block/blk-mq.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Block multiqueue core code
   4  *
   5  * Copyright (C) 2013-2014 Jens Axboe
   6  * Copyright (C) 2013-2014 Christoph Hellwig
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/backing-dev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/kmemleak.h>
  14 #include <linux/mm.h>
  15 #include <linux/init.h>
  16 #include <linux/slab.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/smp.h>
  19 #include <linux/llist.h>
  20 #include <linux/list_sort.h>
  21 #include <linux/cpu.h>
  22 #include <linux/cache.h>
  23 #include <linux/sched/sysctl.h>
  24 #include <linux/sched/topology.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/delay.h>
  27 #include <linux/crash_dump.h>
  28 #include <linux/prefetch.h>
  29
  30 #include <trace/events/block.h>
  31
  32 #include <linux/blk-mq.h>
  33 #include <linux/t10-pi.h>
  34 #include "blk.h"
  35 #include "blk-mq.h"
  36 #include "blk-mq-debugfs.h"
  37 #include "blk-mq-tag.h"
  38 #include "blk-pm.h"
  39 #include "blk-stat.h"
  40 #include "blk-mq-sched.h"
  41 #include "blk-rq-qos.h"
  42
  43 static void blk_mq_poll_stats_start(struct request_queue *q);
  44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  45
  46 static int blk_mq_poll_stats_bkt(const struct request *rq)
  47 {
  48         int ddir, sectors, bucket;
  49
  50         ddir = rq_data_dir(rq);
  51         sectors = blk_rq_stats_sectors(rq);
  52
  53         bucket = ddir + 2 * ilog2(sectors);
  54
  55         if (bucket < 0)
  56                 return -1;
  57         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  58                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  59
  60         return bucket;
  61 }
  62
  63 /*
  64  * Check if any of the ctx, dispatch list or elevator
  65  * have pending work in this hardware queue.
  66  */
  67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  68 {
  69         return !list_empty_careful(&hctx->dispatch) ||
  70                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  71                         blk_mq_sched_has_work(hctx);
  72 }
  73
  74 /*
  75  * Mark this ctx as having pending work in this hardware queue
  76  */
  77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  78                                      struct blk_mq_ctx *ctx)
  79 {
  80         const int bit = ctx->index_hw[hctx->type];
  81
  82         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
  83                 sbitmap_set_bit(&hctx->ctx_map, bit);
  84 }
  85
  86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  87                                       struct blk_mq_ctx *ctx)
  88 {
  89         const int bit = ctx->index_hw[hctx->type];
  90
  91         sbitmap_clear_bit(&hctx->ctx_map, bit);
  92 }
  93
  94 struct mq_inflight {
  95         struct hd_struct *part;
  96         unsigned int inflight[2];
  97 };
  98
  99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
 100                                   struct request *rq, void *priv,
 101                                   bool reserved)
 102 {
 103         struct mq_inflight *mi = priv;
 104
 105         if (rq->part == mi->part)
 106                 mi->inflight[rq_data_dir(rq)]++;
 107
 108         return true;
 109 }
 110
 111 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
 112 {
 113         struct mq_inflight mi = { .part = part };
 114
 115         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 116
 117         return mi.inflight[0] + mi.inflight[1];
 118 }
 119
 120 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 121                          unsigned int inflight[2])
 122 {
 123         struct mq_inflight mi = { .part = part };
 124
 125         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 126         inflight[0] = mi.inflight[0];
 127         inflight[1] = mi.inflight[1];
 128 }
 129
 130 void blk_freeze_queue_start(struct request_queue *q)
 131 {
 132         mutex_lock(&q->mq_freeze_lock);
 133         if (++q->mq_freeze_depth == 1) {
 134                 percpu_ref_kill(&q->q_usage_counter);
 135                 mutex_unlock(&q->mq_freeze_lock);
 136                 if (queue_is_mq(q))
 137                         blk_mq_run_hw_queues(q, false);
 138         } else {
 139                 mutex_unlock(&q->mq_freeze_lock);
 140         }
 141 }
 142 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 143
 144 void blk_mq_freeze_queue_wait(struct request_queue *q)
 145 {
 146         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 147 }
 148 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 149
 150 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 151                                      unsigned long timeout)
 152 {
 153         return wait_event_timeout(q->mq_freeze_wq,
 154                                         percpu_ref_is_zero(&q->q_usage_counter),
 155                                         timeout);
 156 }
 157 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 158
 159 /*
 160  * Guarantee no request is in use, so we can change any data structure of
 161  * the queue afterward.
 162  */
 163 void blk_freeze_queue(struct request_queue *q)
 164 {
 165         /*
 166          * In the !blk_mq case we are only calling this to kill the
 167          * q_usage_counter, otherwise this increases the freeze depth
 168          * and waits for it to return to zero.  For this reason there is
 169          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 170          * exported to drivers as the only user for unfreeze is blk_mq.
 171          */
 172         blk_freeze_queue_start(q);
 173         blk_mq_freeze_queue_wait(q);
 174 }
 175
 176 void blk_mq_freeze_queue(struct request_queue *q)
 177 {
 178         /*
 179          * ...just an alias to keep freeze and unfreeze actions balanced
 180          * in the blk_mq_* namespace
 181          */
 182         blk_freeze_queue(q);
 183 }
 184 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 185
 186 void blk_mq_unfreeze_queue(struct request_queue *q)
 187 {
 188         mutex_lock(&q->mq_freeze_lock);
 189         q->mq_freeze_depth--;
 190         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 191         if (!q->mq_freeze_depth) {
 192                 percpu_ref_resurrect(&q->q_usage_counter);
 193                 wake_up_all(&q->mq_freeze_wq);
 194         }
 195         mutex_unlock(&q->mq_freeze_lock);
 196 }
 197 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 198
 199 /*
 200  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 201  * mpt3sas driver such that this function can be removed.
 202  */
 203 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 204 {
 205         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 206 }
 207 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 208
 209 /**
 210  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 211  * @q: request queue.
 212  *
 213  * Note: this function does not prevent that the struct request end_io()
 214  * callback function is invoked. Once this function is returned, we make
 215  * sure no dispatch can happen until the queue is unquiesced via
 216  * blk_mq_unquiesce_queue().
 217  */
 218 void blk_mq_quiesce_queue(struct request_queue *q)
 219 {
 220         struct blk_mq_hw_ctx *hctx;
 221         unsigned int i;
 222         bool rcu = false;
 223
 224         blk_mq_quiesce_queue_nowait(q);
 225
 226         queue_for_each_hw_ctx(q, hctx, i) {
 227                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 228                         synchronize_srcu(hctx->srcu);
 229                 else
 230                         rcu = true;
 231         }
 232         if (rcu)
 233                 synchronize_rcu();
 234 }
 235 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 236
 237 /*
 238  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 239  * @q: request queue.
 240  *
 241  * This function recovers queue into the state before quiescing
 242  * which is done by blk_mq_quiesce_queue.
 243  */
 244 void blk_mq_unquiesce_queue(struct request_queue *q)
 245 {
 246         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 247
 248         /* dispatch requests which are inserted during quiescing */
 249         blk_mq_run_hw_queues(q, true);
 250 }
 251 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 252
 253 void blk_mq_wake_waiters(struct request_queue *q)
 254 {
 255         struct blk_mq_hw_ctx *hctx;
 256         unsigned int i;
 257
 258         queue_for_each_hw_ctx(q, hctx, i)
 259                 if (blk_mq_hw_queue_mapped(hctx))
 260                         blk_mq_tag_wakeup_all(hctx->tags, true);
 261 }
 262
 263 /*
 264  * Only need start/end time stamping if we have iostat or
 265  * blk stats enabled, or using an IO scheduler.
 266  */
 267 static inline bool blk_mq_need_time_stamp(struct request *rq)
 268 {
 269         return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
 270 }
 271
 272 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 273                 unsigned int tag, unsigned int op, u64 alloc_time_ns)
 274 {
 275         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 276         struct request *rq = tags->static_rqs[tag];
 277         req_flags_t rq_flags = 0;
 278
 279         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 280                 rq->tag = -1;
 281                 rq->internal_tag = tag;
 282         } else {
 283                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
 284                         rq_flags = RQF_MQ_INFLIGHT;
 285                         atomic_inc(&data->hctx->nr_active);
 286                 }
 287                 rq->tag = tag;
 288                 rq->internal_tag = -1;
 289                 data->hctx->tags->rqs[rq->tag] = rq;
 290         }
 291
 292         /* csd/requeue_work/fifo_time is initialized before use */
 293         rq->q = data->q;
 294         rq->mq_ctx = data->ctx;
 295         rq->mq_hctx = data->hctx;
 296         rq->rq_flags = rq_flags;
 297         rq->cmd_flags = op;
 298         if (data->flags & BLK_MQ_REQ_PREEMPT)
 299                 rq->rq_flags |= RQF_PREEMPT;
 300         if (blk_queue_io_stat(data->q))
 301                 rq->rq_flags |= RQF_IO_STAT;
 302         INIT_LIST_HEAD(&rq->queuelist);
 303         INIT_HLIST_NODE(&rq->hash);
 304         RB_CLEAR_NODE(&rq->rb_node);
 305         rq->rq_disk = NULL;
 306         rq->part = NULL;
 307 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
 308         rq->alloc_time_ns = alloc_time_ns;
 309 #endif
 310         if (blk_mq_need_time_stamp(rq))
 311                 rq->start_time_ns = ktime_get_ns();
 312         else
 313                 rq->start_time_ns = 0;
 314         rq->io_start_time_ns = 0;
 315         rq->stats_sectors = 0;
 316         rq->nr_phys_segments = 0;
 317 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 318         rq->nr_integrity_segments = 0;
 319 #endif
 320         /* tag was already set */
 321         rq->extra_len = 0;
 322         WRITE_ONCE(rq->deadline, 0);
 323
 324         rq->timeout = 0;
 325
 326         rq->end_io = NULL;
 327         rq->end_io_data = NULL;
 328
 329         data->ctx->rq_dispatched[op_is_sync(op)]++;
 330         refcount_set(&rq->ref, 1);
 331         return rq;
 332 }
 333
 334 static struct request *blk_mq_get_request(struct request_queue *q,
 335                                           struct bio *bio,
 336                                           struct blk_mq_alloc_data *data)
 337 {
 338         struct elevator_queue *e = q->elevator;
 339         struct request *rq;
 340         unsigned int tag;
 341         bool clear_ctx_on_error = false;
 342         u64 alloc_time_ns = 0;
 343
 344         blk_queue_enter_live(q);
 345
 346         /* alloc_time includes depth and tag waits */
 347         if (blk_queue_rq_alloc_time(q))
 348                 alloc_time_ns = ktime_get_ns();
 349
 350         data->q = q;
 351         if (likely(!data->ctx)) {
 352                 data->ctx = blk_mq_get_ctx(q);
 353                 clear_ctx_on_error = true;
 354         }
 355         if (likely(!data->hctx))
 356                 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
 357                                                 data->ctx);
 358         if (data->cmd_flags & REQ_NOWAIT)
 359                 data->flags |= BLK_MQ_REQ_NOWAIT;
 360
 361         if (e) {
 362                 data->flags |= BLK_MQ_REQ_INTERNAL;
 363
 364                 /*
 365                  * Flush requests are special and go directly to the
 366                  * dispatch list. Don't include reserved tags in the
 367                  * limiting, as it isn't useful.
 368                  */
 369                 if (!op_is_flush(data->cmd_flags) &&
 370                     e->type->ops.limit_depth &&
 371                     !(data->flags & BLK_MQ_REQ_RESERVED))
 372                         e->type->ops.limit_depth(data->cmd_flags, data);
 373         } else {
 374                 blk_mq_tag_busy(data->hctx);
 375         }
 376
 377         tag = blk_mq_get_tag(data);
 378         if (tag == BLK_MQ_TAG_FAIL) {
 379                 if (clear_ctx_on_error)
 380                         data->ctx = NULL;
 381                 blk_queue_exit(q);
 382                 return NULL;
 383         }
 384
 385         rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
 386         if (!op_is_flush(data->cmd_flags)) {
 387                 rq->elv.icq = NULL;
 388                 if (e && e->type->ops.prepare_request) {
 389                         if (e->type->icq_cache)
 390                                 blk_mq_sched_assign_ioc(rq);
 391
 392                         e->type->ops.prepare_request(rq, bio);
 393                         rq->rq_flags |= RQF_ELVPRIV;
 394                 }
 395         }
 396         data->hctx->queued++;
 397         return rq;
 398 }
 399
 400 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 401                 blk_mq_req_flags_t flags)
 402 {
 403         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 404         struct request *rq;
 405         int ret;
 406
 407         ret = blk_queue_enter(q, flags);
 408         if (ret)
 409                 return ERR_PTR(ret);
 410
 411         rq = blk_mq_get_request(q, NULL, &alloc_data);
 412         blk_queue_exit(q);
 413
 414         if (!rq)
 415                 return ERR_PTR(-EWOULDBLOCK);
 416
 417         rq->__data_len = 0;
 418         rq->__sector = (sector_t) -1;
 419         rq->bio = rq->biotail = NULL;
 420         return rq;
 421 }
 422 EXPORT_SYMBOL(blk_mq_alloc_request);
 423
 424 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 425         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 426 {
 427         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 428         struct request *rq;
 429         unsigned int cpu;
 430         int ret;
 431
 432         /*
 433          * If the tag allocator sleeps we could get an allocation for a
 434          * different hardware context.  No need to complicate the low level
 435          * allocator for this for the rare use case of a command tied to
 436          * a specific queue.
