE;
1638         }
1639
1640         wbt_track(&rq->issue_stat, wb_acct);
1641
1642         cookie = request_to_qc_t(data.hctx, rq);
1643
1644         plug = current->plug;
1645         if (unlikely(is_flush_fua)) {
1646                 blk_mq_put_ctx(data.ctx);
1647                 blk_mq_bio_to_request(rq, bio);
1648                 if (q->elevator) {
1649                         blk_mq_sched_insert_request(rq, false, true, true,
1650                                         true);
1651                 } else {
1652                         blk_insert_flush(rq);
1653                         blk_mq_run_hw_queue(data.hctx, true);
1654                 }
1655         } else if (plug && q->nr_hw_queues == 1) {
1656                 struct request *last = NULL;
1657
1658                 blk_mq_put_ctx(data.ctx);
1659                 blk_mq_bio_to_request(rq, bio);
1660
1661                 /*
1662                  * @request_count may become stale because of schedule
1663                  * out, so check the list again.
1664                  */
1665                 if (list_empty(&plug->mq_list))
1666                         request_count = 0;
1667                 else if (blk_queue_nomerges(q))
1668                         request_count = blk_plug_queued_count(q);
1669
1670                 if (!request_count)
1671                         trace_block_plug(q);
1672                 else
1673                         last = list_entry_rq(plug->mq_list.prev);
1674
1675                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1676                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1677                         blk_flush_plug_list(plug, false);
1678                         trace_block_plug(q);
1679                 }
1680
1681                 list_add_tail(&rq->queuelist, &plug->mq_list);
1682         } else if (plug && !blk_queue_nomerges(q)) {
1683                 blk_mq_bio_to_request(rq, bio);
1684
1685                 /*
1686                  * We do limited plugging. If the bio can be merged, do that.
1687                  * Otherwise the existing request in the plug list will be
1688                  * issued. So the plug list will have one request at most
1689                  * The plug list might get flushed before this. If that happens,
1690                  * the plug list is empty, and same_queue_rq is invalid.
1691                  */
1692                 if (list_empty(&plug->mq_list))
1693                         same_queue_rq = NULL;
1694                 if (same_queue_rq)
1695                         list_del_init(&same_queue_rq->queuelist);
1696                 list_add_tail(&rq->queuelist, &plug->mq_list);
1697
1698                 blk_mq_put_ctx(data.ctx);
1699
1700                 if (same_queue_rq) {
1701                         data.hctx = blk_mq_map_queue(q,
1702                                         same_queue_rq->mq_ctx->cpu);
1703                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1704                                         &cookie);
1705                 }
1706         } else if (q->nr_hw_queues > 1 && is_sync) {
1707                 blk_mq_put_ctx(data.ctx);
1708                 blk_mq_bio_to_request(rq, bio);
1709                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1710         } else if (q->elevator) {
1711                 blk_mq_put_ctx(data.ctx);
1712                 blk_mq_bio_to_request(rq, bio);
1713                 blk_mq_sched_insert_request(rq, false, true, true, true);
1714         } else {
1715                 blk_mq_put_ctx(data.ctx);
1716                 blk_mq_bio_to_request(rq, bio);
1717                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1718                 blk_mq_run_hw_queue(data.hctx, true);
1719         }
1720
1721         return cookie;
1722 }
1723
1724 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1725                      unsigned int hctx_idx)
1726 {
1727         struct page *page;
1728
1729         if (tags->rqs && set->ops->exit_request) {
1730                 int i;
1731
1732                 for (i = 0; i < tags->nr_tags; i++) {
1733                         struct request *rq = tags->static_rqs[i];
1734
1735                         if (!rq)
1736                                 continue;
1737                         set->ops->exit_request(set, rq, hctx_idx);
1738                         tags->static_rqs[i] = NULL;
1739                 }
1740         }
1741
1742         while (!list_empty(&tags->page_list)) {
1743                 page = list_first_entry(&tags->page_list, struct page, lru);
1744                 list_del_init(&page->lru);
1745                 /*
1746                  * Remove kmemleak object previously allocated in
1747                  * blk_mq_init_rq_map().
