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 "blk.h"
  34 #include "blk-mq.h"
  35 #include "blk-mq-debugfs.h"
  36 #include "blk-mq-tag.h"
  37 #include "blk-pm.h"
  38 #include "blk-stat.h"
  39 #include "blk-mq-sched.h"
  40 #include "blk-rq-qos.h"
  41
  42 static void blk_mq_poll_stats_start(struct request_queue *q);
  43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  44
  45 static int blk_mq_poll_stats_bkt(const struct request *rq)
  46 {
  47         int ddir, bytes, bucket;
  48
  49         ddir = rq_data_dir(rq);
  50         bytes = blk_rq_bytes(rq);
  51
  52         bucket = ddir + 2*(ilog2(bytes) - 9);
  53
  54         if (bucket < 0)
  55                 return -1;
  56         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  57                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  58
  59         return bucket;
  60 }
  61
  62 /*
  63  * Check if any of the ctx, dispatch list or elevator
  64  * have pending work in this hardware queue.
  65  */
  66 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  67 {
  68         return !list_empty_careful(&hctx->dispatch) ||
  69                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  70                         blk_mq_sched_has_work(hctx);
  71 }
  72
  73 /*
  74  * Mark this ctx as having pending work in this hardware queue
  75  */
  76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  77                                      struct blk_mq_ctx *ctx)
  78 {
  79         const int bit = ctx->index_hw[hctx->type];
  80
  81         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
  82                 sbitmap_set_bit(&hctx->ctx_map, bit);
  83 }
  84
  85 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  86                                       struct blk_mq_ctx *ctx)
  87 {
  88         const int bit = ctx->index_hw[hctx->type];
  89
  90         sbitmap_clear_bit(&hctx->ctx_map, bit);
  91 }
  92
  93 struct mq_inflight {
  94         struct hd_struct *part;
  95         unsigned int *inflight;
  96 };
  97
  98 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  99                                   struct request *rq, void *priv,
 100                                   bool reserved)
 101 {
 102         struct mq_inflight *mi = priv;
 103
 104         /*
 105          * index[0] counts the specific partition that was asked for.
 106          */
 107         if (rq->part == mi->part)
 108                 mi->inflight[0]++;
 109
 110         return true;
 111 }
 112
 113 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
 114 {
 115         unsigned inflight[2];
 116         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 117
 118         inflight[0] = inflight[1] = 0;
 119         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 120
 121         return inflight[0];
 122 }
 123
 124 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
 125                                      struct request *rq, void *priv,
 126                                      bool reserved)
 127 {
 128         struct mq_inflight *mi = priv;
 129
 130         if (rq->part == mi->part)
 131                 mi->inflight[rq_data_dir(rq)]++;
 132
 133         return true;
 134 }
 135
 136 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 137                          unsigned int inflight[2])
 138 {
 139         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 140
 141         inflight[0] = inflight[1] = 0;
 142         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
 143 }
 144
 145 void blk_freeze_queue_start(struct request_queue *q)
 146 {
 147         mutex_lock(&q->mq_freeze_lock);
 148         if (++q->mq_freeze_depth == 1) {
 149                 percpu_ref_kill(&q->q_usage_counter);
 150                 mutex_unlock(&q->mq_freeze_lock);
 151                 if (queue_is_mq(q))
 152                         blk_mq_run_hw_queues(q, false);
 153         } else {
 154                 mutex_unlock(&q->mq_freeze_lock);
 155         }
 156 }
 157 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 158
 159 void blk_mq_freeze_queue_wait(struct request_queue *q)
 160 {
 161         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 162 }
 163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 164
 165 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 166                                      unsigned long timeout)
 167 {
 168         return wait_event_timeout(q->mq_freeze_wq,
 169                                         percpu_ref_is_zero(&q->q_usage_counter),
 170                                         timeout);
 171 }
 172 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 173
 174 /*
 175  * Guarantee no request is in use, so we can change any data structure of
 176  * the queue afterward.
 177  */
 178 void blk_freeze_queue(struct request_queue *q)
 179 {
 180         /*
 181          * In the !blk_mq case we are only calling this to kill the
 182          * q_usage_counter, otherwise this increases the freeze depth
 183          * and waits for it to return to zero.  For this reason there is
 184          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 185          * exported to drivers as the only user for unfreeze is blk_mq.
 186          */
 187         blk_freeze_queue_start(q);
 188         blk_mq_freeze_queue_wait(q);
 189 }
 190
 191 void blk_mq_freeze_queue(struct request_queue *q)
 192 {
 193         /*
 194          * ...just an alias to keep freeze and unfreeze actions balanced
 195          * in the blk_mq_* namespace
 196          */
 197         blk_freeze_queue(q);
 198 }
 199 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 200
 201 void blk_mq_unfreeze_queue(struct request_queue *q)
 202 {
 203         mutex_lock(&q->mq_freeze_lock);
 204         q->mq_freeze_depth--;
 205         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 206         if (!q->mq_freeze_depth) {
 207                 percpu_ref_resurrect(&q->q_usage_counter);
 208                 wake_up_all(&q->mq_freeze_wq);
 209         }
 210         mutex_unlock(&q->mq_freeze_lock);
 211 }
 212 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 213
 214 /*
 215  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 216  * mpt3sas driver such that this function can be removed.
 217  */
 218 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 219 {
 220         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 221 }
 222 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 223
 224 /**
 225  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 226  * @q: request queue.
 227  *
 228  * Note: this function does not prevent that the struct request end_io()
 229  * callback function is invoked. Once this function is returned, we make
 230  * sure no dispatch can happen until the queue is unquiesced via
 231  * blk_mq_unquiesce_queue().
 232  */
 233 void blk_mq_quiesce_queue(struct request_queue *q)
 234 {
 235         struct blk_mq_hw_ctx *hctx;
 236         unsigned int i;
 237         bool rcu = false;
 238
 239         blk_mq_quiesce_queue_nowait(q);
 240
 241         queue_for_each_hw_ctx(q, hctx, i) {
 242                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 243                         synchronize_srcu(hctx->srcu);
 244                 else
 245                         rcu = true;
 246         }
 247         if (rcu)
 248                 synchronize_rcu();
 249 }
 250 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 251
 252 /*
 253  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 254  * @q: request queue.
 255  *
 256  * This function recovers queue into the state before quiescing
 257  * which is done by blk_mq_quiesce_queue.
 258  */
 259 void blk_mq_unquiesce_queue(struct request_queue *q)
 260 {
 261         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 262
 263         /* dispatch requests which are inserted during quiescing */
 264         blk_mq_run_hw_queues(q, true);
 265 }
 266 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 267
 268 void blk_mq_wake_waiters(struct request_queue *q)
 269 {
 270         struct blk_mq_hw_ctx *hctx;
 271         unsigned int i;
 272
 273         queue_for_each_hw_ctx(q, hctx, i)
 274                 if (blk_mq_hw_queue_mapped(hctx))
 275                         blk_mq_tag_wakeup_all(hctx->tags, true);
 276 }
 277
 278 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 279 {
 280         return blk_mq_has_free_tags(hctx->tags);
 281 }
 282 EXPORT_SYMBOL(blk_mq_can_queue);
 283
 284 /*
 285  * Only need start/end time stamping if we have stats enabled, or using
 286  * an IO scheduler.
 287  */
 288 static inline bool blk_mq_need_time_stamp(struct request *rq)
 289 {
 290         return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
 291 }
 292
 293 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 294                 unsigned int tag, unsigned int op)
 295 {
 296         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 297         struct request *rq = tags->static_rqs[tag];
 298         req_flags_t rq_flags = 0;
 299
 300         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 301                 rq->tag = -1;
 302                 rq->internal_tag = tag;
 303         } else {
 304                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
 305                         rq_flags = RQF_MQ_INFLIGHT;
 306                         atomic_inc(&data->hctx->nr_active);
 307                 }
 308                 rq->tag = tag;
 309                 rq->internal_tag = -1;
 310                 data->hctx->tags->rqs[rq->tag] = rq;
 311         }
 312
 313         /* csd/requeue_work/fifo_time is initialized before use */
 314         rq->q = data->q;
 315         rq->mq_ctx = data->ctx;
 316         rq->mq_hctx = data->hctx;
 317         rq->rq_flags = rq_flags;
 318         rq->cmd_flags = op;
 319         if (data->flags & BLK_MQ_REQ_PREEMPT)
 320                 rq->rq_flags |= RQF_PREEMPT;
 321         if (blk_queue_io_stat(data->q))
 322                 rq->rq_flags |= RQF_IO_STAT;
 323         INIT_LIST_HEAD(&rq->queuelist);
 324         INIT_HLIST_NODE(&rq->hash);
 325         RB_CLEAR_NODE(&rq->rb_node);
 326         rq->rq_disk = NULL;
 327         rq->part = NULL;
 328         if (blk_mq_need_time_stamp(rq))
 329                 rq->start_time_ns = ktime_get_ns();
 330         else
 331                 rq->start_time_ns = 0;
 332         rq->io_start_time_ns = 0;
 333         rq->nr_phys_segments = 0;
 334 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 335         rq->nr_integrity_segments = 0;
 336 #endif
 337         /* tag was already set */
 338         rq->extra_len = 0;
 339         WRITE_ONCE(rq->deadline, 0);
 340
 341         rq->timeout = 0;
 342
 343         rq->end_io = NULL;
 344         rq->end_io_data = NULL;
 345
 346         data->ctx->rq_dispatched[op_is_sync(op)]++;
 347         refcount_set(&rq->ref, 1);
 348         return rq;
 349 }
 350
 351 static struct request *blk_mq_get_request(struct request_queue *q,
 352                                           struct bio *bio,
 353                                           struct blk_mq_alloc_data *data)
 354 {
 355         struct elevator_queue *e = q->elevator;
 356         struct request *rq;
 357         unsigned int tag;
 358         bool clear_ctx_on_error = false;
 359
 360         blk_queue_enter_live(q);
 361         data->q = q;
 362         if (likely(!data->ctx)) {
 363                 data->ctx = blk_mq_get_ctx(q);
 364                 clear_ctx_on_error = true;
 365         }
 366         if (likely(!data->hctx))
 367                 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
 368                                                 data->ctx);
 369         if (data->cmd_flags & REQ_NOWAIT)
 370                 data->flags |= BLK_MQ_REQ_NOWAIT;
 371
 372         if (e) {
 373                 data->flags |= BLK_MQ_REQ_INTERNAL;
 374
 375                 /*
 376                  * Flush requests are special and go directly to the
 377                  * dispatch list. Don't include reserved tags in the
 378                  * limiting, as it isn't useful.
 379                  */
 380                 if (!op_is_flush(data->cmd_flags) &&
 381                     e->type->ops.limit_depth &&
 382                     !(data->flags & BLK_MQ_REQ_RESERVED))
 383                         e->type->ops.limit_depth(data->cmd_flags, data);
 384         } else {
 385                 blk_mq_tag_busy(data->hctx);
 386         }
 387
 388         tag = blk_mq_get_tag(data);
 389         if (tag == BLK_MQ_TAG_FAIL) {
 390                 if (clear_ctx_on_error)
 391                         data->ctx = NULL;
 392                 blk_queue_exit(q);
 393                 return NULL;
 394         }
 395
 396         rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags);
 397         if (!op_is_flush(data->cmd_flags)) {
 398                 rq->elv.icq = NULL;
 399                 if (e && e->type->ops.prepare_request) {
 400                         if (e->type->icq_cache)
 401                                 blk_mq_sched_assign_ioc(rq);
 402
 403                         e->type->ops.prepare_request(rq, bio);
 404                         rq->rq_flags |= RQF_ELVPRIV;
 405                 }
 406         }
 407         data->hctx->queued++;
 408         return rq;
 409 }
 410
 411 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 412                 blk_mq_req_flags_t flags)
 413 {
 414         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 415         struct request *rq;
 416         int ret;
 417
 418         ret = blk_queue_enter(q, flags);
 419         if (ret)
 420                 return ERR_PTR(ret);
 421
 422         rq = blk_mq_get_request(q, NULL, &alloc_data);
 423         blk_queue_exit(q);
 424
 425         if (!rq)
 426                 return ERR_PTR(-EWOULDBLOCK);
 427
 428         rq->__data_len = 0;
 429         rq->__sector = (sector_t) -1;
 430         rq->bio = rq->biotail = NULL;
 431         return rq;
 432 }
 433 EXPORT_SYMBOL(blk_mq_alloc_request);
 434
 435 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 436         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 437 {
 438         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 439         struct request *rq;
 440         unsigned int cpu;
 441         int ret;
 442
 443         /*
 444          * If the tag allocator sleeps we could get an allocation for a
 445          * different hardware context.  No need to complicate the low level
 446          * allocator for this for the rare use case of a command tied to
 447          * a specific queue.
 448          */
 449         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 450                 return ERR_PTR(-EINVAL);
 451
 452         if (hctx_idx >= q->nr_hw_queues)
 453                 return ERR_PTR(-EIO);
 454
 455         ret = blk_queue_enter(q, flags);
 456         if (ret)
 457                 return ERR_PTR(ret);
 458
 459         /*
 460          * Check if the hardware context is actually mapped to anything.