 437          */
 438         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 439                 return ERR_PTR(-EINVAL);
 440
 441         if (hctx_idx >= q->nr_hw_queues)
 442                 return ERR_PTR(-EIO);
 443
 444         ret = blk_queue_enter(q, flags);
 445         if (ret)
 446                 return ERR_PTR(ret);
 447
 448         /*
 449          * Check if the hardware context is actually mapped to anything.
 450          * If not tell the caller that it should skip this queue.
 451          */
 452         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 453         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 454                 blk_queue_exit(q);
 455                 return ERR_PTR(-EXDEV);
 456         }
 457         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 458         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 459
 460         rq = blk_mq_get_request(q, NULL, &alloc_data);
 461         blk_queue_exit(q);
 462
 463         if (!rq)
 464                 return ERR_PTR(-EWOULDBLOCK);
 465
 466         return rq;
 467 }
 468 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 469
 470 static void __blk_mq_free_request(struct request *rq)
 471 {
 472         struct request_queue *q = rq->q;
 473         struct blk_mq_ctx *ctx = rq->mq_ctx;
 474         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 475         const int sched_tag = rq->internal_tag;
 476
 477         blk_pm_mark_last_busy(rq);
 478         rq->mq_hctx = NULL;
 479         if (rq->tag != -1)
 480                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 481         if (sched_tag != -1)
 482                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 483         blk_mq_sched_restart(hctx);
 484         blk_queue_exit(q);
 485 }
 486
 487 void blk_mq_free_request(struct request *rq)
 488 {
 489         struct request_queue *q = rq->q;
 490         struct elevator_queue *e = q->elevator;
 491         struct blk_mq_ctx *ctx = rq->mq_ctx;
 492         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 493
 494         if (rq->rq_flags & RQF_ELVPRIV) {
 495                 if (e && e->type->ops.finish_request)
 496                         e->type->ops.finish_request(rq);
 497                 if (rq->elv.icq) {
 498                         put_io_context(rq->elv.icq->ioc);
 499                         rq->elv.icq = NULL;
 500                 }
 501         }
 502
 503         ctx->rq_completed[rq_is_sync(rq)]++;
 504         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 505                 atomic_dec(&hctx->nr_active);
 506
 507         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 508                 laptop_io_completion(q->backing_dev_info);
 509
 510         rq_qos_done(q, rq);
 511
 512         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 513         if (refcount_dec_and_test(&rq->ref))
 514                 __blk_mq_free_request(rq);
 515 }
 516 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 517
 518 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 519 {
 520         u64 now = 0;
 521
 522         if (blk_mq_need_time_stamp(rq))
 523                 now = ktime_get_ns();
 524
 525         if (rq->rq_flags & RQF_STATS) {
 526                 blk_mq_poll_stats_start(rq->q);
 527                 blk_stat_add(rq, now);
 528         }
 529
 530         if (rq->internal_tag != -1)
 531                 blk_mq_sched_completed_request(rq, now);
 532
 533         blk_account_io_done(rq, now);
 534
 535         if (rq->end_io) {
 536                 rq_qos_done(rq->q, rq);
 537                 rq->end_io(rq, error);
 538         } else {
 539                 blk_mq_free_request(rq);
 540         }
 541 }
 542 EXPORT_SYMBOL(__blk_mq_end_request);
 543
 544 void blk_mq_end_request(struct request *rq, blk_status_t error)
 545 {
 546         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 547                 BUG();
 548         __blk_mq_end_request(rq, error);
 549 }
 550 EXPORT_SYMBOL(blk_mq_end_request);
 551
 552 static void __blk_mq_complete_request_remote(void *data)
 553 {
 554         struct request *rq = data;
 555         struct request_queue *q = rq->q;
 556
 557         q->mq_ops->complete(rq);
 558 }
 559
 560 static void __blk_mq_complete_request(struct request *rq)
 561 {
 562         struct blk_mq_ctx *ctx = rq->mq_ctx;
 563         struct request_queue *q = rq->q;
 564         bool shared = false;
 565         int cpu;
 566
 567         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 568         /*
 569          * Most of single queue controllers, there is only one irq vector
 570          * for handling IO completion, and the only irq's affinity is set
 571          * as all possible CPUs. On most of ARCHs, this affinity means the
 572          * irq is handled on one specific CPU.
 573          *
 574          * So complete IO reqeust in softirq context in case of single queue
 575          * for not degrading IO performance by irqsoff latency.
 576          */
 577         if (q->nr_hw_queues == 1) {
 578                 __blk_complete_request(rq);
 579                 return;
 580         }
 581
 582         /*
 583          * For a polled request, always complete locallly, it's pointless
 584          * to redirect the completion.
 585          */
 586         if ((rq->cmd_flags & REQ_HIPRI) ||
 587             !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
 588                 q->mq_ops->complete(rq);
 589                 return;
 590         }
 591
 592         cpu = get_cpu();
 593         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
 594                 shared = cpus_share_cache(cpu, ctx->cpu);
 595
 596         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 597                 rq->csd.func = __blk_mq_complete_request_remote;
 598                 rq->csd.info = rq;
 599                 rq->csd.flags = 0;
 600                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 601         } else {
 602                 q->mq_ops->complete(rq);
 603         }
 604         put_cpu();
 605 }
 606
 607 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 608         __releases(hctx->srcu)
 609 {
 610         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 611                 rcu_read_unlock();
 612         else
 613                 srcu_read_unlock(hctx->srcu, srcu_idx);
 614 }
 615
 616 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 617         __acquires(hctx->srcu)
 618 {
 619         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 620                 /* shut up gcc false positive */
 621                 *srcu_idx = 0;
 622                 rcu_read_lock();
 623         } else
 624                 *srcu_idx = srcu_read_lock(hctx->srcu);
 625 }
 626
 627 /**
 628  * blk_mq_complete_request - end I/O on a request
 629  * @rq:         the request being processed
 630  *
 631  * Description:
 632  *      Ends all I/O on a request. It does not handle partial completions.
 633  *      The actual completion happens out-of-order, through a IPI handler.
 634  **/
 635 bool blk_mq_complete_request(struct request *rq)
 636 {
 637         if (unlikely(blk_should_fake_timeout(rq->q)))
 638                 return false;
 639         __blk_mq_complete_request(rq);
 640         return true;
 641 }
 642 EXPORT_SYMBOL(blk_mq_complete_request);
 643
 644 void blk_mq_start_request(struct request *rq)
 645 {
 646         struct request_queue *q = rq->q;
 647
 648         trace_block_rq_issue(q, rq);
 649
 650         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 651                 rq->io_start_time_ns = ktime_get_ns();
 652                 rq->stats_sectors = blk_rq_sectors(rq);
 653                 rq->rq_flags |= RQF_STATS;
 654                 rq_qos_issue(q, rq);
 655         }
 656
 657         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 658
 659         blk_add_timer(rq);
 660         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 661
 662         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 663                 /*
 664                  * Make sure space for the drain appears.  We know we can do
 665                  * this because max_hw_segments has been adjusted to be one
 666                  * fewer than the device can handle.
 667                  */
 668                 rq->nr_phys_segments++;
 669         }
 670
 671 #ifdef CONFIG_BLK_DEV_INTEGRITY
 672         if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
 673                 q->integrity.profile->prepare_fn(rq);
 674 #endif
 675 }
 676 EXPORT_SYMBOL(blk_mq_start_request);
 677
 678 static void __blk_mq_requeue_request(struct request *rq)
 679 {
 680         struct request_queue *q = rq->q;
 681
 682         blk_mq_put_driver_tag(rq);
 683
 684         trace_block_rq_requeue(q, rq);
 685         rq_qos_requeue(q, rq);
 686
 687         if (blk_mq_request_started(rq)) {
 688                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 689                 rq->rq_flags &= ~RQF_TIMED_OUT;
 690                 if (q->dma_drain_size && blk_rq_bytes(rq))
 691                         rq->nr_phys_segments--;
 692         }
 693 }
 694
 695 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 696 {
 697         __blk_mq_requeue_request(rq);
 698
 699         /* this request will be re-inserted to io scheduler queue */
 700         blk_mq_sched_requeue_request(rq);
 701
 702         BUG_ON(!list_empty(&rq->queuelist));
 703         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 704 }
 705 EXPORT_SYMBOL(blk_mq_requeue_request);
 706
 707 static void blk_mq_requeue_work(struct work_struct *work)
 708 {
 709         struct request_queue *q =
 710                 container_of(work, struct request_queue, requeue_work.work);
 711         LIST_HEAD(rq_list);
 712         struct request *rq, *next;
 713
 714         spin_lock_irq(&q->requeue_lock);
 715         list_splice_init(&q->requeue_list, &rq_list);
 716         spin_unlock_irq(&q->requeue_lock);
 717
 718         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 719                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 720                         continue;
 721
 722                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 723                 list_del_init(&rq->queuelist);
 724                 /*
 725                  * If RQF_DONTPREP, rq has contained some driver specific
 726                  * data, so insert it to hctx dispatch list to avoid any
 727                  * merge.
 728                  */
 729                 if (rq->rq_flags & RQF_DONTPREP)
 730                         blk_mq_request_bypass_insert(rq, false);
 731                 else
 732                         blk_mq_sched_insert_request(rq, true, false, false);
 733         }
 734
 735         while (!list_empty(&rq_list)) {
 736                 rq = list_entry(rq_list.next, struct request, queuelist);
 737                 list_del_init(&rq->queuelist);
 738                 blk_mq_sched_insert_request(rq, false, false, false);
 739         }
 740
 741         blk_mq_run_hw_queues(q, false);
 742 }
 743
 744 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 745                                 bool kick_requeue_list)
 746 {
 747         struct request_queue *q = rq->q;
 748         unsigned long flags;
 749
 750         /*
 751          * We abuse this flag that is otherwise used by the I/O scheduler to
 752          * request head insertion from the workqueue.
 753          */
 754         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 755
 756         spin_lock_irqsave(&q->requeue_lock, flags);
 757         if (at_head) {
 758                 rq->rq_flags |= RQF_SOFTBARRIER;
 759                 list_add(&rq->queuelist, &q->requeue_list);
 760         } else {
 761                 list_add_tail(&rq->queuelist, &q->requeue_list);
 762         }
 763         spin_unlock_irqrestore(&q->requeue_lock, flags);
 764
 765         if (kick_requeue_list)
 766                 blk_mq_kick_requeue_list(q);
 767 }
 768
 769 void blk_mq_kick_requeue_list(struct request_queue *q)
 770 {
 771         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 772 }
 773 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 774
 775 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 776                                     unsigned long msecs)
 777 {
 778         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 779                                     msecs_to_jiffies(msecs));
 780 }
 781 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 782
 783 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 784 {
 785         if (tag < tags->nr_tags) {
 786                 prefetch(tags->rqs[tag]);
 787                 return tags->rqs[tag];
 788         }
 789
 790         return NULL;
 791 }
 792 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 793
 794 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 795                                void *priv, bool reserved)
 796 {
 797         /*
 798          * If we find a request that is inflight and the queue matches,
 799          * we know the queue is busy. Return false to stop the iteration.
 800          */
 801         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 802                 bool *busy = priv;
 803
 804                 *busy = true;
 805                 return false;
 806         }
 807
 808         return true;
 809 }
 810
 811 bool blk_mq_queue_inflight(struct request_queue *q)
 812 {
 813         bool busy = false;
 814
 815         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 816         return busy;
 817 }
 818 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 819
 820 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 821 {
 822         req->rq_flags |= RQF_TIMED_OUT;
 823         if (req->q->mq_ops->timeout) {
 824                 enum blk_eh_timer_return ret;
 825
 826                 ret = req->q->mq_ops->timeout(req, reserved);
 827                 if (ret == BLK_EH_DONE)
 828                         return;
 829                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 830         }
 831
 832         blk_add_timer(req);
 833 }
 834
 835 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 836 {
 837         unsigned long deadline;
 838
 839         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 840                 return false;
 841         if (rq->rq_flags & RQF_TIMED_OUT)
 842                 return false;
 843
 844         deadline = READ_ONCE(rq->deadline);
 845         if (time_after_eq(jiffies, deadline))
 846                 return true;
 847
 848         if (*next == 0)
 849                 *next = deadline;
 850         else if (time_after(*next, deadline))
 851                 *next = deadline;
 852         return false;
 853 }
 854
 855 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 856                 struct request *rq, void *priv, bool reserved)
 857 {
 858         unsigned long *next = priv;
 859
 860         /*
 861          * Just do a quick check if it is expired before locking the request in
 862          * so we're not unnecessarilly synchronizing across CPUs.
 863          */
 864         if (!blk_mq_req_expired(rq, next))
 865                 return true;
 866
 867         /*
 868          * We have reason to believe the request may be expired. Take a
 869          * reference on the request to lock this request lifetime into its
 870          * currently allocated context to prevent it from being reallocated in
 871          * the event the completion by-passes this timeout handler.