1748                  */
1749                 kmemleak_free(page_address(page));
1750                 __free_pages(page, page->private);
1751         }
1752 }
1753
1754 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1755 {
1756         kfree(tags->rqs);
1757         tags->rqs = NULL;
1758         kfree(tags->static_rqs);
1759         tags->static_rqs = NULL;
1760
1761         blk_mq_free_tags(tags);
1762 }
1763
1764 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1765                                         unsigned int hctx_idx,
1766                                         unsigned int nr_tags,
1767                                         unsigned int reserved_tags)
1768 {
1769         struct blk_mq_tags *tags;
1770         int node;
1771
1772         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1773         if (node == NUMA_NO_NODE)
1774                 node = set->numa_node;
1775
1776         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1777                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1778         if (!tags)
1779                 return NULL;
1780
1781         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1782                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1783                                  node);
1784         if (!tags->rqs) {
1785                 blk_mq_free_tags(tags);
1786                 return NULL;
1787         }
1788
1789         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1790                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1791                                  node);
1792         if (!tags->static_rqs) {
1793                 kfree(tags->rqs);
1794                 blk_mq_free_tags(tags);
1795                 return NULL;
1796         }
1797
1798         return tags;
1799 }
1800
1801 static size_t order_to_size(unsigned int order)
1802 {
1803         return (size_t)PAGE_SIZE << order;
1804 }
1805
1806 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1807                      unsigned int hctx_idx, unsigned int depth)
1808 {
1809         unsigned int i, j, entries_per_page, max_order = 4;
1810         size_t rq_size, left;
1811         int node;
1812
1813         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1814         if (node == NUMA_NO_NODE)
1815                 node = set->numa_node;
1816
1817         INIT_LIST_HEAD(&tags->page_list);
1818
1819         /*
1820          * rq_size is the size of the request plus driver payload, rounded
1821          * to the cacheline size
1822          */
1823         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1824                                 cache_line_size());
1825         left = rq_size * depth;
1826
1827         for (i = 0; i < depth; ) {
1828                 int this_order = max_order;
1829                 struct page *page;
1830                 int to_do;
1831                 void *p;
1832
1833                 while (this_order && left < order_to_size(this_order - 1))
1834                         this_order--;
1835
1836                 do {
1837                         page = alloc_pages_node(node,
1838                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1839                                 this_order);
1840                         if (page)
1841                                 break;
1842                         if (!this_order--)
1843                                 break;
1844                         if (order_to_size(this_order) < rq_size)
1845                                 break;
1846                 } while (1);
1847
1848                 if (!page)
1849                         goto fail;
1850
1851                 page->private = this_order;
1852                 list_add_tail(&page->lru, &tags->page_list);
1853
1854                 p = page_address(page);
1855                 /*
1856                  * Allow kmemleak to scan these pages as they contain pointers
1857                  * to additional allocations like via ops->init_request().
1858                  */
1859                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1860                 entries_per_page = order_to_size(this_order) / rq_size;
1861                 to_do = min(entries_per_page, depth - i);
1862                 left -= to_do * rq_size;
1863                 for (j = 0; j < to_do; j++) {
1864                         struct request *rq = p;
1865
1866                         tags->static_rqs[i] = rq;
1867                         if (set->ops->init_request) {
1868                                 if (set->ops->init_request(set, rq, hctx_idx,
1869                                                 node)) {
1870                                         tags->static_rqs[i] = NULL;
1871                                         goto fail;
1872                                 }
1873                         }
1874
1875                         p += rq_size;
1876                         i++;
1877                 }
1878         }
1879         return 0;
1880
1881 fail:
1882         blk_mq_free_rqs(set, tags, hctx_idx);
1883         return -ENOMEM;
1884 }
1885
1886 /*
1887  * 'cpu' is going away. splice any existing rq_list entries from this
1888  * software queue to the hw queue dispatch list, and ensure that it
1889  * gets run.
1890  */
1891 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1892 {
1893         struct blk_mq_hw_ctx *hctx;
1894         struct blk_mq_ctx *ctx;
1895         LIST_HEAD(tmp);
1896
1897         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1898         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1899
1900         spin_lock(&ctx->lock);
1901         if (!list_empty(&ctx->rq_list)) {
1902                 list_splice_init(&ctx->rq_list, &tmp);
1903                 blk_mq_hctx_clear_pending(hctx, ctx);
1904         }
1905         spin_unlock(&ctx->lock);
1906
1907         if (list_empty(&tmp))
1908                 return 0;
1909
1910         spin_lock(&hctx->lock);
1911         list_splice_tail_init(&tmp, &hctx->dispatch);
1912         spin_unlock(&hctx->lock);
1913
1914         blk_mq_run_hw_queue(hctx, true);
1915         return 0;
1916 }
1917
1918 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1919 {
1920         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1921                                             &hctx->cpuhp_dead);
1922 }
1923
1924 /* hctx->ctxs will be freed in queue's release handler */
1925 static void blk_mq_exit_hctx(struct request_queue *q,
1926                 struct blk_mq_tag_set *set,
