block/blk-mq.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Block multiqueue core code
   4  *
   5  * Copyright (C) 2013-2014 Jens Axboe
   6  * Copyright (C) 2013-2014 Christoph Hellwig
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/module.h>
  10 #include <linux/backing-dev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/kmemleak.h>
  14 #include <linux/mm.h>
  15 #include <linux/init.h>
  16 #include <linux/slab.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/smp.h>
  19 #include <linux/llist.h>
  20 #include <linux/list_sort.h>
  21 #include <linux/cpu.h>
  22 #include <linux/cache.h>
  23 #include <linux/sched/sysctl.h>
  24 #include <linux/sched/topology.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/delay.h>
  27 #include <linux/crash_dump.h>
  28 #include <linux/prefetch.h>
  29
  30 #include <trace/events/block.h>
  31
  32 #include <linux/blk-mq.h>
  33 #include "blk.h"
  34 #include "blk-mq.h"
  35 #include "blk-mq-debugfs.h"
  36 #include "blk-mq-tag.h"
  37 #include "blk-pm.h"
  38 #include "blk-stat.h"
  39 #include "blk-mq-sched.h"
  40 #include "blk-rq-qos.h"
  41
  42 static void blk_mq_poll_stats_start(struct request_queue *q);
  43 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb);
  44
  45 static int blk_mq_poll_stats_bkt(const struct request *rq)
  46 {
  47         int ddir, bytes, bucket;
  48
  49         ddir = rq_data_dir(rq);
  50         bytes = blk_rq_bytes(rq);
  51
  52         bucket = ddir + 2*(ilog2(bytes) - 9);
  53
  54         if (bucket < 0)
  55                 return -1;
  56         else if (bucket >= BLK_MQ_POLL_STATS_BKTS)
  57                 return ddir + BLK_MQ_POLL_STATS_BKTS - 2;
  58
  59         return bucket;
  60 }
  61
  62 /*
  63  * Check if any of the ctx, dispatch list or elevator
  64  * have pending work in this hardware queue.
  65  */
  66 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
  67 {
  68         return !list_empty_careful(&hctx->dispatch) ||
  69                 sbitmap_any_bit_set(&hctx->ctx_map) ||
  70                         blk_mq_sched_has_work(hctx);
  71 }
  72
  73 /*
  74  * Mark this ctx as having pending work in this hardware queue
  75  */
  76 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
  77                                      struct blk_mq_ctx *ctx)
  78 {
  79         const int bit = ctx->index_hw[hctx->type];
  80
  81         if (!sbitmap_test_bit(&hctx->ctx_map, bit))
  82                 sbitmap_set_bit(&hctx->ctx_map, bit);
  83 }
  84
  85 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
  86                                       struct blk_mq_ctx *ctx)
  87 {
  88         const int bit = ctx->index_hw[hctx->type];
  89
  90         sbitmap_clear_bit(&hctx->ctx_map, bit);
  91 }
  92
  93 struct mq_inflight {
  94         struct hd_struct *part;
  95         unsigned int *inflight;
  96 };
  97
  98 static bool blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
  99                                   struct request *rq, void *priv,
 100                                   bool reserved)
 101 {
 102         struct mq_inflight *mi = priv;
 103
 104         /*
 105          * index[0] counts the specific partition that was asked for.
 106          */
 107         if (rq->part == mi->part)
 108                 mi->inflight[0]++;
 109
 110         return true;
 111 }
 112
 113 unsigned int blk_mq_in_flight(struct request_queue *q, struct hd_struct *part)
 114 {
 115         unsigned inflight[2];
 116         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 117
 118         inflight[0] = inflight[1] = 0;
 119         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight, &mi);
 120
 121         return inflight[0];
 122 }
 123
 124 static bool blk_mq_check_inflight_rw(struct blk_mq_hw_ctx *hctx,
 125                                      struct request *rq, void *priv,
 126                                      bool reserved)
 127 {
 128         struct mq_inflight *mi = priv;
 129
 130         if (rq->part == mi->part)
 131                 mi->inflight[rq_data_dir(rq)]++;
 132
 133         return true;
 134 }
 135
 136 void blk_mq_in_flight_rw(struct request_queue *q, struct hd_struct *part,
 137                          unsigned int inflight[2])
 138 {
 139         struct mq_inflight mi = { .part = part, .inflight = inflight, };
 140
 141         inflight[0] = inflight[1] = 0;
 142         blk_mq_queue_tag_busy_iter(q, blk_mq_check_inflight_rw, &mi);
 143 }
 144
 145 void blk_freeze_queue_start(struct request_queue *q)
 146 {
 147         mutex_lock(&q->mq_freeze_lock);
 148         if (++q->mq_freeze_depth == 1) {
 149                 percpu_ref_kill(&q->q_usage_counter);
 150                 mutex_unlock(&q->mq_freeze_lock);
 151                 if (queue_is_mq(q))
 152                         blk_mq_run_hw_queues(q, false);
 153         } else {
 154                 mutex_unlock(&q->mq_freeze_lock);
 155         }
 156 }
 157 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
 158
 159 void blk_mq_freeze_queue_wait(struct request_queue *q)
 160 {
 161         wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->q_usage_counter));
 162 }
 163 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait);
 164
 165 int blk_mq_freeze_queue_wait_timeout(struct request_queue *q,
 166                                      unsigned long timeout)
 167 {
 168         return wait_event_timeout(q->mq_freeze_wq,
 169                                         percpu_ref_is_zero(&q->q_usage_counter),
 170                                         timeout);
 171 }
 172 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_wait_timeout);
 173
 174 /*
 175  * Guarantee no request is in use, so we can change any data structure of
 176  * the queue afterward.
 177  */
 178 void blk_freeze_queue(struct request_queue *q)
 179 {
 180         /*
 181          * In the !blk_mq case we are only calling this to kill the
 182          * q_usage_counter, otherwise this increases the freeze depth
 183          * and waits for it to return to zero.  For this reason there is
 184          * no blk_unfreeze_queue(), and blk_freeze_queue() is not
 185          * exported to drivers as the only user for unfreeze is blk_mq.
 186          */
 187         blk_freeze_queue_start(q);
 188         blk_mq_freeze_queue_wait(q);
 189 }
 190
 191 void blk_mq_freeze_queue(struct request_queue *q)
 192 {
 193         /*
 194          * ...just an alias to keep freeze and unfreeze actions balanced
 195          * in the blk_mq_* namespace
 196          */
 197         blk_freeze_queue(q);
 198 }
 199 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 200
 201 void blk_mq_unfreeze_queue(struct request_queue *q)
 202 {
 203         mutex_lock(&q->mq_freeze_lock);
 204         q->mq_freeze_depth--;
 205         WARN_ON_ONCE(q->mq_freeze_depth < 0);
 206         if (!q->mq_freeze_depth) {
 207                 percpu_ref_resurrect(&q->q_usage_counter);
 208                 wake_up_all(&q->mq_freeze_wq);
 209         }
 210         mutex_unlock(&q->mq_freeze_lock);
 211 }
 212 EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
 213
 214 /*
 215  * FIXME: replace the scsi_internal_device_*block_nowait() calls in the
 216  * mpt3sas driver such that this function can be removed.
 217  */
 218 void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 219 {
 220         blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
 221 }
 222 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 223
 224 /**
 225  * blk_mq_quiesce_queue() - wait until all ongoing dispatches have finished
 226  * @q: request queue.
 227  *
 228  * Note: this function does not prevent that the struct request end_io()
 229  * callback function is invoked. Once this function is returned, we make
 230  * sure no dispatch can happen until the queue is unquiesced via
 231  * blk_mq_unquiesce_queue().
 232  */
 233 void blk_mq_quiesce_queue(struct request_queue *q)
 234 {
 235         struct blk_mq_hw_ctx *hctx;
 236         unsigned int i;
 237         bool rcu = false;
 238
 239         blk_mq_quiesce_queue_nowait(q);
 240
 241         queue_for_each_hw_ctx(q, hctx, i) {
 242                 if (hctx->flags & BLK_MQ_F_BLOCKING)
 243                         synchronize_srcu(hctx->srcu);
 244                 else
 245                         rcu = true;
 246         }
 247         if (rcu)
 248                 synchronize_rcu();
 249 }
 250 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue);
 251
 252 /*
 253  * blk_mq_unquiesce_queue() - counterpart of blk_mq_quiesce_queue()
 254  * @q: request queue.
 255  *
 256  * This function recovers queue into the state before quiescing
 257  * which is done by blk_mq_quiesce_queue.
 258  */
 259 void blk_mq_unquiesce_queue(struct request_queue *q)
 260 {
 261         blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 262
 263         /* dispatch requests which are inserted during quiescing */
 264         blk_mq_run_hw_queues(q, true);
 265 }
 266 EXPORT_SYMBOL_GPL(blk_mq_unquiesce_queue);
 267
 268 void blk_mq_wake_waiters(struct request_queue *q)
 269 {
 270         struct blk_mq_hw_ctx *hctx;
 271         unsigned int i;
 272
 273         queue_for_each_hw_ctx(q, hctx, i)
 274                 if (blk_mq_hw_queue_mapped(hctx))
 275                         blk_mq_tag_wakeup_all(hctx->tags, true);
 276 }
 277
 278 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
 279 {
 280         return blk_mq_has_free_tags(hctx->tags);
 281 }
 282 EXPORT_SYMBOL(blk_mq_can_queue);
 283
 284 /*
 285  * Only need start/end time stamping if we have stats enabled, or using
 286  * an IO scheduler.
 287  */
 288 static inline bool blk_mq_need_time_stamp(struct request *rq)
 289 {
 290         return (rq->rq_flags & RQF_IO_STAT) || rq->q->elevator;
 291 }
 292
 293 static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 294                 unsigned int tag, unsigned int op, u64 alloc_time_ns)
 295 {
 296         struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
 297         struct request *rq = tags->static_rqs[tag];
 298         req_flags_t rq_flags = 0;
 299
 300         if (data->flags & BLK_MQ_REQ_INTERNAL) {
 301                 rq->tag = -1;
 302                 rq->internal_tag = tag;
 303         } else {
 304                 if (data->hctx->flags & BLK_MQ_F_TAG_SHARED) {
 305                         rq_flags = RQF_MQ_INFLIGHT;
 306                         atomic_inc(&data->hctx->nr_active);
 307                 }
 308                 rq->tag = tag;
 309                 rq->internal_tag = -1;
 310                 data->hctx->tags->rqs[rq->tag] = rq;
 311         }
 312
 313         /* csd/requeue_work/fifo_time is initialized before use */
 314         rq->q = data->q;
 315         rq->mq_ctx = data->ctx;
 316         rq->mq_hctx = data->hctx;
 317         rq->rq_flags = rq_flags;
 318         rq->cmd_flags = op;
 319         if (data->flags & BLK_MQ_REQ_PREEMPT)
 320                 rq->rq_flags |= RQF_PREEMPT;
 321         if (blk_queue_io_stat(data->q))
 322                 rq->rq_flags |= RQF_IO_STAT;
 323         INIT_LIST_HEAD(&rq->queuelist);
 324         INIT_HLIST_NODE(&rq->hash);
 325         RB_CLEAR_NODE(&rq->rb_node);
 326         rq->rq_disk = NULL;
 327         rq->part = NULL;
 328 #ifdef CONFIG_BLK_RQ_ALLOC_TIME
 329         rq->alloc_time_ns = alloc_time_ns;
 330 #endif
 331         if (blk_mq_need_time_stamp(rq))
 332                 rq->start_time_ns = ktime_get_ns();
 333         else
 334                 rq->start_time_ns = 0;
 335         rq->io_start_time_ns = 0;
 336         rq->nr_phys_segments = 0;
 337 #if defined(CONFIG_BLK_DEV_INTEGRITY)
 338         rq->nr_integrity_segments = 0;
 339 #endif
 340         /* tag was already set */
 341         rq->extra_len = 0;
 342         WRITE_ONCE(rq->deadline, 0);
 343
 344         rq->timeout = 0;
 345
 346         rq->end_io = NULL;
 347         rq->end_io_data = NULL;
 348
 349         data->ctx->rq_dispatched[op_is_sync(op)]++;
 350         refcount_set(&rq->ref, 1);
 351         return rq;
 352 }
 353
 354 static struct request *blk_mq_get_request(struct request_queue *q,
 355                                           struct bio *bio,
 356                                           struct blk_mq_alloc_data *data)
 357 {
 358         struct elevator_queue *e = q->elevator;
 359         struct request *rq;
 360         unsigned int tag;
 361         bool clear_ctx_on_error = false;
 362         u64 alloc_time_ns = 0;
 363
 364         blk_queue_enter_live(q);
 365
 366         /* alloc_time includes depth and tag waits */
 367         if (blk_queue_rq_alloc_time(q))
 368                 alloc_time_ns = ktime_get_ns();
 369
 370         data->q = q;
 371         if (likely(!data->ctx)) {
 372                 data->ctx = blk_mq_get_ctx(q);
 373                 clear_ctx_on_error = true;
 374         }
 375         if (likely(!data->hctx))
 376                 data->hctx = blk_mq_map_queue(q, data->cmd_flags,
 377                                                 data->ctx);
 378         if (data->cmd_flags & REQ_NOWAIT)
 379                 data->flags |= BLK_MQ_REQ_NOWAIT;
 380
 381         if (e) {
 382                 data->flags |= BLK_MQ_REQ_INTERNAL;
 383
 384                 /*
 385                  * Flush requests are special and go directly to the
 386                  * dispatch list. Don't include reserved tags in the
 387                  * limiting, as it isn't useful.