 872          *
 873          * If the reference was already released, then the driver beat the
 874          * timeout handler to posting a natural completion.
 875          */
 876         if (!refcount_inc_not_zero(&rq->ref))
 877                 return true;
 878
 879         /*
 880          * The request is now locked and cannot be reallocated underneath the
 881          * timeout handler's processing. Re-verify this exact request is truly
 882          * expired; if it is not expired, then the request was completed and
 883          * reallocated as a new request.
 884          */
 885         if (blk_mq_req_expired(rq, next))
 886                 blk_mq_rq_timed_out(rq, reserved);
 887
 888         if (is_flush_rq(rq, hctx))
 889                 rq->end_io(rq, 0);
 890         else if (refcount_dec_and_test(&rq->ref))
 891                 __blk_mq_free_request(rq);
 892
 893         return true;
 894 }
 895
 896 static void blk_mq_timeout_work(struct work_struct *work)
 897 {
 898         struct request_queue *q =
 899                 container_of(work, struct request_queue, timeout_work);
 900         unsigned long next = 0;
 901         struct blk_mq_hw_ctx *hctx;
 902         int i;
 903
 904         /* A deadlock might occur if a request is stuck requiring a
 905          * timeout at the same time a queue freeze is waiting
 906          * completion, since the timeout code would not be able to
 907          * acquire the queue reference here.
 908          *
 909          * That's why we don't use blk_queue_enter here; instead, we use
 910          * percpu_ref_tryget directly, because we need to be able to
 911          * obtain a reference even in the short window between the queue
 912          * starting to freeze, by dropping the first reference in
 913          * blk_freeze_queue_start, and the moment the last request is
 914          * consumed, marked by the instant q_usage_counter reaches
 915          * zero.
 916          */
 917         if (!percpu_ref_tryget(&q->q_usage_counter))
 918                 return;
 919
 920         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 921
 922         if (next != 0) {
 923                 mod_timer(&q->timeout, next);
 924         } else {
 925                 /*
 926                  * Request timeouts are handled as a forward rolling timer. If
 927                  * we end up here it means that no requests are pending and
 928                  * also that no request has been pending for a while. Mark
 929                  * each hctx as idle.
 930                  */
 931                 queue_for_each_hw_ctx(q, hctx, i) {
 932                         /* the hctx may be unmapped, so check it here */
 933                         if (blk_mq_hw_queue_mapped(hctx))
 934                                 blk_mq_tag_idle(hctx);
 935                 }
 936         }
 937         blk_queue_exit(q);
 938 }
 939
 940 struct flush_busy_ctx_data {
 941         struct blk_mq_hw_ctx *hctx;
 942         struct list_head *list;
 943 };
 944
 945 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 946 {
 947         struct flush_busy_ctx_data *flush_data = data;
 948         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 949         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 950         enum hctx_type type = hctx->type;
 951
 952         spin_lock(&ctx->lock);
 953         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 954         sbitmap_clear_bit(sb, bitnr);
 955         spin_unlock(&ctx->lock);
 956         return true;
 957 }
 958
 959 /*
 960  * Process software queues that have been marked busy, splicing them
 961  * to the for-dispatch
 962  */
 963 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 964 {
 965         struct flush_busy_ctx_data data = {
 966                 .hctx = hctx,
 967                 .list = list,
 968         };
 969
 970         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 971 }
 972 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 973
 974 struct dispatch_rq_data {
 975         struct blk_mq_hw_ctx *hctx;
 976         struct request *rq;
 977 };
 978
 979 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
 980                 void *data)
 981 {
 982         struct dispatch_rq_data *dispatch_data = data;
 983         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
 984         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 985         enum hctx_type type = hctx->type;
 986
 987         spin_lock(&ctx->lock);
 988         if (!list_empty(&ctx->rq_lists[type])) {
 989                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
 990                 list_del_init(&dispatch_data->rq->queuelist);
 991                 if (list_empty(&ctx->rq_lists[type]))
 992                         sbitmap_clear_bit(sb, bitnr);
 993         }
 994         spin_unlock(&ctx->lock);
 995
 996         return !dispatch_data->rq;
 997 }
 998
 999 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1000                                         struct blk_mq_ctx *start)
1001 {
1002         unsigned off = start ? start->index_hw[hctx->type] : 0;
1003         struct dispatch_rq_data data = {
1004                 .hctx = hctx,
1005                 .rq   = NULL,
1006         };
1007
1008         __sbitmap_for_each_set(&hctx->ctx_map, off,
1009                                dispatch_rq_from_ctx, &data);
1010
1011         return data.rq;
1012 }
1013
1014 static inline unsigned int queued_to_index(unsigned int queued)
1015 {
1016         if (!queued)
1017                 return 0;
1018
1019         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1020 }
1021
1022 bool blk_mq_get_driver_tag(struct request *rq)
1023 {
1024         struct blk_mq_alloc_data data = {
1025                 .q = rq->q,
1026                 .hctx = rq->mq_hctx,
1027                 .flags = BLK_MQ_REQ_NOWAIT,
1028                 .cmd_flags = rq->cmd_flags,
1029         };
1030         bool shared;
1031
1032         if (rq->tag != -1)
1033                 return true;
1034
1035         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1036                 data.flags |= BLK_MQ_REQ_RESERVED;
1037
1038         shared = blk_mq_tag_busy(data.hctx);
1039         rq->tag = blk_mq_get_tag(&data);
1040         if (rq->tag >= 0) {
1041                 if (shared) {
1042                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1043                         atomic_inc(&data.hctx->nr_active);
1044                 }
1045                 data.hctx->tags->rqs[rq->tag] = rq;
1046         }
1047
1048         return rq->tag != -1;
1049 }
1050
1051 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1052                                 int flags, void *key)
1053 {
1054         struct blk_mq_hw_ctx *hctx;
1055
1056         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1057
1058         spin_lock(&hctx->dispatch_wait_lock);
1059         if (!list_empty(&wait->entry)) {
1060                 struct sbitmap_queue *sbq;
1061
1062                 list_del_init(&wait->entry);
1063                 sbq = &hctx->tags->bitmap_tags;
1064                 atomic_dec(&sbq->ws_active);
1065         }
1066         spin_unlock(&hctx->dispatch_wait_lock);
1067
1068         blk_mq_run_hw_queue(hctx, true);
1069         return 1;
1070 }
1071
1072 /*
1073  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1074  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1075  * restart. For both cases, take care to check the condition again after
1076  * marking us as waiting.
1077  */
1078 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1079                                  struct request *rq)
1080 {
1081         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1082         struct wait_queue_head *wq;
1083         wait_queue_entry_t *wait;
1084         bool ret;
1085
1086         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1087                 blk_mq_sched_mark_restart_hctx(hctx);
1088
1089                 /*
1090                  * It's possible that a tag was freed in the window between the
1091                  * allocation failure and adding the hardware queue to the wait
1092                  * queue.
1093                  *
1094                  * Don't clear RESTART here, someone else could have set it.
1095                  * At most this will cost an extra queue run.
1096                  */
1097                 return blk_mq_get_driver_tag(rq);
1098         }
1099
1100         wait = &hctx->dispatch_wait;
1101         if (!list_empty_careful(&wait->entry))
1102                 return false;
1103
1104         wq = &bt_wait_ptr(sbq, hctx)->wait;
1105
1106         spin_lock_irq(&wq->lock);
1107         spin_lock(&hctx->dispatch_wait_lock);
1108         if (!list_empty(&wait->entry)) {
1109                 spin_unlock(&hctx->dispatch_wait_lock);
1110                 spin_unlock_irq(&wq->lock);
1111                 return false;
1112         }
1113
1114         atomic_inc(&sbq->ws_active);
1115         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1116         __add_wait_queue(wq, wait);
1117
1118         /*
1119          * It's possible that a tag was freed in the window between the
1120          * allocation failure and adding the hardware queue to the wait
1121          * queue.
1122          */
1123         ret = blk_mq_get_driver_tag(rq);
1124         if (!ret) {
1125                 spin_unlock(&hctx->dispatch_wait_lock);
1126                 spin_unlock_irq(&wq->lock);
1127                 return false;
1128         }
1129
1130         /*
1131          * We got a tag, remove ourselves from the wait queue to ensure
1132          * someone else gets the wakeup.
1133          */
1134         list_del_init(&wait->entry);
1135         atomic_dec(&sbq->ws_active);
1136         spin_unlock(&hctx->dispatch_wait_lock);
1137         spin_unlock_irq(&wq->lock);
1138
1139         return true;
1140 }
1141
1142 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1143 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1144 /*
1145  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1146  * - EWMA is one simple way to compute running average value
1147  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1148  * - take 4 as factor for avoiding to get too small(0) result, and this
1149  *   factor doesn't matter because EWMA decreases exponentially
1150  */
1151 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1152 {
1153         unsigned int ewma;
1154
1155         if (hctx->queue->elevator)
1156                 return;
1157
1158         ewma = hctx->dispatch_busy;
1159
1160         if (!ewma && !busy)
1161                 return;
1162
1163         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1164         if (busy)
1165                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1166         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1167
1168         hctx->dispatch_busy = ewma;
1169 }
1170
1171 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1172
1173 /*
1174  * Returns true if we did some work AND can potentially do more.
1175  */
1176 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1177                              bool got_budget)
1178 {
1179         struct blk_mq_hw_ctx *hctx;
1180         struct request *rq, *nxt;
1181         bool no_tag = false;
1182         int errors, queued;
1183         blk_status_t ret = BLK_STS_OK;
1184
1185         if (list_empty(list))
1186                 return false;
1187
1188         WARN_ON(!list_is_singular(list) && got_budget);
1189
1190         /*
1191          * Now process all the entries, sending them to the driver.
1192          */
1193         errors = queued = 0;
1194         do {
1195                 struct blk_mq_queue_data bd;
1196
1197                 rq = list_first_entry(list, struct request, queuelist);
1198
1199                 hctx = rq->mq_hctx;
1200                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1201                         break;
1202
1203                 if (!blk_mq_get_driver_tag(rq)) {
1204                         /*
1205                          * The initial allocation attempt failed, so we need to
1206                          * rerun the hardware queue when a tag is freed. The
1207                          * waitqueue takes care of that. If the queue is run
1208                          * before we add this entry back on the dispatch list,
1209                          * we'll re-run it below.
1210                          */
1211                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1212                                 blk_mq_put_dispatch_budget(hctx);
1213                                 /*
1214                                  * For non-shared tags, the RESTART check
1215                                  * will suffice.
1216                                  */
1217                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1218                                         no_tag = true;
1219                                 break;
1220                         }
1221                 }
1222
1223                 list_del_init(&rq->queuelist);
1224
1225                 bd.rq = rq;
1226
1227                 /*
1228                  * Flag last if we have no more requests, or if we have more
1229                  * but can't assign a driver tag to it.
1230                  */
1231                 if (list_empty(list))
1232                         bd.last = true;
1233                 else {
1234                         nxt = list_first_entry(list, struct request, queuelist);
1235                         bd.last = !blk_mq_get_driver_tag(nxt);
1236                 }
1237
1238                 ret = q->mq_ops->queue_rq(hctx, &bd);
1239                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1240                         /*
1241                          * If an I/O scheduler has been configured and we got a
1242                          * driver tag for the next request already, free it
1243                          * again.
1244                          */
1245                         if (!list_empty(list)) {
1246                                 nxt = list_first_entry(list, struct request, queuelist);
1247                                 blk_mq_put_driver_tag(nxt);
1248                         }
1249                         list_add(&rq->queuelist, list);
1250                         __blk_mq_requeue_request(rq);
1251                         break;
1252                 }
1253
1254                 if (unlikely(ret != BLK_STS_OK)) {
1255                         errors++;
1256                         blk_mq_end_request(rq, BLK_STS_IOERR);
1257                         continue;
1258                 }
1259
1260                 queued++;
1261         } while (!list_empty(list));
1262
1263         hctx->dispatched[queued_to_index(queued)]++;
1264
1265         /*
1266          * Any items that need requeuing? Stuff them into hctx->dispatch,
1267          * that is where we will continue on next queue run.
1268          */
1269         if (!list_empty(list)) {
1270                 bool needs_restart;
1271
1272                 /*
1273                  * If we didn't flush the entire list, we could have told
1274                  * the driver there was more coming, but that turned out to
1275                  * be a lie.
1276                  */
1277                 if (q->mq_ops->commit_rqs)
1278                         q->mq_ops->commit_rqs(hctx);
1279
1280                 spin_lock(&hctx->lock);
1281                 list_splice_init(list, &hctx->dispatch);
1282                 spin_unlock(&hctx->lock);
1283
1284                 /*
1285                  * If SCHED_RESTART was set by the caller of this function and
1286                  * it is no longer set that means that it was cleared by another
1287                  * thread and hence that a queue rerun is needed.
1288                  *
1289                  * If 'no_tag' is set, that means that we failed getting
1290                  * a driver tag with an I/O scheduler attached. If our dispatch
1291                  * waitqueue is no longer active, ensure that we run the queue
1292                  * AFTER adding our entries back to the list.
1293                  *
1294                  * If no I/O scheduler has been configured it is possible that
1295                  * the hardware queue got stopped and restarted before requests
1296                  * were pushed back onto the dispatch list. Rerun the queue to
1297                  * avoid starvation. Notes:
1298                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1299                  *   been stopped before rerunning a queue.
