block/blk-mq.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Block multiqueue core code
   4  *
   5  * Copyright (C) 2013-2014 Jens Axboe
   6  * Copyright (C) 2013-2014 Christoph Hellwig
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/backing-dev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/kmemleak.h>
  14 #include <linux/mm.h>
  15 #include <linux/init.h>
  16 #include <linux/slab.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/smp.h>
  19 #include <linux/llist.h>
  20 #include <linux/list_sort.h>
  21 #include <linux/cpu.h>
  22 #include <linux/cache.h>
  23 #include <linux/sched/sysctl.h>
  24 #include <linux/sched/topology.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/delay.h>
  27 #include <linux/crash_dump.h>
  28 #include <linux/prefetch.h>
  29
  30 #include <trace/events/block.h>
  31
  32 #include <linux/blk-mq.h>
  33 #include <linux/t10-pi.h>
  34 #include "blk.h"
  35 #include "blk-mq.h"
  36 #include "blk-mq-debugfs.h"
  37 #include "blk-mq-tag.h"
  38 #include "blk-pm.h"
  39 #include "blk-stat.h"
  40 #include "blk-mq-sched.h"
  41 #include "blk-rq-qos.h"
  42
  43 static void blk_mq_poll_stats_start(struct request_queue *q);
  44 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  45
  46 static int blk_mq_poll_stats_bkt(const struct request *rq)
  47 {
  48         int ddir, sectors, bucket;
  49
  50         ddir = rq_data_dir(rq);
  51         sectors = blk_rq_stats_sectors(rq);
  52
  53         bucket = ddir + 2 * ilog2(sectors);
  54
  55         if (bucket < 0)
  56                 return -1;
  57         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  58                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  59
  60         return bucket;
  61 }
  62
  63 /*
  64  * Check if any of the ctx, dispatch list or elevator
  65  * have pending work in this hardware queue.
  66  */
  67 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  68 {
  69         return !list_empty_careful(&hctx->dispatch) ||
  70                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  71                         blk_mq_sched_has_work(hctx);
  72 }
  73
  74 /*
  75  * Mark this ctx as having pending work in this hardware queue
  76  */
  77 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  78                                      struct blk_mq_ctx *ctx)
  79 {
  80         const int bit = ctx->index_hw[hctx->type];
  81
  82         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
  83                 sbitmap_set_bit(&hctx->ctx_map, bit);
  84 }
  85
  86 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  87                                       struct blk_mq_ctx *ctx)
  88 {
  89         const int bit = ctx->index_hw[hctx->type];
  90
  91         sbitmap_clear_bit(&hctx->ctx_map, bit);
  92 }
  93
  94 struct mq_inflight {
  95         struct hd_struct *part;
  96         unsigned int *inflight;
  97 };
  98
  99 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
 100                                   struct request *rq, void *priv,
 101                                   bool reserved)
 102 {
 103         struct mq_inflight *mi = priv;
 104
 105         /*
 106          * index[0] counts the specific partition that was asked for.
 107          */
 108         if (rq->part == mi->part)
 109                 mi->inflight[0]++;
 110
 111         return true;
 112 }
 113
 114 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
 115 {
 116         unsigned inflight[2];
 117         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 118
 119         inflight[0] = inflight[1] = 0;
 120         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 121
 122         return inflight[0];
 123 }
 124
 125 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
 126                                      struct request *rq, void *priv,
 127                                      bool reserved)
 128 {
 129         struct mq_inflight *mi = priv;
 130
 131         if (rq->part == mi->part)
 132                 mi->inflight[rq_data_dir(rq)]++;
 133
 134         return true;
 135 }
 136
 137 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 138                          unsigned int inflight[2])
 139 {
 140         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 141
 142         inflight[0] = inflight[1] = 0;
 143         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
 144 }
 145
 146 void blk_freeze_queue_start(struct request_queue *q)
 147 {
 148         mutex_lock(&q->mq_freeze_lock);
 149         if (++q->mq_freeze_depth == 1) {
 150                 percpu_ref_kill(&q->q_usage_counter);
 151                 mutex_unlock(&q->mq_freeze_lock);
 152                 if (queue_is_mq(q))
 153                         blk_mq_run_hw_queues(q, false);
 154         } else {
 155                 mutex_unlock(&q->mq_freeze_lock);
 156         }
 157 }
 158 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 159
 160 void blk_mq_freeze_queue_wait(struct request_queue *q)
 161 {
 162         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 163 }
 164 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 165
 166 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 167                                      unsigned long timeout)
 168 {
 169         return wait_event_timeout(q->mq_freeze_wq,
 170                                         percpu_ref_is_zero(&q->q_usage_counter),
 171                                         timeout);
 172 }
 173 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 174
 175 /*
 176  * Guarantee no request is in use, so we can change any data structure of
 177  * the queue afterward.
 178  */
 179 void blk_freeze_queue(struct request_queue *q)
 180 {
 181         /*
 182          * In the !blk_mq case we are only calling this to kill the
 183          * q_usage_counter, otherwise this increases the freeze depth
 184          * and waits for it to return to zero.  For this reason there is
 185          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 186          * exported to drivers as the only user for unfreeze is blk_mq.
 187          */
 188         blk_freeze_queue_start(q);
 189         blk_mq_freeze_queue_wait(q);
 190 }
 191
 192 void blk_mq_freeze_queue(struct request_queue *q)
 193 {
 194         /*
 195          * ...just an alias to keep freeze and unfreeze actions balanced
 196          * in the blk_mq_* namespace
 197          */
 198         blk_freeze_queue(q);
 199 }
 200 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 201
 202 void blk_mq_unfreeze_queue(struct request_queue *q)
 203 {
 204         mutex_lock(&q->mq_freeze_lock);
 205         q->mq_freeze_depth--;
 206         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 207         if (!q->mq_freeze_depth) {
 208                 percpu_ref_resurrect(&q->q_usage_counter);
 209                 wake_up_all(&q->mq_freeze_wq);
 210         }
 211         mutex_unlock(&q->mq_freeze_lock);
 212 }
 213 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 214
 215 /*
 216  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 217  * mpt3sas driver such that this function can be removed.
 218  */
 219 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 220 {
 221         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 222 }
 223 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 224
 225 /**
 226  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 227  * @q: request queue.
 228  *
 229  * Note: this function does not prevent that the struct request end_io()
 230  * callback function is invoked. Once this function is returned, we make
 231  * sure no dispatch can happen until the queue is unquiesced via
 232  * blk_mq_unquiesce_queue().
 233  */
 234 void blk_mq_quiesce_queue(struct request_queue *q)
 235 {
 236         struct blk_mq_hw_ctx *hctx;
 237         unsigned int i;
 238         bool rcu = false;
 239
 240         blk_mq_quiesce_queue_nowait(q);
 241
 242         queue_for_each_hw_ctx(q, hctx, i) {
 243                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 244                         synchronize_srcu(hctx->srcu);
 245                 else
 246                         rcu = true;
 247         }
 248         if (rcu)
 249                 synchronize_rcu();
 250 }
 251 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 252
 253 /*
 254  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 255  * @q: request queue.
 256  *
 257  * This function recovers queue into the state before quiescing
 258  * which is done by blk_mq_quiesce_queue.
 259  */
 260 void blk_mq_unquiesce_queue(struct request_queue *q)
 261 {
 262         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 263
 264         /* dispatch requests which are inserted during quiescing */
 265         blk_mq_run_hw_queues(q, true);
 266 }
 267 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 268
 269 void blk_mq_wake_waiters(struct request_queue *q)
 270 {
 271         struct blk_mq_hw_ctx *hctx;
 272         unsigned int i;
 273
 274         queue_for_each_hw_ctx(q, hctx, i)
 275                 if (blk_mq_hw_queue_mapped(hctx))
 276                         blk_mq_tag_wakeup_all(hctx->tags, true);
 277 }
 278
 279 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 280 {
 281         return blk_mq_has_free_tags(hctx->tags);
 282 }
 283 EXPORT_SYMBOL(blk_mq_can_queue);
 284
 285 /*
 286  * Only need start/end time stamping if we have iostat or
 287  * blk stats enabled, or using an IO scheduler.
 288  */
 289 static inline bool blk_mq_need_time_stamp(struct request *rq)
 290 {
 291         return (rq->rq_flags & (RQF_IO_STAT | RQF_STATS)) || rq->q->elevator;
 292 }
 293
 294 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 295                 unsigned int tag, unsigned int op, u64 alloc_time_ns)
 296 {
 297         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 298         struct request *rq = tags->static_rqs[tag];
 299         req_flags_t rq_flags = 0;
 300
 301         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 302                 rq->tag = -1;
 303                 rq->internal_tag = tag;
 304         } else {
 305                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
 306                         rq_flags = RQF_MQ_INFLIGHT;
 307                         atomic_inc(&data->hctx->nr_active);
 308                 }
 309                 rq->tag = tag;
 310                 rq->internal_tag = -1;
 311                 data->hctx->tags->rqs[rq->tag] = rq;
 312         }
 313
 314         /* csd/requeue_work/fifo_time is initialized before use */
 315         rq->q = data->q;
 316         rq->mq_ctx = data->ctx;
 317         rq->mq_hctx = data->hctx;
 318         rq->rq_flags = rq_flags;
 319         rq->cmd_flags = op;
 320         if (data->flags & BLK_MQ_REQ_PREEMPT)
 321                 rq->rq_flags |= RQF_PREEMPT;
 322         if (blk_queue_io_stat(data->q))
 323                 rq->rq_flags |= RQF_IO_STAT;
 324         INIT_LIST_HEAD(&rq->queuelist);
 325         INIT_HLIST_NODE(&rq->hash);
 326         RB_CLEAR_NODE(&rq->rb_node);
 327         rq->rq_disk = NULL;
 328         rq->part = NULL;
 329 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
 330         rq->alloc_time_ns = alloc_time_ns;
 331 #endif
 332         if (blk_mq_need_time_stamp(rq))
 333                 rq->start_time_ns = ktime_get_ns();
 334         else
 335                 rq->start_time_ns = 0;
 336         rq->io_start_time_ns = 0;
 337         rq->stats_sectors = 0;
 338         rq->nr_phys_segments = 0;
 339 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 340         rq->nr_integrity_segments = 0;
 341 #endif
 342         /* tag was already set */
 343         rq->extra_len = 0;
 344         WRITE_ONCE(rq->deadline, 0);
 345
 346         rq->timeout = 0;
 347
 348         rq->end_io = NULL;
 349         rq->end_io_data = NULL;
 350
 351         data->ctx->rq_dispatched[op_is_sync(op)]++;
 352         refcount_set(&rq->ref, 1);
 353         return rq;
 354 }
 355
 356 static struct request *blk_mq_get_request(struct request_queue *q,
 357                                           struct bio *bio,
 358                                           struct blk_mq_alloc_data *data)
 359 {
 360         struct elevator_queue *e = q->elevator;
 361         struct request *rq;
 362         unsigned int tag;
 363         bool clear_ctx_on_error = false;
 364         u64 alloc_time_ns = 0;
 365
 366         blk_queue_enter_live(q);
 367
 368         /* alloc_time includes depth and tag waits */
 369         if (blk_queue_rq_alloc_time(q))
 370                 alloc_time_ns = ktime_get_ns();
 371
 372         data->q = q;
 373         if (likely(!data->ctx)) {
 374                 data->ctx = blk_mq_get_ctx(q);
 375                 clear_ctx_on_error = true;
 376         }
 377         if (likely(!data->hctx))
 378                 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
 379                                                 data->ctx);
 380         if (data->cmd_flags & REQ_NOWAIT)
 381                 data->flags |= BLK_MQ_REQ_NOWAIT;
 382
 383         if (e) {
 384                 data->flags |= BLK_MQ_REQ_INTERNAL;
 385
 386                 /*
 387                  * Flush requests are special and go directly to the
 388                  * dispatch list. Don't include reserved tags in the
 389                  * limiting, as it isn't useful.
 390                  */
 391                 if (!op_is_flush(data->cmd_flags) &&
 392                     e->type->ops.limit_depth &&
 393                     !(data->flags & BLK_MQ_REQ_RESERVED))
 394                         e->type->ops.limit_depth(data->cmd_flags, data);
 395         } else {
 396                 blk_mq_tag_busy(data->hctx);
 397         }
 398
 399         tag = blk_mq_get_tag(data);
 400         if (tag == BLK_MQ_TAG_FAIL) {
 401                 if (clear_ctx_on_error)
 402                         data->ctx = NULL;
 403                 blk_queue_exit(q);
 404                 return NULL;
 405         }
 406
 407         rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
 408         if (!op_is_flush(data->cmd_flags)) {
 409                 rq->elv.icq = NULL;
 410                 if (e && e->type->ops.prepare_request) {
 411                         if (e->type->icq_cache)
 412                                 blk_mq_sched_assign_ioc(rq);
 413
 414                         e->type->ops.prepare_request(rq, bio);
 415                         rq->rq_flags |= RQF_ELVPRIV;
 416                 }
 417         }
 418         data->hctx->queued++;
 419         return rq;
 420 }
 421
 422 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 423                 blk_mq_req_flags_t flags)
 424 {
 425         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 426         struct request *rq;
 427         int ret;
 428
 429         ret = blk_queue_enter(q, flags);
 430         if (ret)
 431                 return ERR_PTR(ret);
 432
 433         rq = blk_mq_get_request(q, NULL, &alloc_data);
 434         blk_queue_exit(q);
 435
 436         if (!rq)
 437                 return ERR_PTR(-EWOULDBLOCK);
 438
 439         rq->__data_len = 0;
 440         rq->__sector = (sector_t) -1;
 441         rq->bio = rq->biotail = NULL;
 442         return rq;
 443 }
 444 EXPORT_SYMBOL(blk_mq_alloc_request);
 445
 446 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 447         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 448 {
 449         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 450         struct request *rq;
 451         unsigned int cpu;
 452         int ret;
 453
 454         /*
 455          * If the tag allocator sleeps we could get an allocation for a
 456          * different hardware context.  No need to complicate the low level
 457          * allocator for this for the rare use case of a command tied to
 458          * a specific queue.
