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