1927                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1928 {
1929         blk_mq_debugfs_unregister_hctx(hctx);
1930
1931         if (blk_mq_hw_queue_mapped(hctx))
1932                 blk_mq_tag_idle(hctx);
1933
1934         if (set->ops->exit_request)
1935                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1936
1937         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1938
1939         if (set->ops->exit_hctx)
1940                 set->ops->exit_hctx(hctx, hctx_idx);
1941
1942         if (hctx->flags & BLK_MQ_F_BLOCKING)
1943                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1944
1945         blk_mq_remove_cpuhp(hctx);
1946         blk_free_flush_queue(hctx->fq);
1947         sbitmap_free(&hctx->ctx_map);
1948 }
1949
1950 static void blk_mq_exit_hw_queues(struct request_queue *q,
1951                 struct blk_mq_tag_set *set, int nr_queue)
1952 {
1953         struct blk_mq_hw_ctx *hctx;
1954         unsigned int i;
1955
1956         queue_for_each_hw_ctx(q, hctx, i) {
1957                 if (i == nr_queue)
1958                         break;
1959                 blk_mq_exit_hctx(q, set, hctx, i);
1960         }
1961 }
1962
1963 static int blk_mq_init_hctx(struct request_queue *q,
1964                 struct blk_mq_tag_set *set,
1965                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1966 {
1967         int node;
1968
1969         node = hctx->numa_node;
1970         if (node == NUMA_NO_NODE)
1971                 node = hctx->numa_node = set->numa_node;
1972
1973         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1974         spin_lock_init(&hctx->lock);
1975         INIT_LIST_HEAD(&hctx->dispatch);
1976         hctx->queue = q;
1977         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1978
1979         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1980
1981         hctx->tags = set->tags[hctx_idx];
1982
1983         /*
1984          * Allocate space for all possible cpus to avoid allocation at
1985          * runtime
1986          */
1987         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1988                                         GFP_KERNEL, node);
1989         if (!hctx->ctxs)
1990                 goto unregister_cpu_notifier;
1991
1992         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
1993                               node))
1994                 goto free_ctxs;
1995
1996         hctx->nr_ctx = 0;
1997
1998         if (set->ops->init_hctx &&
1999             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2000                 goto free_bitmap;
2001
2002         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2003                 goto exit_hctx;
2004
2005         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2006         if (!hctx->fq)
2007                 goto sched_exit_hctx;
2008
2009         if (set->ops->init_request &&
2010             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
2011                                    node))
2012                 goto free_fq;
2013
2014         if (hctx->flags & BLK_MQ_F_BLOCKING)
2015                 init_srcu_struct(hctx->queue_rq_srcu);
2016
2017         blk_mq_debugfs_register_hctx(q, hctx);
2018
2019         return 0;
2020
2021  free_fq:
2022         kfree(hctx->fq);
2023  sched_exit_hctx:
2024         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2025  exit_hctx:
2026         if (set->ops->exit_hctx)
2027                 set->ops->exit_hctx(hctx, hctx_idx);
2028  free_bitmap:
2029         sbitmap_free(&hctx->ctx_map);
2030  free_ctxs:
2031         kfree(hctx->ctxs);
2032  unregister_cpu_notifier:
2033         blk_mq_remove_cpuhp(hctx);
2034         return -1;
2035 }
2036
2037 static void blk_mq_init_cpu_queues(struct request_queue *q,
2038                                    unsigned int nr_hw_queues)
2039 {
2040         unsigned int i;
2041
2042         for_each_possible_cpu(i) {
2043                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2044                 struct blk_mq_hw_ctx *hctx;
2045
2046                 __ctx->cpu = i;
2047                 spin_lock_init(&__ctx->lock);
2048                 INIT_LIST_HEAD(&__ctx->rq_list);
2049                 __ctx->queue = q;
2050
2051                 /* If the cpu isn't present, the cpu is mapped to first hctx */
2052                 if (!cpu_present(i))
2053                         continue;
2054
2055                 hctx = blk_mq_map_queue(q, i);
2056
2057                 /*
2058                  * Set local node, IFF we have more than one hw queue. If
2059                  * not, we remain on the home node of the device
2060                  */
2061                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2062                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2063         }
2064 }
2065
2066 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2067 {
2068         int ret = 0;
2069
2070         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2071                                         set->queue_depth, set->reserved_tags);
2072         if (!set->tags[hctx_idx])
2073                 return false;
2074
2075         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2076                                 set->queue_depth);
2077         if (!ret)
2078                 return true;
2079
2080         blk_mq_free_rq_map(set->tags[hctx_idx]);
2081         set->tags[hctx_idx] = NULL;
2082         return false;
2083 }
2084
2085 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2086                                          unsigned int hctx_idx)
2087 {
2088         if (set->tags[hctx_idx]) {
2089                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2090                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2091                 set->tags[hctx_idx] = NULL;
2092         }
2093 }
2094
2095 static void blk_mq_map_swqueue(struct request_queue *q)
2096 {
2097         unsigned int i, hctx_idx;
2098         struct blk_mq_hw_ctx *hctx;
2099         struct blk_mq_ctx *ctx;
2100         struct blk_mq_tag_set *set = q->tag_set;
2101
2102         /*
2103          * Avoid others reading imcomplete hctx->cpumask through sysfs
2104          */
2105         mutex_lock(&q->sysfs_lock);
2106
2107         queue_for_each_hw_ctx(q, hctx, i) {
2108                 cpumask_clear(hctx->cpumask);
2109                 hctx->nr_ctx = 0;
2110         }
2111
2112         /*
2113          * Map software to hardware queues.
