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         blk_mq_sched_started_request(rq);
 673
 674         trace_block_rq_issue(q, rq);
 675
 676         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 677                 rq->io_start_time_ns = ktime_get_ns();
 678 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 679                 rq->throtl_size = blk_rq_sectors(rq);
 680 #endif
 681                 rq->rq_flags |= RQF_STATS;
 682                 rq_qos_issue(q, rq);
 683         }
 684
 685         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 686
 687         blk_add_timer(rq);
 688         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 689
 690         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 691                 /*
 692                  * Make sure space for the drain appears.  We know we can do
 693                  * this because max_hw_segments has been adjusted to be one
 694                  * fewer than the device can handle.
 695                  */
 696                 rq->nr_phys_segments++;
 697         }
 698 }
 699 EXPORT_SYMBOL(blk_mq_start_request);
 700
 701 static void __blk_mq_requeue_request(struct request *rq)
 702 {
 703         struct request_queue *q = rq->q;
 704
 705         blk_mq_put_driver_tag(rq);
 706
 707         trace_block_rq_requeue(q, rq);
 708         rq_qos_requeue(q, rq);
 709
 710         if (blk_mq_request_started(rq)) {
 711                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 712                 rq->rq_flags &= ~RQF_TIMED_OUT;
 713                 if (q->dma_drain_size && blk_rq_bytes(rq))
 714                         rq->nr_phys_segments--;
 715         }
 716 }
 717
 718 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 719 {
 720         __blk_mq_requeue_request(rq);
 721
 722         /* this request will be re-inserted to io scheduler queue */
 723         blk_mq_sched_requeue_request(rq);
 724
 725         BUG_ON(!list_empty(&rq->queuelist));
 726         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 727 }
 728 EXPORT_SYMBOL(blk_mq_requeue_request);
 729
 730 static void blk_mq_requeue_work(struct work_struct *work)
 731 {
 732         struct request_queue *q =
 733                 container_of(work, struct request_queue, requeue_work.work);
 734         LIST_HEAD(rq_list);
 735         struct request *rq, *next;
 736
 737         spin_lock_irq(&q->requeue_lock);
 738         list_splice_init(&q->requeue_list, &rq_list);
 739         spin_unlock_irq(&q->requeue_lock);
 740
 741         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 742                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 743                         continue;
 744
 745                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 746                 list_del_init(&rq->queuelist);
 747                 /*
 748                  * If RQF_DONTPREP, rq has contained some driver specific
 749                  * data, so insert it to hctx dispatch list to avoid any
 750                  * merge.
 751                  */
 752                 if (rq->rq_flags & RQF_DONTPREP)
 753                         blk_mq_request_bypass_insert(rq, false);
 754                 else
 755                         blk_mq_sched_insert_request(rq, true, false, false);
 756         }
 757
 758         while (!list_empty(&rq_list)) {
 759                 rq = list_entry(rq_list.next, struct request, queuelist);
 760                 list_del_init(&rq->queuelist);
 761                 blk_mq_sched_insert_request(rq, false, false, false);
 762         }
 763
 764         blk_mq_run_hw_queues(q, false);
 765 }
 766
 767 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 768                                 bool kick_requeue_list)
 769 {
 770         struct request_queue *q = rq->q;
 771         unsigned long flags;
 772
 773         /*
 774          * We abuse this flag that is otherwise used by the I/O scheduler to
 775          * request head insertion from the workqueue.
 776          */
 777         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 778
 779         spin_lock_irqsave(&q->requeue_lock, flags);
 780         if (at_head) {
 781                 rq->rq_flags |= RQF_SOFTBARRIER;
 782                 list_add(&rq->queuelist, &q->requeue_list);
 783         } else {
 784                 list_add_tail(&rq->queuelist, &q->requeue_list);
 785         }
 786         spin_unlock_irqrestore(&q->requeue_lock, flags);
 787
 788         if (kick_requeue_list)
 789                 blk_mq_kick_requeue_list(q);
 790 }
 791
 792 void blk_mq_kick_requeue_list(struct request_queue *q)
 793 {
 794         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 795 }
 796 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 797
 798 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 799                                     unsigned long msecs)
 800 {
 801         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 802                                     msecs_to_jiffies(msecs));
 803 }
 804 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 805
 806 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 807 {
 808         if (tag < tags->nr_tags) {
 809                 prefetch(tags->rqs[tag]);
 810                 return tags->rqs[tag];
 811         }
 812
 813         return NULL;
 814 }
 815 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 816
 817 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 818                                void *priv, bool reserved)
 819 {
 820         /*
 821          * If we find a request that is inflight and the queue matches,
 822          * we know the queue is busy. Return false to stop the iteration.
 823          */
 824         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 825                 bool *busy = priv;
 826
 827                 *busy = true;
 828                 return false;
 829         }
 830
 831         return true;
 832 }
 833
 834 bool blk_mq_queue_inflight(struct request_queue *q)
 835 {
 836         bool busy = false;
 837
 838         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 839         return busy;
 840 }
 841 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 842
 843 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 844 {
 845         req->rq_flags |= RQF_TIMED_OUT;
 846         if (req->q->mq_ops->timeout) {
 847                 enum blk_eh_timer_return ret;
 848
 849                 ret = req->q->mq_ops->timeout(req, reserved);
 850                 if (ret == BLK_EH_DONE)
 851                         return;
 852                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 853         }
 854
 855         blk_add_timer(req);
 856 }
 857
 858 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 859 {
 860         unsigned long deadline;
 861
 862         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 863                 return false;
 864         if (rq->rq_flags & RQF_TIMED_OUT)
 865                 return false;
 866
 867         deadline = READ_ONCE(rq->deadline);
 868         if (time_after_eq(jiffies, deadline))
 869                 return true;
 870
 871         if (*next == 0)
 872                 *next = deadline;
 873         else if (time_after(*next, deadline))
 874                 *next = deadline;
 875         return false;
 876 }
 877
 878 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 879                 struct request *rq, void *priv, bool reserved)
 880 {
 881         unsigned long *next = priv;
 882
 883         /*
 884          * Just do a quick check if it is expired before locking the request in
 885          * so we're not unnecessarilly synchronizing across CPUs.
 886          */
 887         if (!blk_mq_req_expired(rq, next))
 888                 return true;
 889
 890         /*
 891          * We have reason to believe the request may be expired. Take a
 892          * reference on the request to lock this request lifetime into its
 893          * currently allocated context to prevent it from being reallocated in
 894          * the event the completion by-passes this timeout handler.
 895          *
 896          * If the reference was already released, then the driver beat the
 897          * timeout handler to posting a natural completion.
 898          */
 899         if (!refcount_inc_not_zero(&rq->ref))
 900                 return true;
 901
 902         /*
 903          * The request is now locked and cannot be reallocated underneath the
 904          * timeout handler's processing. Re-verify this exact request is truly
 905          * expired; if it is not expired, then the request was completed and
 906          * reallocated as a new request.
 907          */
 908         if (blk_mq_req_expired(rq, next))
 909                 blk_mq_rq_timed_out(rq, reserved);
 910         if (refcount_dec_and_test(&rq->ref))
 911                 __blk_mq_free_request(rq);
 912
 913         return true;
 914 }
 915
 916 static void blk_mq_timeout_work(struct work_struct *work)
 917 {
 918         struct request_queue *q =
 919                 container_of(work, struct request_queue, timeout_work);
 920         unsigned long next = 0;
 921         struct blk_mq_hw_ctx *hctx;
 922         int i;
 923
 924         /* A deadlock might occur if a request is stuck requiring a
 925          * timeout at the same time a queue freeze is waiting
 926          * completion, since the timeout code would not be able to
 927          * acquire the queue reference here.
 928          *
 929          * That's why we don't use blk_queue_enter here; instead, we use
 930          * percpu_ref_tryget directly, because we need to be able to
 931          * obtain a reference even in the short window between the queue
 932          * starting to freeze, by dropping the first reference in
 933          * blk_freeze_queue_start, and the moment the last request is
 934          * consumed, marked by the instant q_usage_counter reaches
 935          * zero.
 936          */
 937         if (!percpu_ref_tryget(&q->q_usage_counter))
 938                 return;
 939
 940         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 941
 942         if (next != 0) {
 943                 mod_timer(&q->timeout, next);
 944         } else {
 945                 /*
 946                  * Request timeouts are handled as a forward rolling timer. If
 947                  * we end up here it means that no requests are pending and
 948                  * also that no request has been pending for a while. Mark
 949                  * each hctx as idle.
 950                  */
 951                 queue_for_each_hw_ctx(q, hctx, i) {
 952                         /* the hctx may be unmapped, so check it here */
 953                         if (blk_mq_hw_queue_mapped(hctx))
 954                                 blk_mq_tag_idle(hctx);
 955                 }
 956         }
 957         blk_queue_exit(q);
 958 }
 959
 960 struct flush_busy_ctx_data {
 961         struct blk_mq_hw_ctx *hctx;
 962         struct list_head *list;
 963 };
 964
 965 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 966 {
 967         struct flush_busy_ctx_data *flush_data = data;
 968         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 969         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 970         enum hctx_type type = hctx->type;
 971
 972         spin_lock(&ctx->lock);
 973         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 974         sbitmap_clear_bit(sb, bitnr);
 975         spin_unlock(&ctx->lock);
 976         return true;
 977 }
 978
 979 /*
 980  * Process software queues that have been marked busy, splicing them
 981  * to the for-dispatch
 982  */
 983 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 984 {
 985         struct flush_busy_ctx_data data = {
 986                 .hctx = hctx,
 987                 .list = list,
 988         };
 989
 990         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 991 }
 992 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 993
 994 struct dispatch_rq_data {
 995         struct blk_mq_hw_ctx *hctx;
 996         struct request *rq;
 997 };
 998
 999 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1000                 void *data)
1001 {
1002         struct dispatch_rq_data *dispatch_data = data;
1003         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1004         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1005         enum hctx_type type = hctx->type;
1006
1007         spin_lock(&ctx->lock);
1008         if (!list_empty(&ctx->rq_lists[type])) {
1009                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1010                 list_del_init(&dispatch_data->rq->queuelist);
1011                 if (list_empty(&ctx->rq_lists[type]))
1012                         sbitmap_clear_bit(sb, bitnr);
1013         }
1014         spin_unlock(&ctx->lock);
1015
1016         return !dispatch_data->rq;
1017 }
1018
1019 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1020                                         struct blk_mq_ctx *start)
1021 {
1022         unsigned off = start ? start->index_hw[hctx->type] : 0;
1023         struct dispatch_rq_data data = {
1024                 .hctx = hctx,
1025                 .rq   = NULL,
1026         };
1027
1028         __sbitmap_for_each_set(&hctx->ctx_map, off,
1029                                dispatch_rq_from_ctx, &data);
1030
1031         return data.rq;
1032 }
1033
1034 static inline unsigned int queued_to_index(unsigned int queued)
1035 {
1036         if (!queued)
1037                 return 0;
1038
1039         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1040 }
1041
1042 bool blk_mq_get_driver_tag(struct request *rq)
1043 {
1044         struct blk_mq_alloc_data data = {
1045                 .q = rq->q,
1046                 .hctx = rq->mq_hctx,
1047                 .flags = BLK_MQ_REQ_NOWAIT,
1048                 .cmd_flags = rq->cmd_flags,
1049         };
1050         bool shared;
1051
1052         if (rq->tag != -1)
1053                 goto done;
1054
1055         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1056                 data.flags |= BLK_MQ_REQ_RESERVED;
1057
1058         shared = blk_mq_tag_busy(data.hctx);
1059         rq->tag = blk_mq_get_tag(&data);
1060         if (rq->tag >= 0) {
1061                 if (shared) {
1062                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1063                         atomic_inc(&data.hctx->nr_active);
1064                 }
1065                 data.hctx->tags->rqs[rq->tag] = rq;
1066         }
1067
1068 done:
1069         return rq->tag != -1;
1070 }
1071
1072 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1073                                 int flags, void *key)
1074 {
1075         struct blk_mq_hw_ctx *hctx;
1076
1077         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1078
1079         spin_lock(&hctx->dispatch_wait_lock);
1080         if (!list_empty(&wait->entry)) {
1081                 struct sbitmap_queue *sbq;
1082
1083                 list_del_init(&wait->entry);
1084                 sbq = &hctx->tags->bitmap_tags;
1085                 atomic_dec(&sbq->ws_active);
1086         }
1087         spin_unlock(&hctx->dispatch_wait_lock);
1088
1089         blk_mq_run_hw_queue(hctx, true);
1090         return 1;
1091 }
1092
1093 /*
1094  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1095  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1096  * restart. For both cases, take care to check the condition again after
1097  * marking us as waiting.