 388                  */
 389                 if (!op_is_flush(data->cmd_flags) &&
 390                     e->type->ops.limit_depth &&
 391                     !(data->flags & BLK_MQ_REQ_RESERVED))
 392                         e->type->ops.limit_depth(data->cmd_flags, data);
 393         } else {
 394                 blk_mq_tag_busy(data->hctx);
 395         }
 396
 397         tag = blk_mq_get_tag(data);
 398         if (tag == BLK_MQ_TAG_FAIL) {
 399                 if (clear_ctx_on_error)
 400                         data->ctx = NULL;
 401                 blk_queue_exit(q);
 402                 return NULL;
 403         }
 404
 405         rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags, alloc_time_ns);
 406         if (!op_is_flush(data->cmd_flags)) {
 407                 rq->elv.icq = NULL;
 408                 if (e && e->type->ops.prepare_request) {
 409                         if (e->type->icq_cache)
 410                                 blk_mq_sched_assign_ioc(rq);
 411
 412                         e->type->ops.prepare_request(rq, bio);
 413                         rq->rq_flags |= RQF_ELVPRIV;
 414                 }
 415         }
 416         data->hctx->queued++;
 417         return rq;
 418 }
 419
 420 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
 421                 blk_mq_req_flags_t flags)
 422 {
 423         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 424         struct request *rq;
 425         int ret;
 426
 427         ret = blk_queue_enter(q, flags);
 428         if (ret)
 429                 return ERR_PTR(ret);
 430
 431         rq = blk_mq_get_request(q, NULL, &alloc_data);
 432         blk_queue_exit(q);
 433
 434         if (!rq)
 435                 return ERR_PTR(-EWOULDBLOCK);
 436
 437         rq->__data_len = 0;
 438         rq->__sector = (sector_t) -1;
 439         rq->bio = rq->biotail = NULL;
 440         return rq;
 441 }
 442 EXPORT_SYMBOL(blk_mq_alloc_request);
 443
 444 struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 445         unsigned int op, blk_mq_req_flags_t flags, unsigned int hctx_idx)
 446 {
 447         struct blk_mq_alloc_data alloc_data = { .flags = flags, .cmd_flags = op };
 448         struct request *rq;
 449         unsigned int cpu;
 450         int ret;
 451
 452         /*
 453          * If the tag allocator sleeps we could get an allocation for a
 454          * different hardware context.  No need to complicate the low level
 455          * allocator for this for the rare use case of a command tied to
 456          * a specific queue.
 457          */
 458         if (WARN_ON_ONCE(!(flags & BLK_MQ_REQ_NOWAIT)))
 459                 return ERR_PTR(-EINVAL);
 460
 461         if (hctx_idx >= q->nr_hw_queues)
 462                 return ERR_PTR(-EIO);
 463
 464         ret = blk_queue_enter(q, flags);
 465         if (ret)
 466                 return ERR_PTR(ret);
 467
 468         /*
 469          * Check if the hardware context is actually mapped to anything.
 470          * If not tell the caller that it should skip this queue.
 471          */
 472         alloc_data.hctx = q->queue_hw_ctx[hctx_idx];
 473         if (!blk_mq_hw_queue_mapped(alloc_data.hctx)) {
 474                 blk_queue_exit(q);
 475                 return ERR_PTR(-EXDEV);
 476         }
 477         cpu = cpumask_first_and(alloc_data.hctx->cpumask, cpu_online_mask);
 478         alloc_data.ctx = __blk_mq_get_ctx(q, cpu);
 479
 480         rq = blk_mq_get_request(q, NULL, &alloc_data);
 481         blk_queue_exit(q);
 482
 483         if (!rq)
 484                 return ERR_PTR(-EWOULDBLOCK);
 485
 486         return rq;
 487 }
 488 EXPORT_SYMBOL_GPL(blk_mq_alloc_request_hctx);
 489
 490 static void __blk_mq_free_request(struct request *rq)
 491 {
 492         struct request_queue *q = rq->q;
 493         struct blk_mq_ctx *ctx = rq->mq_ctx;
 494         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 495         const int sched_tag = rq->internal_tag;
 496
 497         blk_pm_mark_last_busy(rq);
 498         rq->mq_hctx = NULL;
 499         if (rq->tag != -1)
 500                 blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 501         if (sched_tag != -1)
 502                 blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 503         blk_mq_sched_restart(hctx);
 504         blk_queue_exit(q);
 505 }
 506
 507 void blk_mq_free_request(struct request *rq)
 508 {
 509         struct request_queue *q = rq->q;
 510         struct elevator_queue *e = q->elevator;
 511         struct blk_mq_ctx *ctx = rq->mq_ctx;
 512         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
 513
 514         if (rq->rq_flags & RQF_ELVPRIV) {
 515                 if (e && e->type->ops.finish_request)
 516                         e->type->ops.finish_request(rq);
 517                 if (rq->elv.icq) {
 518                         put_io_context(rq->elv.icq->ioc);
 519                         rq->elv.icq = NULL;
 520                 }
 521         }
 522
 523         ctx->rq_completed[rq_is_sync(rq)]++;
 524         if (rq->rq_flags & RQF_MQ_INFLIGHT)
 525                 atomic_dec(&hctx->nr_active);
 526
 527         if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 528                 laptop_io_completion(q->backing_dev_info);
 529
 530         rq_qos_done(q, rq);
 531
 532         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 533         if (refcount_dec_and_test(&rq->ref))
 534                 __blk_mq_free_request(rq);
 535 }
 536 EXPORT_SYMBOL_GPL(blk_mq_free_request);
 537
 538 inline void __blk_mq_end_request(struct request *rq, blk_status_t error)
 539 {
 540         u64 now = 0;
 541
 542         if (blk_mq_need_time_stamp(rq))
 543                 now = ktime_get_ns();
 544
 545         if (rq->rq_flags & RQF_STATS) {
 546                 blk_mq_poll_stats_start(rq->q);
 547                 blk_stat_add(rq, now);
 548         }
 549
 550         if (rq->internal_tag != -1)
 551                 blk_mq_sched_completed_request(rq, now);
 552
 553         blk_account_io_done(rq, now);
 554
 555         if (rq->end_io) {
 556                 rq_qos_done(rq->q, rq);
 557                 rq->end_io(rq, error);
 558         } else {
 559                 blk_mq_free_request(rq);
 560         }
 561 }
 562 EXPORT_SYMBOL(__blk_mq_end_request);
 563
 564 void blk_mq_end_request(struct request *rq, blk_status_t error)
 565 {
 566         if (blk_update_request(rq, error, blk_rq_bytes(rq)))
 567                 BUG();
 568         __blk_mq_end_request(rq, error);
 569 }
 570 EXPORT_SYMBOL(blk_mq_end_request);
 571
 572 static void __blk_mq_complete_request_remote(void *data)
 573 {
 574         struct request *rq = data;
 575         struct request_queue *q = rq->q;
 576
 577         q->mq_ops->complete(rq);
 578 }
 579
 580 static void __blk_mq_complete_request(struct request *rq)
 581 {
 582         struct blk_mq_ctx *ctx = rq->mq_ctx;
 583         struct request_queue *q = rq->q;
 584         bool shared = false;
 585         int cpu;
 586
 587         WRITE_ONCE(rq->state, MQ_RQ_COMPLETE);
 588         /*
 589          * Most of single queue controllers, there is only one irq vector
 590          * for handling IO completion, and the only irq's affinity is set
 591          * as all possible CPUs. On most of ARCHs, this affinity means the
 592          * irq is handled on one specific CPU.
 593          *
 594          * So complete IO reqeust in softirq context in case of single queue
 595          * for not degrading IO performance by irqsoff latency.
 596          */
 597         if (q->nr_hw_queues == 1) {
 598                 __blk_complete_request(rq);
 599                 return;
 600         }
 601
 602         /*
 603          * For a polled request, always complete locallly, it's pointless
 604          * to redirect the completion.
 605          */
 606         if ((rq->cmd_flags & REQ_HIPRI) ||
 607             !test_bit(QUEUE_FLAG_SAME_COMP, &q->queue_flags)) {
 608                 q->mq_ops->complete(rq);
 609                 return;
 610         }
 611
 612         cpu = get_cpu();
 613         if (!test_bit(QUEUE_FLAG_SAME_FORCE, &q->queue_flags))
 614                 shared = cpus_share_cache(cpu, ctx->cpu);
 615
 616         if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
 617                 rq->csd.func = __blk_mq_complete_request_remote;
 618                 rq->csd.info = rq;
 619                 rq->csd.flags = 0;
 620                 smp_call_function_single_async(ctx->cpu, &rq->csd);
 621         } else {
 622                 q->mq_ops->complete(rq);
 623         }
 624         put_cpu();
 625 }
 626
 627 static void hctx_unlock(struct blk_mq_hw_ctx *hctx, int srcu_idx)
 628         __releases(hctx->srcu)
 629 {
 630         if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 631                 rcu_read_unlock();
 632         else
 633                 srcu_read_unlock(hctx->srcu, srcu_idx);
 634 }
 635
 636 static void hctx_lock(struct blk_mq_hw_ctx *hctx, int *srcu_idx)
 637         __acquires(hctx->srcu)
 638 {
 639         if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 640                 /* shut up gcc false positive */
 641                 *srcu_idx = 0;
 642                 rcu_read_lock();
 643         } else
 644                 *srcu_idx = srcu_read_lock(hctx->srcu);
 645 }
 646
 647 /**
 648  * blk_mq_complete_request - end I/O on a request
 649  * @rq:         the request being processed
 650  *
 651  * Description:
 652  *      Ends all I/O on a request. It does not handle partial completions.
 653  *      The actual completion happens out-of-order, through a IPI handler.
 654  **/
 655 bool blk_mq_complete_request(struct request *rq)
 656 {
 657         if (unlikely(blk_should_fake_timeout(rq->q)))
 658                 return false;
 659         __blk_mq_complete_request(rq);
 660         return true;
 661 }
 662 EXPORT_SYMBOL(blk_mq_complete_request);
 663
 664 int blk_mq_request_started(struct request *rq)
 665 {
 666         return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 667 }
 668 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 669
 670 int blk_mq_request_completed(struct request *rq)
 671 {
 672         return blk_mq_rq_state(rq) == MQ_RQ_COMPLETE;
 673 }
 674 EXPORT_SYMBOL_GPL(blk_mq_request_completed);
 675
 676 void blk_mq_start_request(struct request *rq)
 677 {
 678         struct request_queue *q = rq->q;
 679
 680         trace_block_rq_issue(q, rq);
 681
 682         if (test_bit(QUEUE_FLAG_STATS, &q->queue_flags)) {
 683                 rq->io_start_time_ns = ktime_get_ns();
 684 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 685                 rq->throtl_size = blk_rq_sectors(rq);
 686 #endif
 687                 rq->rq_flags |= RQF_STATS;
 688                 rq_qos_issue(q, rq);
 689         }
 690
 691         WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 692
 693         blk_add_timer(rq);
 694         WRITE_ONCE(rq->state, MQ_RQ_IN_FLIGHT);
 695
 696         if (q->dma_drain_size && blk_rq_bytes(rq)) {
 697                 /*
 698                  * Make sure space for the drain appears.  We know we can do
 699                  * this because max_hw_segments has been adjusted to be one
 700                  * fewer than the device can handle.