 461          * If not tell the caller that it should skip this queue.
 462          */
 463         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 464         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 465                 blk_queue_exit(q);
 466                 return ERR_PTR(-EXDEV);
 467         }
 468         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 469         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 470
 471         rq = blk_mq_get_request(q, NULL, &alloc_data);
 472         blk_queue_exit(q);
 473
 474         if (!rq)
 475                 return ERR_PTR(-EWOULDBLOCK);
 476
 477         return rq;
 478 }
 479 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 480
 481 static void __blk_mq_free_request(struct request *rq)
 482 {
 483         struct request_queue *q = rq->q;
 484         struct blk_mq_ctx *ctx = rq->mq_ctx;
 485         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 486         const int sched_tag = rq->internal_tag;
 487
 488         blk_pm_mark_last_busy(rq);
 489         rq->mq_hctx = NULL;
 490         if (rq->tag != -1)
 491                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 492         if (sched_tag != -1)
 493                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 494         blk_mq_sched_restart(hctx);
 495         blk_queue_exit(q);
 496 }
 497
 498 void blk_mq_free_request(struct request *rq)
 499 {
 500         struct request_queue *q = rq->q;
 501         struct elevator_queue *e = q->elevator;
 502         struct blk_mq_ctx *ctx = rq->mq_ctx;
 503         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 504
 505         if (rq->rq_flags & RQF_ELVPRIV) {
 506                 if (e && e->type->ops.finish_request)
 507                         e->type->ops.finish_request(rq);
 508                 if (rq->elv.icq) {
 509                         put_io_context(rq->elv.icq->ioc);
 510                         rq->elv.icq = NULL;
 511                 }
 512         }
 513
 514         ctx->rq_completed[rq_is_sync(rq)]++;
 515         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 516                 atomic_dec(&hctx->nr_active);
 517
 518         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 519                 laptop_io_completion(q->backing_dev_info);
 520
 521         rq_qos_done(q, rq);
 522
 523         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 524         if (refcount_dec_and_test(&rq->ref))
 525                 __blk_mq_free_request(rq);
 526 }
 527 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 528
 529 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 530 {
 531         u64 now = 0;
 532
 533         if (blk_mq_need_time_stamp(rq))
 534                 now = ktime_get_ns();
 535
 536         if (rq->rq_flags & RQF_STATS) {
 537                 blk_mq_poll_stats_start(rq->q);
 538                 blk_stat_add(rq, now);
 539         }
 540
 541         if (rq->internal_tag != -1)
 542                 blk_mq_sched_completed_request(rq, now);
 543
 544         blk_account_io_done(rq, now);
 545
 546         if (rq->end_io) {
 547                 rq_qos_done(rq->q, rq);
 548                 rq->end_io(rq, error);
 549         } else {
 550                 blk_mq_free_request(rq);
 551         }
 552 }
 553 EXPORT_SYMBOL(__blk_mq_end_request);
 554
 555 void blk_mq_end_request(struct request *rq, blk_status_t error)
 556 {
 557         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 558                 BUG();
 559         __blk_mq_end_request(rq, error);
 560 }
 561 EXPORT_SYMBOL(blk_mq_end_request);
 562
 563 static void __blk_mq_complete_request_remote(void *data)
 564 {
 565         struct request *rq = data;
 566         struct request_queue *q = rq->q;
 567
 568         q->mq_ops->complete(rq);
 569 }
 570
 571 static void __blk_mq_complete_request(struct request *rq)
 572 {
 573         struct blk_mq_ctx *ctx = rq->mq_ctx;
 574         struct request_queue *q = rq->q;
 575         bool shared = false;
 576         int cpu;
 577
 578         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 579         /*
 580          * Most of single queue controllers, there is only one irq vector
 581          * for handling IO completion, and the only irq's affinity is set
 582          * as all possible CPUs. On most of ARCHs, this affinity means the
 583          * irq is handled on one specific CPU.
 584          *
 585          * So complete IO reqeust in softirq context in case of single queue
 586          * for not degrading IO performance by irqsoff latency.
 587          */
 588         if (q->nr_hw_queues == 1) {
 589                 __blk_complete_request(rq);
 590                 return;
 591         }
 592
 593         /*
 594          * For a polled request, always complete locallly, it's pointless
 595          * to redirect the completion.
 596          */
 597         if ((rq->cmd_flags & REQ_HIPRI) ||
 598             !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
 599                 q->mq_ops->complete(rq);
 600                 return;
 601         }
 602
 603         cpu = get_cpu();
 604         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
 605                 shared = cpus_share_cache(cpu, ctx->cpu);
 606
 607         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 608                 rq->csd.func = __blk_mq_complete_request_remote;
 609                 rq->csd.info = rq;
 610                 rq->csd.flags = 0;
 611                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 612         } else {
 613                 q->mq_ops->complete(rq);
 614         }
 615         put_cpu();
 616 }
 617
 618 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 619         __releases(hctx->srcu)
 620 {
 621         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 622                 rcu_read_unlock();
 623         else
 624                 srcu_read_unlock(hctx->srcu, srcu_idx);
 625 }
 626
 627 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 628         __acquires(hctx->srcu)
 629 {
 630         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 631                 /* shut up gcc false positive */
 632                 *srcu_idx = 0;
 633                 rcu_read_lock();
 634         } else
 635                 *srcu_idx = srcu_read_lock(hctx->srcu);
 636 }
 637
 638 /**
 639  * blk_mq_complete_request - end I/O on a request
 640  * @rq:         the request being processed
 641  *
 642  * Description:
 643  *      Ends all I/O on a request. It does not handle partial completions.
 644  *      The actual completion happens out-of-order, through a IPI handler.
 645  **/
 646 bool blk_mq_complete_request(struct request *rq)
 647 {
 648         if (unlikely(blk_should_fake_timeout(rq->q)))
 649                 return false;
 650         __blk_mq_complete_request(rq);
 651         return true;
 652 }
 653 EXPORT_SYMBOL(blk_mq_complete_request);
 654
 655 void blk_mq_complete_request_sync(struct request *rq)
 656 {
 657         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 658         rq->q->mq_ops->complete(rq);
 659 }
 660 EXPORT_SYMBOL_GPL(blk_mq_complete_request_sync);
 661
 662 int blk_mq_request_started(struct request *rq)
 663 {
 664         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 665 }
 666 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 667
 668 void blk_mq_start_request(struct request *rq)
 669 {
 670         struct request_queue *q = rq->q;
 671
 672         trace_block_rq_issue(q, rq);
 673
 674         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 675                 rq->io_start_time_ns = ktime_get_ns();
 676 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 677                 rq->throtl_size = blk_rq_sectors(rq);
 678 #endif
 679                 rq->rq_flags |= RQF_STATS;
 680                 rq_qos_issue(q, rq);
 681         }
 682
 683         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 684
 685         blk_add_timer(rq);
 686         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 687
 688         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 689                 /*
 690                  * Make sure space for the drain appears.  We know we can do
 691                  * this because max_hw_segments has been adjusted to be one
 692                  * fewer than the device can handle.
 693                  */
 694                 rq->nr_phys_segments++;
 695         }
 696 }
 697 EXPORT_SYMBOL(blk_mq_start_request);
 698
 699 static void __blk_mq_requeue_request(struct request *rq)
 700 {
 701         struct request_queue *q = rq->q;
 702
 703         blk_mq_put_driver_tag(rq);
 704
 705         trace_block_rq_requeue(q, rq);
 706         rq_qos_requeue(q, rq);
 707
 708         if (blk_mq_request_started(rq)) {
 709                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 710                 rq->rq_flags &= ~RQF_TIMED_OUT;
 711                 if (q->dma_drain_size && blk_rq_bytes(rq))
 712                         rq->nr_phys_segments--;
 713         }
 714 }
 715
 716 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 717 {
 718         __blk_mq_requeue_request(rq);
 719
 720         /* this request will be re-inserted to io scheduler queue */
 721         blk_mq_sched_requeue_request(rq);
 722
 723         BUG_ON(!list_empty(&rq->queuelist));
 724         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 725 }
 726 EXPORT_SYMBOL(blk_mq_requeue_request);
 727
 728 static void blk_mq_requeue_work(struct work_struct *work)
 729 {
 730         struct request_queue *q =
 731                 container_of(work, struct request_queue, requeue_work.work);
 732         LIST_HEAD(rq_list);
 733         struct request *rq, *next;
 734
 735         spin_lock_irq(&q->requeue_lock);
 736         list_splice_init(&q->requeue_list, &rq_list);
 737         spin_unlock_irq(&q->requeue_lock);
 738
 739         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 740                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 741                         continue;
 742
 743                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 744                 list_del_init(&rq->queuelist);
 745                 /*
 746                  * If RQF_DONTPREP, rq has contained some driver specific
 747                  * data, so insert it to hctx dispatch list to avoid any
 748                  * merge.
 749                  */
 750                 if (rq->rq_flags & RQF_DONTPREP)
 751                         blk_mq_request_bypass_insert(rq, false);
 752                 else
 753                         blk_mq_sched_insert_request(rq, true, false, false);
 754         }
 755
 756         while (!list_empty(&rq_list)) {
 757                 rq = list_entry(rq_list.next, struct request, queuelist);
 758                 list_del_init(&rq->queuelist);
 759                 blk_mq_sched_insert_request(rq, false, false, false);
 760         }
 761
 762         blk_mq_run_hw_queues(q, false);
 763 }
 764
 765 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 766                                 bool kick_requeue_list)
 767 {
 768         struct request_queue *q = rq->q;
 769         unsigned long flags;
 770
 771         /*
 772          * We abuse this flag that is otherwise used by the I/O scheduler to
 773          * request head insertion from the workqueue.
 774          */
 775         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 776
 777         spin_lock_irqsave(&q->requeue_lock, flags);
 778         if (at_head) {
 779                 rq->rq_flags |= RQF_SOFTBARRIER;
 780                 list_add(&rq->queuelist, &q->requeue_list);
 781         } else {
 782                 list_add_tail(&rq->queuelist, &q->requeue_list);
 783         }
 784         spin_unlock_irqrestore(&q->requeue_lock, flags);
 785
 786         if (kick_requeue_list)
 787                 blk_mq_kick_requeue_list(q);
 788 }
 789
 790 void blk_mq_kick_requeue_list(struct request_queue *q)
 791 {
 792         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 793 }
 794 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 795
 796 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 797                                     unsigned long msecs)
 798 {
 799         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 800                                     msecs_to_jiffies(msecs));
 801 }
 802 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 803
 804 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 805 {
 806         if (tag < tags->nr_tags) {
 807                 prefetch(tags->rqs[tag]);
 808                 return tags->rqs[tag];
 809         }
 810
 811         return NULL;
 812 }
 813 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 814
 815 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 816                                void *priv, bool reserved)
 817 {
 818         /*
 819          * If we find a request that is inflight and the queue matches,
 820          * we know the queue is busy. Return false to stop the iteration.
 821          */
 822         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 823                 bool *busy = priv;
 824
 825                 *busy = true;
 826                 return false;
 827         }
 828
 829         return true;
 830 }
 831
 832 bool blk_mq_queue_inflight(struct request_queue *q)
 833 {
 834         bool busy = false;
 835
 836         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 837         return busy;
 838 }
 839 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 840
 841 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 842 {
 843         req->rq_flags |= RQF_TIMED_OUT;
 844         if (req->q->mq_ops->timeout) {
 845                 enum blk_eh_timer_return ret;
 846
 847                 ret = req->q->mq_ops->timeout(req, reserved);
 848                 if (ret == BLK_EH_DONE)
 849                         return;
 850                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 851         }
 852
 853         blk_add_timer(req);
 854 }
 855
 856 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 857 {
 858         unsigned long deadline;
 859
 860         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 861                 return false;
 862         if (rq->rq_flags & RQF_TIMED_OUT)
 863                 return false;
 864
 865         deadline = READ_ONCE(rq->deadline);
 866         if (time_after_eq(jiffies, deadline))
 867                 return true;
 868
 869         if (*next == 0)
 870                 *next = deadline;
 871         else if (time_after(*next, deadline))
 872                 *next = deadline;
 873         return false;
 874 }
 875
 876 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 877                 struct request *rq, void *priv, bool reserved)
 878 {
 879         unsigned long *next = priv;
 880
 881         /*
 882          * Just do a quick check if it is expired before locking the request in
 883          * so we're not unnecessarilly synchronizing across CPUs.
 884          */
 885         if (!blk_mq_req_expired(rq, next))
 886                 return true;
 887
 888         /*
 889          * We have reason to believe the request may be expired. Take a
 890          * reference on the request to lock this request lifetime into its
 891          * currently allocated context to prevent it from being reallocated in
 892          * the event the completion by-passes this timeout handler.
 893          *
 894          * If the reference was already released, then the driver beat the
 895          * timeout handler to posting a natural completion.