1300                  * - Some but not all block drivers stop a queue before
1301                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1302                  *   and dm-rq.
1303                  *
1304                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1305                  * bit is set, run queue after a delay to avoid IO stalls
1306                  * that could otherwise occur if the queue is idle.
1307                  */
1308                 needs_restart = blk_mq_sched_needs_restart(hctx);
1309                 if (!needs_restart ||
1310                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1311                         blk_mq_run_hw_queue(hctx, true);
1312                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1313                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1314
1315                 blk_mq_update_dispatch_busy(hctx, true);
1316                 return false;
1317         } else
1318                 blk_mq_update_dispatch_busy(hctx, false);
1319
1320         /*
1321          * If the host/device is unable to accept more work, inform the
1322          * caller of that.
1323          */
1324         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1325                 return false;
1326
1327         return (queued + errors) != 0;
1328 }
1329
1330 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1331 {
1332         int srcu_idx;
1333
1334         /*
1335          * We should be running this queue from one of the CPUs that
1336          * are mapped to it.
1337          *
1338          * There are at least two related races now between setting
1339          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1340          * __blk_mq_run_hw_queue():
1341          *
1342          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1343          *   but later it becomes online, then this warning is harmless
1344          *   at all
1345          *
1346          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1347          *   but later it becomes offline, then the warning can't be
1348          *   triggered, and we depend on blk-mq timeout handler to
1349          *   handle dispatched requests to this hctx
1350          */
1351         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1352                 cpu_online(hctx->next_cpu)) {
1353                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1354                         raw_smp_processor_id(),
1355                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1356                 dump_stack();
1357         }
1358
1359         /*
1360          * We can't run the queue inline with ints disabled. Ensure that
1361          * we catch bad users of this early.
1362          */
1363         WARN_ON_ONCE(in_interrupt());
1364
1365         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1366
1367         hctx_lock(hctx, &srcu_idx);
1368         blk_mq_sched_dispatch_requests(hctx);
1369         hctx_unlock(hctx, srcu_idx);
1370 }
1371
1372 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1373 {
1374         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1375
1376         if (cpu >= nr_cpu_ids)
1377                 cpu = cpumask_first(hctx->cpumask);
1378         return cpu;
1379 }
1380
1381 /*
1382  * It'd be great if the workqueue API had a way to pass
1383  * in a mask and had some smarts for more clever placement.
1384  * For now we just round-robin here, switching for every
1385  * BLK_MQ_CPU_WORK_BATCH queued items.
1386  */
1387 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1388 {
1389         bool tried = false;
1390         int next_cpu = hctx->next_cpu;
1391
1392         if (hctx->queue->nr_hw_queues == 1)
1393                 return WORK_CPU_UNBOUND;
1394
1395         if (--hctx->next_cpu_batch <= 0) {
1396 select_cpu:
1397                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1398                                 cpu_online_mask);
1399                 if (next_cpu >= nr_cpu_ids)
1400                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1401                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1402         }
1403
1404         /*
1405          * Do unbound schedule if we can't find a online CPU for this hctx,
1406          * and it should only happen in the path of handling CPU DEAD.
1407          */
1408         if (!cpu_online(next_cpu)) {
1409                 if (!tried) {
1410                         tried = true;
1411                         goto select_cpu;
1412                 }
1413
1414                 /*
1415                  * Make sure to re-select CPU next time once after CPUs
1416                  * in hctx->cpumask become online again.
1417                  */
1418                 hctx->next_cpu = next_cpu;
1419                 hctx->next_cpu_batch = 1;
1420                 return WORK_CPU_UNBOUND;
1421         }
1422
1423         hctx->next_cpu = next_cpu;
1424         return next_cpu;
1425 }
1426
1427 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1428                                         unsigned long msecs)
1429 {
1430         if (unlikely(blk_mq_hctx_stopped(hctx)))
1431                 return;
1432
1433         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1434                 int cpu = get_cpu();
1435                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1436                         __blk_mq_run_hw_queue(hctx);
1437                         put_cpu();
1438                         return;
1439                 }
1440
1441                 put_cpu();
1442         }
1443
1444         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1445                                     msecs_to_jiffies(msecs));
1446 }
1447
1448 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1449 {
1450         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1451 }
1452 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1453
1454 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1455 {
1456         int srcu_idx;
1457         bool need_run;
1458
1459         /*
1460          * When queue is quiesced, we may be switching io scheduler, or
1461          * updating nr_hw_queues, or other things, and we can't run queue
1462          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1463          *
1464          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1465          * quiesced.
1466          */
1467         hctx_lock(hctx, &srcu_idx);
1468         need_run = !blk_queue_quiesced(hctx->queue) &&
1469                 blk_mq_hctx_has_pending(hctx);
1470         hctx_unlock(hctx, srcu_idx);
1471
1472         if (need_run)
1473                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1474 }
1475 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1476
1477 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1478 {
1479         struct blk_mq_hw_ctx *hctx;
1480         int i;
1481
1482         queue_for_each_hw_ctx(q, hctx, i) {
1483                 if (blk_mq_hctx_stopped(hctx))
1484                         continue;
1485
1486                 blk_mq_run_hw_queue(hctx, async);
1487         }
1488 }
1489 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1490
1491 /**
1492  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1493  * @q: request queue.
1494  *
1495  * The caller is responsible for serializing this function against
1496  * blk_mq_{start,stop}_hw_queue().
1497  */
1498 bool blk_mq_queue_stopped(struct request_queue *q)
1499 {
1500         struct blk_mq_hw_ctx *hctx;
1501         int i;
1502
1503         queue_for_each_hw_ctx(q, hctx, i)
1504                 if (blk_mq_hctx_stopped(hctx))
1505                         return true;
1506
1507         return false;
1508 }
1509 EXPORT_SYMBOL(blk_mq_queue_stopped);
1510
1511 /*
1512  * This function is often used for pausing .queue_rq() by driver when
1513  * there isn't enough resource or some conditions aren't satisfied, and
1514  * BLK_STS_RESOURCE is usually returned.
1515  *
1516  * We do not guarantee that dispatch can be drained or blocked
1517  * after blk_mq_stop_hw_queue() returns. Please use
1518  * blk_mq_quiesce_queue() for that requirement.
1519  */
1520 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1521 {
1522         cancel_delayed_work(&hctx->run_work);
1523
1524         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1525 }
1526 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1527
1528 /*
1529  * This function is often used for pausing .queue_rq() by driver when
1530  * there isn't enough resource or some conditions aren't satisfied, and
1531  * BLK_STS_RESOURCE is usually returned.
1532  *
1533  * We do not guarantee that dispatch can be drained or blocked
1534  * after blk_mq_stop_hw_queues() returns. Please use
1535  * blk_mq_quiesce_queue() for that requirement.
1536  */
1537 void blk_mq_stop_hw_queues(struct request_queue *q)
1538 {
1539         struct blk_mq_hw_ctx *hctx;
1540         int i;
1541
1542         queue_for_each_hw_ctx(q, hctx, i)
1543                 blk_mq_stop_hw_queue(hctx);
1544 }
1545 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1546
1547 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1548 {
1549         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1550
1551         blk_mq_run_hw_queue(hctx, false);
1552 }
1553 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1554
1555 void blk_mq_start_hw_queues(struct request_queue *q)
1556 {
1557         struct blk_mq_hw_ctx *hctx;
1558         int i;
1559
1560         queue_for_each_hw_ctx(q, hctx, i)
1561                 blk_mq_start_hw_queue(hctx);
1562 }
1563 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1564
1565 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1566 {
1567         if (!blk_mq_hctx_stopped(hctx))
1568                 return;
1569
1570         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1571         blk_mq_run_hw_queue(hctx, async);
1572 }
1573 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1574
1575 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1576 {
1577         struct blk_mq_hw_ctx *hctx;
1578         int i;
1579
1580         queue_for_each_hw_ctx(q, hctx, i)
1581                 blk_mq_start_stopped_hw_queue(hctx, async);
1582 }
1583 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1584
1585 static void blk_mq_run_work_fn(struct work_struct *work)
1586 {
1587         struct blk_mq_hw_ctx *hctx;
1588
1589         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1590
1591         /*
1592          * If we are stopped, don't run the queue.
1593          */
1594         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1595                 return;
1596
1597         __blk_mq_run_hw_queue(hctx);
1598 }
1599
1600 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1601                                             struct request *rq,
1602                                             bool at_head)
1603 {
1604         struct blk_mq_ctx *ctx = rq->mq_ctx;
1605         enum hctx_type type = hctx->type;
1606
1607         lockdep_assert_held(&ctx->lock);
1608
1609         trace_block_rq_insert(hctx->queue, rq);
1610
1611         if (at_head)
1612                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1613         else
1614                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1615 }
1616
1617 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1618                              bool at_head)
1619 {
1620         struct blk_mq_ctx *ctx = rq->mq_ctx;
1621
1622         lockdep_assert_held(&ctx->lock);
1623
1624         __blk_mq_insert_req_list(hctx, rq, at_head);
1625         blk_mq_hctx_mark_pending(hctx, ctx);
1626 }
1627
1628 /*
1629  * Should only be used carefully, when the caller knows we want to
1630  * bypass a potential IO scheduler on the target device.
1631  */
1632 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1633 {
1634         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1635
1636         spin_lock(&hctx->lock);
1637         list_add_tail(&rq->queuelist, &hctx->dispatch);
1638         spin_unlock(&hctx->lock);
1639
1640         if (run_queue)
1641                 blk_mq_run_hw_queue(hctx, false);
1642 }
1643
1644 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1645                             struct list_head *list)
1646
1647 {
1648         struct request *rq;
1649         enum hctx_type type = hctx->type;
1650
1651         /*
1652          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1653          * offline now
1654          */
1655         list_for_each_entry(rq, list, queuelist) {
1656                 BUG_ON(rq->mq_ctx != ctx);
1657                 trace_block_rq_insert(hctx->queue, rq);
1658         }
1659
1660         spin_lock(&ctx->lock);
1661         list_splice_tail_init(list, &ctx->rq_lists[type]);
1662         blk_mq_hctx_mark_pending(hctx, ctx);
1663         spin_unlock(&ctx->lock);
1664 }
1665
1666 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1667 {
1668         struct request *rqa = container_of(a, struct request, queuelist);
1669         struct request *rqb = container_of(b, struct request, queuelist);
1670
1671         if (rqa->mq_ctx < rqb->mq_ctx)
1672                 return -1;
1673         else if (rqa->mq_ctx > rqb->mq_ctx)
1674                 return 1;
1675         else if (rqa->mq_hctx < rqb->mq_hctx)
1676                 return -1;
1677         else if (rqa->mq_hctx > rqb->mq_hctx)
1678                 return 1;
1679
1680         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1681 }
1682
1683 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1684 {
1685         struct blk_mq_hw_ctx *this_hctx;
1686         struct blk_mq_ctx *this_ctx;
1687         struct request_queue *this_q;
1688         struct request *rq;
1689         LIST_HEAD(list);
1690         LIST_HEAD(rq_list);
1691         unsigned int depth;
1692
1693         list_splice_init(&plug->mq_list, &list);
1694
1695         if (plug->rq_count > 2 && plug->multiple_queues)
1696                 list_sort(NULL, &list, plug_rq_cmp);
1697
1698         plug->rq_count = 0;
1699
1700         this_q = NULL;
1701         this_hctx = NULL;
1702         this_ctx = NULL;
1703         depth = 0;
1704
1705         while (!list_empty(&list)) {
1706                 rq = list_entry_rq(list.next);
1707                 list_del_init(&rq->queuelist);
1708                 BUG_ON(!rq->q);
1709                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1710                         if (this_hctx) {
1711                                 trace_block_unplug(this_q, depth, !from_schedule);
1712                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1713                                                                 &rq_list,
1714                                                                 from_schedule);
1715                         }
1716
1717                         this_q = rq->q;
1718                         this_ctx = rq->mq_ctx;
1719                         this_hctx = rq->mq_hctx;
1720                         depth = 0;
1721                 }
1722
1723                 depth++;
1724                 list_add_tail(&rq->queuelist, &rq_list);
1725         }
1726
1727         /*
1728          * If 'this_hctx' is set, we know we have entries to complete
1729          * on 'rq_list'. Do those.
1730          */
1731         if (this_hctx) {
1732                 trace_block_unplug(this_q, depth, !from_schedule);
1733                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1734                                                 from_schedule);
1735         }
1736 }
1737
1738 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1739                 unsigned int nr_segs)
1740 {
1741         if (bio->bi_opf & REQ_RAHEAD)
1742                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1743
1744         rq->__sector = bio->bi_iter.bi_sector;
1745         rq->write_hint = bio->bi_write_hint;
1746         blk_rq_bio_prep(rq, bio, nr_segs);
1747
1748         blk_account_io_start(rq, true);
1749 }
1750
1751 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1752                                             struct request *rq,
1753                                             blk_qc_t *cookie, bool last)
1754 {
1755         struct request_queue *q = rq->q;
1756         struct blk_mq_queue_data bd = {
1757                 .rq = rq,
1758                 .last = last,
1759         };
1760         blk_qc_t new_cookie;
1761         blk_status_t ret;
1762
1763         new_cookie = request_to_qc_t(hctx, rq);
1764
1765         /*
1766          * For OK queue, we are done. For error, caller may kill it.