 701                  */
 702                 rq->nr_phys_segments++;
 703         }
 704 }
 705 EXPORT_SYMBOL(blk_mq_start_request);
 706
 707 static void __blk_mq_requeue_request(struct request *rq)
 708 {
 709         struct request_queue *q = rq->q;
 710
 711         blk_mq_put_driver_tag(rq);
 712
 713         trace_block_rq_requeue(q, rq);
 714         rq_qos_requeue(q, rq);
 715
 716         if (blk_mq_request_started(rq)) {
 717                 WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 718                 rq->rq_flags &= ~RQF_TIMED_OUT;
 719                 if (q->dma_drain_size && blk_rq_bytes(rq))
 720                         rq->nr_phys_segments--;
 721         }
 722 }
 723
 724 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 725 {
 726         __blk_mq_requeue_request(rq);
 727
 728         /* this request will be re-inserted to io scheduler queue */
 729         blk_mq_sched_requeue_request(rq);
 730
 731         BUG_ON(!list_empty(&rq->queuelist));
 732         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 733 }
 734 EXPORT_SYMBOL(blk_mq_requeue_request);
 735
 736 static void blk_mq_requeue_work(struct work_struct *work)
 737 {
 738         struct request_queue *q =
 739                 container_of(work, struct request_queue, requeue_work.work);
 740         LIST_HEAD(rq_list);
 741         struct request *rq, *next;
 742
 743         spin_lock_irq(&q->requeue_lock);
 744         list_splice_init(&q->requeue_list, &rq_list);
 745         spin_unlock_irq(&q->requeue_lock);
 746
 747         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 748                 if (!(rq->rq_flags & (RQF_SOFTBARRIER | RQF_DONTPREP)))
 749                         continue;
 750
 751                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 752                 list_del_init(&rq->queuelist);
 753                 /*
 754                  * If RQF_DONTPREP, rq has contained some driver specific
 755                  * data, so insert it to hctx dispatch list to avoid any
 756                  * merge.
 757                  */
 758                 if (rq->rq_flags & RQF_DONTPREP)
 759                         blk_mq_request_bypass_insert(rq, false);
 760                 else
 761                         blk_mq_sched_insert_request(rq, true, false, false);
 762         }
 763
 764         while (!list_empty(&rq_list)) {
 765                 rq = list_entry(rq_list.next, struct request, queuelist);
 766                 list_del_init(&rq->queuelist);
 767                 blk_mq_sched_insert_request(rq, false, false, false);
 768         }
 769
 770         blk_mq_run_hw_queues(q, false);
 771 }
 772
 773 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 774                                 bool kick_requeue_list)
 775 {
 776         struct request_queue *q = rq->q;
 777         unsigned long flags;
 778
 779         /*
 780          * We abuse this flag that is otherwise used by the I/O scheduler to
 781          * request head insertion from the workqueue.
 782          */
 783         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 784
 785         spin_lock_irqsave(&q->requeue_lock, flags);
 786         if (at_head) {
 787                 rq->rq_flags |= RQF_SOFTBARRIER;
 788                 list_add(&rq->queuelist, &q->requeue_list);
 789         } else {
 790                 list_add_tail(&rq->queuelist, &q->requeue_list);
 791         }
 792         spin_unlock_irqrestore(&q->requeue_lock, flags);
 793
 794         if (kick_requeue_list)
 795                 blk_mq_kick_requeue_list(q);
 796 }
 797
 798 void blk_mq_kick_requeue_list(struct request_queue *q)
 799 {
 800         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 801 }
 802 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 803
 804 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 805                                     unsigned long msecs)
 806 {
 807         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 808                                     msecs_to_jiffies(msecs));
 809 }
 810 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 811
 812 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 813 {
 814         if (tag < tags->nr_tags) {
 815                 prefetch(tags->rqs[tag]);
 816                 return tags->rqs[tag];
 817         }
 818
 819         return NULL;
 820 }
 821 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 822
 823 static bool blk_mq_rq_inflight(struct blk_mq_hw_ctx *hctx, struct request *rq,
 824                                void *priv, bool reserved)
 825 {
 826         /*
 827          * If we find a request that is inflight and the queue matches,
 828          * we know the queue is busy. Return false to stop the iteration.
 829          */
 830         if (rq->state == MQ_RQ_IN_FLIGHT && rq->q == hctx->queue) {
 831                 bool *busy = priv;
 832
 833                 *busy = true;
 834                 return false;
 835         }
 836
 837         return true;
 838 }
 839
 840 bool blk_mq_queue_inflight(struct request_queue *q)
 841 {
 842         bool busy = false;
 843
 844         blk_mq_queue_tag_busy_iter(q, blk_mq_rq_inflight, &busy);
 845         return busy;
 846 }
 847 EXPORT_SYMBOL_GPL(blk_mq_queue_inflight);
 848
 849 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 850 {
 851         req->rq_flags |= RQF_TIMED_OUT;
 852         if (req->q->mq_ops->timeout) {
 853                 enum blk_eh_timer_return ret;
 854
 855                 ret = req->q->mq_ops->timeout(req, reserved);
 856                 if (ret == BLK_EH_DONE)
 857                         return;
 858                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 859         }
 860
 861         blk_add_timer(req);
 862 }
 863
 864 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 865 {
 866         unsigned long deadline;
 867
 868         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 869                 return false;
 870         if (rq->rq_flags & RQF_TIMED_OUT)
 871                 return false;
 872
 873         deadline = READ_ONCE(rq->deadline);
 874         if (time_after_eq(jiffies, deadline))
 875                 return true;
 876
 877         if (*next == 0)
 878                 *next = deadline;
 879         else if (time_after(*next, deadline))
 880                 *next = deadline;
 881         return false;
 882 }
 883
 884 static bool blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 885                 struct request *rq, void *priv, bool reserved)
 886 {
 887         unsigned long *next = priv;
 888
 889         /*
 890          * Just do a quick check if it is expired before locking the request in
 891          * so we're not unnecessarilly synchronizing across CPUs.
 892          */
 893         if (!blk_mq_req_expired(rq, next))
 894                 return true;
 895
 896         /*
 897          * We have reason to believe the request may be expired. Take a
 898          * reference on the request to lock this request lifetime into its
 899          * currently allocated context to prevent it from being reallocated in
 900          * the event the completion by-passes this timeout handler.
 901          *
 902          * If the reference was already released, then the driver beat the
 903          * timeout handler to posting a natural completion.
 904          */
 905         if (!refcount_inc_not_zero(&rq->ref))
 906                 return true;
 907
 908         /*
 909          * The request is now locked and cannot be reallocated underneath the
 910          * timeout handler's processing. Re-verify this exact request is truly
 911          * expired; if it is not expired, then the request was completed and
 912          * reallocated as a new request.
 913          */
 914         if (blk_mq_req_expired(rq, next))
 915                 blk_mq_rq_timed_out(rq, reserved);
 916         if (refcount_dec_and_test(&rq->ref))
 917                 __blk_mq_free_request(rq);
 918
 919         return true;
 920 }
 921
 922 static void blk_mq_timeout_work(struct work_struct *work)
 923 {
 924         struct request_queue *q =
 925                 container_of(work, struct request_queue, timeout_work);
 926         unsigned long next = 0;
 927         struct blk_mq_hw_ctx *hctx;
 928         int i;
 929
 930         /* A deadlock might occur if a request is stuck requiring a
 931          * timeout at the same time a queue freeze is waiting
 932          * completion, since the timeout code would not be able to
 933          * acquire the queue reference here.
 934          *
 935          * That's why we don't use blk_queue_enter here; instead, we use
 936          * percpu_ref_tryget directly, because we need to be able to
 937          * obtain a reference even in the short window between the queue
 938          * starting to freeze, by dropping the first reference in
 939          * blk_freeze_queue_start, and the moment the last request is
 940          * consumed, marked by the instant q_usage_counter reaches
 941          * zero.
 942          */
 943         if (!percpu_ref_tryget(&q->q_usage_counter))
 944                 return;
 945
 946         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 947
 948         if (next != 0) {
 949                 mod_timer(&q->timeout, next);
 950         } else {
 951                 /*
 952                  * Request timeouts are handled as a forward rolling timer. If
 953                  * we end up here it means that no requests are pending and
 954                  * also that no request has been pending for a while. Mark
 955                  * each hctx as idle.
 956                  */
 957                 queue_for_each_hw_ctx(q, hctx, i) {
 958                         /* the hctx may be unmapped, so check it here */
 959                         if (blk_mq_hw_queue_mapped(hctx))
 960                                 blk_mq_tag_idle(hctx);
 961                 }
 962         }
 963         blk_queue_exit(q);
 964 }
 965
 966 struct flush_busy_ctx_data {
 967         struct blk_mq_hw_ctx *hctx;
 968         struct list_head *list;
 969 };
 970
 971 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 972 {
 973         struct flush_busy_ctx_data *flush_data = data;
 974         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 975         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 976         enum hctx_type type = hctx->type;
 977
 978         spin_lock(&ctx->lock);
 979         list_splice_tail_init(&ctx->rq_lists[type], flush_data->list);
 980         sbitmap_clear_bit(sb, bitnr);
 981         spin_unlock(&ctx->lock);
 982         return true;
 983 }
 984
 985 /*
 986  * Process software queues that have been marked busy, splicing them
 987  * to the for-dispatch
 988  */
 989 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 990 {
 991         struct flush_busy_ctx_data data = {
 992                 .hctx = hctx,
 993                 .list = list,
 994         };
 995
 996         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 997 }
 998 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 999
1000 struct dispatch_rq_data {
1001         struct blk_mq_hw_ctx *hctx;
1002         struct request *rq;
1003 };
1004
1005 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
1006                 void *data)
1007 {
1008         struct dispatch_rq_data *dispatch_data = data;
1009         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
1010         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
1011         enum hctx_type type = hctx->type;
1012
1013         spin_lock(&ctx->lock);
1014         if (!list_empty(&ctx->rq_lists[type])) {
1015                 dispatch_data->rq = list_entry_rq(ctx->rq_lists[type].next);
1016                 list_del_init(&dispatch_data->rq->queuelist);
1017                 if (list_empty(&ctx->rq_lists[type]))
1018                         sbitmap_clear_bit(sb, bitnr);
1019         }
1020         spin_unlock(&ctx->lock);
1021
1022         return !dispatch_data->rq;
1023 }
1024
1025 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
1026                                         struct blk_mq_ctx *start)
1027 {
1028         unsigned off = start ? start->index_hw[hctx->type] : 0;
1029         struct dispatch_rq_data data = {
1030                 .hctx = hctx,
1031                 .rq   = NULL,
1032         };
1033
1034         __sbitmap_for_each_set(&hctx->ctx_map, off,
1035                                dispatch_rq_from_ctx, &data);
1036
1037         return data.rq;
1038 }
1039
1040 static inline unsigned int queued_to_index(unsigned int queued)
1041 {
1042         if (!queued)
1043                 return 0;
1044
1045         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
1046 }
1047
1048 bool blk_mq_get_driver_tag(struct request *rq)
1049 {
1050         struct blk_mq_alloc_data data = {
1051                 .q = rq->q,
1052                 .hctx = rq->mq_hctx,
1053                 .flags = BLK_MQ_REQ_NOWAIT,
1054                 .cmd_flags = rq->cmd_flags,
1055         };
1056         bool shared;
1057
1058         if (rq->tag != -1)
1059                 goto done;
1060
1061         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
1062                 data.flags |= BLK_MQ_REQ_RESERVED;
1063
1064         shared = blk_mq_tag_busy(data.hctx);
1065         rq->tag = blk_mq_get_tag(&data);
1066         if (rq->tag >= 0) {
1067                 if (shared) {
1068                         rq->rq_flags |= RQF_MQ_INFLIGHT;
1069                         atomic_inc(&data.hctx->nr_active);
1070                 }
1071                 data.hctx->tags->rqs[rq->tag] = rq;
1072         }
1073
1074 done:
1075         return rq->tag != -1;
1076 }
1077
1078 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1079                                 int flags, void *key)
1080 {
1081         struct blk_mq_hw_ctx *hctx;
1082
1083         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1084
1085         spin_lock(&hctx->dispatch_wait_lock);
1086         if (!list_empty(&wait->entry)) {
1087                 struct sbitmap_queue *sbq;
1088
1089                 list_del_init(&wait->entry);
1090                 sbq = &hctx->tags->bitmap_tags;
1091                 atomic_dec(&sbq->ws_active);
1092         }
1093         spin_unlock(&hctx->dispatch_wait_lock);
1094
1095         blk_mq_run_hw_queue(hctx, true);
1096         return 1;
1097 }
1098
1099 /*
1100  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1101  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1102  * restart. For both cases, take care to check the condition again after
1103  * marking us as waiting.
1104  */
1105 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx *hctx,
1106                                  struct request *rq)
1107 {
1108         struct sbitmap_queue *sbq = &hctx->tags->bitmap_tags;
1109         struct wait_queue_head *wq;
1110         wait_queue_entry_t *wait;
1111         bool ret;
1112
1113         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1114                 blk_mq_sched_mark_restart_hctx(hctx);
1115
1116                 /*
1117                  * It's possible that a tag was freed in the window between the
1118                  * allocation failure and adding the hardware queue to the wait
1119                  * queue.