1098  */
1099 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1100                                  struct request *rq)
1101 {
1102         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1103         struct wait_queue_head *wq;
1104         wait_queue_entry_t *wait;
1105         bool ret;
1106
1107         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1108                 blk_mq_sched_mark_restart_hctx(hctx);
1109
1110                 /*
1111                  * It's possible that a tag was freed in the window between the
1112                  * allocation failure and adding the hardware queue to the wait
1113                  * queue.
1114                  *
1115                  * Don't clear RESTART here, someone else could have set it.
1116                  * At most this will cost an extra queue run.
1117                  */
1118                 return blk_mq_get_driver_tag(rq);
1119         }
1120
1121         wait = &hctx->dispatch_wait;
1122         if (!list_empty_careful(&wait->entry))
1123                 return false;
1124
1125         wq = &bt_wait_ptr(sbq, hctx)->wait;
1126
1127         spin_lock_irq(&wq->lock);
1128         spin_lock(&hctx->dispatch_wait_lock);
1129         if (!list_empty(&wait->entry)) {
1130                 spin_unlock(&hctx->dispatch_wait_lock);
1131                 spin_unlock_irq(&wq->lock);
1132                 return false;
1133         }
1134
1135         atomic_inc(&sbq->ws_active);
1136         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1137         __add_wait_queue(wq, wait);
1138
1139         /*
1140          * It's possible that a tag was freed in the window between the
1141          * allocation failure and adding the hardware queue to the wait
1142          * queue.
1143          */
1144         ret = blk_mq_get_driver_tag(rq);
1145         if (!ret) {
1146                 spin_unlock(&hctx->dispatch_wait_lock);
1147                 spin_unlock_irq(&wq->lock);
1148                 return false;
1149         }
1150
1151         /*
1152          * We got a tag, remove ourselves from the wait queue to ensure
1153          * someone else gets the wakeup.
1154          */
1155         list_del_init(&wait->entry);
1156         atomic_dec(&sbq->ws_active);
1157         spin_unlock(&hctx->dispatch_wait_lock);
1158         spin_unlock_irq(&wq->lock);
1159
1160         return true;
1161 }
1162
1163 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1164 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1165 /*
1166  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1167  * - EWMA is one simple way to compute running average value
1168  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1169  * - take 4 as factor for avoiding to get too small(0) result, and this
1170  *   factor doesn't matter because EWMA decreases exponentially
1171  */
1172 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1173 {
1174         unsigned int ewma;
1175
1176         if (hctx->queue->elevator)
1177                 return;
1178
1179         ewma = hctx->dispatch_busy;
1180
1181         if (!ewma && !busy)
1182                 return;
1183
1184         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1185         if (busy)
1186                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1187         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1188
1189         hctx->dispatch_busy = ewma;
1190 }
1191
1192 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1193
1194 /*
1195  * Returns true if we did some work AND can potentially do more.
1196  */
1197 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1198                              bool got_budget)
1199 {
1200         struct blk_mq_hw_ctx *hctx;
1201         struct request *rq, *nxt;
1202         bool no_tag = false;
1203         int errors, queued;
1204         blk_status_t ret = BLK_STS_OK;
1205
1206         if (list_empty(list))
1207                 return false;
1208
1209         WARN_ON(!list_is_singular(list) && got_budget);
1210
1211         /*
1212          * Now process all the entries, sending them to the driver.
1213          */
1214         errors = queued = 0;
1215         do {
1216                 struct blk_mq_queue_data bd;
1217
1218                 rq = list_first_entry(list, struct request, queuelist);
1219
1220                 hctx = rq->mq_hctx;
1221                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1222                         break;
1223
1224                 if (!blk_mq_get_driver_tag(rq)) {
1225                         /*
1226                          * The initial allocation attempt failed, so we need to
1227                          * rerun the hardware queue when a tag is freed. The
1228                          * waitqueue takes care of that. If the queue is run
1229                          * before we add this entry back on the dispatch list,
1230                          * we'll re-run it below.
1231                          */
1232                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1233                                 blk_mq_put_dispatch_budget(hctx);
1234                                 /*
1235                                  * For non-shared tags, the RESTART check
1236                                  * will suffice.
1237                                  */
1238                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1239                                         no_tag = true;
1240                                 break;
1241                         }
1242                 }
1243
1244                 list_del_init(&rq->queuelist);
1245
1246                 bd.rq = rq;
1247
1248                 /*
1249                  * Flag last if we have no more requests, or if we have more
1250                  * but can't assign a driver tag to it.
1251                  */
1252                 if (list_empty(list))
1253                         bd.last = true;
1254                 else {
1255                         nxt = list_first_entry(list, struct request, queuelist);
1256                         bd.last = !blk_mq_get_driver_tag(nxt);
1257                 }
1258
1259                 ret = q->mq_ops->queue_rq(hctx, &bd);
1260                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1261                         /*
1262                          * If an I/O scheduler has been configured and we got a
1263                          * driver tag for the next request already, free it
1264                          * again.
1265                          */
1266                         if (!list_empty(list)) {
1267                                 nxt = list_first_entry(list, struct request, queuelist);
1268                                 blk_mq_put_driver_tag(nxt);
1269                         }
1270                         list_add(&rq->queuelist, list);
1271                         __blk_mq_requeue_request(rq);
1272                         break;
1273                 }
1274
1275                 if (unlikely(ret != BLK_STS_OK)) {
1276                         errors++;
1277                         blk_mq_end_request(rq, BLK_STS_IOERR);
1278                         continue;
1279                 }
1280
1281                 queued++;
1282         } while (!list_empty(list));
1283
1284         hctx->dispatched[queued_to_index(queued)]++;
1285
1286         /*
1287          * Any items that need requeuing? Stuff them into hctx->dispatch,
1288          * that is where we will continue on next queue run.
1289          */
1290         if (!list_empty(list)) {
1291                 bool needs_restart;
1292
1293                 /*
1294                  * If we didn't flush the entire list, we could have told
1295                  * the driver there was more coming, but that turned out to
1296                  * be a lie.
1297                  */
1298                 if (q->mq_ops->commit_rqs)
1299                         q->mq_ops->commit_rqs(hctx);
1300
1301                 spin_lock(&hctx->lock);
1302                 list_splice_init(list, &hctx->dispatch);
1303                 spin_unlock(&hctx->lock);
1304
1305                 /*
1306                  * If SCHED_RESTART was set by the caller of this function and
1307                  * it is no longer set that means that it was cleared by another
1308                  * thread and hence that a queue rerun is needed.
1309                  *
1310                  * If 'no_tag' is set, that means that we failed getting
1311                  * a driver tag with an I/O scheduler attached. If our dispatch
1312                  * waitqueue is no longer active, ensure that we run the queue
1313                  * AFTER adding our entries back to the list.
1314                  *
1315                  * If no I/O scheduler has been configured it is possible that
1316                  * the hardware queue got stopped and restarted before requests
1317                  * were pushed back onto the dispatch list. Rerun the queue to
1318                  * avoid starvation. Notes:
1319                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1320                  *   been stopped before rerunning a queue.
1321                  * - Some but not all block drivers stop a queue before
1322                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1323                  *   and dm-rq.
1324                  *
1325                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1326                  * bit is set, run queue after a delay to avoid IO stalls
1327                  * that could otherwise occur if the queue is idle.
1328                  */
1329                 needs_restart = blk_mq_sched_needs_restart(hctx);
1330                 if (!needs_restart ||
1331                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1332                         blk_mq_run_hw_queue(hctx, true);
1333                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1334                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1335
1336                 blk_mq_update_dispatch_busy(hctx, true);
1337                 return false;
1338         } else
1339                 blk_mq_update_dispatch_busy(hctx, false);
1340
1341         /*
1342          * If the host/device is unable to accept more work, inform the
1343          * caller of that.
1344          */
1345         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1346                 return false;
1347
1348         return (queued + errors) != 0;
1349 }
1350
1351 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1352 {
1353         int srcu_idx;
1354
1355         /*
1356          * We should be running this queue from one of the CPUs that
1357          * are mapped to it.
1358          *
1359          * There are at least two related races now between setting
1360          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1361          * __blk_mq_run_hw_queue():
1362          *
1363          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1364          *   but later it becomes online, then this warning is harmless
1365          *   at all
1366          *
1367          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1368          *   but later it becomes offline, then the warning can't be
1369          *   triggered, and we depend on blk-mq timeout handler to
1370          *   handle dispatched requests to this hctx
1371          */
1372         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1373                 cpu_online(hctx->next_cpu)) {
1374                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1375                         raw_smp_processor_id(),
1376                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1377                 dump_stack();
1378         }
1379
1380         /*
1381          * We can't run the queue inline with ints disabled. Ensure that
1382          * we catch bad users of this early.
