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          * There are at least two related races now between setting
1148          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1149          * __blk_mq_run_hw_queue():
1150          *
1151          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1152          *   but later it becomes online, then this warning is harmless
1153          *   at all
1154          *
1155          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1156          *   but later it becomes offline, then the warning can't be
1157          *   triggered, and we depend on blk-mq timeout handler to
1158          *   handle dispatched requests to this hctx
1159          */
1160         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1161                 cpu_online(hctx->next_cpu)) {
1162                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1163                         raw_smp_processor_id(),
1164                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1165                 dump_stack();
1166         }
1167
1168         /*
1169          * We can't run the queue inline with ints disabled. Ensure that
1170          * we catch bad users of this early.
1171          */
1172         WARN_ON_ONCE(in_interrupt());
1173
1174         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1175                 rcu_read_lock();
1176                 blk_mq_sched_dispatch_requests(hctx);
1177                 rcu_read_unlock();
1178         } else {
1179                 might_sleep();
1180
1181                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1182                 blk_mq_sched_dispatch_requests(hctx);
1183                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1184         }
1185 }
1186
1187 /*
1188  * It'd be great if the workqueue API had a way to pass
1189  * in a mask and had some smarts for more clever placement.
1190  * For now we just round-robin here, switching for every
1191  * BLK_MQ_CPU_WORK_BATCH queued items.
1192  */
1193 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1194 {
1195         if (hctx->queue->nr_hw_queues == 1)
1196                 return WORK_CPU_UNBOUND;
1197
1198         if (--hctx->next_cpu_batch <= 0) {
1199                 int next_cpu;
1200
1201                 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
1202                 if (next_cpu >= nr_cpu_ids)
1203                         next_cpu = cpumask_first(hctx->cpumask);
1204
1205                 hctx->next_cpu = next_cpu;
1206                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1207         }
1208
1209         return hctx->next_cpu;
1210 }
1211
1212 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1213                                         unsigned long msecs)
1214 {
1215         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1216                 return;
1217
1218         if (unlikely(blk_mq_hctx_stopped(hctx)))
1219                 return;
1220
1221         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1222                 int cpu = get_cpu();
1223                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1224                         __blk_mq_run_hw_queue(hctx);
1225                         put_cpu();
1226                         return;
1227                 }
1228
1229                 put_cpu();
1230         }
1231
1232         kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1233                                          &hctx->run_work,
1234                                          msecs_to_jiffies(msecs));
1235 }
1236
1237 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1238 {
1239         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1240 }
1241 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1242
1243 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1244 {
1245         __blk_mq_delay_run_hw_queue(hctx, async, 0);
1246 }
1247 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1248
1249 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1250 {
1251         struct blk_mq_hw_ctx *hctx;
1252         int i;
1253
1254         queue_for_each_hw_ctx(q, hctx, i) {
1255                 if (!blk_mq_hctx_has_pending(hctx) ||
1256                     blk_mq_hctx_stopped(hctx))
1257                         continue;
1258
1259                 blk_mq_run_hw_queue(hctx, async);
1260         }
1261 }
1262 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1263
1264 /**
1265  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1266  * @q: request queue.
1267  *
1268  * The caller is responsible for serializing this function against
1269  * blk_mq_{start,stop}_hw_queue().
1270  */
1271 bool blk_mq_queue_stopped(struct request_queue *q)
1272 {
1273         struct blk_mq_hw_ctx *hctx;
1274         int i;
1275
1276         queue_for_each_hw_ctx(q, hctx, i)
1277                 if (blk_mq_hctx_stopped(hctx))
1278                         return true;
1279
1280         return false;
1281 }
1282 EXPORT_SYMBOL(blk_mq_queue_stopped);
1283
1284 /*
1285  * This function is often used for pausing .queue_rq() by driver when
1286  * there isn't enough resource or some conditions aren't satisfied, and
1287  * BLK_STS_RESOURCE is usually returned.
1288  *
1289  * We do not guarantee that dispatch can be drained or blocked
1290  * after blk_mq_stop_hw_queue() returns. Please use
1291  * blk_mq_quiesce_queue() for that requirement.
1292  */
1293 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1294 {
1295         cancel_delayed_work(&hctx->run_work);
1296
1297         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1298 }
1299 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1300
1301 /*
1302  * This function is often used for pausing .queue_rq() by driver when
1303  * there isn't enough resource or some conditions aren't satisfied, and
1304  * BLK_STS_RESOURCE is usually returned.
1305  *
1306  * We do not guarantee that dispatch can be drained or blocked
1307  * after blk_mq_stop_hw_queues() returns. Please use
1308  * blk_mq_quiesce_queue() for that requirement.
1309  */
1310 void blk_mq_stop_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_stop_hw_queue(hctx);
1317 }
1318 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1319
1320 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1321 {
1322         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1323
1324         blk_mq_run_hw_queue(hctx, false);
1325 }
1326 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1327
1328 void blk_mq_start_hw_queues(struct request_queue *q)
1329 {
1330         struct blk_mq_hw_ctx *hctx;
1331         int i;
1332
1333         queue_for_each_hw_ctx(q, hctx, i)
1334                 blk_mq_start_hw_queue(hctx);
1335 }
1336 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1337
1338 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1339 {
1340         if (!blk_mq_hctx_stopped(hctx))
1341                 return;
1342
1343         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1344         blk_mq_run_hw_queue(hctx, async);
1345 }
1346 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1347
1348 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1349 {
1350         struct blk_mq_hw_ctx *hctx;
1351         int i;
1352
1353         queue_for_each_hw_ctx(q, hctx, i)
1354                 blk_mq_start_stopped_hw_queue(hctx, async);
1355 }
1356 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1357
1358 static void blk_mq_run_work_fn(struct work_struct *work)
1359 {
1360         struct blk_mq_hw_ctx *hctx;
1361
1362         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1363
1364         /*
1365          * If we are stopped, don't run the queue. The exception is if
1366          * BLK_MQ_S_START_ON_RUN is set. For that case, we auto-clear
1367          * the STOPPED bit and run it.
1368          */
1369         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state)) {
1370                 if (!test_bit(BLK_MQ_S_START_ON_RUN, &hctx->state))
1371                         return;
1372
1373                 clear_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1374                 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1375         }
1376
1377         __blk_mq_run_hw_queue(hctx);
1378 }
1379
1380
1381 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1382 {
1383         if (WARN_ON_ONCE(!blk_mq_hw_queue_mapped(hctx)))
1384                 return;
1385
1386         /*
1387          * Stop the hw queue, then modify currently delayed work.
1388          * This should prevent us from running the queue prematurely.
1389          * Mark the queue as auto-clearing STOPPED when it runs.
