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         blk_mq_bio_to_request(rq, bio, nr_segs);
1975
1976         plug = current->plug;
1977         if (unlikely(is_flush_fua)) {
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                 if (!request_count)
1990                         trace_block_plug(q);
1991                 else
1992                         last = list_entry_rq(plug->mq_list.prev);
1993
1994                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1995                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1996                         blk_flush_plug_list(plug, false);
1997                         trace_block_plug(q);
1998                 }
1999
2000                 blk_add_rq_to_plug(plug, rq);
2001         } else if (plug && !blk_queue_nomerges(q)) {
2002                 /*
2003                  * We do limited plugging. If the bio can be merged, do that.
2004                  * Otherwise the existing request in the plug list will be
2005                  * issued. So the plug list will have one request at most
2006                  * The plug list might get flushed before this. If that happens,
2007                  * the plug list is empty, and same_queue_rq is invalid.
2008                  */
2009                 if (list_empty(&plug->mq_list))
2010                         same_queue_rq = NULL;
2011                 if (same_queue_rq) {
2012                         list_del_init(&same_queue_rq->queuelist);
2013                         plug->rq_count--;
2014                 }
2015                 blk_add_rq_to_plug(plug, rq);
2016                 trace_block_plug(q);
2017
2018                 if (same_queue_rq) {
2019                         data.hctx = same_queue_rq->mq_hctx;
2020                         trace_block_unplug(q, 1, true);
2021                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2022                                         &cookie);
2023                 }
2024         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2025                         !data.hctx->dispatch_busy)) {
2026                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2027         } else {
2028                 blk_mq_sched_insert_request(rq, false, true, true);
2029         }
2030
2031         return cookie;
2032 }
2033
2034 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2035                      unsigned int hctx_idx)
2036 {
2037         struct page *page;
2038
2039         if (tags->rqs && set->ops->exit_request) {
2040                 int i;
2041
2042                 for (i = 0; i < tags->nr_tags; i++) {
2043                         struct request *rq = tags->static_rqs[i];
2044
2045                         if (!rq)
2046                                 continue;
2047                         set->ops->exit_request(set, rq, hctx_idx);
2048                         tags->static_rqs[i] = NULL;
2049                 }
2050         }
2051
2052         while (!list_empty(&tags->page_list)) {
2053                 page = list_first_entry(&tags->page_list, struct page, lru);
2054                 list_del_init(&page->lru);
2055                 /*
2056                  * Remove kmemleak object previously allocated in
2057                  * blk_mq_alloc_rqs().
2058                  */
2059                 kmemleak_free(page_address(page));
2060                 __free_pages(page, page->private);
2061         }
2062 }
2063
2064 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2065 {
2066         kfree(tags->rqs);
2067         tags->rqs = NULL;
2068         kfree(tags->static_rqs);
2069         tags->static_rqs = NULL;
2070
2071         blk_mq_free_tags(tags);
2072 }
2073
2074 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2075                                         unsigned int hctx_idx,
2076                                         unsigned int nr_tags,
2077                                         unsigned int reserved_tags)
2078 {
2079         struct blk_mq_tags *tags;
2080         int node;
2081
2082         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2083         if (node == NUMA_NO_NODE)
2084                 node = set->numa_node;
2085
2086         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2087                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2088         if (!tags)
2089                 return NULL;
2090
2091         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2092                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2093                                  node);
2094         if (!tags->rqs) {
2095                 blk_mq_free_tags(tags);
2096                 return NULL;
2097         }
2098
2099         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2100                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2101                                         node);
2102         if (!tags->static_rqs) {
2103                 kfree(tags->rqs);
2104                 blk_mq_free_tags(tags);
2105                 return NULL;
2106         }
2107
2108         return tags;
2109 }
2110
2111 static size_t order_to_size(unsigned int order)
2112 {
2113         return (size_t)PAGE_SIZE << order;
2114 }
2115
2116 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2117                                unsigned int hctx_idx, int node)
2118 {
2119         int ret;
2120
2121         if (set->ops->init_request) {
2122                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2123                 if (ret)
2124                         return ret;
2125         }
2126
2127         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2128         return 0;
2129 }
2130
2131 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2132                      unsigned int hctx_idx, unsigned int depth)
2133 {
2134         unsigned int i, j, entries_per_page, max_order = 4;
2135         size_t rq_size, left;
2136         int node;
2137
2138         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2139         if (node == NUMA_NO_NODE)
2140                 node = set->numa_node;
2141
2142         INIT_LIST_HEAD(&tags->page_list);
2143
2144         /*
2145          * rq_size is the size of the request plus driver payload, rounded
2146          * to the cacheline size
2147          */
2148         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2149                                 cache_line_size());
2150         left = rq_size * depth;
2151
2152         for (i = 0; i < depth; ) {
2153                 int this_order = max_order;
2154                 struct page *page;
2155                 int to_do;
2156                 void *p;
2157
2158                 while (this_order && left < order_to_size(this_order - 1))
2159                         this_order--;
2160
2161                 do {
2162                         page = alloc_pages_node(node,
2163                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2164                                 this_order);
2165                         if (page)
2166                                 break;
2167                         if (!this_order--)
2168                                 break;
2169                         if (order_to_size(this_order) < rq_size)
2170                                 break;
2171                 } while (1);
2172
2173                 if (!page)
2174                         goto fail;
2175
2176                 page->private = this_order;
2177                 list_add_tail(&page->lru, &tags->page_list);
2178
2179                 p = page_address(page);
2180                 /*
2181                  * Allow kmemleak to scan these pages as they contain pointers
2182                  * to additional allocations like via ops->init_request().