1767          * Any other error (busy), just add it to our list as we
1768          * previously would have done.
1769          */
1770         ret = q->mq_ops->queue_rq(hctx, &bd);
1771         switch (ret) {
1772         case BLK_STS_OK:
1773                 blk_mq_update_dispatch_busy(hctx, false);
1774                 *cookie = new_cookie;
1775                 break;
1776         case BLK_STS_RESOURCE:
1777         case BLK_STS_DEV_RESOURCE:
1778                 blk_mq_update_dispatch_busy(hctx, true);
1779                 __blk_mq_requeue_request(rq);
1780                 break;
1781         default:
1782                 blk_mq_update_dispatch_busy(hctx, false);
1783                 *cookie = BLK_QC_T_NONE;
1784                 break;
1785         }
1786
1787         return ret;
1788 }
1789
1790 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1791                                                 struct request *rq,
1792                                                 blk_qc_t *cookie,
1793                                                 bool bypass_insert, bool last)
1794 {
1795         struct request_queue *q = rq->q;
1796         bool run_queue = true;
1797
1798         /*
1799          * RCU or SRCU read lock is needed before checking quiesced flag.
1800          *
1801          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1802          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1803          * and avoid driver to try to dispatch again.
1804          */
1805         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1806                 run_queue = false;
1807                 bypass_insert = false;
1808                 goto insert;
1809         }
1810
1811         if (q->elevator && !bypass_insert)
1812                 goto insert;
1813
1814         if (!blk_mq_get_dispatch_budget(hctx))
1815                 goto insert;
1816
1817         if (!blk_mq_get_driver_tag(rq)) {
1818                 blk_mq_put_dispatch_budget(hctx);
1819                 goto insert;
1820         }
1821
1822         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1823 insert:
1824         if (bypass_insert)
1825                 return BLK_STS_RESOURCE;
1826
1827         blk_mq_request_bypass_insert(rq, run_queue);
1828         return BLK_STS_OK;
1829 }
1830
1831 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1832                 struct request *rq, blk_qc_t *cookie)
1833 {
1834         blk_status_t ret;
1835         int srcu_idx;
1836
1837         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1838
1839         hctx_lock(hctx, &srcu_idx);
1840
1841         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1842         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1843                 blk_mq_request_bypass_insert(rq, true);
1844         else if (ret != BLK_STS_OK)
1845                 blk_mq_end_request(rq, ret);
1846
1847         hctx_unlock(hctx, srcu_idx);
1848 }
1849
1850 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1851 {
1852         blk_status_t ret;
1853         int srcu_idx;
1854         blk_qc_t unused_cookie;
1855         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1856
1857         hctx_lock(hctx, &srcu_idx);
1858         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1859         hctx_unlock(hctx, srcu_idx);
1860
1861         return ret;
1862 }
1863
1864 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1865                 struct list_head *list)
1866 {
1867         while (!list_empty(list)) {
1868                 blk_status_t ret;
1869                 struct request *rq = list_first_entry(list, struct request,
1870                                 queuelist);
1871
1872                 list_del_init(&rq->queuelist);
1873                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1874                 if (ret != BLK_STS_OK) {
1875                         if (ret == BLK_STS_RESOURCE ||
1876                                         ret == BLK_STS_DEV_RESOURCE) {
1877                                 blk_mq_request_bypass_insert(rq,
1878                                                         list_empty(list));
1879                                 break;
1880                         }
1881                         blk_mq_end_request(rq, ret);
1882                 }
1883         }
1884
1885         /*
1886          * If we didn't flush the entire list, we could have told
1887          * the driver there was more coming, but that turned out to
1888          * be a lie.
1889          */
1890         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1891                 hctx->queue->mq_ops->commit_rqs(hctx);
1892 }
1893
1894 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1895 {
1896         list_add_tail(&rq->queuelist, &plug->mq_list);
1897         plug->rq_count++;
1898         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1899                 struct request *tmp;
1900
1901                 tmp = list_first_entry(&plug->mq_list, struct request,
1902                                                 queuelist);
1903                 if (tmp->q != rq->q)
1904                         plug->multiple_queues = true;
1905         }
1906 }
1907
1908 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1909 {
1910         const int is_sync = op_is_sync(bio->bi_opf);
1911         const int is_flush_fua = op_is_flush(bio->bi_opf);
1912         struct blk_mq_alloc_data data = { .flags = 0};
1913         struct request *rq;
1914         struct blk_plug *plug;
1915         struct request *same_queue_rq = NULL;
1916         unsigned int nr_segs;
1917         blk_qc_t cookie;
1918
1919         blk_queue_bounce(q, &bio);
1920         __blk_queue_split(q, &bio, &nr_segs);
1921
1922         if (!bio_integrity_prep(bio))
1923                 return BLK_QC_T_NONE;
1924
1925         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1926             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1927                 return BLK_QC_T_NONE;
1928
1929         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1930                 return BLK_QC_T_NONE;
1931
1932         rq_qos_throttle(q, bio);
1933
1934         data.cmd_flags = bio->bi_opf;
1935         rq = blk_mq_get_request(q, bio, &data);
1936         if (unlikely(!rq)) {
1937                 rq_qos_cleanup(q, bio);
1938                 if (bio->bi_opf & REQ_NOWAIT)
1939                         bio_wouldblock_error(bio);
1940                 return BLK_QC_T_NONE;
1941         }
1942
1943         trace_block_getrq(q, bio, bio->bi_opf);
1944
1945         rq_qos_track(q, rq, bio);
1946
1947         cookie = request_to_qc_t(data.hctx, rq);
1948
1949         blk_mq_bio_to_request(rq, bio, nr_segs);
1950
1951         plug = blk_mq_plug(q, bio);
1952         if (unlikely(is_flush_fua)) {
1953                 /* bypass scheduler for flush rq */
1954                 blk_insert_flush(rq);
1955                 blk_mq_run_hw_queue(data.hctx, true);
1956         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
1957                                 !blk_queue_nonrot(q))) {
1958                 /*
1959                  * Use plugging if we have a ->commit_rqs() hook as well, as
1960                  * we know the driver uses bd->last in a smart fashion.
1961                  *
1962                  * Use normal plugging if this disk is slow HDD, as sequential
1963                  * IO may benefit a lot from plug merging.
1964                  */
1965                 unsigned int request_count = plug->rq_count;
1966                 struct request *last = NULL;
1967
1968                 if (!request_count)
1969                         trace_block_plug(q);
1970                 else
1971                         last = list_entry_rq(plug->mq_list.prev);
1972
1973                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1974                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1975                         blk_flush_plug_list(plug, false);
1976                         trace_block_plug(q);
1977                 }
1978
1979                 blk_add_rq_to_plug(plug, rq);
1980         } else if (q->elevator) {
1981                 blk_mq_sched_insert_request(rq, false, true, true);
1982         } else if (plug && !blk_queue_nomerges(q)) {
1983                 /*
1984                  * We do limited plugging. If the bio can be merged, do that.
1985                  * Otherwise the existing request in the plug list will be
1986                  * issued. So the plug list will have one request at most
1987                  * The plug list might get flushed before this. If that happens,
1988                  * the plug list is empty, and same_queue_rq is invalid.
1989                  */
1990                 if (list_empty(&plug->mq_list))
1991                         same_queue_rq = NULL;
1992                 if (same_queue_rq) {
1993                         list_del_init(&same_queue_rq->queuelist);
1994                         plug->rq_count--;
1995                 }
1996                 blk_add_rq_to_plug(plug, rq);
1997                 trace_block_plug(q);
1998
1999                 if (same_queue_rq) {
2000                         data.hctx = same_queue_rq->mq_hctx;
2001                         trace_block_unplug(q, 1, true);
2002                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2003                                         &cookie);
2004                 }
2005         } else if ((q->nr_hw_queues > 1 && is_sync) ||
2006                         !data.hctx->dispatch_busy) {
2007                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2008         } else {
2009                 blk_mq_sched_insert_request(rq, false, true, true);
2010         }
2011
2012         return cookie;
2013 }
2014
2015 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2016                      unsigned int hctx_idx)
2017 {
2018         struct page *page;
2019
2020         if (tags->rqs && set->ops->exit_request) {
2021                 int i;
2022
2023                 for (i = 0; i < tags->nr_tags; i++) {
2024                         struct request *rq = tags->static_rqs[i];
2025
2026                         if (!rq)
2027                                 continue;
2028                         set->ops->exit_request(set, rq, hctx_idx);
2029                         tags->static_rqs[i] = NULL;
2030                 }
2031         }
2032
2033         while (!list_empty(&tags->page_list)) {
2034                 page = list_first_entry(&tags->page_list, struct page, lru);
2035                 list_del_init(&page->lru);
2036                 /*
2037                  * Remove kmemleak object previously allocated in
2038                  * blk_mq_alloc_rqs().
2039                  */
2040                 kmemleak_free(page_address(page));
2041                 __free_pages(page, page->private);
2042         }
2043 }
2044
2045 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2046 {
2047         kfree(tags->rqs);
2048         tags->rqs = NULL;
2049         kfree(tags->static_rqs);
2050         tags->static_rqs = NULL;
2051
2052         blk_mq_free_tags(tags);
2053 }
2054
2055 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2056                                         unsigned int hctx_idx,
2057                                         unsigned int nr_tags,
2058                                         unsigned int reserved_tags)
2059 {
2060         struct blk_mq_tags *tags;
2061         int node;
2062
2063         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2064         if (node == NUMA_NO_NODE)
2065                 node = set->numa_node;
2066
2067         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2068                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2069         if (!tags)
2070                 return NULL;
2071
2072         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2073                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2074                                  node);
2075         if (!tags->rqs) {
2076                 blk_mq_free_tags(tags);
2077                 return NULL;
2078         }
2079
2080         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2081                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2082                                         node);
2083         if (!tags->static_rqs) {
2084                 kfree(tags->rqs);
2085                 blk_mq_free_tags(tags);
2086                 return NULL;
2087         }
2088
2089         return tags;
2090 }
2091
2092 static size_t order_to_size(unsigned int order)
2093 {
2094         return (size_t)PAGE_SIZE << order;
2095 }
2096
2097 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2098                                unsigned int hctx_idx, int node)
2099 {
2100         int ret;
2101
2102         if (set->ops->init_request) {
2103                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2104                 if (ret)
2105                         return ret;
2106         }
2107
2108         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2109         return 0;
2110 }
2111
2112 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2113                      unsigned int hctx_idx, unsigned int depth)
2114 {
2115         unsigned int i, j, entries_per_page, max_order = 4;
2116         size_t rq_size, left;
2117         int node;
2118
2119         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2120         if (node == NUMA_NO_NODE)
2121                 node = set->numa_node;
2122
2123         INIT_LIST_HEAD(&tags->page_list);
2124
2125         /*
2126          * rq_size is the size of the request plus driver payload, rounded
2127          * to the cacheline size
2128          */
2129         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2130                                 cache_line_size());
2131         left = rq_size * depth;
2132
2133         for (i = 0; i < depth; ) {
2134                 int this_order = max_order;
2135                 struct page *page;
2136                 int to_do;
2137                 void *p;
2138
2139                 while (this_order && left < order_to_size(this_order - 1))
2140                         this_order--;
2141
2142                 do {
2143                         page = alloc_pages_node(node,
2144                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2145                                 this_order);
2146                         if (page)
2147                                 break;
2148                         if (!this_order--)
2149                                 break;
2150                         if (order_to_size(this_order) < rq_size)
2151                                 break;
2152                 } while (1);
2153
2154                 if (!page)
2155                         goto fail;
2156
2157                 page->private = this_order;
2158                 list_add_tail(&page->lru, &tags->page_list);
2159
2160                 p = page_address(page);
2161                 /*
2162                  * Allow kmemleak to scan these pages as they contain pointers
2163                  * to additional allocations like via ops->init_request().
2164                  */
2165                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2166                 entries_per_page = order_to_size(this_order) / rq_size;
2167                 to_do = min(entries_per_page, depth - i);
2168                 left -= to_do * rq_size;
2169                 for (j = 0; j < to_do; j++) {
2170                         struct request *rq = p;
2171
2172                         tags->static_rqs[i] = rq;
2173                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2174                                 tags->static_rqs[i] = NULL;
2175                                 goto fail;
2176                         }
2177
2178                         p += rq_size;
2179                         i++;
2180                 }
2181         }
2182         return 0;
2183
2184 fail:
2185         blk_mq_free_rqs(set, tags, hctx_idx);
2186         return -ENOMEM;
2187 }
2188
2189 /*
2190  * 'cpu' is going away. splice any existing rq_list entries from this
2191  * software queue to the hw queue dispatch list, and ensure that it
2192  * gets run.