1390          */
1391         blk_mq_stop_hw_queue(hctx);
1392         set_bit(BLK_MQ_S_START_ON_RUN, &hctx->state);
1393         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
1394                                         &hctx->run_work,
1395                                         msecs_to_jiffies(msecs));
1396 }
1397 EXPORT_SYMBOL(blk_mq_delay_queue);
1398
1399 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1400                                             struct request *rq,
1401                                             bool at_head)
1402 {
1403         struct blk_mq_ctx *ctx = rq->mq_ctx;
1404
1405         lockdep_assert_held(&ctx->lock);
1406
1407         trace_block_rq_insert(hctx->queue, rq);
1408
1409         if (at_head)
1410                 list_add(&rq->queuelist, &ctx->rq_list);
1411         else
1412                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1413 }
1414
1415 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1416                              bool at_head)
1417 {
1418         struct blk_mq_ctx *ctx = rq->mq_ctx;
1419
1420         lockdep_assert_held(&ctx->lock);
1421
1422         __blk_mq_insert_req_list(hctx, rq, at_head);
1423         blk_mq_hctx_mark_pending(hctx, ctx);
1424 }
1425
1426 /*
1427  * Should only be used carefully, when the caller knows we want to
1428  * bypass a potential IO scheduler on the target device.
1429  */
1430 void blk_mq_request_bypass_insert(struct request *rq)
1431 {
1432         struct blk_mq_ctx *ctx = rq->mq_ctx;
1433         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1434
1435         spin_lock(&hctx->lock);
1436         list_add_tail(&rq->queuelist, &hctx->dispatch);
1437         spin_unlock(&hctx->lock);
1438
1439         blk_mq_run_hw_queue(hctx, false);
1440 }
1441
1442 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1443                             struct list_head *list)
1444
1445 {
1446         /*
1447          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1448          * offline now
1449          */
1450         spin_lock(&ctx->lock);
1451         while (!list_empty(list)) {
1452                 struct request *rq;
1453
1454                 rq = list_first_entry(list, struct request, queuelist);
1455                 BUG_ON(rq->mq_ctx != ctx);
1456                 list_del_init(&rq->queuelist);
1457                 __blk_mq_insert_req_list(hctx, rq, false);
1458         }
1459         blk_mq_hctx_mark_pending(hctx, ctx);
1460         spin_unlock(&ctx->lock);
1461 }
1462
1463 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1464 {
1465         struct request *rqa = container_of(a, struct request, queuelist);
1466         struct request *rqb = container_of(b, struct request, queuelist);
1467
1468         return !(rqa->mq_ctx < rqb->mq_ctx ||
1469                  (rqa->mq_ctx == rqb->mq_ctx &&
1470                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1471 }
1472
1473 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1474 {
1475         struct blk_mq_ctx *this_ctx;
1476         struct request_queue *this_q;
1477         struct request *rq;
1478         LIST_HEAD(list);
1479         LIST_HEAD(ctx_list);
1480         unsigned int depth;
1481
1482         list_splice_init(&plug->mq_list, &list);
1483
1484         list_sort(NULL, &list, plug_ctx_cmp);
1485
1486         this_q = NULL;
1487         this_ctx = NULL;
1488         depth = 0;
1489
1490         while (!list_empty(&list)) {
1491                 rq = list_entry_rq(list.next);
1492                 list_del_init(&rq->queuelist);
1493                 BUG_ON(!rq->q);
1494                 if (rq->mq_ctx != this_ctx) {
1495                         if (this_ctx) {
1496                                 trace_block_unplug(this_q, depth, from_schedule);
1497                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1498                                                                 &ctx_list,
1499                                                                 from_schedule);
1500                         }
1501
1502                         this_ctx = rq->mq_ctx;
1503                         this_q = rq->q;
1504                         depth = 0;
1505                 }
1506
1507                 depth++;
1508                 list_add_tail(&rq->queuelist, &ctx_list);
1509         }
1510
1511         /*
1512          * If 'this_ctx' is set, we know we have entries to complete
1513          * on 'ctx_list'. Do those.
1514          */
1515         if (this_ctx) {
1516                 trace_block_unplug(this_q, depth, from_schedule);
1517                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1518                                                 from_schedule);
1519         }
1520 }
1521
1522 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1523 {
1524         blk_init_request_from_bio(rq, bio);
1525
1526         blk_account_io_start(rq, true);
1527 }
1528
1529 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1530 {
1531         return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1532                 !blk_queue_nomerges(hctx->queue);
1533 }
1534
1535 static inline void blk_mq_queue_io(struct blk_mq_hw_ctx *hctx,
1536                                    struct blk_mq_ctx *ctx,
1537                                    struct request *rq)
1538 {
1539         spin_lock(&ctx->lock);
1540         __blk_mq_insert_request(hctx, rq, false);
1541         spin_unlock(&ctx->lock);
1542 }
1543
1544 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1545 {
1546         if (rq->tag != -1)
1547                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1548
1549         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1550 }
1551
1552 static void __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1553                                         struct request *rq,
1554                                         blk_qc_t *cookie, bool may_sleep)
1555 {
1556         struct request_queue *q = rq->q;
1557         struct blk_mq_queue_data bd = {
1558                 .rq = rq,
1559                 .last = true,
1560         };
1561         blk_qc_t new_cookie;
1562         blk_status_t ret;
1563         bool run_queue = true;
1564
1565         /* RCU or SRCU read lock is needed before checking quiesced flag */
1566         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1567                 run_queue = false;
1568                 goto insert;
1569         }
1570
1571         if (q->elevator)
1572                 goto insert;
1573
1574         if (!blk_mq_get_driver_tag(rq, NULL, false))
1575                 goto insert;
1576
1577         new_cookie = request_to_qc_t(hctx, rq);
1578
1579         /*
1580          * For OK queue, we are done. For error, kill it. Any other
1581          * error (busy), just add it to our list as we previously
1582          * would have done
1583          */
1584         ret = q->mq_ops->queue_rq(hctx, &bd);
1585         switch (ret) {
1586         case BLK_STS_OK:
1587                 *cookie = new_cookie;
1588                 return;
1589         case BLK_STS_RESOURCE:
1590                 __blk_mq_requeue_request(rq);