 459          */
 460         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 461                 return ERR_PTR(-EINVAL);
 462
 463         if (hctx_idx >= q->nr_hw_queues)
 464                 return ERR_PTR(-EIO);
 465
 466         ret = blk_queue_enter(q, flags);
 467         if (ret)
 468                 return ERR_PTR(ret);
 469
 470         /*
 471          * Check if the hardware context is actually mapped to anything.
 472          * If not tell the caller that it should skip this queue.
 473          */
 474         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 475         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 476                 blk_queue_exit(q);
 477                 return ERR_PTR(-EXDEV);
 478         }
 479         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 480         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 481
 482         rq = blk_mq_get_request(q, NULL, &alloc_data);
 483         blk_queue_exit(q);
 484
 485         if (!rq)
 486                 return ERR_PTR(-EWOULDBLOCK);
 487
 488         return rq;
 489 }
 490 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 491
 492 static void __blk_mq_free_request(struct request *rq)
 493 {
 494         struct request_queue *q = rq->q;
 495         struct blk_mq_ctx *ctx = rq->mq_ctx;
 496         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 497         const int sched_tag = rq->internal_tag;
 498
 499         blk_pm_mark_last_busy(rq);
 500         rq->mq_hctx = NULL;
 501         if (rq->tag != -1)
 502                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 503         if (sched_tag != -1)
 504                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 505         blk_mq_sched_restart(hctx);
 506         blk_queue_exit(q);
 507 }
 508
 509 void blk_mq_free_request(struct request *rq)
 510 {
 511         struct request_queue *q = rq->q;
 512         struct elevator_queue *e = q->elevator;
 513         struct blk_mq_ctx *ctx = rq->mq_ctx;
 514         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 515
 516         if (rq->rq_flags & RQF_ELVPRIV) {
 517                 if (e && e->type->ops.finish_request)
 518                         e->type->ops.finish_request(rq);
 519                 if (rq->elv.icq) {
 520                         put_io_context(rq->elv.icq->ioc);
 521                         rq->elv.icq = NULL;
 522                 }
 523         }
 524
 525         ctx->rq_completed[rq_is_sync(rq)]++;
 526         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 527                 atomic_dec(&hctx->nr_active);
 528
 529         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 530                 laptop_io_completion(q->backing_dev_info);
 531
 532         rq_qos_done(q, rq);
 533
 534         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 535         if (refcount_dec_and_test(&rq->ref))
 536                 __blk_mq_free_request(rq);
 537 }
 538 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 539
 540 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 541 {
 542         u64 now = 0;
 543
 544         if (blk_mq_need_time_stamp(rq))
 545                 now = ktime_get_ns();
 546
 547         if (rq->rq_flags & RQF_STATS) {
 548                 blk_mq_poll_stats_start(rq->q);
 549                 blk_stat_add(rq, now);
 550         }
 551
 552         if (rq->internal_tag != -1)
 553                 blk_mq_sched_completed_request(rq, now);
 554
 555         blk_account_io_done(rq, now);
 556
 557         if (rq->end_io) {
 558                 rq_qos_done(rq->q, rq);
 559                 rq->end_io(rq, error);
 560         } else {
 561                 blk_mq_free_request(rq);
 562         }
 563 }
 564 EXPORT_SYMBOL(__blk_mq_end_request);
 565
 566 void blk_mq_end_request(struct request *rq, blk_status_t error)
 567 {
 568         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 569                 BUG();
 570         __blk_mq_end_request(rq, error);
 571 }
 572 EXPORT_SYMBOL(blk_mq_end_request);
 573
 574 static void __blk_mq_complete_request_remote(void *data)
 575 {
 576         struct request *rq = data;
 577         struct request_queue *q = rq->q;
 578
 579         q->mq_ops->complete(rq);
 580 }
 581
 582 static void __blk_mq_complete_request(struct request *rq)
 583 {
 584         struct blk_mq_ctx *ctx = rq->mq_ctx;
 585         struct request_queue *q = rq->q;
 586         bool shared = false;
 587         int cpu;
 588
 589         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 590         /*
 591          * Most of single queue controllers, there is only one irq vector
 592          * for handling IO completion, and the only irq's affinity is set
 593          * as all possible CPUs. On most of ARCHs, this affinity means the
 594          * irq is handled on one specific CPU.
 595          *
 596          * So complete IO reqeust in softirq context in case of single queue
 597          * for not degrading IO performance by irqsoff latency.
 598          */
 599         if (q->nr_hw_queues == 1) {
 600                 __blk_complete_request(rq);
 601                 return;
 602         }
 603
 604         /*
 605          * For a polled request, always complete locallly, it's pointless
 606          * to redirect the completion.
 607          */
 608         if ((rq->cmd_flags & REQ_HIPRI) ||
 609             !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
 610                 q->mq_ops->complete(rq);
 611                 return;
 612         }
 613
 614         cpu = get_cpu();
 615         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
 616                 shared = cpus_share_cache(cpu, ctx->cpu);
 617
 618         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 619                 rq->csd.func = __blk_mq_complete_request_remote;
 620                 rq->csd.info = rq;
 621                 rq->csd.flags = 0;
 622                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 623         } else {
 624                 q->mq_ops->complete(rq);
 625         }
 626         put_cpu();
 627 }
 628
 629 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 630         __releases(hctx->srcu)
 631 {
 632         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 633                 rcu_read_unlock();
 634         else
 635                 srcu_read_unlock(hctx->srcu, srcu_idx);
 636 }
 637
 638 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 639         __acquires(hctx->srcu)
 640 {
 641         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 642                 /* shut up gcc false positive */
 643                 *srcu_idx = 0;
 644                 rcu_read_lock();
 645         } else
 646                 *srcu_idx = srcu_read_lock(hctx->srcu);
 647 }
 648
 649 /**
 650  * blk_mq_complete_request - end I/O on a request
 651  * @rq:         the request being processed
 652  *
 653  * Description:
 654  *      Ends all I/O on a request. It does not handle partial completions.
 655  *      The actual completion happens out-of-order, through a IPI handler.
 656  **/
 657 bool blk_mq_complete_request(struct request *rq)
 658 {
 659         if (unlikely(blk_should_fake_timeout(rq->q)))
 660                 return false;
 661         __blk_mq_complete_request(rq);
 662         return true;
 663 }
 664 EXPORT_SYMBOL(blk_mq_complete_request);
 665
 666 int blk_mq_request_started(struct request *rq)
 667 {
 668         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 669 }
 670 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 671
 672 int blk_mq_request_completed(struct request *rq)
 673 {
 674         return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE;
 675 }
 676 EXPORT_SYMBOL_GPL(blk_mq_request_completed);
 677
 678 void blk_mq_start_request(struct request *rq)
 679 {
 680         struct request_queue *q = rq->q;
 681
 682         trace_block_rq_issue(q, rq);
 683
 684         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 685                 rq->io_start_time_ns = ktime_get_ns();
 686                 rq->stats_sectors = blk_rq_sectors(rq);
 687                 rq->rq_flags |= RQF_STATS;
 688                 rq_qos_issue(q, rq);
 689         }
 690
 691         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 692
 693         blk_add_timer(rq);
 694         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 695
 696         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 697                 /*
 698                  * Make sure space for the drain appears.  We know we can do
 699                  * this because max_hw_segments has been adjusted to be one
 700                  * fewer than the device can handle.
 701                  */
 702                 rq->nr_phys_segments++;
 703         }
 704
 705 #ifdef CONFIG_BLK_DEV_INTEGRITY
 706         if (blk_integrity_rq(rq) && req_op(rq) == REQ_OP_WRITE)
 707                 q->integrity.profile->prepare_fn(rq);
 708 #endif
 709 }
 710 EXPORT_SYMBOL(blk_mq_start_request);
 711
 712 static void __blk_mq_requeue_request(struct request *rq)
 713 {
 714         struct request_queue *q = rq->q;
 715
 716         blk_mq_put_driver_tag(rq);
 717
 718         trace_block_rq_requeue(q, rq);
 719         rq_qos_requeue(q, rq);
 720
 721         if (blk_mq_request_started(rq)) {
 722                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 723                 rq->rq_flags &= ~RQF_TIMED_OUT;
 724                 if (q->dma_drain_size && blk_rq_bytes(rq))
 725                         rq->nr_phys_segments--;
 726         }
 727 }
 728
 729 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 730 {
 731         __blk_mq_requeue_request(rq);
 732
 733         /* this request will be re-inserted to io scheduler queue */
 734         blk_mq_sched_requeue_request(rq);
 735
 736         BUG_ON(!list_empty(&rq->queuelist));
 737         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 738 }
 739 EXPORT_SYMBOL(blk_mq_requeue_request);
 740
 741 static void blk_mq_requeue_work(struct work_struct *work)
 742 {
 743         struct request_queue *q =
 744                 container_of(work, struct request_queue, requeue_work.work);
 745         LIST_HEAD(rq_list);
 746         struct request *rq, *next;
 747
 748         spin_lock_irq(&q->requeue_lock);
 749         list_splice_init(&q->requeue_list, &rq_list);
 750         spin_unlock_irq(&q->requeue_lock);
 751
 752         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 753                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 754                         continue;
 755
 756                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 757                 list_del_init(&rq->queuelist);
 758                 /*
 759                  * If RQF_DONTPREP, rq has contained some driver specific
 760                  * data, so insert it to hctx dispatch list to avoid any
 761                  * merge.
 762                  */
 763                 if (rq->rq_flags & RQF_DONTPREP)
 764                         blk_mq_request_bypass_insert(rq, false, false);
 765                 else
 766                         blk_mq_sched_insert_request(rq, true, false, false);
 767         }
 768
 769         while (!list_empty(&rq_list)) {
 770                 rq = list_entry(rq_list.next, struct request, queuelist);
 771                 list_del_init(&rq->queuelist);
 772                 blk_mq_sched_insert_request(rq, false, false, false);
 773         }
 774
 775         blk_mq_run_hw_queues(q, false);
 776 }
 777
 778 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 779                                 bool kick_requeue_list)
 780 {
 781         struct request_queue *q = rq->q;
 782         unsigned long flags;
 783
 784         /*
 785          * We abuse this flag that is otherwise used by the I/O scheduler to
 786          * request head insertion from the workqueue.
 787          */
 788         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 789
 790         spin_lock_irqsave(&q->requeue_lock, flags);
 791         if (at_head) {
 792                 rq->rq_flags |= RQF_SOFTBARRIER;
 793                 list_add(&rq->queuelist, &q->requeue_list);
 794         } else {
 795                 list_add_tail(&rq->queuelist, &q->requeue_list);
 796         }
 797         spin_unlock_irqrestore(&q->requeue_lock, flags);
 798
 799         if (kick_requeue_list)
 800                 blk_mq_kick_requeue_list(q);
 801 }
 802
 803 void blk_mq_kick_requeue_list(struct request_queue *q)
 804 {
 805         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 806 }
 807 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 808
 809 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 810                                     unsigned long msecs)
 811 {
 812         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 813                                     msecs_to_jiffies(msecs));
 814 }
 815 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 816
 817 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 818 {
 819         if (tag < tags->nr_tags) {
 820                 prefetch(tags->rqs[tag]);
 821                 return tags->rqs[tag];
 822         }
 823
 824         return NULL;
 825 }
 826 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 827
 828 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 829                                void *priv, bool reserved)
 830 {
 831         /*
 832          * If we find a request that is inflight and the queue matches,
 833          * we know the queue is busy. Return false to stop the iteration.
 834          */
 835         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 836                 bool *busy = priv;
 837
 838                 *busy = true;
 839                 return false;
 840         }
 841
 842         return true;
 843 }
 844
 845 bool blk_mq_queue_inflight(struct request_queue *q)
 846 {
 847         bool busy = false;
 848
 849         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 850         return busy;
 851 }
 852 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 853
 854 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 855 {
 856         req->rq_flags |= RQF_TIMED_OUT;
 857         if (req->q->mq_ops->timeout) {
 858                 enum blk_eh_timer_return ret;
 859
 860                 ret = req->q->mq_ops->timeout(req, reserved);
 861                 if (ret == BLK_EH_DONE)
 862                         return;
 863                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 864         }
 865
 866         blk_add_timer(req);
 867 }
 868
 869 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 870 {
 871         unsigned long deadline;
 872
 873         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 874                 return false;
 875         if (rq->rq_flags & RQF_TIMED_OUT)
 876                 return false;
 877
 878         deadline = READ_ONCE(rq->deadline);
 879         if (time_after_eq(jiffies, deadline))
 880                 return true;
 881
 882         if (*next == 0)
 883                 *next = deadline;
 884         else if (time_after(*next, deadline))
 885                 *next = deadline;
 886         return false;
 887 }
 888
 889 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 890                 struct request *rq, void *priv, bool reserved)
 891 {
 892         unsigned long *next = priv;
 893
 894         /*
 895          * Just do a quick check if it is expired before locking the request in
 896          * so we're not unnecessarilly synchronizing across CPUs.