2193  */
2194 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2195 {
2196         struct blk_mq_hw_ctx *hctx;
2197         struct blk_mq_ctx *ctx;
2198         LIST_HEAD(tmp);
2199         enum hctx_type type;
2200
2201         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2202         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2203         type = hctx->type;
2204
2205         spin_lock(&ctx->lock);
2206         if (!list_empty(&ctx->rq_lists[type])) {
2207                 list_splice_init(&ctx->rq_lists[type], &tmp);
2208                 blk_mq_hctx_clear_pending(hctx, ctx);
2209         }
2210         spin_unlock(&ctx->lock);
2211
2212         if (list_empty(&tmp))
2213                 return 0;
2214
2215         spin_lock(&hctx->lock);
2216         list_splice_tail_init(&tmp, &hctx->dispatch);
2217         spin_unlock(&hctx->lock);
2218
2219         blk_mq_run_hw_queue(hctx, true);
2220         return 0;
2221 }
2222
2223 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2224 {
2225         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2226                                             &hctx->cpuhp_dead);
2227 }
2228
2229 /* hctx->ctxs will be freed in queue's release handler */
2230 static void blk_mq_exit_hctx(struct request_queue *q,
2231                 struct blk_mq_tag_set *set,
2232                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2233 {
2234         if (blk_mq_hw_queue_mapped(hctx))
2235                 blk_mq_tag_idle(hctx);
2236
2237         if (set->ops->exit_request)
2238                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2239
2240         if (set->ops->exit_hctx)
2241                 set->ops->exit_hctx(hctx, hctx_idx);
2242
2243         blk_mq_remove_cpuhp(hctx);
2244
2245         spin_lock(&q->unused_hctx_lock);
2246         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2247         spin_unlock(&q->unused_hctx_lock);
2248 }
2249
2250 static void blk_mq_exit_hw_queues(struct request_queue *q,
2251                 struct blk_mq_tag_set *set, int nr_queue)
2252 {
2253         struct blk_mq_hw_ctx *hctx;
2254         unsigned int i;
2255
2256         queue_for_each_hw_ctx(q, hctx, i) {
2257                 if (i == nr_queue)
2258                         break;
2259                 blk_mq_debugfs_unregister_hctx(hctx);
2260                 blk_mq_exit_hctx(q, set, hctx, i);
2261         }
2262 }
2263
2264 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2265 {
2266         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2267
2268         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2269                            __alignof__(struct blk_mq_hw_ctx)) !=
2270                      sizeof(struct blk_mq_hw_ctx));
2271
2272         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2273                 hw_ctx_size += sizeof(struct srcu_struct);
2274
2275         return hw_ctx_size;
2276 }
2277
2278 static int blk_mq_init_hctx(struct request_queue *q,
2279                 struct blk_mq_tag_set *set,
2280                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2281 {
2282         hctx->queue_num = hctx_idx;
2283
2284         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2285
2286         hctx->tags = set->tags[hctx_idx];
2287
2288         if (set->ops->init_hctx &&
2289             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2290                 goto unregister_cpu_notifier;
2291
2292         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2293                                 hctx->numa_node))
2294                 goto exit_hctx;
2295         return 0;
2296
2297  exit_hctx:
2298         if (set->ops->exit_hctx)
2299                 set->ops->exit_hctx(hctx, hctx_idx);
2300  unregister_cpu_notifier:
2301         blk_mq_remove_cpuhp(hctx);
2302         return -1;
2303 }
2304
2305 static struct blk_mq_hw_ctx *
2306 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2307                 int node)
2308 {
2309         struct blk_mq_hw_ctx *hctx;
2310         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2311
2312         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2313         if (!hctx)
2314                 goto fail_alloc_hctx;
2315
2316         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2317                 goto free_hctx;
2318
2319         atomic_set(&hctx->nr_active, 0);
2320         if (node == NUMA_NO_NODE)
2321                 node = set->numa_node;
2322         hctx->numa_node = node;
2323
2324         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2325         spin_lock_init(&hctx->lock);
2326         INIT_LIST_HEAD(&hctx->dispatch);
2327         hctx->queue = q;
2328         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2329
2330         INIT_LIST_HEAD(&hctx->hctx_list);
2331
2332         /*
2333          * Allocate space for all possible cpus to avoid allocation at
2334          * runtime
2335          */
2336         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2337                         gfp, node);
2338         if (!hctx->ctxs)
2339                 goto free_cpumask;
2340
2341         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2342                                 gfp, node))
2343                 goto free_ctxs;
2344         hctx->nr_ctx = 0;
2345
2346         spin_lock_init(&hctx->dispatch_wait_lock);
2347         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2348         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2349
2350         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2351                         gfp);
2352         if (!hctx->fq)
2353                 goto free_bitmap;
2354
2355         if (hctx->flags & BLK_MQ_F_BLOCKING)
2356                 init_srcu_struct(hctx->srcu);
2357         blk_mq_hctx_kobj_init(hctx);
2358
2359         return hctx;
2360
2361  free_bitmap:
2362         sbitmap_free(&hctx->ctx_map);
2363  free_ctxs:
2364         kfree(hctx->ctxs);
2365  free_cpumask:
2366         free_cpumask_var(hctx->cpumask);
2367  free_hctx:
2368         kfree(hctx);
2369  fail_alloc_hctx:
2370         return NULL;
2371 }
2372
2373 static void blk_mq_init_cpu_queues(struct request_queue *q,
2374                                    unsigned int nr_hw_queues)
2375 {
2376         struct blk_mq_tag_set *set = q->tag_set;
2377         unsigned int i, j;
2378
2379         for_each_possible_cpu(i) {
2380                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2381                 struct blk_mq_hw_ctx *hctx;
2382                 int k;
2383
2384                 __ctx->cpu = i;
2385                 spin_lock_init(&__ctx->lock);
2386                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2387                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2388
2389                 __ctx->queue = q;
2390
2391                 /*
2392                  * Set local node, IFF we have more than one hw queue. If
2393                  * not, we remain on the home node of the device
2394                  */
2395                 for (j = 0; j < set->nr_maps; j++) {
2396                         hctx = blk_mq_map_queue_type(q, j, i);
2397                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2398                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2399                 }
2400         }
2401 }
2402
2403 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2404 {
2405         int ret = 0;
2406
2407         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2408                                         set->queue_depth, set->reserved_tags);
2409         if (!set->tags[hctx_idx])
2410                 return false;
2411
2412         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2413                                 set->queue_depth);
2414         if (!ret)
2415                 return true;
2416
2417         blk_mq_free_rq_map(set->tags[hctx_idx]);
2418         set->tags[hctx_idx] = NULL;
2419         return false;
2420 }
2421
2422 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2423                                          unsigned int hctx_idx)
2424 {
2425         if (set->tags && set->tags[hctx_idx]) {
2426                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2427                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2428                 set->tags[hctx_idx] = NULL;
2429         }
2430 }
2431
2432 static void blk_mq_map_swqueue(struct request_queue *q)
2433 {
2434         unsigned int i, j, hctx_idx;
2435         struct blk_mq_hw_ctx *hctx;
2436         struct blk_mq_ctx *ctx;
2437         struct blk_mq_tag_set *set = q->tag_set;
2438
2439         queue_for_each_hw_ctx(q, hctx, i) {
2440                 cpumask_clear(hctx->cpumask);
2441                 hctx->nr_ctx = 0;
2442                 hctx->dispatch_from = NULL;
2443         }
2444
2445         /*
2446          * Map software to hardware queues.
2447          *
2448          * If the cpu isn't present, the cpu is mapped to first hctx.
2449          */
2450         for_each_possible_cpu(i) {
2451                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2452                 /* unmapped hw queue can be remapped after CPU topo changed */
2453                 if (!set->tags[hctx_idx] &&
2454                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2455                         /*
2456                          * If tags initialization fail for some hctx,
2457                          * that hctx won't be brought online.  In this
2458                          * case, remap the current ctx to hctx[0] which
2459                          * is guaranteed to always have tags allocated
2460                          */
2461                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2462                 }
2463
2464                 ctx = per_cpu_ptr(q->queue_ctx, i);
2465                 for (j = 0; j < set->nr_maps; j++) {
2466                         if (!set->map[j].nr_queues) {
2467                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2468                                                 HCTX_TYPE_DEFAULT, i);
2469                                 continue;
2470                         }
2471
2472                         hctx = blk_mq_map_queue_type(q, j, i);
2473                         ctx->hctxs[j] = hctx;
2474                         /*
2475                          * If the CPU is already set in the mask, then we've
2476                          * mapped this one already. This can happen if
2477                          * devices share queues across queue maps.
2478                          */
2479                         if (cpumask_test_cpu(i, hctx->cpumask))
2480                                 continue;
2481
2482                         cpumask_set_cpu(i, hctx->cpumask);
2483                         hctx->type = j;
2484                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2485                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2486
2487                         /*
2488                          * If the nr_ctx type overflows, we have exceeded the
2489                          * amount of sw queues we can support.
2490                          */
2491                         BUG_ON(!hctx->nr_ctx);
2492                 }
2493
2494                 for (; j < HCTX_MAX_TYPES; j++)
2495                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2496                                         HCTX_TYPE_DEFAULT, i);
2497         }
2498
2499         queue_for_each_hw_ctx(q, hctx, i) {
2500                 /*
2501                  * If no software queues are mapped to this hardware queue,
2502                  * disable it and free the request entries.
2503                  */
2504                 if (!hctx->nr_ctx) {
2505                         /* Never unmap queue 0.  We need it as a
2506                          * fallback in case of a new remap fails
2507                          * allocation
2508                          */
2509                         if (i && set->tags[i])
2510                                 blk_mq_free_map_and_requests(set, i);
2511
2512                         hctx->tags = NULL;
2513                         continue;
2514                 }
2515
2516                 hctx->tags = set->tags[i];
2517                 WARN_ON(!hctx->tags);
2518
2519                 /*
2520                  * Set the map size to the number of mapped software queues.
2521                  * This is more accurate and more efficient than looping
2522                  * over all possibly mapped software queues.
2523                  */
2524                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2525
2526                 /*
2527                  * Initialize batch roundrobin counts
2528                  */
2529                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2530                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2531         }
2532 }
2533
2534 /*
2535  * Caller needs to ensure that we're either frozen/quiesced, or that
2536  * the queue isn't live yet.
2537  */
2538 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2539 {
2540         struct blk_mq_hw_ctx *hctx;
2541         int i;
2542
2543         queue_for_each_hw_ctx(q, hctx, i) {
2544                 if (shared)
2545                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2546                 else
2547                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2548         }
2549 }
2550
2551 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2552                                         bool shared)
2553 {
2554         struct request_queue *q;
2555
2556         lockdep_assert_held(&set->tag_list_lock);
2557
2558         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2559                 blk_mq_freeze_queue(q);
2560                 queue_set_hctx_shared(q, shared);
2561                 blk_mq_unfreeze_queue(q);
2562         }
2563 }
2564
2565 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2566 {
2567         struct blk_mq_tag_set *set = q->tag_set;
2568
2569         mutex_lock(&set->tag_list_lock);
2570         list_del_rcu(&q->tag_set_list);
2571         if (list_is_singular(&set->tag_list)) {
2572                 /* just transitioned to unshared */
2573                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2574                 /* update existing queue */
2575                 blk_mq_update_tag_set_depth(set, false);
2576         }
2577         mutex_unlock(&set->tag_list_lock);
2578         INIT_LIST_HEAD(&q->tag_set_list);
2579 }
2580
2581 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2582                                      struct request_queue *q)
2583 {
2584         mutex_lock(&set->tag_list_lock);
2585
2586         /*
2587          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2588          */
2589         if (!list_empty(&set->tag_list) &&
2590             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2591                 set->flags |= BLK_MQ_F_TAG_SHARED;
2592                 /* update existing queue */
2593                 blk_mq_update_tag_set_depth(set, true);
2594         }
2595         if (set->flags & BLK_MQ_F_TAG_SHARED)
2596                 queue_set_hctx_shared(q, true);
2597         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2598
2599         mutex_unlock(&set->tag_list_lock);
2600 }
2601
2602 /* All allocations will be freed in release handler of q->mq_kobj */
2603 static int blk_mq_alloc_ctxs(struct request_queue *q)
2604 {
2605         struct blk_mq_ctxs *ctxs;
2606         int cpu;
2607
2608         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2609         if (!ctxs)
2610                 return -ENOMEM;
2611
2612         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2613         if (!ctxs->queue_ctx)
2614                 goto fail;
2615
2616         for_each_possible_cpu(cpu) {
2617                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2618                 ctx->ctxs = ctxs;
2619         }
2620
2621         q->mq_kobj = &ctxs->kobj;
2622         q->queue_ctx = ctxs->queue_ctx;
2623
2624         return 0;
2625  fail:
2626         kfree(ctxs);
2627         return -ENOMEM;
2628 }
2629
2630 /*
2631  * It is the actual release handler for mq, but we do it from
2632  * request queue's release handler for avoiding use-after-free
2633  * and headache because q->mq_kobj shouldn't have been introduced,
2634  * but we can't group ctx/kctx kobj without it.
2635  */
2636 void blk_mq_release(struct request_queue *q)
2637 {
2638         struct blk_mq_hw_ctx *hctx, *next;
2639         int i;
2640
2641         queue_for_each_hw_ctx(q, hctx, i)
2642                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2643
2644         /* all hctx are in .unused_hctx_list now */
2645         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2646                 list_del_init(&hctx->hctx_list);
2647                 kobject_put(&hctx->kobj);
2648         }
2649
2650         kfree(q->queue_hw_ctx);
2651
2652         /*
2653          * release .mq_kobj and sw queue's kobject now because
2654          * both share lifetime with request queue.