2114          *
2115          * If the cpu isn't present, the cpu is mapped to first hctx.
2116          */
2117         for_each_present_cpu(i) {
2118                 hctx_idx = q->mq_map[i];
2119                 /* unmapped hw queue can be remapped after CPU topo changed */
2120                 if (!set->tags[hctx_idx] &&
2121                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2122                         /*
2123                          * If tags initialization fail for some hctx,
2124                          * that hctx won't be brought online.  In this
2125                          * case, remap the current ctx to hctx[0] which
2126                          * is guaranteed to always have tags allocated
2127                          */
2128                         q->mq_map[i] = 0;
2129                 }
2130
2131                 ctx = per_cpu_ptr(q->queue_ctx, i);
2132                 hctx = blk_mq_map_queue(q, i);
2133
2134                 cpumask_set_cpu(i, hctx->cpumask);
2135                 ctx->index_hw = hctx->nr_ctx;
2136                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2137         }
2138
2139         mutex_unlock(&q->sysfs_lock);
2140
2141         queue_for_each_hw_ctx(q, hctx, i) {
2142                 /*
2143                  * If no software queues are mapped to this hardware queue,
2144                  * disable it and free the request entries.
2145                  */
2146                 if (!hctx->nr_ctx) {
2147                         /* Never unmap queue 0.  We need it as a
2148                          * fallback in case of a new remap fails
2149                          * allocation
2150                          */
2151                         if (i && set->tags[i])
2152                                 blk_mq_free_map_and_requests(set, i);
2153
2154                         hctx->tags = NULL;
2155                         continue;
2156                 }
2157
2158                 hctx->tags = set->tags[i];
2159                 WARN_ON(!hctx->tags);
2160
2161                 /*
2162                  * Set the map size to the number of mapped software queues.
2163                  * This is more accurate and more efficient than looping
2164                  * over all possibly mapped software queues.
2165                  */
2166                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2167
2168                 /*
2169                  * Initialize batch roundrobin counts
2170                  */
2171                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2172                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2173         }
2174 }
2175
2176 /*
2177  * Caller needs to ensure that we're either frozen/quiesced, or that
2178  * the queue isn't live yet.
2179  */
2180 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2181 {
2182         struct blk_mq_hw_ctx *hctx;
2183         int i;
2184
2185         queue_for_each_hw_ctx(q, hctx, i) {
2186                 if (shared) {
2187                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2188                                 atomic_inc(&q->shared_hctx_restart);
2189                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2190                 } else {
2191                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2192                                 atomic_dec(&q->shared_hctx_restart);
2193                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2194                 }
2195         }
2196 }
2197
2198 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2199                                         bool shared)
2200 {
2201         struct request_queue *q;
2202
2203         lockdep_assert_held(&set->tag_list_lock);
2204
2205         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2206                 blk_mq_freeze_queue(q);
2207                 queue_set_hctx_shared(q, shared);
2208                 blk_mq_unfreeze_queue(q);
2209         }
2210 }
2211
2212 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2213 {
2214         struct blk_mq_tag_set *set = q->tag_set;
2215
2216         mutex_lock(&set->tag_list_lock);
2217         list_del_rcu(&q->tag_set_list);
2218         INIT_LIST_HEAD(&q->tag_set_list);
2219         if (list_is_singular(&set->tag_list)) {
2220                 /* just transitioned to unshared */
2221                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2222                 /* update existing queue */
2223                 blk_mq_update_tag_set_depth(set, false);
2224         }
2225         mutex_unlock(&set->tag_list_lock);
2226
2227         synchronize_rcu();
2228 }
2229
2230 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2231                                      struct request_queue *q)
2232 {
2233         q->tag_set = set;
2234
2235         mutex_lock(&set->tag_list_lock);
2236
2237         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2238         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2239                 set->flags |= BLK_MQ_F_TAG_SHARED;
2240                 /* update existing queue */
2241                 blk_mq_update_tag_set_depth(set, true);
2242         }
2243         if (set->flags & BLK_MQ_F_TAG_SHARED)
2244                 queue_set_hctx_shared(q, true);
2245         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2246
2247         mutex_unlock(&set->tag_list_lock);
2248 }
2249
2250 /*
2251  * It is the actual release handler for mq, but we do it from
2252  * request queue's release handler for avoiding use-after-free
2253  * and headache because q->mq_kobj shouldn't have been introduced,
2254  * but we can't group ctx/kctx kobj without it.
