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);
 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                         break;
1237
1238                 if (!blk_mq_get_driver_tag(rq)) {
1239                         /*
1240                          * The initial allocation attempt failed, so we need to
1241                          * rerun the hardware queue when a tag is freed. The
1242                          * waitqueue takes care of that. If the queue is run
1243                          * before we add this entry back on the dispatch list,
1244                          * we'll re-run it below.
1245                          */
1246                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1247                                 blk_mq_put_dispatch_budget(hctx);
1248                                 /*
1249                                  * For non-shared tags, the RESTART check
1250                                  * will suffice.
1251                                  */
1252                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1253                                         no_tag = true;
1254                                 break;
1255                         }
1256                 }
1257
1258                 list_del_init(&rq->queuelist);
1259
1260                 bd.rq = rq;
1261
1262                 /*
1263                  * Flag last if we have no more requests, or if we have more
1264                  * but can't assign a driver tag to it.
1265                  */
1266                 if (list_empty(list))
1267                         bd.last = true;
1268                 else {
1269                         nxt = list_first_entry(list, struct request, queuelist);
1270                         bd.last = !blk_mq_get_driver_tag(nxt);
1271                 }
1272
1273                 ret = q->mq_ops->queue_rq(hctx, &bd);
1274                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1275                         /*
1276                          * If an I/O scheduler has been configured and we got a
1277                          * driver tag for the next request already, free it
1278                          * again.
1279                          */
1280                         if (!list_empty(list)) {
1281                                 nxt = list_first_entry(list, struct request, queuelist);
1282                                 blk_mq_put_driver_tag(nxt);
1283                         }
1284                         list_add(&rq->queuelist, list);
1285                         __blk_mq_requeue_request(rq);
1286                         break;
1287                 }
1288
1289                 if (unlikely(ret != BLK_STS_OK)) {
1290                         errors++;
1291                         blk_mq_end_request(rq, BLK_STS_IOERR);
1292                         continue;
1293                 }
1294
1295                 queued++;
1296         } while (!list_empty(list));
1297
1298         hctx->dispatched[queued_to_index(queued)]++;
1299
1300         /*
1301          * Any items that need requeuing? Stuff them into hctx->dispatch,
1302          * that is where we will continue on next queue run.
1303          */
1304         if (!list_empty(list)) {
1305                 bool needs_restart;
1306
1307                 /*
1308                  * If we didn't flush the entire list, we could have told
1309                  * the driver there was more coming, but that turned out to
1310                  * be a lie.
1311                  */
1312                 if (q->mq_ops->commit_rqs)
1313                         q->mq_ops->commit_rqs(hctx);
1314
1315                 spin_lock(&hctx->lock);
1316                 list_splice_init(list, &hctx->dispatch);
1317                 spin_unlock(&hctx->lock);
1318
1319                 /*
1320                  * If SCHED_RESTART was set by the caller of this function and
1321                  * it is no longer set that means that it was cleared by another
1322                  * thread and hence that a queue rerun is needed.
1323                  *
1324                  * If 'no_tag' is set, that means that we failed getting
1325                  * a driver tag with an I/O scheduler attached. If our dispatch
1326                  * waitqueue is no longer active, ensure that we run the queue
1327                  * AFTER adding our entries back to the list.
1328                  *
1329                  * If no I/O scheduler has been configured it is possible that
1330                  * the hardware queue got stopped and restarted before requests
1331                  * were pushed back onto the dispatch list. Rerun the queue to
1332                  * avoid starvation. Notes:
1333                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1334                  *   been stopped before rerunning a queue.
1335                  * - Some but not all block drivers stop a queue before
1336                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1337                  *   and dm-rq.
1338                  *
1339                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1340                  * bit is set, run queue after a delay to avoid IO stalls
1341                  * that could otherwise occur if the queue is idle.
1342                  */
1343                 needs_restart = blk_mq_sched_needs_restart(hctx);
1344                 if (!needs_restart ||
1345                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1346                         blk_mq_run_hw_queue(hctx, true);
1347                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1348                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1349
1350                 blk_mq_update_dispatch_busy(hctx, true);
1351                 return false;
1352         } else
1353                 blk_mq_update_dispatch_busy(hctx, false);
1354
1355         /*
1356          * If the host/device is unable to accept more work, inform the
1357          * caller of that.
1358          */
1359         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1360                 return false;
1361
1362         return (queued + errors) != 0;
1363 }
1364
1365 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1366 {
1367         int srcu_idx;
1368
1369         /*
1370          * We should be running this queue from one of the CPUs that
1371          * are mapped to it.
1372          *
1373          * There are at least two related races now between setting
1374          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1375          * __blk_mq_run_hw_queue():
1376          *
1377          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1378          *   but later it becomes online, then this warning is harmless
1379          *   at all
1380          *
1381          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1382          *   but later it becomes offline, then the warning can't be
1383          *   triggered, and we depend on blk-mq timeout handler to
1384          *   handle dispatched requests to this hctx
1385          */
1386         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1387                 cpu_online(hctx->next_cpu)) {
1388                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1389                         raw_smp_processor_id(),
1390                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1391                 dump_stack();
1392         }
1393
1394         /*
1395          * We can't run the queue inline with ints disabled. Ensure that
1396          * we catch bad users of this early.
1397          */
1398         WARN_ON_ONCE(in_interrupt());
1399
1400         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1401
1402         hctx_lock(hctx, &srcu_idx);
1403         blk_mq_sched_dispatch_requests(hctx);
1404         hctx_unlock(hctx, srcu_idx);
1405 }
1406
1407 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1408 {
1409         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1410
1411         if (cpu >= nr_cpu_ids)
1412                 cpu = cpumask_first(hctx->cpumask);
1413         return cpu;
1414 }
1415
1416 /*
1417  * It'd be great if the workqueue API had a way to pass
1418  * in a mask and had some smarts for more clever placement.
1419  * For now we just round-robin here, switching for every
1420  * BLK_MQ_CPU_WORK_BATCH queued items.
1421  */
1422 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1423 {
1424         bool tried = false;
1425         int next_cpu = hctx->next_cpu;
1426
1427         if (hctx->queue->nr_hw_queues == 1)
1428                 return WORK_CPU_UNBOUND;
1429
1430         if (--hctx->next_cpu_batch <= 0) {
1431 select_cpu:
1432                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1433                                 cpu_online_mask);
1434                 if (next_cpu >= nr_cpu_ids)
1435                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1436                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1437         }
1438
1439         /*
1440          * Do unbound schedule if we can't find a online CPU for this hctx,
1441          * and it should only happen in the path of handling CPU DEAD.
1442          */
1443         if (!cpu_online(next_cpu)) {
1444                 if (!tried) {
1445                         tried = true;
1446                         goto select_cpu;
1447                 }
1448
1449                 /*
1450                  * Make sure to re-select CPU next time once after CPUs
1451                  * in hctx->cpumask become online again.
1452                  */
1453                 hctx->next_cpu = next_cpu;
1454                 hctx->next_cpu_batch = 1;
1455                 return WORK_CPU_UNBOUND;
1456         }
1457
1458         hctx->next_cpu = next_cpu;
1459         return next_cpu;
1460 }
1461
1462 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1463                                         unsigned long msecs)
1464 {
1465         if (unlikely(blk_mq_hctx_stopped(hctx)))
1466                 return;
1467
1468         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1469                 int cpu = get_cpu();
1470                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1471                         __blk_mq_run_hw_queue(hctx);
1472                         put_cpu();
1473                         return;
1474                 }
1475
1476                 put_cpu();
1477         }
1478
1479         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1480                                     msecs_to_jiffies(msecs));
1481 }
1482
1483 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1484 {
1485         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1486 }
1487 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1488
1489 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1490 {
1491         int srcu_idx;
1492         bool need_run;
1493
1494         /*
1495          * When queue is quiesced, we may be switching io scheduler, or
1496          * updating nr_hw_queues, or other things, and we can't run queue
1497          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1498          *
1499          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1500          * quiesced.
1501          */
1502         hctx_lock(hctx, &srcu_idx);
1503         need_run = !blk_queue_quiesced(hctx->queue) &&
1504                 blk_mq_hctx_has_pending(hctx);
1505         hctx_unlock(hctx, srcu_idx);
1506
1507         if (need_run) {
1508                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1509                 return true;
1510         }
1511
1512         return false;
1513 }
1514 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1515
1516 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1517 {
1518         struct blk_mq_hw_ctx *hctx;
1519         int i;
1520
1521         queue_for_each_hw_ctx(q, hctx, i) {
1522                 if (blk_mq_hctx_stopped(hctx))
1523                         continue;
1524
1525                 blk_mq_run_hw_queue(hctx, async);
1526         }
1527 }
1528 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1529
1530 /**
1531  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1532  * @q: request queue.