1383          */
1384         WARN_ON_ONCE(in_interrupt());
1385
1386         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1387
1388         hctx_lock(hctx, &srcu_idx);
1389         blk_mq_sched_dispatch_requests(hctx);
1390         hctx_unlock(hctx, srcu_idx);
1391 }
1392
1393 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1394 {
1395         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1396
1397         if (cpu >= nr_cpu_ids)
1398                 cpu = cpumask_first(hctx->cpumask);
1399         return cpu;
1400 }
1401
1402 /*
1403  * It'd be great if the workqueue API had a way to pass
1404  * in a mask and had some smarts for more clever placement.
1405  * For now we just round-robin here, switching for every
1406  * BLK_MQ_CPU_WORK_BATCH queued items.
1407  */
1408 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1409 {
1410         bool tried = false;
1411         int next_cpu = hctx->next_cpu;
1412
1413         if (hctx->queue->nr_hw_queues == 1)
1414                 return WORK_CPU_UNBOUND;
1415
1416         if (--hctx->next_cpu_batch <= 0) {
1417 select_cpu:
1418                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1419                                 cpu_online_mask);
1420                 if (next_cpu >= nr_cpu_ids)
1421                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1422                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1423         }
1424
1425         /*
1426          * Do unbound schedule if we can't find a online CPU for this hctx,
1427          * and it should only happen in the path of handling CPU DEAD.
1428          */
1429         if (!cpu_online(next_cpu)) {
1430                 if (!tried) {
1431                         tried = true;
1432                         goto select_cpu;
1433                 }
1434
1435                 /*
1436                  * Make sure to re-select CPU next time once after CPUs
1437                  * in hctx->cpumask become online again.
1438                  */
1439                 hctx->next_cpu = next_cpu;
1440                 hctx->next_cpu_batch = 1;
1441                 return WORK_CPU_UNBOUND;
1442         }
1443
1444         hctx->next_cpu = next_cpu;
1445         return next_cpu;
1446 }
1447
1448 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1449                                         unsigned long msecs)
1450 {
1451         if (unlikely(blk_mq_hctx_stopped(hctx)))
1452                 return;
1453
1454         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1455                 int cpu = get_cpu();
1456                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1457                         __blk_mq_run_hw_queue(hctx);
1458                         put_cpu();
1459                         return;
1460                 }
1461
1462                 put_cpu();
1463         }
1464
1465         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1466                                     msecs_to_jiffies(msecs));
1467 }
1468
1469 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1470 {
1471         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1472 }
1473 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1474
1475 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1476 {
1477         int srcu_idx;
1478         bool need_run;
1479
1480         /*
1481          * When queue is quiesced, we may be switching io scheduler, or
1482          * updating nr_hw_queues, or other things, and we can't run queue
1483          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1484          *
1485          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1486          * quiesced.
1487          */
1488         hctx_lock(hctx, &srcu_idx);
1489         need_run = !blk_queue_quiesced(hctx->queue) &&
1490                 blk_mq_hctx_has_pending(hctx);
1491         hctx_unlock(hctx, srcu_idx);
1492
1493         if (need_run) {
1494                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1495                 return true;
1496         }
1497
1498         return false;
1499 }
1500 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1501
1502 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1503 {
1504         struct blk_mq_hw_ctx *hctx;
1505         int i;
1506
1507         queue_for_each_hw_ctx(q, hctx, i) {
1508                 if (blk_mq_hctx_stopped(hctx))
1509                         continue;
1510
1511                 blk_mq_run_hw_queue(hctx, async);
1512         }
1513 }
1514 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1515
1516 /**
1517  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1518  * @q: request queue.
1519  *
1520  * The caller is responsible for serializing this function against
1521  * blk_mq_{start,stop}_hw_queue().
1522  */
1523 bool blk_mq_queue_stopped(struct request_queue *q)
1524 {
1525         struct blk_mq_hw_ctx *hctx;
1526         int i;
1527
1528         queue_for_each_hw_ctx(q, hctx, i)
1529                 if (blk_mq_hctx_stopped(hctx))
1530                         return true;
1531
1532         return false;
1533 }
1534 EXPORT_SYMBOL(blk_mq_queue_stopped);
1535
1536 /*
1537  * This function is often used for pausing .queue_rq() by driver when
1538  * there isn't enough resource or some conditions aren't satisfied, and
1539  * BLK_STS_RESOURCE is usually returned.
1540  *
1541  * We do not guarantee that dispatch can be drained or blocked
1542  * after blk_mq_stop_hw_queue() returns. Please use
1543  * blk_mq_quiesce_queue() for that requirement.
1544  */
1545 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1546 {
1547         cancel_delayed_work(&hctx->run_work);
1548
1549         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1550 }
1551 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1552
1553 /*
1554  * This function is often used for pausing .queue_rq() by driver when
1555  * there isn't enough resource or some conditions aren't satisfied, and
1556  * BLK_STS_RESOURCE is usually returned.
1557  *
1558  * We do not guarantee that dispatch can be drained or blocked
1559  * after blk_mq_stop_hw_queues() returns. Please use
1560  * blk_mq_quiesce_queue() for that requirement.
1561  */
1562 void blk_mq_stop_hw_queues(struct request_queue *q)
1563 {
1564         struct blk_mq_hw_ctx *hctx;
1565         int i;
1566
1567         queue_for_each_hw_ctx(q, hctx, i)
1568                 blk_mq_stop_hw_queue(hctx);
1569 }
1570 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1571
1572 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1573 {
1574         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1575
1576         blk_mq_run_hw_queue(hctx, false);
1577 }
1578 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1579
1580 void blk_mq_start_hw_queues(struct request_queue *q)
1581 {
1582         struct blk_mq_hw_ctx *hctx;
1583         int i;
1584
1585         queue_for_each_hw_ctx(q, hctx, i)
1586                 blk_mq_start_hw_queue(hctx);
1587 }
1588 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1589
1590 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1591 {
1592         if (!blk_mq_hctx_stopped(hctx))
1593                 return;
1594
1595         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1596         blk_mq_run_hw_queue(hctx, async);
1597 }
1598 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1599
1600 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1601 {
1602         struct blk_mq_hw_ctx *hctx;
1603         int i;
1604
1605         queue_for_each_hw_ctx(q, hctx, i)
1606                 blk_mq_start_stopped_hw_queue(hctx, async);
1607 }
1608 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1609
1610 static void blk_mq_run_work_fn(struct work_struct *work)
1611 {
1612         struct blk_mq_hw_ctx *hctx;
1613
1614         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1615
1616         /*
1617          * If we are stopped, don't run the queue.
1618          */
1619         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1620                 return;
1621
1622         __blk_mq_run_hw_queue(hctx);
1623 }
1624
1625 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1626                                             struct request *rq,
1627                                             bool at_head)
1628 {
1629         struct blk_mq_ctx *ctx = rq->mq_ctx;
1630         enum hctx_type type = hctx->type;
1631
1632         lockdep_assert_held(&ctx->lock);
1633
1634         trace_block_rq_insert(hctx->queue, rq);
1635
1636         if (at_head)
1637                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1638         else
1639                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1640 }
1641
1642 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1643                              bool at_head)
1644 {
1645         struct blk_mq_ctx *ctx = rq->mq_ctx;
1646
1647         lockdep_assert_held(&ctx->lock);
1648
1649         __blk_mq_insert_req_list(hctx, rq, at_head);
1650         blk_mq_hctx_mark_pending(hctx, ctx);
1651 }
1652
1653 /*
1654  * Should only be used carefully, when the caller knows we want to
1655  * bypass a potential IO scheduler on the target device.
1656  */
1657 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1658 {
1659         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1660
1661         spin_lock(&hctx->lock);
1662         list_add_tail(&rq->queuelist, &hctx->dispatch);
1663         spin_unlock(&hctx->lock);
1664
1665         if (run_queue)
1666                 blk_mq_run_hw_queue(hctx, false);
1667 }
1668
1669 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1670                             struct list_head *list)
1671
1672 {
1673         struct request *rq;
1674         enum hctx_type type = hctx->type;
1675
1676         /*
1677          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1678          * offline now
1679          */
1680         list_for_each_entry(rq, list, queuelist) {
1681                 BUG_ON(rq->mq_ctx != ctx);
1682                 trace_block_rq_insert(hctx->queue, rq);
1683         }
1684
1685         spin_lock(&ctx->lock);
1686         list_splice_tail_init(list, &ctx->rq_lists[type]);
1687         blk_mq_hctx_mark_pending(hctx, ctx);
1688         spin_unlock(&ctx->lock);
1689 }
1690
1691 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1692 {
1693         struct request *rqa = container_of(a, struct request, queuelist);
1694         struct request *rqb = container_of(b, struct request, queuelist);
1695
1696         if (rqa->mq_ctx < rqb->mq_ctx)
1697                 return -1;
1698         else if (rqa->mq_ctx > rqb->mq_ctx)
1699                 return 1;
1700         else if (rqa->mq_hctx < rqb->mq_hctx)
1701                 return -1;
1702         else if (rqa->mq_hctx > rqb->mq_hctx)
1703                 return 1;
1704
1705         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1706 }
1707
1708 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1709 {
1710         struct blk_mq_hw_ctx *this_hctx;
1711         struct blk_mq_ctx *this_ctx;
1712         struct request_queue *this_q;
1713         struct request *rq;
1714         LIST_HEAD(list);
1715         LIST_HEAD(rq_list);
1716         unsigned int depth;
1717
1718         list_splice_init(&plug->mq_list, &list);
1719
1720         if (plug->rq_count > 2 && plug->multiple_queues)
1721                 list_sort(NULL, &list, plug_rq_cmp);
1722
1723         plug->rq_count = 0;
1724
1725         this_q = NULL;
1726         this_hctx = NULL;
1727         this_ctx = NULL;
1728         depth = 0;
1729
1730         while (!list_empty(&list)) {
1731                 rq = list_entry_rq(list.next);
1732                 list_del_init(&rq->queuelist);
1733                 BUG_ON(!rq->q);
1734                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1735                         if (this_hctx) {
1736                                 trace_block_unplug(this_q, depth, !from_schedule);
1737                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1738                                                                 &rq_list,
1739                                                                 from_schedule);
1740                         }
1741
1742                         this_q = rq->q;
1743                         this_ctx = rq->mq_ctx;
1744                         this_hctx = rq->mq_hctx;
1745                         depth = 0;
1746                 }
1747
1748                 depth++;
1749                 list_add_tail(&rq->queuelist, &rq_list);
1750         }
1751
1752         /*
1753          * If 'this_hctx' is set, we know we have entries to complete
1754          * on 'rq_list'. Do those.