 897          */
 898         if (!blk_mq_req_expired(rq, next))
 899                 return true;
 900
 901         /*
 902          * We have reason to believe the request may be expired. Take a
 903          * reference on the request to lock this request lifetime into its
 904          * currently allocated context to prevent it from being reallocated in
 905          * the event the completion by-passes this timeout handler.
 906          *
 907          * If the reference was already released, then the driver beat the
 908          * timeout handler to posting a natural completion.
 909          */
 910         if (!refcount_inc_not_zero(&rq->ref))
 911                 return true;
 912
 913         /*
 914          * The request is now locked and cannot be reallocated underneath the
 915          * timeout handler's processing. Re-verify this exact request is truly
 916          * expired; if it is not expired, then the request was completed and
 917          * reallocated as a new request.
 918          */
 919         if (blk_mq_req_expired(rq, next))
 920                 blk_mq_rq_timed_out(rq, reserved);
 921
 922         if (is_flush_rq(rq, hctx))
 923                 rq->end_io(rq, 0);
 924         else if (refcount_dec_and_test(&rq->ref))
 925                 __blk_mq_free_request(rq);
 926
 927         return true;
 928 }
 929
 930 static void blk_mq_timeout_work(struct work_struct *work)
 931 {
 932         struct request_queue *q =
 933                 container_of(work, struct request_queue, timeout_work);
 934         unsigned long next = 0;
 935         struct blk_mq_hw_ctx *hctx;
 936         int i;
 937
 938         /* A deadlock might occur if a request is stuck requiring a
 939          * timeout at the same time a queue freeze is waiting
 940          * completion, since the timeout code would not be able to
 941          * acquire the queue reference here.
 942          *
 943          * That's why we don't use blk_queue_enter here; instead, we use
 944          * percpu_ref_tryget directly, because we need to be able to
 945          * obtain a reference even in the short window between the queue
 946          * starting to freeze, by dropping the first reference in
 947          * blk_freeze_queue_start, and the moment the last request is
 948          * consumed, marked by the instant q_usage_counter reaches
 949          * zero.
 950          */
 951         if (!percpu_ref_tryget(&q->q_usage_counter))
 952                 return;
 953
 954         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 955
 956         if (next != 0) {
 957                 mod_timer(&q->timeout, next);
 958         } else {
 959                 /*
 960                  * Request timeouts are handled as a forward rolling timer. If
 961                  * we end up here it means that no requests are pending and
 962                  * also that no request has been pending for a while. Mark
 963                  * each hctx as idle.
 964                  */
 965                 queue_for_each_hw_ctx(q, hctx, i) {
 966                         /* the hctx may be unmapped, so check it here */
 967                         if (blk_mq_hw_queue_mapped(hctx))
 968                                 blk_mq_tag_idle(hctx);
 969                 }
 970         }
 971         blk_queue_exit(q);
 972 }
 973
 974 struct flush_busy_ctx_data {
 975         struct blk_mq_hw_ctx *hctx;
 976         struct list_head *list;
 977 };
 978
 979 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 980 {
 981         struct flush_busy_ctx_data *flush_data = data;
 982         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 983         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 984         enum hctx_type type = hctx->type;
 985
 986         spin_lock(&ctx->lock);
 987         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 988         sbitmap_clear_bit(sb, bitnr);
 989         spin_unlock(&ctx->lock);
 990         return true;
 991 }
 992
 993 /*
 994  * Process software queues that have been marked busy, splicing them
 995  * to the for-dispatch
 996  */
 997 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 998 {
 999         struct flush_busy_ctx_data data = {
1000                 .hctx = hctx,
1001                 .list = list,
1002         };
1003
1004         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1005 }
1006 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1007
1008 struct dispatch_rq_data {
1009         struct blk_mq_hw_ctx *hctx;
1010         struct request *rq;
1011 };
1012
1013 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1014                 void *data)
1015 {
1016         struct dispatch_rq_data *dispatch_data = data;
1017         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1018         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1019         enum hctx_type type = hctx->type;
1020
1021         spin_lock(&ctx->lock);
1022         if (!list_empty(&ctx->rq_lists[type])) {
1023                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1024                 list_del_init(&dispatch_data->rq->queuelist);
1025                 if (list_empty(&ctx->rq_lists[type]))
1026                         sbitmap_clear_bit(sb, bitnr);
1027         }
1028         spin_unlock(&ctx->lock);
1029
1030         return !dispatch_data->rq;
1031 }
1032
1033 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1034                                         struct blk_mq_ctx *start)
1035 {
1036         unsigned off = start ? start->index_hw[hctx->type] : 0;
1037         struct dispatch_rq_data data = {
1038                 .hctx = hctx,
1039                 .rq   = NULL,
1040         };
1041
1042         __sbitmap_for_each_set(&hctx->ctx_map, off,
1043                                dispatch_rq_from_ctx, &data);
1044
1045         return data.rq;
1046 }
1047
1048 static inline unsigned int queued_to_index(unsigned int queued)
1049 {
1050         if (!queued)
1051                 return 0;
1052
1053         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1054 }
1055
1056 bool blk_mq_get_driver_tag(struct request *rq)
1057 {
1058         struct blk_mq_alloc_data data = {
1059                 .q = rq->q,
1060                 .hctx = rq->mq_hctx,
1061                 .flags = BLK_MQ_REQ_NOWAIT,
1062                 .cmd_flags = rq->cmd_flags,
1063         };
1064         bool shared;
1065
1066         if (rq->tag != -1)
1067                 goto done;
1068
1069         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1070                 data.flags |= BLK_MQ_REQ_RESERVED;
1071
1072         shared = blk_mq_tag_busy(data.hctx);
1073         rq->tag = blk_mq_get_tag(&data);
1074         if (rq->tag >= 0) {
1075                 if (shared) {
1076                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1077                         atomic_inc(&data.hctx->nr_active);
1078                 }
1079                 data.hctx->tags->rqs[rq->tag] = rq;
1080         }
1081
1082 done:
1083         return rq->tag != -1;
1084 }
1085
1086 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1087                                 int flags, void *key)
1088 {
1089         struct blk_mq_hw_ctx *hctx;
1090
1091         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1092
1093         spin_lock(&hctx->dispatch_wait_lock);
1094         if (!list_empty(&wait->entry)) {
1095                 struct sbitmap_queue *sbq;
1096
1097                 list_del_init(&wait->entry);
1098                 sbq = &hctx->tags->bitmap_tags;
1099                 atomic_dec(&sbq->ws_active);
1100         }
1101         spin_unlock(&hctx->dispatch_wait_lock);
1102
1103         blk_mq_run_hw_queue(hctx, true);
1104         return 1;
1105 }
1106
1107 /*
1108  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1109  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1110  * restart. For both cases, take care to check the condition again after
1111  * marking us as waiting.
1112  */
1113 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1114                                  struct request *rq)
1115 {
1116         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1117         struct wait_queue_head *wq;
1118         wait_queue_entry_t *wait;
1119         bool ret;
1120
1121         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1122                 blk_mq_sched_mark_restart_hctx(hctx);
1123
1124                 /*
1125                  * It's possible that a tag was freed in the window between the
1126                  * allocation failure and adding the hardware queue to the wait
1127                  * queue.
1128                  *
1129                  * Don't clear RESTART here, someone else could have set it.
1130                  * At most this will cost an extra queue run.
1131                  */
1132                 return blk_mq_get_driver_tag(rq);
1133         }
1134
1135         wait = &hctx->dispatch_wait;
1136         if (!list_empty_careful(&wait->entry))
1137                 return false;
1138
1139         wq = &bt_wait_ptr(sbq, hctx)->wait;
1140
1141         spin_lock_irq(&wq->lock);
1142         spin_lock(&hctx->dispatch_wait_lock);
1143         if (!list_empty(&wait->entry)) {
1144                 spin_unlock(&hctx->dispatch_wait_lock);
1145                 spin_unlock_irq(&wq->lock);
1146                 return false;
1147         }
1148
1149         atomic_inc(&sbq->ws_active);
1150         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1151         __add_wait_queue(wq, wait);
1152
1153         /*
1154          * It's possible that a tag was freed in the window between the
1155          * allocation failure and adding the hardware queue to the wait
1156          * queue.
1157          */
1158         ret = blk_mq_get_driver_tag(rq);
1159         if (!ret) {
1160                 spin_unlock(&hctx->dispatch_wait_lock);
1161                 spin_unlock_irq(&wq->lock);
1162                 return false;
1163         }
1164
1165         /*
1166          * We got a tag, remove ourselves from the wait queue to ensure
1167          * someone else gets the wakeup.
1168          */
1169         list_del_init(&wait->entry);
1170         atomic_dec(&sbq->ws_active);
1171         spin_unlock(&hctx->dispatch_wait_lock);
1172         spin_unlock_irq(&wq->lock);
1173
1174         return true;
1175 }
1176
1177 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1178 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1179 /*
1180  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1181  * - EWMA is one simple way to compute running average value
1182  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1183  * - take 4 as factor for avoiding to get too small(0) result, and this
1184  *   factor doesn't matter because EWMA decreases exponentially
1185  */
1186 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1187 {
1188         unsigned int ewma;
1189
1190         if (hctx->queue->elevator)
1191                 return;
1192
1193         ewma = hctx->dispatch_busy;
1194
1195         if (!ewma && !busy)
1196                 return;
1197
1198         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1199         if (busy)
1200                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1201         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1202
1203         hctx->dispatch_busy = ewma;
1204 }
1205
1206 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1207
1208 /*
1209  * Returns true if we did some work AND can potentially do more.
1210  */
1211 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1212                              bool got_budget)
1213 {
1214         struct blk_mq_hw_ctx *hctx;
1215         struct request *rq, *nxt;
1216         bool no_tag = false;
1217         int errors, queued;
1218         blk_status_t ret = BLK_STS_OK;
1219
1220         if (list_empty(list))
1221                 return false;
1222
1223         WARN_ON(!list_is_singular(list) && got_budget);
1224
1225         /*
1226          * Now process all the entries, sending them to the driver.
1227          */
1228         errors = queued = 0;
1229         do {
1230                 struct blk_mq_queue_data bd;
1231
1232                 rq = list_first_entry(list, struct request, queuelist);
1233
1234                 hctx = rq->mq_hctx;
1235                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1236                         blk_mq_put_driver_tag(rq);
1237                         break;
1238                 }
1239
1240                 if (!blk_mq_get_driver_tag(rq)) {
1241                         /*
1242                          * The initial allocation attempt failed, so we need to
1243                          * rerun the hardware queue when a tag is freed. The
1244                          * waitqueue takes care of that. If the queue is run
1245                          * before we add this entry back on the dispatch list,
1246                          * we'll re-run it below.
1247                          */
1248                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1249                                 blk_mq_put_dispatch_budget(hctx);
1250                                 /*
1251                                  * For non-shared tags, the RESTART check
1252                                  * will suffice.
1253                                  */
1254                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1255                                         no_tag = true;
1256                                 break;
1257                         }
1258                 }
1259
1260                 list_del_init(&rq->queuelist);
1261
1262                 bd.rq = rq;
1263
1264                 /*
1265                  * Flag last if we have no more requests, or if we have more
1266                  * but can't assign a driver tag to it.
1267                  */
1268                 if (list_empty(list))
1269                         bd.last = true;
1270                 else {
1271                         nxt = list_first_entry(list, struct request, queuelist);
1272                         bd.last = !blk_mq_get_driver_tag(nxt);
1273                 }
1274
1275                 ret = q->mq_ops->queue_rq(hctx, &bd);
1276                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1277                         /*
1278                          * If an I/O scheduler has been configured and we got a
1279                          * driver tag for the next request already, free it
1280                          * again.
1281                          */
1282                         if (!list_empty(list)) {
1283                                 nxt = list_first_entry(list, struct request, queuelist);
1284                                 blk_mq_put_driver_tag(nxt);
1285                         }
1286                         list_add(&rq->queuelist, list);
1287                         __blk_mq_requeue_request(rq);
1288                         break;
1289                 }
1290
1291                 if (unlikely(ret != BLK_STS_OK)) {
1292                         errors++;
1293                         blk_mq_end_request(rq, BLK_STS_IOERR);
1294                         continue;
1295                 }
1296
1297                 queued++;
1298         } while (!list_empty(list));
1299
1300         hctx->dispatched[queued_to_index(queued)]++;
1301
1302         /*
1303          * Any items that need requeuing? Stuff them into hctx->dispatch,
1304          * that is where we will continue on next queue run.
1305          */
1306         if (!list_empty(list)) {
1307                 bool needs_restart;
1308
1309                 /*
1310                  * If we didn't flush the entire list, we could have told
1311                  * the driver there was more coming, but that turned out to
1312                  * be a lie.