1533  *
1534  * The caller is responsible for serializing this function against
1535  * blk_mq_{start,stop}_hw_queue().
1536  */
1537 bool blk_mq_queue_stopped(struct request_queue *q)
1538 {
1539         struct blk_mq_hw_ctx *hctx;
1540         int i;
1541
1542         queue_for_each_hw_ctx(q, hctx, i)
1543                 if (blk_mq_hctx_stopped(hctx))
1544                         return true;
1545
1546         return false;
1547 }
1548 EXPORT_SYMBOL(blk_mq_queue_stopped);
1549
1550 /*
1551  * This function is often used for pausing .queue_rq() by driver when
1552  * there isn't enough resource or some conditions aren't satisfied, and
1553  * BLK_STS_RESOURCE is usually returned.
1554  *
1555  * We do not guarantee that dispatch can be drained or blocked
1556  * after blk_mq_stop_hw_queue() returns. Please use
1557  * blk_mq_quiesce_queue() for that requirement.
1558  */
1559 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1560 {
1561         cancel_delayed_work(&hctx->run_work);
1562
1563         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1564 }
1565 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1566
1567 /*
1568  * This function is often used for pausing .queue_rq() by driver when
1569  * there isn't enough resource or some conditions aren't satisfied, and
1570  * BLK_STS_RESOURCE is usually returned.
1571  *
1572  * We do not guarantee that dispatch can be drained or blocked
1573  * after blk_mq_stop_hw_queues() returns. Please use
1574  * blk_mq_quiesce_queue() for that requirement.
1575  */
1576 void blk_mq_stop_hw_queues(struct request_queue *q)
1577 {
1578         struct blk_mq_hw_ctx *hctx;
1579         int i;
1580
1581         queue_for_each_hw_ctx(q, hctx, i)
1582                 blk_mq_stop_hw_queue(hctx);
1583 }
1584 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1585
1586 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1587 {
1588         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1589
1590         blk_mq_run_hw_queue(hctx, false);
1591 }
1592 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1593
1594 void blk_mq_start_hw_queues(struct request_queue *q)
1595 {
1596         struct blk_mq_hw_ctx *hctx;
1597         int i;
1598
1599         queue_for_each_hw_ctx(q, hctx, i)
1600                 blk_mq_start_hw_queue(hctx);
1601 }
1602 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1603
1604 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1605 {
1606         if (!blk_mq_hctx_stopped(hctx))
1607                 return;
1608
1609         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1610         blk_mq_run_hw_queue(hctx, async);
1611 }
1612 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1613
1614 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1615 {
1616         struct blk_mq_hw_ctx *hctx;
1617         int i;
1618
1619         queue_for_each_hw_ctx(q, hctx, i)
1620                 blk_mq_start_stopped_hw_queue(hctx, async);
1621 }
1622 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1623
1624 static void blk_mq_run_work_fn(struct work_struct *work)
1625 {
1626         struct blk_mq_hw_ctx *hctx;
1627
1628         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1629
1630         /*
1631          * If we are stopped, don't run the queue.
1632          */
1633         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1634                 return;
1635
1636         __blk_mq_run_hw_queue(hctx);
1637 }
1638
1639 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1640                                             struct request *rq,
1641                                             bool at_head)
1642 {
1643         struct blk_mq_ctx *ctx = rq->mq_ctx;
1644         enum hctx_type type = hctx->type;
1645
1646         lockdep_assert_held(&ctx->lock);
1647
1648         trace_block_rq_insert(hctx->queue, rq);
1649
1650         if (at_head)
1651                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1652         else
1653                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1654 }
1655
1656 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1657                              bool at_head)
1658 {
1659         struct blk_mq_ctx *ctx = rq->mq_ctx;
1660
1661         lockdep_assert_held(&ctx->lock);
1662
1663         __blk_mq_insert_req_list(hctx, rq, at_head);
1664         blk_mq_hctx_mark_pending(hctx, ctx);
1665 }
1666
1667 /*
1668  * Should only be used carefully, when the caller knows we want to
1669  * bypass a potential IO scheduler on the target device.
1670  */
1671 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1672 {
1673         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1674
1675         spin_lock(&hctx->lock);
1676         list_add_tail(&rq->queuelist, &hctx->dispatch);
1677         spin_unlock(&hctx->lock);
1678
1679         if (run_queue)
1680                 blk_mq_run_hw_queue(hctx, false);
1681 }
1682
1683 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1684                             struct list_head *list)
1685
1686 {
1687         struct request *rq;
1688         enum hctx_type type = hctx->type;
1689
1690         /*
1691          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1692          * offline now
1693          */
1694         list_for_each_entry(rq, list, queuelist) {
1695                 BUG_ON(rq->mq_ctx != ctx);
1696                 trace_block_rq_insert(hctx->queue, rq);
1697         }
1698
1699         spin_lock(&ctx->lock);
1700         list_splice_tail_init(list, &ctx->rq_lists[type]);
1701         blk_mq_hctx_mark_pending(hctx, ctx);
1702         spin_unlock(&ctx->lock);
1703 }
1704
1705 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1706 {
1707         struct request *rqa = container_of(a, struct request, queuelist);
1708         struct request *rqb = container_of(b, struct request, queuelist);
1709
1710         if (rqa->mq_ctx < rqb->mq_ctx)
1711                 return -1;
1712         else if (rqa->mq_ctx > rqb->mq_ctx)
1713                 return 1;
1714         else if (rqa->mq_hctx < rqb->mq_hctx)
1715                 return -1;
1716         else if (rqa->mq_hctx > rqb->mq_hctx)
1717                 return 1;
1718
1719         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1720 }
1721
1722 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1723 {
1724         struct blk_mq_hw_ctx *this_hctx;
1725         struct blk_mq_ctx *this_ctx;
1726         struct request_queue *this_q;
1727         struct request *rq;
1728         LIST_HEAD(list);
1729         LIST_HEAD(rq_list);
1730         unsigned int depth;
1731
1732         list_splice_init(&plug->mq_list, &list);
1733
1734         if (plug->rq_count > 2 && plug->multiple_queues)
1735                 list_sort(NULL, &list, plug_rq_cmp);
1736
1737         plug->rq_count = 0;
1738
1739         this_q = NULL;
1740         this_hctx = NULL;
1741         this_ctx = NULL;
1742         depth = 0;
1743
1744         while (!list_empty(&list)) {
1745                 rq = list_entry_rq(list.next);
1746                 list_del_init(&rq->queuelist);
1747                 BUG_ON(!rq->q);
1748                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1749                         if (this_hctx) {
1750                                 trace_block_unplug(this_q, depth, !from_schedule);
1751                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1752                                                                 &rq_list,
1753                                                                 from_schedule);
1754                         }
1755
1756                         this_q = rq->q;
1757                         this_ctx = rq->mq_ctx;
1758                         this_hctx = rq->mq_hctx;
1759                         depth = 0;
1760                 }
1761
1762                 depth++;
1763                 list_add_tail(&rq->queuelist, &rq_list);
1764         }
1765
1766         /*
1767          * If 'this_hctx' is set, we know we have entries to complete
1768          * on 'rq_list'. Do those.
1769          */
1770         if (this_hctx) {
1771                 trace_block_unplug(this_q, depth, !from_schedule);
1772                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1773                                                 from_schedule);
1774         }
1775 }
1776
1777 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1778                 unsigned int nr_segs)
1779 {
1780         if (bio->bi_opf & REQ_RAHEAD)
1781                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1782
1783         rq->__sector = bio->bi_iter.bi_sector;
1784         rq->write_hint = bio->bi_write_hint;
1785         blk_rq_bio_prep(rq, bio, nr_segs);
1786
1787         blk_account_io_start(rq, true);
1788 }
1789
1790 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1791                                             struct request *rq,
1792                                             blk_qc_t *cookie, bool last)
1793 {
1794         struct request_queue *q = rq->q;
1795         struct blk_mq_queue_data bd = {
1796                 .rq = rq,
1797                 .last = last,
1798         };
1799         blk_qc_t new_cookie;
1800         blk_status_t ret;
1801
1802         new_cookie = request_to_qc_t(hctx, rq);
1803
1804         /*
1805          * For OK queue, we are done. For error, caller may kill it.
1806          * Any other error (busy), just add it to our list as we
1807          * previously would have done.