1755          */
1756         if (this_hctx) {
1757                 trace_block_unplug(this_q, depth, !from_schedule);
1758                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1759                                                 from_schedule);
1760         }
1761 }
1762
1763 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1764                 unsigned int nr_segs)
1765 {
1766         if (bio->bi_opf & REQ_RAHEAD)
1767                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1768
1769         rq->__sector = bio->bi_iter.bi_sector;
1770         rq->write_hint = bio->bi_write_hint;
1771         blk_rq_bio_prep(rq, bio, nr_segs);
1772
1773         blk_account_io_start(rq, true);
1774 }
1775
1776 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1777                                             struct request *rq,
1778                                             blk_qc_t *cookie, bool last)
1779 {
1780         struct request_queue *q = rq->q;
1781         struct blk_mq_queue_data bd = {
1782                 .rq = rq,
1783                 .last = last,
1784         };
1785         blk_qc_t new_cookie;
1786         blk_status_t ret;
1787
1788         new_cookie = request_to_qc_t(hctx, rq);
1789
1790         /*
1791          * For OK queue, we are done. For error, caller may kill it.
1792          * Any other error (busy), just add it to our list as we
1793          * previously would have done.
1794          */
1795         ret = q->mq_ops->queue_rq(hctx, &bd);
1796         switch (ret) {
1797         case BLK_STS_OK:
1798                 blk_mq_update_dispatch_busy(hctx, false);
1799                 *cookie = new_cookie;
1800                 break;
1801         case BLK_STS_RESOURCE:
1802         case BLK_STS_DEV_RESOURCE:
1803                 blk_mq_update_dispatch_busy(hctx, true);
1804                 __blk_mq_requeue_request(rq);
1805                 break;
1806         default:
1807                 blk_mq_update_dispatch_busy(hctx, false);
1808                 *cookie = BLK_QC_T_NONE;
1809                 break;
1810         }
1811
1812         return ret;
1813 }
1814
1815 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1816                                                 struct request *rq,
1817                                                 blk_qc_t *cookie,
1818                                                 bool bypass_insert, bool last)
1819 {
1820         struct request_queue *q = rq->q;
1821         bool run_queue = true;
1822
1823         /*
1824          * RCU or SRCU read lock is needed before checking quiesced flag.
1825          *
1826          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1827          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1828          * and avoid driver to try to dispatch again.
1829          */
1830         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1831                 run_queue = false;
1832                 bypass_insert = false;
1833                 goto insert;
1834         }
1835
1836         if (q->elevator && !bypass_insert)
1837                 goto insert;
1838
1839         if (!blk_mq_get_dispatch_budget(hctx))
1840                 goto insert;
1841
1842         if (!blk_mq_get_driver_tag(rq)) {
1843                 blk_mq_put_dispatch_budget(hctx);
1844                 goto insert;
1845         }
1846
1847         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1848 insert:
1849         if (bypass_insert)
1850                 return BLK_STS_RESOURCE;
1851
1852         blk_mq_request_bypass_insert(rq, run_queue);
1853         return BLK_STS_OK;
1854 }
1855
1856 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1857                 struct request *rq, blk_qc_t *cookie)
1858 {
1859         blk_status_t ret;
1860         int srcu_idx;
1861
1862         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1863
1864         hctx_lock(hctx, &srcu_idx);
1865
1866         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1867         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1868                 blk_mq_request_bypass_insert(rq, true);
1869         else if (ret != BLK_STS_OK)
1870                 blk_mq_end_request(rq, ret);
1871
1872         hctx_unlock(hctx, srcu_idx);
1873 }
1874
1875 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1876 {
1877         blk_status_t ret;
1878         int srcu_idx;
1879         blk_qc_t unused_cookie;
1880         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1881
1882         hctx_lock(hctx, &srcu_idx);
1883         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1884         hctx_unlock(hctx, srcu_idx);
1885
1886         return ret;
1887 }
1888
1889 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1890                 struct list_head *list)
1891 {
1892         while (!list_empty(list)) {
1893                 blk_status_t ret;
1894                 struct request *rq = list_first_entry(list, struct request,
1895                                 queuelist);
1896
1897                 list_del_init(&rq->queuelist);
1898                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1899                 if (ret != BLK_STS_OK) {
1900                         if (ret == BLK_STS_RESOURCE ||
1901                                         ret == BLK_STS_DEV_RESOURCE) {
1902                                 blk_mq_request_bypass_insert(rq,
1903                                                         list_empty(list));
1904                                 break;
1905                         }
1906                         blk_mq_end_request(rq, ret);
1907                 }
1908         }
1909
1910         /*
1911          * If we didn't flush the entire list, we could have told
1912          * the driver there was more coming, but that turned out to
1913          * be a lie.
1914          */
1915         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1916                 hctx->queue->mq_ops->commit_rqs(hctx);
1917 }
1918
1919 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1920 {
1921         list_add_tail(&rq->queuelist, &plug->mq_list);
1922         plug->rq_count++;
1923         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1924                 struct request *tmp;
1925
1926                 tmp = list_first_entry(&plug->mq_list, struct request,
1927                                                 queuelist);
1928                 if (tmp->q != rq->q)
1929                         plug->multiple_queues = true;
1930         }
1931 }
1932
1933 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1934 {
1935         const int is_sync = op_is_sync(bio->bi_opf);
1936         const int is_flush_fua = op_is_flush(bio->bi_opf);
1937         struct blk_mq_alloc_data data = { .flags = 0};
1938         struct request *rq;
1939         struct blk_plug *plug;
1940         struct request *same_queue_rq = NULL;
1941         unsigned int nr_segs;
1942         blk_qc_t cookie;
1943
1944         blk_queue_bounce(q, &bio);
1945         __blk_queue_split(q, &bio, &nr_segs);
1946
1947         if (!bio_integrity_prep(bio))
1948                 return BLK_QC_T_NONE;
1949
1950         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1951             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1952                 return BLK_QC_T_NONE;
1953
1954         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1955                 return BLK_QC_T_NONE;
1956
1957         rq_qos_throttle(q, bio);
1958
1959         data.cmd_flags = bio->bi_opf;
1960         rq = blk_mq_get_request(q, bio, &data);
1961         if (unlikely(!rq)) {
1962                 rq_qos_cleanup(q, bio);
1963                 if (bio->bi_opf & REQ_NOWAIT)
1964                         bio_wouldblock_error(bio);
1965                 return BLK_QC_T_NONE;
1966         }
1967
1968         trace_block_getrq(q, bio, bio->bi_opf);
1969
1970         rq_qos_track(q, rq, bio);
1971
1972         cookie = request_to_qc_t(data.hctx, rq);
1973
1974         plug = current->plug;
1975         if (unlikely(is_flush_fua)) {
1976                 blk_mq_bio_to_request(rq, bio, nr_segs);
1977
1978                 /* bypass scheduler for flush rq */
1979                 blk_insert_flush(rq);
1980                 blk_mq_run_hw_queue(data.hctx, true);
1981         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1982                 /*
1983                  * Use plugging if we have a ->commit_rqs() hook as well, as
1984                  * we know the driver uses bd->last in a smart fashion.
1985                  */
1986                 unsigned int request_count = plug->rq_count;
1987                 struct request *last = NULL;
1988
1989                 blk_mq_bio_to_request(rq, bio, nr_segs);
1990
1991                 if (!request_count)
1992                         trace_block_plug(q);
1993                 else
1994                         last = list_entry_rq(plug->mq_list.prev);
1995
1996                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1997                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1998                         blk_flush_plug_list(plug, false);
1999                         trace_block_plug(q);
2000                 }
2001
2002                 blk_add_rq_to_plug(plug, rq);
2003         } else if (plug && !blk_queue_nomerges(q)) {
2004                 blk_mq_bio_to_request(rq, bio, nr_segs);
2005
2006                 /*
2007                  * We do limited plugging. If the bio can be merged, do that.
2008                  * Otherwise the existing request in the plug list will be
2009                  * issued. So the plug list will have one request at most
2010                  * The plug list might get flushed before this. If that happens,
2011                  * the plug list is empty, and same_queue_rq is invalid.
2012                  */
2013                 if (list_empty(&plug->mq_list))
2014                         same_queue_rq = NULL;
2015                 if (same_queue_rq) {
2016                         list_del_init(&same_queue_rq->queuelist);
2017                         plug->rq_count--;
2018                 }
2019                 blk_add_rq_to_plug(plug, rq);
2020                 trace_block_plug(q);
2021
2022                 if (same_queue_rq) {
2023                         data.hctx = same_queue_rq->mq_hctx;
2024                         trace_block_unplug(q, 1, true);
2025                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2026                                         &cookie);
2027                 }
2028         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2029                         !data.hctx->dispatch_busy)) {
2030                 blk_mq_bio_to_request(rq, bio, nr_segs);
2031                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2032         } else {
2033                 blk_mq_bio_to_request(rq, bio, nr_segs);
2034                 blk_mq_sched_insert_request(rq, false, true, true);
2035         }
2036
2037         return cookie;
2038 }
2039
2040 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2041                      unsigned int hctx_idx)
2042 {
2043         struct page *page;
2044
2045         if (tags->rqs && set->ops->exit_request) {
2046                 int i;
2047
2048                 for (i = 0; i < tags->nr_tags; i++) {
2049                         struct request *rq = tags->static_rqs[i];
2050
2051                         if (!rq)
2052                                 continue;
2053                         set->ops->exit_request(set, rq, hctx_idx);
2054                         tags->static_rqs[i] = NULL;
2055                 }
2056         }
2057
2058         while (!list_empty(&tags->page_list)) {
2059                 page = list_first_entry(&tags->page_list, struct page, lru);
2060                 list_del_init(&page->lru);
2061                 /*
2062                  * Remove kmemleak object previously allocated in
2063                  * blk_mq_alloc_rqs().