1591                 goto insert;
1592         default:
1593                 *cookie = BLK_QC_T_NONE;
1594                 blk_mq_end_request(rq, ret);
1595                 return;
1596         }
1597
1598 insert:
1599         blk_mq_sched_insert_request(rq, false, run_queue, false, may_sleep);
1600 }
1601
1602 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1603                 struct request *rq, blk_qc_t *cookie)
1604 {
1605         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
1606                 rcu_read_lock();
1607                 __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1608                 rcu_read_unlock();
1609         } else {
1610                 unsigned int srcu_idx;
1611
1612                 might_sleep();
1613
1614                 srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
1615                 __blk_mq_try_issue_directly(hctx, rq, cookie, true);
1616                 srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
1617         }
1618 }
1619
1620 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1621 {
1622         const int is_sync = op_is_sync(bio->bi_opf);
1623         const int is_flush_fua = op_is_flush(bio->bi_opf);
1624         struct blk_mq_alloc_data data = { .flags = 0 };
1625         struct request *rq;
1626         unsigned int request_count = 0;
1627         struct blk_plug *plug;
1628         struct request *same_queue_rq = NULL;
1629         blk_qc_t cookie;
1630         unsigned int wb_acct;
1631
1632         blk_queue_bounce(q, &bio);
1633
1634         blk_queue_split(q, &bio);
1635
1636         if (!bio_integrity_prep(bio))
1637                 return BLK_QC_T_NONE;
1638
1639         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1640             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1641                 return BLK_QC_T_NONE;
1642
1643         if (blk_mq_sched_bio_merge(q, bio))
1644                 return BLK_QC_T_NONE;
1645
1646         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1647
1648         trace_block_getrq(q, bio, bio->bi_opf);
1649
1650         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1651         if (unlikely(!rq)) {
1652                 __wbt_done(q->rq_wb, wb_acct);
1653                 if (bio->bi_opf & REQ_NOWAIT)
1654                         bio_wouldblock_error(bio);
1655                 return BLK_QC_T_NONE;
1656         }
1657
1658         wbt_track(&rq->issue_stat, wb_acct);
1659
1660         cookie = request_to_qc_t(data.hctx, rq);
1661
1662         plug = current->plug;
1663         if (unlikely(is_flush_fua)) {
1664                 blk_mq_put_ctx(data.ctx);
1665                 blk_mq_bio_to_request(rq, bio);
1666                 if (q->elevator) {
1667                         blk_mq_sched_insert_request(rq, false, true, true,
1668                                         true);
1669                 } else {
1670                         blk_insert_flush(rq);
1671                         blk_mq_run_hw_queue(data.hctx, true);
1672                 }
1673         } else if (plug && q->nr_hw_queues == 1) {
1674                 struct request *last = NULL;
1675
1676                 blk_mq_put_ctx(data.ctx);
1677                 blk_mq_bio_to_request(rq, bio);
1678
1679                 /*
1680                  * @request_count may become stale because of schedule
1681                  * out, so check the list again.
1682                  */
1683                 if (list_empty(&plug->mq_list))
1684                         request_count = 0;
1685                 else if (blk_queue_nomerges(q))
1686                         request_count = blk_plug_queued_count(q);
1687
1688                 if (!request_count)
1689                         trace_block_plug(q);
1690                 else
1691                         last = list_entry_rq(plug->mq_list.prev);
1692
1693                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1694                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1695                         blk_flush_plug_list(plug, false);
1696                         trace_block_plug(q);
1697                 }
1698
1699                 list_add_tail(&rq->queuelist, &plug->mq_list);
1700         } else if (plug && !blk_queue_nomerges(q)) {
1701                 blk_mq_bio_to_request(rq, bio);
1702
1703                 /*
1704                  * We do limited plugging. If the bio can be merged, do that.
1705                  * Otherwise the existing request in the plug list will be
1706                  * issued. So the plug list will have one request at most
1707                  * The plug list might get flushed before this. If that happens,
1708                  * the plug list is empty, and same_queue_rq is invalid.
1709                  */
1710                 if (list_empty(&plug->mq_list))
1711                         same_queue_rq = NULL;
1712                 if (same_queue_rq)
1713                         list_del_init(&same_queue_rq->queuelist);
1714                 list_add_tail(&rq->queuelist, &plug->mq_list);
1715
1716                 blk_mq_put_ctx(data.ctx);
1717
1718                 if (same_queue_rq) {
1719                         data.hctx = blk_mq_map_queue(q,
1720                                         same_queue_rq->mq_ctx->cpu);
1721                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1722                                         &cookie);
1723                 }
1724         } else if (q->nr_hw_queues > 1 && is_sync) {
1725                 blk_mq_put_ctx(data.ctx);
1726                 blk_mq_bio_to_request(rq, bio);
1727                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1728         } else if (q->elevator) {
1729                 blk_mq_put_ctx(data.ctx);
1730                 blk_mq_bio_to_request(rq, bio);
1731                 blk_mq_sched_insert_request(rq, false, true, true, true);
1732         } else {
1733                 blk_mq_put_ctx(data.ctx);
1734                 blk_mq_bio_to_request(rq, bio);
1735                 blk_mq_queue_io(data.hctx, data.ctx, rq);
1736                 blk_mq_run_hw_queue(data.hctx, true);
1737         }
1738
1739         return cookie;
1740 }
1741
1742 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1743                      unsigned int hctx_idx)
1744 {
1745         struct page *page;
1746
1747         if (tags->rqs && set->ops->exit_request) {
1748                 int i;
1749
1750                 for (i = 0; i < tags->nr_tags; i++) {
1751                         struct request *rq = tags->static_rqs[i];
1752
1753                         if (!rq)
1754                                 continue;
1755                         set->ops->exit_request(set, rq, hctx_idx);
1756                         tags->static_rqs[i] = NULL;
1757                 }
1758         }
1759
1760         while (!list_empty(&tags->page_list)) {
1761                 page = list_first_entry(&tags->page_list, struct page, lru);
1762                 list_del_init(&page->lru);
1763                 /*
1764                  * Remove kmemleak object previously allocated in
1765                  * blk_mq_init_rq_map().
