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         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 737 }
 738 EXPORT_SYMBOL(blk_mq_requeue_request);
 739
 740 static void blk_mq_requeue_work(struct work_struct *work)
 741 {
 742         struct request_queue *q =
 743                 container_of(work, struct request_queue, requeue_work.work);
 744         LIST_HEAD(rq_list);
 745         struct request *rq, *next;
 746
 747         spin_lock_irq(&q->requeue_lock);
 748         list_splice_init(&q->requeue_list, &rq_list);
 749         spin_unlock_irq(&q->requeue_lock);
 750
 751         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 752                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 753                         continue;
 754
 755                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 756                 list_del_init(&rq->queuelist);
 757                 /*
 758                  * If RQF_DONTPREP, rq has contained some driver specific
 759                  * data, so insert it to hctx dispatch list to avoid any
 760                  * merge.
 761                  */
 762                 if (rq->rq_flags & RQF_DONTPREP)
 763                         blk_mq_request_bypass_insert(rq, false, false);
 764                 else
 765                         blk_mq_sched_insert_request(rq, true, false, false);
 766         }
 767
 768         while (!list_empty(&rq_list)) {
 769                 rq = list_entry(rq_list.next, struct request, queuelist);
 770                 list_del_init(&rq->queuelist);
 771                 blk_mq_sched_insert_request(rq, false, false, false);
 772         }
 773
 774         blk_mq_run_hw_queues(q, false);
 775 }
 776
 777 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 778                                 bool kick_requeue_list)
 779 {
 780         struct request_queue *q = rq->q;
 781         unsigned long flags;
 782
 783         /*
 784          * We abuse this flag that is otherwise used by the I/O scheduler to
 785          * request head insertion from the workqueue.
 786          */
 787         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 788
 789         spin_lock_irqsave(&q->requeue_lock, flags);
 790         if (at_head) {
 791                 rq->rq_flags |= RQF_SOFTBARRIER;
 792                 list_add(&rq->queuelist, &q->requeue_list);
 793         } else {
 794                 list_add_tail(&rq->queuelist, &q->requeue_list);
 795         }
 796         spin_unlock_irqrestore(&q->requeue_lock, flags);
 797
 798         if (kick_requeue_list)
 799                 blk_mq_kick_requeue_list(q);
 800 }
 801
 802 void blk_mq_kick_requeue_list(struct request_queue *q)
 803 {
 804         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 805 }
 806 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 807
 808 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 809                                     unsigned long msecs)
 810 {
 811         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 812                                     msecs_to_jiffies(msecs));
 813 }
 814 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 815
 816 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 817 {
 818         if (tag < tags->nr_tags) {
 819                 prefetch(tags->rqs[tag]);
 820                 return tags->rqs[tag];
 821         }
 822
 823         return NULL;
 824 }
 825 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 826
 827 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 828                                void *priv, bool reserved)
 829 {
 830         /*
 831          * If we find a request that isn't idle and the queue matches,
 832          * we know the queue is busy. Return false to stop the iteration.
 833          */
 834         if (blk_mq_request_started(rq) && rq->q == hctx->queue) {
 835                 bool *busy = priv;
 836
 837                 *busy = true;
 838                 return false;
 839         }
 840
 841         return true;
 842 }
 843
 844 bool blk_mq_queue_inflight(struct request_queue *q)
 845 {
 846         bool busy = false;
 847
 848         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 849         return busy;
 850 }
 851 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 852
 853 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 854 {
 855         req->rq_flags |= RQF_TIMED_OUT;
 856         if (req->q->mq_ops->timeout) {
 857                 enum blk_eh_timer_return ret;
 858
 859                 ret = req->q->mq_ops->timeout(req, reserved);
 860                 if (ret == BLK_EH_DONE)
 861                         return;
 862                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 863         }
 864
 865         blk_add_timer(req);
 866 }
 867
 868 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 869 {
 870         unsigned long deadline;
 871
 872         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 873                 return false;
 874         if (rq->rq_flags & RQF_TIMED_OUT)
 875                 return false;
 876
 877         deadline = READ_ONCE(rq->deadline);
 878         if (time_after_eq(jiffies, deadline))
 879                 return true;
 880
 881         if (*next == 0)
 882                 *next = deadline;
 883         else if (time_after(*next, deadline))
 884                 *next = deadline;
 885         return false;
 886 }
 887
 888 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 889                 struct request *rq, void *priv, bool reserved)
 890 {
 891         unsigned long *next = priv;
 892
 893         /*
 894          * Just do a quick check if it is expired before locking the request in
 895          * so we're not unnecessarilly synchronizing across CPUs.
 896          */
 897         if (!blk_mq_req_expired(rq, next))
 898                 return true;
 899
 900         /*
 901          * We have reason to believe the request may be expired. Take a
 902          * reference on the request to lock this request lifetime into its
 903          * currently allocated context to prevent it from being reallocated in
 904          * the event the completion by-passes this timeout handler.
 905          *
 906          * If the reference was already released, then the driver beat the
 907          * timeout handler to posting a natural completion.
 908          */
 909         if (!refcount_inc_not_zero(&rq->ref))
 910                 return true;
 911
 912         /*
 913          * The request is now locked and cannot be reallocated underneath the
 914          * timeout handler's processing. Re-verify this exact request is truly
 915          * expired; if it is not expired, then the request was completed and
 916          * reallocated as a new request.
 917          */
 918         if (blk_mq_req_expired(rq, next))
 919                 blk_mq_rq_timed_out(rq, reserved);
 920
 921         if (is_flush_rq(rq, hctx))
 922                 rq->end_io(rq, 0);
 923         else if (refcount_dec_and_test(&rq->ref))
 924                 __blk_mq_free_request(rq);
 925
 926         return true;
 927 }
 928
 929 static void blk_mq_timeout_work(struct work_struct *work)
 930 {
 931         struct request_queue *q =
 932                 container_of(work, struct request_queue, timeout_work);
 933         unsigned long next = 0;
 934         struct blk_mq_hw_ctx *hctx;
 935         int i;
 936
 937         /* A deadlock might occur if a request is stuck requiring a
 938          * timeout at the same time a queue freeze is waiting
 939          * completion, since the timeout code would not be able to
 940          * acquire the queue reference here.
 941          *
 942          * That's why we don't use blk_queue_enter here; instead, we use
 943          * percpu_ref_tryget directly, because we need to be able to
 944          * obtain a reference even in the short window between the queue
 945          * starting to freeze, by dropping the first reference in
 946          * blk_freeze_queue_start, and the moment the last request is
 947          * consumed, marked by the instant q_usage_counter reaches
 948          * zero.
 949          */
 950         if (!percpu_ref_tryget(&q->q_usage_counter))
 951                 return;
 952
 953         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 954
 955         if (next != 0) {
 956                 mod_timer(&q->timeout, next);
 957         } else {
 958                 /*
 959                  * Request timeouts are handled as a forward rolling timer. If
 960                  * we end up here it means that no requests are pending and
 961                  * also that no request has been pending for a while. Mark
 962                  * each hctx as idle.
 963                  */
 964                 queue_for_each_hw_ctx(q, hctx, i) {
 965                         /* the hctx may be unmapped, so check it here */
 966                         if (blk_mq_hw_queue_mapped(hctx))
 967                                 blk_mq_tag_idle(hctx);
 968                 }
 969         }
 970         blk_queue_exit(q);
 971 }
 972
 973 struct flush_busy_ctx_data {
 974         struct blk_mq_hw_ctx *hctx;
 975         struct list_head *list;
 976 };
 977
 978 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 979 {
 980         struct flush_busy_ctx_data *flush_data = data;
 981         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 982         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 983         enum hctx_type type = hctx->type;
 984
 985         spin_lock(&ctx->lock);
 986         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 987         sbitmap_clear_bit(sb, bitnr);
 988         spin_unlock(&ctx->lock);
 989         return true;
 990 }
 991
 992 /*
 993  * Process software queues that have been marked busy, splicing them
 994  * to the for-dispatch
 995  */
 996 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 997 {
 998         struct flush_busy_ctx_data data = {
 999                 .hctx = hctx,
1000                 .list = list,
1001         };
1002
1003         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
1004 }
1005 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
1006
1007 struct dispatch_rq_data {
1008         struct blk_mq_hw_ctx *hctx;
1009         struct request *rq;
1010 };
1011
1012 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1013                 void *data)
1014 {
1015         struct dispatch_rq_data *dispatch_data = data;
1016         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1017         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1018         enum hctx_type type = hctx->type;
1019
1020         spin_lock(&ctx->lock);
1021         if (!list_empty(&ctx->rq_lists[type])) {
1022                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1023                 list_del_init(&dispatch_data->rq->queuelist);
1024                 if (list_empty(&ctx->rq_lists[type]))
1025                         sbitmap_clear_bit(sb, bitnr);
1026         }
1027         spin_unlock(&ctx->lock);
1028
1029         return !dispatch_data->rq;
1030 }
1031
1032 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1033                                         struct blk_mq_ctx *start)
1034 {
1035         unsigned off = start ? start->index_hw[hctx->type] : 0;
1036         struct dispatch_rq_data data = {
1037                 .hctx = hctx,
1038                 .rq   = NULL,
1039         };
1040
1041         __sbitmap_for_each_set(&hctx->ctx_map, off,
1042                                dispatch_rq_from_ctx, &data);
1043
1044         return data.rq;
1045 }
1046
1047 static inline unsigned int queued_to_index(unsigned int queued)
1048 {
1049         if (!queued)
1050                 return 0;
1051
1052         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1053 }
1054
1055 bool blk_mq_get_driver_tag(struct request *rq)
1056 {
1057         struct blk_mq_alloc_data data = {
1058                 .q = rq->q,
1059                 .hctx = rq->mq_hctx,
1060                 .flags = BLK_MQ_REQ_NOWAIT,
1061                 .cmd_flags = rq->cmd_flags,
1062         };
1063         bool shared;
1064
1065         if (rq->tag != -1)
1066                 goto done;
1067
1068         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1069                 data.flags |= BLK_MQ_REQ_RESERVED;
1070
1071         shared = blk_mq_tag_busy(data.hctx);
1072         rq->tag = blk_mq_get_tag(&data);
1073         if (rq->tag >= 0) {
1074                 if (shared) {
1075                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1076                         atomic_inc(&data.hctx->nr_active);
1077                 }
1078                 data.hctx->tags->rqs[rq->tag] = rq;
1079         }
1080
1081 done:
1082         return rq->tag != -1;
1083 }
1084
1085 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1086                                 int flags, void *key)
1087 {
1088         struct blk_mq_hw_ctx *hctx;
1089
1090         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1091
1092         spin_lock(&hctx->dispatch_wait_lock);
1093         if (!list_empty(&wait->entry)) {
1094                 struct sbitmap_queue *sbq;
1095
1096                 list_del_init(&wait->entry);
1097                 sbq = &hctx->tags->bitmap_tags;
1098                 atomic_dec(&sbq->ws_active);
1099         }
1100         spin_unlock(&hctx->dispatch_wait_lock);
1101
1102         blk_mq_run_hw_queue(hctx, true);
1103         return 1;
1104 }
1105
1106 /*
1107  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1108  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1109  * restart. For both cases, take care to check the condition again after
1110  * marking us as waiting.