1313                  */
1314                 if (q->mq_ops->commit_rqs)
1315                         q->mq_ops->commit_rqs(hctx);
1316
1317                 spin_lock(&hctx->lock);
1318                 list_splice_tail_init(list, &hctx->dispatch);
1319                 spin_unlock(&hctx->lock);
1320
1321                 /*
1322                  * If SCHED_RESTART was set by the caller of this function and
1323                  * it is no longer set that means that it was cleared by another
1324                  * thread and hence that a queue rerun is needed.
1325                  *
1326                  * If 'no_tag' is set, that means that we failed getting
1327                  * a driver tag with an I/O scheduler attached. If our dispatch
1328                  * waitqueue is no longer active, ensure that we run the queue
1329                  * AFTER adding our entries back to the list.
1330                  *
1331                  * If no I/O scheduler has been configured it is possible that
1332                  * the hardware queue got stopped and restarted before requests
1333                  * were pushed back onto the dispatch list. Rerun the queue to
1334                  * avoid starvation. Notes:
1335                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1336                  *   been stopped before rerunning a queue.
1337                  * - Some but not all block drivers stop a queue before
1338                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1339                  *   and dm-rq.
1340                  *
1341                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1342                  * bit is set, run queue after a delay to avoid IO stalls
1343                  * that could otherwise occur if the queue is idle.
1344                  */
1345                 needs_restart = blk_mq_sched_needs_restart(hctx);
1346                 if (!needs_restart ||
1347                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1348                         blk_mq_run_hw_queue(hctx, true);
1349                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1350                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1351
1352                 blk_mq_update_dispatch_busy(hctx, true);
1353                 return false;
1354         } else
1355                 blk_mq_update_dispatch_busy(hctx, false);
1356
1357         /*
1358          * If the host/device is unable to accept more work, inform the
1359          * caller of that.
1360          */
1361         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1362                 return false;
1363
1364         return (queued + errors) != 0;
1365 }
1366
1367 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1368 {
1369         int srcu_idx;
1370
1371         /*
1372          * We should be running this queue from one of the CPUs that
1373          * are mapped to it.
1374          *
1375          * There are at least two related races now between setting
1376          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1377          * __blk_mq_run_hw_queue():
1378          *
1379          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1380          *   but later it becomes online, then this warning is harmless
1381          *   at all
1382          *
1383          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1384          *   but later it becomes offline, then the warning can't be
1385          *   triggered, and we depend on blk-mq timeout handler to
1386          *   handle dispatched requests to this hctx
1387          */
1388         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1389                 cpu_online(hctx->next_cpu)) {
1390                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1391                         raw_smp_processor_id(),
1392                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1393                 dump_stack();
1394         }
1395
1396         /*
1397          * We can't run the queue inline with ints disabled. Ensure that
1398          * we catch bad users of this early.
1399          */
1400         WARN_ON_ONCE(in_interrupt());
1401
1402         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1403
1404         hctx_lock(hctx, &srcu_idx);
1405         blk_mq_sched_dispatch_requests(hctx);
1406         hctx_unlock(hctx, srcu_idx);
1407 }
1408
1409 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1410 {
1411         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1412
1413         if (cpu >= nr_cpu_ids)
1414                 cpu = cpumask_first(hctx->cpumask);
1415         return cpu;
1416 }
1417
1418 /*
1419  * It'd be great if the workqueue API had a way to pass
1420  * in a mask and had some smarts for more clever placement.
1421  * For now we just round-robin here, switching for every
1422  * BLK_MQ_CPU_WORK_BATCH queued items.
1423  */
1424 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1425 {
1426         bool tried = false;
1427         int next_cpu = hctx->next_cpu;
1428
1429         if (hctx->queue->nr_hw_queues == 1)
1430                 return WORK_CPU_UNBOUND;
1431
1432         if (--hctx->next_cpu_batch <= 0) {
1433 select_cpu:
1434                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1435                                 cpu_online_mask);
1436                 if (next_cpu >= nr_cpu_ids)
1437                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1438                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1439         }
1440
1441         /*
1442          * Do unbound schedule if we can't find a online CPU for this hctx,
1443          * and it should only happen in the path of handling CPU DEAD.
1444          */
1445         if (!cpu_online(next_cpu)) {
1446                 if (!tried) {
1447                         tried = true;
1448                         goto select_cpu;
1449                 }
1450
1451                 /*
1452                  * Make sure to re-select CPU next time once after CPUs
1453                  * in hctx->cpumask become online again.
1454                  */
1455                 hctx->next_cpu = next_cpu;
1456                 hctx->next_cpu_batch = 1;
1457                 return WORK_CPU_UNBOUND;
1458         }
1459
1460         hctx->next_cpu = next_cpu;
1461         return next_cpu;
1462 }
1463
1464 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1465                                         unsigned long msecs)
1466 {
1467         if (unlikely(blk_mq_hctx_stopped(hctx)))
1468                 return;
1469
1470         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1471                 int cpu = get_cpu();
1472                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1473                         __blk_mq_run_hw_queue(hctx);
1474                         put_cpu();
1475                         return;
1476                 }
1477
1478                 put_cpu();
1479         }
1480
1481         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1482                                     msecs_to_jiffies(msecs));
1483 }
1484
1485 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1486 {
1487         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1488 }
1489 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1490
1491 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1492 {
1493         int srcu_idx;
1494         bool need_run;
1495
1496         /*
1497          * When queue is quiesced, we may be switching io scheduler, or
1498          * updating nr_hw_queues, or other things, and we can't run queue
1499          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1500          *
1501          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1502          * quiesced.
1503          */
1504         hctx_lock(hctx, &srcu_idx);
1505         need_run = !blk_queue_quiesced(hctx->queue) &&
1506                 blk_mq_hctx_has_pending(hctx);
1507         hctx_unlock(hctx, srcu_idx);
1508
1509         if (need_run) {
1510                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1511                 return true;
1512         }
1513
1514         return false;
1515 }
1516 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1517
1518 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1519 {
1520         struct blk_mq_hw_ctx *hctx;
1521         int i;
1522
1523         queue_for_each_hw_ctx(q, hctx, i) {
1524                 if (blk_mq_hctx_stopped(hctx))
1525                         continue;
1526
1527                 blk_mq_run_hw_queue(hctx, async);
1528         }
1529 }
1530 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1531
1532 /**
1533  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1534  * @q: request queue.
1535  *
1536  * The caller is responsible for serializing this function against
1537  * blk_mq_{start,stop}_hw_queue().
1538  */
1539 bool blk_mq_queue_stopped(struct request_queue *q)
1540 {
1541         struct blk_mq_hw_ctx *hctx;
1542         int i;
1543
1544         queue_for_each_hw_ctx(q, hctx, i)
1545                 if (blk_mq_hctx_stopped(hctx))
1546                         return true;
1547
1548         return false;
1549 }
1550 EXPORT_SYMBOL(blk_mq_queue_stopped);
1551
1552 /*
1553  * This function is often used for pausing .queue_rq() by driver when
1554  * there isn't enough resource or some conditions aren't satisfied, and
1555  * BLK_STS_RESOURCE is usually returned.
1556  *
1557  * We do not guarantee that dispatch can be drained or blocked
1558  * after blk_mq_stop_hw_queue() returns. Please use
1559  * blk_mq_quiesce_queue() for that requirement.
1560  */
1561 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1562 {
1563         cancel_delayed_work(&hctx->run_work);
1564
1565         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1566 }
1567 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1568
1569 /*
1570  * This function is often used for pausing .queue_rq() by driver when
1571  * there isn't enough resource or some conditions aren't satisfied, and
1572  * BLK_STS_RESOURCE is usually returned.
1573  *
1574  * We do not guarantee that dispatch can be drained or blocked
1575  * after blk_mq_stop_hw_queues() returns. Please use
1576  * blk_mq_quiesce_queue() for that requirement.
1577  */
1578 void blk_mq_stop_hw_queues(struct request_queue *q)
1579 {
1580         struct blk_mq_hw_ctx *hctx;
1581         int i;
1582
1583         queue_for_each_hw_ctx(q, hctx, i)
1584                 blk_mq_stop_hw_queue(hctx);
1585 }
1586 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1587
1588 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1589 {
1590         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1591
1592         blk_mq_run_hw_queue(hctx, false);
1593 }
1594 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1595
1596 void blk_mq_start_hw_queues(struct request_queue *q)
1597 {
1598         struct blk_mq_hw_ctx *hctx;
1599         int i;
1600
1601         queue_for_each_hw_ctx(q, hctx, i)
1602                 blk_mq_start_hw_queue(hctx);
1603 }
1604 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1605
1606 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1607 {
1608         if (!blk_mq_hctx_stopped(hctx))
1609                 return;
1610
1611         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1612         blk_mq_run_hw_queue(hctx, async);
1613 }
1614 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1615
1616 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1617 {
1618         struct blk_mq_hw_ctx *hctx;
1619         int i;
1620
1621         queue_for_each_hw_ctx(q, hctx, i)
1622                 blk_mq_start_stopped_hw_queue(hctx, async);
1623 }
1624 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1625
1626 static void blk_mq_run_work_fn(struct work_struct *work)
1627 {
1628         struct blk_mq_hw_ctx *hctx;
1629
1630         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1631
1632         /*
1633          * If we are stopped, don't run the queue.
1634          */
1635         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1636                 return;
1637
1638         __blk_mq_run_hw_queue(hctx);
1639 }
1640
1641 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1642                                             struct request *rq,
1643                                             bool at_head)
1644 {
1645         struct blk_mq_ctx *ctx = rq->mq_ctx;
1646         enum hctx_type type = hctx->type;
1647
1648         lockdep_assert_held(&ctx->lock);
1649
1650         trace_block_rq_insert(hctx->queue, rq);
1651
1652         if (at_head)
1653                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1654         else
1655                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1656 }
1657
1658 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1659                              bool at_head)
1660 {
1661         struct blk_mq_ctx *ctx = rq->mq_ctx;
1662
1663         lockdep_assert_held(&ctx->lock);
1664
1665         __blk_mq_insert_req_list(hctx, rq, at_head);
1666         blk_mq_hctx_mark_pending(hctx, ctx);
1667 }
1668
1669 /*
1670  * Should only be used carefully, when the caller knows we want to
1671  * bypass a potential IO scheduler on the target device.
1672  */
1673 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1674                                   bool run_queue)
1675 {
1676         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1677
1678         spin_lock(&hctx->lock);
1679         if (at_head)
1680                 list_add(&rq->queuelist, &hctx->dispatch);
1681         else
1682                 list_add_tail(&rq->queuelist, &hctx->dispatch);
1683         spin_unlock(&hctx->lock);
1684
1685         if (run_queue)
1686                 blk_mq_run_hw_queue(hctx, false);
1687 }
1688
1689 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1690                             struct list_head *list)
1691
1692 {
1693         struct request *rq;
1694         enum hctx_type type = hctx->type;
1695
1696         /*
1697          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1698          * offline now
1699          */
1700         list_for_each_entry(rq, list, queuelist) {
1701                 BUG_ON(rq->mq_ctx != ctx);
1702                 trace_block_rq_insert(hctx->queue, rq);
1703         }
1704
1705         spin_lock(&ctx->lock);
1706         list_splice_tail_init(list, &ctx->rq_lists[type]);
1707         blk_mq_hctx_mark_pending(hctx, ctx);
1708         spin_unlock(&ctx->lock);
1709 }
1710
1711 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1712 {
1713         struct request *rqa = container_of(a, struct request, queuelist);
1714         struct request *rqb = container_of(b, struct request, queuelist);
1715
1716         if (rqa->mq_ctx < rqb->mq_ctx)
1717                 return -1;
1718         else if (rqa->mq_ctx > rqb->mq_ctx)
1719                 return 1;
1720         else if (rqa->mq_hctx < rqb->mq_hctx)
1721                 return -1;
1722         else if (rqa->mq_hctx > rqb->mq_hctx)
1723                 return 1;
1724
1725         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1726 }
1727
1728 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1729 {
1730         struct blk_mq_hw_ctx *this_hctx;
1731         struct blk_mq_ctx *this_ctx;
1732         struct request_queue *this_q;
1733         struct request *rq;
1734         LIST_HEAD(list);
1735         LIST_HEAD(rq_list);
1736         unsigned int depth;
1737
1738         list_splice_init(&plug->mq_list, &list);
1739
1740         if (plug->rq_count > 2 && plug->multiple_queues)
1741                 list_sort(NULL, &list, plug_rq_cmp);
1742
1743         plug->rq_count = 0;
1744
1745         this_q = NULL;
1746         this_hctx = NULL;
1747         this_ctx = NULL;
1748         depth = 0;
1749
1750         while (!list_empty(&list)) {
1751                 rq = list_entry_rq(list.next);
1752                 list_del_init(&rq->queuelist);
1753                 BUG_ON(!rq->q);
1754                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1755                         if (this_hctx) {
1756                                 trace_block_unplug(this_q, depth, !from_schedule);
1757                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1758                                                                 &rq_list,
1759                                                                 from_schedule);
1760                         }
1761
1762                         this_q = rq->q;
1763                         this_ctx = rq->mq_ctx;
1764                         this_hctx = rq->mq_hctx;
1765                         depth = 0;
1766                 }
1767
1768                 depth++;
1769                 list_add_tail(&rq->queuelist, &rq_list);
1770         }
1771
1772         /*
1773          * If 'this_hctx' is set, we know we have entries to complete
1774          * on 'rq_list'. Do those.