1766                  */
1767                 kmemleak_free(page_address(page));
1768                 __free_pages(page, page->private);
1769         }
1770 }
1771
1772 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1773 {
1774         kfree(tags->rqs);
1775         tags->rqs = NULL;
1776         kfree(tags->static_rqs);
1777         tags->static_rqs = NULL;
1778
1779         blk_mq_free_tags(tags);
1780 }
1781
1782 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1783                                         unsigned int hctx_idx,
1784                                         unsigned int nr_tags,
1785                                         unsigned int reserved_tags)
1786 {
1787         struct blk_mq_tags *tags;
1788         int node;
1789
1790         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1791         if (node == NUMA_NO_NODE)
1792                 node = set->numa_node;
1793
1794         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1795                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1796         if (!tags)
1797                 return NULL;
1798
1799         tags->rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1800                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1801                                  node);
1802         if (!tags->rqs) {
1803                 blk_mq_free_tags(tags);
1804                 return NULL;
1805         }
1806
1807         tags->static_rqs = kzalloc_node(nr_tags * sizeof(struct request *),
1808                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1809                                  node);
1810         if (!tags->static_rqs) {
1811                 kfree(tags->rqs);
1812                 blk_mq_free_tags(tags);
1813                 return NULL;
1814         }
1815
1816         return tags;
1817 }
1818
1819 static size_t order_to_size(unsigned int order)
1820 {
1821         return (size_t)PAGE_SIZE << order;
1822 }
1823
1824 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1825                      unsigned int hctx_idx, unsigned int depth)
1826 {
1827         unsigned int i, j, entries_per_page, max_order = 4;
1828         size_t rq_size, left;
1829         int node;
1830
1831         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1832         if (node == NUMA_NO_NODE)
1833                 node = set->numa_node;
1834
1835         INIT_LIST_HEAD(&tags->page_list);
1836
1837         /*
1838          * rq_size is the size of the request plus driver payload, rounded
1839          * to the cacheline size
1840          */
1841         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1842                                 cache_line_size());
1843         left = rq_size * depth;
1844
1845         for (i = 0; i < depth; ) {
1846                 int this_order = max_order;
1847                 struct page *page;
1848                 int to_do;
1849                 void *p;
1850
1851                 while (this_order && left < order_to_size(this_order - 1))
1852                         this_order--;
1853
1854                 do {
1855                         page = alloc_pages_node(node,
1856                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1857                                 this_order);
1858                         if (page)
1859                                 break;
1860                         if (!this_order--)
1861                                 break;
1862                         if (order_to_size(this_order) < rq_size)
1863                                 break;
1864                 } while (1);
1865
1866                 if (!page)
1867                         goto fail;
1868
1869                 page->private = this_order;
1870                 list_add_tail(&page->lru, &tags->page_list);
1871
1872                 p = page_address(page);
1873                 /*
1874                  * Allow kmemleak to scan these pages as they contain pointers
1875                  * to additional allocations like via ops->init_request().
1876                  */
1877                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
1878                 entries_per_page = order_to_size(this_order) / rq_size;
1879                 to_do = min(entries_per_page, depth - i);
1880                 left -= to_do * rq_size;
1881                 for (j = 0; j < to_do; j++) {
1882                         struct request *rq = p;
1883
1884                         tags->static_rqs[i] = rq;
1885                         if (set->ops->init_request) {
1886                                 if (set->ops->init_request(set, rq, hctx_idx,
1887                                                 node)) {
1888                                         tags->static_rqs[i] = NULL;
1889                                         goto fail;
1890                                 }
1891                         }
1892
1893                         p += rq_size;
1894                         i++;
1895                 }
1896         }
1897         return 0;
1898
1899 fail:
1900         blk_mq_free_rqs(set, tags, hctx_idx);
1901         return -ENOMEM;
1902 }
1903
1904 /*
1905  * 'cpu' is going away. splice any existing rq_list entries from this
1906  * software queue to the hw queue dispatch list, and ensure that it
1907  * gets run.
1908  */
1909 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
1910 {
1911         struct blk_mq_hw_ctx *hctx;
1912         struct blk_mq_ctx *ctx;
1913         LIST_HEAD(tmp);
1914
1915         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
1916         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
1917
1918         spin_lock(&ctx->lock);
1919         if (!list_empty(&ctx->rq_list)) {
1920                 list_splice_init(&ctx->rq_list, &tmp);
1921                 blk_mq_hctx_clear_pending(hctx, ctx);
1922         }
1923         spin_unlock(&ctx->lock);
1924
1925         if (list_empty(&tmp))
1926                 return 0;
1927
1928         spin_lock(&hctx->lock);
1929         list_splice_tail_init(&tmp, &hctx->dispatch);
1930         spin_unlock(&hctx->lock);
1931
1932         blk_mq_run_hw_queue(hctx, true);
1933         return 0;
1934 }
1935
1936 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
1937 {
1938         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
1939                                             &hctx->cpuhp_dead);
1940 }
1941
1942 /* hctx->ctxs will be freed in queue's release handler */
1943 static void blk_mq_exit_hctx(struct request_queue *q,
1944                 struct blk_mq_tag_set *set,
1945                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
1946 {
1947         blk_mq_debugfs_unregister_hctx(hctx);
1948
1949         if (blk_mq_hw_queue_mapped(hctx))
1950                 blk_mq_tag_idle(hctx);
1951
1952         if (set->ops->exit_request)
1953                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
1954
1955         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
1956
1957         if (set->ops->exit_hctx)
1958                 set->ops->exit_hctx(hctx, hctx_idx);
1959
1960         if (hctx->flags & BLK_MQ_F_BLOCKING)
1961                 cleanup_srcu_struct(hctx->queue_rq_srcu);
1962
1963         blk_mq_remove_cpuhp(hctx);
1964         blk_free_flush_queue(hctx->fq);
1965         sbitmap_free(&hctx->ctx_map);
1966 }
1967
1968 static void blk_mq_exit_hw_queues(struct request_queue *q,
1969                 struct blk_mq_tag_set *set, int nr_queue)
1970 {
1971         struct blk_mq_hw_ctx *hctx;
1972         unsigned int i;
1973
1974         queue_for_each_hw_ctx(q, hctx, i) {
1975                 if (i == nr_queue)
1976                         break;
1977                 blk_mq_exit_hctx(q, set, hctx, i);
1978         }
1979 }
1980
1981 static int blk_mq_init_hctx(struct request_queue *q,
1982                 struct blk_mq_tag_set *set,
1983                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
1984 {
1985         int node;
1986
1987         node = hctx->numa_node;
1988         if (node == NUMA_NO_NODE)
1989                 node = hctx->numa_node = set->numa_node;
1990
1991         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1992         spin_lock_init(&hctx->lock);
1993         INIT_LIST_HEAD(&hctx->dispatch);
1994         hctx->queue = q;
1995         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
1996
1997         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
1998
1999         hctx->tags = set->tags[hctx_idx];
2000
2001         /*
2002          * Allocate space for all possible cpus to avoid allocation at
2003          * runtime
2004          */
2005         hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
2006                                         GFP_KERNEL, node);
2007         if (!hctx->ctxs)
2008                 goto unregister_cpu_notifier;
2009
2010         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2011                               node))
2012                 goto free_ctxs;
2013
2014         hctx->nr_ctx = 0;
2015
2016         if (set->ops->init_hctx &&
2017             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2018                 goto free_bitmap;
2019
2020         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2021                 goto exit_hctx;
2022
2023         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2024         if (!hctx->fq)
2025                 goto sched_exit_hctx;