1111  */
1112 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1113                                  struct request *rq)
1114 {
1115         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1116         struct wait_queue_head *wq;
1117         wait_queue_entry_t *wait;
1118         bool ret;
1119
1120         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1121                 blk_mq_sched_mark_restart_hctx(hctx);
1122
1123                 /*
1124                  * It's possible that a tag was freed in the window between the
1125                  * allocation failure and adding the hardware queue to the wait
1126                  * queue.
1127                  *
1128                  * Don't clear RESTART here, someone else could have set it.
1129                  * At most this will cost an extra queue run.
1130                  */
1131                 return blk_mq_get_driver_tag(rq);
1132         }
1133
1134         wait = &hctx->dispatch_wait;
1135         if (!list_empty_careful(&wait->entry))
1136                 return false;
1137
1138         wq = &bt_wait_ptr(sbq, hctx)->wait;
1139
1140         spin_lock_irq(&wq->lock);
1141         spin_lock(&hctx->dispatch_wait_lock);
1142         if (!list_empty(&wait->entry)) {
1143                 spin_unlock(&hctx->dispatch_wait_lock);
1144                 spin_unlock_irq(&wq->lock);
1145                 return false;
1146         }
1147
1148         atomic_inc(&sbq->ws_active);
1149         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1150         __add_wait_queue(wq, wait);
1151
1152         /*
1153          * It's possible that a tag was freed in the window between the
1154          * allocation failure and adding the hardware queue to the wait
1155          * queue.
1156          */
1157         ret = blk_mq_get_driver_tag(rq);
1158         if (!ret) {
1159                 spin_unlock(&hctx->dispatch_wait_lock);
1160                 spin_unlock_irq(&wq->lock);
1161                 return false;
1162         }
1163
1164         /*
1165          * We got a tag, remove ourselves from the wait queue to ensure
1166          * someone else gets the wakeup.
1167          */
1168         list_del_init(&wait->entry);
1169         atomic_dec(&sbq->ws_active);
1170         spin_unlock(&hctx->dispatch_wait_lock);
1171         spin_unlock_irq(&wq->lock);
1172
1173         return true;
1174 }
1175
1176 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1177 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1178 /*
1179  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1180  * - EWMA is one simple way to compute running average value
1181  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1182  * - take 4 as factor for avoiding to get too small(0) result, and this
1183  *   factor doesn't matter because EWMA decreases exponentially
1184  */
1185 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1186 {
1187         unsigned int ewma;
1188
1189         if (hctx->queue->elevator)
1190                 return;
1191
1192         ewma = hctx->dispatch_busy;
1193
1194         if (!ewma && !busy)
1195                 return;
1196
1197         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1198         if (busy)
1199                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1200         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1201
1202         hctx->dispatch_busy = ewma;
1203 }
1204
1205 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1206
1207 static void blk_mq_handle_dev_resource(struct request *rq,
1208                                        struct list_head *list)
1209 {
1210         struct request *next =
1211                 list_first_entry_or_null(list, struct request, queuelist);
1212
1213         /*
1214          * If an I/O scheduler has been configured and we got a driver tag for
1215          * the next request already, free it.
1216          */
1217         if (next)
1218                 blk_mq_put_driver_tag(next);
1219
1220         list_add(&rq->queuelist, list);
1221         __blk_mq_requeue_request(rq);
1222 }
1223
1224 /*
1225  * Returns true if we did some work AND can potentially do more.
1226  */
1227 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1228                              bool got_budget)
1229 {
1230         struct blk_mq_hw_ctx *hctx;
1231         struct request *rq, *nxt;
1232         bool no_tag = false;
1233         int errors, queued;
1234         blk_status_t ret = BLK_STS_OK;
1235         bool no_budget_avail = false;
1236
1237         if (list_empty(list))
1238                 return false;
1239
1240         WARN_ON(!list_is_singular(list) && got_budget);
1241
1242         /*
1243          * Now process all the entries, sending them to the driver.
1244          */
1245         errors = queued = 0;
1246         do {
1247                 struct blk_mq_queue_data bd;
1248
1249                 rq = list_first_entry(list, struct request, queuelist);
1250
1251                 hctx = rq->mq_hctx;
1252                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx)) {
1253                         blk_mq_put_driver_tag(rq);
1254                         no_budget_avail = true;
1255                         break;
1256                 }
1257
1258                 if (!blk_mq_get_driver_tag(rq)) {
1259                         /*
1260                          * The initial allocation attempt failed, so we need to
1261                          * rerun the hardware queue when a tag is freed. The
1262                          * waitqueue takes care of that. If the queue is run
1263                          * before we add this entry back on the dispatch list,
1264                          * we'll re-run it below.
1265                          */
1266                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1267                                 blk_mq_put_dispatch_budget(hctx);
1268                                 /*
1269                                  * For non-shared tags, the RESTART check
1270                                  * will suffice.
1271                                  */
1272                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1273                                         no_tag = true;
1274                                 break;
1275                         }
1276                 }
1277
1278                 list_del_init(&rq->queuelist);
1279
1280                 bd.rq = rq;
1281
1282                 /*
1283                  * Flag last if we have no more requests, or if we have more
1284                  * but can't assign a driver tag to it.
1285                  */
1286                 if (list_empty(list))
1287                         bd.last = true;
1288                 else {
1289                         nxt = list_first_entry(list, struct request, queuelist);
1290                         bd.last = !blk_mq_get_driver_tag(nxt);
1291                 }
1292
1293                 ret = q->mq_ops->queue_rq(hctx, &bd);
1294                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1295                         blk_mq_handle_dev_resource(rq, list);
1296                         break;
1297                 }
1298
1299                 if (unlikely(ret != BLK_STS_OK)) {
1300                         errors++;
1301                         blk_mq_end_request(rq, BLK_STS_IOERR);
1302                         continue;
1303                 }
1304
1305                 queued++;
1306         } while (!list_empty(list));
1307
1308         hctx->dispatched[queued_to_index(queued)]++;
1309
1310         /*
1311          * Any items that need requeuing? Stuff them into hctx->dispatch,
1312          * that is where we will continue on next queue run.
1313          */
1314         if (!list_empty(list)) {
1315                 bool needs_restart;
1316
1317                 /*
1318                  * If we didn't flush the entire list, we could have told
1319                  * the driver there was more coming, but that turned out to
1320                  * be a lie.
1321                  */
1322                 if (q->mq_ops->commit_rqs)
1323                         q->mq_ops->commit_rqs(hctx);
1324
1325                 spin_lock(&hctx->lock);
1326                 list_splice_tail_init(list, &hctx->dispatch);
1327                 spin_unlock(&hctx->lock);
1328
1329                 /*
1330                  * Order adding requests to hctx->dispatch and checking
1331                  * SCHED_RESTART flag. The pair of this smp_mb() is the one
1332                  * in blk_mq_sched_restart(). Avoid restart code path to
1333                  * miss the new added requests to hctx->dispatch, meantime
1334                  * SCHED_RESTART is observed here.
1335                  */
1336                 smp_mb();
1337
1338                 /*
1339                  * If SCHED_RESTART was set by the caller of this function and
1340                  * it is no longer set that means that it was cleared by another
1341                  * thread and hence that a queue rerun is needed.
1342                  *
1343                  * If 'no_tag' is set, that means that we failed getting
1344                  * a driver tag with an I/O scheduler attached. If our dispatch
1345                  * waitqueue is no longer active, ensure that we run the queue
1346                  * AFTER adding our entries back to the list.
1347                  *
1348                  * If no I/O scheduler has been configured it is possible that
1349                  * the hardware queue got stopped and restarted before requests
1350                  * were pushed back onto the dispatch list. Rerun the queue to
1351                  * avoid starvation. Notes:
1352                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1353                  *   been stopped before rerunning a queue.
1354                  * - Some but not all block drivers stop a queue before
1355                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1356                  *   and dm-rq.
1357                  *
1358                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1359                  * bit is set, run queue after a delay to avoid IO stalls
1360                  * that could otherwise occur if the queue is idle.  We'll do
1361                  * similar if we couldn't get budget and SCHED_RESTART is set.
1362                  */
1363                 needs_restart = blk_mq_sched_needs_restart(hctx);
1364                 if (!needs_restart ||
1365                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1366                         blk_mq_run_hw_queue(hctx, true);
1367                 else if (needs_restart && (ret == BLK_STS_RESOURCE ||
1368                                            no_budget_avail))
1369                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1370
1371                 blk_mq_update_dispatch_busy(hctx, true);
1372                 return false;
1373         } else
1374                 blk_mq_update_dispatch_busy(hctx, false);
1375
1376         /*
1377          * If the host/device is unable to accept more work, inform the
1378          * caller of that.
1379          */
1380         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1381                 return false;
1382
1383         return (queued + errors) != 0;
1384 }
1385
1386 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1387 {
1388         int srcu_idx;
1389
1390         /*
1391          * We should be running this queue from one of the CPUs that
1392          * are mapped to it.
1393          *
1394          * There are at least two related races now between setting
1395          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1396          * __blk_mq_run_hw_queue():
1397          *
1398          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1399          *   but later it becomes online, then this warning is harmless
1400          *   at all
1401          *
1402          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1403          *   but later it becomes offline, then the warning can't be
1404          *   triggered, and we depend on blk-mq timeout handler to
1405          *   handle dispatched requests to this hctx
1406          */
1407         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1408                 cpu_online(hctx->next_cpu)) {
1409                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1410                         raw_smp_processor_id(),
1411                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1412                 dump_stack();
1413         }
1414
1415         /*
1416          * We can't run the queue inline with ints disabled. Ensure that
1417          * we catch bad users of this early.
1418          */
1419         WARN_ON_ONCE(in_interrupt());
1420
1421         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1422
1423         hctx_lock(hctx, &srcu_idx);
1424         blk_mq_sched_dispatch_requests(hctx);
1425         hctx_unlock(hctx, srcu_idx);
1426 }
1427
1428 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1429 {
1430         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1431
1432         if (cpu >= nr_cpu_ids)
1433                 cpu = cpumask_first(hctx->cpumask);
1434         return cpu;
1435 }
1436
1437 /*
1438  * It'd be great if the workqueue API had a way to pass
1439  * in a mask and had some smarts for more clever placement.