2655          */
2656         blk_mq_sysfs_deinit(q);
2657 }
2658
2659 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2660 {
2661         struct request_queue *uninit_q, *q;
2662
2663         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2664         if (!uninit_q)
2665                 return ERR_PTR(-ENOMEM);
2666
2667         /*
2668          * Initialize the queue without an elevator. device_add_disk() will do
2669          * the initialization.
2670          */
2671         q = blk_mq_init_allocated_queue(set, uninit_q, false);
2672         if (IS_ERR(q))
2673                 blk_cleanup_queue(uninit_q);
2674
2675         return q;
2676 }
2677 EXPORT_SYMBOL(blk_mq_init_queue);
2678
2679 /*
2680  * Helper for setting up a queue with mq ops, given queue depth, and
2681  * the passed in mq ops flags.
2682  */
2683 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2684                                            const struct blk_mq_ops *ops,
2685                                            unsigned int queue_depth,
2686                                            unsigned int set_flags)
2687 {
2688         struct request_queue *q;
2689         int ret;
2690
2691         memset(set, 0, sizeof(*set));
2692         set->ops = ops;
2693         set->nr_hw_queues = 1;
2694         set->nr_maps = 1;
2695         set->queue_depth = queue_depth;
2696         set->numa_node = NUMA_NO_NODE;
2697         set->flags = set_flags;
2698
2699         ret = blk_mq_alloc_tag_set(set);
2700         if (ret)
2701                 return ERR_PTR(ret);
2702
2703         q = blk_mq_init_queue(set);
2704         if (IS_ERR(q)) {
2705                 blk_mq_free_tag_set(set);
2706                 return q;
2707         }
2708
2709         return q;
2710 }
2711 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2712
2713 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2714                 struct blk_mq_tag_set *set, struct request_queue *q,
2715                 int hctx_idx, int node)
2716 {
2717         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2718
2719         /* reuse dead hctx first */
2720         spin_lock(&q->unused_hctx_lock);
2721         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2722                 if (tmp->numa_node == node) {
2723                         hctx = tmp;
2724                         break;
2725                 }
2726         }
2727         if (hctx)
2728                 list_del_init(&hctx->hctx_list);
2729         spin_unlock(&q->unused_hctx_lock);
2730
2731         if (!hctx)
2732                 hctx = blk_mq_alloc_hctx(q, set, node);
2733         if (!hctx)
2734                 goto fail;
2735
2736         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2737                 goto free_hctx;
2738
2739         return hctx;
2740
2741  free_hctx:
2742         kobject_put(&hctx->kobj);
2743  fail:
2744         return NULL;
2745 }
2746
2747 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2748                                                 struct request_queue *q)
2749 {
2750         int i, j, end;
2751         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2752
2753         if (q->nr_hw_queues < set->nr_hw_queues) {
2754                 struct blk_mq_hw_ctx **new_hctxs;
2755
2756                 new_hctxs = kcalloc_node(set->nr_hw_queues,
2757                                        sizeof(*new_hctxs), GFP_KERNEL,
2758                                        set->numa_node);
2759                 if (!new_hctxs)
2760                         return;
2761                 if (hctxs)
2762                         memcpy(new_hctxs, hctxs, q->nr_hw_queues *
2763                                sizeof(*hctxs));
2764                 q->queue_hw_ctx = new_hctxs;
2765                 q->nr_hw_queues = set->nr_hw_queues;
2766                 kfree(hctxs);
2767                 hctxs = new_hctxs;
2768         }
2769
2770         /* protect against switching io scheduler  */
2771         mutex_lock(&q->sysfs_lock);
2772         for (i = 0; i < set->nr_hw_queues; i++) {
2773                 int node;
2774                 struct blk_mq_hw_ctx *hctx;
2775
2776                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2777                 /*
2778                  * If the hw queue has been mapped to another numa node,
2779                  * we need to realloc the hctx. If allocation fails, fallback
2780                  * to use the previous one.
2781                  */
2782                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2783                         continue;
2784
2785                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2786                 if (hctx) {
2787                         if (hctxs[i])
2788                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2789                         hctxs[i] = hctx;
2790                 } else {
2791                         if (hctxs[i])
2792                                 pr_warn("Allocate new hctx on node %d fails,\
2793                                                 fallback to previous one on node %d\n",
2794                                                 node, hctxs[i]->numa_node);
2795                         else
2796                                 break;
2797                 }
2798         }
2799         /*
2800          * Increasing nr_hw_queues fails. Free the newly allocated
2801          * hctxs and keep the previous q->nr_hw_queues.
2802          */
2803         if (i != set->nr_hw_queues) {
2804                 j = q->nr_hw_queues;
2805                 end = i;
2806         } else {
2807                 j = i;
2808                 end = q->nr_hw_queues;
2809                 q->nr_hw_queues = set->nr_hw_queues;
2810         }
2811
2812         for (; j < end; j++) {
2813                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2814
2815                 if (hctx) {
2816                         if (hctx->tags)
2817                                 blk_mq_free_map_and_requests(set, j);
2818                         blk_mq_exit_hctx(q, set, hctx, j);
2819                         hctxs[j] = NULL;
2820                 }
2821         }
2822         mutex_unlock(&q->sysfs_lock);
2823 }
2824
2825 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2826                                                   struct request_queue *q,
2827                                                   bool elevator_init)
2828 {
2829         /* mark the queue as mq asap */
2830         q->mq_ops = set->ops;
2831
2832         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2833                                              blk_mq_poll_stats_bkt,
2834                                              BLK_MQ_POLL_STATS_BKTS, q);
2835         if (!q->poll_cb)
2836                 goto err_exit;
2837
2838         if (blk_mq_alloc_ctxs(q))
2839                 goto err_poll;
2840
2841         /* init q->mq_kobj and sw queues' kobjects */
2842         blk_mq_sysfs_init(q);
2843
2844         INIT_LIST_HEAD(&q->unused_hctx_list);
2845         spin_lock_init(&q->unused_hctx_lock);
2846
2847         blk_mq_realloc_hw_ctxs(set, q);
2848         if (!q->nr_hw_queues)
2849                 goto err_hctxs;
2850
2851         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2852         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2853
2854         q->tag_set = set;
2855
2856         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2857         if (set->nr_maps > HCTX_TYPE_POLL &&
2858             set->map[HCTX_TYPE_POLL].nr_queues)
2859                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2860
2861         q->sg_reserved_size = INT_MAX;
2862
2863         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2864         INIT_LIST_HEAD(&q->requeue_list);
2865         spin_lock_init(&q->requeue_lock);
2866
2867         blk_queue_make_request(q, blk_mq_make_request);
2868
2869         /*
2870          * Do this after blk_queue_make_request() overrides it...
2871          */
2872         q->nr_requests = set->queue_depth;
2873
2874         /*
2875          * Default to classic polling
2876          */
2877         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2878
2879         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2880         blk_mq_add_queue_tag_set(set, q);
2881         blk_mq_map_swqueue(q);
2882
2883         if (elevator_init)
2884                 elevator_init_mq(q);
2885
2886         return q;
2887
2888 err_hctxs:
2889         kfree(q->queue_hw_ctx);
2890         q->nr_hw_queues = 0;
2891         blk_mq_sysfs_deinit(q);
2892 err_poll:
2893         blk_stat_free_callback(q->poll_cb);
2894         q->poll_cb = NULL;
2895 err_exit:
2896         q->mq_ops = NULL;
2897         return ERR_PTR(-ENOMEM);
2898 }
2899 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2900
2901 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2902 void blk_mq_exit_queue(struct request_queue *q)
2903 {
2904         struct blk_mq_tag_set   *set = q->tag_set;
2905
2906         blk_mq_del_queue_tag_set(q);
2907         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2908 }
2909
2910 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2911 {
2912         int i;
2913
2914         for (i = 0; i < set->nr_hw_queues; i++)
2915                 if (!__blk_mq_alloc_rq_map(set, i))
2916                         goto out_unwind;
2917
2918         return 0;
2919
2920 out_unwind:
2921         while (--i >= 0)
2922                 blk_mq_free_rq_map(set->tags[i]);
2923
2924         return -ENOMEM;
2925 }
2926
2927 /*
2928  * Allocate the request maps associated with this tag_set. Note that this
2929  * may reduce the depth asked for, if memory is tight. set->queue_depth
2930  * will be updated to reflect the allocated depth.
2931  */
2932 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2933 {
2934         unsigned int depth;
2935         int err;
2936
2937         depth = set->queue_depth;
2938         do {
2939                 err = __blk_mq_alloc_rq_maps(set);
2940                 if (!err)
2941                         break;
2942
2943                 set->queue_depth >>= 1;
2944                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2945                         err = -ENOMEM;
2946                         break;
2947                 }
2948         } while (set->queue_depth);
2949
2950         if (!set->queue_depth || err) {
2951                 pr_err("blk-mq: failed to allocate request map\n");
2952                 return -ENOMEM;
2953         }
2954
2955         if (depth != set->queue_depth)
2956                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2957                                                 depth, set->queue_depth);
2958
2959         return 0;
2960 }
2961
2962 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2963 {
2964         if (set->ops->map_queues && !is_kdump_kernel()) {
2965                 int i;
2966
2967                 /*
2968                  * transport .map_queues is usually done in the following
2969                  * way:
2970                  *
2971                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2972                  *      mask = get_cpu_mask(queue)
2973                  *      for_each_cpu(cpu, mask)
2974                  *              set->map[x].mq_map[cpu] = queue;
2975                  * }
2976                  *
2977                  * When we need to remap, the table has to be cleared for
2978                  * killing stale mapping since one CPU may not be mapped
2979                  * to any hw queue.
2980                  */
2981                 for (i = 0; i < set->nr_maps; i++)
2982                         blk_mq_clear_mq_map(&set->map[i]);
2983
2984                 return set->ops->map_queues(set);
2985         } else {
2986                 BUG_ON(set->nr_maps > 1);
2987                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
2988         }
2989 }
2990
2991 static int blk_mq_realloc_tag_set_tags(struct blk_mq_tag_set *set,
2992                                   int cur_nr_hw_queues, int new_nr_hw_queues)
2993 {
2994         struct blk_mq_tags **new_tags;
2995
2996         if (cur_nr_hw_queues >= new_nr_hw_queues)
2997                 return 0;
2998
2999         new_tags = kcalloc_node(new_nr_hw_queues, sizeof(struct blk_mq_tags *),
3000                                 GFP_KERNEL, set->numa_node);
3001         if (!new_tags)
3002                 return -ENOMEM;
3003
3004         if (set->tags)
3005                 memcpy(new_tags, set->tags, cur_nr_hw_queues *
3006                        sizeof(*set->tags));
3007         kfree(set->tags);
3008         set->tags = new_tags;
3009         set->nr_hw_queues = new_nr_hw_queues;
3010
3011         return 0;
3012 }
3013
3014 /*
3015  * Alloc a tag set to be associated with one or more request queues.
3016  * May fail with EINVAL for various error conditions. May adjust the
3017  * requested depth down, if it's too large. In that case, the set
3018  * value will be stored in set->queue_depth.
3019  */
3020 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3021 {
3022         int i, ret;
3023
3024         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3025
3026         if (!set->nr_hw_queues)
3027                 return -EINVAL;
3028         if (!set->queue_depth)
3029                 return -EINVAL;
3030         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3031                 return -EINVAL;
3032
3033         if (!set->ops->queue_rq)
3034                 return -EINVAL;
3035
3036         if (!set->ops->get_budget ^ !set->ops->put_budget)
3037                 return -EINVAL;
3038
3039         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3040                 pr_info("blk-mq: reduced tag depth to %u\n",
3041                         BLK_MQ_MAX_DEPTH);
3042                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3043         }
3044
3045         if (!set->nr_maps)
3046                 set->nr_maps = 1;
3047         else if (set->nr_maps > HCTX_MAX_TYPES)
3048                 return -EINVAL;
3049
3050         /*
3051          * If a crashdump is active, then we are potentially in a very
3052          * memory constrained environment. Limit us to 1 queue and
3053          * 64 tags to prevent using too much memory.