2064                  */
2065                 kmemleak_free(page_address(page));
2066                 __free_pages(page, page->private);
2067         }
2068 }
2069
2070 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2071 {
2072         kfree(tags->rqs);
2073         tags->rqs = NULL;
2074         kfree(tags->static_rqs);
2075         tags->static_rqs = NULL;
2076
2077         blk_mq_free_tags(tags);
2078 }
2079
2080 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2081                                         unsigned int hctx_idx,
2082                                         unsigned int nr_tags,
2083                                         unsigned int reserved_tags)
2084 {
2085         struct blk_mq_tags *tags;
2086         int node;
2087
2088         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2089         if (node == NUMA_NO_NODE)
2090                 node = set->numa_node;
2091
2092         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2093                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2094         if (!tags)
2095                 return NULL;
2096
2097         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2098                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2099                                  node);
2100         if (!tags->rqs) {
2101                 blk_mq_free_tags(tags);
2102                 return NULL;
2103         }
2104
2105         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2106                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2107                                         node);
2108         if (!tags->static_rqs) {
2109                 kfree(tags->rqs);
2110                 blk_mq_free_tags(tags);
2111                 return NULL;
2112         }
2113
2114         return tags;
2115 }
2116
2117 static size_t order_to_size(unsigned int order)
2118 {
2119         return (size_t)PAGE_SIZE << order;
2120 }
2121
2122 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2123                                unsigned int hctx_idx, int node)
2124 {
2125         int ret;
2126
2127         if (set->ops->init_request) {
2128                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2129                 if (ret)
2130                         return ret;
2131         }
2132
2133         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2134         return 0;
2135 }
2136
2137 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2138                      unsigned int hctx_idx, unsigned int depth)
2139 {
2140         unsigned int i, j, entries_per_page, max_order = 4;
2141         size_t rq_size, left;
2142         int node;
2143
2144         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2145         if (node == NUMA_NO_NODE)
2146                 node = set->numa_node;
2147
2148         INIT_LIST_HEAD(&tags->page_list);
2149
2150         /*
2151          * rq_size is the size of the request plus driver payload, rounded
2152          * to the cacheline size
2153          */
2154         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2155                                 cache_line_size());
2156         left = rq_size * depth;
2157
2158         for (i = 0; i < depth; ) {
2159                 int this_order = max_order;
2160                 struct page *page;
2161                 int to_do;
2162                 void *p;
2163
2164                 while (this_order && left < order_to_size(this_order - 1))
2165                         this_order--;
2166
2167                 do {
2168                         page = alloc_pages_node(node,
2169                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2170                                 this_order);
2171                         if (page)
2172                                 break;
2173                         if (!this_order--)
2174                                 break;
2175                         if (order_to_size(this_order) < rq_size)
2176                                 break;
2177                 } while (1);
2178
2179                 if (!page)
2180                         goto fail;
2181
2182                 page->private = this_order;
2183                 list_add_tail(&page->lru, &tags->page_list);
2184
2185                 p = page_address(page);
2186                 /*
2187                  * Allow kmemleak to scan these pages as they contain pointers
2188                  * to additional allocations like via ops->init_request().
2189                  */
2190                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2191                 entries_per_page = order_to_size(this_order) / rq_size;
2192                 to_do = min(entries_per_page, depth - i);
2193                 left -= to_do * rq_size;
2194                 for (j = 0; j < to_do; j++) {
2195                         struct request *rq = p;
2196
2197                         tags->static_rqs[i] = rq;
2198                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2199                                 tags->static_rqs[i] = NULL;
2200                                 goto fail;
2201                         }
2202
2203                         p += rq_size;
2204                         i++;
2205                 }
2206         }
2207         return 0;
2208
2209 fail:
2210         blk_mq_free_rqs(set, tags, hctx_idx);
2211         return -ENOMEM;
2212 }
2213
2214 /*
2215  * 'cpu' is going away. splice any existing rq_list entries from this
2216  * software queue to the hw queue dispatch list, and ensure that it
2217  * gets run.
2218  */
2219 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2220 {
2221         struct blk_mq_hw_ctx *hctx;
2222         struct blk_mq_ctx *ctx;
2223         LIST_HEAD(tmp);
2224         enum hctx_type type;
2225
2226         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2227         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2228         type = hctx->type;
2229
2230         spin_lock(&ctx->lock);
2231         if (!list_empty(&ctx->rq_lists[type])) {
2232                 list_splice_init(&ctx->rq_lists[type], &tmp);
2233                 blk_mq_hctx_clear_pending(hctx, ctx);
2234         }
2235         spin_unlock(&ctx->lock);
2236
2237         if (list_empty(&tmp))
2238                 return 0;
2239
2240         spin_lock(&hctx->lock);
2241         list_splice_tail_init(&tmp, &hctx->dispatch);
2242         spin_unlock(&hctx->lock);
2243
2244         blk_mq_run_hw_queue(hctx, true);
2245         return 0;
2246 }
2247
2248 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2249 {
2250         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2251                                             &hctx->cpuhp_dead);
2252 }
2253
2254 /* hctx->ctxs will be freed in queue's release handler */
2255 static void blk_mq_exit_hctx(struct request_queue *q,
2256                 struct blk_mq_tag_set *set,
2257                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2258 {
2259         if (blk_mq_hw_queue_mapped(hctx))
2260                 blk_mq_tag_idle(hctx);
2261
2262         if (set->ops->exit_request)
2263                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2264
2265         if (set->ops->exit_hctx)
2266                 set->ops->exit_hctx(hctx, hctx_idx);
2267
2268         blk_mq_remove_cpuhp(hctx);
2269
2270         spin_lock(&q->unused_hctx_lock);
2271         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2272         spin_unlock(&q->unused_hctx_lock);
2273 }
2274
2275 static void blk_mq_exit_hw_queues(struct request_queue *q,
2276                 struct blk_mq_tag_set *set, int nr_queue)
2277 {
2278         struct blk_mq_hw_ctx *hctx;
2279         unsigned int i;
2280
2281         queue_for_each_hw_ctx(q, hctx, i) {
2282                 if (i == nr_queue)
2283                         break;
2284                 blk_mq_debugfs_unregister_hctx(hctx);
2285                 blk_mq_exit_hctx(q, set, hctx, i);
2286         }
2287 }
2288
2289 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2290 {
2291         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2292
2293         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2294                            __alignof__(struct blk_mq_hw_ctx)) !=
2295                      sizeof(struct blk_mq_hw_ctx));
2296
2297         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2298                 hw_ctx_size += sizeof(struct srcu_struct);
2299
2300         return hw_ctx_size;
2301 }
2302
2303 static int blk_mq_init_hctx(struct request_queue *q,
2304                 struct blk_mq_tag_set *set,
2305                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2306 {
2307         hctx->queue_num = hctx_idx;
2308
2309         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2310
2311         hctx->tags = set->tags[hctx_idx];
2312
2313         if (set->ops->init_hctx &&
2314             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2315                 goto unregister_cpu_notifier;
2316
2317         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2318                                 hctx->numa_node))
2319                 goto exit_hctx;
2320         return 0;
2321
2322  exit_hctx:
2323         if (set->ops->exit_hctx)
2324                 set->ops->exit_hctx(hctx, hctx_idx);
2325  unregister_cpu_notifier:
2326         blk_mq_remove_cpuhp(hctx);
2327         return -1;
2328 }
2329
2330 static struct blk_mq_hw_ctx *
2331 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2332                 int node)
2333 {
2334         struct blk_mq_hw_ctx *hctx;
2335         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2336
2337         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2338         if (!hctx)
2339                 goto fail_alloc_hctx;
2340
2341         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2342                 goto free_hctx;
2343
2344         atomic_set(&hctx->nr_active, 0);
2345         if (node == NUMA_NO_NODE)
2346                 node = set->numa_node;
2347         hctx->numa_node = node;
2348
2349         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2350         spin_lock_init(&hctx->lock);
2351         INIT_LIST_HEAD(&hctx->dispatch);
2352         hctx->queue = q;
2353         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2354
2355         INIT_LIST_HEAD(&hctx->hctx_list);
2356
2357         /*
2358          * Allocate space for all possible cpus to avoid allocation at
2359          * runtime
2360          */
2361         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2362                         gfp, node);
2363         if (!hctx->ctxs)
2364                 goto free_cpumask;
2365
2366         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2367                                 gfp, node))
2368                 goto free_ctxs;
2369         hctx->nr_ctx = 0;
2370
2371         spin_lock_init(&hctx->dispatch_wait_lock);
2372         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2373         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2374
2375         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2376                         gfp);
2377         if (!hctx->fq)
2378                 goto free_bitmap;
2379
2380         if (hctx->flags & BLK_MQ_F_BLOCKING)
2381                 init_srcu_struct(hctx->srcu);
2382         blk_mq_hctx_kobj_init(hctx);
2383
2384         return hctx;
2385
2386  free_bitmap:
2387         sbitmap_free(&hctx->ctx_map);
2388  free_ctxs:
2389         kfree(hctx->ctxs);
2390  free_cpumask:
2391         free_cpumask_var(hctx->cpumask);
2392  free_hctx:
2393         kfree(hctx);
2394  fail_alloc_hctx:
2395         return NULL;
2396 }
2397
2398 static void blk_mq_init_cpu_queues(struct request_queue *q,
2399                                    unsigned int nr_hw_queues)
2400 {
2401         struct blk_mq_tag_set *set = q->tag_set;
2402         unsigned int i, j;
2403
2404         for_each_possible_cpu(i) {
2405                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2406                 struct blk_mq_hw_ctx *hctx;
2407                 int k;
2408
2409                 __ctx->cpu = i;
2410                 spin_lock_init(&__ctx->lock);
2411                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2412                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2413
2414                 __ctx->queue = q;
2415
2416                 /*
2417                  * Set local node, IFF we have more than one hw queue. If
2418                  * not, we remain on the home node of the device
2419                  */
2420                 for (j = 0; j < set->nr_maps; j++) {
2421                         hctx = blk_mq_map_queue_type(q, j, i);
2422                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2423                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2424                 }
2425         }
2426 }
2427
2428 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2429 {
2430         int ret = 0;
2431
2432         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2433                                         set->queue_depth, set->reserved_tags);
2434         if (!set->tags[hctx_idx])
2435                 return false;
2436
2437         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2438                                 set->queue_depth);
2439         if (!ret)
2440                 return true;
2441
2442         blk_mq_free_rq_map(set->tags[hctx_idx]);
2443         set->tags[hctx_idx] = NULL;
2444         return false;
2445 }
2446
2447 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2448                                          unsigned int hctx_idx)
2449 {
2450         if (set->tags && set->tags[hctx_idx]) {
2451                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2452                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2453                 set->tags[hctx_idx] = NULL;
2454         }
2455 }
2456
2457 static void blk_mq_map_swqueue(struct request_queue *q)
2458 {
2459         unsigned int i, j, hctx_idx;
2460         struct blk_mq_hw_ctx *hctx;
2461         struct blk_mq_ctx *ctx;
2462         struct blk_mq_tag_set *set = q->tag_set;
2463
2464         /*
2465          * Avoid others reading imcomplete hctx->cpumask through sysfs
2466          */
2467         mutex_lock(&q->sysfs_lock);
2468
2469         queue_for_each_hw_ctx(q, hctx, i) {
2470                 cpumask_clear(hctx->cpumask);
2471                 hctx->nr_ctx = 0;
2472                 hctx->dispatch_from = NULL;
2473         }
2474
2475         /*
2476          * Map software to hardware queues.