1120                  *
1121                  * Don't clear RESTART here, someone else could have set it.
1122                  * At most this will cost an extra queue run.
1123                  */
1124                 return blk_mq_get_driver_tag(rq);
1125         }
1126
1127         wait = &hctx->dispatch_wait;
1128         if (!list_empty_careful(&wait->entry))
1129                 return false;
1130
1131         wq = &bt_wait_ptr(sbq, hctx)->wait;
1132
1133         spin_lock_irq(&wq->lock);
1134         spin_lock(&hctx->dispatch_wait_lock);
1135         if (!list_empty(&wait->entry)) {
1136                 spin_unlock(&hctx->dispatch_wait_lock);
1137                 spin_unlock_irq(&wq->lock);
1138                 return false;
1139         }
1140
1141         atomic_inc(&sbq->ws_active);
1142         wait->flags &= ~WQ_FLAG_EXCLUSIVE;
1143         __add_wait_queue(wq, wait);
1144
1145         /*
1146          * It's possible that a tag was freed in the window between the
1147          * allocation failure and adding the hardware queue to the wait
1148          * queue.
1149          */
1150         ret = blk_mq_get_driver_tag(rq);
1151         if (!ret) {
1152                 spin_unlock(&hctx->dispatch_wait_lock);
1153                 spin_unlock_irq(&wq->lock);
1154                 return false;
1155         }
1156
1157         /*
1158          * We got a tag, remove ourselves from the wait queue to ensure
1159          * someone else gets the wakeup.
1160          */
1161         list_del_init(&wait->entry);
1162         atomic_dec(&sbq->ws_active);
1163         spin_unlock(&hctx->dispatch_wait_lock);
1164         spin_unlock_irq(&wq->lock);
1165
1166         return true;
1167 }
1168
1169 #define BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT  8
1170 #define BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR  4
1171 /*
1172  * Update dispatch busy with the Exponential Weighted Moving Average(EWMA):
1173  * - EWMA is one simple way to compute running average value
1174  * - weight(7/8 and 1/8) is applied so that it can decrease exponentially
1175  * - take 4 as factor for avoiding to get too small(0) result, and this
1176  *   factor doesn't matter because EWMA decreases exponentially
1177  */
1178 static void blk_mq_update_dispatch_busy(struct blk_mq_hw_ctx *hctx, bool busy)
1179 {
1180         unsigned int ewma;
1181
1182         if (hctx->queue->elevator)
1183                 return;
1184
1185         ewma = hctx->dispatch_busy;
1186
1187         if (!ewma && !busy)
1188                 return;
1189
1190         ewma *= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT - 1;
1191         if (busy)
1192                 ewma += 1 << BLK_MQ_DISPATCH_BUSY_EWMA_FACTOR;
1193         ewma /= BLK_MQ_DISPATCH_BUSY_EWMA_WEIGHT;
1194
1195         hctx->dispatch_busy = ewma;
1196 }
1197
1198 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1199
1200 /*
1201  * Returns true if we did some work AND can potentially do more.
1202  */
1203 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1204                              bool got_budget)
1205 {
1206         struct blk_mq_hw_ctx *hctx;
1207         struct request *rq, *nxt;
1208         bool no_tag = false;
1209         int errors, queued;
1210         blk_status_t ret = BLK_STS_OK;
1211
1212         if (list_empty(list))
1213                 return false;
1214
1215         WARN_ON(!list_is_singular(list) && got_budget);
1216
1217         /*
1218          * Now process all the entries, sending them to the driver.
1219          */
1220         errors = queued = 0;
1221         do {
1222                 struct blk_mq_queue_data bd;
1223
1224                 rq = list_first_entry(list, struct request, queuelist);
1225
1226                 hctx = rq->mq_hctx;
1227                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1228                         break;
1229
1230                 if (!blk_mq_get_driver_tag(rq)) {
1231                         /*
1232                          * The initial allocation attempt failed, so we need to
1233                          * rerun the hardware queue when a tag is freed. The
1234                          * waitqueue takes care of that. If the queue is run
1235                          * before we add this entry back on the dispatch list,
1236                          * we'll re-run it below.
1237                          */
1238                         if (!blk_mq_mark_tag_wait(hctx, rq)) {
1239                                 blk_mq_put_dispatch_budget(hctx);
1240                                 /*
1241                                  * For non-shared tags, the RESTART check
1242                                  * will suffice.
1243                                  */
1244                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1245                                         no_tag = true;
1246                                 break;
1247                         }
1248                 }
1249
1250                 list_del_init(&rq->queuelist);
1251
1252                 bd.rq = rq;
1253
1254                 /*
1255                  * Flag last if we have no more requests, or if we have more
1256                  * but can't assign a driver tag to it.
1257                  */
1258                 if (list_empty(list))
1259                         bd.last = true;
1260                 else {
1261                         nxt = list_first_entry(list, struct request, queuelist);
1262                         bd.last = !blk_mq_get_driver_tag(nxt);
1263                 }
1264
1265                 ret = q->mq_ops->queue_rq(hctx, &bd);
1266                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1267                         /*
1268                          * If an I/O scheduler has been configured and we got a
1269                          * driver tag for the next request already, free it
1270                          * again.
1271                          */
1272                         if (!list_empty(list)) {
1273                                 nxt = list_first_entry(list, struct request, queuelist);
1274                                 blk_mq_put_driver_tag(nxt);
1275                         }
1276                         list_add(&rq->queuelist, list);
1277                         __blk_mq_requeue_request(rq);
1278                         break;
1279                 }
1280
1281                 if (unlikely(ret != BLK_STS_OK)) {
1282                         errors++;
1283                         blk_mq_end_request(rq, BLK_STS_IOERR);
1284                         continue;
1285                 }
1286
1287                 queued++;
1288         } while (!list_empty(list));
1289
1290         hctx->dispatched[queued_to_index(queued)]++;
1291
1292         /*
1293          * Any items that need requeuing? Stuff them into hctx->dispatch,
1294          * that is where we will continue on next queue run.
1295          */
1296         if (!list_empty(list)) {
1297                 bool needs_restart;
1298
1299                 /*
1300                  * If we didn't flush the entire list, we could have told
1301                  * the driver there was more coming, but that turned out to
1302                  * be a lie.
1303                  */
1304                 if (q->mq_ops->commit_rqs)
1305                         q->mq_ops->commit_rqs(hctx);
1306
1307                 spin_lock(&hctx->lock);
1308                 list_splice_init(list, &hctx->dispatch);
1309                 spin_unlock(&hctx->lock);
1310
1311                 /*
1312                  * If SCHED_RESTART was set by the caller of this function and
1313                  * it is no longer set that means that it was cleared by another
1314                  * thread and hence that a queue rerun is needed.
1315                  *
1316                  * If 'no_tag' is set, that means that we failed getting
1317                  * a driver tag with an I/O scheduler attached. If our dispatch
1318                  * waitqueue is no longer active, ensure that we run the queue
1319                  * AFTER adding our entries back to the list.
1320                  *
1321                  * If no I/O scheduler has been configured it is possible that
1322                  * the hardware queue got stopped and restarted before requests
1323                  * were pushed back onto the dispatch list. Rerun the queue to
1324                  * avoid starvation. Notes:
1325                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1326                  *   been stopped before rerunning a queue.
1327                  * - Some but not all block drivers stop a queue before
1328                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1329                  *   and dm-rq.
1330                  *
1331                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1332                  * bit is set, run queue after a delay to avoid IO stalls
1333                  * that could otherwise occur if the queue is idle.
1334                  */
1335                 needs_restart = blk_mq_sched_needs_restart(hctx);
1336                 if (!needs_restart ||
1337                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1338                         blk_mq_run_hw_queue(hctx, true);
1339                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1340                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1341
1342                 blk_mq_update_dispatch_busy(hctx, true);
1343                 return false;
1344         } else
1345                 blk_mq_update_dispatch_busy(hctx, false);
1346
1347         /*
1348          * If the host/device is unable to accept more work, inform the
1349          * caller of that.
1350          */
1351         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1352                 return false;
1353
1354         return (queued + errors) != 0;
1355 }
1356
1357 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1358 {
1359         int srcu_idx;
1360
1361         /*
1362          * We should be running this queue from one of the CPUs that
1363          * are mapped to it.
1364          *
1365          * There are at least two related races now between setting
1366          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1367          * __blk_mq_run_hw_queue():
1368          *
1369          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1370          *   but later it becomes online, then this warning is harmless
1371          *   at all
1372          *
1373          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1374          *   but later it becomes offline, then the warning can't be
1375          *   triggered, and we depend on blk-mq timeout handler to
1376          *   handle dispatched requests to this hctx
1377          */
1378         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1379                 cpu_online(hctx->next_cpu)) {
1380                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1381                         raw_smp_processor_id(),
1382                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1383                 dump_stack();
1384         }
1385
1386         /*
1387          * We can't run the queue inline with ints disabled. Ensure that
1388          * we catch bad users of this early.
1389          */
1390         WARN_ON_ONCE(in_interrupt());
1391
1392         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1393
1394         hctx_lock(hctx, &srcu_idx);
1395         blk_mq_sched_dispatch_requests(hctx);
1396         hctx_unlock(hctx, srcu_idx);
1397 }
1398
1399 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1400 {
1401         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1402
1403         if (cpu >= nr_cpu_ids)
1404                 cpu = cpumask_first(hctx->cpumask);
1405         return cpu;
1406 }
1407
1408 /*
1409  * It'd be great if the workqueue API had a way to pass
1410  * in a mask and had some smarts for more clever placement.
1411  * For now we just round-robin here, switching for every
1412  * BLK_MQ_CPU_WORK_BATCH queued items.
1413  */
1414 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1415 {
1416         bool tried = false;
1417         int next_cpu = hctx->next_cpu;
1418
1419         if (hctx->queue->nr_hw_queues == 1)
1420                 return WORK_CPU_UNBOUND;
1421
1422         if (--hctx->next_cpu_batch <= 0) {
1423 select_cpu:
1424                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1425                                 cpu_online_mask);
1426                 if (next_cpu >= nr_cpu_ids)
1427                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1428                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1429         }
1430
1431         /*
1432          * Do unbound schedule if we can't find a online CPU for this hctx,
1433          * and it should only happen in the path of handling CPU DEAD.
1434          */
1435         if (!cpu_online(next_cpu)) {
1436                 if (!tried) {
1437                         tried = true;
1438                         goto select_cpu;
1439                 }
1440
1441                 /*
1442                  * Make sure to re-select CPU next time once after CPUs
1443                  * in hctx->cpumask become online again.
1444                  */
1445                 hctx->next_cpu = next_cpu;
1446                 hctx->next_cpu_batch = 1;
1447                 return WORK_CPU_UNBOUND;
1448         }
1449
1450         hctx->next_cpu = next_cpu;
1451         return next_cpu;
1452 }
1453
1454 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1455                                         unsigned long msecs)
1456 {
1457         if (unlikely(blk_mq_hctx_stopped(hctx)))
1458                 return;
1459
1460         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1461                 int cpu = get_cpu();
1462                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1463                         __blk_mq_run_hw_queue(hctx);
1464                         put_cpu();
1465                         return;
1466                 }
1467
1468                 put_cpu();
1469         }
1470
1471         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1472                                     msecs_to_jiffies(msecs));
1473 }
1474
1475 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1476 {
1477         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1478 }
1479 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1480
1481 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1482 {
1483         int srcu_idx;
1484         bool need_run;
1485
1486         /*
1487          * When queue is quiesced, we may be switching io scheduler, or
1488          * updating nr_hw_queues, or other things, and we can't run queue
1489          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1490          *
1491          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1492          * quiesced.
1493          */
1494         hctx_lock(hctx, &srcu_idx);
1495         need_run = !blk_queue_quiesced(hctx->queue) &&
1496                 blk_mq_hctx_has_pending(hctx);
1497         hctx_unlock(hctx, srcu_idx);
1498
1499         if (need_run) {
1500                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1501                 return true;
1502         }
1503
1504         return false;
1505 }
1506 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1507
1508 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1509 {
1510         struct blk_mq_hw_ctx *hctx;
1511         int i;
1512
1513         queue_for_each_hw_ctx(q, hctx, i) {
1514                 if (blk_mq_hctx_stopped(hctx))
1515                         continue;
1516
1517                 blk_mq_run_hw_queue(hctx, async);
1518         }
1519 }
1520 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1521
1522 /**
1523  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1524  * @q: request queue.