1808          */
1809         ret = q->mq_ops->queue_rq(hctx, &bd);
1810         switch (ret) {
1811         case BLK_STS_OK:
1812                 blk_mq_update_dispatch_busy(hctx, false);
1813                 *cookie = new_cookie;
1814                 break;
1815         case BLK_STS_RESOURCE:
1816         case BLK_STS_DEV_RESOURCE:
1817                 blk_mq_update_dispatch_busy(hctx, true);
1818                 __blk_mq_requeue_request(rq);
1819                 break;
1820         default:
1821                 blk_mq_update_dispatch_busy(hctx, false);
1822                 *cookie = BLK_QC_T_NONE;
1823                 break;
1824         }
1825
1826         return ret;
1827 }
1828
1829 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1830                                                 struct request *rq,
1831                                                 blk_qc_t *cookie,
1832                                                 bool bypass_insert, bool last)
1833 {
1834         struct request_queue *q = rq->q;
1835         bool run_queue = true;
1836
1837         /*
1838          * RCU or SRCU read lock is needed before checking quiesced flag.
1839          *
1840          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1841          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1842          * and avoid driver to try to dispatch again.
1843          */
1844         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1845                 run_queue = false;
1846                 bypass_insert = false;
1847                 goto insert;
1848         }
1849
1850         if (q->elevator && !bypass_insert)
1851                 goto insert;
1852
1853         if (!blk_mq_get_dispatch_budget(hctx))
1854                 goto insert;
1855
1856         if (!blk_mq_get_driver_tag(rq)) {
1857                 blk_mq_put_dispatch_budget(hctx);
1858                 goto insert;
1859         }
1860
1861         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1862 insert:
1863         if (bypass_insert)
1864                 return BLK_STS_RESOURCE;
1865
1866         blk_mq_request_bypass_insert(rq, run_queue);
1867         return BLK_STS_OK;
1868 }
1869
1870 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1871                 struct request *rq, blk_qc_t *cookie)
1872 {
1873         blk_status_t ret;
1874         int srcu_idx;
1875
1876         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1877
1878         hctx_lock(hctx, &srcu_idx);
1879
1880         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1881         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1882                 blk_mq_request_bypass_insert(rq, true);
1883         else if (ret != BLK_STS_OK)
1884                 blk_mq_end_request(rq, ret);
1885
1886         hctx_unlock(hctx, srcu_idx);
1887 }
1888
1889 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1890 {
1891         blk_status_t ret;
1892         int srcu_idx;
1893         blk_qc_t unused_cookie;
1894         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1895
1896         hctx_lock(hctx, &srcu_idx);
1897         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1898         hctx_unlock(hctx, srcu_idx);
1899
1900         return ret;
1901 }
1902
1903 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1904                 struct list_head *list)
1905 {
1906         while (!list_empty(list)) {
1907                 blk_status_t ret;
1908                 struct request *rq = list_first_entry(list, struct request,
1909                                 queuelist);
1910
1911                 list_del_init(&rq->queuelist);
1912                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1913                 if (ret != BLK_STS_OK) {
1914                         if (ret == BLK_STS_RESOURCE ||
1915                                         ret == BLK_STS_DEV_RESOURCE) {
1916                                 blk_mq_request_bypass_insert(rq,
1917                                                         list_empty(list));
1918                                 break;
1919                         }
1920                         blk_mq_end_request(rq, ret);
1921                 }
1922         }
1923
1924         /*
1925          * If we didn't flush the entire list, we could have told
1926          * the driver there was more coming, but that turned out to
1927          * be a lie.
1928          */
1929         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1930                 hctx->queue->mq_ops->commit_rqs(hctx);
1931 }
1932
1933 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1934 {
1935         list_add_tail(&rq->queuelist, &plug->mq_list);
1936         plug->rq_count++;
1937         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1938                 struct request *tmp;
1939
1940                 tmp = list_first_entry(&plug->mq_list, struct request,
1941                                                 queuelist);
1942                 if (tmp->q != rq->q)
1943                         plug->multiple_queues = true;
1944         }
1945 }
1946
1947 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1948 {
1949         const int is_sync = op_is_sync(bio->bi_opf);
1950         const int is_flush_fua = op_is_flush(bio->bi_opf);
1951         struct blk_mq_alloc_data data = { .flags = 0};
1952         struct request *rq;
1953         struct blk_plug *plug;
1954         struct request *same_queue_rq = NULL;
1955         unsigned int nr_segs;
1956         blk_qc_t cookie;
1957
1958         blk_queue_bounce(q, &bio);
1959         __blk_queue_split(q, &bio, &nr_segs);
1960
1961         if (!bio_integrity_prep(bio))
1962                 return BLK_QC_T_NONE;
1963
1964         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1965             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1966                 return BLK_QC_T_NONE;
1967
1968         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1969                 return BLK_QC_T_NONE;
1970
1971         rq_qos_throttle(q, bio);
1972
1973         data.cmd_flags = bio->bi_opf;
1974         rq = blk_mq_get_request(q, bio, &data);
1975         if (unlikely(!rq)) {
1976                 rq_qos_cleanup(q, bio);
1977                 if (bio->bi_opf & REQ_NOWAIT)
1978                         bio_wouldblock_error(bio);
1979                 return BLK_QC_T_NONE;
1980         }
1981
1982         trace_block_getrq(q, bio, bio->bi_opf);
1983
1984         rq_qos_track(q, rq, bio);
1985
1986         cookie = request_to_qc_t(data.hctx, rq);
1987
1988         blk_mq_bio_to_request(rq, bio, nr_segs);
1989
1990         plug = blk_mq_plug(q, bio);
1991         if (unlikely(is_flush_fua)) {
1992                 /* bypass scheduler for flush rq */
1993                 blk_insert_flush(rq);
1994                 blk_mq_run_hw_queue(data.hctx, true);
1995         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1996                 /*
1997                  * Use plugging if we have a ->commit_rqs() hook as well, as
1998                  * we know the driver uses bd->last in a smart fashion.
1999                  */
2000                 unsigned int request_count = plug->rq_count;
2001                 struct request *last = NULL;
2002
2003                 if (!request_count)
2004                         trace_block_plug(q);
2005                 else
2006                         last = list_entry_rq(plug->mq_list.prev);
2007
2008                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2009                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2010                         blk_flush_plug_list(plug, false);
2011                         trace_block_plug(q);
2012                 }
2013
2014                 blk_add_rq_to_plug(plug, rq);
2015         } else if (q->elevator) {
2016                 blk_mq_sched_insert_request(rq, false, true, true);
2017         } else if (plug && !blk_queue_nomerges(q)) {
2018                 /*
2019                  * We do limited plugging. If the bio can be merged, do that.
2020                  * Otherwise the existing request in the plug list will be
2021                  * issued. So the plug list will have one request at most
2022                  * The plug list might get flushed before this. If that happens,
2023                  * the plug list is empty, and same_queue_rq is invalid.
2024                  */
2025                 if (list_empty(&plug->mq_list))
2026                         same_queue_rq = NULL;
2027                 if (same_queue_rq) {
2028                         list_del_init(&same_queue_rq->queuelist);
2029                         plug->rq_count--;
2030                 }
2031                 blk_add_rq_to_plug(plug, rq);
2032                 trace_block_plug(q);
2033
2034                 if (same_queue_rq) {
2035                         data.hctx = same_queue_rq->mq_hctx;
2036                         trace_block_unplug(q, 1, true);
2037                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2038                                         &cookie);
2039                 }
2040         } else if ((q->nr_hw_queues > 1 && is_sync) ||
2041                         !data.hctx->dispatch_busy) {
2042                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2043         } else {
2044                 blk_mq_sched_insert_request(rq, false, true, true);
2045         }
2046
2047         return cookie;
2048 }
2049
2050 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2051                      unsigned int hctx_idx)
2052 {
2053         struct page *page;
2054
2055         if (tags->rqs && set->ops->exit_request) {
2056                 int i;
2057
2058                 for (i = 0; i < tags->nr_tags; i++) {
2059                         struct request *rq = tags->static_rqs[i];
2060
2061                         if (!rq)
2062                                 continue;
2063                         set->ops->exit_request(set, rq, hctx_idx);
2064                         tags->static_rqs[i] = NULL;
2065                 }
2066         }
2067
2068         while (!list_empty(&tags->page_list)) {
2069                 page = list_first_entry(&tags->page_list, struct page, lru);
2070                 list_del_init(&page->lru);
2071                 /*
2072                  * Remove kmemleak object previously allocated in
2073                  * blk_mq_alloc_rqs().