2255  */
2256 void blk_mq_release(struct request_queue *q)
2257 {
2258         struct blk_mq_hw_ctx *hctx;
2259         unsigned int i;
2260
2261         /* hctx kobj stays in hctx */
2262         queue_for_each_hw_ctx(q, hctx, i) {
2263                 if (!hctx)
2264                         continue;
2265                 kobject_put(&hctx->kobj);
2266         }
2267
2268         q->mq_map = NULL;
2269
2270         kfree(q->queue_hw_ctx);
2271
2272         /*
2273          * release .mq_kobj and sw queue's kobject now because
2274          * both share lifetime with request queue.
2275          */
2276         blk_mq_sysfs_deinit(q);
2277
2278         free_percpu(q->queue_ctx);
2279 }
2280
2281 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2282 {
2283         struct request_queue *uninit_q, *q;
2284
2285         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2286         if (!uninit_q)
2287                 return ERR_PTR(-ENOMEM);
2288
2289         q = blk_mq_init_allocated_queue(set, uninit_q);
2290         if (IS_ERR(q))
2291                 blk_cleanup_queue(uninit_q);
2292
2293         return q;
2294 }
2295 EXPORT_SYMBOL(blk_mq_init_queue);
2296
2297 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2298 {
2299         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2300
2301         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2302                            __alignof__(struct blk_mq_hw_ctx)) !=
2303                      sizeof(struct blk_mq_hw_ctx));
2304
2305         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2306                 hw_ctx_size += sizeof(struct srcu_struct);
2307
2308         return hw_ctx_size;
2309 }
2310
2311 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2312                                                 struct request_queue *q)
2313 {
2314         int i, j;
2315         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2316
2317         blk_mq_sysfs_unregister(q);
2318
2319         /* protect against switching io scheduler  */
2320         mutex_lock(&q->sysfs_lock);
2321         for (i = 0; i < set->nr_hw_queues; i++) {
2322                 int node;
2323
2324                 if (hctxs[i])
2325                         continue;
2326
2327                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2328                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2329                                         GFP_KERNEL, node);
2330                 if (!hctxs[i])
2331                         break;
2332
2333                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2334                                                 node)) {
2335                         kfree(hctxs[i]);
2336                         hctxs[i] = NULL;
2337                         break;
2338                 }
2339
2340                 atomic_set(&hctxs[i]->nr_active, 0);
2341                 hctxs[i]->numa_node = node;
2342                 hctxs[i]->queue_num = i;
2343
2344                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2345                         free_cpumask_var(hctxs[i]->cpumask);
2346                         kfree(hctxs[i]);
2347                         hctxs[i] = NULL;
2348                         break;
2349                 }
2350                 blk_mq_hctx_kobj_init(hctxs[i]);
2351         }
2352         for (j = i; j < q->nr_hw_queues; j++) {
2353                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2354
2355                 if (hctx) {
2356                         if (hctx->tags)
2357                                 blk_mq_free_map_and_requests(set, j);
2358                         blk_mq_exit_hctx(q, set, hctx, j);
2359                         kobject_put(&hctx->kobj);
2360                         hctxs[j] = NULL;
2361
2362                 }
2363         }
2364         q->nr_hw_queues = i;
2365         mutex_unlock(&q->sysfs_lock);
2366         blk_mq_sysfs_register(q);
2367 }
2368
2369 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2370                                                   struct request_queue *q)
2371 {
2372         /* mark the queue as mq asap */
2373         q->mq_ops = set->ops;
2374
2375         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2376                                              blk_mq_poll_stats_bkt,
2377                                              BLK_MQ_POLL_STATS_BKTS, q);
2378         if (!q->poll_cb)
2379                 goto err_exit;
2380
2381         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2382         if (!q->queue_ctx)
2383                 goto err_exit;
2384
2385         /* init q->mq_kobj and sw queues' kobjects */
2386         blk_mq_sysfs_init(q);
2387
2388         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2389                                                 GFP_KERNEL, set->numa_node);
2390         if (!q->queue_hw_ctx)
2391                 goto err_percpu;
2392
2393         q->mq_map = set->mq_map;
2394
2395         blk_mq_realloc_hw_ctxs(set, q);
2396         if (!q->nr_hw_queues)
2397                 goto err_hctxs;
2398
2399         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2400         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2401
2402         q->nr_queues = nr_cpu_ids;
2403
2404         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2405
2406         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2407                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2408
2409         q->sg_reserved_size = INT_MAX;
2410
2411         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2412         INIT_LIST_HEAD(&q->requeue_list);
2413         spin_lock_init(&q->requeue_lock);
2414
2415         blk_queue_make_request(q, blk_mq_make_request);
2416
2417         /*
2418          * Do this after blk_queue_make_request() overrides it...