 896          */
 897         if (!refcount_inc_not_zero(&rq->ref))
 898                 return true;
 899
 900         /*
 901          * The request is now locked and cannot be reallocated underneath the
 902          * timeout handler's processing. Re-verify this exact request is truly
 903          * expired; if it is not expired, then the request was completed and
 904          * reallocated as a new request.
 905          */
 906         if (blk_mq_req_expired(rq, next))
 907                 blk_mq_rq_timed_out(rq, reserved);
 908         if (refcount_dec_and_test(&rq->ref))
 909                 __blk_mq_free_request(rq);
 910
 911         return true;
 912 }
 913
 914 static void blk_mq_timeout_work(struct work_struct *work)
 915 {
 916         struct request_queue *q =
 917                 container_of(work, struct request_queue, timeout_work);
 918         unsigned long next = 0;
 919         struct blk_mq_hw_ctx *hctx;
 920         int i;
 921
 922         /* A deadlock might occur if a request is stuck requiring a
 923          * timeout at the same time a queue freeze is waiting
 924          * completion, since the timeout code would not be able to
 925          * acquire the queue reference here.
 926          *
 927          * That's why we don't use blk_queue_enter here; instead, we use
 928          * percpu_ref_tryget directly, because we need to be able to
 929          * obtain a reference even in the short window between the queue
 930          * starting to freeze, by dropping the first reference in
 931          * blk_freeze_queue_start, and the moment the last request is
 932          * consumed, marked by the instant q_usage_counter reaches
 933          * zero.
 934          */
 935         if (!percpu_ref_tryget(&q->q_usage_counter))
 936                 return;
 937
 938         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 939
 940         if (next != 0) {
 941                 mod_timer(&q->timeout, next);
 942         } else {
 943                 /*
 944                  * Request timeouts are handled as a forward rolling timer. If
 945                  * we end up here it means that no requests are pending and
 946                  * also that no request has been pending for a while. Mark
 947                  * each hctx as idle.
 948                  */
 949                 queue_for_each_hw_ctx(q, hctx, i) {
 950                         /* the hctx may be unmapped, so check it here */
 951                         if (blk_mq_hw_queue_mapped(hctx))
 952                                 blk_mq_tag_idle(hctx);
 953                 }
 954         }
 955         blk_queue_exit(q);
 956 }
 957
 958 struct flush_busy_ctx_data {
 959         struct blk_mq_hw_ctx *hctx;
 960         struct list_head *list;
 961 };
 962
 963 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 964 {
 965         struct flush_busy_ctx_data *flush_data = data;
 966         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 967         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 968         enum hctx_type type = hctx->type;
 969
 970         spin_lock(&ctx->lock);
 971         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 972         sbitmap_clear_bit(sb, bitnr);
 973         spin_unlock(&ctx->lock);
 974         return true;
 975 }
 976
 977 /*
 978  * Process software queues that have been marked busy, splicing them
 979  * to the for-dispatch
 980  */
 981 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 982 {
 983         struct flush_busy_ctx_data data = {
 984                 .hctx = hctx,
 985                 .list = list,
 986         };
 987
 988         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 989 }
 990 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 991
 992 struct dispatch_rq_data {
 993         struct blk_mq_hw_ctx *hctx;
 994         struct request *rq;
 995 };
 996
 997 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
 998                 void *data)
 999 {
1000         struct dispatch_rq_data *dispatch_data = data;
1001         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1002         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1003         enum hctx_type type = hctx->type;
1004
1005         spin_lock(&ctx->lock);
1006         if (!list_empty(&ctx->rq_lists[type])) {
1007                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1008                 list_del_init(&dispatch_data->rq->queuelist);
1009                 if (list_empty(&ctx->rq_lists[type]))
1010                         sbitmap_clear_bit(sb, bitnr);
1011         }
1012         spin_unlock(&ctx->lock);
1013
1014         return !dispatch_data->rq;
1015 }
1016
1017 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1018                                         struct blk_mq_ctx *start)
1019 {
1020         unsigned off = start ? start->index_hw[hctx->type] : 0;
1021         struct dispatch_rq_data data = {
1022                 .hctx = hctx,
1023                 .rq   = NULL,
1024         };
1025
1026         __sbitmap_for_each_set(&hctx->ctx_map, off,
1027                                dispatch_rq_from_ctx, &data);
1028
1029         return data.rq;
1030 }
1031
1032 static inline unsigned int queued_to_index(unsigned int queued)
1033 {
1034         if (!queued)
1035                 return 0;
1036
1037         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1038 }
1039
1040 bool blk_mq_get_driver_tag(struct request *rq)
1041 {
1042         struct blk_mq_alloc_data data = {
1043                 .q = rq->q,
1044                 .hctx = rq->mq_hctx,
1045                 .flags = BLK_MQ_REQ_NOWAIT,
1046                 .cmd_flags = rq->cmd_flags,
1047         };
1048         bool shared;
1049
1050         if (rq->tag != -1)
1051                 goto done;
1052
1053         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1054                 data.flags |= BLK_MQ_REQ_RESERVED;
1055
1056         shared = blk_mq_tag_busy(data.hctx);
1057         rq->tag = blk_mq_get_tag(&data);
1058         if (rq->tag >= 0) {
1059                 if (shared) {
1060                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1061                         atomic_inc(&data.hctx->nr_active);
1062                 }
1063                 data.hctx->tags->rqs[rq->tag] = rq;
1064         }
1065
1066 done:
1067         return rq->tag != -1;
1068 }
1069
1070 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1071                                 int flags, void *key)
1072 {
1073         struct blk_mq_hw_ctx *hctx;
1074
1075         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1076
1077         spin_lock(&hctx->dispatch_wait_lock);
1078         if (!list_empty(&wait->entry)) {
1079                 struct sbitmap_queue *sbq;
1080
1081                 list_del_init(&wait->entry);
1082                 sbq = &hctx->tags->bitmap_tags;
1083                 atomic_dec(&sbq->ws_active);
1084         }
1085         spin_unlock(&hctx->dispatch_wait_lock);
1086
1087         blk_mq_run_hw_queue(hctx, true);
1088         return 1;
1089 }
1090
1091 /*
1092  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1093  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1094  * restart. For both cases, take care to check the condition again after
1095  * marking us as waiting.
1096  */
1097 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1098                                  struct request *rq)
1099 {
1100         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1101         struct wait_queue_head *wq;
1102         wait_queue_entry_t *wait;
1103         bool ret;
1104
1105         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1106                 blk_mq_sched_mark_restart_hctx(hctx);
1107
1108                 /*
1109                  * It's possible that a tag was freed in the window between the
1110                  * allocation failure and adding the hardware queue to the wait
1111                  * queue.
1112                  *
1113                  * Don't clear RESTART here, someone else could have set it.
1114                  * At most this will cost an extra queue run.
1115                  */
1116                 return blk_mq_get_driver_tag(rq);
1117         }
1118
1119         wait = &hctx->dispatch_wait;
1120         if (!list_empty_careful(&wait->entry))
1121                 return false;
1122
1123         wq = &bt_wait_ptr(sbq, hctx)->wait;
1124
1125         spin_lock_irq(&wq->lock);
1126         spin_lock(&hctx->dispatch_wait_lock);
1127         if (!list_empty(&wait->entry)) {
1128                 spin_unlock(&hctx->dispatch_wait_lock);
1129                 spin_unlock_irq(&wq->lock);
1130                 return false;
1131         }
1132
1133         atomic_inc(&sbq->ws_active);
1134         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1135         __add_wait_queue(wq, wait);
1136
1137         /*
1138          * It's possible that a tag was freed in the window between the
1139          * allocation failure and adding the hardware queue to the wait
1140          * queue.
1141          */
1142         ret = blk_mq_get_driver_tag(rq);
1143         if (!ret) {
1144                 spin_unlock(&hctx->dispatch_wait_lock);
1145                 spin_unlock_irq(&wq->lock);
1146                 return false;
1147         }
1148
1149         /*
1150          * We got a tag, remove ourselves from the wait queue to ensure
1151          * someone else gets the wakeup.
1152          */
1153         list_del_init(&wait->entry);
1154         atomic_dec(&sbq->ws_active);
1155         spin_unlock(&hctx->dispatch_wait_lock);
1156         spin_unlock_irq(&wq->lock);
1157
1158         return true;
1159 }
1160
1161 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1162 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1163 /*
1164  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1165  * - EWMA is one simple way to compute running average value
1166  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1167  * - take 4 as factor for avoiding to get too small(0) result, and this
1168  *   factor doesn't matter because EWMA decreases exponentially
1169  */
1170 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1171 {
1172         unsigned int ewma;
1173
1174         if (hctx->queue->elevator)
1175                 return;
1176
1177         ewma = hctx->dispatch_busy;
1178
1179         if (!ewma && !busy)
1180                 return;
1181
1182         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1183         if (busy)
1184                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1185         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1186
1187         hctx->dispatch_busy = ewma;
1188 }
1189
1190 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1191
1192 /*
1193  * Returns true if we did some work AND can potentially do more.
1194  */
1195 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1196                              bool got_budget)
1197 {
1198         struct blk_mq_hw_ctx *hctx;
1199         struct request *rq, *nxt;
1200         bool no_tag = false;
1201         int errors, queued;
1202         blk_status_t ret = BLK_STS_OK;
1203
1204         if (list_empty(list))
1205                 return false;
1206
1207         WARN_ON(!list_is_singular(list) && got_budget);
1208
1209         /*
1210          * Now process all the entries, sending them to the driver.
1211          */
1212         errors = queued = 0;
1213         do {
1214                 struct blk_mq_queue_data bd;
1215
1216                 rq = list_first_entry(list, struct request, queuelist);
1217
1218                 hctx = rq->mq_hctx;
1219                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1220                         break;
1221
1222                 if (!blk_mq_get_driver_tag(rq)) {
1223                         /*
1224                          * The initial allocation attempt failed, so we need to
1225                          * rerun the hardware queue when a tag is freed. The
1226                          * waitqueue takes care of that. If the queue is run
1227                          * before we add this entry back on the dispatch list,
1228                          * we'll re-run it below.
1229                          */
1230                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1231                                 blk_mq_put_dispatch_budget(hctx);
1232                                 /*
1233                                  * For non-shared tags, the RESTART check
1234                                  * will suffice.
1235                                  */
1236                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1237                                         no_tag = true;
1238                                 break;
1239                         }
1240                 }
1241
1242                 list_del_init(&rq->queuelist);
1243
1244                 bd.rq = rq;
1245
1246                 /*
1247                  * Flag last if we have no more requests, or if we have more
1248                  * but can't assign a driver tag to it.
1249                  */
1250                 if (list_empty(list))
1251                         bd.last = true;
1252                 else {
1253                         nxt = list_first_entry(list, struct request, queuelist);
1254                         bd.last = !blk_mq_get_driver_tag(nxt);
1255                 }
1256
1257                 ret = q->mq_ops->queue_rq(hctx, &bd);
1258                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1259                         /*
1260                          * If an I/O scheduler has been configured and we got a
1261                          * driver tag for the next request already, free it
1262                          * again.
1263                          */
1264                         if (!list_empty(list)) {
1265                                 nxt = list_first_entry(list, struct request, queuelist);
1266                                 blk_mq_put_driver_tag(nxt);
1267                         }
1268                         list_add(&rq->queuelist, list);
1269                         __blk_mq_requeue_request(rq);
1270                         break;
1271                 }
1272
1273                 if (unlikely(ret != BLK_STS_OK)) {
1274                         errors++;
1275                         blk_mq_end_request(rq, BLK_STS_IOERR);
1276                         continue;
1277                 }
1278
1279                 queued++;
1280         } while (!list_empty(list));
1281
1282         hctx->dispatched[queued_to_index(queued)]++;
1283
1284         /*
1285          * Any items that need requeuing? Stuff them into hctx->dispatch,
1286          * that is where we will continue on next queue run.
1287          */
1288         if (!list_empty(list)) {
1289                 bool needs_restart;
1290
1291                 /*
1292                  * If we didn't flush the entire list, we could have told
1293                  * the driver there was more coming, but that turned out to
1294                  * be a lie.
1295                  */
1296                 if (q->mq_ops->commit_rqs)
1297                         q->mq_ops->commit_rqs(hctx);
1298
1299                 spin_lock(&hctx->lock);
1300                 list_splice_init(list, &hctx->dispatch);
1301                 spin_unlock(&hctx->lock);
1302
1303                 /*
1304                  * If SCHED_RESTART was set by the caller of this function and
1305                  * it is no longer set that means that it was cleared by another
1306                  * thread and hence that a queue rerun is needed.
1307                  *
1308                  * If 'no_tag' is set, that means that we failed getting
1309                  * a driver tag with an I/O scheduler attached. If our dispatch
1310                  * waitqueue is no longer active, ensure that we run the queue
1311                  * AFTER adding our entries back to the list.
1312                  *
1313                  * If no I/O scheduler has been configured it is possible that
1314                  * the hardware queue got stopped and restarted before requests
1315                  * were pushed back onto the dispatch list. Rerun the queue to
1316                  * avoid starvation. Notes:
1317                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1318                  *   been stopped before rerunning a queue.