2183                  */
2184                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2185                 entries_per_page = order_to_size(this_order) / rq_size;
2186                 to_do = min(entries_per_page, depth - i);
2187                 left -= to_do * rq_size;
2188                 for (j = 0; j < to_do; j++) {
2189                         struct request *rq = p;
2190
2191                         tags->static_rqs[i] = rq;
2192                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2193                                 tags->static_rqs[i] = NULL;
2194                                 goto fail;
2195                         }
2196
2197                         p += rq_size;
2198                         i++;
2199                 }
2200         }
2201         return 0;
2202
2203 fail:
2204         blk_mq_free_rqs(set, tags, hctx_idx);
2205         return -ENOMEM;
2206 }
2207
2208 /*
2209  * 'cpu' is going away. splice any existing rq_list entries from this
2210  * software queue to the hw queue dispatch list, and ensure that it
2211  * gets run.
2212  */
2213 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2214 {
2215         struct blk_mq_hw_ctx *hctx;
2216         struct blk_mq_ctx *ctx;
2217         LIST_HEAD(tmp);
2218         enum hctx_type type;
2219
2220         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2221         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2222         type = hctx->type;
2223
2224         spin_lock(&ctx->lock);
2225         if (!list_empty(&ctx->rq_lists[type])) {
2226                 list_splice_init(&ctx->rq_lists[type], &tmp);
2227                 blk_mq_hctx_clear_pending(hctx, ctx);
2228         }
2229         spin_unlock(&ctx->lock);
2230
2231         if (list_empty(&tmp))
2232                 return 0;
2233
2234         spin_lock(&hctx->lock);
2235         list_splice_tail_init(&tmp, &hctx->dispatch);
2236         spin_unlock(&hctx->lock);
2237
2238         blk_mq_run_hw_queue(hctx, true);
2239         return 0;
2240 }
2241
2242 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2243 {
2244         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2245                                             &hctx->cpuhp_dead);
2246 }
2247
2248 /* hctx->ctxs will be freed in queue's release handler */
2249 static void blk_mq_exit_hctx(struct request_queue *q,
2250                 struct blk_mq_tag_set *set,
2251                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2252 {
2253         if (blk_mq_hw_queue_mapped(hctx))
2254                 blk_mq_tag_idle(hctx);
2255
2256         if (set->ops->exit_request)
2257                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2258
2259         if (set->ops->exit_hctx)
2260                 set->ops->exit_hctx(hctx, hctx_idx);
2261
2262         blk_mq_remove_cpuhp(hctx);
2263
2264         spin_lock(&q->unused_hctx_lock);
2265         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2266         spin_unlock(&q->unused_hctx_lock);
2267 }
2268
2269 static void blk_mq_exit_hw_queues(struct request_queue *q,
2270                 struct blk_mq_tag_set *set, int nr_queue)
2271 {
2272         struct blk_mq_hw_ctx *hctx;
2273         unsigned int i;
2274
2275         queue_for_each_hw_ctx(q, hctx, i) {
2276                 if (i == nr_queue)
2277                         break;
2278                 blk_mq_debugfs_unregister_hctx(hctx);
2279                 blk_mq_exit_hctx(q, set, hctx, i);
2280         }
2281 }
2282
2283 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2284 {
2285         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2286
2287         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2288                            __alignof__(struct blk_mq_hw_ctx)) !=
2289                      sizeof(struct blk_mq_hw_ctx));
2290
2291         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2292                 hw_ctx_size += sizeof(struct srcu_struct);
2293
2294         return hw_ctx_size;
2295 }
2296
2297 static int blk_mq_init_hctx(struct request_queue *q,
2298                 struct blk_mq_tag_set *set,
2299                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2300 {
2301         hctx->queue_num = hctx_idx;
2302
2303         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2304
2305         hctx->tags = set->tags[hctx_idx];
2306
2307         if (set->ops->init_hctx &&
2308             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2309                 goto unregister_cpu_notifier;
2310
2311         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2312                                 hctx->numa_node))
2313                 goto exit_hctx;
2314         return 0;
2315
2316  exit_hctx:
2317         if (set->ops->exit_hctx)
2318                 set->ops->exit_hctx(hctx, hctx_idx);
2319  unregister_cpu_notifier:
2320         blk_mq_remove_cpuhp(hctx);
2321         return -1;
2322 }
2323
2324 static struct blk_mq_hw_ctx *
2325 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2326                 int node)
2327 {
2328         struct blk_mq_hw_ctx *hctx;
2329         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2330
2331         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2332         if (!hctx)
2333                 goto fail_alloc_hctx;
2334
2335         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2336                 goto free_hctx;
2337
2338         atomic_set(&hctx->nr_active, 0);
2339         if (node == NUMA_NO_NODE)
2340                 node = set->numa_node;
2341         hctx->numa_node = node;
2342
2343         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2344         spin_lock_init(&hctx->lock);
2345         INIT_LIST_HEAD(&hctx->dispatch);
2346         hctx->queue = q;
2347         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2348
2349         INIT_LIST_HEAD(&hctx->hctx_list);
2350
2351         /*
2352          * Allocate space for all possible cpus to avoid allocation at
2353          * runtime
2354          */
2355         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2356                         gfp, node);
2357         if (!hctx->ctxs)
2358                 goto free_cpumask;
2359
2360         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2361                                 gfp, node))
2362                 goto free_ctxs;
2363         hctx->nr_ctx = 0;
2364
2365         spin_lock_init(&hctx->dispatch_wait_lock);
2366         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2367         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2368
2369         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2370                         gfp);
2371         if (!hctx->fq)
2372                 goto free_bitmap;
2373
2374         if (hctx->flags & BLK_MQ_F_BLOCKING)
2375                 init_srcu_struct(hctx->srcu);
2376         blk_mq_hctx_kobj_init(hctx);
2377
2378         return hctx;
2379
2380  free_bitmap:
2381         sbitmap_free(&hctx->ctx_map);
2382  free_ctxs:
2383         kfree(hctx->ctxs);
2384  free_cpumask:
2385         free_cpumask_var(hctx->cpumask);
2386  free_hctx:
2387         kfree(hctx);
2388  fail_alloc_hctx:
2389         return NULL;