2026
2027         if (set->ops->init_request &&
2028             set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
2029                                    node))
2030                 goto free_fq;
2031
2032         if (hctx->flags & BLK_MQ_F_BLOCKING)
2033                 init_srcu_struct(hctx->queue_rq_srcu);
2034
2035         blk_mq_debugfs_register_hctx(q, hctx);
2036
2037         return 0;
2038
2039  free_fq:
2040         kfree(hctx->fq);
2041  sched_exit_hctx:
2042         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2043  exit_hctx:
2044         if (set->ops->exit_hctx)
2045                 set->ops->exit_hctx(hctx, hctx_idx);
2046  free_bitmap:
2047         sbitmap_free(&hctx->ctx_map);
2048  free_ctxs:
2049         kfree(hctx->ctxs);
2050  unregister_cpu_notifier:
2051         blk_mq_remove_cpuhp(hctx);
2052         return -1;
2053 }
2054
2055 static void blk_mq_init_cpu_queues(struct request_queue *q,
2056                                    unsigned int nr_hw_queues)
2057 {
2058         unsigned int i;
2059
2060         for_each_possible_cpu(i) {
2061                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2062                 struct blk_mq_hw_ctx *hctx;
2063
2064                 __ctx->cpu = i;
2065                 spin_lock_init(&__ctx->lock);
2066                 INIT_LIST_HEAD(&__ctx->rq_list);
2067                 __ctx->queue = q;
2068
2069                 /* If the cpu isn't present, the cpu is mapped to first hctx */
2070                 if (!cpu_present(i))
2071                         continue;
2072
2073                 hctx = blk_mq_map_queue(q, i);
2074
2075                 /*
2076                  * Set local node, IFF we have more than one hw queue. If
2077                  * not, we remain on the home node of the device
2078                  */
2079                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2080                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2081         }
2082 }
2083
2084 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2085 {
2086         int ret = 0;
2087
2088         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2089                                         set->queue_depth, set->reserved_tags);
2090         if (!set->tags[hctx_idx])
2091                 return false;
2092
2093         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2094                                 set->queue_depth);
2095         if (!ret)
2096                 return true;
2097
2098         blk_mq_free_rq_map(set->tags[hctx_idx]);
2099         set->tags[hctx_idx] = NULL;
2100         return false;
2101 }
2102
2103 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2104                                          unsigned int hctx_idx)
2105 {
2106         if (set->tags[hctx_idx]) {
2107                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2108                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2109                 set->tags[hctx_idx] = NULL;
2110         }
2111 }
2112
2113 static void blk_mq_map_swqueue(struct request_queue *q)
2114 {
2115         unsigned int i, hctx_idx;
2116         struct blk_mq_hw_ctx *hctx;
2117         struct blk_mq_ctx *ctx;
2118         struct blk_mq_tag_set *set = q->tag_set;
2119
2120         /*
2121          * Avoid others reading imcomplete hctx->cpumask through sysfs
2122          */
2123         mutex_lock(&q->sysfs_lock);
2124
2125         queue_for_each_hw_ctx(q, hctx, i) {
2126                 cpumask_clear(hctx->cpumask);
2127                 hctx->nr_ctx = 0;
2128         }
2129
2130         /*
2131          * Map software to hardware queues.
2132          *
2133          * If the cpu isn't present, the cpu is mapped to first hctx.
2134          */
2135         for_each_present_cpu(i) {
2136                 hctx_idx = q->mq_map[i];
2137                 /* unmapped hw queue can be remapped after CPU topo changed */
2138                 if (!set->tags[hctx_idx] &&
2139                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2140                         /*
2141                          * If tags initialization fail for some hctx,
2142                          * that hctx won't be brought online.  In this
2143                          * case, remap the current ctx to hctx[0] which
2144                          * is guaranteed to always have tags allocated
2145                          */
2146                         q->mq_map[i] = 0;
2147                 }
2148
2149                 ctx = per_cpu_ptr(q->queue_ctx, i);
2150                 hctx = blk_mq_map_queue(q, i);
2151
2152                 cpumask_set_cpu(i, hctx->cpumask);
2153                 ctx->index_hw = hctx->nr_ctx;
2154                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2155         }
2156
2157         mutex_unlock(&q->sysfs_lock);
2158
2159         queue_for_each_hw_ctx(q, hctx, i) {
2160                 /*
2161                  * If no software queues are mapped to this hardware queue,
2162                  * disable it and free the request entries.
2163                  */
2164                 if (!hctx->nr_ctx) {
2165                         /* Never unmap queue 0.  We need it as a
2166                          * fallback in case of a new remap fails
2167                          * allocation
2168                          */
2169                         if (i && set->tags[i])
2170                                 blk_mq_free_map_and_requests(set, i);
2171
2172                         hctx->tags = NULL;
2173                         continue;
2174                 }
2175
2176                 hctx->tags = set->tags[i];
2177                 WARN_ON(!hctx->tags);
2178
2179                 /*
2180                  * Set the map size to the number of mapped software queues.
2181                  * This is more accurate and more efficient than looping
2182                  * over all possibly mapped software queues.
2183                  */
2184                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2185
2186                 /*
2187                  * Initialize batch roundrobin counts
2188                  */
2189                 hctx->next_cpu = cpumask_first(hctx->cpumask);
2190                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2191         }
2192 }
2193
2194 /*
2195  * Caller needs to ensure that we're either frozen/quiesced, or that
2196  * the queue isn't live yet.
2197  */
2198 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2199 {
2200         struct blk_mq_hw_ctx *hctx;
2201         int i;
2202
2203         queue_for_each_hw_ctx(q, hctx, i) {
2204                 if (shared) {
2205                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2206                                 atomic_inc(&q->shared_hctx_restart);
2207                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2208                 } else {
2209                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2210                                 atomic_dec(&q->shared_hctx_restart);
2211                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2212                 }
2213         }
2214 }
2215
2216 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2217                                         bool shared)
2218 {
2219         struct request_queue *q;
2220
2221         lockdep_assert_held(&set->tag_list_lock);
2222
2223         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2224                 blk_mq_freeze_queue(q);
2225                 queue_set_hctx_shared(q, shared);
2226                 blk_mq_unfreeze_queue(q);
2227         }
2228 }
2229
2230 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2231 {
2232         struct blk_mq_tag_set *set = q->tag_set;
2233
2234         mutex_lock(&set->tag_list_lock);
2235         list_del_rcu(&q->tag_set_list);
2236         INIT_LIST_HEAD(&q->tag_set_list);
2237         if (list_is_singular(&set->tag_list)) {
2238                 /* just transitioned to unshared */
2239                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2240                 /* update existing queue */
2241                 blk_mq_update_tag_set_depth(set, false);
2242         }
2243         mutex_unlock(&set->tag_list_lock);
2244
2245         synchronize_rcu();
2246 }
2247
2248 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2249                                      struct request_queue *q)
2250 {
2251         q->tag_set = set;
2252
2253         mutex_lock(&set->tag_list_lock);
2254
2255         /* Check to see if we're transitioning to shared (from 1 to 2 queues). */
2256         if (!list_empty(&set->tag_list) && !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2257                 set->flags |= BLK_MQ_F_TAG_SHARED;
2258                 /* update existing queue */
2259                 blk_mq_update_tag_set_depth(set, true);
2260         }
2261         if (set->flags & BLK_MQ_F_TAG_SHARED)
2262                 queue_set_hctx_shared(q, true);
2263         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2264
2265         mutex_unlock(&set->tag_list_lock);
2266 }
2267
2268 /*
2269  * It is the actual release handler for mq, but we do it from
2270  * request queue's release handler for avoiding use-after-free
2271  * and headache because q->mq_kobj shouldn't have been introduced,
2272  * but we can't group ctx/kctx kobj without it.