2074                  */
2075                 kmemleak_free(page_address(page));
2076                 __free_pages(page, page->private);
2077         }
2078 }
2079
2080 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2081 {
2082         kfree(tags->rqs);
2083         tags->rqs = NULL;
2084         kfree(tags->static_rqs);
2085         tags->static_rqs = NULL;
2086
2087         blk_mq_free_tags(tags);
2088 }
2089
2090 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2091                                         unsigned int hctx_idx,
2092                                         unsigned int nr_tags,
2093                                         unsigned int reserved_tags)
2094 {
2095         struct blk_mq_tags *tags;
2096         int node;
2097
2098         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2099         if (node == NUMA_NO_NODE)
2100                 node = set->numa_node;
2101
2102         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2103                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2104         if (!tags)
2105                 return NULL;
2106
2107         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2108                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2109                                  node);
2110         if (!tags->rqs) {
2111                 blk_mq_free_tags(tags);
2112                 return NULL;
2113         }
2114
2115         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2116                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2117                                         node);
2118         if (!tags->static_rqs) {
2119                 kfree(tags->rqs);
2120                 blk_mq_free_tags(tags);
2121                 return NULL;
2122         }
2123
2124         return tags;
2125 }
2126
2127 static size_t order_to_size(unsigned int order)
2128 {
2129         return (size_t)PAGE_SIZE << order;
2130 }
2131
2132 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2133                                unsigned int hctx_idx, int node)
2134 {
2135         int ret;
2136
2137         if (set->ops->init_request) {
2138                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2139                 if (ret)
2140                         return ret;
2141         }
2142
2143         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2144         return 0;
2145 }
2146
2147 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2148                      unsigned int hctx_idx, unsigned int depth)
2149 {
2150         unsigned int i, j, entries_per_page, max_order = 4;
2151         size_t rq_size, left;
2152         int node;
2153
2154         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2155         if (node == NUMA_NO_NODE)
2156                 node = set->numa_node;
2157
2158         INIT_LIST_HEAD(&tags->page_list);
2159
2160         /*
2161          * rq_size is the size of the request plus driver payload, rounded
2162          * to the cacheline size
2163          */
2164         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2165                                 cache_line_size());
2166         left = rq_size * depth;
2167
2168         for (i = 0; i < depth; ) {
2169                 int this_order = max_order;
2170                 struct page *page;
2171                 int to_do;
2172                 void *p;
2173
2174                 while (this_order && left < order_to_size(this_order - 1))
2175                         this_order--;
2176
2177                 do {
2178                         page = alloc_pages_node(node,
2179                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2180                                 this_order);
2181                         if (page)
2182                                 break;
2183                         if (!this_order--)
2184                                 break;
2185                         if (order_to_size(this_order) < rq_size)
2186                                 break;
2187                 } while (1);
2188
2189                 if (!page)
2190                         goto fail;
2191
2192                 page->private = this_order;
2193                 list_add_tail(&page->lru, &tags->page_list);
2194
2195                 p = page_address(page);
2196                 /*
2197                  * Allow kmemleak to scan these pages as they contain pointers
2198                  * to additional allocations like via ops->init_request().
2199                  */
2200                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2201                 entries_per_page = order_to_size(this_order) / rq_size;
2202                 to_do = min(entries_per_page, depth - i);
2203                 left -= to_do * rq_size;
2204                 for (j = 0; j < to_do; j++) {
2205                         struct request *rq = p;
2206
2207                         tags->static_rqs[i] = rq;
2208                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2209                                 tags->static_rqs[i] = NULL;
2210                                 goto fail;
2211                         }
2212
2213                         p += rq_size;
2214                         i++;
2215                 }
2216         }
2217         return 0;
2218
2219 fail:
2220         blk_mq_free_rqs(set, tags, hctx_idx);
2221         return -ENOMEM;
2222 }
2223
2224 /*
2225  * 'cpu' is going away. splice any existing rq_list entries from this
2226  * software queue to the hw queue dispatch list, and ensure that it
2227  * gets run.
2228  */
2229 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2230 {
2231         struct blk_mq_hw_ctx *hctx;
2232         struct blk_mq_ctx *ctx;
2233         LIST_HEAD(tmp);
2234         enum hctx_type type;
2235
2236         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2237         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2238         type = hctx->type;
2239
2240         spin_lock(&ctx->lock);
2241         if (!list_empty(&ctx->rq_lists[type])) {
2242                 list_splice_init(&ctx->rq_lists[type], &tmp);
2243                 blk_mq_hctx_clear_pending(hctx, ctx);
2244         }
2245         spin_unlock(&ctx->lock);
2246
2247         if (list_empty(&tmp))
2248                 return 0;
2249
2250         spin_lock(&hctx->lock);
2251         list_splice_tail_init(&tmp, &hctx->dispatch);
2252         spin_unlock(&hctx->lock);
2253
2254         blk_mq_run_hw_queue(hctx, true);
2255         return 0;
2256 }
2257
2258 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2259 {
2260         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2261                                             &hctx->cpuhp_dead);
2262 }
2263
2264 /* hctx->ctxs will be freed in queue's release handler */
2265 static void blk_mq_exit_hctx(struct request_queue *q,
2266                 struct blk_mq_tag_set *set,
2267                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2268 {
2269         if (blk_mq_hw_queue_mapped(hctx))
2270                 blk_mq_tag_idle(hctx);
2271
2272         if (set->ops->exit_request)
2273                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2274
2275         if (set->ops->exit_hctx)
2276                 set->ops->exit_hctx(hctx, hctx_idx);
2277
2278         blk_mq_remove_cpuhp(hctx);
2279
2280         spin_lock(&q->unused_hctx_lock);
2281         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2282         spin_unlock(&q->unused_hctx_lock);
2283 }
2284
2285 static void blk_mq_exit_hw_queues(struct request_queue *q,
2286                 struct blk_mq_tag_set *set, int nr_queue)
2287 {
2288         struct blk_mq_hw_ctx *hctx;
2289         unsigned int i;
2290
2291         queue_for_each_hw_ctx(q, hctx, i) {
2292                 if (i == nr_queue)
2293                         break;
2294                 blk_mq_debugfs_unregister_hctx(hctx);
2295                 blk_mq_exit_hctx(q, set, hctx, i);
2296         }
2297 }
2298
2299 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2300 {
2301         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2302
2303         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2304                            __alignof__(struct blk_mq_hw_ctx)) !=
2305                      sizeof(struct blk_mq_hw_ctx));
2306
2307         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2308                 hw_ctx_size += sizeof(struct srcu_struct);
2309
2310         return hw_ctx_size;
2311 }
2312
2313 static int blk_mq_init_hctx(struct request_queue *q,
2314                 struct blk_mq_tag_set *set,
2315                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2316 {
2317         hctx->queue_num = hctx_idx;
2318
2319         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2320
2321         hctx->tags = set->tags[hctx_idx];
2322
2323         if (set->ops->init_hctx &&
2324             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2325                 goto unregister_cpu_notifier;
2326
2327         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2328                                 hctx->numa_node))
2329                 goto exit_hctx;
2330         return 0;
2331
2332  exit_hctx:
2333         if (set->ops->exit_hctx)
2334                 set->ops->exit_hctx(hctx, hctx_idx);
2335  unregister_cpu_notifier:
2336         blk_mq_remove_cpuhp(hctx);
2337         return -1;
2338 }
2339
2340 static struct blk_mq_hw_ctx *
2341 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2342                 int node)
2343 {
2344         struct blk_mq_hw_ctx *hctx;
2345         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2346
2347         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2348         if (!hctx)
2349                 goto fail_alloc_hctx;
2350
2351         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2352                 goto free_hctx;
2353
2354         atomic_set(&hctx->nr_active, 0);
2355         if (node == NUMA_NO_NODE)
2356                 node = set->numa_node;
2357         hctx->numa_node = node;
2358
2359         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2360         spin_lock_init(&hctx->lock);
2361         INIT_LIST_HEAD(&hctx->dispatch);
2362         hctx->queue = q;
2363         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2364
2365         INIT_LIST_HEAD(&hctx->hctx_list);
2366
2367         /*
2368          * Allocate space for all possible cpus to avoid allocation at
2369          * runtime
2370          */
2371         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2372                         gfp, node);
2373         if (!hctx->ctxs)
2374                 goto free_cpumask;
2375
2376         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2377                                 gfp, node))
2378                 goto free_ctxs;
2379         hctx->nr_ctx = 0;
2380
2381         spin_lock_init(&hctx->dispatch_wait_lock);
2382         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2383         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2384
2385         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2386                         gfp);
2387         if (!hctx->fq)
2388                 goto free_bitmap;
2389
2390         if (hctx->flags & BLK_MQ_F_BLOCKING)
2391                 init_srcu_struct(hctx->srcu);
2392         blk_mq_hctx_kobj_init(hctx);
2393
2394         return hctx;
2395
2396  free_bitmap:
2397         sbitmap_free(&hctx->ctx_map);
2398  free_ctxs:
2399         kfree(hctx->ctxs);
2400  free_cpumask:
2401         free_cpumask_var(hctx->cpumask);
2402  free_hctx:
2403         kfree(hctx);
2404  fail_alloc_hctx:
2405         return NULL;
2406 }
2407
2408 static void blk_mq_init_cpu_queues(struct request_queue *q,
2409                                    unsigned int nr_hw_queues)
2410 {
2411         struct blk_mq_tag_set *set = q->tag_set;
2412         unsigned int i, j;
2413
2414         for_each_possible_cpu(i) {
2415                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2416                 struct blk_mq_hw_ctx *hctx;
2417                 int k;
2418
2419                 __ctx->cpu = i;
2420                 spin_lock_init(&__ctx->lock);
2421                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2422                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2423
2424                 __ctx->queue = q;
2425
2426                 /*
2427                  * Set local node, IFF we have more than one hw queue. If
2428                  * not, we remain on the home node of the device
2429                  */
2430                 for (j = 0; j < set->nr_maps; j++) {
2431                         hctx = blk_mq_map_queue_type(q, j, i);
2432                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2433                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2434                 }
2435         }
2436 }
2437
2438 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2439 {
2440         int ret = 0;
2441
2442         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2443                                         set->queue_depth, set->reserved_tags);
2444         if (!set->tags[hctx_idx])
2445                 return false;
2446
2447         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2448                                 set->queue_depth);
2449         if (!ret)
2450                 return true;
2451
2452         blk_mq_free_rq_map(set->tags[hctx_idx]);
2453         set->tags[hctx_idx] = NULL;
2454         return false;
2455 }
2456
2457 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2458                                          unsigned int hctx_idx)
2459 {
2460         if (set->tags && set->tags[hctx_idx]) {
2461                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2462                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2463                 set->tags[hctx_idx] = NULL;
2464         }
2465 }
2466
2467 static void blk_mq_map_swqueue(struct request_queue *q)
2468 {
2469         unsigned int i, j, hctx_idx;
2470         struct blk_mq_hw_ctx *hctx;
2471         struct blk_mq_ctx *ctx;
2472         struct blk_mq_tag_set *set = q->tag_set;
2473
2474         queue_for_each_hw_ctx(q, hctx, i) {
2475                 cpumask_clear(hctx->cpumask);
2476                 hctx->nr_ctx = 0;
2477                 hctx->dispatch_from = NULL;
2478         }
2479
2480         /*
2481          * Map software to hardware queues.