2477          *
2478          * If the cpu isn't present, the cpu is mapped to first hctx.
2479          */
2480         for_each_possible_cpu(i) {
2481                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2482                 /* unmapped hw queue can be remapped after CPU topo changed */
2483                 if (!set->tags[hctx_idx] &&
2484                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2485                         /*
2486                          * If tags initialization fail for some hctx,
2487                          * that hctx won't be brought online.  In this
2488                          * case, remap the current ctx to hctx[0] which
2489                          * is guaranteed to always have tags allocated
2490                          */
2491                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2492                 }
2493
2494                 ctx = per_cpu_ptr(q->queue_ctx, i);
2495                 for (j = 0; j < set->nr_maps; j++) {
2496                         if (!set->map[j].nr_queues) {
2497                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2498                                                 HCTX_TYPE_DEFAULT, i);
2499                                 continue;
2500                         }
2501
2502                         hctx = blk_mq_map_queue_type(q, j, i);
2503                         ctx->hctxs[j] = hctx;
2504                         /*
2505                          * If the CPU is already set in the mask, then we've
2506                          * mapped this one already. This can happen if
2507                          * devices share queues across queue maps.
2508                          */
2509                         if (cpumask_test_cpu(i, hctx->cpumask))
2510                                 continue;
2511
2512                         cpumask_set_cpu(i, hctx->cpumask);
2513                         hctx->type = j;
2514                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2515                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2516
2517                         /*
2518                          * If the nr_ctx type overflows, we have exceeded the
2519                          * amount of sw queues we can support.
2520                          */
2521                         BUG_ON(!hctx->nr_ctx);
2522                 }
2523
2524                 for (; j < HCTX_MAX_TYPES; j++)
2525                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2526                                         HCTX_TYPE_DEFAULT, i);
2527         }
2528
2529         mutex_unlock(&q->sysfs_lock);
2530
2531         queue_for_each_hw_ctx(q, hctx, i) {
2532                 /*
2533                  * If no software queues are mapped to this hardware queue,
2534                  * disable it and free the request entries.
2535                  */
2536                 if (!hctx->nr_ctx) {
2537                         /* Never unmap queue 0.  We need it as a
2538                          * fallback in case of a new remap fails
2539                          * allocation
2540                          */
2541                         if (i && set->tags[i])
2542                                 blk_mq_free_map_and_requests(set, i);
2543
2544                         hctx->tags = NULL;
2545                         continue;
2546                 }
2547
2548                 hctx->tags = set->tags[i];
2549                 WARN_ON(!hctx->tags);
2550
2551                 /*
2552                  * Set the map size to the number of mapped software queues.
2553                  * This is more accurate and more efficient than looping
2554                  * over all possibly mapped software queues.
2555                  */
2556                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2557
2558                 /*
2559                  * Initialize batch roundrobin counts
2560                  */
2561                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2562                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2563         }
2564 }
2565
2566 /*
2567  * Caller needs to ensure that we're either frozen/quiesced, or that
2568  * the queue isn't live yet.
2569  */
2570 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2571 {
2572         struct blk_mq_hw_ctx *hctx;
2573         int i;
2574
2575         queue_for_each_hw_ctx(q, hctx, i) {
2576                 if (shared)
2577                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2578                 else
2579                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2580         }
2581 }
2582
2583 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2584                                         bool shared)
2585 {
2586         struct request_queue *q;
2587
2588         lockdep_assert_held(&set->tag_list_lock);
2589
2590         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2591                 blk_mq_freeze_queue(q);
2592                 queue_set_hctx_shared(q, shared);
2593                 blk_mq_unfreeze_queue(q);
2594         }
2595 }
2596
2597 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2598 {
2599         struct blk_mq_tag_set *set = q->tag_set;
2600
2601         mutex_lock(&set->tag_list_lock);
2602         list_del_rcu(&q->tag_set_list);
2603         if (list_is_singular(&set->tag_list)) {
2604                 /* just transitioned to unshared */
2605                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2606                 /* update existing queue */
2607                 blk_mq_update_tag_set_depth(set, false);
2608         }
2609         mutex_unlock(&set->tag_list_lock);
2610         INIT_LIST_HEAD(&q->tag_set_list);
2611 }
2612
2613 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2614                                      struct request_queue *q)
2615 {
2616         mutex_lock(&set->tag_list_lock);
2617
2618         /*
2619          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2620          */
2621         if (!list_empty(&set->tag_list) &&
2622             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2623                 set->flags |= BLK_MQ_F_TAG_SHARED;
2624                 /* update existing queue */
2625                 blk_mq_update_tag_set_depth(set, true);
2626         }
2627         if (set->flags & BLK_MQ_F_TAG_SHARED)
2628                 queue_set_hctx_shared(q, true);
2629         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2630
2631         mutex_unlock(&set->tag_list_lock);
2632 }
2633
2634 /* All allocations will be freed in release handler of q->mq_kobj */
2635 static int blk_mq_alloc_ctxs(struct request_queue *q)
2636 {
2637         struct blk_mq_ctxs *ctxs;
2638         int cpu;
2639
2640         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2641         if (!ctxs)
2642                 return -ENOMEM;
2643
2644         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2645         if (!ctxs->queue_ctx)
2646                 goto fail;
2647
2648         for_each_possible_cpu(cpu) {
2649                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2650                 ctx->ctxs = ctxs;
2651         }
2652
2653         q->mq_kobj = &ctxs->kobj;
2654         q->queue_ctx = ctxs->queue_ctx;
2655
2656         return 0;
2657  fail:
2658         kfree(ctxs);
2659         return -ENOMEM;
2660 }
2661
2662 /*
2663  * It is the actual release handler for mq, but we do it from
2664  * request queue's release handler for avoiding use-after-free
2665  * and headache because q->mq_kobj shouldn't have been introduced,
2666  * but we can't group ctx/kctx kobj without it.
2667  */
2668 void blk_mq_release(struct request_queue *q)
2669 {
2670         struct blk_mq_hw_ctx *hctx, *next;
2671         int i;
2672
2673         cancel_delayed_work_sync(&q->requeue_work);
2674
2675         queue_for_each_hw_ctx(q, hctx, i)
2676                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2677
2678         /* all hctx are in .unused_hctx_list now */
2679         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2680                 list_del_init(&hctx->hctx_list);
2681                 kobject_put(&hctx->kobj);
2682         }
2683
2684         kfree(q->queue_hw_ctx);
2685
2686         /*
2687          * release .mq_kobj and sw queue's kobject now because
2688          * both share lifetime with request queue.
2689          */
2690         blk_mq_sysfs_deinit(q);
2691 }
2692
2693 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2694 {
2695         struct request_queue *uninit_q, *q;
2696
2697         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2698         if (!uninit_q)
2699                 return ERR_PTR(-ENOMEM);
2700
2701         q = blk_mq_init_allocated_queue(set, uninit_q);
2702         if (IS_ERR(q))
2703                 blk_cleanup_queue(uninit_q);
2704
2705         return q;
2706 }
2707 EXPORT_SYMBOL(blk_mq_init_queue);
2708
2709 /*
2710  * Helper for setting up a queue with mq ops, given queue depth, and
2711  * the passed in mq ops flags.
2712  */
2713 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2714                                            const struct blk_mq_ops *ops,
2715                                            unsigned int queue_depth,
2716                                            unsigned int set_flags)
2717 {
2718         struct request_queue *q;
2719         int ret;
2720
2721         memset(set, 0, sizeof(*set));
2722         set->ops = ops;
2723         set->nr_hw_queues = 1;
2724         set->nr_maps = 1;
2725         set->queue_depth = queue_depth;
2726         set->numa_node = NUMA_NO_NODE;
2727         set->flags = set_flags;
2728
2729         ret = blk_mq_alloc_tag_set(set);
2730         if (ret)
2731                 return ERR_PTR(ret);
2732
2733         q = blk_mq_init_queue(set);
2734         if (IS_ERR(q)) {
2735                 blk_mq_free_tag_set(set);
2736                 return q;
2737         }
2738
2739         return q;
2740 }
2741 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2742
2743 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2744                 struct blk_mq_tag_set *set, struct request_queue *q,
2745                 int hctx_idx, int node)
2746 {
2747         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2748
2749         /* reuse dead hctx first */
2750         spin_lock(&q->unused_hctx_lock);
2751         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2752                 if (tmp->numa_node == node) {
2753                         hctx = tmp;
2754                         break;
2755                 }
2756         }
2757         if (hctx)
2758                 list_del_init(&hctx->hctx_list);
2759         spin_unlock(&q->unused_hctx_lock);
2760
2761         if (!hctx)
2762                 hctx = blk_mq_alloc_hctx(q, set, node);
2763         if (!hctx)
2764                 goto fail;
2765
2766         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2767                 goto free_hctx;
2768
2769         return hctx;
2770
2771  free_hctx:
2772         kobject_put(&hctx->kobj);
2773  fail:
2774         return NULL;
2775 }
2776
2777 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2778                                                 struct request_queue *q)
2779 {
2780         int i, j, end;
2781         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2782
2783         /* protect against switching io scheduler  */
2784         mutex_lock(&q->sysfs_lock);
2785         for (i = 0; i < set->nr_hw_queues; i++) {
2786                 int node;
2787                 struct blk_mq_hw_ctx *hctx;
2788
2789                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2790                 /*
2791                  * If the hw queue has been mapped to another numa node,
2792                  * we need to realloc the hctx. If allocation fails, fallback
2793                  * to use the previous one.
2794                  */
2795                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2796                         continue;
2797
2798                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2799                 if (hctx) {
2800                         if (hctxs[i])
2801                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2802                         hctxs[i] = hctx;
2803                 } else {
2804                         if (hctxs[i])
2805                                 pr_warn("Allocate new hctx on node %d fails,\
2806                                                 fallback to previous one on node %d\n",
2807                                                 node, hctxs[i]->numa_node);
2808                         else
2809                                 break;
2810                 }
2811         }
2812         /*
2813          * Increasing nr_hw_queues fails. Free the newly allocated
2814          * hctxs and keep the previous q->nr_hw_queues.
2815          */
2816         if (i != set->nr_hw_queues) {
2817                 j = q->nr_hw_queues;
2818                 end = i;
2819         } else {
2820                 j = i;
2821                 end = q->nr_hw_queues;
2822                 q->nr_hw_queues = set->nr_hw_queues;
2823         }
2824
2825         for (; j < end; j++) {
2826                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2827
2828                 if (hctx) {
2829                         if (hctx->tags)
2830                                 blk_mq_free_map_and_requests(set, j);
2831                         blk_mq_exit_hctx(q, set, hctx, j);
2832                         hctxs[j] = NULL;
2833                 }
2834         }
2835         mutex_unlock(&q->sysfs_lock);
2836 }
2837
2838 /*
2839  * Maximum number of hardware queues we support. For single sets, we'll never
2840  * have more than the CPUs (software queues). For multiple sets, the tag_set
2841  * user may have set ->nr_hw_queues larger.