1319                  * - Some but not all block drivers stop a queue before
1320                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1321                  *   and dm-rq.
1322                  *
1323                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1324                  * bit is set, run queue after a delay to avoid IO stalls
1325                  * that could otherwise occur if the queue is idle.
1326                  */
1327                 needs_restart = blk_mq_sched_needs_restart(hctx);
1328                 if (!needs_restart ||
1329                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1330                         blk_mq_run_hw_queue(hctx, true);
1331                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1332                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1333
1334                 blk_mq_update_dispatch_busy(hctx, true);
1335                 return false;
1336         } else
1337                 blk_mq_update_dispatch_busy(hctx, false);
1338
1339         /*
1340          * If the host/device is unable to accept more work, inform the
1341          * caller of that.
1342          */
1343         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1344                 return false;
1345
1346         return (queued + errors) != 0;
1347 }
1348
1349 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1350 {
1351         int srcu_idx;
1352
1353         /*
1354          * We should be running this queue from one of the CPUs that
1355          * are mapped to it.
1356          *
1357          * There are at least two related races now between setting
1358          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1359          * __blk_mq_run_hw_queue():
1360          *
1361          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1362          *   but later it becomes online, then this warning is harmless
1363          *   at all
1364          *
1365          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1366          *   but later it becomes offline, then the warning can't be
1367          *   triggered, and we depend on blk-mq timeout handler to
1368          *   handle dispatched requests to this hctx
1369          */
1370         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1371                 cpu_online(hctx->next_cpu)) {
1372                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1373                         raw_smp_processor_id(),
1374                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1375                 dump_stack();
1376         }
1377
1378         /*
1379          * We can't run the queue inline with ints disabled. Ensure that
1380          * we catch bad users of this early.
1381          */
1382         WARN_ON_ONCE(in_interrupt());
1383
1384         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1385
1386         hctx_lock(hctx, &srcu_idx);
1387         blk_mq_sched_dispatch_requests(hctx);
1388         hctx_unlock(hctx, srcu_idx);
1389 }
1390
1391 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1392 {
1393         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1394
1395         if (cpu >= nr_cpu_ids)
1396                 cpu = cpumask_first(hctx->cpumask);
1397         return cpu;
1398 }
1399
1400 /*
1401  * It'd be great if the workqueue API had a way to pass
1402  * in a mask and had some smarts for more clever placement.
1403  * For now we just round-robin here, switching for every
1404  * BLK_MQ_CPU_WORK_BATCH queued items.
1405  */
1406 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1407 {
1408         bool tried = false;
1409         int next_cpu = hctx->next_cpu;
1410
1411         if (hctx->queue->nr_hw_queues == 1)
1412                 return WORK_CPU_UNBOUND;
1413
1414         if (--hctx->next_cpu_batch <= 0) {
1415 select_cpu:
1416                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1417                                 cpu_online_mask);
1418                 if (next_cpu >= nr_cpu_ids)
1419                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1420                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1421         }
1422
1423         /*
1424          * Do unbound schedule if we can't find a online CPU for this hctx,
1425          * and it should only happen in the path of handling CPU DEAD.
1426          */
1427         if (!cpu_online(next_cpu)) {
1428                 if (!tried) {
1429                         tried = true;
1430                         goto select_cpu;
1431                 }
1432
1433                 /*
1434                  * Make sure to re-select CPU next time once after CPUs
1435                  * in hctx->cpumask become online again.
1436                  */
1437                 hctx->next_cpu = next_cpu;
1438                 hctx->next_cpu_batch = 1;
1439                 return WORK_CPU_UNBOUND;
1440         }
1441
1442         hctx->next_cpu = next_cpu;
1443         return next_cpu;
1444 }
1445
1446 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1447                                         unsigned long msecs)
1448 {
1449         if (unlikely(blk_mq_hctx_stopped(hctx)))
1450                 return;
1451
1452         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1453                 int cpu = get_cpu();
1454                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1455                         __blk_mq_run_hw_queue(hctx);
1456                         put_cpu();
1457                         return;
1458                 }
1459
1460                 put_cpu();
1461         }
1462
1463         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1464                                     msecs_to_jiffies(msecs));
1465 }
1466
1467 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1468 {
1469         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1470 }
1471 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1472
1473 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1474 {
1475         int srcu_idx;
1476         bool need_run;
1477
1478         /*
1479          * When queue is quiesced, we may be switching io scheduler, or
1480          * updating nr_hw_queues, or other things, and we can't run queue
1481          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1482          *
1483          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1484          * quiesced.
1485          */
1486         hctx_lock(hctx, &srcu_idx);
1487         need_run = !blk_queue_quiesced(hctx->queue) &&
1488                 blk_mq_hctx_has_pending(hctx);
1489         hctx_unlock(hctx, srcu_idx);
1490
1491         if (need_run) {
1492                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1493                 return true;
1494         }
1495
1496         return false;
1497 }
1498 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1499
1500 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1501 {
1502         struct blk_mq_hw_ctx *hctx;
1503         int i;
1504
1505         queue_for_each_hw_ctx(q, hctx, i) {
1506                 if (blk_mq_hctx_stopped(hctx))
1507                         continue;
1508
1509                 blk_mq_run_hw_queue(hctx, async);
1510         }
1511 }
1512 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1513
1514 /**
1515  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1516  * @q: request queue.
1517  *
1518  * The caller is responsible for serializing this function against
1519  * blk_mq_{start,stop}_hw_queue().
1520  */
1521 bool blk_mq_queue_stopped(struct request_queue *q)
1522 {
1523         struct blk_mq_hw_ctx *hctx;
1524         int i;
1525
1526         queue_for_each_hw_ctx(q, hctx, i)
1527                 if (blk_mq_hctx_stopped(hctx))
1528                         return true;
1529
1530         return false;
1531 }
1532 EXPORT_SYMBOL(blk_mq_queue_stopped);
1533
1534 /*
1535  * This function is often used for pausing .queue_rq() by driver when
1536  * there isn't enough resource or some conditions aren't satisfied, and
1537  * BLK_STS_RESOURCE is usually returned.
1538  *
1539  * We do not guarantee that dispatch can be drained or blocked
1540  * after blk_mq_stop_hw_queue() returns. Please use
1541  * blk_mq_quiesce_queue() for that requirement.
1542  */
1543 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1544 {
1545         cancel_delayed_work(&hctx->run_work);
1546
1547         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1548 }
1549 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1550
1551 /*
1552  * This function is often used for pausing .queue_rq() by driver when
1553  * there isn't enough resource or some conditions aren't satisfied, and
1554  * BLK_STS_RESOURCE is usually returned.
1555  *
1556  * We do not guarantee that dispatch can be drained or blocked
1557  * after blk_mq_stop_hw_queues() returns. Please use
1558  * blk_mq_quiesce_queue() for that requirement.
1559  */
1560 void blk_mq_stop_hw_queues(struct request_queue *q)
1561 {
1562         struct blk_mq_hw_ctx *hctx;
1563         int i;
1564
1565         queue_for_each_hw_ctx(q, hctx, i)
1566                 blk_mq_stop_hw_queue(hctx);
1567 }
1568 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1569
1570 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1571 {
1572         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1573
1574         blk_mq_run_hw_queue(hctx, false);
1575 }
1576 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1577
1578 void blk_mq_start_hw_queues(struct request_queue *q)
1579 {
1580         struct blk_mq_hw_ctx *hctx;
1581         int i;
1582
1583         queue_for_each_hw_ctx(q, hctx, i)
1584                 blk_mq_start_hw_queue(hctx);
1585 }
1586 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1587
1588 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1589 {
1590         if (!blk_mq_hctx_stopped(hctx))
1591                 return;
1592
1593         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1594         blk_mq_run_hw_queue(hctx, async);
1595 }
1596 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1597
1598 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1599 {
1600         struct blk_mq_hw_ctx *hctx;
1601         int i;
1602
1603         queue_for_each_hw_ctx(q, hctx, i)
1604                 blk_mq_start_stopped_hw_queue(hctx, async);
1605 }
1606 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1607
1608 static void blk_mq_run_work_fn(struct work_struct *work)
1609 {
1610         struct blk_mq_hw_ctx *hctx;
1611
1612         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1613
1614         /*
1615          * If we are stopped, don't run the queue.
1616          */
1617         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1618                 return;
1619
1620         __blk_mq_run_hw_queue(hctx);
1621 }
1622
1623 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1624                                             struct request *rq,
1625                                             bool at_head)
1626 {
1627         struct blk_mq_ctx *ctx = rq->mq_ctx;
1628         enum hctx_type type = hctx->type;
1629
1630         lockdep_assert_held(&ctx->lock);
1631
1632         trace_block_rq_insert(hctx->queue, rq);
1633
1634         if (at_head)
1635                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1636         else
1637                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1638 }
1639
1640 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1641                              bool at_head)
1642 {
1643         struct blk_mq_ctx *ctx = rq->mq_ctx;
1644
1645         lockdep_assert_held(&ctx->lock);
1646
1647         __blk_mq_insert_req_list(hctx, rq, at_head);
1648         blk_mq_hctx_mark_pending(hctx, ctx);
1649 }
1650
1651 /*
1652  * Should only be used carefully, when the caller knows we want to
1653  * bypass a potential IO scheduler on the target device.
1654  */
1655 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1656 {
1657         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1658
1659         spin_lock(&hctx->lock);
1660         list_add_tail(&rq->queuelist, &hctx->dispatch);
1661         spin_unlock(&hctx->lock);
1662
1663         if (run_queue)
1664                 blk_mq_run_hw_queue(hctx, false);
1665 }
1666
1667 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1668                             struct list_head *list)
1669
1670 {
1671         struct request *rq;
1672         enum hctx_type type = hctx->type;
1673
1674         /*
1675          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1676          * offline now
1677          */
1678         list_for_each_entry(rq, list, queuelist) {
1679                 BUG_ON(rq->mq_ctx != ctx);
1680                 trace_block_rq_insert(hctx->queue, rq);
1681         }
1682
1683         spin_lock(&ctx->lock);
1684         list_splice_tail_init(list, &ctx->rq_lists[type]);
1685         blk_mq_hctx_mark_pending(hctx, ctx);
1686         spin_unlock(&ctx->lock);
1687 }
1688
1689 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1690 {
1691         struct request *rqa = container_of(a, struct request, queuelist);
1692         struct request *rqb = container_of(b, struct request, queuelist);
1693
1694         if (rqa->mq_ctx < rqb->mq_ctx)
1695                 return -1;
1696         else if (rqa->mq_ctx > rqb->mq_ctx)
1697                 return 1;
1698         else if (rqa->mq_hctx < rqb->mq_hctx)
1699                 return -1;
1700         else if (rqa->mq_hctx > rqb->mq_hctx)
1701                 return 1;
1702
1703         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1704 }
1705
1706 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1707 {
1708         struct blk_mq_hw_ctx *this_hctx;
1709         struct blk_mq_ctx *this_ctx;
1710         struct request_queue *this_q;
1711         struct request *rq;
1712         LIST_HEAD(list);
1713         LIST_HEAD(rq_list);
1714         unsigned int depth;
1715
1716         list_splice_init(&plug->mq_list, &list);
1717
1718         if (plug->rq_count > 2 && plug->multiple_queues)
1719                 list_sort(NULL, &list, plug_rq_cmp);
1720
1721         plug->rq_count = 0;
1722
1723         this_q = NULL;
1724         this_hctx = NULL;
1725         this_ctx = NULL;
1726         depth = 0;
1727
1728         while (!list_empty(&list)) {
1729                 rq = list_entry_rq(list.next);
1730                 list_del_init(&rq->queuelist);
1731                 BUG_ON(!rq->q);
1732                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1733                         if (this_hctx) {
1734                                 trace_block_unplug(this_q, depth, !from_schedule);
1735                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1736                                                                 &rq_list,
1737                                                                 from_schedule);
1738                         }
1739
1740                         this_q = rq->q;
1741                         this_ctx = rq->mq_ctx;
1742                         this_hctx = rq->mq_hctx;
1743                         depth = 0;
1744                 }
1745
1746                 depth++;
1747                 list_add_tail(&rq->queuelist, &rq_list);
1748         }
1749
1750         /*
1751          * If 'this_hctx' is set, we know we have entries to complete
1752          * on 'rq_list'. Do those.
1753          */
1754         if (this_hctx) {
1755                 trace_block_unplug(this_q, depth, !from_schedule);
1756                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1757                                                 from_schedule);
1758         }
1759 }
1760
1761 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1762                 unsigned int nr_segs)
1763 {
1764         if (bio->bi_opf & REQ_RAHEAD)
1765                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1766
1767         rq->__sector = bio->bi_iter.bi_sector;
1768         rq->write_hint = bio->bi_write_hint;
1769         blk_rq_bio_prep(rq, bio, nr_segs);
1770
1771         blk_account_io_start(rq, true);
1772 }
1773
1774 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1775                                             struct request *rq,
1776                                             blk_qc_t *cookie, bool last)
1777 {
1778         struct request_queue *q = rq->q;
1779         struct blk_mq_queue_data bd = {
1780                 .rq = rq,
1781                 .last = last,
1782         };
1783         blk_qc_t new_cookie;
1784         blk_status_t ret;
1785
1786         new_cookie = request_to_qc_t(hctx, rq);
1787
1788         /*
1789          * For OK queue, we are done. For error, caller may kill it.