2482          *
2483          * If the cpu isn't present, the cpu is mapped to first hctx.
2484          */
2485         for_each_possible_cpu(i) {
2486                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2487                 /* unmapped hw queue can be remapped after CPU topo changed */
2488                 if (!set->tags[hctx_idx] &&
2489                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2490                         /*
2491                          * If tags initialization fail for some hctx,
2492                          * that hctx won't be brought online.  In this
2493                          * case, remap the current ctx to hctx[0] which
2494                          * is guaranteed to always have tags allocated
2495                          */
2496                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2497                 }
2498
2499                 ctx = per_cpu_ptr(q->queue_ctx, i);
2500                 for (j = 0; j < set->nr_maps; j++) {
2501                         if (!set->map[j].nr_queues) {
2502                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2503                                                 HCTX_TYPE_DEFAULT, i);
2504                                 continue;
2505                         }
2506
2507                         hctx = blk_mq_map_queue_type(q, j, i);
2508                         ctx->hctxs[j] = hctx;
2509                         /*
2510                          * If the CPU is already set in the mask, then we've
2511                          * mapped this one already. This can happen if
2512                          * devices share queues across queue maps.
2513                          */
2514                         if (cpumask_test_cpu(i, hctx->cpumask))
2515                                 continue;
2516
2517                         cpumask_set_cpu(i, hctx->cpumask);
2518                         hctx->type = j;
2519                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2520                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2521
2522                         /*
2523                          * If the nr_ctx type overflows, we have exceeded the
2524                          * amount of sw queues we can support.
2525                          */
2526                         BUG_ON(!hctx->nr_ctx);
2527                 }
2528
2529                 for (; j < HCTX_MAX_TYPES; j++)
2530                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2531                                         HCTX_TYPE_DEFAULT, i);
2532         }
2533
2534         queue_for_each_hw_ctx(q, hctx, i) {
2535                 /*
2536                  * If no software queues are mapped to this hardware queue,
2537                  * disable it and free the request entries.
2538                  */
2539                 if (!hctx->nr_ctx) {
2540                         /* Never unmap queue 0.  We need it as a
2541                          * fallback in case of a new remap fails
2542                          * allocation
2543                          */
2544                         if (i && set->tags[i])
2545                                 blk_mq_free_map_and_requests(set, i);
2546
2547                         hctx->tags = NULL;
2548                         continue;
2549                 }
2550
2551                 hctx->tags = set->tags[i];
2552                 WARN_ON(!hctx->tags);
2553
2554                 /*
2555                  * Set the map size to the number of mapped software queues.
2556                  * This is more accurate and more efficient than looping
2557                  * over all possibly mapped software queues.
2558                  */
2559                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2560
2561                 /*
2562                  * Initialize batch roundrobin counts
2563                  */
2564                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2565                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2566         }
2567 }
2568
2569 /*
2570  * Caller needs to ensure that we're either frozen/quiesced, or that
2571  * the queue isn't live yet.
2572  */
2573 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2574 {
2575         struct blk_mq_hw_ctx *hctx;
2576         int i;
2577
2578         queue_for_each_hw_ctx(q, hctx, i) {
2579                 if (shared)
2580                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2581                 else
2582                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2583         }
2584 }
2585
2586 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2587                                         bool shared)
2588 {
2589         struct request_queue *q;
2590
2591         lockdep_assert_held(&set->tag_list_lock);
2592
2593         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2594                 blk_mq_freeze_queue(q);
2595                 queue_set_hctx_shared(q, shared);
2596                 blk_mq_unfreeze_queue(q);
2597         }
2598 }
2599
2600 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2601 {
2602         struct blk_mq_tag_set *set = q->tag_set;
2603
2604         mutex_lock(&set->tag_list_lock);
2605         list_del_rcu(&q->tag_set_list);
2606         if (list_is_singular(&set->tag_list)) {
2607                 /* just transitioned to unshared */
2608                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2609                 /* update existing queue */
2610                 blk_mq_update_tag_set_depth(set, false);
2611         }
2612         mutex_unlock(&set->tag_list_lock);
2613         INIT_LIST_HEAD(&q->tag_set_list);
2614 }
2615
2616 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2617                                      struct request_queue *q)
2618 {
2619         mutex_lock(&set->tag_list_lock);
2620
2621         /*
2622          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2623          */
2624         if (!list_empty(&set->tag_list) &&
2625             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2626                 set->flags |= BLK_MQ_F_TAG_SHARED;
2627                 /* update existing queue */
2628                 blk_mq_update_tag_set_depth(set, true);
2629         }
2630         if (set->flags & BLK_MQ_F_TAG_SHARED)
2631                 queue_set_hctx_shared(q, true);
2632         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2633
2634         mutex_unlock(&set->tag_list_lock);
2635 }
2636
2637 /* All allocations will be freed in release handler of q->mq_kobj */
2638 static int blk_mq_alloc_ctxs(struct request_queue *q)
2639 {
2640         struct blk_mq_ctxs *ctxs;
2641         int cpu;
2642
2643         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2644         if (!ctxs)
2645                 return -ENOMEM;
2646
2647         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2648         if (!ctxs->queue_ctx)
2649                 goto fail;
2650
2651         for_each_possible_cpu(cpu) {
2652                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2653                 ctx->ctxs = ctxs;
2654         }
2655
2656         q->mq_kobj = &ctxs->kobj;
2657         q->queue_ctx = ctxs->queue_ctx;
2658
2659         return 0;
2660  fail:
2661         kfree(ctxs);
2662         return -ENOMEM;
2663 }
2664
2665 /*
2666  * It is the actual release handler for mq, but we do it from
2667  * request queue's release handler for avoiding use-after-free
2668  * and headache because q->mq_kobj shouldn't have been introduced,
2669  * but we can't group ctx/kctx kobj without it.
2670  */
2671 void blk_mq_release(struct request_queue *q)
2672 {
2673         struct blk_mq_hw_ctx *hctx, *next;
2674         int i;
2675
2676         queue_for_each_hw_ctx(q, hctx, i)
2677                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2678
2679         /* all hctx are in .unused_hctx_list now */
2680         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2681                 list_del_init(&hctx->hctx_list);
2682                 kobject_put(&hctx->kobj);
2683         }
2684
2685         kfree(q->queue_hw_ctx);
2686
2687         /*
2688          * release .mq_kobj and sw queue's kobject now because
2689          * both share lifetime with request queue.