2273  */
2274 void blk_mq_release(struct request_queue *q)
2275 {
2276         struct blk_mq_hw_ctx *hctx;
2277         unsigned int i;
2278
2279         /* hctx kobj stays in hctx */
2280         queue_for_each_hw_ctx(q, hctx, i) {
2281                 if (!hctx)
2282                         continue;
2283                 kobject_put(&hctx->kobj);
2284         }
2285
2286         q->mq_map = NULL;
2287
2288         kfree(q->queue_hw_ctx);
2289
2290         /*
2291          * release .mq_kobj and sw queue's kobject now because
2292          * both share lifetime with request queue.
2293          */
2294         blk_mq_sysfs_deinit(q);
2295
2296         free_percpu(q->queue_ctx);
2297 }
2298
2299 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2300 {
2301         struct request_queue *uninit_q, *q;
2302
2303         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2304         if (!uninit_q)
2305                 return ERR_PTR(-ENOMEM);
2306
2307         q = blk_mq_init_allocated_queue(set, uninit_q);
2308         if (IS_ERR(q))
2309                 blk_cleanup_queue(uninit_q);
2310
2311         return q;
2312 }
2313 EXPORT_SYMBOL(blk_mq_init_queue);
2314
2315 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2316 {
2317         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2318
2319         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
2320                            __alignof__(struct blk_mq_hw_ctx)) !=
2321                      sizeof(struct blk_mq_hw_ctx));
2322
2323         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2324                 hw_ctx_size += sizeof(struct srcu_struct);
2325
2326         return hw_ctx_size;
2327 }
2328
2329 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2330                                                 struct request_queue *q)
2331 {
2332         int i, j;
2333         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2334
2335         blk_mq_sysfs_unregister(q);
2336
2337         /* protect against switching io scheduler  */
2338         mutex_lock(&q->sysfs_lock);
2339         for (i = 0; i < set->nr_hw_queues; i++) {
2340                 int node;
2341
2342                 if (hctxs[i])
2343                         continue;
2344
2345                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2346                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2347                                         GFP_KERNEL, node);
2348                 if (!hctxs[i])
2349                         break;
2350
2351                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2352                                                 node)) {
2353                         kfree(hctxs[i]);
2354                         hctxs[i] = NULL;
2355                         break;
2356                 }
2357
2358                 atomic_set(&hctxs[i]->nr_active, 0);
2359                 hctxs[i]->numa_node = node;
2360                 hctxs[i]->queue_num = i;
2361
2362                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2363                         free_cpumask_var(hctxs[i]->cpumask);
2364                         kfree(hctxs[i]);
2365                         hctxs[i] = NULL;
2366                         break;
2367                 }
2368                 blk_mq_hctx_kobj_init(hctxs[i]);
2369         }
2370         for (j = i; j < q->nr_hw_queues; j++) {
2371                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2372
2373                 if (hctx) {
2374                         if (hctx->tags)
2375                                 blk_mq_free_map_and_requests(set, j);
2376                         blk_mq_exit_hctx(q, set, hctx, j);
2377                         kobject_put(&hctx->kobj);
2378                         hctxs[j] = NULL;
2379
2380                 }
2381         }
2382         q->nr_hw_queues = i;
2383         mutex_unlock(&q->sysfs_lock);
2384         blk_mq_sysfs_register(q);
2385 }
2386
2387 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2388                                                   struct request_queue *q)
2389 {
2390         /* mark the queue as mq asap */
2391         q->mq_ops = set->ops;
2392
2393         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2394                                              blk_mq_poll_stats_bkt,
2395                                              BLK_MQ_POLL_STATS_BKTS, q);
2396         if (!q->poll_cb)
2397                 goto err_exit;
2398
2399         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2400         if (!q->queue_ctx)
2401                 goto err_exit;
2402
2403         /* init q->mq_kobj and sw queues' kobjects */
2404         blk_mq_sysfs_init(q);
2405
2406         q->queue_hw_ctx = kzalloc_node(nr_cpu_ids * sizeof(*(q->queue_hw_ctx)),
2407                                                 GFP_KERNEL, set->numa_node);
2408         if (!q->queue_hw_ctx)
2409                 goto err_percpu;
2410
2411         q->mq_map = set->mq_map;
2412
2413         blk_mq_realloc_hw_ctxs(set, q);
2414         if (!q->nr_hw_queues)
2415                 goto err_hctxs;
2416
2417         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2418         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2419
2420         q->nr_queues = nr_cpu_ids;
2421
2422         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2423
2424         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2425                 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
2426
2427         q->sg_reserved_size = INT_MAX;
2428
2429         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2430         INIT_LIST_HEAD(&q->requeue_list);
2431         spin_lock_init(&q->requeue_lock);
2432
2433         blk_queue_make_request(q, blk_mq_make_request);
2434
2435         /*
2436          * Do this after blk_queue_make_request() overrides it...
2437          */
2438         q->nr_requests = set->queue_depth;
2439
2440         /*
2441          * Default to classic polling
2442          */
2443         q->poll_nsec = -1;
2444
2445         if (set->ops->complete)
2446                 blk_queue_softirq_done(q, set->ops->complete);
2447
2448         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2449         blk_mq_add_queue_tag_set(set, q);
2450         blk_mq_map_swqueue(q);
2451
2452         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2453                 int ret;
2454
2455                 ret = blk_mq_sched_init(q);
2456                 if (ret)
2457                         return ERR_PTR(ret);
2458         }
2459
2460         return q;
2461
2462 err_hctxs:
2463         kfree(q->queue_hw_ctx);
2464 err_percpu:
2465         free_percpu(q->queue_ctx);
2466 err_exit:
2467         q->mq_ops = NULL;
2468         return ERR_PTR(-ENOMEM);
2469 }
2470 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2471
2472 void blk_mq_free_queue(struct request_queue *q)
2473 {
2474         struct blk_mq_tag_set   *set = q->tag_set;
2475
2476         blk_mq_del_queue_tag_set(q);
2477         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2478 }
2479
2480 /* Basically redo blk_mq_init_queue with queue frozen */
2481 static void blk_mq_queue_reinit(struct request_queue *q)
2482 {
2483         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2484
2485         blk_mq_debugfs_unregister_hctxs(q);
2486         blk_mq_sysfs_unregister(q);
2487
2488         /*
2489          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2490          * we should change hctx numa_node according to new topology (this
2491          * involves free and re-allocate memory, worthy doing?)