2390 }
2391
2392 static void blk_mq_init_cpu_queues(struct request_queue *q,
2393                                    unsigned int nr_hw_queues)
2394 {
2395         struct blk_mq_tag_set *set = q->tag_set;
2396         unsigned int i, j;
2397
2398         for_each_possible_cpu(i) {
2399                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2400                 struct blk_mq_hw_ctx *hctx;
2401                 int k;
2402
2403                 __ctx->cpu = i;
2404                 spin_lock_init(&__ctx->lock);
2405                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2406                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2407
2408                 __ctx->queue = q;
2409
2410                 /*
2411                  * Set local node, IFF we have more than one hw queue. If
2412                  * not, we remain on the home node of the device
2413                  */
2414                 for (j = 0; j < set->nr_maps; j++) {
2415                         hctx = blk_mq_map_queue_type(q, j, i);
2416                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2417                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2418                 }
2419         }
2420 }
2421
2422 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2423 {
2424         int ret = 0;
2425
2426         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2427                                         set->queue_depth, set->reserved_tags);
2428         if (!set->tags[hctx_idx])
2429                 return false;
2430
2431         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2432                                 set->queue_depth);
2433         if (!ret)
2434                 return true;
2435
2436         blk_mq_free_rq_map(set->tags[hctx_idx]);
2437         set->tags[hctx_idx] = NULL;
2438         return false;
2439 }
2440
2441 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2442                                          unsigned int hctx_idx)
2443 {
2444         if (set->tags && set->tags[hctx_idx]) {
2445                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2446                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2447                 set->tags[hctx_idx] = NULL;
2448         }
2449 }
2450
2451 static void blk_mq_map_swqueue(struct request_queue *q)
2452 {
2453         unsigned int i, j, hctx_idx;
2454         struct blk_mq_hw_ctx *hctx;
2455         struct blk_mq_ctx *ctx;
2456         struct blk_mq_tag_set *set = q->tag_set;
2457
2458         /*
2459          * Avoid others reading imcomplete hctx->cpumask through sysfs
2460          */
2461         mutex_lock(&q->sysfs_lock);
2462
2463         queue_for_each_hw_ctx(q, hctx, i) {
2464                 cpumask_clear(hctx->cpumask);
2465                 hctx->nr_ctx = 0;
2466                 hctx->dispatch_from = NULL;
2467         }
2468
2469         /*
2470          * Map software to hardware queues.
2471          *
2472          * If the cpu isn't present, the cpu is mapped to first hctx.
2473          */
2474         for_each_possible_cpu(i) {
2475                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2476                 /* unmapped hw queue can be remapped after CPU topo changed */
2477                 if (!set->tags[hctx_idx] &&
2478                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2479                         /*
2480                          * If tags initialization fail for some hctx,
2481                          * that hctx won't be brought online.  In this
2482                          * case, remap the current ctx to hctx[0] which
2483                          * is guaranteed to always have tags allocated
2484                          */
2485                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2486                 }
2487
2488                 ctx = per_cpu_ptr(q->queue_ctx, i);
2489                 for (j = 0; j < set->nr_maps; j++) {
2490                         if (!set->map[j].nr_queues) {
2491                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2492                                                 HCTX_TYPE_DEFAULT, i);
2493                                 continue;
2494                         }
2495
2496                         hctx = blk_mq_map_queue_type(q, j, i);
2497                         ctx->hctxs[j] = hctx;
2498                         /*
2499                          * If the CPU is already set in the mask, then we've
2500                          * mapped this one already. This can happen if
2501                          * devices share queues across queue maps.
2502                          */
2503                         if (cpumask_test_cpu(i, hctx->cpumask))
2504                                 continue;
2505
2506                         cpumask_set_cpu(i, hctx->cpumask);
2507                         hctx->type = j;
2508                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2509                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2510
2511                         /*
2512                          * If the nr_ctx type overflows, we have exceeded the
2513                          * amount of sw queues we can support.
2514                          */
2515                         BUG_ON(!hctx->nr_ctx);
2516                 }
2517
2518                 for (; j < HCTX_MAX_TYPES; j++)
2519                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2520                                         HCTX_TYPE_DEFAULT, i);
2521         }
2522
2523         mutex_unlock(&q->sysfs_lock);
2524
2525         queue_for_each_hw_ctx(q, hctx, i) {
2526                 /*
2527                  * If no software queues are mapped to this hardware queue,
2528                  * disable it and free the request entries.
2529                  */
2530                 if (!hctx->nr_ctx) {
2531                         /* Never unmap queue 0.  We need it as a
2532                          * fallback in case of a new remap fails
2533                          * allocation
2534                          */
2535                         if (i && set->tags[i])
2536                                 blk_mq_free_map_and_requests(set, i);
2537
2538                         hctx->tags = NULL;
2539                         continue;
2540                 }
2541
2542                 hctx->tags = set->tags[i];
2543                 WARN_ON(!hctx->tags);
2544
2545                 /*
2546                  * Set the map size to the number of mapped software queues.
2547                  * This is more accurate and more efficient than looping
2548                  * over all possibly mapped software queues.
2549                  */
2550                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2551
2552                 /*
2553                  * Initialize batch roundrobin counts
2554                  */
2555                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2556                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2557         }
2558 }
2559
2560 /*
2561  * Caller needs to ensure that we're either frozen/quiesced, or that
2562  * the queue isn't live yet.