2419          */
2420         q->nr_requests = set->queue_depth;
2421
2422         /*
2423          * Default to classic polling
2424          */
2425         q->poll_nsec = -1;
2426
2427         if (set->ops->complete)
2428                 blk_queue_softirq_done(q, set->ops->complete);
2429
2430         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2431         blk_mq_add_queue_tag_set(set, q);
2432         blk_mq_map_swqueue(q);
2433
2434         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2435                 int ret;
2436
2437                 ret = blk_mq_sched_init(q);
2438                 if (ret)
2439                         return ERR_PTR(ret);
2440         }
2441
2442         return q;
2443
2444 err_hctxs:
2445         kfree(q->queue_hw_ctx);
2446 err_percpu:
2447         free_percpu(q->queue_ctx);
2448 err_exit:
2449         q->mq_ops = NULL;
2450         return ERR_PTR(-ENOMEM);
2451 }
2452 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2453
2454 void blk_mq_free_queue(struct request_queue *q)
2455 {
2456         struct blk_mq_tag_set   *set = q->tag_set;
2457
2458         blk_mq_del_queue_tag_set(q);
2459         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2460 }
2461
2462 /* Basically redo blk_mq_init_queue with queue frozen */
2463 static void blk_mq_queue_reinit(struct request_queue *q)
2464 {
2465         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2466
2467         blk_mq_debugfs_unregister_hctxs(q);
2468         blk_mq_sysfs_unregister(q);
2469
2470         /*
2471          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2472          * we should change hctx numa_node according to new topology (this
2473          * involves free and re-allocate memory, worthy doing?)
2474          */
2475
2476         blk_mq_map_swqueue(q);
2477
2478         blk_mq_sysfs_register(q);
2479         blk_mq_debugfs_register_hctxs(q);
2480 }
2481
2482 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2483 {
2484         int i;
2485
2486         for (i = 0; i < set->nr_hw_queues; i++)
2487                 if (!__blk_mq_alloc_rq_map(set, i))
2488                         goto out_unwind;
2489
2490         return 0;
2491
2492 out_unwind:
2493         while (--i >= 0)
2494                 blk_mq_free_rq_map(set->tags[i]);
2495
2496         return -ENOMEM;
2497 }
2498
2499 /*
2500  * Allocate the request maps associated with this tag_set. Note that this
2501  * may reduce the depth asked for, if memory is tight. set->queue_depth
2502  * will be updated to reflect the allocated depth.
2503  */
2504 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2505 {
2506         unsigned int depth;
2507         int err;
2508
2509         depth = set->queue_depth;
2510         do {
2511                 err = __blk_mq_alloc_rq_maps(set);
2512                 if (!err)
2513                         break;
2514
2515                 set->queue_depth >>= 1;
2516                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2517                         err = -ENOMEM;
2518                         break;
2519                 }
2520         } while (set->queue_depth);
2521
2522         if (!set->queue_depth || err) {
2523                 pr_err("blk-mq: failed to allocate request map\n");
2524                 return -ENOMEM;
2525         }
2526
2527         if (depth != set->queue_depth)
2528                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2529                                                 depth, set->queue_depth);
2530
2531         return 0;
2532 }
2533
2534 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2535 {
2536         if (set->ops->map_queues) {
2537                 int cpu;
2538                 /*
2539                  * transport .map_queues is usually done in the following
2540                  * way:
2541                  *
2542                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2543                  *      mask = get_cpu_mask(queue)
2544                  *      for_each_cpu(cpu, mask)
2545                  *              set->mq_map[cpu] = queue;
2546                  * }
2547                  *
2548                  * When we need to remap, the table has to be cleared for
2549                  * killing stale mapping since one CPU may not be mapped
2550                  * to any hw queue.
2551                  */
2552                 for_each_possible_cpu(cpu)
2553                         set->mq_map[cpu] = 0;
2554
2555                 return set->ops->map_queues(set);
2556         } else
2557                 return blk_mq_map_queues(set);
2558 }
2559
2560 /*
2561  * Alloc a tag set to be associated with one or more request queues.
2562  * May fail with EINVAL for various error conditions. May adjust the
2563  * requested depth down, if if it too large. In that case, the set
2564  * value will be stored in set->queue_depth.
2565  */
2566 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2567 {
2568         int ret;
2569
2570         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2571
2572         if (!set->nr_hw_queues)
2573                 return -EINVAL;
2574         if (!set->queue_depth)
2575                 return -EINVAL;
2576         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2577                 return -EINVAL;
2578
2579         if (!set->ops->queue_rq)
2580                 return -EINVAL;
2581
2582         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2583                 pr_info("blk-mq: reduced tag depth to %u\n",
2584                         BLK_MQ_MAX_DEPTH);
2585                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2586         }
2587
2588         /*
2589          * If a crashdump is active, then we are potentially in a very
2590          * memory constrained environment. Limit us to 1 queue and
2591          * 64 tags to prevent using too much memory.
2592          */
2593         if (is_kdump_kernel()) {
2594                 set->nr_hw_queues = 1;
2595                 set->queue_depth = min(64U, set->queue_depth);
2596         }
2597         /*
2598          * There is no use for more h/w queues than cpus.