1440  * For now we just round-robin here, switching for every
1441  * BLK_MQ_CPU_WORK_BATCH queued items.
1442  */
1443 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1444 {
1445         bool tried = false;
1446         int next_cpu = hctx->next_cpu;
1447
1448         if (hctx->queue->nr_hw_queues == 1)
1449                 return WORK_CPU_UNBOUND;
1450
1451         if (--hctx->next_cpu_batch <= 0) {
1452 select_cpu:
1453                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1454                                 cpu_online_mask);
1455                 if (next_cpu >= nr_cpu_ids)
1456                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1457                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1458         }
1459
1460         /*
1461          * Do unbound schedule if we can't find a online CPU for this hctx,
1462          * and it should only happen in the path of handling CPU DEAD.
1463          */
1464         if (!cpu_online(next_cpu)) {
1465                 if (!tried) {
1466                         tried = true;
1467                         goto select_cpu;
1468                 }
1469
1470                 /*
1471                  * Make sure to re-select CPU next time once after CPUs
1472                  * in hctx->cpumask become online again.
1473                  */
1474                 hctx->next_cpu = next_cpu;
1475                 hctx->next_cpu_batch = 1;
1476                 return WORK_CPU_UNBOUND;
1477         }
1478
1479         hctx->next_cpu = next_cpu;
1480         return next_cpu;
1481 }
1482
1483 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1484                                         unsigned long msecs)
1485 {
1486         if (unlikely(blk_mq_hctx_stopped(hctx)))
1487                 return;
1488
1489         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1490                 int cpu = get_cpu();
1491                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1492                         __blk_mq_run_hw_queue(hctx);
1493                         put_cpu();
1494                         return;
1495                 }
1496
1497                 put_cpu();
1498         }
1499
1500         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1501                                     msecs_to_jiffies(msecs));
1502 }
1503
1504 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1505 {
1506         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1507 }
1508 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1509
1510 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1511 {
1512         int srcu_idx;
1513         bool need_run;
1514
1515         /*
1516          * When queue is quiesced, we may be switching io scheduler, or
1517          * updating nr_hw_queues, or other things, and we can't run queue
1518          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1519          *
1520          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1521          * quiesced.
1522          */
1523         hctx_lock(hctx, &srcu_idx);
1524         need_run = !blk_queue_quiesced(hctx->queue) &&
1525                 blk_mq_hctx_has_pending(hctx);
1526         hctx_unlock(hctx, srcu_idx);
1527
1528         if (need_run) {
1529                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1530                 return true;
1531         }
1532
1533         return false;
1534 }
1535 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1536
1537 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
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                         continue;
1545
1546                 blk_mq_run_hw_queue(hctx, async);
1547         }
1548 }
1549 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1550
1551 /**
1552  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1553  * @q: request queue.
1554  *
1555  * The caller is responsible for serializing this function against
1556  * blk_mq_{start,stop}_hw_queue().
1557  */
1558 bool blk_mq_queue_stopped(struct request_queue *q)
1559 {
1560         struct blk_mq_hw_ctx *hctx;
1561         int i;
1562
1563         queue_for_each_hw_ctx(q, hctx, i)
1564                 if (blk_mq_hctx_stopped(hctx))
1565                         return true;
1566
1567         return false;
1568 }
1569 EXPORT_SYMBOL(blk_mq_queue_stopped);
1570
1571 /*
1572  * This function is often used for pausing .queue_rq() by driver when
1573  * there isn't enough resource or some conditions aren't satisfied, and
1574  * BLK_STS_RESOURCE is usually returned.
1575  *
1576  * We do not guarantee that dispatch can be drained or blocked
1577  * after blk_mq_stop_hw_queue() returns. Please use
1578  * blk_mq_quiesce_queue() for that requirement.
1579  */
1580 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1581 {
1582         cancel_delayed_work(&hctx->run_work);
1583
1584         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1585 }
1586 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1587
1588 /*
1589  * This function is often used for pausing .queue_rq() by driver when
1590  * there isn't enough resource or some conditions aren't satisfied, and
1591  * BLK_STS_RESOURCE is usually returned.
1592  *
1593  * We do not guarantee that dispatch can be drained or blocked
1594  * after blk_mq_stop_hw_queues() returns. Please use
1595  * blk_mq_quiesce_queue() for that requirement.
1596  */
1597 void blk_mq_stop_hw_queues(struct request_queue *q)
1598 {
1599         struct blk_mq_hw_ctx *hctx;
1600         int i;
1601
1602         queue_for_each_hw_ctx(q, hctx, i)
1603                 blk_mq_stop_hw_queue(hctx);
1604 }
1605 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1606
1607 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1608 {
1609         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1610
1611         blk_mq_run_hw_queue(hctx, false);
1612 }
1613 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1614
1615 void blk_mq_start_hw_queues(struct request_queue *q)
1616 {
1617         struct blk_mq_hw_ctx *hctx;
1618         int i;
1619
1620         queue_for_each_hw_ctx(q, hctx, i)
1621                 blk_mq_start_hw_queue(hctx);
1622 }
1623 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1624
1625 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1626 {
1627         if (!blk_mq_hctx_stopped(hctx))
1628                 return;
1629
1630         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1631         blk_mq_run_hw_queue(hctx, async);
1632 }
1633 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1634
1635 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1636 {
1637         struct blk_mq_hw_ctx *hctx;
1638         int i;
1639
1640         queue_for_each_hw_ctx(q, hctx, i)
1641                 blk_mq_start_stopped_hw_queue(hctx, async);
1642 }
1643 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1644
1645 static void blk_mq_run_work_fn(struct work_struct *work)
1646 {
1647         struct blk_mq_hw_ctx *hctx;
1648
1649         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1650
1651         /*
1652          * If we are stopped, don't run the queue.
1653          */
1654         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1655                 return;
1656
1657         __blk_mq_run_hw_queue(hctx);
1658 }
1659
1660 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1661                                             struct request *rq,
1662                                             bool at_head)
1663 {
1664         struct blk_mq_ctx *ctx = rq->mq_ctx;
1665         enum hctx_type type = hctx->type;
1666
1667         lockdep_assert_held(&ctx->lock);
1668
1669         trace_block_rq_insert(hctx->queue, rq);
1670
1671         if (at_head)
1672                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1673         else
1674                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1675 }
1676
1677 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1678                              bool at_head)
1679 {
1680         struct blk_mq_ctx *ctx = rq->mq_ctx;
1681
1682         lockdep_assert_held(&ctx->lock);
1683
1684         __blk_mq_insert_req_list(hctx, rq, at_head);
1685         blk_mq_hctx_mark_pending(hctx, ctx);
1686 }
1687
1688 /*
1689  * Should only be used carefully, when the caller knows we want to
1690  * bypass a potential IO scheduler on the target device.
1691  */
1692 void blk_mq_request_bypass_insert(struct request *rq, bool at_head,
1693                                   bool run_queue)
1694 {
1695         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1696
1697         spin_lock(&hctx->lock);
1698         if (at_head)
1699                 list_add(&rq->queuelist, &hctx->dispatch);
1700         else
1701                 list_add_tail(&rq->queuelist, &hctx->dispatch);
1702         spin_unlock(&hctx->lock);
1703
1704         if (run_queue)
1705                 blk_mq_run_hw_queue(hctx, false);
1706 }
1707
1708 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1709                             struct list_head *list)
1710
1711 {
1712         struct request *rq;
1713         enum hctx_type type = hctx->type;
1714
1715         /*
1716          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1717          * offline now
1718          */
1719         list_for_each_entry(rq, list, queuelist) {
1720                 BUG_ON(rq->mq_ctx != ctx);
1721                 trace_block_rq_insert(hctx->queue, rq);
1722         }
1723
1724         spin_lock(&ctx->lock);
1725         list_splice_tail_init(list, &ctx->rq_lists[type]);
1726         blk_mq_hctx_mark_pending(hctx, ctx);
1727         spin_unlock(&ctx->lock);
1728 }
1729
1730 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1731 {
1732         struct request *rqa = container_of(a, struct request, queuelist);
1733         struct request *rqb = container_of(b, struct request, queuelist);
1734
1735         if (rqa->mq_ctx < rqb->mq_ctx)
1736                 return -1;
1737         else if (rqa->mq_ctx > rqb->mq_ctx)
1738                 return 1;
1739         else if (rqa->mq_hctx < rqb->mq_hctx)
1740                 return -1;
1741         else if (rqa->mq_hctx > rqb->mq_hctx)
1742                 return 1;
1743
1744         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1745 }
1746
1747 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1748 {
1749         struct blk_mq_hw_ctx *this_hctx;
1750         struct blk_mq_ctx *this_ctx;
1751         struct request_queue *this_q;
1752         struct request *rq;
1753         LIST_HEAD(list);
1754         LIST_HEAD(rq_list);
1755         unsigned int depth;
1756
1757         list_splice_init(&plug->mq_list, &list);
1758
1759         if (plug->rq_count > 2 && plug->multiple_queues)
1760                 list_sort(NULL, &list, plug_rq_cmp);
1761
1762         plug->rq_count = 0;
1763
1764         this_q = NULL;
1765         this_hctx = NULL;
1766         this_ctx = NULL;
1767         depth = 0;
1768
1769         while (!list_empty(&list)) {
1770                 rq = list_entry_rq(list.next);
1771                 list_del_init(&rq->queuelist);
1772                 BUG_ON(!rq->q);
1773                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1774                         if (this_hctx) {
1775                                 trace_block_unplug(this_q, depth, !from_schedule);
1776                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1777                                                                 &rq_list,
1778                                                                 from_schedule);
1779                         }
1780
1781                         this_q = rq->q;
1782                         this_ctx = rq->mq_ctx;
1783                         this_hctx = rq->mq_hctx;
1784                         depth = 0;
1785                 }
1786
1787                 depth++;
1788                 list_add_tail(&rq->queuelist, &rq_list);
1789         }
1790
1791         /*
1792          * If 'this_hctx' is set, we know we have entries to complete
1793          * on 'rq_list'. Do those.