1790          * Any other error (busy), just add it to our list as we
1791          * previously would have done.
1792          */
1793         ret = q->mq_ops->queue_rq(hctx, &bd);
1794         switch (ret) {
1795         case BLK_STS_OK:
1796                 blk_mq_update_dispatch_busy(hctx, false);
1797                 *cookie = new_cookie;
1798                 break;
1799         case BLK_STS_RESOURCE:
1800         case BLK_STS_DEV_RESOURCE:
1801                 blk_mq_update_dispatch_busy(hctx, true);
1802                 __blk_mq_requeue_request(rq);
1803                 break;
1804         default:
1805                 blk_mq_update_dispatch_busy(hctx, false);
1806                 *cookie = BLK_QC_T_NONE;
1807                 break;
1808         }
1809
1810         return ret;
1811 }
1812
1813 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1814                                                 struct request *rq,
1815                                                 blk_qc_t *cookie,
1816                                                 bool bypass_insert, bool last)
1817 {
1818         struct request_queue *q = rq->q;
1819         bool run_queue = true;
1820
1821         /*
1822          * RCU or SRCU read lock is needed before checking quiesced flag.
1823          *
1824          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1825          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1826          * and avoid driver to try to dispatch again.
1827          */
1828         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1829                 run_queue = false;
1830                 bypass_insert = false;
1831                 goto insert;
1832         }
1833
1834         if (q->elevator && !bypass_insert)
1835                 goto insert;
1836
1837         if (!blk_mq_get_dispatch_budget(hctx))
1838                 goto insert;
1839
1840         if (!blk_mq_get_driver_tag(rq)) {
1841                 blk_mq_put_dispatch_budget(hctx);
1842                 goto insert;
1843         }
1844
1845         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1846 insert:
1847         if (bypass_insert)
1848                 return BLK_STS_RESOURCE;
1849
1850         blk_mq_request_bypass_insert(rq, run_queue);
1851         return BLK_STS_OK;
1852 }
1853
1854 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1855                 struct request *rq, blk_qc_t *cookie)
1856 {
1857         blk_status_t ret;
1858         int srcu_idx;
1859
1860         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1861
1862         hctx_lock(hctx, &srcu_idx);
1863
1864         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1865         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1866                 blk_mq_request_bypass_insert(rq, true);
1867         else if (ret != BLK_STS_OK)
1868                 blk_mq_end_request(rq, ret);
1869
1870         hctx_unlock(hctx, srcu_idx);
1871 }
1872
1873 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1874 {
1875         blk_status_t ret;
1876         int srcu_idx;
1877         blk_qc_t unused_cookie;
1878         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1879
1880         hctx_lock(hctx, &srcu_idx);
1881         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1882         hctx_unlock(hctx, srcu_idx);
1883
1884         return ret;
1885 }
1886
1887 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1888                 struct list_head *list)
1889 {
1890         while (!list_empty(list)) {
1891                 blk_status_t ret;
1892                 struct request *rq = list_first_entry(list, struct request,
1893                                 queuelist);
1894
1895                 list_del_init(&rq->queuelist);
1896                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1897                 if (ret != BLK_STS_OK) {
1898                         if (ret == BLK_STS_RESOURCE ||
1899                                         ret == BLK_STS_DEV_RESOURCE) {
1900                                 blk_mq_request_bypass_insert(rq,
1901                                                         list_empty(list));
1902                                 break;
1903                         }
1904                         blk_mq_end_request(rq, ret);
1905                 }
1906         }
1907
1908         /*
1909          * If we didn't flush the entire list, we could have told
1910          * the driver there was more coming, but that turned out to
1911          * be a lie.
1912          */
1913         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1914                 hctx->queue->mq_ops->commit_rqs(hctx);
1915 }
1916
1917 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1918 {
1919         list_add_tail(&rq->queuelist, &plug->mq_list);
1920         plug->rq_count++;
1921         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1922                 struct request *tmp;
1923
1924                 tmp = list_first_entry(&plug->mq_list, struct request,
1925                                                 queuelist);
1926                 if (tmp->q != rq->q)
1927                         plug->multiple_queues = true;
1928         }
1929 }
1930
1931 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1932 {
1933         const int is_sync = op_is_sync(bio->bi_opf);
1934         const int is_flush_fua = op_is_flush(bio->bi_opf);
1935         struct blk_mq_alloc_data data = { .flags = 0};
1936         struct request *rq;
1937         struct blk_plug *plug;
1938         struct request *same_queue_rq = NULL;
1939         unsigned int nr_segs;
1940         blk_qc_t cookie;
1941
1942         blk_queue_bounce(q, &bio);
1943         __blk_queue_split(q, &bio, &nr_segs);
1944
1945         if (!bio_integrity_prep(bio))
1946                 return BLK_QC_T_NONE;
1947
1948         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1949             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1950                 return BLK_QC_T_NONE;
1951
1952         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1953                 return BLK_QC_T_NONE;
1954
1955         rq_qos_throttle(q, bio);
1956
1957         data.cmd_flags = bio->bi_opf;
1958         rq = blk_mq_get_request(q, bio, &data);
1959         if (unlikely(!rq)) {
1960                 rq_qos_cleanup(q, bio);
1961                 if (bio->bi_opf & REQ_NOWAIT)
1962                         bio_wouldblock_error(bio);
1963                 return BLK_QC_T_NONE;
1964         }
1965
1966         trace_block_getrq(q, bio, bio->bi_opf);
1967
1968         rq_qos_track(q, rq, bio);
1969
1970         cookie = request_to_qc_t(data.hctx, rq);
1971
1972         blk_mq_bio_to_request(rq, bio, nr_segs);
1973
1974         plug = blk_mq_plug(q, bio);
1975         if (unlikely(is_flush_fua)) {
1976                 /* bypass scheduler for flush rq */
1977                 blk_insert_flush(rq);
1978                 blk_mq_run_hw_queue(data.hctx, true);
1979         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1980                 /*
1981                  * Use plugging if we have a ->commit_rqs() hook as well, as
1982                  * we know the driver uses bd->last in a smart fashion.
1983                  */
1984                 unsigned int request_count = plug->rq_count;
1985                 struct request *last = NULL;
1986
1987                 if (!request_count)
1988                         trace_block_plug(q);
1989                 else
1990                         last = list_entry_rq(plug->mq_list.prev);
1991
1992                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1993                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1994                         blk_flush_plug_list(plug, false);
1995                         trace_block_plug(q);
1996                 }
1997
1998                 blk_add_rq_to_plug(plug, rq);
1999         } else if (plug && !blk_queue_nomerges(q)) {
2000                 /*
2001                  * We do limited plugging. If the bio can be merged, do that.
2002                  * Otherwise the existing request in the plug list will be
2003                  * issued. So the plug list will have one request at most
2004                  * The plug list might get flushed before this. If that happens,
2005                  * the plug list is empty, and same_queue_rq is invalid.
2006                  */
2007                 if (list_empty(&plug->mq_list))
2008                         same_queue_rq = NULL;
2009                 if (same_queue_rq) {
2010                         list_del_init(&same_queue_rq->queuelist);
2011                         plug->rq_count--;
2012                 }
2013                 blk_add_rq_to_plug(plug, rq);
2014                 trace_block_plug(q);
2015
2016                 if (same_queue_rq) {
2017                         data.hctx = same_queue_rq->mq_hctx;
2018                         trace_block_unplug(q, 1, true);
2019                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2020                                         &cookie);
2021                 }
2022         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2023                         !data.hctx->dispatch_busy)) {
2024                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2025         } else {
2026                 blk_mq_sched_insert_request(rq, false, true, true);
2027         }
2028
2029         return cookie;
2030 }
2031
2032 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2033                      unsigned int hctx_idx)
2034 {
2035         struct page *page;
2036
2037         if (tags->rqs && set->ops->exit_request) {
2038                 int i;
2039
2040                 for (i = 0; i < tags->nr_tags; i++) {
2041                         struct request *rq = tags->static_rqs[i];
2042
2043                         if (!rq)
2044                                 continue;
2045                         set->ops->exit_request(set, rq, hctx_idx);
2046                         tags->static_rqs[i] = NULL;
2047                 }
2048         }
2049
2050         while (!list_empty(&tags->page_list)) {
2051                 page = list_first_entry(&tags->page_list, struct page, lru);
2052                 list_del_init(&page->lru);
2053                 /*
2054                  * Remove kmemleak object previously allocated in
2055                  * blk_mq_alloc_rqs().
2056                  */
2057                 kmemleak_free(page_address(page));
2058                 __free_pages(page, page->private);
2059         }
2060 }
2061
2062 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2063 {
2064         kfree(tags->rqs);
2065         tags->rqs = NULL;
2066         kfree(tags->static_rqs);
2067         tags->static_rqs = NULL;
2068
2069         blk_mq_free_tags(tags);
2070 }
2071
2072 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2073                                         unsigned int hctx_idx,
2074                                         unsigned int nr_tags,
2075                                         unsigned int reserved_tags)
2076 {
2077         struct blk_mq_tags *tags;
2078         int node;
2079
2080         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2081         if (node == NUMA_NO_NODE)
2082                 node = set->numa_node;
2083
2084         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2085                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2086         if (!tags)
2087                 return NULL;
2088
2089         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2090                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2091                                  node);
2092         if (!tags->rqs) {
2093                 blk_mq_free_tags(tags);
2094                 return NULL;
2095         }
2096
2097         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2098                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2099                                         node);
2100         if (!tags->static_rqs) {
2101                 kfree(tags->rqs);
2102                 blk_mq_free_tags(tags);
2103                 return NULL;
2104         }
2105
2106         return tags;
2107 }
2108
2109 static size_t order_to_size(unsigned int order)
2110 {
2111         return (size_t)PAGE_SIZE << order;
2112 }
2113
2114 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2115                                unsigned int hctx_idx, int node)
2116 {
2117         int ret;
2118
2119         if (set->ops->init_request) {
2120                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2121                 if (ret)
2122                         return ret;
2123         }
2124
2125         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2126         return 0;
2127 }
2128
2129 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2130                      unsigned int hctx_idx, unsigned int depth)
2131 {
2132         unsigned int i, j, entries_per_page, max_order = 4;
2133         size_t rq_size, left;
2134         int node;
2135
2136         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2137         if (node == NUMA_NO_NODE)
2138                 node = set->numa_node;
2139
2140         INIT_LIST_HEAD(&tags->page_list);
2141
2142         /*
2143          * rq_size is the size of the request plus driver payload, rounded
2144          * to the cacheline size
2145          */
2146         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2147                                 cache_line_size());
2148         left = rq_size * depth;
2149
2150         for (i = 0; i < depth; ) {
2151                 int this_order = max_order;
2152                 struct page *page;
2153                 int to_do;
2154                 void *p;
2155
2156                 while (this_order && left < order_to_size(this_order - 1))
2157                         this_order--;
2158
2159                 do {
2160                         page = alloc_pages_node(node,
2161                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2162                                 this_order);
2163                         if (page)
2164                                 break;
2165                         if (!this_order--)
2166                                 break;
2167                         if (order_to_size(this_order) < rq_size)
2168                                 break;
2169                 } while (1);
2170
2171                 if (!page)
2172                         goto fail;
2173
2174                 page->private = this_order;
2175                 list_add_tail(&page->lru, &tags->page_list);
2176
2177                 p = page_address(page);
2178                 /*
2179                  * Allow kmemleak to scan these pages as they contain pointers
2180                  * to additional allocations like via ops->init_request().
2181                  */
2182                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2183                 entries_per_page = order_to_size(this_order) / rq_size;
2184                 to_do = min(entries_per_page, depth - i);
2185                 left -= to_do * rq_size;
2186                 for (j = 0; j < to_do; j++) {
2187                         struct request *rq = p;
2188
2189                         tags->static_rqs[i] = rq;
2190                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2191                                 tags->static_rqs[i] = NULL;
2192                                 goto fail;
2193                         }
2194
2195                         p += rq_size;
2196                         i++;
2197                 }
2198         }
2199         return 0;
2200
2201 fail:
2202         blk_mq_free_rqs(set, tags, hctx_idx);
2203         return -ENOMEM;
2204 }
2205
2206 /*
2207  * 'cpu' is going away. splice any existing rq_list entries from this
2208  * software queue to the hw queue dispatch list, and ensure that it
2209  * gets run.