1775          */
1776         if (this_hctx) {
1777                 trace_block_unplug(this_q, depth, !from_schedule);
1778                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1779                                                 from_schedule);
1780         }
1781 }
1782
1783 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1784                 unsigned int nr_segs)
1785 {
1786         if (bio->bi_opf & REQ_RAHEAD)
1787                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1788
1789         rq->__sector = bio->bi_iter.bi_sector;
1790         rq->write_hint = bio->bi_write_hint;
1791         blk_rq_bio_prep(rq, bio, nr_segs);
1792
1793         blk_account_io_start(rq, true);
1794 }
1795
1796 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1797                                             struct request *rq,
1798                                             blk_qc_t *cookie, bool last)
1799 {
1800         struct request_queue *q = rq->q;
1801         struct blk_mq_queue_data bd = {
1802                 .rq = rq,
1803                 .last = last,
1804         };
1805         blk_qc_t new_cookie;
1806         blk_status_t ret;
1807
1808         new_cookie = request_to_qc_t(hctx, rq);
1809
1810         /*
1811          * For OK queue, we are done. For error, caller may kill it.
1812          * Any other error (busy), just add it to our list as we
1813          * previously would have done.
1814          */
1815         ret = q->mq_ops->queue_rq(hctx, &bd);
1816         switch (ret) {
1817         case BLK_STS_OK:
1818                 blk_mq_update_dispatch_busy(hctx, false);
1819                 *cookie = new_cookie;
1820                 break;
1821         case BLK_STS_RESOURCE:
1822         case BLK_STS_DEV_RESOURCE:
1823                 blk_mq_update_dispatch_busy(hctx, true);
1824                 __blk_mq_requeue_request(rq);
1825                 break;
1826         default:
1827                 blk_mq_update_dispatch_busy(hctx, false);
1828                 *cookie = BLK_QC_T_NONE;
1829                 break;
1830         }
1831
1832         return ret;
1833 }
1834
1835 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1836                                                 struct request *rq,
1837                                                 blk_qc_t *cookie,
1838                                                 bool bypass_insert, bool last)
1839 {
1840         struct request_queue *q = rq->q;
1841         bool run_queue = true;
1842
1843         /*
1844          * RCU or SRCU read lock is needed before checking quiesced flag.
1845          *
1846          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1847          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1848          * and avoid driver to try to dispatch again.
1849          */
1850         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1851                 run_queue = false;
1852                 bypass_insert = false;
1853                 goto insert;
1854         }
1855
1856         if (q->elevator && !bypass_insert)
1857                 goto insert;
1858
1859         if (!blk_mq_get_dispatch_budget(hctx))
1860                 goto insert;
1861
1862         if (!blk_mq_get_driver_tag(rq)) {
1863                 blk_mq_put_dispatch_budget(hctx);
1864                 goto insert;
1865         }
1866
1867         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1868 insert:
1869         if (bypass_insert)
1870                 return BLK_STS_RESOURCE;
1871
1872         blk_mq_request_bypass_insert(rq, false, run_queue);
1873         return BLK_STS_OK;
1874 }
1875
1876 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1877                 struct request *rq, blk_qc_t *cookie)
1878 {
1879         blk_status_t ret;
1880         int srcu_idx;
1881
1882         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1883
1884         hctx_lock(hctx, &srcu_idx);
1885
1886         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1887         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1888                 blk_mq_request_bypass_insert(rq, false, true);
1889         else if (ret != BLK_STS_OK)
1890                 blk_mq_end_request(rq, ret);
1891
1892         hctx_unlock(hctx, srcu_idx);
1893 }
1894
1895 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1896 {
1897         blk_status_t ret;
1898         int srcu_idx;
1899         blk_qc_t unused_cookie;
1900         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1901
1902         hctx_lock(hctx, &srcu_idx);
1903         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1904         hctx_unlock(hctx, srcu_idx);
1905
1906         return ret;
1907 }
1908
1909 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1910                 struct list_head *list)
1911 {
1912         while (!list_empty(list)) {
1913                 blk_status_t ret;
1914                 struct request *rq = list_first_entry(list, struct request,
1915                                 queuelist);
1916
1917                 list_del_init(&rq->queuelist);
1918                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1919                 if (ret != BLK_STS_OK) {
1920                         if (ret == BLK_STS_RESOURCE ||
1921                                         ret == BLK_STS_DEV_RESOURCE) {
1922                                 blk_mq_request_bypass_insert(rq, false,
1923                                                         list_empty(list));
1924                                 break;
1925                         }
1926                         blk_mq_end_request(rq, ret);
1927                 }
1928         }
1929
1930         /*
1931          * If we didn't flush the entire list, we could have told
1932          * the driver there was more coming, but that turned out to
1933          * be a lie.
1934          */
1935         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1936                 hctx->queue->mq_ops->commit_rqs(hctx);
1937 }
1938
1939 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1940 {
1941         list_add_tail(&rq->queuelist, &plug->mq_list);
1942         plug->rq_count++;
1943         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1944                 struct request *tmp;
1945
1946                 tmp = list_first_entry(&plug->mq_list, struct request,
1947                                                 queuelist);
1948                 if (tmp->q != rq->q)
1949                         plug->multiple_queues = true;
1950         }
1951 }
1952
1953 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1954 {
1955         const int is_sync = op_is_sync(bio->bi_opf);
1956         const int is_flush_fua = op_is_flush(bio->bi_opf);
1957         struct blk_mq_alloc_data data = { .flags = 0};
1958         struct request *rq;
1959         struct blk_plug *plug;
1960         struct request *same_queue_rq = NULL;
1961         unsigned int nr_segs;
1962         blk_qc_t cookie;
1963
1964         blk_queue_bounce(q, &bio);
1965         __blk_queue_split(q, &bio, &nr_segs);
1966
1967         if (!bio_integrity_prep(bio))
1968                 return BLK_QC_T_NONE;
1969
1970         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1971             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1972                 return BLK_QC_T_NONE;
1973
1974         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1975                 return BLK_QC_T_NONE;
1976
1977         rq_qos_throttle(q, bio);
1978
1979         data.cmd_flags = bio->bi_opf;
1980         rq = blk_mq_get_request(q, bio, &data);
1981         if (unlikely(!rq)) {
1982                 rq_qos_cleanup(q, bio);
1983                 if (bio->bi_opf & REQ_NOWAIT)
1984                         bio_wouldblock_error(bio);
1985                 return BLK_QC_T_NONE;
1986         }
1987
1988         trace_block_getrq(q, bio, bio->bi_opf);
1989
1990         rq_qos_track(q, rq, bio);
1991
1992         cookie = request_to_qc_t(data.hctx, rq);
1993
1994         blk_mq_bio_to_request(rq, bio, nr_segs);
1995
1996         plug = blk_mq_plug(q, bio);
1997         if (unlikely(is_flush_fua)) {
1998                 /* bypass scheduler for flush rq */
1999                 blk_insert_flush(rq);
2000                 blk_mq_run_hw_queue(data.hctx, true);
2001         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2002                                 !blk_queue_nonrot(q))) {
2003                 /*
2004                  * Use plugging if we have a ->commit_rqs() hook as well, as
2005                  * we know the driver uses bd->last in a smart fashion.
2006                  *
2007                  * Use normal plugging if this disk is slow HDD, as sequential
2008                  * IO may benefit a lot from plug merging.
2009                  */
2010                 unsigned int request_count = plug->rq_count;
2011                 struct request *last = NULL;
2012
2013                 if (!request_count)
2014                         trace_block_plug(q);
2015                 else
2016                         last = list_entry_rq(plug->mq_list.prev);
2017
2018                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2019                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2020                         blk_flush_plug_list(plug, false);
2021                         trace_block_plug(q);
2022                 }
2023
2024                 blk_add_rq_to_plug(plug, rq);
2025         } else if (q->elevator) {
2026                 blk_mq_sched_insert_request(rq, false, true, true);
2027         } else if (plug && !blk_queue_nomerges(q)) {
2028                 /*
2029                  * We do limited plugging. If the bio can be merged, do that.
2030                  * Otherwise the existing request in the plug list will be
2031                  * issued. So the plug list will have one request at most
2032                  * The plug list might get flushed before this. If that happens,
2033                  * the plug list is empty, and same_queue_rq is invalid.
2034                  */
2035                 if (list_empty(&plug->mq_list))
2036                         same_queue_rq = NULL;
2037                 if (same_queue_rq) {
2038                         list_del_init(&same_queue_rq->queuelist);
2039                         plug->rq_count--;
2040                 }
2041                 blk_add_rq_to_plug(plug, rq);
2042                 trace_block_plug(q);
2043
2044                 if (same_queue_rq) {
2045                         data.hctx = same_queue_rq->mq_hctx;
2046                         trace_block_unplug(q, 1, true);
2047                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2048                                         &cookie);
2049                 }
2050         } else if ((q->nr_hw_queues > 1 && is_sync) ||
2051                         !data.hctx->dispatch_busy) {
2052                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2053         } else {
2054                 blk_mq_sched_insert_request(rq, false, true, true);
2055         }
2056
2057         return cookie;
2058 }
2059
2060 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2061                      unsigned int hctx_idx)
2062 {
2063         struct page *page;
2064
2065         if (tags->rqs && set->ops->exit_request) {
2066                 int i;
2067
2068                 for (i = 0; i < tags->nr_tags; i++) {
2069                         struct request *rq = tags->static_rqs[i];
2070
2071                         if (!rq)
2072                                 continue;
2073                         set->ops->exit_request(set, rq, hctx_idx);
2074                         tags->static_rqs[i] = NULL;
2075                 }
2076         }
2077
2078         while (!list_empty(&tags->page_list)) {
2079                 page = list_first_entry(&tags->page_list, struct page, lru);
2080                 list_del_init(&page->lru);
2081                 /*
2082                  * Remove kmemleak object previously allocated in
2083                  * blk_mq_alloc_rqs().
2084                  */
2085                 kmemleak_free(page_address(page));
2086                 __free_pages(page, page->private);
2087         }
2088 }
2089
2090 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2091 {
2092         kfree(tags->rqs);
2093         tags->rqs = NULL;
2094         kfree(tags->static_rqs);
2095         tags->static_rqs = NULL;
2096
2097         blk_mq_free_tags(tags);
2098 }
2099
2100 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2101                                         unsigned int hctx_idx,
2102                                         unsigned int nr_tags,
2103                                         unsigned int reserved_tags)
2104 {
2105         struct blk_mq_tags *tags;
2106         int node;
2107
2108         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2109         if (node == NUMA_NO_NODE)
2110                 node = set->numa_node;
2111
2112         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2113                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2114         if (!tags)
2115                 return NULL;
2116
2117         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2118                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2119                                  node);
2120         if (!tags->rqs) {
2121                 blk_mq_free_tags(tags);
2122                 return NULL;
2123         }
2124
2125         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2126                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2127                                         node);
2128         if (!tags->static_rqs) {
2129                 kfree(tags->rqs);
2130                 blk_mq_free_tags(tags);
2131                 return NULL;
2132         }
2133
2134         return tags;
2135 }
2136
2137 static size_t order_to_size(unsigned int order)
2138 {
2139         return (size_t)PAGE_SIZE << order;
2140 }
2141
2142 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2143                                unsigned int hctx_idx, int node)
2144 {
2145         int ret;
2146
2147         if (set->ops->init_request) {
2148                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2149                 if (ret)
2150                         return ret;
2151         }
2152
2153         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2154         return 0;
2155 }
2156
2157 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2158                      unsigned int hctx_idx, unsigned int depth)
2159 {
2160         unsigned int i, j, entries_per_page, max_order = 4;
2161         size_t rq_size, left;
2162         int node;
2163
2164         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2165         if (node == NUMA_NO_NODE)
2166                 node = set->numa_node;
2167
2168         INIT_LIST_HEAD(&tags->page_list);
2169
2170         /*
2171          * rq_size is the size of the request plus driver payload, rounded
2172          * to the cacheline size
2173          */
2174         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2175                                 cache_line_size());
2176         left = rq_size * depth;
2177
2178         for (i = 0; i < depth; ) {
2179                 int this_order = max_order;
2180                 struct page *page;
2181                 int to_do;
2182                 void *p;
2183
2184                 while (this_order && left < order_to_size(this_order - 1))
2185                         this_order--;
2186
2187                 do {
2188                         page = alloc_pages_node(node,
2189                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2190                                 this_order);
2191                         if (page)
2192                                 break;
2193                         if (!this_order--)
2194                                 break;
2195                         if (order_to_size(this_order) < rq_size)
2196                                 break;
2197                 } while (1);
2198
2199                 if (!page)
2200                         goto fail;
2201
2202                 page->private = this_order;
2203                 list_add_tail(&page->lru, &tags->page_list);
2204
2205                 p = page_address(page);
2206                 /*
2207                  * Allow kmemleak to scan these pages as they contain pointers
2208                  * to additional allocations like via ops->init_request().