3054          */
3055         if (is_kdump_kernel()) {
3056                 set->nr_hw_queues = 1;
3057                 set->nr_maps = 1;
3058                 set->queue_depth = min(64U, set->queue_depth);
3059         }
3060         /*
3061          * There is no use for more h/w queues than cpus if we just have
3062          * a single map
3063          */
3064         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3065                 set->nr_hw_queues = nr_cpu_ids;
3066
3067         if (blk_mq_realloc_tag_set_tags(set, 0, set->nr_hw_queues) < 0)
3068                 return -ENOMEM;
3069
3070         ret = -ENOMEM;
3071         for (i = 0; i < set->nr_maps; i++) {
3072                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3073                                                   sizeof(set->map[i].mq_map[0]),
3074                                                   GFP_KERNEL, set->numa_node);
3075                 if (!set->map[i].mq_map)
3076                         goto out_free_mq_map;
3077                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3078         }
3079
3080         ret = blk_mq_update_queue_map(set);
3081         if (ret)
3082                 goto out_free_mq_map;
3083
3084         ret = blk_mq_alloc_rq_maps(set);
3085         if (ret)
3086                 goto out_free_mq_map;
3087
3088         mutex_init(&set->tag_list_lock);
3089         INIT_LIST_HEAD(&set->tag_list);
3090
3091         return 0;
3092
3093 out_free_mq_map:
3094         for (i = 0; i < set->nr_maps; i++) {
3095                 kfree(set->map[i].mq_map);
3096                 set->map[i].mq_map = NULL;
3097         }
3098         kfree(set->tags);
3099         set->tags = NULL;
3100         return ret;
3101 }
3102 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3103
3104 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3105 {
3106         int i, j;
3107
3108         for (i = 0; i < set->nr_hw_queues; i++)
3109                 blk_mq_free_map_and_requests(set, i);
3110
3111         for (j = 0; j < set->nr_maps; j++) {
3112                 kfree(set->map[j].mq_map);
3113                 set->map[j].mq_map = NULL;
3114         }
3115
3116         kfree(set->tags);
3117         set->tags = NULL;
3118 }
3119 EXPORT_SYMBOL(blk_mq_free_tag_set);
3120
3121 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3122 {
3123         struct blk_mq_tag_set *set = q->tag_set;
3124         struct blk_mq_hw_ctx *hctx;
3125         int i, ret;
3126
3127         if (!set)
3128                 return -EINVAL;
3129
3130         if (q->nr_requests == nr)
3131                 return 0;
3132
3133         blk_mq_freeze_queue(q);
3134         blk_mq_quiesce_queue(q);
3135
3136         ret = 0;
3137         queue_for_each_hw_ctx(q, hctx, i) {
3138                 if (!hctx->tags)
3139                         continue;
3140                 /*
3141                  * If we're using an MQ scheduler, just update the scheduler
3142                  * queue depth. This is similar to what the old code would do.
3143                  */
3144                 if (!hctx->sched_tags) {
3145                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3146                                                         false);
3147                 } else {
3148                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3149                                                         nr, true);
3150                 }
3151                 if (ret)
3152                         break;
3153                 if (q->elevator && q->elevator->type->ops.depth_updated)
3154                         q->elevator->type->ops.depth_updated(hctx);
3155         }
3156
3157         if (!ret)
3158                 q->nr_requests = nr;
3159
3160         blk_mq_unquiesce_queue(q);
3161         blk_mq_unfreeze_queue(q);
3162
3163         return ret;
3164 }
3165
3166 /*
3167  * request_queue and elevator_type pair.
3168  * It is just used by __blk_mq_update_nr_hw_queues to cache
3169  * the elevator_type associated with a request_queue.
3170  */
3171 struct blk_mq_qe_pair {
3172         struct list_head node;
3173         struct request_queue *q;
3174         struct elevator_type *type;
3175 };
3176
3177 /*
3178  * Cache the elevator_type in qe pair list and switch the
3179  * io scheduler to 'none'
3180  */
3181 static bool blk_mq_elv_switch_none(struct list_head *head,
3182                 struct request_queue *q)
3183 {
3184         struct blk_mq_qe_pair *qe;
3185
3186         if (!q->elevator)
3187                 return true;
3188
3189         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3190         if (!qe)
3191                 return false;
3192
3193         INIT_LIST_HEAD(&qe->node);
3194         qe->q = q;
3195         qe->type = q->elevator->type;
3196         list_add(&qe->node, head);
3197
3198         mutex_lock(&q->sysfs_lock);
3199         /*
3200          * After elevator_switch_mq, the previous elevator_queue will be
3201          * released by elevator_release. The reference of the io scheduler
3202          * module get by elevator_get will also be put. So we need to get
3203          * a reference of the io scheduler module here to prevent it to be
3204          * removed.
3205          */
3206         __module_get(qe->type->elevator_owner);
3207         elevator_switch_mq(q, NULL);
3208         mutex_unlock(&q->sysfs_lock);
3209
3210         return true;
3211 }
3212
3213 static void blk_mq_elv_switch_back(struct list_head *head,
3214                 struct request_queue *q)
3215 {
3216         struct blk_mq_qe_pair *qe;
3217         struct elevator_type *t = NULL;
3218
3219         list_for_each_entry(qe, head, node)
3220                 if (qe->q == q) {
3221                         t = qe->type;
3222                         break;
3223                 }
3224
3225         if (!t)
3226                 return;
3227
3228         list_del(&qe->node);
3229         kfree(qe);
3230
3231         mutex_lock(&q->sysfs_lock);
3232         elevator_switch_mq(q, t);
3233         mutex_unlock(&q->sysfs_lock);
3234 }
3235
3236 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3237                                                         int nr_hw_queues)
3238 {
3239         struct request_queue *q;
3240         LIST_HEAD(head);
3241         int prev_nr_hw_queues;
3242
3243         lockdep_assert_held(&set->tag_list_lock);
3244
3245         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3246                 nr_hw_queues = nr_cpu_ids;
3247         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3248                 return;
3249
3250         list_for_each_entry(q, &set->tag_list, tag_set_list)
3251                 blk_mq_freeze_queue(q);
3252         /*
3253          * Switch IO scheduler to 'none', cleaning up the data associated
3254          * with the previous scheduler. We will switch back once we are done
3255          * updating the new sw to hw queue mappings.
3256          */
3257         list_for_each_entry(q, &set->tag_list, tag_set_list)
3258                 if (!blk_mq_elv_switch_none(&head, q))
3259                         goto switch_back;
3260
3261         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3262                 blk_mq_debugfs_unregister_hctxs(q);
3263                 blk_mq_sysfs_unregister(q);
3264         }
3265
3266         if (blk_mq_realloc_tag_set_tags(set, set->nr_hw_queues, nr_hw_queues) <
3267             0)
3268                 goto reregister;
3269
3270         prev_nr_hw_queues = set->nr_hw_queues;
3271         set->nr_hw_queues = nr_hw_queues;
3272         blk_mq_update_queue_map(set);
3273 fallback:
3274         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3275                 blk_mq_realloc_hw_ctxs(set, q);
3276                 if (q->nr_hw_queues != set->nr_hw_queues) {
3277                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3278                                         nr_hw_queues, prev_nr_hw_queues);
3279                         set->nr_hw_queues = prev_nr_hw_queues;
3280                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3281                         goto fallback;
3282                 }
3283                 blk_mq_map_swqueue(q);
3284         }
3285
3286 reregister:
3287         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3288                 blk_mq_sysfs_register(q);
3289                 blk_mq_debugfs_register_hctxs(q);
3290         }
3291
3292 switch_back:
3293         list_for_each_entry(q, &set->tag_list, tag_set_list)
3294                 blk_mq_elv_switch_back(&head, q);
3295
3296         list_for_each_entry(q, &set->tag_list, tag_set_list)
3297                 blk_mq_unfreeze_queue(q);
3298 }
3299
3300 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3301 {
3302         mutex_lock(&set->tag_list_lock);
3303         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3304         mutex_unlock(&set->tag_list_lock);
3305 }
3306 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3307
3308 /* Enable polling stats and return whether they were already enabled. */
3309 static bool blk_poll_stats_enable(struct request_queue *q)
3310 {
3311         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3312             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3313                 return true;
3314         blk_stat_add_callback(q, q->poll_cb);
3315         return false;
3316 }
3317
3318 static void blk_mq_poll_stats_start(struct request_queue *q)
3319 {
3320         /*
3321          * We don't arm the callback if polling stats are not enabled or the
3322          * callback is already active.
3323          */
3324         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3325             blk_stat_is_active(q->poll_cb))
3326                 return;
3327
3328         blk_stat_activate_msecs(q->poll_cb, 100);
3329 }
3330
3331 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3332 {
3333         struct request_queue *q = cb->data;
3334         int bucket;
3335
3336         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3337                 if (cb->stat[bucket].nr_samples)
3338                         q->poll_stat[bucket] = cb->stat[bucket];
3339         }
3340 }
3341
3342 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3343                                        struct blk_mq_hw_ctx *hctx,
3344                                        struct request *rq)
3345 {
3346         unsigned long ret = 0;
3347         int bucket;
3348
3349         /*
3350          * If stats collection isn't on, don't sleep but turn it on for
3351          * future users
3352          */
3353         if (!blk_poll_stats_enable(q))
3354                 return 0;
3355
3356         /*
3357          * As an optimistic guess, use half of the mean service time
3358          * for this type of request. We can (and should) make this smarter.
3359          * For instance, if the completion latencies are tight, we can
3360          * get closer than just half the mean. This is especially
3361          * important on devices where the completion latencies are longer
3362          * than ~10 usec. We do use the stats for the relevant IO size
3363          * if available which does lead to better estimates.
3364          */
3365         bucket = blk_mq_poll_stats_bkt(rq);
3366         if (bucket < 0)
3367                 return ret;
3368
3369         if (q->poll_stat[bucket].nr_samples)
3370                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3371
3372         return ret;
3373 }
3374
3375 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3376                                      struct blk_mq_hw_ctx *hctx,
3377                                      struct request *rq)
3378 {
3379         struct hrtimer_sleeper hs;
3380         enum hrtimer_mode mode;
3381         unsigned int nsecs;
3382         ktime_t kt;
3383
3384         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3385                 return false;
3386
3387         /*
3388          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3389          *
3390          *  0:  use half of prev avg
3391          * >0:  use this specific value
3392          */
3393         if (q->poll_nsec > 0)
3394                 nsecs = q->poll_nsec;
3395         else
3396                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3397
3398         if (!nsecs)
3399                 return false;
3400
3401         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3402
3403         /*
3404          * This will be replaced with the stats tracking code, using
3405          * 'avg_completion_time / 2' as the pre-sleep target.
3406          */
3407         kt = nsecs;
3408
3409         mode = HRTIMER_MODE_REL;
3410         hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3411         hrtimer_set_expires(&hs.timer, kt);
3412
3413         do {
3414                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3415                         break;
3416                 set_current_state(TASK_UNINTERRUPTIBLE);
3417                 hrtimer_sleeper_start_expires(&hs, mode);
3418                 if (hs.task)
3419                         io_schedule();
3420                 hrtimer_cancel(&hs.timer);
3421                 mode = HRTIMER_MODE_ABS;
3422         } while (hs.task && !signal_pending(current));
3423
3424         __set_current_state(TASK_RUNNING);
3425         destroy_hrtimer_on_stack(&hs.timer);
3426         return true;
3427 }
3428
3429 static bool blk_mq_poll_hybrid(struct request_queue *q,
3430                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3431 {
3432         struct request *rq;
3433
3434         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3435                 return false;
3436
3437         if (!blk_qc_t_is_internal(cookie))
3438                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3439         else {
3440                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3441                 /*
3442                  * With scheduling, if the request has completed, we'll
3443                  * get a NULL return here, as we clear the sched tag when
3444                  * that happens. The request still remains valid, like always,
3445                  * so we should be safe with just the NULL check.
3446                  */
3447                 if (!rq)
3448                         return false;
3449         }
3450
3451         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3452 }
3453
3454 /**
3455  * blk_poll - poll for IO completions
3456  * @q:  the queue
3457  * @cookie: cookie passed back at IO submission time
3458  * @spin: whether to spin for completions
3459  *
3460  * Description:
3461  *    Poll for completions on the passed in queue. Returns number of
3462  *    completed entries found. If @spin is true, then blk_poll will continue
3463  *    looping until at least one completion is found, unless the task is
3464  *    otherwise marked running (or we need to reschedule).
3465  */
3466 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3467 {
3468         struct blk_mq_hw_ctx *hctx;
3469         long state;
3470
3471         if (!blk_qc_t_valid(cookie) ||
3472             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3473                 return 0;
3474
3475         if (current->plug)
3476                 blk_flush_plug_list(current->plug, false);
3477
3478         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3479
3480         /*
3481          * If we sleep, have the caller restart the poll loop to reset
3482          * the state. Like for the other success return cases, the
3483          * caller is responsible for checking if the IO completed. If
3484          * the IO isn't complete, we'll get called again and will go
3485          * straight to the busy poll loop.
3486          */
3487         if (blk_mq_poll_hybrid(q, hctx, cookie))
3488                 return 1;
3489
3490         hctx->poll_considered++;
3491
3492         state = current->state;
3493         do {
3494                 int ret;
3495
3496                 hctx->poll_invoked++;
3497
3498                 ret = q->mq_ops->poll(hctx);
3499                 if (ret > 0) {
3500                         hctx->poll_success++;
3501                         __set_current_state(TASK_RUNNING);
3502                         return ret;
3503                 }
3504
3505                 if (signal_pending_state(state, current))
3506                         __set_current_state(TASK_RUNNING);
3507
3508                 if (current->state == TASK_RUNNING)
3509                         return 1;
3510                 if (ret < 0 || !spin)
3511                         break;
3512                 cpu_relax();
3513         } while (!need_resched());
3514
3515         __set_current_state(TASK_RUNNING);
3516         return 0;
3517 }
3518 EXPORT_SYMBOL_GPL(blk_poll);
3519
3520 unsigned int blk_mq_rq_cpu(struct request *rq)
3521 {
3522         return rq->mq_ctx->cpu;
3523 }
3524 EXPORT_SYMBOL(blk_mq_rq_cpu);
3525
3526 static int __init blk_mq_init(void)
3527 {
3528         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3529                                 blk_mq_hctx_notify_dead);
3530         return 0;
3531 }
3532 subsys_initcall(blk_mq_init);