1794          */
1795         if (this_hctx) {
1796                 trace_block_unplug(this_q, depth, !from_schedule);
1797                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1798                                                 from_schedule);
1799         }
1800 }
1801
1802 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1803                 unsigned int nr_segs)
1804 {
1805         if (bio->bi_opf & REQ_RAHEAD)
1806                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1807
1808         rq->__sector = bio->bi_iter.bi_sector;
1809         rq->write_hint = bio->bi_write_hint;
1810         blk_rq_bio_prep(rq, bio, nr_segs);
1811
1812         blk_account_io_start(rq, true);
1813 }
1814
1815 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1816                                             struct request *rq,
1817                                             blk_qc_t *cookie, bool last)
1818 {
1819         struct request_queue *q = rq->q;
1820         struct blk_mq_queue_data bd = {
1821                 .rq = rq,
1822                 .last = last,
1823         };
1824         blk_qc_t new_cookie;
1825         blk_status_t ret;
1826
1827         new_cookie = request_to_qc_t(hctx, rq);
1828
1829         /*
1830          * For OK queue, we are done. For error, caller may kill it.
1831          * Any other error (busy), just add it to our list as we
1832          * previously would have done.
1833          */
1834         ret = q->mq_ops->queue_rq(hctx, &bd);
1835         switch (ret) {
1836         case BLK_STS_OK:
1837                 blk_mq_update_dispatch_busy(hctx, false);
1838                 *cookie = new_cookie;
1839                 break;
1840         case BLK_STS_RESOURCE:
1841         case BLK_STS_DEV_RESOURCE:
1842                 blk_mq_update_dispatch_busy(hctx, true);
1843                 __blk_mq_requeue_request(rq);
1844                 break;
1845         default:
1846                 blk_mq_update_dispatch_busy(hctx, false);
1847                 *cookie = BLK_QC_T_NONE;
1848                 break;
1849         }
1850
1851         return ret;
1852 }
1853
1854 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1855                                                 struct request *rq,
1856                                                 blk_qc_t *cookie,
1857                                                 bool bypass_insert, bool last)
1858 {
1859         struct request_queue *q = rq->q;
1860         bool run_queue = true;
1861
1862         /*
1863          * RCU or SRCU read lock is needed before checking quiesced flag.
1864          *
1865          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1866          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1867          * and avoid driver to try to dispatch again.
1868          */
1869         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1870                 run_queue = false;
1871                 bypass_insert = false;
1872                 goto insert;
1873         }
1874
1875         if (q->elevator && !bypass_insert)
1876                 goto insert;
1877
1878         if (!blk_mq_get_dispatch_budget(hctx))
1879                 goto insert;
1880
1881         if (!blk_mq_get_driver_tag(rq)) {
1882                 blk_mq_put_dispatch_budget(hctx);
1883                 goto insert;
1884         }
1885
1886         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1887 insert:
1888         if (bypass_insert)
1889                 return BLK_STS_RESOURCE;
1890
1891         blk_mq_sched_insert_request(rq, false, run_queue, false);
1892
1893         return BLK_STS_OK;
1894 }
1895
1896 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1897                 struct request *rq, blk_qc_t *cookie)
1898 {
1899         blk_status_t ret;
1900         int srcu_idx;
1901
1902         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1903
1904         hctx_lock(hctx, &srcu_idx);
1905
1906         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1907         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1908                 blk_mq_request_bypass_insert(rq, false, true);
1909         else if (ret != BLK_STS_OK)
1910                 blk_mq_end_request(rq, ret);
1911
1912         hctx_unlock(hctx, srcu_idx);
1913 }
1914
1915 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1916 {
1917         blk_status_t ret;
1918         int srcu_idx;
1919         blk_qc_t unused_cookie;
1920         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1921
1922         hctx_lock(hctx, &srcu_idx);
1923         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1924         hctx_unlock(hctx, srcu_idx);
1925
1926         return ret;
1927 }
1928
1929 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1930                 struct list_head *list)
1931 {
1932         while (!list_empty(list)) {
1933                 blk_status_t ret;
1934                 struct request *rq = list_first_entry(list, struct request,
1935                                 queuelist);
1936
1937                 list_del_init(&rq->queuelist);
1938                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1939                 if (ret != BLK_STS_OK) {
1940                         if (ret == BLK_STS_RESOURCE ||
1941                                         ret == BLK_STS_DEV_RESOURCE) {
1942                                 blk_mq_request_bypass_insert(rq, false,
1943                                                         list_empty(list));
1944                                 break;
1945                         }
1946                         blk_mq_end_request(rq, ret);
1947                 }
1948         }
1949
1950         /*
1951          * If we didn't flush the entire list, we could have told
1952          * the driver there was more coming, but that turned out to
1953          * be a lie.
1954          */
1955         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1956                 hctx->queue->mq_ops->commit_rqs(hctx);
1957 }
1958
1959 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1960 {
1961         list_add_tail(&rq->queuelist, &plug->mq_list);
1962         plug->rq_count++;
1963         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1964                 struct request *tmp;
1965
1966                 tmp = list_first_entry(&plug->mq_list, struct request,
1967                                                 queuelist);
1968                 if (tmp->q != rq->q)
1969                         plug->multiple_queues = true;
1970         }
1971 }
1972
1973 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1974 {
1975         const int is_sync = op_is_sync(bio->bi_opf);
1976         const int is_flush_fua = op_is_flush(bio->bi_opf);
1977         struct blk_mq_alloc_data data = { .flags = 0};
1978         struct request *rq;
1979         struct blk_plug *plug;
1980         struct request *same_queue_rq = NULL;
1981         unsigned int nr_segs;
1982         blk_qc_t cookie;
1983
1984         blk_queue_bounce(q, &bio);
1985         __blk_queue_split(q, &bio, &nr_segs);
1986
1987         if (!bio_integrity_prep(bio))
1988                 return BLK_QC_T_NONE;
1989
1990         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1991             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1992                 return BLK_QC_T_NONE;
1993
1994         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1995                 return BLK_QC_T_NONE;
1996
1997         rq_qos_throttle(q, bio);
1998
1999         data.cmd_flags = bio->bi_opf;
2000         rq = blk_mq_get_request(q, bio, &data);
2001         if (unlikely(!rq)) {
2002                 rq_qos_cleanup(q, bio);
2003                 if (bio->bi_opf & REQ_NOWAIT)
2004                         bio_wouldblock_error(bio);
2005                 return BLK_QC_T_NONE;
2006         }
2007
2008         trace_block_getrq(q, bio, bio->bi_opf);
2009
2010         rq_qos_track(q, rq, bio);
2011
2012         cookie = request_to_qc_t(data.hctx, rq);
2013
2014         blk_mq_bio_to_request(rq, bio, nr_segs);
2015
2016         plug = blk_mq_plug(q, bio);
2017         if (unlikely(is_flush_fua)) {
2018                 /* bypass scheduler for flush rq */
2019                 blk_insert_flush(rq);
2020                 blk_mq_run_hw_queue(data.hctx, true);
2021         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs ||
2022                                 !blk_queue_nonrot(q))) {
2023                 /*
2024                  * Use plugging if we have a ->commit_rqs() hook as well, as
2025                  * we know the driver uses bd->last in a smart fashion.
2026                  *
2027                  * Use normal plugging if this disk is slow HDD, as sequential
2028                  * IO may benefit a lot from plug merging.
2029                  */
2030                 unsigned int request_count = plug->rq_count;
2031                 struct request *last = NULL;
2032
2033                 if (!request_count)
2034                         trace_block_plug(q);
2035                 else
2036                         last = list_entry_rq(plug->mq_list.prev);
2037
2038                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2039                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2040                         blk_flush_plug_list(plug, false);
2041                         trace_block_plug(q);
2042                 }
2043
2044                 blk_add_rq_to_plug(plug, rq);
2045         } else if (q->elevator) {
2046                 blk_mq_sched_insert_request(rq, false, true, true);
2047         } else if (plug && !blk_queue_nomerges(q)) {
2048                 /*
2049                  * We do limited plugging. If the bio can be merged, do that.
2050                  * Otherwise the existing request in the plug list will be
2051                  * issued. So the plug list will have one request at most
2052                  * The plug list might get flushed before this. If that happens,
2053                  * the plug list is empty, and same_queue_rq is invalid.
2054                  */
2055                 if (list_empty(&plug->mq_list))
2056                         same_queue_rq = NULL;
2057                 if (same_queue_rq) {
2058                         list_del_init(&same_queue_rq->queuelist);
2059                         plug->rq_count--;
2060                 }
2061                 blk_add_rq_to_plug(plug, rq);
2062                 trace_block_plug(q);
2063
2064                 if (same_queue_rq) {
2065                         data.hctx = same_queue_rq->mq_hctx;
2066                         trace_block_unplug(q, 1, true);
2067                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2068                                         &cookie);
2069                 }
2070         } else if ((q->nr_hw_queues > 1 && is_sync) ||
2071                         !data.hctx->dispatch_busy) {
2072                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2073         } else {
2074                 blk_mq_sched_insert_request(rq, false, true, true);
2075         }
2076
2077         return cookie;
2078 }
2079
2080 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2081                      unsigned int hctx_idx)
2082 {
2083         struct page *page;
2084
2085         if (tags->rqs && set->ops->exit_request) {
2086                 int i;
2087
2088                 for (i = 0; i < tags->nr_tags; i++) {
2089                         struct request *rq = tags->static_rqs[i];
2090
2091                         if (!rq)
2092                                 continue;
2093                         set->ops->exit_request(set, rq, hctx_idx);
2094                         tags->static_rqs[i] = NULL;
2095                 }
2096         }
2097
2098         while (!list_empty(&tags->page_list)) {
2099                 page = list_first_entry(&tags->page_list, struct page, lru);
2100                 list_del_init(&page->lru);
2101                 /*
2102                  * Remove kmemleak object previously allocated in
2103                  * blk_mq_alloc_rqs().