2842  */
2843 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2844 {
2845         if (set->nr_maps == 1)
2846                 return nr_cpu_ids;
2847
2848         return max(set->nr_hw_queues, nr_cpu_ids);
2849 }
2850
2851 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2852                                                   struct request_queue *q)
2853 {
2854         /* mark the queue as mq asap */
2855         q->mq_ops = set->ops;
2856
2857         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2858                                              blk_mq_poll_stats_bkt,
2859                                              BLK_MQ_POLL_STATS_BKTS, q);
2860         if (!q->poll_cb)
2861                 goto err_exit;
2862
2863         if (blk_mq_alloc_ctxs(q))
2864                 goto err_poll;
2865
2866         /* init q->mq_kobj and sw queues' kobjects */
2867         blk_mq_sysfs_init(q);
2868
2869         q->nr_queues = nr_hw_queues(set);
2870         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2871                                                 GFP_KERNEL, set->numa_node);
2872         if (!q->queue_hw_ctx)
2873                 goto err_sys_init;
2874
2875         INIT_LIST_HEAD(&q->unused_hctx_list);
2876         spin_lock_init(&q->unused_hctx_lock);
2877
2878         blk_mq_realloc_hw_ctxs(set, q);
2879         if (!q->nr_hw_queues)
2880                 goto err_hctxs;
2881
2882         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2883         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2884
2885         q->tag_set = set;
2886
2887         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2888         if (set->nr_maps > HCTX_TYPE_POLL &&
2889             set->map[HCTX_TYPE_POLL].nr_queues)
2890                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2891
2892         q->sg_reserved_size = INT_MAX;
2893
2894         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2895         INIT_LIST_HEAD(&q->requeue_list);
2896         spin_lock_init(&q->requeue_lock);
2897
2898         blk_queue_make_request(q, blk_mq_make_request);
2899
2900         /*
2901          * Do this after blk_queue_make_request() overrides it...
2902          */
2903         q->nr_requests = set->queue_depth;
2904
2905         /*
2906          * Default to classic polling
2907          */
2908         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2909
2910         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2911         blk_mq_add_queue_tag_set(set, q);
2912         blk_mq_map_swqueue(q);
2913
2914         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2915                 int ret;
2916
2917                 ret = elevator_init_mq(q);
2918                 if (ret)
2919                         return ERR_PTR(ret);
2920         }
2921
2922         return q;
2923
2924 err_hctxs:
2925         kfree(q->queue_hw_ctx);
2926 err_sys_init:
2927         blk_mq_sysfs_deinit(q);
2928 err_poll:
2929         blk_stat_free_callback(q->poll_cb);
2930         q->poll_cb = NULL;
2931 err_exit:
2932         q->mq_ops = NULL;
2933         return ERR_PTR(-ENOMEM);
2934 }
2935 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2936
2937 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2938 void blk_mq_exit_queue(struct request_queue *q)
2939 {
2940         struct blk_mq_tag_set   *set = q->tag_set;
2941
2942         blk_mq_del_queue_tag_set(q);
2943         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2944 }
2945
2946 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2947 {
2948         int i;
2949
2950         for (i = 0; i < set->nr_hw_queues; i++)
2951                 if (!__blk_mq_alloc_rq_map(set, i))
2952                         goto out_unwind;
2953
2954         return 0;
2955
2956 out_unwind:
2957         while (--i >= 0)
2958                 blk_mq_free_rq_map(set->tags[i]);
2959
2960         return -ENOMEM;
2961 }
2962
2963 /*
2964  * Allocate the request maps associated with this tag_set. Note that this
2965  * may reduce the depth asked for, if memory is tight. set->queue_depth
2966  * will be updated to reflect the allocated depth.
2967  */
2968 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2969 {
2970         unsigned int depth;
2971         int err;
2972
2973         depth = set->queue_depth;
2974         do {
2975                 err = __blk_mq_alloc_rq_maps(set);
2976                 if (!err)
2977                         break;
2978
2979                 set->queue_depth >>= 1;
2980                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2981                         err = -ENOMEM;
2982                         break;
2983                 }
2984         } while (set->queue_depth);
2985
2986         if (!set->queue_depth || err) {
2987                 pr_err("blk-mq: failed to allocate request map\n");
2988                 return -ENOMEM;
2989         }
2990
2991         if (depth != set->queue_depth)
2992                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2993                                                 depth, set->queue_depth);
2994
2995         return 0;
2996 }
2997
2998 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2999 {
3000         if (set->ops->map_queues && !is_kdump_kernel()) {
3001                 int i;
3002
3003                 /*
3004                  * transport .map_queues is usually done in the following
3005                  * way:
3006                  *
3007                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3008                  *      mask = get_cpu_mask(queue)
3009                  *      for_each_cpu(cpu, mask)
3010                  *              set->map[x].mq_map[cpu] = queue;
3011                  * }
3012                  *
3013                  * When we need to remap, the table has to be cleared for
3014                  * killing stale mapping since one CPU may not be mapped
3015                  * to any hw queue.
3016                  */
3017                 for (i = 0; i < set->nr_maps; i++)
3018                         blk_mq_clear_mq_map(&set->map[i]);
3019
3020                 return set->ops->map_queues(set);
3021         } else {
3022                 BUG_ON(set->nr_maps > 1);
3023                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3024         }
3025 }
3026
3027 /*
3028  * Alloc a tag set to be associated with one or more request queues.
3029  * May fail with EINVAL for various error conditions. May adjust the
3030  * requested depth down, if it's too large. In that case, the set
3031  * value will be stored in set->queue_depth.
3032  */
3033 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3034 {
3035         int i, ret;
3036
3037         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3038
3039         if (!set->nr_hw_queues)
3040                 return -EINVAL;
3041         if (!set->queue_depth)
3042                 return -EINVAL;
3043         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3044                 return -EINVAL;
3045
3046         if (!set->ops->queue_rq)
3047                 return -EINVAL;
3048
3049         if (!set->ops->get_budget ^ !set->ops->put_budget)
3050                 return -EINVAL;
3051
3052         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3053                 pr_info("blk-mq: reduced tag depth to %u\n",
3054                         BLK_MQ_MAX_DEPTH);
3055                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3056         }
3057
3058         if (!set->nr_maps)
3059                 set->nr_maps = 1;
3060         else if (set->nr_maps > HCTX_MAX_TYPES)
3061                 return -EINVAL;
3062
3063         /*
3064          * If a crashdump is active, then we are potentially in a very
3065          * memory constrained environment. Limit us to 1 queue and
3066          * 64 tags to prevent using too much memory.
3067          */
3068         if (is_kdump_kernel()) {
3069                 set->nr_hw_queues = 1;
3070                 set->nr_maps = 1;
3071                 set->queue_depth = min(64U, set->queue_depth);
3072         }
3073         /*
3074          * There is no use for more h/w queues than cpus if we just have
3075          * a single map
3076          */
3077         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3078                 set->nr_hw_queues = nr_cpu_ids;
3079
3080         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3081                                  GFP_KERNEL, set->numa_node);
3082         if (!set->tags)
3083                 return -ENOMEM;
3084
3085         ret = -ENOMEM;
3086         for (i = 0; i < set->nr_maps; i++) {
3087                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3088                                                   sizeof(set->map[i].mq_map[0]),
3089                                                   GFP_KERNEL, set->numa_node);
3090                 if (!set->map[i].mq_map)
3091                         goto out_free_mq_map;
3092                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3093         }
3094
3095         ret = blk_mq_update_queue_map(set);
3096         if (ret)
3097                 goto out_free_mq_map;
3098
3099         ret = blk_mq_alloc_rq_maps(set);
3100         if (ret)
3101                 goto out_free_mq_map;
3102
3103         mutex_init(&set->tag_list_lock);
3104         INIT_LIST_HEAD(&set->tag_list);
3105
3106         return 0;
3107
3108 out_free_mq_map:
3109         for (i = 0; i < set->nr_maps; i++) {
3110                 kfree(set->map[i].mq_map);
3111                 set->map[i].mq_map = NULL;
3112         }
3113         kfree(set->tags);
3114         set->tags = NULL;
3115         return ret;
3116 }
3117 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3118
3119 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3120 {
3121         int i, j;
3122
3123         for (i = 0; i < nr_hw_queues(set); i++)
3124                 blk_mq_free_map_and_requests(set, i);
3125
3126         for (j = 0; j < set->nr_maps; j++) {
3127                 kfree(set->map[j].mq_map);
3128                 set->map[j].mq_map = NULL;
3129         }
3130
3131         kfree(set->tags);
3132         set->tags = NULL;
3133 }
3134 EXPORT_SYMBOL(blk_mq_free_tag_set);
3135
3136 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3137 {
3138         struct blk_mq_tag_set *set = q->tag_set;
3139         struct blk_mq_hw_ctx *hctx;
3140         int i, ret;
3141
3142         if (!set)
3143                 return -EINVAL;
3144
3145         if (q->nr_requests == nr)
3146                 return 0;
3147
3148         blk_mq_freeze_queue(q);
3149         blk_mq_quiesce_queue(q);
3150
3151         ret = 0;
3152         queue_for_each_hw_ctx(q, hctx, i) {
3153                 if (!hctx->tags)
3154                         continue;
3155                 /*
3156                  * If we're using an MQ scheduler, just update the scheduler
3157                  * queue depth. This is similar to what the old code would do.
3158                  */
3159                 if (!hctx->sched_tags) {
3160                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3161                                                         false);
3162                 } else {
3163                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3164                                                         nr, true);
3165                 }
3166                 if (ret)
3167                         break;
3168                 if (q->elevator && q->elevator->type->ops.depth_updated)
3169                         q->elevator->type->ops.depth_updated(hctx);
3170         }
3171
3172         if (!ret)
3173                 q->nr_requests = nr;
3174
3175         blk_mq_unquiesce_queue(q);
3176         blk_mq_unfreeze_queue(q);
3177
3178         return ret;
3179 }
3180
3181 /*
3182  * request_queue and elevator_type pair.
3183  * It is just used by __blk_mq_update_nr_hw_queues to cache
3184  * the elevator_type associated with a request_queue.