2563  */
2564 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2565 {
2566         struct blk_mq_hw_ctx *hctx;
2567         int i;
2568
2569         queue_for_each_hw_ctx(q, hctx, i) {
2570                 if (shared)
2571                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2572                 else
2573                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2574         }
2575 }
2576
2577 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2578                                         bool shared)
2579 {
2580         struct request_queue *q;
2581
2582         lockdep_assert_held(&set->tag_list_lock);
2583
2584         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2585                 blk_mq_freeze_queue(q);
2586                 queue_set_hctx_shared(q, shared);
2587                 blk_mq_unfreeze_queue(q);
2588         }
2589 }
2590
2591 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2592 {
2593         struct blk_mq_tag_set *set = q->tag_set;
2594
2595         mutex_lock(&set->tag_list_lock);
2596         list_del_rcu(&q->tag_set_list);
2597         if (list_is_singular(&set->tag_list)) {
2598                 /* just transitioned to unshared */
2599                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2600                 /* update existing queue */
2601                 blk_mq_update_tag_set_depth(set, false);
2602         }
2603         mutex_unlock(&set->tag_list_lock);
2604         INIT_LIST_HEAD(&q->tag_set_list);
2605 }
2606
2607 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2608                                      struct request_queue *q)
2609 {
2610         mutex_lock(&set->tag_list_lock);
2611
2612         /*
2613          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2614          */
2615         if (!list_empty(&set->tag_list) &&
2616             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2617                 set->flags |= BLK_MQ_F_TAG_SHARED;
2618                 /* update existing queue */
2619                 blk_mq_update_tag_set_depth(set, true);
2620         }
2621         if (set->flags & BLK_MQ_F_TAG_SHARED)
2622                 queue_set_hctx_shared(q, true);
2623         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2624
2625         mutex_unlock(&set->tag_list_lock);
2626 }
2627
2628 /* All allocations will be freed in release handler of q->mq_kobj */
2629 static int blk_mq_alloc_ctxs(struct request_queue *q)
2630 {
2631         struct blk_mq_ctxs *ctxs;
2632         int cpu;
2633
2634         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2635         if (!ctxs)
2636                 return -ENOMEM;
2637
2638         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2639         if (!ctxs->queue_ctx)
2640                 goto fail;
2641
2642         for_each_possible_cpu(cpu) {
2643                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2644                 ctx->ctxs = ctxs;
2645         }
2646
2647         q->mq_kobj = &ctxs->kobj;
2648         q->queue_ctx = ctxs->queue_ctx;
2649
2650         return 0;
2651  fail:
2652         kfree(ctxs);
2653         return -ENOMEM;
2654 }
2655
2656 /*
2657  * It is the actual release handler for mq, but we do it from
2658  * request queue's release handler for avoiding use-after-free
2659  * and headache because q->mq_kobj shouldn't have been introduced,
2660  * but we can't group ctx/kctx kobj without it.
2661  */
2662 void blk_mq_release(struct request_queue *q)
2663 {
2664         struct blk_mq_hw_ctx *hctx, *next;
2665         int i;
2666
2667         cancel_delayed_work_sync(&q->requeue_work);
2668
2669         queue_for_each_hw_ctx(q, hctx, i)
2670                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2671
2672         /* all hctx are in .unused_hctx_list now */
2673         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2674                 list_del_init(&hctx->hctx_list);
2675                 kobject_put(&hctx->kobj);
2676         }
2677
2678         kfree(q->queue_hw_ctx);
2679
2680         /*
2681          * release .mq_kobj and sw queue's kobject now because
2682          * both share lifetime with request queue.
2683          */
2684         blk_mq_sysfs_deinit(q);
2685 }
2686
2687 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2688 {
2689         struct request_queue *uninit_q, *q;
2690
2691         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2692         if (!uninit_q)
2693                 return ERR_PTR(-ENOMEM);
2694
2695         q = blk_mq_init_allocated_queue(set, uninit_q);
2696         if (IS_ERR(q))
2697                 blk_cleanup_queue(uninit_q);
2698
2699         return q;
2700 }
2701 EXPORT_SYMBOL(blk_mq_init_queue);
2702
2703 /*
2704  * Helper for setting up a queue with mq ops, given queue depth, and
2705  * the passed in mq ops flags.
2706  */
2707 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2708                                            const struct blk_mq_ops *ops,
2709                                            unsigned int queue_depth,
2710                                            unsigned int set_flags)
2711 {
2712         struct request_queue *q;
2713         int ret;
2714
2715         memset(set, 0, sizeof(*set));
2716         set->ops = ops;
2717         set->nr_hw_queues = 1;
2718         set->nr_maps = 1;
2719         set->queue_depth = queue_depth;
2720         set->numa_node = NUMA_NO_NODE;
2721         set->flags = set_flags;
2722
2723         ret = blk_mq_alloc_tag_set(set);
2724         if (ret)
2725                 return ERR_PTR(ret);
2726
2727         q = blk_mq_init_queue(set);
2728         if (IS_ERR(q)) {
2729                 blk_mq_free_tag_set(set);
2730                 return q;
2731         }
2732
2733         return q;
2734 }
2735 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2736
2737 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2738                 struct blk_mq_tag_set *set, struct request_queue *q,
2739                 int hctx_idx, int node)
2740 {
2741         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2742
2743         /* reuse dead hctx first */
2744         spin_lock(&q->unused_hctx_lock);
2745         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2746                 if (tmp->numa_node == node) {
2747                         hctx = tmp;
2748                         break;
2749                 }
2750         }
2751         if (hctx)
2752                 list_del_init(&hctx->hctx_list);
2753         spin_unlock(&q->unused_hctx_lock);
2754
2755         if (!hctx)
2756                 hctx = blk_mq_alloc_hctx(q, set, node);
2757         if (!hctx)
2758                 goto fail;
2759
2760         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2761                 goto free_hctx;
2762
2763         return hctx;
2764
2765  free_hctx:
2766         kobject_put(&hctx->kobj);
2767  fail:
2768         return NULL;
2769 }
2770
2771 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2772                                                 struct request_queue *q)
2773 {
2774         int i, j, end;
2775         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2776
2777         /* protect against switching io scheduler  */
2778         mutex_lock(&q->sysfs_lock);
2779         for (i = 0; i < set->nr_hw_queues; i++) {
2780                 int node;
2781                 struct blk_mq_hw_ctx *hctx;
2782
2783                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2784                 /*
2785                  * If the hw queue has been mapped to another numa node,
2786                  * we need to realloc the hctx. If allocation fails, fallback
2787                  * to use the previous one.