2104                  */
2105                 kmemleak_free(page_address(page));
2106                 __free_pages(page, page->private);
2107         }
2108 }
2109
2110 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2111 {
2112         kfree(tags->rqs);
2113         tags->rqs = NULL;
2114         kfree(tags->static_rqs);
2115         tags->static_rqs = NULL;
2116
2117         blk_mq_free_tags(tags);
2118 }
2119
2120 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2121                                         unsigned int hctx_idx,
2122                                         unsigned int nr_tags,
2123                                         unsigned int reserved_tags)
2124 {
2125         struct blk_mq_tags *tags;
2126         int node;
2127
2128         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2129         if (node == NUMA_NO_NODE)
2130                 node = set->numa_node;
2131
2132         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2133                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2134         if (!tags)
2135                 return NULL;
2136
2137         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2138                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2139                                  node);
2140         if (!tags->rqs) {
2141                 blk_mq_free_tags(tags);
2142                 return NULL;
2143         }
2144
2145         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2146                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2147                                         node);
2148         if (!tags->static_rqs) {
2149                 kfree(tags->rqs);
2150                 blk_mq_free_tags(tags);
2151                 return NULL;
2152         }
2153
2154         return tags;
2155 }
2156
2157 static size_t order_to_size(unsigned int order)
2158 {
2159         return (size_t)PAGE_SIZE << order;
2160 }
2161
2162 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2163                                unsigned int hctx_idx, int node)
2164 {
2165         int ret;
2166
2167         if (set->ops->init_request) {
2168                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2169                 if (ret)
2170                         return ret;
2171         }
2172
2173         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2174         return 0;
2175 }
2176
2177 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2178                      unsigned int hctx_idx, unsigned int depth)
2179 {
2180         unsigned int i, j, entries_per_page, max_order = 4;
2181         size_t rq_size, left;
2182         int node;
2183
2184         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2185         if (node == NUMA_NO_NODE)
2186                 node = set->numa_node;
2187
2188         INIT_LIST_HEAD(&tags->page_list);
2189
2190         /*
2191          * rq_size is the size of the request plus driver payload, rounded
2192          * to the cacheline size
2193          */
2194         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2195                                 cache_line_size());
2196         left = rq_size * depth;
2197
2198         for (i = 0; i < depth; ) {
2199                 int this_order = max_order;
2200                 struct page *page;
2201                 int to_do;
2202                 void *p;
2203
2204                 while (this_order && left < order_to_size(this_order - 1))
2205                         this_order--;
2206
2207                 do {
2208                         page = alloc_pages_node(node,
2209                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2210                                 this_order);
2211                         if (page)
2212                                 break;
2213                         if (!this_order--)
2214                                 break;
2215                         if (order_to_size(this_order) < rq_size)
2216                                 break;
2217                 } while (1);
2218
2219                 if (!page)
2220                         goto fail;
2221
2222                 page->private = this_order;
2223                 list_add_tail(&page->lru, &tags->page_list);
2224
2225                 p = page_address(page);
2226                 /*
2227                  * Allow kmemleak to scan these pages as they contain pointers
2228                  * to additional allocations like via ops->init_request().
2229                  */
2230                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2231                 entries_per_page = order_to_size(this_order) / rq_size;
2232                 to_do = min(entries_per_page, depth - i);
2233                 left -= to_do * rq_size;
2234                 for (j = 0; j < to_do; j++) {
2235                         struct request *rq = p;
2236
2237                         tags->static_rqs[i] = rq;
2238                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2239                                 tags->static_rqs[i] = NULL;
2240                                 goto fail;
2241                         }
2242
2243                         p += rq_size;
2244                         i++;
2245                 }
2246         }
2247         return 0;
2248
2249 fail:
2250         blk_mq_free_rqs(set, tags, hctx_idx);
2251         return -ENOMEM;
2252 }
2253
2254 /*
2255  * 'cpu' is going away. splice any existing rq_list entries from this
2256  * software queue to the hw queue dispatch list, and ensure that it
2257  * gets run.
2258  */
2259 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2260 {
2261         struct blk_mq_hw_ctx *hctx;
2262         struct blk_mq_ctx *ctx;
2263         LIST_HEAD(tmp);
2264         enum hctx_type type;
2265
2266         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2267         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2268         type = hctx->type;
2269
2270         spin_lock(&ctx->lock);
2271         if (!list_empty(&ctx->rq_lists[type])) {
2272                 list_splice_init(&ctx->rq_lists[type], &tmp);
2273                 blk_mq_hctx_clear_pending(hctx, ctx);
2274         }
2275         spin_unlock(&ctx->lock);
2276
2277         if (list_empty(&tmp))
2278                 return 0;
2279
2280         spin_lock(&hctx->lock);
2281         list_splice_tail_init(&tmp, &hctx->dispatch);
2282         spin_unlock(&hctx->lock);
2283
2284         blk_mq_run_hw_queue(hctx, true);
2285         return 0;
2286 }
2287
2288 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2289 {
2290         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2291                                             &hctx->cpuhp_dead);
2292 }
2293
2294 /* hctx->ctxs will be freed in queue's release handler */
2295 static void blk_mq_exit_hctx(struct request_queue *q,
2296                 struct blk_mq_tag_set *set,
2297                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2298 {
2299         if (blk_mq_hw_queue_mapped(hctx))
2300                 blk_mq_tag_idle(hctx);
2301
2302         if (set->ops->exit_request)
2303                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2304
2305         if (set->ops->exit_hctx)
2306                 set->ops->exit_hctx(hctx, hctx_idx);
2307
2308         blk_mq_remove_cpuhp(hctx);
2309
2310         spin_lock(&q->unused_hctx_lock);
2311         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2312         spin_unlock(&q->unused_hctx_lock);
2313 }
2314
2315 static void blk_mq_exit_hw_queues(struct request_queue *q,
2316                 struct blk_mq_tag_set *set, int nr_queue)
2317 {
2318         struct blk_mq_hw_ctx *hctx;
2319         unsigned int i;
2320
2321         queue_for_each_hw_ctx(q, hctx, i) {
2322                 if (i == nr_queue)
2323                         break;
2324                 blk_mq_debugfs_unregister_hctx(hctx);
2325                 blk_mq_exit_hctx(q, set, hctx, i);
2326         }
2327 }
2328
2329 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2330 {
2331         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2332
2333         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2334                            __alignof__(struct blk_mq_hw_ctx)) !=
2335                      sizeof(struct blk_mq_hw_ctx));
2336
2337         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2338                 hw_ctx_size += sizeof(struct srcu_struct);
2339
2340         return hw_ctx_size;
2341 }
2342
2343 static int blk_mq_init_hctx(struct request_queue *q,
2344                 struct blk_mq_tag_set *set,
2345                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2346 {
2347         hctx->queue_num = hctx_idx;
2348
2349         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2350
2351         hctx->tags = set->tags[hctx_idx];
2352
2353         if (set->ops->init_hctx &&
2354             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2355                 goto unregister_cpu_notifier;
2356
2357         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2358                                 hctx->numa_node))
2359                 goto exit_hctx;
2360         return 0;
2361
2362  exit_hctx:
2363         if (set->ops->exit_hctx)
2364                 set->ops->exit_hctx(hctx, hctx_idx);
2365  unregister_cpu_notifier:
2366         blk_mq_remove_cpuhp(hctx);
2367         return -1;
2368 }
2369
2370 static struct blk_mq_hw_ctx *
2371 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2372                 int node)
2373 {
2374         struct blk_mq_hw_ctx *hctx;
2375         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2376
2377         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2378         if (!hctx)
2379                 goto fail_alloc_hctx;
2380
2381         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2382                 goto free_hctx;
2383
2384         atomic_set(&hctx->nr_active, 0);
2385         if (node == NUMA_NO_NODE)
2386                 node = set->numa_node;
2387         hctx->numa_node = node;
2388
2389         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2390         spin_lock_init(&hctx->lock);
2391         INIT_LIST_HEAD(&hctx->dispatch);
2392         hctx->queue = q;
2393         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2394
2395         INIT_LIST_HEAD(&hctx->hctx_list);
2396
2397         /*
2398          * Allocate space for all possible cpus to avoid allocation at
2399          * runtime
2400          */
2401         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2402                         gfp, node);
2403         if (!hctx->ctxs)
2404                 goto free_cpumask;
2405
2406         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2407                                 gfp, node))
2408                 goto free_ctxs;
2409         hctx->nr_ctx = 0;
2410
2411         spin_lock_init(&hctx->dispatch_wait_lock);
2412         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2413         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2414
2415         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2416                         gfp);
2417         if (!hctx->fq)
2418                 goto free_bitmap;
2419
2420         if (hctx->flags & BLK_MQ_F_BLOCKING)
2421                 init_srcu_struct(hctx->srcu);
2422         blk_mq_hctx_kobj_init(hctx);
2423
2424         return hctx;
2425
2426  free_bitmap:
2427         sbitmap_free(&hctx->ctx_map);
2428  free_ctxs:
2429         kfree(hctx->ctxs);
2430  free_cpumask:
2431         free_cpumask_var(hctx->cpumask);
2432  free_hctx:
2433         kfree(hctx);
2434  fail_alloc_hctx:
2435         return NULL;
2436 }
2437
2438 static void blk_mq_init_cpu_queues(struct request_queue *q,
2439                                    unsigned int nr_hw_queues)
2440 {
2441         struct blk_mq_tag_set *set = q->tag_set;
2442         unsigned int i, j;
2443
2444         for_each_possible_cpu(i) {
2445                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2446                 struct blk_mq_hw_ctx *hctx;
2447                 int k;
2448
2449                 __ctx->cpu = i;
2450                 spin_lock_init(&__ctx->lock);
2451                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2452                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2453
2454                 __ctx->queue = q;
2455
2456                 /*
2457                  * Set local node, IFF we have more than one hw queue. If
2458                  * not, we remain on the home node of the device
2459                  */
2460                 for (j = 0; j < set->nr_maps; j++) {
2461                         hctx = blk_mq_map_queue_type(q, j, i);
2462                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2463                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2464                 }
2465         }
2466 }
2467
2468 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2469 {
2470         int ret = 0;
2471
2472         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2473                                         set->queue_depth, set->reserved_tags);
2474         if (!set->tags[hctx_idx])
2475                 return false;
2476
2477         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2478                                 set->queue_depth);
2479         if (!ret)
2480                 return true;
2481
2482         blk_mq_free_rq_map(set->tags[hctx_idx]);
2483         set->tags[hctx_idx] = NULL;
2484         return false;
2485 }
2486
2487 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2488                                          unsigned int hctx_idx)
2489 {
2490         if (set->tags && set->tags[hctx_idx]) {
2491                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2492                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2493                 set->tags[hctx_idx] = NULL;
2494         }
2495 }
2496
2497 static void blk_mq_map_swqueue(struct request_queue *q)
2498 {
2499         unsigned int i, j, hctx_idx;
2500         struct blk_mq_hw_ctx *hctx;
2501         struct blk_mq_ctx *ctx;
2502         struct blk_mq_tag_set *set = q->tag_set;
2503
2504         queue_for_each_hw_ctx(q, hctx, i) {
2505                 cpumask_clear(hctx->cpumask);
2506                 hctx->nr_ctx = 0;
2507                 hctx->dispatch_from = NULL;
2508         }
2509
2510         /*
2511          * Map software to hardware queues.
2512          *
2513          * If the cpu isn't present, the cpu is mapped to first hctx.