2492          */
2493
2494         blk_mq_map_swqueue(q);
2495
2496         blk_mq_sysfs_register(q);
2497         blk_mq_debugfs_register_hctxs(q);
2498 }
2499
2500 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2501 {
2502         int i;
2503
2504         for (i = 0; i < set->nr_hw_queues; i++)
2505                 if (!__blk_mq_alloc_rq_map(set, i))
2506                         goto out_unwind;
2507
2508         return 0;
2509
2510 out_unwind:
2511         while (--i >= 0)
2512                 blk_mq_free_rq_map(set->tags[i]);
2513
2514         return -ENOMEM;
2515 }
2516
2517 /*
2518  * Allocate the request maps associated with this tag_set. Note that this
2519  * may reduce the depth asked for, if memory is tight. set->queue_depth
2520  * will be updated to reflect the allocated depth.
2521  */
2522 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2523 {
2524         unsigned int depth;
2525         int err;
2526
2527         depth = set->queue_depth;
2528         do {
2529                 err = __blk_mq_alloc_rq_maps(set);
2530                 if (!err)
2531                         break;
2532
2533                 set->queue_depth >>= 1;
2534                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2535                         err = -ENOMEM;
2536                         break;
2537                 }
2538         } while (set->queue_depth);
2539
2540         if (!set->queue_depth || err) {
2541                 pr_err("blk-mq: failed to allocate request map\n");
2542                 return -ENOMEM;
2543         }
2544
2545         if (depth != set->queue_depth)
2546                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2547                                                 depth, set->queue_depth);
2548
2549         return 0;
2550 }
2551
2552 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2553 {
2554         if (set->ops->map_queues) {
2555                 int cpu;
2556                 /*
2557                  * transport .map_queues is usually done in the following
2558                  * way:
2559                  *
2560                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2561                  *      mask = get_cpu_mask(queue)
2562                  *      for_each_cpu(cpu, mask)
2563                  *              set->mq_map[cpu] = queue;
2564                  * }
2565                  *
2566                  * When we need to remap, the table has to be cleared for
2567                  * killing stale mapping since one CPU may not be mapped
2568                  * to any hw queue.
2569                  */
2570                 for_each_possible_cpu(cpu)
2571                         set->mq_map[cpu] = 0;
2572
2573                 return set->ops->map_queues(set);
2574         } else
2575                 return blk_mq_map_queues(set);
2576 }
2577
2578 /*
2579  * Alloc a tag set to be associated with one or more request queues.
2580  * May fail with EINVAL for various error conditions. May adjust the
2581  * requested depth down, if if it too large. In that case, the set
2582  * value will be stored in set->queue_depth.
2583  */
2584 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2585 {
2586         int ret;
2587
2588         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2589
2590         if (!set->nr_hw_queues)
2591                 return -EINVAL;
2592         if (!set->queue_depth)
2593                 return -EINVAL;
2594         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2595                 return -EINVAL;
2596
2597         if (!set->ops->queue_rq)
2598                 return -EINVAL;
2599
2600         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2601                 pr_info("blk-mq: reduced tag depth to %u\n",
2602                         BLK_MQ_MAX_DEPTH);
2603                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2604         }
2605
2606         /*
2607          * If a crashdump is active, then we are potentially in a very
2608          * memory constrained environment. Limit us to 1 queue and
2609          * 64 tags to prevent using too much memory.
2610          */
2611         if (is_kdump_kernel()) {
2612                 set->nr_hw_queues = 1;
2613                 set->queue_depth = min(64U, set->queue_depth);
2614         }
2615         /*
2616          * There is no use for more h/w queues than cpus.
2617          */
2618         if (set->nr_hw_queues > nr_cpu_ids)
2619                 set->nr_hw_queues = nr_cpu_ids;
2620
2621         set->tags = kzalloc_node(nr_cpu_ids * sizeof(struct blk_mq_tags *),
2622                                  GFP_KERNEL, set->numa_node);
2623         if (!set->tags)
2624                 return -ENOMEM;
2625
2626         ret = -ENOMEM;
2627         set->mq_map = kzalloc_node(sizeof(*set->mq_map) * nr_cpu_ids,
2628                         GFP_KERNEL, set->numa_node);
2629         if (!set->mq_map)
2630                 goto out_free_tags;
2631
2632         ret = blk_mq_update_queue_map(set);
2633         if (ret)
2634                 goto out_free_mq_map;
2635
2636         ret = blk_mq_alloc_rq_maps(set);
2637         if (ret)
2638                 goto out_free_mq_map;
2639
2640         mutex_init(&set->tag_list_lock);
2641         INIT_LIST_HEAD(&set->tag_list);
2642
2643         return 0;
2644
2645 out_free_mq_map:
2646         kfree(set->mq_map);
2647         set->mq_map = NULL;
2648 out_free_tags:
2649         kfree(set->tags);
2650         set->tags = NULL;
2651         return ret;
2652 }
2653 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2654
2655 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2656 {
2657         int i;
2658
2659         for (i = 0; i < nr_cpu_ids; i++)
2660                 blk_mq_free_map_and_requests(set, i);
2661
2662         kfree(set->mq_map);
2663         set->mq_map = NULL;
2664
2665         kfree(set->tags);
2666         set->tags = NULL;
2667 }
2668 EXPORT_SYMBOL(blk_mq_free_tag_set);
2669
2670 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2671 {
2672         struct blk_mq_tag_set *set = q->tag_set;
2673         struct blk_mq_hw_ctx *hctx;
2674         int i, ret;
2675
2676         if (!set)
2677                 return -EINVAL;
2678
2679         blk_mq_freeze_queue(q);
2680
2681         ret = 0;
2682         queue_for_each_hw_ctx(q, hctx, i) {
2683                 if (!hctx->tags)
2684                         continue;
2685                 /*
2686                  * If we're using an MQ scheduler, just update the scheduler
2687                  * queue depth. This is similar to what the old code would do.