2690          */
2691         blk_mq_sysfs_deinit(q);
2692 }
2693
2694 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2695 {
2696         struct request_queue *uninit_q, *q;
2697
2698         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2699         if (!uninit_q)
2700                 return ERR_PTR(-ENOMEM);
2701
2702         /*
2703          * Initialize the queue without an elevator. device_add_disk() will do
2704          * the initialization.
2705          */
2706         q = blk_mq_init_allocated_queue(set, uninit_q, false);
2707         if (IS_ERR(q))
2708                 blk_cleanup_queue(uninit_q);
2709
2710         return q;
2711 }
2712 EXPORT_SYMBOL(blk_mq_init_queue);
2713
2714 /*
2715  * Helper for setting up a queue with mq ops, given queue depth, and
2716  * the passed in mq ops flags.
2717  */
2718 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2719                                            const struct blk_mq_ops *ops,
2720                                            unsigned int queue_depth,
2721                                            unsigned int set_flags)
2722 {
2723         struct request_queue *q;
2724         int ret;
2725
2726         memset(set, 0, sizeof(*set));
2727         set->ops = ops;
2728         set->nr_hw_queues = 1;
2729         set->nr_maps = 1;
2730         set->queue_depth = queue_depth;
2731         set->numa_node = NUMA_NO_NODE;
2732         set->flags = set_flags;
2733
2734         ret = blk_mq_alloc_tag_set(set);
2735         if (ret)
2736                 return ERR_PTR(ret);
2737
2738         q = blk_mq_init_queue(set);
2739         if (IS_ERR(q)) {
2740                 blk_mq_free_tag_set(set);
2741                 return q;
2742         }
2743
2744         return q;
2745 }
2746 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2747
2748 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2749                 struct blk_mq_tag_set *set, struct request_queue *q,
2750                 int hctx_idx, int node)
2751 {
2752         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2753
2754         /* reuse dead hctx first */
2755         spin_lock(&q->unused_hctx_lock);
2756         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2757                 if (tmp->numa_node == node) {
2758                         hctx = tmp;
2759                         break;
2760                 }
2761         }
2762         if (hctx)
2763                 list_del_init(&hctx->hctx_list);
2764         spin_unlock(&q->unused_hctx_lock);
2765
2766         if (!hctx)
2767                 hctx = blk_mq_alloc_hctx(q, set, node);
2768         if (!hctx)
2769                 goto fail;
2770
2771         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2772                 goto free_hctx;
2773
2774         return hctx;
2775
2776  free_hctx:
2777         kobject_put(&hctx->kobj);
2778  fail:
2779         return NULL;
2780 }
2781
2782 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2783                                                 struct request_queue *q)
2784 {
2785         int i, j, end;
2786         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2787
2788         /* protect against switching io scheduler  */
2789         mutex_lock(&q->sysfs_lock);
2790         for (i = 0; i < set->nr_hw_queues; i++) {
2791                 int node;
2792                 struct blk_mq_hw_ctx *hctx;
2793
2794                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2795                 /*
2796                  * If the hw queue has been mapped to another numa node,
2797                  * we need to realloc the hctx. If allocation fails, fallback
2798                  * to use the previous one.
2799                  */
2800                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2801                         continue;
2802
2803                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2804                 if (hctx) {
2805                         if (hctxs[i])
2806                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2807                         hctxs[i] = hctx;
2808                 } else {
2809                         if (hctxs[i])
2810                                 pr_warn("Allocate new hctx on node %d fails,\
2811                                                 fallback to previous one on node %d\n",
2812                                                 node, hctxs[i]->numa_node);
2813                         else
2814                                 break;
2815                 }
2816         }
2817         /*
2818          * Increasing nr_hw_queues fails. Free the newly allocated
2819          * hctxs and keep the previous q->nr_hw_queues.
2820          */
2821         if (i != set->nr_hw_queues) {
2822                 j = q->nr_hw_queues;
2823                 end = i;
2824         } else {
2825                 j = i;
2826                 end = q->nr_hw_queues;
2827                 q->nr_hw_queues = set->nr_hw_queues;
2828         }
2829
2830         for (; j < end; j++) {
2831                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2832
2833                 if (hctx) {
2834                         if (hctx->tags)
2835                                 blk_mq_free_map_and_requests(set, j);
2836                         blk_mq_exit_hctx(q, set, hctx, j);
2837                         hctxs[j] = NULL;
2838                 }
2839         }
2840         mutex_unlock(&q->sysfs_lock);
2841 }
2842
2843 /*
2844  * Maximum number of hardware queues we support. For single sets, we'll never
2845  * have more than the CPUs (software queues). For multiple sets, the tag_set
2846  * user may have set ->nr_hw_queues larger.
2847  */
2848 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2849 {
2850         if (set->nr_maps == 1)
2851                 return nr_cpu_ids;
2852
2853         return max(set->nr_hw_queues, nr_cpu_ids);
2854 }
2855
2856 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2857                                                   struct request_queue *q,
2858                                                   bool elevator_init)
2859 {
2860         /* mark the queue as mq asap */
2861         q->mq_ops = set->ops;
2862
2863         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2864                                              blk_mq_poll_stats_bkt,
2865                                              BLK_MQ_POLL_STATS_BKTS, q);
2866         if (!q->poll_cb)
2867                 goto err_exit;
2868
2869         if (blk_mq_alloc_ctxs(q))
2870                 goto err_poll;
2871
2872         /* init q->mq_kobj and sw queues' kobjects */
2873         blk_mq_sysfs_init(q);
2874
2875         q->nr_queues = nr_hw_queues(set);
2876         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2877                                                 GFP_KERNEL, set->numa_node);
2878         if (!q->queue_hw_ctx)
2879                 goto err_sys_init;
2880
2881         INIT_LIST_HEAD(&q->unused_hctx_list);
2882         spin_lock_init(&q->unused_hctx_lock);
2883
2884         blk_mq_realloc_hw_ctxs(set, q);
2885         if (!q->nr_hw_queues)
2886                 goto err_hctxs;
2887
2888         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2889         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2890
2891         q->tag_set = set;
2892
2893         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2894         if (set->nr_maps > HCTX_TYPE_POLL &&
2895             set->map[HCTX_TYPE_POLL].nr_queues)
2896                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2897
2898         q->sg_reserved_size = INT_MAX;
2899
2900         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2901         INIT_LIST_HEAD(&q->requeue_list);
2902         spin_lock_init(&q->requeue_lock);
2903
2904         blk_queue_make_request(q, blk_mq_make_request);
2905
2906         /*
2907          * Do this after blk_queue_make_request() overrides it...
2908          */
2909         q->nr_requests = set->queue_depth;
2910
2911         /*
2912          * Default to classic polling
2913          */
2914         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2915
2916         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2917         blk_mq_add_queue_tag_set(set, q);
2918         blk_mq_map_swqueue(q);
2919
2920         if (elevator_init)
2921                 elevator_init_mq(q);
2922
2923         return q;
2924
2925 err_hctxs:
2926         kfree(q->queue_hw_ctx);
2927         q->nr_hw_queues = 0;
2928 err_sys_init:
2929         blk_mq_sysfs_deinit(q);
2930 err_poll:
2931         blk_stat_free_callback(q->poll_cb);
2932         q->poll_cb = NULL;
2933 err_exit:
2934         q->mq_ops = NULL;
2935         return ERR_PTR(-ENOMEM);
2936 }
2937 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2938
2939 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2940 void blk_mq_exit_queue(struct request_queue *q)
2941 {
2942         struct blk_mq_tag_set   *set = q->tag_set;
2943
2944         blk_mq_del_queue_tag_set(q);
2945         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2946 }
2947
2948 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2949 {
2950         int i;
2951
2952         for (i = 0; i < set->nr_hw_queues; i++)
2953                 if (!__blk_mq_alloc_rq_map(set, i))
2954                         goto out_unwind;
2955
2956         return 0;
2957
2958 out_unwind:
2959         while (--i >= 0)
2960                 blk_mq_free_rq_map(set->tags[i]);
2961
2962         return -ENOMEM;
2963 }
2964
2965 /*
2966  * Allocate the request maps associated with this tag_set. Note that this
2967  * may reduce the depth asked for, if memory is tight. set->queue_depth
2968  * will be updated to reflect the allocated depth.