2788                  */
2789                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2790                         continue;
2791
2792                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2793                 if (hctx) {
2794                         if (hctxs[i])
2795                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2796                         hctxs[i] = hctx;
2797                 } else {
2798                         if (hctxs[i])
2799                                 pr_warn("Allocate new hctx on node %d fails,\
2800                                                 fallback to previous one on node %d\n",
2801                                                 node, hctxs[i]->numa_node);
2802                         else
2803                                 break;
2804                 }
2805         }
2806         /*
2807          * Increasing nr_hw_queues fails. Free the newly allocated
2808          * hctxs and keep the previous q->nr_hw_queues.
2809          */
2810         if (i != set->nr_hw_queues) {
2811                 j = q->nr_hw_queues;
2812                 end = i;
2813         } else {
2814                 j = i;
2815                 end = q->nr_hw_queues;
2816                 q->nr_hw_queues = set->nr_hw_queues;
2817         }
2818
2819         for (; j < end; j++) {
2820                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2821
2822                 if (hctx) {
2823                         if (hctx->tags)
2824                                 blk_mq_free_map_and_requests(set, j);
2825                         blk_mq_exit_hctx(q, set, hctx, j);
2826                         hctxs[j] = NULL;
2827                 }
2828         }
2829         mutex_unlock(&q->sysfs_lock);
2830 }
2831
2832 /*
2833  * Maximum number of hardware queues we support. For single sets, we'll never
2834  * have more than the CPUs (software queues). For multiple sets, the tag_set
2835  * user may have set ->nr_hw_queues larger.
2836  */
2837 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2838 {
2839         if (set->nr_maps == 1)
2840                 return nr_cpu_ids;
2841
2842         return max(set->nr_hw_queues, nr_cpu_ids);
2843 }
2844
2845 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2846                                                   struct request_queue *q)
2847 {
2848         /* mark the queue as mq asap */
2849         q->mq_ops = set->ops;
2850
2851         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2852                                              blk_mq_poll_stats_bkt,
2853                                              BLK_MQ_POLL_STATS_BKTS, q);
2854         if (!q->poll_cb)
2855                 goto err_exit;
2856
2857         if (blk_mq_alloc_ctxs(q))
2858                 goto err_poll;
2859
2860         /* init q->mq_kobj and sw queues' kobjects */
2861         blk_mq_sysfs_init(q);
2862
2863         q->nr_queues = nr_hw_queues(set);
2864         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2865                                                 GFP_KERNEL, set->numa_node);
2866         if (!q->queue_hw_ctx)
2867                 goto err_sys_init;
2868
2869         INIT_LIST_HEAD(&q->unused_hctx_list);
2870         spin_lock_init(&q->unused_hctx_lock);
2871
2872         blk_mq_realloc_hw_ctxs(set, q);
2873         if (!q->nr_hw_queues)
2874                 goto err_hctxs;
2875
2876         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2877         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2878
2879         q->tag_set = set;
2880
2881         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2882         if (set->nr_maps > HCTX_TYPE_POLL &&
2883             set->map[HCTX_TYPE_POLL].nr_queues)
2884                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2885
2886         q->sg_reserved_size = INT_MAX;
2887
2888         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2889         INIT_LIST_HEAD(&q->requeue_list);
2890         spin_lock_init(&q->requeue_lock);
2891
2892         blk_queue_make_request(q, blk_mq_make_request);
2893
2894         /*
2895          * Do this after blk_queue_make_request() overrides it...
2896          */
2897         q->nr_requests = set->queue_depth;
2898
2899         /*
2900          * Default to classic polling
2901          */
2902         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2903
2904         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2905         blk_mq_add_queue_tag_set(set, q);
2906         blk_mq_map_swqueue(q);
2907
2908         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2909                 int ret;
2910
2911                 ret = elevator_init_mq(q);
2912                 if (ret)
2913                         return ERR_PTR(ret);
2914         }
2915
2916         return q;
2917
2918 err_hctxs:
2919         kfree(q->queue_hw_ctx);
2920 err_sys_init:
2921         blk_mq_sysfs_deinit(q);
2922 err_poll:
2923         blk_stat_free_callback(q->poll_cb);
2924         q->poll_cb = NULL;
2925 err_exit:
2926         q->mq_ops = NULL;
2927         return ERR_PTR(-ENOMEM);
2928 }
2929 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2930
2931 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2932 void blk_mq_exit_queue(struct request_queue *q)
2933 {
2934         struct blk_mq_tag_set   *set = q->tag_set;
2935
2936         blk_mq_del_queue_tag_set(q);
2937         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2938 }
2939
2940 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2941 {
2942         int i;
2943
2944         for (i = 0; i < set->nr_hw_queues; i++)
2945                 if (!__blk_mq_alloc_rq_map(set, i))
2946                         goto out_unwind;
2947
2948         return 0;
2949
2950 out_unwind:
2951         while (--i >= 0)
2952                 blk_mq_free_rq_map(set->tags[i]);
2953
2954         return -ENOMEM;
2955 }
2956
2957 /*
2958  * Allocate the request maps associated with this tag_set. Note that this
2959  * may reduce the depth asked for, if memory is tight. set->queue_depth
2960  * will be updated to reflect the allocated depth.