2688                  */
2689                 if (!hctx->sched_tags) {
2690                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags,
2691                                                         min(nr, set->queue_depth),
2692                                                         false);
2693                 } else {
2694                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2695                                                         nr, true);
2696                 }
2697                 if (ret)
2698                         break;
2699         }
2700
2701         if (!ret)
2702                 q->nr_requests = nr;
2703
2704         blk_mq_unfreeze_queue(q);
2705
2706         return ret;
2707 }
2708
2709 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2710                                                         int nr_hw_queues)
2711 {
2712         struct request_queue *q;
2713
2714         lockdep_assert_held(&set->tag_list_lock);
2715
2716         if (nr_hw_queues > nr_cpu_ids)
2717                 nr_hw_queues = nr_cpu_ids;
2718         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2719                 return;
2720
2721         list_for_each_entry(q, &set->tag_list, tag_set_list)
2722                 blk_mq_freeze_queue(q);
2723
2724         set->nr_hw_queues = nr_hw_queues;
2725         blk_mq_update_queue_map(set);
2726         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2727                 blk_mq_realloc_hw_ctxs(set, q);
2728                 blk_mq_queue_reinit(q);
2729         }
2730
2731         list_for_each_entry(q, &set->tag_list, tag_set_list)
2732                 blk_mq_unfreeze_queue(q);
2733 }
2734
2735 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2736 {
2737         mutex_lock(&set->tag_list_lock);
2738         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2739         mutex_unlock(&set->tag_list_lock);
2740 }
2741 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2742
2743 /* Enable polling stats and return whether they were already enabled. */
2744 static bool blk_poll_stats_enable(struct request_queue *q)
2745 {
2746         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2747             test_and_set_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
2748                 return true;
2749         blk_stat_add_callback(q, q->poll_cb);
2750         return false;
2751 }
2752
2753 static void blk_mq_poll_stats_start(struct request_queue *q)
2754 {
2755         /*
2756          * We don't arm the callback if polling stats are not enabled or the
2757          * callback is already active.
2758          */
2759         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2760             blk_stat_is_active(q->poll_cb))
2761                 return;
2762
2763         blk_stat_activate_msecs(q->poll_cb, 100);
2764 }
2765
2766 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2767 {
2768         struct request_queue *q = cb->data;
2769         int bucket;
2770
2771         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2772                 if (cb->stat[bucket].nr_samples)
2773                         q->poll_stat[bucket] = cb->stat[bucket];
2774         }
2775 }
2776
2777 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2778                                        struct blk_mq_hw_ctx *hctx,
2779                                        struct request *rq)
2780 {
2781         unsigned long ret = 0;
2782         int bucket;
2783
2784         /*
2785          * If stats collection isn't on, don't sleep but turn it on for
2786          * future users
2787          */
2788         if (!blk_poll_stats_enable(q))
2789                 return 0;
2790
2791         /*
2792          * As an optimistic guess, use half of the mean service time
2793          * for this type of request. We can (and should) make this smarter.
2794          * For instance, if the completion latencies are tight, we can
2795          * get closer than just half the mean. This is especially
2796          * important on devices where the completion latencies are longer
2797          * than ~10 usec. We do use the stats for the relevant IO size
2798          * if available which does lead to better estimates.
2799          */
2800         bucket = blk_mq_poll_stats_bkt(rq);
2801         if (bucket < 0)
2802                 return ret;
2803
2804         if (q->poll_stat[bucket].nr_samples)
2805                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2806
2807         return ret;
2808 }
2809
2810 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2811                                      struct blk_mq_hw_ctx *hctx,
2812                                      struct request *rq)
2813 {
2814         struct hrtimer_sleeper hs;
2815         enum hrtimer_mode mode;
2816         unsigned int nsecs;
2817         ktime_t kt;
2818
2819         if (test_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags))
2820                 return false;
2821
2822         /*
2823          * poll_nsec can be:
2824          *
2825          * -1:  don't ever hybrid sleep
2826          *  0:  use half of prev avg
2827          * >0:  use this specific value
2828          */
2829         if (q->poll_nsec == -1)
2830                 return false;
2831         else if (q->poll_nsec > 0)
2832                 nsecs = q->poll_nsec;
2833         else
2834                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2835
2836         if (!nsecs)
2837                 return false;
2838
2839         set_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
2840
2841         /*
2842          * This will be replaced with the stats tracking code, using
2843          * 'avg_completion_time / 2' as the pre-sleep target.
2844          */
2845         kt = nsecs;
2846
2847         mode = HRTIMER_MODE_REL;
2848         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2849         hrtimer_set_expires(&hs.timer, kt);
2850
2851         hrtimer_init_sleeper(&hs, current);
2852         do {
2853                 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
2854                         break;
2855                 set_current_state(TASK_UNINTERRUPTIBLE);
2856                 hrtimer_start_expires(&hs.timer, mode);
2857                 if (hs.task)
2858                         io_schedule();
2859                 hrtimer_cancel(&hs.timer);
2860                 mode = HRTIMER_MODE_ABS;
2861         } while (hs.task && !signal_pending(current));
2862
2863         __set_current_state(TASK_RUNNING);
2864         destroy_hrtimer_on_stack(&hs.timer);
2865         return true;
2866 }
2867
2868 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
2869 {
2870         struct request_queue *q = hctx->queue;
2871         long state;
2872
2873         /*
2874          * If we sleep, have the caller restart the poll loop to reset
2875          * the state. Like for the other success return cases, the
2876          * caller is responsible for checking if the IO completed. If
2877          * the IO isn't complete, we'll get called again and will go
2878          * straight to the busy poll loop.
2879          */
2880         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
2881                 return true;
2882
2883         hctx->poll_considered++;
2884
2885         state = current->state;
2886         while (!need_resched()) {
2887                 int ret;
2888
2889                 hctx->poll_invoked++;
2890
2891                 ret = q->mq_ops->poll(hctx, rq->tag);
2892                 if (ret > 0) {
2893                         hctx->poll_success++;
2894                         set_current_state(TASK_RUNNING);
2895                         return true;
2896                 }
2897
2898                 if (signal_pending_state(state, current))
2899                         set_current_state(TASK_RUNNING);
2900
2901                 if (current->state == TASK_RUNNING)
2902                         return true;
2903                 if (ret < 0)
2904                         break;
2905                 cpu_relax();
2906         }
2907
2908         return false;
2909 }
2910
2911 bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
2912 {
2913         struct blk_mq_hw_ctx *hctx;
2914         struct blk_plug *plug;
2915         struct request *rq;
2916
2917         if (!q->mq_ops || !q->mq_ops->poll || !blk_qc_t_valid(cookie) ||
2918             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
2919                 return false;
2920
2921         plug = current->plug;
2922         if (plug)
2923                 blk_flush_plug_list(plug, false);
2924
2925         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
2926         if (!blk_qc_t_is_internal(cookie))
2927                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
2928         else {
2929                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
2930                 /*
2931                  * With scheduling, if the request has completed, we'll
2932                  * get a NULL return here, as we clear the sched tag when
2933                  * that happens. The request still remains valid, like always,
2934                  * so we should be safe with just the NULL check.
2935                  */
2936                 if (!rq)
2937                         return false;
2938         }
2939
2940         return __blk_mq_poll(hctx, rq);
2941 }
2942 EXPORT_SYMBOL_GPL(blk_mq_poll);
2943
2944 static int __init blk_mq_init(void)
2945 {
2946         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
2947                                 blk_mq_hctx_notify_dead);
2948         return 0;
2949 }
2950 subsys_initcall(blk_mq_init);