2969  */
2970 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2971 {
2972         unsigned int depth;
2973         int err;
2974
2975         depth = set->queue_depth;
2976         do {
2977                 err = __blk_mq_alloc_rq_maps(set);
2978                 if (!err)
2979                         break;
2980
2981                 set->queue_depth >>= 1;
2982                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2983                         err = -ENOMEM;
2984                         break;
2985                 }
2986         } while (set->queue_depth);
2987
2988         if (!set->queue_depth || err) {
2989                 pr_err("blk-mq: failed to allocate request map\n");
2990                 return -ENOMEM;
2991         }
2992
2993         if (depth != set->queue_depth)
2994                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2995                                                 depth, set->queue_depth);
2996
2997         return 0;
2998 }
2999
3000 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3001 {
3002         if (set->ops->map_queues && !is_kdump_kernel()) {
3003                 int i;
3004
3005                 /*
3006                  * transport .map_queues is usually done in the following
3007                  * way:
3008                  *
3009                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3010                  *      mask = get_cpu_mask(queue)
3011                  *      for_each_cpu(cpu, mask)
3012                  *              set->map[x].mq_map[cpu] = queue;
3013                  * }
3014                  *
3015                  * When we need to remap, the table has to be cleared for
3016                  * killing stale mapping since one CPU may not be mapped
3017                  * to any hw queue.
3018                  */
3019                 for (i = 0; i < set->nr_maps; i++)
3020                         blk_mq_clear_mq_map(&set->map[i]);
3021
3022                 return set->ops->map_queues(set);
3023         } else {
3024                 BUG_ON(set->nr_maps > 1);
3025                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3026         }
3027 }
3028
3029 /*
3030  * Alloc a tag set to be associated with one or more request queues.
3031  * May fail with EINVAL for various error conditions. May adjust the
3032  * requested depth down, if it's too large. In that case, the set
3033  * value will be stored in set->queue_depth.
3034  */
3035 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3036 {
3037         int i, ret;
3038
3039         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3040
3041         if (!set->nr_hw_queues)
3042                 return -EINVAL;
3043         if (!set->queue_depth)
3044                 return -EINVAL;
3045         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3046                 return -EINVAL;
3047
3048         if (!set->ops->queue_rq)
3049                 return -EINVAL;
3050
3051         if (!set->ops->get_budget ^ !set->ops->put_budget)
3052                 return -EINVAL;
3053
3054         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3055                 pr_info("blk-mq: reduced tag depth to %u\n",
3056                         BLK_MQ_MAX_DEPTH);
3057                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3058         }
3059
3060         if (!set->nr_maps)
3061                 set->nr_maps = 1;
3062         else if (set->nr_maps > HCTX_MAX_TYPES)
3063                 return -EINVAL;
3064
3065         /*
3066          * If a crashdump is active, then we are potentially in a very
3067          * memory constrained environment. Limit us to 1 queue and
3068          * 64 tags to prevent using too much memory.
3069          */
3070         if (is_kdump_kernel()) {
3071                 set->nr_hw_queues = 1;
3072                 set->nr_maps = 1;
3073                 set->queue_depth = min(64U, set->queue_depth);
3074         }
3075         /*
3076          * There is no use for more h/w queues than cpus if we just have
3077          * a single map
3078          */
3079         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3080                 set->nr_hw_queues = nr_cpu_ids;
3081
3082         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3083                                  GFP_KERNEL, set->numa_node);
3084         if (!set->tags)
3085                 return -ENOMEM;
3086
3087         ret = -ENOMEM;
3088         for (i = 0; i < set->nr_maps; i++) {
3089                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3090                                                   sizeof(set->map[i].mq_map[0]),
3091                                                   GFP_KERNEL, set->numa_node);
3092                 if (!set->map[i].mq_map)
3093                         goto out_free_mq_map;
3094                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3095         }
3096
3097         ret = blk_mq_update_queue_map(set);
3098         if (ret)
3099                 goto out_free_mq_map;
3100
3101         ret = blk_mq_alloc_rq_maps(set);
3102         if (ret)
3103                 goto out_free_mq_map;
3104
3105         mutex_init(&set->tag_list_lock);
3106         INIT_LIST_HEAD(&set->tag_list);
3107
3108         return 0;
3109
3110 out_free_mq_map:
3111         for (i = 0; i < set->nr_maps; i++) {
3112                 kfree(set->map[i].mq_map);
3113                 set->map[i].mq_map = NULL;
3114         }
3115         kfree(set->tags);
3116         set->tags = NULL;
3117         return ret;
3118 }
3119 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3120
3121 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3122 {
3123         int i, j;
3124
3125         for (i = 0; i < nr_hw_queues(set); i++)
3126                 blk_mq_free_map_and_requests(set, i);
3127
3128         for (j = 0; j < set->nr_maps; j++) {
3129                 kfree(set->map[j].mq_map);
3130                 set->map[j].mq_map = NULL;
3131         }
3132
3133         kfree(set->tags);
3134         set->tags = NULL;
3135 }
3136 EXPORT_SYMBOL(blk_mq_free_tag_set);
3137
3138 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3139 {
3140         struct blk_mq_tag_set *set = q->tag_set;
3141         struct blk_mq_hw_ctx *hctx;
3142         int i, ret;
3143
3144         if (!set)
3145                 return -EINVAL;
3146
3147         if (q->nr_requests == nr)
3148                 return 0;
3149
3150         blk_mq_freeze_queue(q);
3151         blk_mq_quiesce_queue(q);
3152
3153         ret = 0;
3154         queue_for_each_hw_ctx(q, hctx, i) {
3155                 if (!hctx->tags)
3156                         continue;
3157                 /*
3158                  * If we're using an MQ scheduler, just update the scheduler
3159                  * queue depth. This is similar to what the old code would do.
3160                  */
3161                 if (!hctx->sched_tags) {
3162                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3163                                                         false);
3164                 } else {
3165                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3166                                                         nr, true);
3167                 }
3168                 if (ret)
3169                         break;
3170                 if (q->elevator && q->elevator->type->ops.depth_updated)
3171                         q->elevator->type->ops.depth_updated(hctx);
3172         }
3173
3174         if (!ret)
3175                 q->nr_requests = nr;
3176
3177         blk_mq_unquiesce_queue(q);
3178         blk_mq_unfreeze_queue(q);
3179
3180         return ret;
3181 }
3182
3183 /*
3184  * request_queue and elevator_type pair.
3185  * It is just used by __blk_mq_update_nr_hw_queues to cache
3186  * the elevator_type associated with a request_queue.
3187  */
3188 struct blk_mq_qe_pair {
3189         struct list_head node;
3190         struct request_queue *q;
3191         struct elevator_type *type;
3192 };
3193
3194 /*
3195  * Cache the elevator_type in qe pair list and switch the
3196  * io scheduler to 'none'
3197  */
3198 static bool blk_mq_elv_switch_none(struct list_head *head,
3199                 struct request_queue *q)
3200 {
3201         struct blk_mq_qe_pair *qe;
3202
3203         if (!q->elevator)
3204                 return true;
3205
3206         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3207         if (!qe)
3208                 return false;
3209
3210         INIT_LIST_HEAD(&qe->node);
3211         qe->q = q;
3212         qe->type = q->elevator->type;
3213         list_add(&qe->node, head);
3214
3215         mutex_lock(&q->sysfs_lock);
3216         /*
3217          * After elevator_switch_mq, the previous elevator_queue will be
3218          * released by elevator_release. The reference of the io scheduler
3219          * module get by elevator_get will also be put. So we need to get
3220          * a reference of the io scheduler module here to prevent it to be
3221          * removed.
3222          */
3223         __module_get(qe->type->elevator_owner);
3224         elevator_switch_mq(q, NULL);
3225         mutex_unlock(&q->sysfs_lock);
3226
3227         return true;
3228 }
3229
3230 static void blk_mq_elv_switch_back(struct list_head *head,
3231                 struct request_queue *q)
3232 {
3233         struct blk_mq_qe_pair *qe;
3234         struct elevator_type *t = NULL;
3235
3236         list_for_each_entry(qe, head, node)
3237                 if (qe->q == q) {
3238                         t = qe->type;
3239                         break;
3240                 }
3241
3242         if (!t)
3243                 return;
3244
3245         list_del(&qe->node);
3246         kfree(qe);
3247
3248         mutex_lock(&q->sysfs_lock);
3249         elevator_switch_mq(q, t);
3250         mutex_unlock(&q->sysfs_lock);
3251 }
3252
3253 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3254                                                         int nr_hw_queues)
3255 {
3256         struct request_queue *q;
3257         LIST_HEAD(head);
3258         int prev_nr_hw_queues;
3259
3260         lockdep_assert_held(&set->tag_list_lock);
3261
3262         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3263                 nr_hw_queues = nr_cpu_ids;
3264         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3265                 return;
3266
3267         list_for_each_entry(q, &set->tag_list, tag_set_list)
3268                 blk_mq_freeze_queue(q);
3269         /*
3270          * Sync with blk_mq_queue_tag_busy_iter.