1525  *
1526  * The caller is responsible for serializing this function against
1527  * blk_mq_{start,stop}_hw_queue().
1528  */
1529 bool blk_mq_queue_stopped(struct request_queue *q)
1530 {
1531         struct blk_mq_hw_ctx *hctx;
1532         int i;
1533
1534         queue_for_each_hw_ctx(q, hctx, i)
1535                 if (blk_mq_hctx_stopped(hctx))
1536                         return true;
1537
1538         return false;
1539 }
1540 EXPORT_SYMBOL(blk_mq_queue_stopped);
1541
1542 /*
1543  * This function is often used for pausing .queue_rq() by driver when
1544  * there isn't enough resource or some conditions aren't satisfied, and
1545  * BLK_STS_RESOURCE is usually returned.
1546  *
1547  * We do not guarantee that dispatch can be drained or blocked
1548  * after blk_mq_stop_hw_queue() returns. Please use
1549  * blk_mq_quiesce_queue() for that requirement.
1550  */
1551 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1552 {
1553         cancel_delayed_work(&hctx->run_work);
1554
1555         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1556 }
1557 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1558
1559 /*
1560  * This function is often used for pausing .queue_rq() by driver when
1561  * there isn't enough resource or some conditions aren't satisfied, and
1562  * BLK_STS_RESOURCE is usually returned.
1563  *
1564  * We do not guarantee that dispatch can be drained or blocked
1565  * after blk_mq_stop_hw_queues() returns. Please use
1566  * blk_mq_quiesce_queue() for that requirement.
1567  */
1568 void blk_mq_stop_hw_queues(struct request_queue *q)
1569 {
1570         struct blk_mq_hw_ctx *hctx;
1571         int i;
1572
1573         queue_for_each_hw_ctx(q, hctx, i)
1574                 blk_mq_stop_hw_queue(hctx);
1575 }
1576 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1577
1578 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1579 {
1580         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1581
1582         blk_mq_run_hw_queue(hctx, false);
1583 }
1584 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1585
1586 void blk_mq_start_hw_queues(struct request_queue *q)
1587 {
1588         struct blk_mq_hw_ctx *hctx;
1589         int i;
1590
1591         queue_for_each_hw_ctx(q, hctx, i)
1592                 blk_mq_start_hw_queue(hctx);
1593 }
1594 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1595
1596 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1597 {
1598         if (!blk_mq_hctx_stopped(hctx))
1599                 return;
1600
1601         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1602         blk_mq_run_hw_queue(hctx, async);
1603 }
1604 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1605
1606 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1607 {
1608         struct blk_mq_hw_ctx *hctx;
1609         int i;
1610
1611         queue_for_each_hw_ctx(q, hctx, i)
1612                 blk_mq_start_stopped_hw_queue(hctx, async);
1613 }
1614 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1615
1616 static void blk_mq_run_work_fn(struct work_struct *work)
1617 {
1618         struct blk_mq_hw_ctx *hctx;
1619
1620         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1621
1622         /*
1623          * If we are stopped, don't run the queue.
1624          */
1625         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1626                 return;
1627
1628         __blk_mq_run_hw_queue(hctx);
1629 }
1630
1631 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1632                                             struct request *rq,
1633                                             bool at_head)
1634 {
1635         struct blk_mq_ctx *ctx = rq->mq_ctx;
1636         enum hctx_type type = hctx->type;
1637
1638         lockdep_assert_held(&ctx->lock);
1639
1640         trace_block_rq_insert(hctx->queue, rq);
1641
1642         if (at_head)
1643                 list_add(&rq->queuelist, &ctx->rq_lists[type]);
1644         else
1645                 list_add_tail(&rq->queuelist, &ctx->rq_lists[type]);
1646 }
1647
1648 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1649                              bool at_head)
1650 {
1651         struct blk_mq_ctx *ctx = rq->mq_ctx;
1652
1653         lockdep_assert_held(&ctx->lock);
1654
1655         __blk_mq_insert_req_list(hctx, rq, at_head);
1656         blk_mq_hctx_mark_pending(hctx, ctx);
1657 }
1658
1659 /*
1660  * Should only be used carefully, when the caller knows we want to
1661  * bypass a potential IO scheduler on the target device.
1662  */
1663 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1664 {
1665         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1666
1667         spin_lock(&hctx->lock);
1668         list_add_tail(&rq->queuelist, &hctx->dispatch);
1669         spin_unlock(&hctx->lock);
1670
1671         if (run_queue)
1672                 blk_mq_run_hw_queue(hctx, false);
1673 }
1674
1675 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1676                             struct list_head *list)
1677
1678 {
1679         struct request *rq;
1680         enum hctx_type type = hctx->type;
1681
1682         /*
1683          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1684          * offline now
1685          */
1686         list_for_each_entry(rq, list, queuelist) {
1687                 BUG_ON(rq->mq_ctx != ctx);
1688                 trace_block_rq_insert(hctx->queue, rq);
1689         }
1690
1691         spin_lock(&ctx->lock);
1692         list_splice_tail_init(list, &ctx->rq_lists[type]);
1693         blk_mq_hctx_mark_pending(hctx, ctx);
1694         spin_unlock(&ctx->lock);
1695 }
1696
1697 static int plug_rq_cmp(void *priv, struct list_head *a, struct list_head *b)
1698 {
1699         struct request *rqa = container_of(a, struct request, queuelist);
1700         struct request *rqb = container_of(b, struct request, queuelist);
1701
1702         if (rqa->mq_ctx < rqb->mq_ctx)
1703                 return -1;
1704         else if (rqa->mq_ctx > rqb->mq_ctx)
1705                 return 1;
1706         else if (rqa->mq_hctx < rqb->mq_hctx)
1707                 return -1;
1708         else if (rqa->mq_hctx > rqb->mq_hctx)
1709                 return 1;
1710
1711         return blk_rq_pos(rqa) > blk_rq_pos(rqb);
1712 }
1713
1714 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1715 {
1716         struct blk_mq_hw_ctx *this_hctx;
1717         struct blk_mq_ctx *this_ctx;
1718         struct request_queue *this_q;
1719         struct request *rq;
1720         LIST_HEAD(list);
1721         LIST_HEAD(rq_list);
1722         unsigned int depth;
1723
1724         list_splice_init(&plug->mq_list, &list);
1725
1726         if (plug->rq_count > 2 && plug->multiple_queues)
1727                 list_sort(NULL, &list, plug_rq_cmp);
1728
1729         plug->rq_count = 0;
1730
1731         this_q = NULL;
1732         this_hctx = NULL;
1733         this_ctx = NULL;
1734         depth = 0;
1735
1736         while (!list_empty(&list)) {
1737                 rq = list_entry_rq(list.next);
1738                 list_del_init(&rq->queuelist);
1739                 BUG_ON(!rq->q);
1740                 if (rq->mq_hctx != this_hctx || rq->mq_ctx != this_ctx) {
1741                         if (this_hctx) {
1742                                 trace_block_unplug(this_q, depth, !from_schedule);
1743                                 blk_mq_sched_insert_requests(this_hctx, this_ctx,
1744                                                                 &rq_list,
1745                                                                 from_schedule);
1746                         }
1747
1748                         this_q = rq->q;
1749                         this_ctx = rq->mq_ctx;
1750                         this_hctx = rq->mq_hctx;
1751                         depth = 0;
1752                 }
1753
1754                 depth++;
1755                 list_add_tail(&rq->queuelist, &rq_list);
1756         }
1757
1758         /*
1759          * If 'this_hctx' is set, we know we have entries to complete
1760          * on 'rq_list'. Do those.
1761          */
1762         if (this_hctx) {
1763                 trace_block_unplug(this_q, depth, !from_schedule);
1764                 blk_mq_sched_insert_requests(this_hctx, this_ctx, &rq_list,
1765                                                 from_schedule);
1766         }
1767 }
1768
1769 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio,
1770                 unsigned int nr_segs)
1771 {
1772         if (bio->bi_opf & REQ_RAHEAD)
1773                 rq->cmd_flags |= REQ_FAILFAST_MASK;
1774
1775         rq->__sector = bio->bi_iter.bi_sector;
1776         rq->write_hint = bio->bi_write_hint;
1777         blk_rq_bio_prep(rq, bio, nr_segs);
1778
1779         blk_account_io_start(rq, true);
1780 }
1781
1782 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1783                                             struct request *rq,
1784                                             blk_qc_t *cookie, bool last)
1785 {
1786         struct request_queue *q = rq->q;
1787         struct blk_mq_queue_data bd = {
1788                 .rq = rq,
1789                 .last = last,
1790         };
1791         blk_qc_t new_cookie;
1792         blk_status_t ret;
1793
1794         new_cookie = request_to_qc_t(hctx, rq);
1795
1796         /*
1797          * For OK queue, we are done. For error, caller may kill it.
1798          * Any other error (busy), just add it to our list as we
1799          * previously would have done.
1800          */
1801         ret = q->mq_ops->queue_rq(hctx, &bd);
1802         switch (ret) {
1803         case BLK_STS_OK:
1804                 blk_mq_update_dispatch_busy(hctx, false);
1805                 *cookie = new_cookie;
1806                 break;
1807         case BLK_STS_RESOURCE:
1808         case BLK_STS_DEV_RESOURCE:
1809                 blk_mq_update_dispatch_busy(hctx, true);
1810                 __blk_mq_requeue_request(rq);
1811                 break;
1812         default:
1813                 blk_mq_update_dispatch_busy(hctx, false);
1814                 *cookie = BLK_QC_T_NONE;
1815                 break;
1816         }
1817
1818         return ret;
1819 }
1820
1821 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1822                                                 struct request *rq,
1823                                                 blk_qc_t *cookie,
1824                                                 bool bypass_insert, bool last)
1825 {
1826         struct request_queue *q = rq->q;
1827         bool run_queue = true;
1828
1829         /*
1830          * RCU or SRCU read lock is needed before checking quiesced flag.
1831          *
1832          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1833          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1834          * and avoid driver to try to dispatch again.
1835          */
1836         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1837                 run_queue = false;
1838                 bypass_insert = false;
1839                 goto insert;
1840         }
1841
1842         if (q->elevator && !bypass_insert)
1843                 goto insert;
1844
1845         if (!blk_mq_get_dispatch_budget(hctx))
1846                 goto insert;
1847
1848         if (!blk_mq_get_driver_tag(rq)) {
1849                 blk_mq_put_dispatch_budget(hctx);
1850                 goto insert;
1851         }
1852
1853         return __blk_mq_issue_directly(hctx, rq, cookie, last);
1854 insert:
1855         if (bypass_insert)
1856                 return BLK_STS_RESOURCE;
1857
1858         blk_mq_request_bypass_insert(rq, run_queue);
1859         return BLK_STS_OK;
1860 }
1861
1862 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1863                 struct request *rq, blk_qc_t *cookie)
1864 {
1865         blk_status_t ret;
1866         int srcu_idx;
1867
1868         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1869
1870         hctx_lock(hctx, &srcu_idx);
1871
1872         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false, true);
1873         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1874                 blk_mq_request_bypass_insert(rq, true);
1875         else if (ret != BLK_STS_OK)
1876                 blk_mq_end_request(rq, ret);
1877
1878         hctx_unlock(hctx, srcu_idx);
1879 }
1880
1881 blk_status_t blk_mq_request_issue_directly(struct request *rq, bool last)
1882 {
1883         blk_status_t ret;
1884         int srcu_idx;
1885         blk_qc_t unused_cookie;
1886         struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
1887
1888         hctx_lock(hctx, &srcu_idx);
1889         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true, last);
1890         hctx_unlock(hctx, srcu_idx);
1891
1892         return ret;
1893 }
1894
1895 void blk_mq_try_issue_list_directly(struct blk_mq_hw_ctx *hctx,
1896                 struct list_head *list)
1897 {
1898         while (!list_empty(list)) {
1899                 blk_status_t ret;
1900                 struct request *rq = list_first_entry(list, struct request,
1901                                 queuelist);
1902
1903                 list_del_init(&rq->queuelist);
1904                 ret = blk_mq_request_issue_directly(rq, list_empty(list));
1905                 if (ret != BLK_STS_OK) {
1906                         if (ret == BLK_STS_RESOURCE ||
1907                                         ret == BLK_STS_DEV_RESOURCE) {
1908                                 blk_mq_request_bypass_insert(rq,
1909                                                         list_empty(list));
1910                                 break;
1911                         }
1912                         blk_mq_end_request(rq, ret);
1913                 }
1914         }
1915
1916         /*
1917          * If we didn't flush the entire list, we could have told
1918          * the driver there was more coming, but that turned out to
1919          * be a lie.