2961  */
2962 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2963 {
2964         unsigned int depth;
2965         int err;
2966
2967         depth = set->queue_depth;
2968         do {
2969                 err = __blk_mq_alloc_rq_maps(set);
2970                 if (!err)
2971                         break;
2972
2973                 set->queue_depth >>= 1;
2974                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2975                         err = -ENOMEM;
2976                         break;
2977                 }
2978         } while (set->queue_depth);
2979
2980         if (!set->queue_depth || err) {
2981                 pr_err("blk-mq: failed to allocate request map\n");
2982                 return -ENOMEM;
2983         }
2984
2985         if (depth != set->queue_depth)
2986                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2987                                                 depth, set->queue_depth);
2988
2989         return 0;
2990 }
2991
2992 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2993 {
2994         if (set->ops->map_queues && !is_kdump_kernel()) {
2995                 int i;
2996
2997                 /*
2998                  * transport .map_queues is usually done in the following
2999                  * way:
3000                  *
3001                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3002                  *      mask = get_cpu_mask(queue)
3003                  *      for_each_cpu(cpu, mask)
3004                  *              set->map[x].mq_map[cpu] = queue;
3005                  * }
3006                  *
3007                  * When we need to remap, the table has to be cleared for
3008                  * killing stale mapping since one CPU may not be mapped
3009                  * to any hw queue.
3010                  */
3011                 for (i = 0; i < set->nr_maps; i++)
3012                         blk_mq_clear_mq_map(&set->map[i]);
3013
3014                 return set->ops->map_queues(set);
3015         } else {
3016                 BUG_ON(set->nr_maps > 1);
3017                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3018         }
3019 }
3020
3021 /*
3022  * Alloc a tag set to be associated with one or more request queues.
3023  * May fail with EINVAL for various error conditions. May adjust the
3024  * requested depth down, if it's too large. In that case, the set
3025  * value will be stored in set->queue_depth.
3026  */
3027 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3028 {
3029         int i, ret;
3030
3031         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3032
3033         if (!set->nr_hw_queues)
3034                 return -EINVAL;
3035         if (!set->queue_depth)
3036                 return -EINVAL;
3037         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3038                 return -EINVAL;
3039
3040         if (!set->ops->queue_rq)
3041                 return -EINVAL;
3042
3043         if (!set->ops->get_budget ^ !set->ops->put_budget)
3044                 return -EINVAL;
3045
3046         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3047                 pr_info("blk-mq: reduced tag depth to %u\n",
3048                         BLK_MQ_MAX_DEPTH);
3049                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3050         }
3051
3052         if (!set->nr_maps)
3053                 set->nr_maps = 1;
3054         else if (set->nr_maps > HCTX_MAX_TYPES)
3055                 return -EINVAL;
3056
3057         /*
3058          * If a crashdump is active, then we are potentially in a very
3059          * memory constrained environment. Limit us to 1 queue and
3060          * 64 tags to prevent using too much memory.
3061          */
3062         if (is_kdump_kernel()) {
3063                 set->nr_hw_queues = 1;
3064                 set->nr_maps = 1;
3065                 set->queue_depth = min(64U, set->queue_depth);
3066         }
3067         /*
3068          * There is no use for more h/w queues than cpus if we just have
3069          * a single map
3070          */
3071         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3072                 set->nr_hw_queues = nr_cpu_ids;
3073
3074         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3075                                  GFP_KERNEL, set->numa_node);
3076         if (!set->tags)
3077                 return -ENOMEM;
3078
3079         ret = -ENOMEM;
3080         for (i = 0; i < set->nr_maps; i++) {
3081                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3082                                                   sizeof(set->map[i].mq_map[0]),
3083                                                   GFP_KERNEL, set->numa_node);
3084                 if (!set->map[i].mq_map)
3085                         goto out_free_mq_map;
3086                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3087         }
3088
3089         ret = blk_mq_update_queue_map(set);
3090         if (ret)
3091                 goto out_free_mq_map;
3092
3093         ret = blk_mq_alloc_rq_maps(set);
3094         if (ret)
3095                 goto out_free_mq_map;
3096
3097         mutex_init(&set->tag_list_lock);
3098         INIT_LIST_HEAD(&set->tag_list);
3099
3100         return 0;
3101
3102 out_free_mq_map:
3103         for (i = 0; i < set->nr_maps; i++) {
3104                 kfree(set->map[i].mq_map);
3105                 set->map[i].mq_map = NULL;
3106         }
3107         kfree(set->tags);
3108         set->tags = NULL;
3109         return ret;
3110 }
3111 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3112
3113 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3114 {
3115         int i, j;
3116
3117         for (i = 0; i < nr_hw_queues(set); i++)
3118                 blk_mq_free_map_and_requests(set, i);
3119
3120         for (j = 0; j < set->nr_maps; j++) {
3121                 kfree(set->map[j].mq_map);
3122                 set->map[j].mq_map = NULL;
3123         }
3124
3125         kfree(set->tags);
3126         set->tags = NULL;
3127 }
3128 EXPORT_SYMBOL(blk_mq_free_tag_set);
3129
3130 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3131 {
3132         struct blk_mq_tag_set *set = q->tag_set;
3133         struct blk_mq_hw_ctx *hctx;
3134         int i, ret;
3135
3136         if (!set)
3137                 return -EINVAL;
3138
3139         if (q->nr_requests == nr)
3140                 return 0;
3141
3142         blk_mq_freeze_queue(q);
3143         blk_mq_quiesce_queue(q);
3144
3145         ret = 0;
3146         queue_for_each_hw_ctx(q, hctx, i) {
3147                 if (!hctx->tags)
3148                         continue;
3149                 /*
3150                  * If we're using an MQ scheduler, just update the scheduler
3151                  * queue depth. This is similar to what the old code would do.