2514          */
2515         for_each_possible_cpu(i) {
2516
2517                 ctx = per_cpu_ptr(q->queue_ctx, i);
2518                 for (j = 0; j < set->nr_maps; j++) {
2519                         if (!set->map[j].nr_queues) {
2520                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2521                                                 HCTX_TYPE_DEFAULT, i);
2522                                 continue;
2523                         }
2524                         hctx_idx = set->map[j].mq_map[i];
2525                         /* unmapped hw queue can be remapped after CPU topo changed */
2526                         if (!set->tags[hctx_idx] &&
2527                             !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2528                                 /*
2529                                  * If tags initialization fail for some hctx,
2530                                  * that hctx won't be brought online.  In this
2531                                  * case, remap the current ctx to hctx[0] which
2532                                  * is guaranteed to always have tags allocated
2533                                  */
2534                                 set->map[j].mq_map[i] = 0;
2535                         }
2536
2537                         hctx = blk_mq_map_queue_type(q, j, i);
2538                         ctx->hctxs[j] = hctx;
2539                         /*
2540                          * If the CPU is already set in the mask, then we've
2541                          * mapped this one already. This can happen if
2542                          * devices share queues across queue maps.
2543                          */
2544                         if (cpumask_test_cpu(i, hctx->cpumask))
2545                                 continue;
2546
2547                         cpumask_set_cpu(i, hctx->cpumask);
2548                         hctx->type = j;
2549                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2550                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2551
2552                         /*
2553                          * If the nr_ctx type overflows, we have exceeded the
2554                          * amount of sw queues we can support.
2555                          */
2556                         BUG_ON(!hctx->nr_ctx);
2557                 }
2558
2559                 for (; j < HCTX_MAX_TYPES; j++)
2560                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2561                                         HCTX_TYPE_DEFAULT, i);
2562         }
2563
2564         queue_for_each_hw_ctx(q, hctx, i) {
2565                 /*
2566                  * If no software queues are mapped to this hardware queue,
2567                  * disable it and free the request entries.
2568                  */
2569                 if (!hctx->nr_ctx) {
2570                         /* Never unmap queue 0.  We need it as a
2571                          * fallback in case of a new remap fails
2572                          * allocation
2573                          */
2574                         if (i && set->tags[i])
2575                                 blk_mq_free_map_and_requests(set, i);
2576
2577                         hctx->tags = NULL;
2578                         continue;
2579                 }
2580
2581                 hctx->tags = set->tags[i];
2582                 WARN_ON(!hctx->tags);
2583
2584                 /*
2585                  * Set the map size to the number of mapped software queues.
2586                  * This is more accurate and more efficient than looping
2587                  * over all possibly mapped software queues.
2588                  */
2589                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2590
2591                 /*
2592                  * Initialize batch roundrobin counts
2593                  */
2594                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2595                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2596         }
2597 }
2598
2599 /*
2600  * Caller needs to ensure that we're either frozen/quiesced, or that
2601  * the queue isn't live yet.
2602  */
2603 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2604 {
2605         struct blk_mq_hw_ctx *hctx;
2606         int i;
2607
2608         queue_for_each_hw_ctx(q, hctx, i) {
2609                 if (shared)
2610                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2611                 else
2612                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2613         }
2614 }
2615
2616 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2617                                         bool shared)
2618 {
2619         struct request_queue *q;
2620
2621         lockdep_assert_held(&set->tag_list_lock);
2622
2623         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2624                 blk_mq_freeze_queue(q);
2625                 queue_set_hctx_shared(q, shared);
2626                 blk_mq_unfreeze_queue(q);
2627         }
2628 }
2629
2630 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2631 {
2632         struct blk_mq_tag_set *set = q->tag_set;
2633
2634         mutex_lock(&set->tag_list_lock);
2635         list_del_rcu(&q->tag_set_list);
2636         if (list_is_singular(&set->tag_list)) {
2637                 /* just transitioned to unshared */
2638                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2639                 /* update existing queue */
2640                 blk_mq_update_tag_set_depth(set, false);
2641         }
2642         mutex_unlock(&set->tag_list_lock);
2643         INIT_LIST_HEAD(&q->tag_set_list);
2644 }
2645
2646 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2647                                      struct request_queue *q)
2648 {
2649         mutex_lock(&set->tag_list_lock);
2650
2651         /*
2652          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2653          */
2654         if (!list_empty(&set->tag_list) &&
2655             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2656                 set->flags |= BLK_MQ_F_TAG_SHARED;
2657                 /* update existing queue */
2658                 blk_mq_update_tag_set_depth(set, true);
2659         }
2660         if (set->flags & BLK_MQ_F_TAG_SHARED)
2661                 queue_set_hctx_shared(q, true);
2662         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2663
2664         mutex_unlock(&set->tag_list_lock);
2665 }
2666
2667 /* All allocations will be freed in release handler of q->mq_kobj */
2668 static int blk_mq_alloc_ctxs(struct request_queue *q)
2669 {
2670         struct blk_mq_ctxs *ctxs;
2671         int cpu;
2672
2673         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2674         if (!ctxs)
2675                 return -ENOMEM;
2676
2677         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2678         if (!ctxs->queue_ctx)
2679                 goto fail;
2680
2681         for_each_possible_cpu(cpu) {
2682                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2683                 ctx->ctxs = ctxs;
2684         }
2685
2686         q->mq_kobj = &ctxs->kobj;
2687         q->queue_ctx = ctxs->queue_ctx;
2688
2689         return 0;
2690  fail:
2691         kfree(ctxs);
2692         return -ENOMEM;
2693 }
2694
2695 /*
2696  * It is the actual release handler for mq, but we do it from
2697  * request queue's release handler for avoiding use-after-free
2698  * and headache because q->mq_kobj shouldn't have been introduced,
2699  * but we can't group ctx/kctx kobj without it.
2700  */
2701 void blk_mq_release(struct request_queue *q)
2702 {
2703         struct blk_mq_hw_ctx *hctx, *next;
2704         int i;
2705
2706         queue_for_each_hw_ctx(q, hctx, i)
2707                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2708
2709         /* all hctx are in .unused_hctx_list now */
2710         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2711                 list_del_init(&hctx->hctx_list);
2712                 kobject_put(&hctx->kobj);
2713         }
2714
2715         kfree(q->queue_hw_ctx);
2716
2717         /*
2718          * release .mq_kobj and sw queue's kobject now because
2719          * both share lifetime with request queue.
2720          */
2721         blk_mq_sysfs_deinit(q);
2722 }
2723
2724 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2725 {
2726         struct request_queue *uninit_q, *q;
2727
2728         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2729         if (!uninit_q)
2730                 return ERR_PTR(-ENOMEM);
2731
2732         /*
2733          * Initialize the queue without an elevator. device_add_disk() will do
2734          * the initialization.
2735          */
2736         q = blk_mq_init_allocated_queue(set, uninit_q, false);
2737         if (IS_ERR(q))
2738                 blk_cleanup_queue(uninit_q);
2739
2740         return q;
2741 }
2742 EXPORT_SYMBOL(blk_mq_init_queue);
2743
2744 /*
2745  * Helper for setting up a queue with mq ops, given queue depth, and
2746  * the passed in mq ops flags.
2747  */
2748 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2749                                            const struct blk_mq_ops *ops,
2750                                            unsigned int queue_depth,
2751                                            unsigned int set_flags)
2752 {
2753         struct request_queue *q;
2754         int ret;
2755
2756         memset(set, 0, sizeof(*set));
2757         set->ops = ops;
2758         set->nr_hw_queues = 1;
2759         set->nr_maps = 1;
2760         set->queue_depth = queue_depth;
2761         set->numa_node = NUMA_NO_NODE;
2762         set->flags = set_flags;
2763
2764         ret = blk_mq_alloc_tag_set(set);
2765         if (ret)
2766                 return ERR_PTR(ret);
2767
2768         q = blk_mq_init_queue(set);
2769         if (IS_ERR(q)) {
2770                 blk_mq_free_tag_set(set);
2771                 return q;
2772         }
2773
2774         return q;
2775 }
2776 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2777
2778 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2779                 struct blk_mq_tag_set *set, struct request_queue *q,
2780                 int hctx_idx, int node)
2781 {
2782         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2783
2784         /* reuse dead hctx first */
2785         spin_lock(&q->unused_hctx_lock);
2786         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2787                 if (tmp->numa_node == node) {
2788                         hctx = tmp;
2789                         break;
2790                 }
2791         }
2792         if (hctx)
2793                 list_del_init(&hctx->hctx_list);
2794         spin_unlock(&q->unused_hctx_lock);
2795
2796         if (!hctx)
2797                 hctx = blk_mq_alloc_hctx(q, set, node);
2798         if (!hctx)
2799                 goto fail;
2800
2801         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2802                 goto free_hctx;
2803
2804         return hctx;
2805
2806  free_hctx:
2807         kobject_put(&hctx->kobj);
2808  fail:
2809         return NULL;
2810 }
2811
2812 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2813                                                 struct request_queue *q)
2814 {
2815         int i, j, end;
2816         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2817
2818         /* protect against switching io scheduler  */
2819         mutex_lock(&q->sysfs_lock);
2820         for (i = 0; i < set->nr_hw_queues; i++) {
2821                 int node;
2822                 struct blk_mq_hw_ctx *hctx;
2823
2824                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2825                 /*
2826                  * If the hw queue has been mapped to another numa node,
2827                  * we need to realloc the hctx. If allocation fails, fallback
2828                  * to use the previous one.
2829                  */
2830                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2831                         continue;
2832
2833                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2834                 if (hctx) {
2835                         if (hctxs[i])
2836                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2837                         hctxs[i] = hctx;
2838                 } else {
2839                         if (hctxs[i])
2840                                 pr_warn("Allocate new hctx on node %d fails,\
2841                                                 fallback to previous one on node %d\n",
2842                                                 node, hctxs[i]->numa_node);
2843                         else
2844                                 break;
2845                 }
2846         }
2847         /*
2848          * Increasing nr_hw_queues fails. Free the newly allocated
2849          * hctxs and keep the previous q->nr_hw_queues.
2850          */
2851         if (i != set->nr_hw_queues) {
2852                 j = q->nr_hw_queues;
2853                 end = i;
2854         } else {
2855                 j = i;
2856                 end = q->nr_hw_queues;
2857                 q->nr_hw_queues = set->nr_hw_queues;
2858         }
2859
2860         for (; j < end; j++) {
2861                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2862
2863                 if (hctx) {
2864                         if (hctx->tags)
2865                                 blk_mq_free_map_and_requests(set, j);
2866                         blk_mq_exit_hctx(q, set, hctx, j);
2867                         hctxs[j] = NULL;
2868                 }
2869         }
2870         mutex_unlock(&q->sysfs_lock);
2871 }
2872
2873 /*
2874  * Maximum number of hardware queues we support. For single sets, we'll never
2875  * have more than the CPUs (software queues). For multiple sets, the tag_set
2876  * user may have set ->nr_hw_queues larger.