1920          */
1921         if (!list_empty(list) && hctx->queue->mq_ops->commit_rqs)
1922                 hctx->queue->mq_ops->commit_rqs(hctx);
1923 }
1924
1925 static void blk_add_rq_to_plug(struct blk_plug *plug, struct request *rq)
1926 {
1927         list_add_tail(&rq->queuelist, &plug->mq_list);
1928         plug->rq_count++;
1929         if (!plug->multiple_queues && !list_is_singular(&plug->mq_list)) {
1930                 struct request *tmp;
1931
1932                 tmp = list_first_entry(&plug->mq_list, struct request,
1933                                                 queuelist);
1934                 if (tmp->q != rq->q)
1935                         plug->multiple_queues = true;
1936         }
1937 }
1938
1939 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1940 {
1941         const int is_sync = op_is_sync(bio->bi_opf);
1942         const int is_flush_fua = op_is_flush(bio->bi_opf);
1943         struct blk_mq_alloc_data data = { .flags = 0};
1944         struct request *rq;
1945         struct blk_plug *plug;
1946         struct request *same_queue_rq = NULL;
1947         unsigned int nr_segs;
1948         blk_qc_t cookie;
1949
1950         blk_queue_bounce(q, &bio);
1951         __blk_queue_split(q, &bio, &nr_segs);
1952
1953         if (!bio_integrity_prep(bio))
1954                 return BLK_QC_T_NONE;
1955
1956         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1957             blk_attempt_plug_merge(q, bio, nr_segs, &same_queue_rq))
1958                 return BLK_QC_T_NONE;
1959
1960         if (blk_mq_sched_bio_merge(q, bio, nr_segs))
1961                 return BLK_QC_T_NONE;
1962
1963         rq_qos_throttle(q, bio);
1964
1965         data.cmd_flags = bio->bi_opf;
1966         rq = blk_mq_get_request(q, bio, &data);
1967         if (unlikely(!rq)) {
1968                 rq_qos_cleanup(q, bio);
1969
1970                 cookie = BLK_QC_T_NONE;
1971                 if (bio->bi_opf & REQ_NOWAIT_INLINE)
1972                         cookie = BLK_QC_T_EAGAIN;
1973                 else if (bio->bi_opf & REQ_NOWAIT)
1974                         bio_wouldblock_error(bio);
1975                 return cookie;
1976         }
1977
1978         trace_block_getrq(q, bio, bio->bi_opf);
1979
1980         rq_qos_track(q, rq, bio);
1981
1982         cookie = request_to_qc_t(data.hctx, rq);
1983
1984         blk_mq_bio_to_request(rq, bio, nr_segs);
1985
1986         plug = blk_mq_plug(q, bio);
1987         if (unlikely(is_flush_fua)) {
1988                 /* bypass scheduler for flush rq */
1989                 blk_insert_flush(rq);
1990                 blk_mq_run_hw_queue(data.hctx, true);
1991         } else if (plug && (q->nr_hw_queues == 1 || q->mq_ops->commit_rqs)) {
1992                 /*
1993                  * Use plugging if we have a ->commit_rqs() hook as well, as
1994                  * we know the driver uses bd->last in a smart fashion.
1995                  */
1996                 unsigned int request_count = plug->rq_count;
1997                 struct request *last = NULL;
1998
1999                 if (!request_count)
2000                         trace_block_plug(q);
2001                 else
2002                         last = list_entry_rq(plug->mq_list.prev);
2003
2004                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
2005                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
2006                         blk_flush_plug_list(plug, false);
2007                         trace_block_plug(q);
2008                 }
2009
2010                 blk_add_rq_to_plug(plug, rq);
2011         } else if (plug && !blk_queue_nomerges(q)) {
2012                 /*
2013                  * We do limited plugging. If the bio can be merged, do that.
2014                  * Otherwise the existing request in the plug list will be
2015                  * issued. So the plug list will have one request at most
2016                  * The plug list might get flushed before this. If that happens,
2017                  * the plug list is empty, and same_queue_rq is invalid.
2018                  */
2019                 if (list_empty(&plug->mq_list))
2020                         same_queue_rq = NULL;
2021                 if (same_queue_rq) {
2022                         list_del_init(&same_queue_rq->queuelist);
2023                         plug->rq_count--;
2024                 }
2025                 blk_add_rq_to_plug(plug, rq);
2026                 trace_block_plug(q);
2027
2028                 if (same_queue_rq) {
2029                         data.hctx = same_queue_rq->mq_hctx;
2030                         trace_block_unplug(q, 1, true);
2031                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
2032                                         &cookie);
2033                 }
2034         } else if ((q->nr_hw_queues > 1 && is_sync) || (!q->elevator &&
2035                         !data.hctx->dispatch_busy)) {
2036                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
2037         } else {
2038                 blk_mq_sched_insert_request(rq, false, true, true);
2039         }
2040
2041         return cookie;
2042 }
2043
2044 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2045                      unsigned int hctx_idx)
2046 {
2047         struct page *page;
2048
2049         if (tags->rqs && set->ops->exit_request) {
2050                 int i;
2051
2052                 for (i = 0; i < tags->nr_tags; i++) {
2053                         struct request *rq = tags->static_rqs[i];
2054
2055                         if (!rq)
2056                                 continue;
2057                         set->ops->exit_request(set, rq, hctx_idx);
2058                         tags->static_rqs[i] = NULL;
2059                 }
2060         }
2061
2062         while (!list_empty(&tags->page_list)) {
2063                 page = list_first_entry(&tags->page_list, struct page, lru);
2064                 list_del_init(&page->lru);
2065                 /*
2066                  * Remove kmemleak object previously allocated in
2067                  * blk_mq_alloc_rqs().
2068                  */
2069                 kmemleak_free(page_address(page));
2070                 __free_pages(page, page->private);
2071         }
2072 }
2073
2074 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
2075 {
2076         kfree(tags->rqs);
2077         tags->rqs = NULL;
2078         kfree(tags->static_rqs);
2079         tags->static_rqs = NULL;
2080
2081         blk_mq_free_tags(tags);
2082 }
2083
2084 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
2085                                         unsigned int hctx_idx,
2086                                         unsigned int nr_tags,
2087                                         unsigned int reserved_tags)
2088 {
2089         struct blk_mq_tags *tags;
2090         int node;
2091
2092         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2093         if (node == NUMA_NO_NODE)
2094                 node = set->numa_node;
2095
2096         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
2097                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
2098         if (!tags)
2099                 return NULL;
2100
2101         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2102                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2103                                  node);
2104         if (!tags->rqs) {
2105                 blk_mq_free_tags(tags);
2106                 return NULL;
2107         }
2108
2109         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
2110                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
2111                                         node);
2112         if (!tags->static_rqs) {
2113                 kfree(tags->rqs);
2114                 blk_mq_free_tags(tags);
2115                 return NULL;
2116         }
2117
2118         return tags;
2119 }
2120
2121 static size_t order_to_size(unsigned int order)
2122 {
2123         return (size_t)PAGE_SIZE << order;
2124 }
2125
2126 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2127                                unsigned int hctx_idx, int node)
2128 {
2129         int ret;
2130
2131         if (set->ops->init_request) {
2132                 ret = set->ops->init_request(set, rq, hctx_idx, node);
2133                 if (ret)
2134                         return ret;
2135         }
2136
2137         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
2138         return 0;
2139 }
2140
2141 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
2142                      unsigned int hctx_idx, unsigned int depth)
2143 {
2144         unsigned int i, j, entries_per_page, max_order = 4;
2145         size_t rq_size, left;
2146         int node;
2147
2148         node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], hctx_idx);
2149         if (node == NUMA_NO_NODE)
2150                 node = set->numa_node;
2151
2152         INIT_LIST_HEAD(&tags->page_list);
2153
2154         /*
2155          * rq_size is the size of the request plus driver payload, rounded
2156          * to the cacheline size
2157          */
2158         rq_size = round_up(sizeof(struct request) + set->cmd_size,
2159                                 cache_line_size());
2160         left = rq_size * depth;
2161
2162         for (i = 0; i < depth; ) {
2163                 int this_order = max_order;
2164                 struct page *page;
2165                 int to_do;
2166                 void *p;
2167
2168                 while (this_order && left < order_to_size(this_order - 1))
2169                         this_order--;
2170
2171                 do {
2172                         page = alloc_pages_node(node,
2173                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
2174                                 this_order);
2175                         if (page)
2176                                 break;
2177                         if (!this_order--)
2178                                 break;
2179                         if (order_to_size(this_order) < rq_size)
2180                                 break;
2181                 } while (1);
2182
2183                 if (!page)
2184                         goto fail;
2185
2186                 page->private = this_order;
2187                 list_add_tail(&page->lru, &tags->page_list);
2188
2189                 p = page_address(page);
2190                 /*
2191                  * Allow kmemleak to scan these pages as they contain pointers
2192                  * to additional allocations like via ops->init_request().
2193                  */
2194                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2195                 entries_per_page = order_to_size(this_order) / rq_size;
2196                 to_do = min(entries_per_page, depth - i);
2197                 left -= to_do * rq_size;
2198                 for (j = 0; j < to_do; j++) {
2199                         struct request *rq = p;
2200
2201                         tags->static_rqs[i] = rq;
2202                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2203                                 tags->static_rqs[i] = NULL;
2204                                 goto fail;
2205                         }
2206
2207                         p += rq_size;
2208                         i++;
2209                 }
2210         }
2211         return 0;
2212
2213 fail:
2214         blk_mq_free_rqs(set, tags, hctx_idx);
2215         return -ENOMEM;
2216 }
2217
2218 /*
2219  * 'cpu' is going away. splice any existing rq_list entries from this
2220  * software queue to the hw queue dispatch list, and ensure that it
2221  * gets run.