2599          */
2600         if (set->nr_hw_queues > nr_cpu_ids)
2601                 set->nr_hw_queues = nr_cpu_ids;
2602
2603         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2604                                  GFP_KERNEL, set->numa_node);
2605         if (!set->tags)
2606                 return -ENOMEM;
2607
2608         ret = -ENOMEM;
2609         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2610                         GFP_KERNEL, set->numa_node);
2611         if (!set->mq_map)
2612                 goto out_free_tags;
2613
2614         ret = blk_mq_update_queue_map(set);
2615         if (ret)
2616                 goto out_free_mq_map;
2617
2618         ret = blk_mq_alloc_rq_maps(set);
2619         if (ret)
2620                 goto out_free_mq_map;
2621
2622         mutex_init(&set->tag_list_lock);
2623         INIT_LIST_HEAD(&set->tag_list);
2624
2625         return 0;
2626
2627 out_free_mq_map:
2628         kfree(set->mq_map);
2629         set->mq_map = NULL;
2630 out_free_tags:
2631         kfree(set->tags);
2632         set->tags = NULL;
2633         return ret;
2634 }
2635 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2636
2637 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2638 {
2639         int i;
2640
2641         for (i = 0; i < nr_cpu_ids; i++)
2642                 blk_mq_free_map_and_requests(set, i);
2643
2644         kfree(set->mq_map);
2645         set->mq_map = NULL;
2646
2647         kfree(set->tags);
2648         set->tags = NULL;
2649 }
2650 EXPORT_SYMBOL(blk_mq_free_tag_set);
2651
2652 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2653 {
2654         struct blk_mq_tag_set *set = q->tag_set;
2655         struct blk_mq_hw_ctx *hctx;
2656         int i, ret;
2657
2658         if (!set)
2659                 return -EINVAL;
2660
2661         blk_mq_freeze_queue(q);
2662
2663         ret = 0;
2664         queue_for_each_hw_ctx(q, hctx, i) {
2665                 if (!hctx->tags)
2666                         continue;
2667                 /*
2668                  * If we're using an MQ scheduler, just update the scheduler
2669                  * queue depth. This is similar to what the old code would do.
2670                  */
2671                 if (!hctx->sched_tags) {
2672                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2673                                                         min(nr, set->queue_depth),
2674                                                         false);
2675                 } else {
2676                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2677                                                         nr, true);
2678                 }
2679                 if (ret)
2680                         break;
2681         }
2682
2683         if (!ret)
2684                 q->nr_requests = nr;
2685
2686         blk_mq_unfreeze_queue(q);
2687
2688         return ret;
2689 }
2690
2691 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2692                                                         int nr_hw_queues)
2693 {
2694         struct request_queue *q;
2695
2696         lockdep_assert_held(&set->tag_list_lock);
2697
2698         if (nr_hw_queues > nr_cpu_ids)
2699                 nr_hw_queues = nr_cpu_ids;
2700         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2701                 return;
2702
2703         list_for_each_entry(q, &set->tag_list, tag_set_list)
2704                 blk_mq_freeze_queue(q);
2705
2706         set->nr_hw_queues = nr_hw_queues;
2707         blk_mq_update_queue_map(set);
2708         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2709                 blk_mq_realloc_hw_ctxs(set, q);
2710                 blk_mq_queue_reinit(q);
2711         }
2712
2713         list_for_each_entry(q, &set->tag_list, tag_set_list)
2714                 blk_mq_unfreeze_queue(q);
2715 }
2716
2717 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2718 {
2719         mutex_lock(&set->tag_list_lock);
2720         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2721         mutex_unlock(&set->tag_list_lock);
2722 }
2723 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2724
2725 /* Enable polling stats and return whether they were already enabled. */
2726 static bool blk_poll_stats_enable(struct request_queue *q)
2727 {
2728         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2729             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2730                 return true;
2731         blk_stat_add_callback(q, q->poll_cb);
2732         return false;
2733 }
2734
2735 static void blk_mq_poll_stats_start(struct request_queue *q)
2736 {
2737         /*
2738          * We don't arm the callback if polling stats are not enabled or the
2739          * callback is already active.
2740          */
2741         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2742             blk_stat_is_active(q->poll_cb))
2743                 return;
2744
2745         blk_stat_activate_msecs(q->poll_cb, 100);
2746 }
2747
2748 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2749 {
2750         struct request_queue *q = cb->data;
2751         int bucket;
2752
2753         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2754                 if (cb->stat[bucket].nr_samples)
2755                         q->poll_stat[bucket] = cb->stat[bucket];
2756         }
2757 }
2758
2759 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2760                                        struct blk_mq_hw_ctx *hctx,
2761                                        struct request *rq)
2762 {
2763         unsigned long ret = 0;
2764         int bucket;
2765
2766         /*
2767          * If stats collection isn't on, don't sleep but turn it on for
2768          * future users
2769          */
2770         if (!blk_poll_stats_enable(q))
2771                 return 0;
2772
2773         /*
2774          * As an optimistic guess, use half of the mean service time
2775          * for this type of request. We can (and should) make this smarter.