3152                  */
3153                 if (!hctx->sched_tags) {
3154                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3155                                                         false);
3156                 } else {
3157                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3158                                                         nr, true);
3159                 }
3160                 if (ret)
3161                         break;
3162                 if (q->elevator && q->elevator->type->ops.depth_updated)
3163                         q->elevator->type->ops.depth_updated(hctx);
3164         }
3165
3166         if (!ret)
3167                 q->nr_requests = nr;
3168
3169         blk_mq_unquiesce_queue(q);
3170         blk_mq_unfreeze_queue(q);
3171
3172         return ret;
3173 }
3174
3175 /*
3176  * request_queue and elevator_type pair.
3177  * It is just used by __blk_mq_update_nr_hw_queues to cache
3178  * the elevator_type associated with a request_queue.
3179  */
3180 struct blk_mq_qe_pair {
3181         struct list_head node;
3182         struct request_queue *q;
3183         struct elevator_type *type;
3184 };
3185
3186 /*
3187  * Cache the elevator_type in qe pair list and switch the
3188  * io scheduler to 'none'
3189  */
3190 static bool blk_mq_elv_switch_none(struct list_head *head,
3191                 struct request_queue *q)
3192 {
3193         struct blk_mq_qe_pair *qe;
3194
3195         if (!q->elevator)
3196                 return true;
3197
3198         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3199         if (!qe)
3200                 return false;
3201
3202         INIT_LIST_HEAD(&qe->node);
3203         qe->q = q;
3204         qe->type = q->elevator->type;
3205         list_add(&qe->node, head);
3206
3207         mutex_lock(&q->sysfs_lock);
3208         /*
3209          * After elevator_switch_mq, the previous elevator_queue will be
3210          * released by elevator_release. The reference of the io scheduler
3211          * module get by elevator_get will also be put. So we need to get
3212          * a reference of the io scheduler module here to prevent it to be
3213          * removed.
3214          */
3215         __module_get(qe->type->elevator_owner);
3216         elevator_switch_mq(q, NULL);
3217         mutex_unlock(&q->sysfs_lock);
3218
3219         return true;
3220 }
3221
3222 static void blk_mq_elv_switch_back(struct list_head *head,
3223                 struct request_queue *q)
3224 {
3225         struct blk_mq_qe_pair *qe;
3226         struct elevator_type *t = NULL;
3227
3228         list_for_each_entry(qe, head, node)
3229                 if (qe->q == q) {
3230                         t = qe->type;
3231                         break;
3232                 }
3233
3234         if (!t)
3235                 return;
3236
3237         list_del(&qe->node);
3238         kfree(qe);
3239
3240         mutex_lock(&q->sysfs_lock);
3241         elevator_switch_mq(q, t);
3242         mutex_unlock(&q->sysfs_lock);
3243 }
3244
3245 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3246                                                         int nr_hw_queues)
3247 {
3248         struct request_queue *q;
3249         LIST_HEAD(head);
3250         int prev_nr_hw_queues;
3251
3252         lockdep_assert_held(&set->tag_list_lock);
3253
3254         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3255                 nr_hw_queues = nr_cpu_ids;
3256         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3257                 return;
3258
3259         list_for_each_entry(q, &set->tag_list, tag_set_list)
3260                 blk_mq_freeze_queue(q);
3261         /*
3262          * Sync with blk_mq_queue_tag_busy_iter.
3263          */
3264         synchronize_rcu();
3265         /*
3266          * Switch IO scheduler to 'none', cleaning up the data associated
3267          * with the previous scheduler. We will switch back once we are done
3268          * updating the new sw to hw queue mappings.
3269          */
3270         list_for_each_entry(q, &set->tag_list, tag_set_list)
3271                 if (!blk_mq_elv_switch_none(&head, q))
3272                         goto switch_back;
3273
3274         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3275                 blk_mq_debugfs_unregister_hctxs(q);
3276                 blk_mq_sysfs_unregister(q);
3277         }
3278
3279         prev_nr_hw_queues = set->nr_hw_queues;
3280         set->nr_hw_queues = nr_hw_queues;
3281         blk_mq_update_queue_map(set);
3282 fallback:
3283         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3284                 blk_mq_realloc_hw_ctxs(set, q);
3285                 if (q->nr_hw_queues != set->nr_hw_queues) {
3286                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3287                                         nr_hw_queues, prev_nr_hw_queues);
3288                         set->nr_hw_queues = prev_nr_hw_queues;
3289                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3290                         goto fallback;
3291                 }
3292                 blk_mq_map_swqueue(q);
3293         }
3294
3295         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3296                 blk_mq_sysfs_register(q);
3297                 blk_mq_debugfs_register_hctxs(q);
3298         }
3299
3300 switch_back:
3301         list_for_each_entry(q, &set->tag_list, tag_set_list)
3302                 blk_mq_elv_switch_back(&head, q);
3303
3304         list_for_each_entry(q, &set->tag_list, tag_set_list)
3305                 blk_mq_unfreeze_queue(q);
3306 }
3307
3308 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3309 {
3310         mutex_lock(&set->tag_list_lock);
3311         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3312         mutex_unlock(&set->tag_list_lock);
3313 }
3314 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3315
3316 /* Enable polling stats and return whether they were already enabled. */
3317 static bool blk_poll_stats_enable(struct request_queue *q)
3318 {
3319         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3320             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3321                 return true;
3322         blk_stat_add_callback(q, q->poll_cb);
3323         return false;
3324 }
3325
3326 static void blk_mq_poll_stats_start(struct request_queue *q)
3327 {
3328         /*
3329          * We don't arm the callback if polling stats are not enabled or the
3330          * callback is already active.
