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