2209                  */
2210                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2211                 entries_per_page = order_to_size(this_order) / rq_size;
2212                 to_do = min(entries_per_page, depth - i);
2213                 left -= to_do * rq_size;
2214                 for (j = 0; j < to_do; j++) {
2215                         struct request *rq = p;
2216
2217                         tags->static_rqs[i] = rq;
2218                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2219                                 tags->static_rqs[i] = NULL;
2220                                 goto fail;
2221                         }
2222
2223                         p += rq_size;
2224                         i++;
2225                 }
2226         }
2227         return 0;
2228
2229 fail:
2230         blk_mq_free_rqs(set, tags, hctx_idx);
2231         return -ENOMEM;
2232 }
2233
2234 /*
2235  * 'cpu' is going away. splice any existing rq_list entries from this
2236  * software queue to the hw queue dispatch list, and ensure that it
2237  * gets run.
2238  */
2239 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2240 {
2241         struct blk_mq_hw_ctx *hctx;
2242         struct blk_mq_ctx *ctx;
2243         LIST_HEAD(tmp);
2244         enum hctx_type type;
2245
2246         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2247         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2248         type = hctx->type;
2249
2250         spin_lock(&ctx->lock);
2251         if (!list_empty(&ctx->rq_lists[type])) {
2252                 list_splice_init(&ctx->rq_lists[type], &tmp);
2253                 blk_mq_hctx_clear_pending(hctx, ctx);
2254         }
2255         spin_unlock(&ctx->lock);
2256
2257         if (list_empty(&tmp))
2258                 return 0;
2259
2260         spin_lock(&hctx->lock);
2261         list_splice_tail_init(&tmp, &hctx->dispatch);
2262         spin_unlock(&hctx->lock);
2263
2264         blk_mq_run_hw_queue(hctx, true);
2265         return 0;
2266 }
2267
2268 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2269 {
2270         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2271                                             &hctx->cpuhp_dead);
2272 }
2273
2274 /* hctx->ctxs will be freed in queue's release handler */
2275 static void blk_mq_exit_hctx(struct request_queue *q,
2276                 struct blk_mq_tag_set *set,
2277                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2278 {
2279         if (blk_mq_hw_queue_mapped(hctx))
2280                 blk_mq_tag_idle(hctx);
2281
2282         if (set->ops->exit_request)
2283                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2284
2285         if (set->ops->exit_hctx)
2286                 set->ops->exit_hctx(hctx, hctx_idx);
2287
2288         blk_mq_remove_cpuhp(hctx);
2289
2290         spin_lock(&q->unused_hctx_lock);
2291         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2292         spin_unlock(&q->unused_hctx_lock);
2293 }
2294
2295 static void blk_mq_exit_hw_queues(struct request_queue *q,
2296                 struct blk_mq_tag_set *set, int nr_queue)
2297 {
2298         struct blk_mq_hw_ctx *hctx;
2299         unsigned int i;
2300
2301         queue_for_each_hw_ctx(q, hctx, i) {
2302                 if (i == nr_queue)
2303                         break;
2304                 blk_mq_debugfs_unregister_hctx(hctx);
2305                 blk_mq_exit_hctx(q, set, hctx, i);
2306         }
2307 }
2308
2309 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2310 {
2311         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2312
2313         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2314                            __alignof__(struct blk_mq_hw_ctx)) !=
2315                      sizeof(struct blk_mq_hw_ctx));
2316
2317         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2318                 hw_ctx_size += sizeof(struct srcu_struct);
2319
2320         return hw_ctx_size;
2321 }
2322
2323 static int blk_mq_init_hctx(struct request_queue *q,
2324                 struct blk_mq_tag_set *set,
2325                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2326 {
2327         hctx->queue_num = hctx_idx;
2328
2329         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2330
2331         hctx->tags = set->tags[hctx_idx];
2332
2333         if (set->ops->init_hctx &&
2334             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2335                 goto unregister_cpu_notifier;
2336
2337         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2338                                 hctx->numa_node))
2339                 goto exit_hctx;
2340         return 0;
2341
2342  exit_hctx:
2343         if (set->ops->exit_hctx)
2344                 set->ops->exit_hctx(hctx, hctx_idx);
2345  unregister_cpu_notifier:
2346         blk_mq_remove_cpuhp(hctx);
2347         return -1;
2348 }
2349
2350 static struct blk_mq_hw_ctx *
2351 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2352                 int node)
2353 {
2354         struct blk_mq_hw_ctx *hctx;
2355         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2356
2357         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2358         if (!hctx)
2359                 goto fail_alloc_hctx;
2360
2361         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2362                 goto free_hctx;
2363
2364         atomic_set(&hctx->nr_active, 0);
2365         if (node == NUMA_NO_NODE)
2366                 node = set->numa_node;
2367         hctx->numa_node = node;
2368
2369         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2370         spin_lock_init(&hctx->lock);
2371         INIT_LIST_HEAD(&hctx->dispatch);
2372         hctx->queue = q;
2373         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2374
2375         INIT_LIST_HEAD(&hctx->hctx_list);
2376
2377         /*
2378          * Allocate space for all possible cpus to avoid allocation at
2379          * runtime
2380          */
2381         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2382                         gfp, node);
2383         if (!hctx->ctxs)
2384                 goto free_cpumask;
2385
2386         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2387                                 gfp, node))
2388                 goto free_ctxs;
2389         hctx->nr_ctx = 0;
2390
2391         spin_lock_init(&hctx->dispatch_wait_lock);
2392         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2393         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2394
2395         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2396                         gfp);
2397         if (!hctx->fq)
2398                 goto free_bitmap;
2399
2400         if (hctx->flags & BLK_MQ_F_BLOCKING)
2401                 init_srcu_struct(hctx->srcu);
2402         blk_mq_hctx_kobj_init(hctx);
2403
2404         return hctx;
2405
2406  free_bitmap:
2407         sbitmap_free(&hctx->ctx_map);
2408  free_ctxs:
2409         kfree(hctx->ctxs);
2410  free_cpumask:
2411         free_cpumask_var(hctx->cpumask);
2412  free_hctx:
2413         kfree(hctx);
2414  fail_alloc_hctx:
2415         return NULL;
2416 }
2417
2418 static void blk_mq_init_cpu_queues(struct request_queue *q,
2419                                    unsigned int nr_hw_queues)
2420 {
2421         struct blk_mq_tag_set *set = q->tag_set;
2422         unsigned int i, j;
2423
2424         for_each_possible_cpu(i) {
2425                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2426                 struct blk_mq_hw_ctx *hctx;
2427                 int k;
2428
2429                 __ctx->cpu = i;
2430                 spin_lock_init(&__ctx->lock);
2431                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2432                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2433
2434                 __ctx->queue = q;
2435
2436                 /*
2437                  * Set local node, IFF we have more than one hw queue. If
2438                  * not, we remain on the home node of the device
2439                  */
2440                 for (j = 0; j < set->nr_maps; j++) {
2441                         hctx = blk_mq_map_queue_type(q, j, i);
2442                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2443                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2444                 }
2445         }
2446 }
2447
2448 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2449 {
2450         int ret = 0;
2451
2452         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2453                                         set->queue_depth, set->reserved_tags);
2454         if (!set->tags[hctx_idx])
2455                 return false;
2456
2457         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2458                                 set->queue_depth);
2459         if (!ret)
2460                 return true;
2461
2462         blk_mq_free_rq_map(set->tags[hctx_idx]);
2463         set->tags[hctx_idx] = NULL;
2464         return false;
2465 }
2466
2467 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2468                                          unsigned int hctx_idx)
2469 {
2470         if (set->tags && set->tags[hctx_idx]) {
2471                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2472                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2473                 set->tags[hctx_idx] = NULL;
2474         }
2475 }
2476
2477 static void blk_mq_map_swqueue(struct request_queue *q)
2478 {
2479         unsigned int i, j, hctx_idx;
2480         struct blk_mq_hw_ctx *hctx;
2481         struct blk_mq_ctx *ctx;
2482         struct blk_mq_tag_set *set = q->tag_set;
2483
2484         queue_for_each_hw_ctx(q, hctx, i) {
2485                 cpumask_clear(hctx->cpumask);
2486                 hctx->nr_ctx = 0;
2487                 hctx->dispatch_from = NULL;
2488         }
2489
2490         /*
2491          * Map software to hardware queues.
2492          *
2493          * If the cpu isn't present, the cpu is mapped to first hctx.
2494          */
2495         for_each_possible_cpu(i) {
2496
2497                 ctx = per_cpu_ptr(q->queue_ctx, i);
2498                 for (j = 0; j < set->nr_maps; j++) {
2499                         if (!set->map[j].nr_queues) {
2500                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2501                                                 HCTX_TYPE_DEFAULT, i);
2502                                 continue;
2503                         }
2504                         hctx_idx = set->map[j].mq_map[i];
2505                         /* unmapped hw queue can be remapped after CPU topo changed */
2506                         if (!set->tags[hctx_idx] &&
2507                             !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2508                                 /*
2509                                  * If tags initialization fail for some hctx,
2510                                  * that hctx won't be brought online.  In this
2511                                  * case, remap the current ctx to hctx[0] which
2512                                  * is guaranteed to always have tags allocated
2513                                  */
2514                                 set->map[j].mq_map[i] = 0;
2515                         }
2516
2517                         hctx = blk_mq_map_queue_type(q, j, i);
2518                         ctx->hctxs[j] = hctx;
2519                         /*
2520                          * If the CPU is already set in the mask, then we've
2521                          * mapped this one already. This can happen if
2522                          * devices share queues across queue maps.
2523                          */
2524                         if (cpumask_test_cpu(i, hctx->cpumask))
2525                                 continue;
2526
2527                         cpumask_set_cpu(i, hctx->cpumask);
2528                         hctx->type = j;
2529                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2530                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2531
2532                         /*
2533                          * If the nr_ctx type overflows, we have exceeded the
2534                          * amount of sw queues we can support.
2535                          */
2536                         BUG_ON(!hctx->nr_ctx);
2537                 }
2538
2539                 for (; j < HCTX_MAX_TYPES; j++)
2540                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2541                                         HCTX_TYPE_DEFAULT, i);
2542         }
2543
2544         queue_for_each_hw_ctx(q, hctx, i) {
2545                 /*
2546                  * If no software queues are mapped to this hardware queue,
2547                  * disable it and free the request entries.
2548                  */
2549                 if (!hctx->nr_ctx) {
2550                         /* Never unmap queue 0.  We need it as a
2551                          * fallback in case of a new remap fails
2552                          * allocation
2553                          */
2554                         if (i && set->tags[i])
2555                                 blk_mq_free_map_and_requests(set, i);
2556
2557                         hctx->tags = NULL;
2558                         continue;
2559                 }
2560
2561                 hctx->tags = set->tags[i];
2562                 WARN_ON(!hctx->tags);
2563
2564                 /*
2565                  * Set the map size to the number of mapped software queues.
2566                  * This is more accurate and more efficient than looping
2567                  * over all possibly mapped software queues.
2568                  */
2569                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2570
2571                 /*
2572                  * Initialize batch roundrobin counts
2573                  */
2574                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2575                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2576         }
2577 }
2578
2579 /*
2580  * Caller needs to ensure that we're either frozen/quiesced, or that
2581  * the queue isn't live yet.
2582  */
2583 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2584 {
2585         struct blk_mq_hw_ctx *hctx;
2586         int i;
2587
2588         queue_for_each_hw_ctx(q, hctx, i) {
2589                 if (shared)
2590                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2591                 else
2592                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2593         }
2594 }
2595
2596 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2597                                         bool shared)
2598 {
2599         struct request_queue *q;
2600
2601         lockdep_assert_held(&set->tag_list_lock);
2602
2603         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2604                 blk_mq_freeze_queue(q);
2605                 queue_set_hctx_shared(q, shared);
2606                 blk_mq_unfreeze_queue(q);
2607         }
2608 }
2609
2610 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2611 {
2612         struct blk_mq_tag_set *set = q->tag_set;
2613
2614         mutex_lock(&set->tag_list_lock);
2615         list_del_rcu(&q->tag_set_list);
2616         if (list_is_singular(&set->tag_list)) {
2617                 /* just transitioned to unshared */
2618                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2619                 /* update existing queue */
2620                 blk_mq_update_tag_set_depth(set, false);
2621         }
2622         mutex_unlock(&set->tag_list_lock);
2623         INIT_LIST_HEAD(&q->tag_set_list);
2624 }
2625
2626 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2627                                      struct request_queue *q)
2628 {
2629         mutex_lock(&set->tag_list_lock);
2630
2631         /*
2632          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2633          */
2634         if (!list_empty(&set->tag_list) &&
2635             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2636                 set->flags |= BLK_MQ_F_TAG_SHARED;
2637                 /* update existing queue */
2638                 blk_mq_update_tag_set_depth(set, true);
2639         }
2640         if (set->flags & BLK_MQ_F_TAG_SHARED)
2641                 queue_set_hctx_shared(q, true);
2642         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2643
2644         mutex_unlock(&set->tag_list_lock);
2645 }
2646
2647 /* All allocations will be freed in release handler of q->mq_kobj */
2648 static int blk_mq_alloc_ctxs(struct request_queue *q)
2649 {
2650         struct blk_mq_ctxs *ctxs;
2651         int cpu;
2652
2653         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2654         if (!ctxs)
2655                 return -ENOMEM;
2656
2657         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2658         if (!ctxs->queue_ctx)
2659                 goto fail;
2660
2661         for_each_possible_cpu(cpu) {
2662                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2663                 ctx->ctxs = ctxs;
2664         }
2665
2666         q->mq_kobj = &ctxs->kobj;
2667         q->queue_ctx = ctxs->queue_ctx;
2668
2669         return 0;
2670  fail:
2671         kfree(ctxs);
2672         return -ENOMEM;
2673 }
2674
2675 /*
2676  * It is the actual release handler for mq, but we do it from
2677  * request queue's release handler for avoiding use-after-free
2678  * and headache because q->mq_kobj shouldn't have been introduced,
2679  * but we can't group ctx/kctx kobj without it.