3185  */
3186 struct blk_mq_qe_pair {
3187         struct list_head node;
3188         struct request_queue *q;
3189         struct elevator_type *type;
3190 };
3191
3192 /*
3193  * Cache the elevator_type in qe pair list and switch the
3194  * io scheduler to 'none'
3195  */
3196 static bool blk_mq_elv_switch_none(struct list_head *head,
3197                 struct request_queue *q)
3198 {
3199         struct blk_mq_qe_pair *qe;
3200
3201         if (!q->elevator)
3202                 return true;
3203
3204         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3205         if (!qe)
3206                 return false;
3207
3208         INIT_LIST_HEAD(&qe->node);
3209         qe->q = q;
3210         qe->type = q->elevator->type;
3211         list_add(&qe->node, head);
3212
3213         mutex_lock(&q->sysfs_lock);
3214         /*
3215          * After elevator_switch_mq, the previous elevator_queue will be
3216          * released by elevator_release. The reference of the io scheduler
3217          * module get by elevator_get will also be put. So we need to get
3218          * a reference of the io scheduler module here to prevent it to be
3219          * removed.
3220          */
3221         __module_get(qe->type->elevator_owner);
3222         elevator_switch_mq(q, NULL);
3223         mutex_unlock(&q->sysfs_lock);
3224
3225         return true;
3226 }
3227
3228 static void blk_mq_elv_switch_back(struct list_head *head,
3229                 struct request_queue *q)
3230 {
3231         struct blk_mq_qe_pair *qe;
3232         struct elevator_type *t = NULL;
3233
3234         list_for_each_entry(qe, head, node)
3235                 if (qe->q == q) {
3236                         t = qe->type;
3237                         break;
3238                 }
3239
3240         if (!t)
3241                 return;
3242
3243         list_del(&qe->node);
3244         kfree(qe);
3245
3246         mutex_lock(&q->sysfs_lock);
3247         elevator_switch_mq(q, t);
3248         mutex_unlock(&q->sysfs_lock);
3249 }
3250
3251 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3252                                                         int nr_hw_queues)
3253 {
3254         struct request_queue *q;
3255         LIST_HEAD(head);
3256         int prev_nr_hw_queues;
3257
3258         lockdep_assert_held(&set->tag_list_lock);
3259
3260         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3261                 nr_hw_queues = nr_cpu_ids;
3262         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3263                 return;
3264
3265         list_for_each_entry(q, &set->tag_list, tag_set_list)
3266                 blk_mq_freeze_queue(q);
3267         /*
3268          * Sync with blk_mq_queue_tag_busy_iter.
3269          */
3270         synchronize_rcu();
3271         /*
3272          * Switch IO scheduler to 'none', cleaning up the data associated
3273          * with the previous scheduler. We will switch back once we are done
3274          * updating the new sw to hw queue mappings.
3275          */
3276         list_for_each_entry(q, &set->tag_list, tag_set_list)
3277                 if (!blk_mq_elv_switch_none(&head, q))
3278                         goto switch_back;
3279
3280         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3281                 blk_mq_debugfs_unregister_hctxs(q);
3282                 blk_mq_sysfs_unregister(q);
3283         }
3284
3285         prev_nr_hw_queues = set->nr_hw_queues;
3286         set->nr_hw_queues = nr_hw_queues;
3287         blk_mq_update_queue_map(set);
3288 fallback:
3289         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3290                 blk_mq_realloc_hw_ctxs(set, q);
3291                 if (q->nr_hw_queues != set->nr_hw_queues) {
3292                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3293                                         nr_hw_queues, prev_nr_hw_queues);
3294                         set->nr_hw_queues = prev_nr_hw_queues;
3295                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3296                         goto fallback;
3297                 }
3298                 blk_mq_map_swqueue(q);
3299         }
3300
3301         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3302                 blk_mq_sysfs_register(q);
3303                 blk_mq_debugfs_register_hctxs(q);
3304         }
3305
3306 switch_back:
3307         list_for_each_entry(q, &set->tag_list, tag_set_list)
3308                 blk_mq_elv_switch_back(&head, q);
3309
3310         list_for_each_entry(q, &set->tag_list, tag_set_list)
3311                 blk_mq_unfreeze_queue(q);
3312 }
3313
3314 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3315 {
3316         mutex_lock(&set->tag_list_lock);
3317         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3318         mutex_unlock(&set->tag_list_lock);
3319 }
3320 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3321
3322 /* Enable polling stats and return whether they were already enabled. */
3323 static bool blk_poll_stats_enable(struct request_queue *q)
3324 {
3325         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3326             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3327                 return true;
3328         blk_stat_add_callback(q, q->poll_cb);
3329         return false;
3330 }
3331
3332 static void blk_mq_poll_stats_start(struct request_queue *q)
3333 {
3334         /*
3335          * We don't arm the callback if polling stats are not enabled or the
3336          * callback is already active.
3337          */
3338         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3339             blk_stat_is_active(q->poll_cb))
3340                 return;
3341
3342         blk_stat_activate_msecs(q->poll_cb, 100);
3343 }
3344
3345 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3346 {
3347         struct request_queue *q = cb->data;
3348         int bucket;
3349
3350         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3351                 if (cb->stat[bucket].nr_samples)
3352                         q->poll_stat[bucket] = cb->stat[bucket];
3353         }
3354 }
3355
3356 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3357                                        struct blk_mq_hw_ctx *hctx,
3358                                        struct request *rq)
3359 {
3360         unsigned long ret = 0;
3361         int bucket;
3362
3363         /*
3364          * If stats collection isn't on, don't sleep but turn it on for
3365          * future users
3366          */
3367         if (!blk_poll_stats_enable(q))
3368                 return 0;
3369
3370         /*
3371          * As an optimistic guess, use half of the mean service time
3372          * for this type of request. We can (and should) make this smarter.
3373          * For instance, if the completion latencies are tight, we can
3374          * get closer than just half the mean. This is especially
3375          * important on devices where the completion latencies are longer
3376          * than ~10 usec. We do use the stats for the relevant IO size
3377          * if available which does lead to better estimates.
3378          */
3379         bucket = blk_mq_poll_stats_bkt(rq);
3380         if (bucket < 0)
3381                 return ret;
3382
3383         if (q->poll_stat[bucket].nr_samples)
3384                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3385
3386         return ret;
3387 }
3388
3389 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3390                                      struct blk_mq_hw_ctx *hctx,
3391                                      struct request *rq)
3392 {
3393         struct hrtimer_sleeper hs;
3394         enum hrtimer_mode mode;
3395         unsigned int nsecs;
3396         ktime_t kt;
3397
3398         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3399                 return false;
3400
3401         /*
3402          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3403          *
3404          *  0:  use half of prev avg
3405          * >0:  use this specific value
3406          */
3407         if (q->poll_nsec > 0)
3408                 nsecs = q->poll_nsec;
3409         else
3410                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3411
3412         if (!nsecs)
3413                 return false;
3414
3415         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3416
3417         /*
3418          * This will be replaced with the stats tracking code, using
3419          * 'avg_completion_time / 2' as the pre-sleep target.
3420          */
3421         kt = nsecs;
3422
3423         mode = HRTIMER_MODE_REL;
3424         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3425         hrtimer_set_expires(&hs.timer, kt);
3426
3427         hrtimer_init_sleeper(&hs, current);
3428         do {
3429                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3430                         break;
3431                 set_current_state(TASK_UNINTERRUPTIBLE);
3432                 hrtimer_start_expires(&hs.timer, mode);
3433                 if (hs.task)
3434                         io_schedule();
3435                 hrtimer_cancel(&hs.timer);
3436                 mode = HRTIMER_MODE_ABS;
3437         } while (hs.task && !signal_pending(current));
3438
3439         __set_current_state(TASK_RUNNING);
3440         destroy_hrtimer_on_stack(&hs.timer);
3441         return true;
3442 }
3443
3444 static bool blk_mq_poll_hybrid(struct request_queue *q,
3445                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3446 {
3447         struct request *rq;
3448
3449         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3450                 return false;
3451
3452         if (!blk_qc_t_is_internal(cookie))
3453                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3454         else {
3455                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3456                 /*
3457                  * With scheduling, if the request has completed, we'll
3458                  * get a NULL return here, as we clear the sched tag when
3459                  * that happens. The request still remains valid, like always,
3460                  * so we should be safe with just the NULL check.
3461                  */
3462                 if (!rq)
3463                         return false;
3464         }
3465
3466         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3467 }
3468
3469 /**
3470  * blk_poll - poll for IO completions
3471  * @q:  the queue
3472  * @cookie: cookie passed back at IO submission time
3473  * @spin: whether to spin for completions
3474  *
3475  * Description:
3476  *    Poll for completions on the passed in queue. Returns number of
3477  *    completed entries found. If @spin is true, then blk_poll will continue
3478  *    looping until at least one completion is found, unless the task is
3479  *    otherwise marked running (or we need to reschedule).
3480  */
3481 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3482 {
3483         struct blk_mq_hw_ctx *hctx;
3484         long state;
3485
3486         if (!blk_qc_t_valid(cookie) ||
3487             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3488                 return 0;
3489
3490         if (current->plug)
3491                 blk_flush_plug_list(current->plug, false);
3492
3493         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3494
3495         /*
3496          * If we sleep, have the caller restart the poll loop to reset
3497          * the state. Like for the other success return cases, the
3498          * caller is responsible for checking if the IO completed. If
3499          * the IO isn't complete, we'll get called again and will go
3500          * straight to the busy poll loop.
3501          */
3502         if (blk_mq_poll_hybrid(q, hctx, cookie))
3503                 return 1;
3504
3505         hctx->poll_considered++;
3506
3507         state = current->state;
3508         do {
3509                 int ret;
3510
3511                 hctx->poll_invoked++;
3512
3513                 ret = q->mq_ops->poll(hctx);
3514                 if (ret > 0) {
3515                         hctx->poll_success++;
3516                         __set_current_state(TASK_RUNNING);
3517                         return ret;
3518                 }
3519
3520                 if (signal_pending_state(state, current))
3521                         __set_current_state(TASK_RUNNING);
3522
3523                 if (current->state == TASK_RUNNING)
3524                         return 1;
3525                 if (ret < 0 || !spin)
3526                         break;
3527                 cpu_relax();
3528         } while (!need_resched());
3529
3530         __set_current_state(TASK_RUNNING);
3531         return 0;
3532 }
3533 EXPORT_SYMBOL_GPL(blk_poll);
3534
3535 unsigned int blk_mq_rq_cpu(struct request *rq)
3536 {
3537         return rq->mq_ctx->cpu;
3538 }
3539 EXPORT_SYMBOL(blk_mq_rq_cpu);
3540
3541 static int __init blk_mq_init(void)
3542 {
3543         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3544                                 blk_mq_hctx_notify_dead);
3545         return 0;
3546 }
3547 subsys_initcall(blk_mq_init);