2222  */
2223 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2224 {
2225         struct blk_mq_hw_ctx *hctx;
2226         struct blk_mq_ctx *ctx;
2227         LIST_HEAD(tmp);
2228         enum hctx_type type;
2229
2230         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2231         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2232         type = hctx->type;
2233
2234         spin_lock(&ctx->lock);
2235         if (!list_empty(&ctx->rq_lists[type])) {
2236                 list_splice_init(&ctx->rq_lists[type], &tmp);
2237                 blk_mq_hctx_clear_pending(hctx, ctx);
2238         }
2239         spin_unlock(&ctx->lock);
2240
2241         if (list_empty(&tmp))
2242                 return 0;
2243
2244         spin_lock(&hctx->lock);
2245         list_splice_tail_init(&tmp, &hctx->dispatch);
2246         spin_unlock(&hctx->lock);
2247
2248         blk_mq_run_hw_queue(hctx, true);
2249         return 0;
2250 }
2251
2252 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2253 {
2254         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2255                                             &hctx->cpuhp_dead);
2256 }
2257
2258 /* hctx->ctxs will be freed in queue's release handler */
2259 static void blk_mq_exit_hctx(struct request_queue *q,
2260                 struct blk_mq_tag_set *set,
2261                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2262 {
2263         if (blk_mq_hw_queue_mapped(hctx))
2264                 blk_mq_tag_idle(hctx);
2265
2266         if (set->ops->exit_request)
2267                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2268
2269         if (set->ops->exit_hctx)
2270                 set->ops->exit_hctx(hctx, hctx_idx);
2271
2272         blk_mq_remove_cpuhp(hctx);
2273
2274         spin_lock(&q->unused_hctx_lock);
2275         list_add(&hctx->hctx_list, &q->unused_hctx_list);
2276         spin_unlock(&q->unused_hctx_lock);
2277 }
2278
2279 static void blk_mq_exit_hw_queues(struct request_queue *q,
2280                 struct blk_mq_tag_set *set, int nr_queue)
2281 {
2282         struct blk_mq_hw_ctx *hctx;
2283         unsigned int i;
2284
2285         queue_for_each_hw_ctx(q, hctx, i) {
2286                 if (i == nr_queue)
2287                         break;
2288                 blk_mq_debugfs_unregister_hctx(hctx);
2289                 blk_mq_exit_hctx(q, set, hctx, i);
2290         }
2291 }
2292
2293 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2294 {
2295         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2296
2297         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2298                            __alignof__(struct blk_mq_hw_ctx)) !=
2299                      sizeof(struct blk_mq_hw_ctx));
2300
2301         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2302                 hw_ctx_size += sizeof(struct srcu_struct);
2303
2304         return hw_ctx_size;
2305 }
2306
2307 static int blk_mq_init_hctx(struct request_queue *q,
2308                 struct blk_mq_tag_set *set,
2309                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2310 {
2311         hctx->queue_num = hctx_idx;
2312
2313         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2314
2315         hctx->tags = set->tags[hctx_idx];
2316
2317         if (set->ops->init_hctx &&
2318             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2319                 goto unregister_cpu_notifier;
2320
2321         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx,
2322                                 hctx->numa_node))
2323                 goto exit_hctx;
2324         return 0;
2325
2326  exit_hctx:
2327         if (set->ops->exit_hctx)
2328                 set->ops->exit_hctx(hctx, hctx_idx);
2329  unregister_cpu_notifier:
2330         blk_mq_remove_cpuhp(hctx);
2331         return -1;
2332 }
2333
2334 static struct blk_mq_hw_ctx *
2335 blk_mq_alloc_hctx(struct request_queue *q, struct blk_mq_tag_set *set,
2336                 int node)
2337 {
2338         struct blk_mq_hw_ctx *hctx;
2339         gfp_t gfp = GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY;
2340
2341         hctx = kzalloc_node(blk_mq_hw_ctx_size(set), gfp, node);
2342         if (!hctx)
2343                 goto fail_alloc_hctx;
2344
2345         if (!zalloc_cpumask_var_node(&hctx->cpumask, gfp, node))
2346                 goto free_hctx;
2347
2348         atomic_set(&hctx->nr_active, 0);
2349         if (node == NUMA_NO_NODE)
2350                 node = set->numa_node;
2351         hctx->numa_node = node;
2352
2353         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2354         spin_lock_init(&hctx->lock);
2355         INIT_LIST_HEAD(&hctx->dispatch);
2356         hctx->queue = q;
2357         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2358
2359         INIT_LIST_HEAD(&hctx->hctx_list);
2360
2361         /*
2362          * Allocate space for all possible cpus to avoid allocation at
2363          * runtime
2364          */
2365         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2366                         gfp, node);
2367         if (!hctx->ctxs)
2368                 goto free_cpumask;
2369
2370         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8),
2371                                 gfp, node))
2372                 goto free_ctxs;
2373         hctx->nr_ctx = 0;
2374
2375         spin_lock_init(&hctx->dispatch_wait_lock);
2376         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2377         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2378
2379         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size,
2380                         gfp);
2381         if (!hctx->fq)
2382                 goto free_bitmap;
2383
2384         if (hctx->flags & BLK_MQ_F_BLOCKING)
2385                 init_srcu_struct(hctx->srcu);
2386         blk_mq_hctx_kobj_init(hctx);
2387
2388         return hctx;
2389
2390  free_bitmap:
2391         sbitmap_free(&hctx->ctx_map);
2392  free_ctxs:
2393         kfree(hctx->ctxs);
2394  free_cpumask:
2395         free_cpumask_var(hctx->cpumask);
2396  free_hctx:
2397         kfree(hctx);
2398  fail_alloc_hctx:
2399         return NULL;
2400 }
2401
2402 static void blk_mq_init_cpu_queues(struct request_queue *q,
2403                                    unsigned int nr_hw_queues)
2404 {
2405         struct blk_mq_tag_set *set = q->tag_set;
2406         unsigned int i, j;
2407
2408         for_each_possible_cpu(i) {
2409                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2410                 struct blk_mq_hw_ctx *hctx;
2411                 int k;
2412
2413                 __ctx->cpu = i;
2414                 spin_lock_init(&__ctx->lock);
2415                 for (k = HCTX_TYPE_DEFAULT; k < HCTX_MAX_TYPES; k++)
2416                         INIT_LIST_HEAD(&__ctx->rq_lists[k]);
2417
2418                 __ctx->queue = q;
2419
2420                 /*
2421                  * Set local node, IFF we have more than one hw queue. If
2422                  * not, we remain on the home node of the device
2423                  */
2424                 for (j = 0; j < set->nr_maps; j++) {
2425                         hctx = blk_mq_map_queue_type(q, j, i);
2426                         if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2427                                 hctx->numa_node = local_memory_node(cpu_to_node(i));
2428                 }
2429         }
2430 }
2431
2432 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2433 {
2434         int ret = 0;
2435
2436         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2437                                         set->queue_depth, set->reserved_tags);
2438         if (!set->tags[hctx_idx])
2439                 return false;
2440
2441         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2442                                 set->queue_depth);
2443         if (!ret)
2444                 return true;
2445
2446         blk_mq_free_rq_map(set->tags[hctx_idx]);
2447         set->tags[hctx_idx] = NULL;
2448         return false;
2449 }
2450
2451 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2452                                          unsigned int hctx_idx)
2453 {
2454         if (set->tags && set->tags[hctx_idx]) {
2455                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2456                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2457                 set->tags[hctx_idx] = NULL;
2458         }
2459 }
2460
2461 static void blk_mq_map_swqueue(struct request_queue *q)
2462 {
2463         unsigned int i, j, hctx_idx;
2464         struct blk_mq_hw_ctx *hctx;
2465         struct blk_mq_ctx *ctx;
2466         struct blk_mq_tag_set *set = q->tag_set;
2467
2468         queue_for_each_hw_ctx(q, hctx, i) {
2469                 cpumask_clear(hctx->cpumask);
2470                 hctx->nr_ctx = 0;
2471                 hctx->dispatch_from = NULL;
2472         }
2473
2474         /*
2475          * Map software to hardware queues.
2476          *
2477          * If the cpu isn't present, the cpu is mapped to first hctx.
2478          */
2479         for_each_possible_cpu(i) {
2480                 hctx_idx = set->map[HCTX_TYPE_DEFAULT].mq_map[i];
2481                 /* unmapped hw queue can be remapped after CPU topo changed */
2482                 if (!set->tags[hctx_idx] &&
2483                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2484                         /*
2485                          * If tags initialization fail for some hctx,
2486                          * that hctx won't be brought online.  In this
2487                          * case, remap the current ctx to hctx[0] which
2488                          * is guaranteed to always have tags allocated
2489                          */
2490                         set->map[HCTX_TYPE_DEFAULT].mq_map[i] = 0;
2491                 }
2492
2493                 ctx = per_cpu_ptr(q->queue_ctx, i);
2494                 for (j = 0; j < set->nr_maps; j++) {
2495                         if (!set->map[j].nr_queues) {
2496                                 ctx->hctxs[j] = blk_mq_map_queue_type(q,
2497                                                 HCTX_TYPE_DEFAULT, i);
2498                                 continue;
2499                         }
2500
2501                         hctx = blk_mq_map_queue_type(q, j, i);
2502                         ctx->hctxs[j] = hctx;
2503                         /*
2504                          * If the CPU is already set in the mask, then we've
2505                          * mapped this one already. This can happen if
2506                          * devices share queues across queue maps.
2507                          */
2508                         if (cpumask_test_cpu(i, hctx->cpumask))
2509                                 continue;
2510
2511                         cpumask_set_cpu(i, hctx->cpumask);
2512                         hctx->type = j;
2513                         ctx->index_hw[hctx->type] = hctx->nr_ctx;
2514                         hctx->ctxs[hctx->nr_ctx++] = ctx;
2515
2516                         /*
2517                          * If the nr_ctx type overflows, we have exceeded the
2518                          * amount of sw queues we can support.
2519                          */
2520                         BUG_ON(!hctx->nr_ctx);
2521                 }
2522
2523                 for (; j < HCTX_MAX_TYPES; j++)
2524                         ctx->hctxs[j] = blk_mq_map_queue_type(q,
2525                                         HCTX_TYPE_DEFAULT, i);
2526         }
2527
2528         queue_for_each_hw_ctx(q, hctx, i) {
2529                 /*
2530                  * If no software queues are mapped to this hardware queue,
2531                  * disable it and free the request entries.
2532                  */
2533                 if (!hctx->nr_ctx) {
2534                         /* Never unmap queue 0.  We need it as a
2535                          * fallback in case of a new remap fails
2536                          * allocation
2537                          */
2538                         if (i && set->tags[i])
2539                                 blk_mq_free_map_and_requests(set, i);
2540
2541                         hctx->tags = NULL;
2542                         continue;
2543                 }
2544
2545                 hctx->tags = set->tags[i];
2546                 WARN_ON(!hctx->tags);
2547
2548                 /*
2549                  * Set the map size to the number of mapped software queues.
2550                  * This is more accurate and more efficient than looping
2551                  * over all possibly mapped software queues.
2552                  */
2553                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2554
2555                 /*
2556                  * Initialize batch roundrobin counts
2557                  */
2558                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2559                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2560         }
2561 }
2562
2563 /*
2564  * Caller needs to ensure that we're either frozen/quiesced, or that
2565  * the queue isn't live yet.
2566  */
2567 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2568 {
2569         struct blk_mq_hw_ctx *hctx;
2570         int i;
2571
2572         queue_for_each_hw_ctx(q, hctx, i) {
2573                 if (shared)
2574                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2575                 else
2576                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2577         }
2578 }
2579
2580 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2581                                         bool shared)
2582 {
2583         struct request_queue *q;
2584
2585         lockdep_assert_held(&set->tag_list_lock);
2586
2587         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2588                 blk_mq_freeze_queue(q);
2589                 queue_set_hctx_shared(q, shared);
2590                 blk_mq_unfreeze_queue(q);
2591         }
2592 }
2593
2594 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2595 {
2596         struct blk_mq_tag_set *set = q->tag_set;
2597
2598         mutex_lock(&set->tag_list_lock);
2599         list_del_rcu(&q->tag_set_list);
2600         if (list_is_singular(&set->tag_list)) {
2601                 /* just transitioned to unshared */
2602                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2603                 /* update existing queue */
2604                 blk_mq_update_tag_set_depth(set, false);
2605         }
2606         mutex_unlock(&set->tag_list_lock);
2607         INIT_LIST_HEAD(&q->tag_set_list);
2608 }
2609
2610 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2611                                      struct request_queue *q)
2612 {
2613         mutex_lock(&set->tag_list_lock);
2614
2615         /*
2616          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2617          */
2618         if (!list_empty(&set->tag_list) &&
2619             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2620                 set->flags |= BLK_MQ_F_TAG_SHARED;
2621                 /* update existing queue */
2622                 blk_mq_update_tag_set_depth(set, true);
2623         }
2624         if (set->flags & BLK_MQ_F_TAG_SHARED)
2625                 queue_set_hctx_shared(q, true);
2626         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2627
2628         mutex_unlock(&set->tag_list_lock);
2629 }
2630
2631 /* All allocations will be freed in release handler of q->mq_kobj */
2632 static int blk_mq_alloc_ctxs(struct request_queue *q)
2633 {
2634         struct blk_mq_ctxs *ctxs;
2635         int cpu;
2636
2637         ctxs = kzalloc(sizeof(*ctxs), GFP_KERNEL);
2638         if (!ctxs)
2639                 return -ENOMEM;
2640
2641         ctxs->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2642         if (!ctxs->queue_ctx)
2643                 goto fail;
2644
2645         for_each_possible_cpu(cpu) {
2646                 struct blk_mq_ctx *ctx = per_cpu_ptr(ctxs->queue_ctx, cpu);
2647                 ctx->ctxs = ctxs;
2648         }
2649
2650         q->mq_kobj = &ctxs->kobj;
2651         q->queue_ctx = ctxs->queue_ctx;
2652
2653         return 0;
2654  fail:
2655         kfree(ctxs);
2656         return -ENOMEM;
2657 }
2658
2659 /*
2660  * It is the actual release handler for mq, but we do it from
2661  * request queue's release handler for avoiding use-after-free
2662  * and headache because q->mq_kobj shouldn't have been introduced,
2663  * but we can't group ctx/kctx kobj without it.
2664  */
2665 void blk_mq_release(struct request_queue *q)
2666 {
2667         struct blk_mq_hw_ctx *hctx, *next;
2668         int i;
2669
2670         cancel_delayed_work_sync(&q->requeue_work);
2671
2672         queue_for_each_hw_ctx(q, hctx, i)
2673                 WARN_ON_ONCE(hctx && list_empty(&hctx->hctx_list));
2674
2675         /* all hctx are in .unused_hctx_list now */
2676         list_for_each_entry_safe(hctx, next, &q->unused_hctx_list, hctx_list) {
2677                 list_del_init(&hctx->hctx_list);
2678                 kobject_put(&hctx->kobj);
2679         }
2680
2681         kfree(q->queue_hw_ctx);
2682
2683         /*
2684          * release .mq_kobj and sw queue's kobject now because
2685          * both share lifetime with request queue.
2686          */
2687         blk_mq_sysfs_deinit(q);
2688 }
2689
2690 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2691 {
2692         struct request_queue *uninit_q, *q;
2693
2694         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
2695         if (!uninit_q)
2696                 return ERR_PTR(-ENOMEM);
2697
2698         q = blk_mq_init_allocated_queue(set, uninit_q);
2699         if (IS_ERR(q))
2700                 blk_cleanup_queue(uninit_q);
2701
2702         return q;
2703 }
2704 EXPORT_SYMBOL(blk_mq_init_queue);
2705
2706 /*
2707  * Helper for setting up a queue with mq ops, given queue depth, and
2708  * the passed in mq ops flags.