2877  */
2878 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2879 {
2880         if (set->nr_maps == 1)
2881                 return nr_cpu_ids;
2882
2883         return max(set->nr_hw_queues, nr_cpu_ids);
2884 }
2885
2886 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2887                                                   struct request_queue *q,
2888                                                   bool elevator_init)
2889 {
2890         /* mark the queue as mq asap */
2891         q->mq_ops = set->ops;
2892
2893         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2894                                              blk_mq_poll_stats_bkt,
2895                                              BLK_MQ_POLL_STATS_BKTS, q);
2896         if (!q->poll_cb)
2897                 goto err_exit;
2898
2899         if (blk_mq_alloc_ctxs(q))
2900                 goto err_poll;
2901
2902         /* init q->mq_kobj and sw queues' kobjects */
2903         blk_mq_sysfs_init(q);
2904
2905         q->nr_queues = nr_hw_queues(set);
2906         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2907                                                 GFP_KERNEL, set->numa_node);
2908         if (!q->queue_hw_ctx)
2909                 goto err_sys_init;
2910
2911         INIT_LIST_HEAD(&q->unused_hctx_list);
2912         spin_lock_init(&q->unused_hctx_lock);
2913
2914         blk_mq_realloc_hw_ctxs(set, q);
2915         if (!q->nr_hw_queues)
2916                 goto err_hctxs;
2917
2918         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2919         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2920
2921         q->tag_set = set;
2922
2923         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2924         if (set->nr_maps > HCTX_TYPE_POLL &&
2925             set->map[HCTX_TYPE_POLL].nr_queues)
2926                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2927
2928         q->sg_reserved_size = INT_MAX;
2929
2930         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2931         INIT_LIST_HEAD(&q->requeue_list);
2932         spin_lock_init(&q->requeue_lock);
2933
2934         blk_queue_make_request(q, blk_mq_make_request);
2935
2936         /*
2937          * Do this after blk_queue_make_request() overrides it...
2938          */
2939         q->nr_requests = set->queue_depth;
2940
2941         /*
2942          * Default to classic polling
2943          */
2944         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2945
2946         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2947         blk_mq_add_queue_tag_set(set, q);
2948         blk_mq_map_swqueue(q);
2949
2950         if (elevator_init)
2951                 elevator_init_mq(q);
2952
2953         return q;
2954
2955 err_hctxs:
2956         kfree(q->queue_hw_ctx);
2957         q->nr_hw_queues = 0;
2958 err_sys_init:
2959         blk_mq_sysfs_deinit(q);
2960 err_poll:
2961         blk_stat_free_callback(q->poll_cb);
2962         q->poll_cb = NULL;
2963 err_exit:
2964         q->mq_ops = NULL;
2965         return ERR_PTR(-ENOMEM);
2966 }
2967 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2968
2969 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2970 void blk_mq_exit_queue(struct request_queue *q)
2971 {
2972         struct blk_mq_tag_set *set = q->tag_set;
2973
2974         /* Checks hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED. */
2975         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2976         /* May clear BLK_MQ_F_TAG_QUEUE_SHARED in hctx->flags. */
2977         blk_mq_del_queue_tag_set(q);
2978 }
2979
2980 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2981 {
2982         int i;
2983
2984         for (i = 0; i < set->nr_hw_queues; i++)
2985                 if (!__blk_mq_alloc_rq_map(set, i))
2986                         goto out_unwind;
2987
2988         return 0;
2989
2990 out_unwind:
2991         while (--i >= 0)
2992                 blk_mq_free_rq_map(set->tags[i]);
2993
2994         return -ENOMEM;
2995 }
2996
2997 /*
2998  * Allocate the request maps associated with this tag_set. Note that this
2999  * may reduce the depth asked for, if memory is tight. set->queue_depth
3000  * will be updated to reflect the allocated depth.
3001  */
3002 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
3003 {
3004         unsigned int depth;
3005         int err;
3006
3007         depth = set->queue_depth;
3008         do {
3009                 err = __blk_mq_alloc_rq_maps(set);
3010                 if (!err)
3011                         break;
3012
3013                 set->queue_depth >>= 1;
3014                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
3015                         err = -ENOMEM;
3016                         break;
3017                 }
3018         } while (set->queue_depth);
3019
3020         if (!set->queue_depth || err) {
3021                 pr_err("blk-mq: failed to allocate request map\n");
3022                 return -ENOMEM;
3023         }
3024
3025         if (depth != set->queue_depth)
3026                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
3027                                                 depth, set->queue_depth);
3028
3029         return 0;
3030 }
3031
3032 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
3033 {
3034         /*
3035          * blk_mq_map_queues() and multiple .map_queues() implementations
3036          * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
3037          * number of hardware queues.
3038          */
3039         if (set->nr_maps == 1)
3040                 set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
3041
3042         if (set->ops->map_queues && !is_kdump_kernel()) {
3043                 int i;
3044
3045                 /*
3046                  * transport .map_queues is usually done in the following
3047                  * way:
3048                  *
3049                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3050                  *      mask = get_cpu_mask(queue)
3051                  *      for_each_cpu(cpu, mask)
3052                  *              set->map[x].mq_map[cpu] = queue;
3053                  * }
3054                  *
3055                  * When we need to remap, the table has to be cleared for
3056                  * killing stale mapping since one CPU may not be mapped
3057                  * to any hw queue.
3058                  */
3059                 for (i = 0; i < set->nr_maps; i++)
3060                         blk_mq_clear_mq_map(&set->map[i]);
3061
3062                 return set->ops->map_queues(set);
3063         } else {
3064                 BUG_ON(set->nr_maps > 1);
3065                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3066         }
3067 }
3068
3069 /*
3070  * Alloc a tag set to be associated with one or more request queues.
3071  * May fail with EINVAL for various error conditions. May adjust the
3072  * requested depth down, if it's too large. In that case, the set
3073  * value will be stored in set->queue_depth.
3074  */
3075 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3076 {
3077         int i, ret;
3078
3079         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3080
3081         if (!set->nr_hw_queues)
3082                 return -EINVAL;
3083         if (!set->queue_depth)
3084                 return -EINVAL;
3085         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3086                 return -EINVAL;
3087
3088         if (!set->ops->queue_rq)
3089                 return -EINVAL;
3090
3091         if (!set->ops->get_budget ^ !set->ops->put_budget)
3092                 return -EINVAL;
3093
3094         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3095                 pr_info("blk-mq: reduced tag depth to %u\n",
3096                         BLK_MQ_MAX_DEPTH);
3097                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3098         }
3099
3100         if (!set->nr_maps)
3101                 set->nr_maps = 1;
3102         else if (set->nr_maps > HCTX_MAX_TYPES)
3103                 return -EINVAL;
3104
3105         /*
3106          * If a crashdump is active, then we are potentially in a very
3107          * memory constrained environment. Limit us to 1 queue and
3108          * 64 tags to prevent using too much memory.
3109          */
3110         if (is_kdump_kernel()) {
3111                 set->nr_hw_queues = 1;
3112                 set->nr_maps = 1;
3113                 set->queue_depth = min(64U, set->queue_depth);
3114         }
3115         /*
3116          * There is no use for more h/w queues than cpus if we just have
3117          * a single map
3118          */
3119         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3120                 set->nr_hw_queues = nr_cpu_ids;
3121
3122         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3123                                  GFP_KERNEL, set->numa_node);
3124         if (!set->tags)
3125                 return -ENOMEM;
3126
3127         ret = -ENOMEM;
3128         for (i = 0; i < set->nr_maps; i++) {
3129                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3130                                                   sizeof(set->map[i].mq_map[0]),
3131                                                   GFP_KERNEL, set->numa_node);
3132                 if (!set->map[i].mq_map)
3133                         goto out_free_mq_map;
3134                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3135         }
3136
3137         ret = blk_mq_update_queue_map(set);
3138         if (ret)
3139                 goto out_free_mq_map;
3140
3141         ret = blk_mq_alloc_rq_maps(set);
3142         if (ret)
3143                 goto out_free_mq_map;
3144
3145         mutex_init(&set->tag_list_lock);
3146         INIT_LIST_HEAD(&set->tag_list);
3147
3148         return 0;
3149
3150 out_free_mq_map:
3151         for (i = 0; i < set->nr_maps; i++) {
3152                 kfree(set->map[i].mq_map);
3153                 set->map[i].mq_map = NULL;
3154         }
3155         kfree(set->tags);
3156         set->tags = NULL;
3157         return ret;
3158 }
3159 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3160
3161 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3162 {
3163         int i, j;
3164
3165         for (i = 0; i < nr_hw_queues(set); i++)
3166                 blk_mq_free_map_and_requests(set, i);
3167
3168         for (j = 0; j < set->nr_maps; j++) {
3169                 kfree(set->map[j].mq_map);
3170                 set->map[j].mq_map = NULL;
3171         }
3172
3173         kfree(set->tags);
3174         set->tags = NULL;
3175 }
3176 EXPORT_SYMBOL(blk_mq_free_tag_set);
3177
3178 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3179 {
3180         struct blk_mq_tag_set *set = q->tag_set;
3181         struct blk_mq_hw_ctx *hctx;
3182         int i, ret;
3183
3184         if (!set)
3185                 return -EINVAL;
3186
3187         if (q->nr_requests == nr)
3188                 return 0;
3189
3190         blk_mq_freeze_queue(q);
3191         blk_mq_quiesce_queue(q);
3192
3193         ret = 0;
3194         queue_for_each_hw_ctx(q, hctx, i) {
3195                 if (!hctx->tags)
3196                         continue;
3197                 /*
3198                  * If we're using an MQ scheduler, just update the scheduler
3199                  * queue depth. This is similar to what the old code would do.
3200                  */
3201                 if (!hctx->sched_tags) {
3202                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3203                                                         false);
3204                 } else {
3205                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3206                                                         nr, true);
3207                 }
3208                 if (ret)
3209                         break;
3210                 if (q->elevator && q->elevator->type->ops.depth_updated)
3211                         q->elevator->type->ops.depth_updated(hctx);
3212         }
3213
3214         if (!ret)
3215                 q->nr_requests = nr;
3216
3217         blk_mq_unquiesce_queue(q);
3218         blk_mq_unfreeze_queue(q);
3219
3220         return ret;
3221 }
3222
3223 /*
3224  * request_queue and elevator_type pair.
3225  * It is just used by __blk_mq_update_nr_hw_queues to cache
3226  * the elevator_type associated with a request_queue.
