block/blk-mq.c

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