2776          * For instance, if the completion latencies are tight, we can
2777          * get closer than just half the mean. This is especially
2778          * important on devices where the completion latencies are longer
2779          * than ~10 usec. We do use the stats for the relevant IO size
2780          * if available which does lead to better estimates.
2781          */
2782         bucket = blk_mq_poll_stats_bkt(rq);
2783         if (bucket < 0)
2784                 return ret;
2785
2786         if (q->poll_stat[bucket].nr_samples)
2787                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2788
2789         return ret;
2790 }
2791
2792 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2793                                      struct blk_mq_hw_ctx *hctx,
2794                                      struct request *rq)
2795 {
2796         struct hrtimer_sleeper hs;
2797         enum hrtimer_mode mode;
2798         unsigned int nsecs;
2799         ktime_t kt;
2800
2801         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2802                 return false;
2803
2804         /*
2805          * poll_nsec can be:
2806          *
2807          * -1:  don't ever hybrid sleep
2808          *  0:  use half of prev avg
2809          * >0:  use this specific value
2810          */
2811         if (q->poll_nsec == -1)
2812                 return false;
2813         else if (q->poll_nsec > 0)
2814                 nsecs = q->poll_nsec;
2815         else
2816                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2817
2818         if (!nsecs)
2819                 return false;
2820
2821         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2822
2823         /*
2824          * This will be replaced with the stats tracking code, using
2825          * 'avg_completion_time / 2' as the pre-sleep target.
2826          */
2827         kt = nsecs;
2828
2829         mode = HRTIMER_MODE_REL;
2830         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2831         hrtimer_set_expires(&hs.timer, kt);
2832
2833         hrtimer_init_sleeper(&hs, current);
2834         do {
2835                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2836                         break;
2837                 set_current_state(TASK_UNINTERRUPTIBLE);
2838                 hrtimer_start_expires(&hs.timer, mode);
2839                 if (hs.task)
2840                         io_schedule();
2841                 hrtimer_cancel(&hs.timer);
2842                 mode = HRTIMER_MODE_ABS;
2843         } while (hs.task && !signal_pending(current));
2844
2845         __set_current_state(TASK_RUNNING);
2846         destroy_hrtimer_on_stack(&hs.timer);
2847         return true;
2848 }
2849
2850 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2851 {
2852         struct request_queue *q = hctx->queue;
2853         long state;
2854
2855         /*
2856          * If we sleep, have the caller restart the poll loop to reset
2857          * the state. Like for the other success return cases, the
2858          * caller is responsible for checking if the IO completed. If
2859          * the IO isn't complete, we'll get called again and will go
2860          * straight to the busy poll loop.
2861          */
2862         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2863                 return true;
2864
2865         hctx->poll_considered++;
2866
2867         state = current->state;
2868         while (!need_resched()) {
2869                 int ret;
2870
2871                 hctx->poll_invoked++;
2872
2873                 ret = q->mq_ops->poll(hctx, rq->tag);
2874                 if (ret > 0) {
2875                         hctx->poll_success++;
2876                         set_current_state(TASK_RUNNING);
2877                         return true;
2878                 }
2879
2880                 if (signal_pending_state(state, current))
2881                         set_current_state(TASK_RUNNING);
2882
2883                 if (current->state == TASK_RUNNING)
2884                         return true;
2885                 if (ret < 0)
2886                         break;
2887                 cpu_relax();
2888         }
2889
2890         return false;
2891 }
2892
2893 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2894 {
2895         struct blk_mq_hw_ctx *hctx;
2896         struct blk_plug *plug;
2897         struct request *rq;
2898
2899         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2900             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2901                 return false;
2902
2903         plug = current->plug;
2904         if (plug)
2905                 blk_flush_plug_list(plug, false);
2906
2907         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2908         if (!blk_qc_t_is_internal(cookie))
2909                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2910         else {
2911                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2912                 /*
2913                  * With scheduling, if the request has completed, we'll
2914                  * get a NULL return here, as we clear the sched tag when
2915                  * that happens. The request still remains valid, like always,
2916                  * so we should be safe with just the NULL check.
2917                  */
2918                 if (!rq)
2919                         return false;
2920         }
2921
2922         return __blk_mq_poll(hctx, rq);
2923 }
2924 EXPORT_SYMBOL_GPL(blk_mq_poll);
2925
2926 static int __init blk_mq_init(void)
2927 {
2928         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2929                                 blk_mq_hctx_notify_dead);
2930         return 0;
2931 }
2932 subsys_initcall(blk_mq_init);