2210  */
2211 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2212 {
2213         struct blk_mq_hw_ctx *hctx;
2214         struct blk_mq_ctx *ctx;
2215         LIST_HEAD(tmp);
2216         enum hctx_type type;
2217
2218         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2219         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2220         type = hctx->type;
2221
2222         spin_lock(&ctx->lock);
2223         if (!list_empty(&ctx->rq_lists[type])) {
2224                 list_splice_init(&ctx->rq_lists[type], &tmp);
2225                 blk_mq_hctx_clear_pending(hctx, ctx);
2226         }
2227         spin_unlock(&ctx->lock);
2228
2229         if (list_empty(&tmp))
2230                 return 0;
2231
2232         spin_lock(&hctx->lock);
2233         list_splice_tail_init(&tmp, &hctx->dispatch);
2234         spin_unlock(&hctx->lock);
2235
2236         blk_mq_run_hw_queue(hctx, true);
2237         return 0;
2238 }
2239
2240 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2241 {
2242         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2243                                             &hctx->cpuhp_dead);
2244 }
2245
2246 /* hctx->ctxs will be freed in queue's release handler */
2247 static void blk_mq_exit_hctx(struct request_queue *q,
2248                 struct blk_mq_tag_set *set,
2249                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2250 {
2251         if (blk_mq_hw_queue_mapped(hctx))
2252                 blk_mq_tag_idle(hctx);
2253
2254         if (set->ops->exit_request)
2255                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2256
2257         if (set->ops->exit_hctx)
2258                 set->ops->exit_hctx(hctx, hctx_idx);
2259
2260         blk_mq_remove_cpuhp(hctx);
2261
2262         spin_lock(&q->unused_hctx_lock);
2263         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2264         spin_unlock(&q->unused_hctx_lock);
2265 }
2266
2267 static void blk_mq_exit_hw_queues(struct request_queue *q,
2268                 struct blk_mq_tag_set *set, int nr_queue)
2269 {
2270         struct blk_mq_hw_ctx *hctx;
2271         unsigned int i;
2272
2273         queue_for_each_hw_ctx(q, hctx, i) {
2274                 if (i == nr_queue)
2275                         break;
2276                 blk_mq_debugfs_unregister_hctx(hctx);
2277                 blk_mq_exit_hctx(q, set, hctx, i);
2278         }
2279 }
2280
2281 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2282 {
2283         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2284
2285         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2286                            __alignof__(struct blk_mq_hw_ctx)) !=
2287                      sizeof(struct blk_mq_hw_ctx));
2288
2289         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2290                 hw_ctx_size += sizeof(struct srcu_struct);
2291
2292         return hw_ctx_size;
2293 }
2294
2295 static int blk_mq_init_hctx(struct request_queue *q,
2296                 struct blk_mq_tag_set *set,
2297                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2298 {
2299         hctx->queue_num = hctx_idx;
2300
2301         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2302
2303         hctx->tags = set->tags[hctx_idx];
2304
2305         if (set->ops->init_hctx &&
2306             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2307                 goto unregister_cpu_notifier;
2308
2309         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2310                                 hctx->numa_node))
2311                 goto exit_hctx;
2312         return 0;
2313
2314  exit_hctx:
2315         if (set->ops->exit_hctx)
2316                 set->ops->exit_hctx(hctx, hctx_idx);
2317  unregister_cpu_notifier:
2318         blk_mq_remove_cpuhp(hctx);
2319         return -1;
2320 }
2321
2322 static struct blk_mq_hw_ctx *
2323 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2324                 int node)
2325 {
2326         struct blk_mq_hw_ctx *hctx;
2327         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2328
2329         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2330         if (!hctx)
2331                 goto fail_alloc_hctx;
2332
2333         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2334                 goto free_hctx;
2335
2336         atomic_set(&hctx->nr_active, 0);
2337         if (node == NUMA_NO_NODE)
2338                 node = set->numa_node;
2339         hctx->numa_node = node;
2340
2341         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2342         spin_lock_init(&hctx->lock);
2343         INIT_LIST_HEAD(&hctx->dispatch);
2344         hctx->queue = q;
2345         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2346
2347         INIT_LIST_HEAD(&hctx->hctx_list);
2348
2349         /*
2350          * Allocate space for all possible cpus to avoid allocation at
2351          * runtime
2352          */
2353         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2354                         gfp, node);
2355         if (!hctx->ctxs)
2356                 goto free_cpumask;
2357
2358         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2359                                 gfp, node))
2360                 goto free_ctxs;
2361         hctx->nr_ctx = 0;
2362
2363         spin_lock_init(&hctx->dispatch_wait_lock);
2364         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2365         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2366
2367         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2368                         gfp);
2369         if (!hctx->fq)
2370                 goto free_bitmap;
2371
2372         if (hctx->flags & BLK_MQ_F_BLOCKING)
2373                 init_srcu_struct(hctx->srcu);
2374         blk_mq_hctx_kobj_init(hctx);
2375
2376         return hctx;
2377
2378  free_bitmap:
2379         sbitmap_free(&hctx->ctx_map);
2380  free_ctxs:
2381         kfree(hctx->ctxs);
2382  free_cpumask:
2383         free_cpumask_var(hctx->cpumask);
2384  free_hctx:
2385         kfree(hctx);
2386  fail_alloc_hctx:
2387         return NULL;
2388 }
2389
2390 static void blk_mq_init_cpu_queues(struct request_queue *q,
2391                                    unsigned int nr_hw_queues)
2392 {
2393         struct blk_mq_tag_set *set = q->tag_set;
2394         unsigned int i, j;
2395
2396         for_each_possible_cpu(i) {
2397                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2398                 struct blk_mq_hw_ctx *hctx;
2399                 int k;
2400
2401                 __ctx->cpu = i;
2402                 spin_lock_init(&__ctx->lock);
2403                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2404                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2405
2406                 __ctx->queue = q;
2407
2408                 /*
2409                  * Set local node, IFF we have more than one hw queue. If
2410                  * not, we remain on the home node of the device
2411                  */
2412                 for (j = 0; j < set->nr_maps; j++) {
2413                         hctx = blk_mq_map_queue_type(q, j, i);
2414                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2415                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2416                 }
2417         }
2418 }
2419
2420 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2421 {
2422         int ret = 0;
2423
2424         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2425                                         set->queue_depth, set->reserved_tags);
2426         if (!set->tags[hctx_idx])
2427                 return false;
2428
2429         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2430                                 set->queue_depth);
2431         if (!ret)
2432                 return true;
2433
2434         blk_mq_free_rq_map(set->tags[hctx_idx]);
2435         set->tags[hctx_idx] = NULL;
2436         return false;
2437 }
2438
2439 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2440                                          unsigned int hctx_idx)
2441 {
2442         if (set->tags && set->tags[hctx_idx]) {
2443                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2444                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2445                 set->tags[hctx_idx] = NULL;
2446         }
2447 }
2448
2449 static void blk_mq_map_swqueue(struct request_queue *q)
2450 {
2451         unsigned int i, j, hctx_idx;
2452         struct blk_mq_hw_ctx *hctx;
2453         struct blk_mq_ctx *ctx;
2454         struct blk_mq_tag_set *set = q->tag_set;
2455
2456         /*
2457          * Avoid others reading imcomplete hctx->cpumask through sysfs
2458          */
2459         mutex_lock(&q->sysfs_lock);
2460
2461         queue_for_each_hw_ctx(q, hctx, i) {
2462                 cpumask_clear(hctx->cpumask);
2463                 hctx->nr_ctx = 0;
2464                 hctx->dispatch_from = NULL;
2465         }
2466
2467         /*
2468          * Map software to hardware queues.
2469          *
2470          * If the cpu isn't present, the cpu is mapped to first hctx.
2471          */
2472         for_each_possible_cpu(i) {
2473                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2474                 /* unmapped hw queue can be remapped after CPU topo changed */
2475                 if (!set->tags[hctx_idx] &&
2476                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2477                         /*
2478                          * If tags initialization fail for some hctx,
2479                          * that hctx won't be brought online.  In this
2480                          * case, remap the current ctx to hctx[0] which
2481                          * is guaranteed to always have tags allocated
2482                          */
2483                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2484                 }
2485
2486                 ctx = per_cpu_ptr(q->queue_ctx, i);
2487                 for (j = 0; j < set->nr_maps; j++) {
2488                         if (!set->map[j].nr_queues) {
2489                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2490                                                 HCTX_TYPE_DEFAULT, i);
2491                                 continue;
2492                         }
2493
2494                         hctx = blk_mq_map_queue_type(q, j, i);
2495                         ctx->hctxs[j] = hctx;
2496                         /*
2497                          * If the CPU is already set in the mask, then we've
2498                          * mapped this one already. This can happen if
2499                          * devices share queues across queue maps.
2500                          */
2501                         if (cpumask_test_cpu(i, hctx->cpumask))
2502                                 continue;
2503
2504                         cpumask_set_cpu(i, hctx->cpumask);
2505                         hctx->type = j;
2506                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2507                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2508
2509                         /*
2510                          * If the nr_ctx type overflows, we have exceeded the
2511                          * amount of sw queues we can support.
2512                          */
2513                         BUG_ON(!hctx->nr_ctx);
2514                 }
2515
2516                 for (; j < HCTX_MAX_TYPES; j++)
2517                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2518                                         HCTX_TYPE_DEFAULT, i);
2519         }
2520
2521         mutex_unlock(&q->sysfs_lock);
2522
2523         queue_for_each_hw_ctx(q, hctx, i) {
2524                 /*
2525                  * If no software queues are mapped to this hardware queue,
2526                  * disable it and free the request entries.
2527                  */
2528                 if (!hctx->nr_ctx) {
2529                         /* Never unmap queue 0.  We need it as a
2530                          * fallback in case of a new remap fails
2531                          * allocation
2532                          */
2533                         if (i && set->tags[i])
2534                                 blk_mq_free_map_and_requests(set, i);
2535
2536                         hctx->tags = NULL;
2537                         continue;
2538                 }
2539
2540                 hctx->tags = set->tags[i];
2541                 WARN_ON(!hctx->tags);
2542
2543                 /*
2544                  * Set the map size to the number of mapped software queues.
2545                  * This is more accurate and more efficient than looping
2546                  * over all possibly mapped software queues.
2547                  */
2548                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2549
2550                 /*
2551                  * Initialize batch roundrobin counts
2552                  */
2553                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2554                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2555         }
2556 }
2557
2558 /*
2559  * Caller needs to ensure that we're either frozen/quiesced, or that
2560  * the queue isn't live yet.
2561  */
2562 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2563 {
2564         struct blk_mq_hw_ctx *hctx;
2565         int i;
2566
2567         queue_for_each_hw_ctx(q, hctx, i) {
2568                 if (shared)
2569                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2570                 else
2571                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2572         }
2573 }
2574
2575 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2576                                         bool shared)
2577 {
2578         struct request_queue *q;
2579
2580         lockdep_assert_held(&set->tag_list_lock);
2581
2582         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2583                 blk_mq_freeze_queue(q);
2584                 queue_set_hctx_shared(q, shared);
2585                 blk_mq_unfreeze_queue(q);
2586         }
2587 }
2588
2589 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2590 {
2591         struct blk_mq_tag_set *set = q->tag_set;
2592
2593         mutex_lock(&set->tag_list_lock);
2594         list_del_rcu(&q->tag_set_list);
2595         if (list_is_singular(&set->tag_list)) {
2596                 /* just transitioned to unshared */
2597                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2598                 /* update existing queue */
2599                 blk_mq_update_tag_set_depth(set, false);
2600         }
2601         mutex_unlock(&set->tag_list_lock);
2602         INIT_LIST_HEAD(&q->tag_set_list);
2603 }
2604
2605 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2606                                      struct request_queue *q)
2607 {
2608         mutex_lock(&set->tag_list_lock);
2609
2610         /*
2611          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2612          */
2613         if (!list_empty(&set->tag_list) &&
2614             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2615                 set->flags |= BLK_MQ_F_TAG_SHARED;
2616                 /* update existing queue */
2617                 blk_mq_update_tag_set_depth(set, true);
2618         }
2619         if (set->flags & BLK_MQ_F_TAG_SHARED)
2620                 queue_set_hctx_shared(q, true);
2621         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2622
2623         mutex_unlock(&set->tag_list_lock);
2624 }
2625
2626 /* All allocations will be freed in release handler of q->mq_kobj */
2627 static int blk_mq_alloc_ctxs(struct request_queue *q)
2628 {
2629         struct blk_mq_ctxs *ctxs;
2630         int cpu;
2631
2632         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2633         if (!ctxs)
2634                 return -ENOMEM;
2635
2636         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2637         if (!ctxs->queue_ctx)
2638                 goto fail;
2639
2640         for_each_possible_cpu(cpu) {
2641                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2642                 ctx->ctxs = ctxs;
2643         }
2644
2645         q->mq_kobj = &ctxs->kobj;
2646         q->queue_ctx = ctxs->queue_ctx;
2647
2648         return 0;
2649  fail:
2650         kfree(ctxs);
2651         return -ENOMEM;
2652 }
2653
2654 /*
2655  * It is the actual release handler for mq, but we do it from
2656  * request queue's release handler for avoiding use-after-free
2657  * and headache because q->mq_kobj shouldn't have been introduced,
2658  * but we can't group ctx/kctx kobj without it.
