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