3227  */
3228 struct blk_mq_qe_pair {
3229         struct list_head node;
3230         struct request_queue *q;
3231         struct elevator_type *type;
3232 };
3233
3234 /*
3235  * Cache the elevator_type in qe pair list and switch the
3236  * io scheduler to 'none'
3237  */
3238 static bool blk_mq_elv_switch_none(struct list_head *head,
3239                 struct request_queue *q)
3240 {
3241         struct blk_mq_qe_pair *qe;
3242
3243         if (!q->elevator)
3244                 return true;
3245
3246         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3247         if (!qe)
3248                 return false;
3249
3250         INIT_LIST_HEAD(&qe->node);
3251         qe->q = q;
3252         qe->type = q->elevator->type;
3253         list_add(&qe->node, head);
3254
3255         mutex_lock(&q->sysfs_lock);
3256         /*
3257          * After elevator_switch_mq, the previous elevator_queue will be
3258          * released by elevator_release. The reference of the io scheduler
3259          * module get by elevator_get will also be put. So we need to get
3260          * a reference of the io scheduler module here to prevent it to be
3261          * removed.
3262          */
3263         __module_get(qe->type->elevator_owner);
3264         elevator_switch_mq(q, NULL);
3265         mutex_unlock(&q->sysfs_lock);
3266
3267         return true;
3268 }
3269
3270 static void blk_mq_elv_switch_back(struct list_head *head,
3271                 struct request_queue *q)
3272 {
3273         struct blk_mq_qe_pair *qe;
3274         struct elevator_type *t = NULL;
3275
3276         list_for_each_entry(qe, head, node)
3277                 if (qe->q == q) {
3278                         t = qe->type;
3279                         break;
3280                 }
3281
3282         if (!t)
3283                 return;
3284
3285         list_del(&qe->node);
3286         kfree(qe);
3287
3288         mutex_lock(&q->sysfs_lock);
3289         elevator_switch_mq(q, t);
3290         mutex_unlock(&q->sysfs_lock);
3291 }
3292
3293 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3294                                                         int nr_hw_queues)
3295 {
3296         struct request_queue *q;
3297         LIST_HEAD(head);
3298         int prev_nr_hw_queues;
3299
3300         lockdep_assert_held(&set->tag_list_lock);
3301
3302         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3303                 nr_hw_queues = nr_cpu_ids;
3304         if (nr_hw_queues < 1)
3305                 return;
3306         if (set->nr_maps == 1 && nr_hw_queues == set->nr_hw_queues)
3307                 return;
3308
3309         list_for_each_entry(q, &set->tag_list, tag_set_list)
3310                 blk_mq_freeze_queue(q);
3311         /*
3312          * Sync with blk_mq_queue_tag_busy_iter.
3313          */
3314         synchronize_rcu();
3315         /*
3316          * Switch IO scheduler to 'none', cleaning up the data associated
3317          * with the previous scheduler. We will switch back once we are done
3318          * updating the new sw to hw queue mappings.
3319          */
3320         list_for_each_entry(q, &set->tag_list, tag_set_list)
3321                 if (!blk_mq_elv_switch_none(&head, q))
3322                         goto switch_back;
3323
3324         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3325                 blk_mq_debugfs_unregister_hctxs(q);
3326                 blk_mq_sysfs_unregister(q);
3327         }
3328
3329         prev_nr_hw_queues = set->nr_hw_queues;
3330         set->nr_hw_queues = nr_hw_queues;
3331 fallback:
3332         blk_mq_update_queue_map(set);
3333         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3334                 blk_mq_realloc_hw_ctxs(set, q);
3335                 if (q->nr_hw_queues != set->nr_hw_queues) {
3336                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3337                                         nr_hw_queues, prev_nr_hw_queues);
3338                         set->nr_hw_queues = prev_nr_hw_queues;
3339                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3340                         goto fallback;
3341                 }
3342                 blk_mq_map_swqueue(q);
3343         }
3344
3345         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3346                 blk_mq_sysfs_register(q);
3347                 blk_mq_debugfs_register_hctxs(q);
3348         }
3349
3350 switch_back:
3351         list_for_each_entry(q, &set->tag_list, tag_set_list)
3352                 blk_mq_elv_switch_back(&head, q);
3353
3354         list_for_each_entry(q, &set->tag_list, tag_set_list)
3355                 blk_mq_unfreeze_queue(q);
3356 }
3357
3358 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3359 {
3360         mutex_lock(&set->tag_list_lock);
3361         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3362         mutex_unlock(&set->tag_list_lock);
3363 }
3364 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3365
3366 /* Enable polling stats and return whether they were already enabled. */
3367 static bool blk_poll_stats_enable(struct request_queue *q)
3368 {
3369         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3370             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3371                 return true;
3372         blk_stat_add_callback(q, q->poll_cb);
3373         return false;
3374 }
3375
3376 static void blk_mq_poll_stats_start(struct request_queue *q)
3377 {
3378         /*
3379          * We don't arm the callback if polling stats are not enabled or the
3380          * callback is already active.
3381          */
3382         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3383             blk_stat_is_active(q->poll_cb))
3384                 return;
3385
3386         blk_stat_activate_msecs(q->poll_cb, 100);
3387 }
3388
3389 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3390 {
3391         struct request_queue *q = cb->data;
3392         int bucket;
3393
3394         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3395                 if (cb->stat[bucket].nr_samples)
3396                         q->poll_stat[bucket] = cb->stat[bucket];
3397         }
3398 }
3399
3400 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3401                                        struct blk_mq_hw_ctx *hctx,
3402                                        struct request *rq)
3403 {
3404         unsigned long ret = 0;
3405         int bucket;
3406
3407         /*
3408          * If stats collection isn't on, don't sleep but turn it on for
3409          * future users
3410          */
3411         if (!blk_poll_stats_enable(q))
3412                 return 0;
3413
3414         /*
3415          * As an optimistic guess, use half of the mean service time
3416          * for this type of request. We can (and should) make this smarter.
3417          * For instance, if the completion latencies are tight, we can
3418          * get closer than just half the mean. This is especially
3419          * important on devices where the completion latencies are longer
3420          * than ~10 usec. We do use the stats for the relevant IO size
3421          * if available which does lead to better estimates.
3422          */
3423         bucket = blk_mq_poll_stats_bkt(rq);
3424         if (bucket < 0)
3425                 return ret;
3426
3427         if (q->poll_stat[bucket].nr_samples)
3428                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3429
3430         return ret;
3431 }
3432
3433 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3434                                      struct blk_mq_hw_ctx *hctx,
3435                                      struct request *rq)
3436 {
3437         struct hrtimer_sleeper hs;
3438         enum hrtimer_mode mode;
3439         unsigned int nsecs;
3440         ktime_t kt;
3441
3442         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3443                 return false;
3444
3445         /*
3446          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3447          *
3448          *  0:  use half of prev avg
3449          * >0:  use this specific value
3450          */
3451         if (q->poll_nsec > 0)
3452                 nsecs = q->poll_nsec;
3453         else
3454                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3455
3456         if (!nsecs)
3457                 return false;
3458
3459         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3460
3461         /*
3462          * This will be replaced with the stats tracking code, using
3463          * 'avg_completion_time / 2' as the pre-sleep target.
3464          */
3465         kt = nsecs;
3466
3467         mode = HRTIMER_MODE_REL;
3468         hrtimer_init_sleeper_on_stack(&hs, CLOCK_MONOTONIC, mode);
3469         hrtimer_set_expires(&hs.timer, kt);
3470
3471         do {
3472                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3473                         break;
3474                 set_current_state(TASK_UNINTERRUPTIBLE);
3475                 hrtimer_sleeper_start_expires(&hs, mode);
3476                 if (hs.task)
3477                         io_schedule();
3478                 hrtimer_cancel(&hs.timer);
3479                 mode = HRTIMER_MODE_ABS;
3480         } while (hs.task && !signal_pending(current));
3481
3482         __set_current_state(TASK_RUNNING);
3483         destroy_hrtimer_on_stack(&hs.timer);
3484         return true;
3485 }
3486
3487 static bool blk_mq_poll_hybrid(struct request_queue *q,
3488                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3489 {
3490         struct request *rq;
3491
3492         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3493                 return false;
3494
3495         if (!blk_qc_t_is_internal(cookie))
3496                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3497         else {
3498                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3499                 /*
3500                  * With scheduling, if the request has completed, we'll
3501                  * get a NULL return here, as we clear the sched tag when
3502                  * that happens. The request still remains valid, like always,
3503                  * so we should be safe with just the NULL check.
3504                  */
3505                 if (!rq)
3506                         return false;
3507         }
3508
3509         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3510 }
3511
3512 /**
3513  * blk_poll - poll for IO completions
3514  * @q:  the queue
3515  * @cookie: cookie passed back at IO submission time
3516  * @spin: whether to spin for completions
3517  *
3518  * Description:
3519  *    Poll for completions on the passed in queue. Returns number of
3520  *    completed entries found. If @spin is true, then blk_poll will continue
3521  *    looping until at least one completion is found, unless the task is
3522  *    otherwise marked running (or we need to reschedule).
3523  */
3524 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3525 {
3526         struct blk_mq_hw_ctx *hctx;
3527         long state;
3528
3529         if (!blk_qc_t_valid(cookie) ||
3530             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3531                 return 0;
3532
3533         if (current->plug)
3534                 blk_flush_plug_list(current->plug, false);
3535
3536         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3537
3538         /*
3539          * If we sleep, have the caller restart the poll loop to reset
3540          * the state. Like for the other success return cases, the
3541          * caller is responsible for checking if the IO completed. If
3542          * the IO isn't complete, we'll get called again and will go
3543          * straight to the busy poll loop.
3544          */
3545         if (blk_mq_poll_hybrid(q, hctx, cookie))
3546                 return 1;
3547
3548         hctx->poll_considered++;
3549
3550         state = current->state;
3551         do {
3552                 int ret;
3553
3554                 hctx->poll_invoked++;
3555
3556                 ret = q->mq_ops->poll(hctx);
3557                 if (ret > 0) {
3558                         hctx->poll_success++;
3559                         __set_current_state(TASK_RUNNING);
3560                         return ret;
3561                 }
3562
3563                 if (signal_pending_state(state, current))
3564                         __set_current_state(TASK_RUNNING);
3565
3566                 if (current->state == TASK_RUNNING)
3567                         return 1;
3568                 if (ret < 0 || !spin)
3569                         break;
3570                 cpu_relax();
3571         } while (!need_resched());
3572
3573         __set_current_state(TASK_RUNNING);
3574         return 0;
3575 }
3576 EXPORT_SYMBOL_GPL(blk_poll);
3577
3578 unsigned int blk_mq_rq_cpu(struct request *rq)
3579 {
3580         return rq->mq_ctx->cpu;
3581 }
3582 EXPORT_SYMBOL(blk_mq_rq_cpu);
3583
3584 static int __init blk_mq_init(void)
3585 {
3586         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3587                                 blk_mq_hctx_notify_dead);
3588         return 0;
3589 }
3590 subsys_initcall(blk_mq_init);