2659  */
2660 void blk_mq_release(struct request_queue *q)
2661 {
2662         struct blk_mq_hw_ctx *hctx, *next;
2663         int i;
2664
2665         queue_for_each_hw_ctx(q, hctx, i)
2666                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2667
2668         /* all hctx are in .unused_hctx_list now */
2669         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2670                 list_del_init(&hctx->hctx_list);
2671                 kobject_put(&hctx->kobj);
2672         }
2673
2674         kfree(q->queue_hw_ctx);
2675
2676         /*
2677          * release .mq_kobj and sw queue's kobject now because
2678          * both share lifetime with request queue.
2679          */
2680         blk_mq_sysfs_deinit(q);
2681 }
2682
2683 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2684 {
2685         struct request_queue *uninit_q, *q;
2686
2687         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2688         if (!uninit_q)
2689                 return ERR_PTR(-ENOMEM);
2690
2691         q = blk_mq_init_allocated_queue(set, uninit_q);
2692         if (IS_ERR(q))
2693                 blk_cleanup_queue(uninit_q);
2694
2695         return q;
2696 }
2697 EXPORT_SYMBOL(blk_mq_init_queue);
2698
2699 /*
2700  * Helper for setting up a queue with mq ops, given queue depth, and
2701  * the passed in mq ops flags.
2702  */
2703 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2704                                            const struct blk_mq_ops *ops,
2705                                            unsigned int queue_depth,
2706                                            unsigned int set_flags)
2707 {
2708         struct request_queue *q;
2709         int ret;
2710
2711         memset(set, 0, sizeof(*set));
2712         set->ops = ops;
2713         set->nr_hw_queues = 1;
2714         set->nr_maps = 1;
2715         set->queue_depth = queue_depth;
2716         set->numa_node = NUMA_NO_NODE;
2717         set->flags = set_flags;
2718
2719         ret = blk_mq_alloc_tag_set(set);
2720         if (ret)
2721                 return ERR_PTR(ret);
2722
2723         q = blk_mq_init_queue(set);
2724         if (IS_ERR(q)) {
2725                 blk_mq_free_tag_set(set);
2726                 return q;
2727         }
2728
2729         return q;
2730 }
2731 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2732
2733 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2734                 struct blk_mq_tag_set *set, struct request_queue *q,
2735                 int hctx_idx, int node)
2736 {
2737         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2738
2739         /* reuse dead hctx first */
2740         spin_lock(&q->unused_hctx_lock);
2741         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2742                 if (tmp->numa_node == node) {
2743                         hctx = tmp;
2744                         break;
2745                 }
2746         }
2747         if (hctx)
2748                 list_del_init(&hctx->hctx_list);
2749         spin_unlock(&q->unused_hctx_lock);
2750
2751         if (!hctx)
2752                 hctx = blk_mq_alloc_hctx(q, set, node);
2753         if (!hctx)
2754                 goto fail;
2755
2756         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2757                 goto free_hctx;
2758
2759         return hctx;
2760
2761  free_hctx:
2762         kobject_put(&hctx->kobj);
2763  fail:
2764         return NULL;
2765 }
2766
2767 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2768                                                 struct request_queue *q)
2769 {
2770         int i, j, end;
2771         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2772
2773         /* protect against switching io scheduler  */
2774         mutex_lock(&q->sysfs_lock);
2775         for (i = 0; i < set->nr_hw_queues; i++) {
2776                 int node;
2777                 struct blk_mq_hw_ctx *hctx;
2778
2779                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2780                 /*
2781                  * If the hw queue has been mapped to another numa node,
2782                  * we need to realloc the hctx. If allocation fails, fallback
2783                  * to use the previous one.
2784                  */
2785                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2786                         continue;
2787
2788                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2789                 if (hctx) {
2790                         if (hctxs[i])
2791                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2792                         hctxs[i] = hctx;
2793                 } else {
2794                         if (hctxs[i])
2795                                 pr_warn("Allocate new hctx on node %d fails,\
2796                                                 fallback to previous one on node %d\n",
2797                                                 node, hctxs[i]->numa_node);
2798                         else
2799                                 break;
2800                 }
2801         }
2802         /*
2803          * Increasing nr_hw_queues fails. Free the newly allocated
2804          * hctxs and keep the previous q->nr_hw_queues.
2805          */
2806         if (i != set->nr_hw_queues) {
2807                 j = q->nr_hw_queues;
2808                 end = i;
2809         } else {
2810                 j = i;
2811                 end = q->nr_hw_queues;
2812                 q->nr_hw_queues = set->nr_hw_queues;
2813         }
2814
2815         for (; j < end; j++) {
2816                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2817
2818                 if (hctx) {
2819                         if (hctx->tags)
2820                                 blk_mq_free_map_and_requests(set, j);
2821                         blk_mq_exit_hctx(q, set, hctx, j);
2822                         hctxs[j] = NULL;
2823                 }
2824         }
2825         mutex_unlock(&q->sysfs_lock);
2826 }
2827
2828 /*
2829  * Maximum number of hardware queues we support. For single sets, we'll never
2830  * have more than the CPUs (software queues). For multiple sets, the tag_set
2831  * user may have set ->nr_hw_queues larger.
2832  */
2833 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2834 {
2835         if (set->nr_maps == 1)
2836                 return nr_cpu_ids;
2837
2838         return max(set->nr_hw_queues, nr_cpu_ids);
2839 }
2840
2841 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2842                                                   struct request_queue *q)
2843 {
2844         /* mark the queue as mq asap */
2845         q->mq_ops = set->ops;
2846
2847         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2848                                              blk_mq_poll_stats_bkt,
2849                                              BLK_MQ_POLL_STATS_BKTS, q);
2850         if (!q->poll_cb)
2851                 goto err_exit;
2852
2853         if (blk_mq_alloc_ctxs(q))
2854                 goto err_poll;
2855
2856         /* init q->mq_kobj and sw queues' kobjects */
2857         blk_mq_sysfs_init(q);
2858
2859         q->nr_queues = nr_hw_queues(set);
2860         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2861                                                 GFP_KERNEL, set->numa_node);
2862         if (!q->queue_hw_ctx)
2863                 goto err_sys_init;
2864
2865         INIT_LIST_HEAD(&q->unused_hctx_list);
2866         spin_lock_init(&q->unused_hctx_lock);
2867
2868         blk_mq_realloc_hw_ctxs(set, q);
2869         if (!q->nr_hw_queues)
2870                 goto err_hctxs;
2871
2872         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2873         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2874
2875         q->tag_set = set;
2876
2877         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2878         if (set->nr_maps > HCTX_TYPE_POLL &&
2879             set->map[HCTX_TYPE_POLL].nr_queues)
2880                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2881
2882         q->sg_reserved_size = INT_MAX;
2883
2884         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2885         INIT_LIST_HEAD(&q->requeue_list);
2886         spin_lock_init(&q->requeue_lock);
2887
2888         blk_queue_make_request(q, blk_mq_make_request);
2889
2890         /*
2891          * Do this after blk_queue_make_request() overrides it...
2892          */
2893         q->nr_requests = set->queue_depth;
2894
2895         /*
2896          * Default to classic polling
2897          */
2898         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2899
2900         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2901         blk_mq_add_queue_tag_set(set, q);
2902         blk_mq_map_swqueue(q);
2903
2904         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2905                 int ret;
2906
2907                 ret = elevator_init_mq(q);
2908                 if (ret)
2909                         return ERR_PTR(ret);
2910         }
2911
2912         return q;
2913
2914 err_hctxs:
2915         kfree(q->queue_hw_ctx);
2916 err_sys_init:
2917         blk_mq_sysfs_deinit(q);
2918 err_poll:
2919         blk_stat_free_callback(q->poll_cb);
2920         q->poll_cb = NULL;
2921 err_exit:
2922         q->mq_ops = NULL;
2923         return ERR_PTR(-ENOMEM);
2924 }
2925 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2926
2927 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2928 void blk_mq_exit_queue(struct request_queue *q)
2929 {
2930         struct blk_mq_tag_set   *set = q->tag_set;
2931
2932         blk_mq_del_queue_tag_set(q);
2933         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2934 }
2935
2936 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2937 {
2938         int i;
2939
2940         for (i = 0; i < set->nr_hw_queues; i++)
2941                 if (!__blk_mq_alloc_rq_map(set, i))
2942                         goto out_unwind;
2943
2944         return 0;
2945
2946 out_unwind:
2947         while (--i >= 0)
2948                 blk_mq_free_rq_map(set->tags[i]);
2949
2950         return -ENOMEM;
2951 }
2952
2953 /*
2954  * Allocate the request maps associated with this tag_set. Note that this
2955  * may reduce the depth asked for, if memory is tight. set->queue_depth
2956  * will be updated to reflect the allocated depth.
2957  */
2958 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2959 {
2960         unsigned int depth;
2961         int err;
2962
2963         depth = set->queue_depth;
2964         do {
2965                 err = __blk_mq_alloc_rq_maps(set);
2966                 if (!err)
2967                         break;
2968
2969                 set->queue_depth >>= 1;
2970                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2971                         err = -ENOMEM;
2972                         break;
2973                 }
2974         } while (set->queue_depth);
2975
2976         if (!set->queue_depth || err) {
2977                 pr_err("blk-mq: failed to allocate request map\n");
2978                 return -ENOMEM;
2979         }
2980
2981         if (depth != set->queue_depth)
2982                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2983                                                 depth, set->queue_depth);
2984
2985         return 0;
2986 }
2987
2988 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2989 {
2990         if (set->ops->map_queues && !is_kdump_kernel()) {
2991                 int i;
2992
2993                 /*
2994                  * transport .map_queues is usually done in the following
2995                  * way:
2996                  *
2997                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2998                  *      mask = get_cpu_mask(queue)
2999                  *      for_each_cpu(cpu, mask)
3000                  *              set->map[x].mq_map[cpu] = queue;
3001                  * }
3002                  *
3003                  * When we need to remap, the table has to be cleared for
3004                  * killing stale mapping since one CPU may not be mapped
3005                  * to any hw queue.
3006                  */
3007                 for (i = 0; i < set->nr_maps; i++)
3008                         blk_mq_clear_mq_map(&set->map[i]);
3009
3010                 return set->ops->map_queues(set);
3011         } else {
3012                 BUG_ON(set->nr_maps > 1);
3013                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3014         }
3015 }
3016
3017 /*
3018  * Alloc a tag set to be associated with one or more request queues.
3019  * May fail with EINVAL for various error conditions. May adjust the
3020  * requested depth down, if it's too large. In that case, the set
3021  * value will be stored in set->queue_depth.
3022  */
3023 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3024 {
3025         int i, ret;
3026
3027         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3028
3029         if (!set->nr_hw_queues)
3030                 return -EINVAL;
3031         if (!set->queue_depth)
3032                 return -EINVAL;
3033         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3034                 return -EINVAL;
3035
3036         if (!set->ops->queue_rq)
3037                 return -EINVAL;
3038
3039         if (!set->ops->get_budget ^ !set->ops->put_budget)
3040                 return -EINVAL;
3041
3042         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3043                 pr_info("blk-mq: reduced tag depth to %u\n",
3044                         BLK_MQ_MAX_DEPTH);
3045                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3046         }
3047
3048         if (!set->nr_maps)
3049                 set->nr_maps = 1;
3050         else if (set->nr_maps > HCTX_MAX_TYPES)
3051                 return -EINVAL;
3052
3053         /*
3054          * If a crashdump is active, then we are potentially in a very
3055          * memory constrained environment. Limit us to 1 queue and
3056          * 64 tags to prevent using too much memory.