3331          */
3332         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3333             blk_stat_is_active(q->poll_cb))
3334                 return;
3335
3336         blk_stat_activate_msecs(q->poll_cb, 100);
3337 }
3338
3339 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3340 {
3341         struct request_queue *q = cb->data;
3342         int bucket;
3343
3344         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3345                 if (cb->stat[bucket].nr_samples)
3346                         q->poll_stat[bucket] = cb->stat[bucket];
3347         }
3348 }
3349
3350 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3351                                        struct blk_mq_hw_ctx *hctx,
3352                                        struct request *rq)
3353 {
3354         unsigned long ret = 0;
3355         int bucket;
3356
3357         /*
3358          * If stats collection isn't on, don't sleep but turn it on for
3359          * future users
3360          */
3361         if (!blk_poll_stats_enable(q))
3362                 return 0;
3363
3364         /*
3365          * As an optimistic guess, use half of the mean service time
3366          * for this type of request. We can (and should) make this smarter.
3367          * For instance, if the completion latencies are tight, we can
3368          * get closer than just half the mean. This is especially
3369          * important on devices where the completion latencies are longer
3370          * than ~10 usec. We do use the stats for the relevant IO size
3371          * if available which does lead to better estimates.
3372          */
3373         bucket = blk_mq_poll_stats_bkt(rq);
3374         if (bucket < 0)
3375                 return ret;
3376
3377         if (q->poll_stat[bucket].nr_samples)
3378                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3379
3380         return ret;
3381 }
3382
3383 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3384                                      struct blk_mq_hw_ctx *hctx,
3385                                      struct request *rq)
3386 {
3387         struct hrtimer_sleeper hs;
3388         enum hrtimer_mode mode;
3389         unsigned int nsecs;
3390         ktime_t kt;
3391
3392         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3393                 return false;
3394
3395         /*
3396          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3397          *
3398          *  0:  use half of prev avg
3399          * >0:  use this specific value
3400          */
3401         if (q->poll_nsec > 0)
3402                 nsecs = q->poll_nsec;
3403         else
3404                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3405
3406         if (!nsecs)
3407                 return false;
3408
3409         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3410
3411         /*
3412          * This will be replaced with the stats tracking code, using
3413          * 'avg_completion_time / 2' as the pre-sleep target.
3414          */
3415         kt = nsecs;
3416
3417         mode = HRTIMER_MODE_REL;
3418         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3419         hrtimer_set_expires(&hs.timer, kt);
3420
3421         hrtimer_init_sleeper(&hs, current);
3422         do {
3423                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3424                         break;
3425                 set_current_state(TASK_UNINTERRUPTIBLE);
3426                 hrtimer_start_expires(&hs.timer, mode);
3427                 if (hs.task)
3428                         io_schedule();
3429                 hrtimer_cancel(&hs.timer);
3430                 mode = HRTIMER_MODE_ABS;
3431         } while (hs.task && !signal_pending(current));
3432
3433         __set_current_state(TASK_RUNNING);
3434         destroy_hrtimer_on_stack(&hs.timer);
3435         return true;
3436 }
3437
3438 static bool blk_mq_poll_hybrid(struct request_queue *q,
3439                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3440 {
3441         struct request *rq;
3442
3443         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3444                 return false;
3445
3446         if (!blk_qc_t_is_internal(cookie))
3447                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3448         else {
3449                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3450                 /*
3451                  * With scheduling, if the request has completed, we'll
3452                  * get a NULL return here, as we clear the sched tag when
3453                  * that happens. The request still remains valid, like always,
3454                  * so we should be safe with just the NULL check.
3455                  */
3456                 if (!rq)
3457                         return false;
3458         }
3459
3460         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3461 }
3462
3463 /**
3464  * blk_poll - poll for IO completions
3465  * @q:  the queue
3466  * @cookie: cookie passed back at IO submission time
3467  * @spin: whether to spin for completions
3468  *
3469  * Description:
3470  *    Poll for completions on the passed in queue. Returns number of
3471  *    completed entries found. If @spin is true, then blk_poll will continue
3472  *    looping until at least one completion is found, unless the task is
3473  *    otherwise marked running (or we need to reschedule).
3474  */
3475 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3476 {
3477         struct blk_mq_hw_ctx *hctx;
3478         long state;
3479
3480         if (!blk_qc_t_valid(cookie) ||
3481             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3482                 return 0;
3483
3484         if (current->plug)
3485                 blk_flush_plug_list(current->plug, false);
3486
3487         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3488
3489         /*
3490          * If we sleep, have the caller restart the poll loop to reset
3491          * the state. Like for the other success return cases, the
3492          * caller is responsible for checking if the IO completed. If
3493          * the IO isn't complete, we'll get called again and will go
3494          * straight to the busy poll loop.
3495          */
3496         if (blk_mq_poll_hybrid(q, hctx, cookie))
3497                 return 1;
3498
3499         hctx->poll_considered++;
3500
3501         state = current->state;
3502         do {
3503                 int ret;
3504
3505                 hctx->poll_invoked++;
3506
3507                 ret = q->mq_ops->poll(hctx);
3508                 if (ret > 0) {
3509                         hctx->poll_success++;
3510                         __set_current_state(TASK_RUNNING);
3511                         return ret;
3512                 }
3513
3514                 if (signal_pending_state(state, current))
3515                         __set_current_state(TASK_RUNNING);
3516
3517                 if (current->state == TASK_RUNNING)
3518                         return 1;
3519                 if (ret < 0 || !spin)
3520                         break;
3521                 cpu_relax();
3522         } while (!need_resched());
3523
3524         __set_current_state(TASK_RUNNING);
3525         return 0;
3526 }
3527 EXPORT_SYMBOL_GPL(blk_poll);
3528
3529 unsigned int blk_mq_rq_cpu(struct request *rq)
3530 {
3531         return rq->mq_ctx->cpu;
3532 }
3533 EXPORT_SYMBOL(blk_mq_rq_cpu);
3534
3535 static int __init blk_mq_init(void)
3536 {
3537         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3538                                 blk_mq_hctx_notify_dead);
3539         return 0;
3540 }
3541 subsys_initcall(blk_mq_init);