3271          */
3272         synchronize_rcu();
3273         /*
3274          * Switch IO scheduler to 'none', cleaning up the data associated
3275          * with the previous scheduler. We will switch back once we are done
3276          * updating the new sw to hw queue mappings.
3277          */
3278         list_for_each_entry(q, &set->tag_list, tag_set_list)
3279                 if (!blk_mq_elv_switch_none(&head, q))
3280                         goto switch_back;
3281
3282         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3283                 blk_mq_debugfs_unregister_hctxs(q);
3284                 blk_mq_sysfs_unregister(q);
3285         }
3286
3287         prev_nr_hw_queues = set->nr_hw_queues;
3288         set->nr_hw_queues = nr_hw_queues;
3289         blk_mq_update_queue_map(set);
3290 fallback:
3291         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3292                 blk_mq_realloc_hw_ctxs(set, q);
3293                 if (q->nr_hw_queues != set->nr_hw_queues) {
3294                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3295                                         nr_hw_queues, prev_nr_hw_queues);
3296                         set->nr_hw_queues = prev_nr_hw_queues;
3297                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3298                         goto fallback;
3299                 }
3300                 blk_mq_map_swqueue(q);
3301         }
3302
3303         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3304                 blk_mq_sysfs_register(q);
3305                 blk_mq_debugfs_register_hctxs(q);
3306         }
3307
3308 switch_back:
3309         list_for_each_entry(q, &set->tag_list, tag_set_list)
3310                 blk_mq_elv_switch_back(&head, q);
3311
3312         list_for_each_entry(q, &set->tag_list, tag_set_list)
3313                 blk_mq_unfreeze_queue(q);
3314 }
3315
3316 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3317 {
3318         mutex_lock(&set->tag_list_lock);
3319         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3320         mutex_unlock(&set->tag_list_lock);
3321 }
3322 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3323
3324 /* Enable polling stats and return whether they were already enabled. */
3325 static bool blk_poll_stats_enable(struct request_queue *q)
3326 {
3327         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3328             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3329                 return true;
3330         blk_stat_add_callback(q, q->poll_cb);
3331         return false;
3332 }
3333
3334 static void blk_mq_poll_stats_start(struct request_queue *q)
3335 {
3336         /*
3337          * We don't arm the callback if polling stats are not enabled or the
3338          * callback is already active.
3339          */
3340         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3341             blk_stat_is_active(q->poll_cb))
3342                 return;
3343
3344         blk_stat_activate_msecs(q->poll_cb, 100);
3345 }
3346
3347 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3348 {
3349         struct request_queue *q = cb->data;
3350         int bucket;
3351
3352         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3353                 if (cb->stat[bucket].nr_samples)
3354                         q->poll_stat[bucket] = cb->stat[bucket];
3355         }
3356 }
3357
3358 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3359                                        struct blk_mq_hw_ctx *hctx,
3360                                        struct request *rq)
3361 {
3362         unsigned long ret = 0;
3363         int bucket;
3364
3365         /*
3366          * If stats collection isn't on, don't sleep but turn it on for
3367          * future users
3368          */
3369         if (!blk_poll_stats_enable(q))
3370                 return 0;
3371
3372         /*
3373          * As an optimistic guess, use half of the mean service time
3374          * for this type of request. We can (and should) make this smarter.
3375          * For instance, if the completion latencies are tight, we can
3376          * get closer than just half the mean. This is especially
3377          * important on devices where the completion latencies are longer
3378          * than ~10 usec. We do use the stats for the relevant IO size
3379          * if available which does lead to better estimates.
3380          */
3381         bucket = blk_mq_poll_stats_bkt(rq);
3382         if (bucket < 0)
3383                 return ret;
3384
3385         if (q->poll_stat[bucket].nr_samples)
3386                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3387
3388         return ret;
3389 }
3390
3391 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3392                                      struct blk_mq_hw_ctx *hctx,
3393                                      struct request *rq)
3394 {
3395         struct hrtimer_sleeper hs;
3396         enum hrtimer_mode mode;
3397         unsigned int nsecs;
3398         ktime_t kt;
3399
3400         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3401                 return false;
3402
3403         /*
3404          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3405          *
3406          *  0:  use half of prev avg
3407          * >0:  use this specific value
3408          */
3409         if (q->poll_nsec > 0)
3410                 nsecs = q->poll_nsec;
3411         else
3412                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3413
3414         if (!nsecs)
3415                 return false;
3416
3417         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3418
3419         /*
3420          * This will be replaced with the stats tracking code, using
3421          * 'avg_completion_time / 2' as the pre-sleep target.
3422          */
3423         kt = nsecs;
3424
3425         mode = HRTIMER_MODE_REL;
3426         hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3427         hrtimer_set_expires(&hs.timer, kt);
3428
3429         do {
3430                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3431                         break;
3432                 set_current_state(TASK_UNINTERRUPTIBLE);
3433                 hrtimer_sleeper_start_expires(&hs, mode);
3434                 if (hs.task)
3435                         io_schedule();
3436                 hrtimer_cancel(&hs.timer);
3437                 mode = HRTIMER_MODE_ABS;
3438         } while (hs.task && !signal_pending(current));
3439
3440         __set_current_state(TASK_RUNNING);
3441         destroy_hrtimer_on_stack(&hs.timer);
3442         return true;
3443 }
3444
3445 static bool blk_mq_poll_hybrid(struct request_queue *q,
3446                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3447 {
3448         struct request *rq;
3449
3450         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3451                 return false;
3452
3453         if (!blk_qc_t_is_internal(cookie))
3454                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3455         else {
3456                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3457                 /*
3458                  * With scheduling, if the request has completed, we'll
3459                  * get a NULL return here, as we clear the sched tag when
3460                  * that happens. The request still remains valid, like always,
3461                  * so we should be safe with just the NULL check.
3462                  */
3463                 if (!rq)
3464                         return false;
3465         }
3466
3467         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3468 }
3469
3470 /**
3471  * blk_poll - poll for IO completions
3472  * @q:  the queue
3473  * @cookie: cookie passed back at IO submission time
3474  * @spin: whether to spin for completions
3475  *
3476  * Description:
3477  *    Poll for completions on the passed in queue. Returns number of
3478  *    completed entries found. If @spin is true, then blk_poll will continue
3479  *    looping until at least one completion is found, unless the task is
3480  *    otherwise marked running (or we need to reschedule).
3481  */
3482 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3483 {
3484         struct blk_mq_hw_ctx *hctx;
3485         long state;
3486
3487         if (!blk_qc_t_valid(cookie) ||
3488             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3489                 return 0;
3490
3491         if (current->plug)
3492                 blk_flush_plug_list(current->plug, false);
3493
3494         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3495
3496         /*
3497          * If we sleep, have the caller restart the poll loop to reset
3498          * the state. Like for the other success return cases, the
3499          * caller is responsible for checking if the IO completed. If
3500          * the IO isn't complete, we'll get called again and will go
3501          * straight to the busy poll loop.
3502          */
3503         if (blk_mq_poll_hybrid(q, hctx, cookie))
3504                 return 1;
3505
3506         hctx->poll_considered++;
3507
3508         state = current->state;
3509         do {
3510                 int ret;
3511
3512                 hctx->poll_invoked++;
3513
3514                 ret = q->mq_ops->poll(hctx);
3515                 if (ret > 0) {
3516                         hctx->poll_success++;
3517                         __set_current_state(TASK_RUNNING);
3518                         return ret;
3519                 }
3520
3521                 if (signal_pending_state(state, current))
3522                         __set_current_state(TASK_RUNNING);
3523
3524                 if (current->state == TASK_RUNNING)
3525                         return 1;
3526                 if (ret < 0 || !spin)
3527                         break;
3528                 cpu_relax();
3529         } while (!need_resched());
3530
3531         __set_current_state(TASK_RUNNING);
3532         return 0;
3533 }
3534 EXPORT_SYMBOL_GPL(blk_poll);
3535
3536 unsigned int blk_mq_rq_cpu(struct request *rq)
3537 {
3538         return rq->mq_ctx->cpu;
3539 }
3540 EXPORT_SYMBOL(blk_mq_rq_cpu);
3541
3542 static int __init blk_mq_init(void)
3543 {
3544         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3545                                 blk_mq_hctx_notify_dead);
3546         return 0;
3547 }
3548 subsys_initcall(blk_mq_init);