3057          */
3058         if (is_kdump_kernel()) {
3059                 set->nr_hw_queues = 1;
3060                 set->nr_maps = 1;
3061                 set->queue_depth = min(64U, set->queue_depth);
3062         }
3063         /*
3064          * There is no use for more h/w queues than cpus if we just have
3065          * a single map
3066          */
3067         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3068                 set->nr_hw_queues = nr_cpu_ids;
3069
3070         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3071                                  GFP_KERNEL, set->numa_node);
3072         if (!set->tags)
3073                 return -ENOMEM;
3074
3075         ret = -ENOMEM;
3076         for (i = 0; i < set->nr_maps; i++) {
3077                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3078                                                   sizeof(set->map[i].mq_map[0]),
3079                                                   GFP_KERNEL, set->numa_node);
3080                 if (!set->map[i].mq_map)
3081                         goto out_free_mq_map;
3082                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3083         }
3084
3085         ret = blk_mq_update_queue_map(set);
3086         if (ret)
3087                 goto out_free_mq_map;
3088
3089         ret = blk_mq_alloc_rq_maps(set);
3090         if (ret)
3091                 goto out_free_mq_map;
3092
3093         mutex_init(&set->tag_list_lock);
3094         INIT_LIST_HEAD(&set->tag_list);
3095
3096         return 0;
3097
3098 out_free_mq_map:
3099         for (i = 0; i < set->nr_maps; i++) {
3100                 kfree(set->map[i].mq_map);
3101                 set->map[i].mq_map = NULL;
3102         }
3103         kfree(set->tags);
3104         set->tags = NULL;
3105         return ret;
3106 }
3107 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3108
3109 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3110 {
3111         int i, j;
3112
3113         for (i = 0; i < nr_hw_queues(set); i++)
3114                 blk_mq_free_map_and_requests(set, i);
3115
3116         for (j = 0; j < set->nr_maps; j++) {
3117                 kfree(set->map[j].mq_map);
3118                 set->map[j].mq_map = NULL;
3119         }
3120
3121         kfree(set->tags);
3122         set->tags = NULL;
3123 }
3124 EXPORT_SYMBOL(blk_mq_free_tag_set);
3125
3126 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3127 {
3128         struct blk_mq_tag_set *set = q->tag_set;
3129         struct blk_mq_hw_ctx *hctx;
3130         int i, ret;
3131
3132         if (!set)
3133                 return -EINVAL;
3134
3135         if (q->nr_requests == nr)
3136                 return 0;
3137
3138         blk_mq_freeze_queue(q);
3139         blk_mq_quiesce_queue(q);
3140
3141         ret = 0;
3142         queue_for_each_hw_ctx(q, hctx, i) {
3143                 if (!hctx->tags)
3144                         continue;
3145                 /*
3146                  * If we're using an MQ scheduler, just update the scheduler
3147                  * queue depth. This is similar to what the old code would do.
3148                  */
3149                 if (!hctx->sched_tags) {
3150                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3151                                                         false);
3152                 } else {
3153                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3154                                                         nr, true);
3155                 }
3156                 if (ret)
3157                         break;
3158                 if (q->elevator && q->elevator->type->ops.depth_updated)
3159                         q->elevator->type->ops.depth_updated(hctx);
3160         }
3161
3162         if (!ret)
3163                 q->nr_requests = nr;
3164
3165         blk_mq_unquiesce_queue(q);
3166         blk_mq_unfreeze_queue(q);
3167
3168         return ret;
3169 }
3170
3171 /*
3172  * request_queue and elevator_type pair.
3173  * It is just used by __blk_mq_update_nr_hw_queues to cache
3174  * the elevator_type associated with a request_queue.
3175  */
3176 struct blk_mq_qe_pair {
3177         struct list_head node;
3178         struct request_queue *q;
3179         struct elevator_type *type;
3180 };
3181
3182 /*
3183  * Cache the elevator_type in qe pair list and switch the
3184  * io scheduler to 'none'
3185  */
3186 static bool blk_mq_elv_switch_none(struct list_head *head,
3187                 struct request_queue *q)
3188 {
3189         struct blk_mq_qe_pair *qe;
3190
3191         if (!q->elevator)
3192                 return true;
3193
3194         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3195         if (!qe)
3196                 return false;
3197
3198         INIT_LIST_HEAD(&qe->node);
3199         qe->q = q;
3200         qe->type = q->elevator->type;
3201         list_add(&qe->node, head);
3202
3203         mutex_lock(&q->sysfs_lock);
3204         /*
3205          * After elevator_switch_mq, the previous elevator_queue will be
3206          * released by elevator_release. The reference of the io scheduler
3207          * module get by elevator_get will also be put. So we need to get
3208          * a reference of the io scheduler module here to prevent it to be
3209          * removed.
3210          */
3211         __module_get(qe->type->elevator_owner);
3212         elevator_switch_mq(q, NULL);
3213         mutex_unlock(&q->sysfs_lock);
3214
3215         return true;
3216 }
3217
3218 static void blk_mq_elv_switch_back(struct list_head *head,
3219                 struct request_queue *q)
3220 {
3221         struct blk_mq_qe_pair *qe;
3222         struct elevator_type *t = NULL;
3223
3224         list_for_each_entry(qe, head, node)
3225                 if (qe->q == q) {
3226                         t = qe->type;
3227                         break;
3228                 }
3229
3230         if (!t)
3231                 return;
3232
3233         list_del(&qe->node);
3234         kfree(qe);
3235
3236         mutex_lock(&q->sysfs_lock);
3237         elevator_switch_mq(q, t);
3238         mutex_unlock(&q->sysfs_lock);
3239 }
3240
3241 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3242                                                         int nr_hw_queues)
3243 {
3244         struct request_queue *q;
3245         LIST_HEAD(head);
3246         int prev_nr_hw_queues;
3247
3248         lockdep_assert_held(&set->tag_list_lock);
3249
3250         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3251                 nr_hw_queues = nr_cpu_ids;
3252         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3253                 return;
3254
3255         list_for_each_entry(q, &set->tag_list, tag_set_list)
3256                 blk_mq_freeze_queue(q);
3257         /*
3258          * Sync with blk_mq_queue_tag_busy_iter.
3259          */
3260         synchronize_rcu();
3261         /*
3262          * Switch IO scheduler to 'none', cleaning up the data associated
3263          * with the previous scheduler. We will switch back once we are done
3264          * updating the new sw to hw queue mappings.
3265          */
3266         list_for_each_entry(q, &set->tag_list, tag_set_list)
3267                 if (!blk_mq_elv_switch_none(&head, q))
3268                         goto switch_back;
3269
3270         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3271                 blk_mq_debugfs_unregister_hctxs(q);
3272                 blk_mq_sysfs_unregister(q);
3273         }
3274
3275         prev_nr_hw_queues = set->nr_hw_queues;
3276         set->nr_hw_queues = nr_hw_queues;
3277         blk_mq_update_queue_map(set);
3278 fallback:
3279         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3280                 blk_mq_realloc_hw_ctxs(set, q);
3281                 if (q->nr_hw_queues != set->nr_hw_queues) {
3282                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3283                                         nr_hw_queues, prev_nr_hw_queues);
3284                         set->nr_hw_queues = prev_nr_hw_queues;
3285                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3286                         goto fallback;
3287                 }
3288                 blk_mq_map_swqueue(q);
3289         }
3290
3291         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3292                 blk_mq_sysfs_register(q);
3293                 blk_mq_debugfs_register_hctxs(q);
3294         }
3295
3296 switch_back:
3297         list_for_each_entry(q, &set->tag_list, tag_set_list)
3298                 blk_mq_elv_switch_back(&head, q);
3299
3300         list_for_each_entry(q, &set->tag_list, tag_set_list)
3301                 blk_mq_unfreeze_queue(q);
3302 }
3303
3304 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3305 {
3306         mutex_lock(&set->tag_list_lock);
3307         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3308         mutex_unlock(&set->tag_list_lock);
3309 }
3310 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3311
3312 /* Enable polling stats and return whether they were already enabled. */
3313 static bool blk_poll_stats_enable(struct request_queue *q)
3314 {
3315         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3316             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3317                 return true;
3318         blk_stat_add_callback(q, q->poll_cb);
3319         return false;
3320 }
3321
3322 static void blk_mq_poll_stats_start(struct request_queue *q)
3323 {
3324         /*
3325          * We don't arm the callback if polling stats are not enabled or the
3326          * callback is already active.
3327          */
3328         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3329             blk_stat_is_active(q->poll_cb))
3330                 return;
3331
3332         blk_stat_activate_msecs(q->poll_cb, 100);
3333 }
3334
3335 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3336 {
3337         struct request_queue *q = cb->data;
3338         int bucket;
3339
3340         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3341                 if (cb->stat[bucket].nr_samples)
3342                         q->poll_stat[bucket] = cb->stat[bucket];
3343         }
3344 }
3345
3346 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3347                                        struct blk_mq_hw_ctx *hctx,
3348                                        struct request *rq)
3349 {
3350         unsigned long ret = 0;
3351         int bucket;
3352
3353         /*
3354          * If stats collection isn't on, don't sleep but turn it on for
3355          * future users
3356          */
3357         if (!blk_poll_stats_enable(q))
3358                 return 0;
3359
3360         /*
3361          * As an optimistic guess, use half of the mean service time
3362          * for this type of request. We can (and should) make this smarter.
3363          * For instance, if the completion latencies are tight, we can
3364          * get closer than just half the mean. This is especially
3365          * important on devices where the completion latencies are longer
3366          * than ~10 usec. We do use the stats for the relevant IO size
3367          * if available which does lead to better estimates.
3368          */
3369         bucket = blk_mq_poll_stats_bkt(rq);
3370         if (bucket < 0)
3371                 return ret;
3372
3373         if (q->poll_stat[bucket].nr_samples)
3374                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3375
3376         return ret;
3377 }
3378
3379 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3380                                      struct blk_mq_hw_ctx *hctx,
3381                                      struct request *rq)
3382 {
3383         struct hrtimer_sleeper hs;
3384         enum hrtimer_mode mode;
3385         unsigned int nsecs;
3386         ktime_t kt;
3387
3388         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3389                 return false;
3390
3391         /*
3392          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3393          *
3394          *  0:  use half of prev avg
3395          * >0:  use this specific value
3396          */
3397         if (q->poll_nsec > 0)
3398                 nsecs = q->poll_nsec;
3399         else
3400                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3401
3402         if (!nsecs)
3403                 return false;
3404
3405         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3406
3407         /*
3408          * This will be replaced with the stats tracking code, using
3409          * 'avg_completion_time / 2' as the pre-sleep target.
3410          */
3411         kt = nsecs;
3412
3413         mode = HRTIMER_MODE_REL;
3414         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3415         hrtimer_set_expires(&hs.timer, kt);
3416
3417         hrtimer_init_sleeper(&hs, current);
3418         do {
3419                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3420                         break;
3421                 set_current_state(TASK_UNINTERRUPTIBLE);
3422                 hrtimer_start_expires(&hs.timer, mode);
3423                 if (hs.task)
3424                         io_schedule();
3425                 hrtimer_cancel(&hs.timer);
3426                 mode = HRTIMER_MODE_ABS;
3427         } while (hs.task && !signal_pending(current));
3428
3429         __set_current_state(TASK_RUNNING);
3430         destroy_hrtimer_on_stack(&hs.timer);
3431         return true;
3432 }
3433
3434 static bool blk_mq_poll_hybrid(struct request_queue *q,
3435                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3436 {
3437         struct request *rq;
3438
3439         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3440                 return false;
3441
3442         if (!blk_qc_t_is_internal(cookie))
3443                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3444         else {
3445                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3446                 /*
3447                  * With scheduling, if the request has completed, we'll
3448                  * get a NULL return here, as we clear the sched tag when
3449                  * that happens. The request still remains valid, like always,
3450                  * so we should be safe with just the NULL check.
3451                  */
3452                 if (!rq)
3453                         return false;
3454         }
3455
3456         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3457 }
3458
3459 /**
3460  * blk_poll - poll for IO completions
3461  * @q:  the queue
3462  * @cookie: cookie passed back at IO submission time
3463  * @spin: whether to spin for completions
3464  *
3465  * Description:
3466  *    Poll for completions on the passed in queue. Returns number of
3467  *    completed entries found. If @spin is true, then blk_poll will continue
3468  *    looping until at least one completion is found, unless the task is
3469  *    otherwise marked running (or we need to reschedule).
3470  */
3471 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3472 {
3473         struct blk_mq_hw_ctx *hctx;
3474         long state;
3475
3476         if (!blk_qc_t_valid(cookie) ||
3477             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3478                 return 0;
3479
3480         if (current->plug)
3481                 blk_flush_plug_list(current->plug, false);
3482
3483         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3484
3485         /*
3486          * If we sleep, have the caller restart the poll loop to reset
3487          * the state. Like for the other success return cases, the
3488          * caller is responsible for checking if the IO completed. If
3489          * the IO isn't complete, we'll get called again and will go
3490          * straight to the busy poll loop.
3491          */
3492         if (blk_mq_poll_hybrid(q, hctx, cookie))
3493                 return 1;
3494
3495         hctx->poll_considered++;
3496
3497         state = current->state;
3498         do {
3499                 int ret;
3500
3501                 hctx->poll_invoked++;
3502
3503                 ret = q->mq_ops->poll(hctx);
3504                 if (ret > 0) {
3505                         hctx->poll_success++;
3506                         __set_current_state(TASK_RUNNING);
3507                         return ret;
3508                 }
3509
3510                 if (signal_pending_state(state, current))
3511                         __set_current_state(TASK_RUNNING);
3512
3513                 if (current->state == TASK_RUNNING)
3514                         return 1;
3515                 if (ret < 0 || !spin)
3516                         break;
3517                 cpu_relax();
3518         } while (!need_resched());
3519
3520         __set_current_state(TASK_RUNNING);
3521         return 0;
3522 }
3523 EXPORT_SYMBOL_GPL(blk_poll);
3524
3525 unsigned int blk_mq_rq_cpu(struct request *rq)
3526 {
3527         return rq->mq_ctx->cpu;
3528 }
3529 EXPORT_SYMBOL(blk_mq_rq_cpu);
3530
3531 static int __init blk_mq_init(void)
3532 {
3533         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3534                                 blk_mq_hctx_notify_dead);
3535         return 0;
3536 }
3537 subsys_initcall(blk_mq_init);