2680  */
2681 void blk_mq_release(struct request_queue *q)
2682 {
2683         struct blk_mq_hw_ctx *hctx, *next;
2684         int i;
2685
2686         queue_for_each_hw_ctx(q, hctx, i)
2687                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2688
2689         /* all hctx are in .unused_hctx_list now */
2690         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2691                 list_del_init(&hctx->hctx_list);
2692                 kobject_put(&hctx->kobj);
2693         }
2694
2695         kfree(q->queue_hw_ctx);
2696
2697         /*
2698          * release .mq_kobj and sw queue's kobject now because
2699          * both share lifetime with request queue.
2700          */
2701         blk_mq_sysfs_deinit(q);
2702 }
2703
2704 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2705 {
2706         struct request_queue *uninit_q, *q;
2707
2708         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2709         if (!uninit_q)
2710                 return ERR_PTR(-ENOMEM);
2711
2712         /*
2713          * Initialize the queue without an elevator. device_add_disk() will do
2714          * the initialization.
2715          */
2716         q = blk_mq_init_allocated_queue(set, uninit_q, false);
2717         if (IS_ERR(q))
2718                 blk_cleanup_queue(uninit_q);
2719
2720         return q;
2721 }
2722 EXPORT_SYMBOL(blk_mq_init_queue);
2723
2724 /*
2725  * Helper for setting up a queue with mq ops, given queue depth, and
2726  * the passed in mq ops flags.
2727  */
2728 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2729                                            const struct blk_mq_ops *ops,
2730                                            unsigned int queue_depth,
2731                                            unsigned int set_flags)
2732 {
2733         struct request_queue *q;
2734         int ret;
2735
2736         memset(set, 0, sizeof(*set));
2737         set->ops = ops;
2738         set->nr_hw_queues = 1;
2739         set->nr_maps = 1;
2740         set->queue_depth = queue_depth;
2741         set->numa_node = NUMA_NO_NODE;
2742         set->flags = set_flags;
2743
2744         ret = blk_mq_alloc_tag_set(set);
2745         if (ret)
2746                 return ERR_PTR(ret);
2747
2748         q = blk_mq_init_queue(set);
2749         if (IS_ERR(q)) {
2750                 blk_mq_free_tag_set(set);
2751                 return q;
2752         }
2753
2754         return q;
2755 }
2756 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2757
2758 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2759                 struct blk_mq_tag_set *set, struct request_queue *q,
2760                 int hctx_idx, int node)
2761 {
2762         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2763
2764         /* reuse dead hctx first */
2765         spin_lock(&q->unused_hctx_lock);
2766         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2767                 if (tmp->numa_node == node) {
2768                         hctx = tmp;
2769                         break;
2770                 }
2771         }
2772         if (hctx)
2773                 list_del_init(&hctx->hctx_list);
2774         spin_unlock(&q->unused_hctx_lock);
2775
2776         if (!hctx)
2777                 hctx = blk_mq_alloc_hctx(q, set, node);
2778         if (!hctx)
2779                 goto fail;
2780
2781         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2782                 goto free_hctx;
2783
2784         return hctx;
2785
2786  free_hctx:
2787         kobject_put(&hctx->kobj);
2788  fail:
2789         return NULL;
2790 }
2791
2792 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2793                                                 struct request_queue *q)
2794 {
2795         int i, j, end;
2796         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2797
2798         /* protect against switching io scheduler  */
2799         mutex_lock(&q->sysfs_lock);
2800         for (i = 0; i < set->nr_hw_queues; i++) {
2801                 int node;
2802                 struct blk_mq_hw_ctx *hctx;
2803
2804                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2805                 /*
2806                  * If the hw queue has been mapped to another numa node,
2807                  * we need to realloc the hctx. If allocation fails, fallback
2808                  * to use the previous one.
2809                  */
2810                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2811                         continue;
2812
2813                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2814                 if (hctx) {
2815                         if (hctxs[i])
2816                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2817                         hctxs[i] = hctx;
2818                 } else {
2819                         if (hctxs[i])
2820                                 pr_warn("Allocate new hctx on node %d fails,\
2821                                                 fallback to previous one on node %d\n",
2822                                                 node, hctxs[i]->numa_node);
2823                         else
2824                                 break;
2825                 }
2826         }
2827         /*
2828          * Increasing nr_hw_queues fails. Free the newly allocated
2829          * hctxs and keep the previous q->nr_hw_queues.
2830          */
2831         if (i != set->nr_hw_queues) {
2832                 j = q->nr_hw_queues;
2833                 end = i;
2834         } else {
2835                 j = i;
2836                 end = q->nr_hw_queues;
2837                 q->nr_hw_queues = set->nr_hw_queues;
2838         }
2839
2840         for (; j < end; j++) {
2841                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2842
2843                 if (hctx) {
2844                         if (hctx->tags)
2845                                 blk_mq_free_map_and_requests(set, j);
2846                         blk_mq_exit_hctx(q, set, hctx, j);
2847                         hctxs[j] = NULL;
2848                 }
2849         }
2850         mutex_unlock(&q->sysfs_lock);
2851 }
2852
2853 /*
2854  * Maximum number of hardware queues we support. For single sets, we'll never
2855  * have more than the CPUs (software queues). For multiple sets, the tag_set
2856  * user may have set ->nr_hw_queues larger.
2857  */
2858 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2859 {
2860         if (set->nr_maps == 1)
2861                 return nr_cpu_ids;
2862
2863         return max(set->nr_hw_queues, nr_cpu_ids);
2864 }
2865
2866 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2867                                                   struct request_queue *q,
2868                                                   bool elevator_init)
2869 {
2870         /* mark the queue as mq asap */
2871         q->mq_ops = set->ops;
2872
2873         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2874                                              blk_mq_poll_stats_bkt,
2875                                              BLK_MQ_POLL_STATS_BKTS, q);
2876         if (!q->poll_cb)
2877                 goto err_exit;
2878
2879         if (blk_mq_alloc_ctxs(q))
2880                 goto err_poll;
2881
2882         /* init q->mq_kobj and sw queues' kobjects */
2883         blk_mq_sysfs_init(q);
2884
2885         q->nr_queues = nr_hw_queues(set);
2886         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2887                                                 GFP_KERNEL, set->numa_node);
2888         if (!q->queue_hw_ctx)
2889                 goto err_sys_init;
2890
2891         INIT_LIST_HEAD(&q->unused_hctx_list);
2892         spin_lock_init(&q->unused_hctx_lock);
2893
2894         blk_mq_realloc_hw_ctxs(set, q);
2895         if (!q->nr_hw_queues)
2896                 goto err_hctxs;
2897
2898         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2899         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2900
2901         q->tag_set = set;
2902
2903         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2904         if (set->nr_maps > HCTX_TYPE_POLL &&
2905             set->map[HCTX_TYPE_POLL].nr_queues)
2906                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2907
2908         q->sg_reserved_size = INT_MAX;
2909
2910         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2911         INIT_LIST_HEAD(&q->requeue_list);
2912         spin_lock_init(&q->requeue_lock);
2913
2914         blk_queue_make_request(q, blk_mq_make_request);
2915
2916         /*
2917          * Do this after blk_queue_make_request() overrides it...
2918          */
2919         q->nr_requests = set->queue_depth;
2920
2921         /*
2922          * Default to classic polling
2923          */
2924         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2925
2926         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2927         blk_mq_add_queue_tag_set(set, q);
2928         blk_mq_map_swqueue(q);
2929
2930         if (elevator_init)
2931                 elevator_init_mq(q);
2932
2933         return q;
2934
2935 err_hctxs:
2936         kfree(q->queue_hw_ctx);
2937         q->nr_hw_queues = 0;
2938 err_sys_init:
2939         blk_mq_sysfs_deinit(q);
2940 err_poll:
2941         blk_stat_free_callback(q->poll_cb);
2942         q->poll_cb = NULL;
2943 err_exit:
2944         q->mq_ops = NULL;
2945         return ERR_PTR(-ENOMEM);
2946 }
2947 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2948
2949 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2950 void blk_mq_exit_queue(struct request_queue *q)
2951 {
2952         struct blk_mq_tag_set   *set = q->tag_set;
2953
2954         blk_mq_del_queue_tag_set(q);
2955         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2956 }
2957
2958 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2959 {
2960         int i;
2961
2962         for (i = 0; i < set->nr_hw_queues; i++)
2963                 if (!__blk_mq_alloc_rq_map(set, i))
2964                         goto out_unwind;
2965
2966         return 0;
2967
2968 out_unwind:
2969         while (--i >= 0)
2970                 blk_mq_free_rq_map(set->tags[i]);
2971
2972         return -ENOMEM;
2973 }
2974
2975 /*
2976  * Allocate the request maps associated with this tag_set. Note that this
2977  * may reduce the depth asked for, if memory is tight. set->queue_depth
2978  * will be updated to reflect the allocated depth.
2979  */
2980 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2981 {
2982         unsigned int depth;
2983         int err;
2984
2985         depth = set->queue_depth;
2986         do {
2987                 err = __blk_mq_alloc_rq_maps(set);
2988                 if (!err)
2989                         break;
2990
2991                 set->queue_depth >>= 1;
2992                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2993                         err = -ENOMEM;
2994                         break;
2995                 }
2996         } while (set->queue_depth);
2997
2998         if (!set->queue_depth || err) {
2999                 pr_err("blk-mq: failed to allocate request map\n");
3000                 return -ENOMEM;
3001         }
3002
3003         if (depth != set->queue_depth)
3004                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3005                                                 depth, set->queue_depth);
3006
3007         return 0;
3008 }
3009
3010 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3011 {
3012         /*
3013          * blk_mq_map_queues() and multiple .map_queues() implementations
3014          * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3015          * number of hardware queues.
3016          */
3017         if (set->nr_maps == 1)
3018                 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3019
3020         if (set->ops->map_queues && !is_kdump_kernel()) {
3021                 int i;
3022
3023                 /*
3024                  * transport .map_queues is usually done in the following
3025                  * way:
3026                  *
3027                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3028                  *      mask = get_cpu_mask(queue)
3029                  *      for_each_cpu(cpu, mask)
3030                  *              set->map[x].mq_map[cpu] = queue;
3031                  * }
3032                  *
3033                  * When we need to remap, the table has to be cleared for
3034                  * killing stale mapping since one CPU may not be mapped
3035                  * to any hw queue.
3036                  */
3037                 for (i = 0; i < set->nr_maps; i++)
3038                         blk_mq_clear_mq_map(&set->map[i]);
3039
3040                 return set->ops->map_queues(set);
3041         } else {
3042                 BUG_ON(set->nr_maps > 1);
3043                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3044         }
3045 }
3046
3047 /*
3048  * Alloc a tag set to be associated with one or more request queues.
3049  * May fail with EINVAL for various error conditions. May adjust the
3050  * requested depth down, if it's too large. In that case, the set
3051  * value will be stored in set->queue_depth.
3052  */
3053 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3054 {
3055         int i, ret;
3056
3057         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3058
3059         if (!set->nr_hw_queues)
3060                 return -EINVAL;
3061         if (!set->queue_depth)
3062                 return -EINVAL;
3063         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3064                 return -EINVAL;
3065
3066         if (!set->ops->queue_rq)
3067                 return -EINVAL;
3068
3069         if (!set->ops->get_budget ^ !set->ops->put_budget)
3070                 return -EINVAL;
3071
3072         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3073                 pr_info("blk-mq: reduced tag depth to %u\n",
3074                         BLK_MQ_MAX_DEPTH);
3075                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3076         }
3077
3078         if (!set->nr_maps)
3079                 set->nr_maps = 1;
3080         else if (set->nr_maps > HCTX_MAX_TYPES)
3081                 return -EINVAL;
3082
3083         /*
3084          * If a crashdump is active, then we are potentially in a very
3085          * memory constrained environment. Limit us to 1 queue and
3086          * 64 tags to prevent using too much memory.
3087          */
3088         if (is_kdump_kernel()) {
3089                 set->nr_hw_queues = 1;
3090                 set->nr_maps = 1;
3091                 set->queue_depth = min(64U, set->queue_depth);
3092         }
3093         /*
3094          * There is no use for more h/w queues than cpus if we just have
3095          * a single map
3096          */
3097         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3098                 set->nr_hw_queues = nr_cpu_ids;
3099
3100         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3101                                  GFP_KERNEL, set->numa_node);
3102         if (!set->tags)
3103                 return -ENOMEM;
3104
3105         ret = -ENOMEM;
3106         for (i = 0; i < set->nr_maps; i++) {
3107                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3108                                                   sizeof(set->map[i].mq_map[0]),
3109                                                   GFP_KERNEL, set->numa_node);
3110                 if (!set->map[i].mq_map)
3111                         goto out_free_mq_map;
3112                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3113         }
3114
3115         ret = blk_mq_update_queue_map(set);
3116         if (ret)
3117                 goto out_free_mq_map;
3118
3119         ret = blk_mq_alloc_rq_maps(set);
3120         if (ret)
3121                 goto out_free_mq_map;
3122
3123         mutex_init(&set->tag_list_lock);
3124         INIT_LIST_HEAD(&set->tag_list);
3125
3126         return 0;
3127
3128 out_free_mq_map:
3129         for (i = 0; i < set->nr_maps; i++) {
3130                 kfree(set->map[i].mq_map);
3131                 set->map[i].mq_map = NULL;
3132         }
3133         kfree(set->tags);
3134         set->tags = NULL;
3135         return ret;
3136 }
3137 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3138
3139 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3140 {
3141         int i, j;
3142
3143         for (i = 0; i < nr_hw_queues(set); i++)
3144                 blk_mq_free_map_and_requests(set, i);
3145
3146         for (j = 0; j < set->nr_maps; j++) {
3147                 kfree(set->map[j].mq_map);
3148                 set->map[j].mq_map = NULL;
3149         }
3150
3151         kfree(set->tags);
3152         set->tags = NULL;
3153 }
3154 EXPORT_SYMBOL(blk_mq_free_tag_set);
3155
3156 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3157 {
3158         struct blk_mq_tag_set *set = q->tag_set;
3159         struct blk_mq_hw_ctx *hctx;
3160         int i, ret;
3161
3162         if (!set)
3163                 return -EINVAL;
3164
3165         if (q->nr_requests == nr)
3166                 return 0;
3167
3168         blk_mq_freeze_queue(q);
3169         blk_mq_quiesce_queue(q);
3170
3171         ret = 0;
3172         queue_for_each_hw_ctx(q, hctx, i) {
3173                 if (!hctx->tags)
3174                         continue;
3175                 /*
3176                  * If we're using an MQ scheduler, just update the scheduler
3177                  * queue depth. This is similar to what the old code would do.
3178                  */
3179                 if (!hctx->sched_tags) {
3180                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3181                                                         false);
3182                 } else {
3183                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3184                                                         nr, true);
3185                 }
3186                 if (ret)
3187                         break;
3188                 if (q->elevator && q->elevator->type->ops.depth_updated)
3189                         q->elevator->type->ops.depth_updated(hctx);
3190         }
3191
3192         if (!ret)
3193                 q->nr_requests = nr;
3194
3195         blk_mq_unquiesce_queue(q);
3196         blk_mq_unfreeze_queue(q);
3197
3198         return ret;
3199 }
3200
3201 /*
3202  * request_queue and elevator_type pair.
3203  * It is just used by __blk_mq_update_nr_hw_queues to cache
3204  * the elevator_type associated with a request_queue.
3205  */
3206 struct blk_mq_qe_pair {
3207         struct list_head node;
3208         struct request_queue *q;
3209         struct elevator_type *type;
3210 };
3211
3212 /*
3213  * Cache the elevator_type in qe pair list and switch the
3214  * io scheduler to 'none'
3215  */
3216 static bool blk_mq_elv_switch_none(struct list_head *head,
3217                 struct request_queue *q)
3218 {
3219         struct blk_mq_qe_pair *qe;
3220
3221         if (!q->elevator)
3222                 return true;
3223
3224         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3225         if (!qe)
3226                 return false;
3227
3228         INIT_LIST_HEAD(&qe->node);
3229         qe->q = q;
3230         qe->type = q->elevator->type;
3231         list_add(&qe->node, head);
3232
3233         mutex_lock(&q->sysfs_lock);
3234         /*
3235          * After elevator_switch_mq, the previous elevator_queue will be
3236          * released by elevator_release. The reference of the io scheduler
3237          * module get by elevator_get will also be put. So we need to get
3238          * a reference of the io scheduler module here to prevent it to be
3239          * removed.
3240          */
3241         __module_get(qe->type->elevator_owner);
3242         elevator_switch_mq(q, NULL);
3243         mutex_unlock(&q->sysfs_lock);
3244
3245         return true;
3246 }
3247
3248 static void blk_mq_elv_switch_back(struct list_head *head,
3249                 struct request_queue *q)
3250 {
3251         struct blk_mq_qe_pair *qe;
3252         struct elevator_type *t = NULL;
3253
3254         list_for_each_entry(qe, head, node)
3255                 if (qe->q == q) {
3256                         t = qe->type;
3257                         break;
3258                 }
3259
3260         if (!t)
3261                 return;
3262
3263         list_del(&qe->node);
3264         kfree(qe);
3265
3266         mutex_lock(&q->sysfs_lock);
3267         elevator_switch_mq(q, t);
3268         mutex_unlock(&q->sysfs_lock);
3269 }
3270
3271 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3272                                                         int nr_hw_queues)
3273 {
3274         struct request_queue *q;
3275         LIST_HEAD(head);
3276         int prev_nr_hw_queues;
3277
3278         lockdep_assert_held(&set->tag_list_lock);
3279
3280         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3281                 nr_hw_queues = nr_cpu_ids;
3282         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3283                 return;
3284
3285         list_for_each_entry(q, &set->tag_list, tag_set_list)
3286                 blk_mq_freeze_queue(q);
3287         /*
3288          * Sync with blk_mq_queue_tag_busy_iter.
3289          */
3290         synchronize_rcu();
3291         /*
3292          * Switch IO scheduler to 'none', cleaning up the data associated
3293          * with the previous scheduler. We will switch back once we are done
3294          * updating the new sw to hw queue mappings.
3295          */
3296         list_for_each_entry(q, &set->tag_list, tag_set_list)
3297                 if (!blk_mq_elv_switch_none(&head, q))
3298                         goto switch_back;
3299
3300         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3301                 blk_mq_debugfs_unregister_hctxs(q);
3302                 blk_mq_sysfs_unregister(q);
3303         }
3304
3305         prev_nr_hw_queues = set->nr_hw_queues;
3306         set->nr_hw_queues = nr_hw_queues;
3307 fallback:
3308         blk_mq_update_queue_map(set);
3309         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3310                 blk_mq_realloc_hw_ctxs(set, q);
3311                 if (q->nr_hw_queues != set->nr_hw_queues) {
3312                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3313                                         nr_hw_queues, prev_nr_hw_queues);
3314                         set->nr_hw_queues = prev_nr_hw_queues;
3315                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3316                         goto fallback;
3317                 }
3318                 blk_mq_map_swqueue(q);
3319         }
3320
3321         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3322                 blk_mq_sysfs_register(q);
3323                 blk_mq_debugfs_register_hctxs(q);
3324         }
3325
3326 switch_back:
3327         list_for_each_entry(q, &set->tag_list, tag_set_list)
3328                 blk_mq_elv_switch_back(&head, q);
3329
3330         list_for_each_entry(q, &set->tag_list, tag_set_list)
3331                 blk_mq_unfreeze_queue(q);
3332 }
3333
3334 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3335 {
3336         mutex_lock(&set->tag_list_lock);
3337         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3338         mutex_unlock(&set->tag_list_lock);
3339 }
3340 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3341
3342 /* Enable polling stats and return whether they were already enabled. */
3343 static bool blk_poll_stats_enable(struct request_queue *q)
3344 {
3345         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3346             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3347                 return true;
3348         blk_stat_add_callback(q, q->poll_cb);
3349         return false;
3350 }
3351
3352 static void blk_mq_poll_stats_start(struct request_queue *q)
3353 {
3354         /*
3355          * We don't arm the callback if polling stats are not enabled or the
3356          * callback is already active.
3357          */
3358         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3359             blk_stat_is_active(q->poll_cb))
3360                 return;
3361
3362         blk_stat_activate_msecs(q->poll_cb, 100);
3363 }
3364
3365 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3366 {
3367         struct request_queue *q = cb->data;
3368         int bucket;
3369
3370         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3371                 if (cb->stat[bucket].nr_samples)
3372                         q->poll_stat[bucket] = cb->stat[bucket];
3373         }
3374 }
3375
3376 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3377                                        struct blk_mq_hw_ctx *hctx,
3378                                        struct request *rq)
3379 {
3380         unsigned long ret = 0;
3381         int bucket;
3382
3383         /*
3384          * If stats collection isn't on, don't sleep but turn it on for
3385          * future users
3386          */
3387         if (!blk_poll_stats_enable(q))
3388                 return 0;
3389
3390         /*
3391          * As an optimistic guess, use half of the mean service time
3392          * for this type of request. We can (and should) make this smarter.
3393          * For instance, if the completion latencies are tight, we can
3394          * get closer than just half the mean. This is especially
3395          * important on devices where the completion latencies are longer
3396          * than ~10 usec. We do use the stats for the relevant IO size
3397          * if available which does lead to better estimates.
3398          */
3399         bucket = blk_mq_poll_stats_bkt(rq);
3400         if (bucket < 0)
3401                 return ret;
3402
3403         if (q->poll_stat[bucket].nr_samples)
3404                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3405
3406         return ret;
3407 }
3408
3409 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3410                                      struct blk_mq_hw_ctx *hctx,
3411                                      struct request *rq)
3412 {
3413         struct hrtimer_sleeper hs;
3414         enum hrtimer_mode mode;
3415         unsigned int nsecs;
3416         ktime_t kt;
3417
3418         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3419                 return false;
3420
3421         /*
3422          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3423          *
3424          *  0:  use half of prev avg
3425          * >0:  use this specific value
3426          */
3427         if (q->poll_nsec > 0)
3428                 nsecs = q->poll_nsec;
3429         else
3430                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3431
3432         if (!nsecs)
3433                 return false;
3434
3435         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3436
3437         /*
3438          * This will be replaced with the stats tracking code, using
3439          * 'avg_completion_time / 2' as the pre-sleep target.
3440          */
3441         kt = nsecs;
3442
3443         mode = HRTIMER_MODE_REL;
3444         hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3445         hrtimer_set_expires(&hs.timer, kt);
3446
3447         do {
3448                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3449                         break;
3450                 set_current_state(TASK_UNINTERRUPTIBLE);
3451                 hrtimer_sleeper_start_expires(&hs, mode);
3452                 if (hs.task)
3453                         io_schedule();
3454                 hrtimer_cancel(&hs.timer);
3455                 mode = HRTIMER_MODE_ABS;
3456         } while (hs.task && !signal_pending(current));
3457
3458         __set_current_state(TASK_RUNNING);
3459         destroy_hrtimer_on_stack(&hs.timer);
3460         return true;
3461 }
3462
3463 static bool blk_mq_poll_hybrid(struct request_queue *q,
3464                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3465 {
3466         struct request *rq;
3467
3468         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3469                 return false;
3470
3471         if (!blk_qc_t_is_internal(cookie))
3472                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3473         else {
3474                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3475                 /*
3476                  * With scheduling, if the request has completed, we'll
3477                  * get a NULL return here, as we clear the sched tag when
3478                  * that happens. The request still remains valid, like always,
3479                  * so we should be safe with just the NULL check.
3480                  */
3481                 if (!rq)
3482                         return false;
3483         }
3484
3485         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3486 }
3487
3488 /**
3489  * blk_poll - poll for IO completions
3490  * @q:  the queue
3491  * @cookie: cookie passed back at IO submission time
3492  * @spin: whether to spin for completions
3493  *
3494  * Description:
3495  *    Poll for completions on the passed in queue. Returns number of
3496  *    completed entries found. If @spin is true, then blk_poll will continue
3497  *    looping until at least one completion is found, unless the task is
3498  *    otherwise marked running (or we need to reschedule).
3499  */
3500 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3501 {
3502         struct blk_mq_hw_ctx *hctx;
3503         long state;
3504
3505         if (!blk_qc_t_valid(cookie) ||
3506             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3507                 return 0;
3508
3509         if (current->plug)
3510                 blk_flush_plug_list(current->plug, false);
3511
3512         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3513
3514         /*
3515          * If we sleep, have the caller restart the poll loop to reset
3516          * the state. Like for the other success return cases, the
3517          * caller is responsible for checking if the IO completed. If
3518          * the IO isn't complete, we'll get called again and will go
3519          * straight to the busy poll loop.
3520          */
3521         if (blk_mq_poll_hybrid(q, hctx, cookie))
3522                 return 1;
3523
3524         hctx->poll_considered++;
3525
3526         state = current->state;
3527         do {
3528                 int ret;
3529
3530                 hctx->poll_invoked++;
3531
3532                 ret = q->mq_ops->poll(hctx);
3533                 if (ret > 0) {
3534                         hctx->poll_success++;
3535                         __set_current_state(TASK_RUNNING);
3536                         return ret;
3537                 }
3538
3539                 if (signal_pending_state(state, current))
3540                         __set_current_state(TASK_RUNNING);
3541
3542                 if (current->state == TASK_RUNNING)
3543                         return 1;
3544                 if (ret < 0 || !spin)
3545                         break;
3546                 cpu_relax();
3547         } while (!need_resched());
3548
3549         __set_current_state(TASK_RUNNING);
3550         return 0;
3551 }
3552 EXPORT_SYMBOL_GPL(blk_poll);
3553
3554 unsigned int blk_mq_rq_cpu(struct request *rq)
3555 {
3556         return rq->mq_ctx->cpu;
3557 }
3558 EXPORT_SYMBOL(blk_mq_rq_cpu);
3559
3560 static int __init blk_mq_init(void)
3561 {
3562         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3563                                 blk_mq_hctx_notify_dead);
3564         return 0;
3565 }
3566 subsys_initcall(blk_mq_init);