2709  */
2710 struct request_queue *blk_mq_init_sq_queue(struct blk_mq_tag_set *set,
2711                                            const struct blk_mq_ops *ops,
2712                                            unsigned int queue_depth,
2713                                            unsigned int set_flags)
2714 {
2715         struct request_queue *q;
2716         int ret;
2717
2718         memset(set, 0, sizeof(*set));
2719         set->ops = ops;
2720         set->nr_hw_queues = 1;
2721         set->nr_maps = 1;
2722         set->queue_depth = queue_depth;
2723         set->numa_node = NUMA_NO_NODE;
2724         set->flags = set_flags;
2725
2726         ret = blk_mq_alloc_tag_set(set);
2727         if (ret)
2728                 return ERR_PTR(ret);
2729
2730         q = blk_mq_init_queue(set);
2731         if (IS_ERR(q)) {
2732                 blk_mq_free_tag_set(set);
2733                 return q;
2734         }
2735
2736         return q;
2737 }
2738 EXPORT_SYMBOL(blk_mq_init_sq_queue);
2739
2740 static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
2741                 struct blk_mq_tag_set *set, struct request_queue *q,
2742                 int hctx_idx, int node)
2743 {
2744         struct blk_mq_hw_ctx *hctx = NULL, *tmp;
2745
2746         /* reuse dead hctx first */
2747         spin_lock(&q->unused_hctx_lock);
2748         list_for_each_entry(tmp, &q->unused_hctx_list, hctx_list) {
2749                 if (tmp->numa_node == node) {
2750                         hctx = tmp;
2751                         break;
2752                 }
2753         }
2754         if (hctx)
2755                 list_del_init(&hctx->hctx_list);
2756         spin_unlock(&q->unused_hctx_lock);
2757
2758         if (!hctx)
2759                 hctx = blk_mq_alloc_hctx(q, set, node);
2760         if (!hctx)
2761                 goto fail;
2762
2763         if (blk_mq_init_hctx(q, set, hctx, hctx_idx))
2764                 goto free_hctx;
2765
2766         return hctx;
2767
2768  free_hctx:
2769         kobject_put(&hctx->kobj);
2770  fail:
2771         return NULL;
2772 }
2773
2774 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2775                                                 struct request_queue *q)
2776 {
2777         int i, j, end;
2778         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2779
2780         /* protect against switching io scheduler  */
2781         mutex_lock(&q->sysfs_lock);
2782         for (i = 0; i < set->nr_hw_queues; i++) {
2783                 int node;
2784                 struct blk_mq_hw_ctx *hctx;
2785
2786                 node = blk_mq_hw_queue_to_node(&set->map[HCTX_TYPE_DEFAULT], i);
2787                 /*
2788                  * If the hw queue has been mapped to another numa node,
2789                  * we need to realloc the hctx. If allocation fails, fallback
2790                  * to use the previous one.
2791                  */
2792                 if (hctxs[i] && (hctxs[i]->numa_node == node))
2793                         continue;
2794
2795                 hctx = blk_mq_alloc_and_init_hctx(set, q, i, node);
2796                 if (hctx) {
2797                         if (hctxs[i])
2798                                 blk_mq_exit_hctx(q, set, hctxs[i], i);
2799                         hctxs[i] = hctx;
2800                 } else {
2801                         if (hctxs[i])
2802                                 pr_warn("Allocate new hctx on node %d fails,\
2803                                                 fallback to previous one on node %d\n",
2804                                                 node, hctxs[i]->numa_node);
2805                         else
2806                                 break;
2807                 }
2808         }
2809         /*
2810          * Increasing nr_hw_queues fails. Free the newly allocated
2811          * hctxs and keep the previous q->nr_hw_queues.
2812          */
2813         if (i != set->nr_hw_queues) {
2814                 j = q->nr_hw_queues;
2815                 end = i;
2816         } else {
2817                 j = i;
2818                 end = q->nr_hw_queues;
2819                 q->nr_hw_queues = set->nr_hw_queues;
2820         }
2821
2822         for (; j < end; j++) {
2823                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2824
2825                 if (hctx) {
2826                         if (hctx->tags)
2827                                 blk_mq_free_map_and_requests(set, j);
2828                         blk_mq_exit_hctx(q, set, hctx, j);
2829                         hctxs[j] = NULL;
2830                 }
2831         }
2832         mutex_unlock(&q->sysfs_lock);
2833 }
2834
2835 /*
2836  * Maximum number of hardware queues we support. For single sets, we'll never
2837  * have more than the CPUs (software queues). For multiple sets, the tag_set
2838  * user may have set ->nr_hw_queues larger.
2839  */
2840 static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
2841 {
2842         if (set->nr_maps == 1)
2843                 return nr_cpu_ids;
2844
2845         return max(set->nr_hw_queues, nr_cpu_ids);
2846 }
2847
2848 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2849                                                   struct request_queue *q)
2850 {
2851         int ret = -ENOMEM;
2852
2853         /* mark the queue as mq asap */
2854         q->mq_ops = set->ops;
2855
2856         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2857                                              blk_mq_poll_stats_bkt,
2858                                              BLK_MQ_POLL_STATS_BKTS, q);
2859         if (!q->poll_cb)
2860                 goto err_exit;
2861
2862         if (blk_mq_alloc_ctxs(q))
2863                 goto err_poll;
2864
2865         /* init q->mq_kobj and sw queues' kobjects */
2866         blk_mq_sysfs_init(q);
2867
2868         q->nr_queues = nr_hw_queues(set);
2869         q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
2870                                                 GFP_KERNEL, set->numa_node);
2871         if (!q->queue_hw_ctx)
2872                 goto err_sys_init;
2873
2874         INIT_LIST_HEAD(&q->unused_hctx_list);
2875         spin_lock_init(&q->unused_hctx_lock);
2876
2877         blk_mq_realloc_hw_ctxs(set, q);
2878         if (!q->nr_hw_queues)
2879                 goto err_hctxs;
2880
2881         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2882         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2883
2884         q->tag_set = set;
2885
2886         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2887         if (set->nr_maps > HCTX_TYPE_POLL &&
2888             set->map[HCTX_TYPE_POLL].nr_queues)
2889                 blk_queue_flag_set(QUEUE_FLAG_POLL, q);
2890
2891         q->sg_reserved_size = INT_MAX;
2892
2893         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2894         INIT_LIST_HEAD(&q->requeue_list);
2895         spin_lock_init(&q->requeue_lock);
2896
2897         blk_queue_make_request(q, blk_mq_make_request);
2898
2899         /*
2900          * Do this after blk_queue_make_request() overrides it...
2901          */
2902         q->nr_requests = set->queue_depth;
2903
2904         /*
2905          * Default to classic polling
2906          */
2907         q->poll_nsec = BLK_MQ_POLL_CLASSIC;
2908
2909         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2910         blk_mq_add_queue_tag_set(set, q);
2911         blk_mq_map_swqueue(q);
2912
2913         ret = elevator_init_mq(q);
2914         if (ret)
2915                 goto err_tag_set;
2916
2917         return q;
2918
2919 err_tag_set:
2920         blk_mq_del_queue_tag_set(q);
2921 err_hctxs:
2922         kfree(q->queue_hw_ctx);
2923         q->nr_hw_queues = 0;
2924 err_sys_init:
2925         blk_mq_sysfs_deinit(q);
2926 err_poll:
2927         blk_stat_free_callback(q->poll_cb);
2928         q->poll_cb = NULL;
2929 err_exit:
2930         q->mq_ops = NULL;
2931         return ERR_PTR(-ENOMEM);
2932 }
2933 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2934
2935 /* tags can _not_ be used after returning from blk_mq_exit_queue */
2936 void blk_mq_exit_queue(struct request_queue *q)
2937 {
2938         struct blk_mq_tag_set   *set = q->tag_set;
2939
2940         blk_mq_del_queue_tag_set(q);
2941         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2942 }
2943
2944 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2945 {
2946         int i;
2947
2948         for (i = 0; i < set->nr_hw_queues; i++)
2949                 if (!__blk_mq_alloc_rq_map(set, i))
2950                         goto out_unwind;
2951
2952         return 0;
2953
2954 out_unwind:
2955         while (--i >= 0)
2956                 blk_mq_free_rq_map(set->tags[i]);
2957
2958         return -ENOMEM;
2959 }
2960
2961 /*
2962  * Allocate the request maps associated with this tag_set. Note that this
2963  * may reduce the depth asked for, if memory is tight. set->queue_depth
2964  * will be updated to reflect the allocated depth.
2965  */
2966 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2967 {
2968         unsigned int depth;
2969         int err;
2970
2971         depth = set->queue_depth;
2972         do {
2973                 err = __blk_mq_alloc_rq_maps(set);
2974                 if (!err)
2975                         break;
2976
2977                 set->queue_depth >>= 1;
2978                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2979                         err = -ENOMEM;
2980                         break;
2981                 }
2982         } while (set->queue_depth);
2983
2984         if (!set->queue_depth || err) {
2985                 pr_err("blk-mq: failed to allocate request map\n");
2986                 return -ENOMEM;
2987         }
2988
2989         if (depth != set->queue_depth)
2990                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2991                                                 depth, set->queue_depth);
2992
2993         return 0;
2994 }
2995
2996 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2997 {
2998         if (set->ops->map_queues && !is_kdump_kernel()) {
2999                 int i;
3000
3001                 /*
3002                  * transport .map_queues is usually done in the following
3003                  * way:
3004                  *
3005                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
3006                  *      mask = get_cpu_mask(queue)
3007                  *      for_each_cpu(cpu, mask)
3008                  *              set->map[x].mq_map[cpu] = queue;
3009                  * }
3010                  *
3011                  * When we need to remap, the table has to be cleared for
3012                  * killing stale mapping since one CPU may not be mapped
3013                  * to any hw queue.
3014                  */
3015                 for (i = 0; i < set->nr_maps; i++)
3016                         blk_mq_clear_mq_map(&set->map[i]);
3017
3018                 return set->ops->map_queues(set);
3019         } else {
3020                 BUG_ON(set->nr_maps > 1);
3021                 return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3022         }
3023 }
3024
3025 /*
3026  * Alloc a tag set to be associated with one or more request queues.
3027  * May fail with EINVAL for various error conditions. May adjust the
3028  * requested depth down, if it's too large. In that case, the set
3029  * value will be stored in set->queue_depth.
3030  */
3031 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
3032 {
3033         int i, ret;
3034
3035         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
3036
3037         if (!set->nr_hw_queues)
3038                 return -EINVAL;
3039         if (!set->queue_depth)
3040                 return -EINVAL;
3041         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
3042                 return -EINVAL;
3043
3044         if (!set->ops->queue_rq)
3045                 return -EINVAL;
3046
3047         if (!set->ops->get_budget ^ !set->ops->put_budget)
3048                 return -EINVAL;
3049
3050         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
3051                 pr_info("blk-mq: reduced tag depth to %u\n",
3052                         BLK_MQ_MAX_DEPTH);
3053                 set->queue_depth = BLK_MQ_MAX_DEPTH;
3054         }
3055
3056         if (!set->nr_maps)
3057                 set->nr_maps = 1;
3058         else if (set->nr_maps > HCTX_MAX_TYPES)
3059                 return -EINVAL;
3060
3061         /*
3062          * If a crashdump is active, then we are potentially in a very
3063          * memory constrained environment. Limit us to 1 queue and
3064          * 64 tags to prevent using too much memory.
3065          */
3066         if (is_kdump_kernel()) {
3067                 set->nr_hw_queues = 1;
3068                 set->nr_maps = 1;
3069                 set->queue_depth = min(64U, set->queue_depth);
3070         }
3071         /*
3072          * There is no use for more h/w queues than cpus if we just have
3073          * a single map
3074          */
3075         if (set->nr_maps == 1 && set->nr_hw_queues > nr_cpu_ids)
3076                 set->nr_hw_queues = nr_cpu_ids;
3077
3078         set->tags = kcalloc_node(nr_hw_queues(set), sizeof(struct blk_mq_tags *),
3079                                  GFP_KERNEL, set->numa_node);
3080         if (!set->tags)
3081                 return -ENOMEM;
3082
3083         ret = -ENOMEM;
3084         for (i = 0; i < set->nr_maps; i++) {
3085                 set->map[i].mq_map = kcalloc_node(nr_cpu_ids,
3086                                                   sizeof(set->map[i].mq_map[0]),
3087                                                   GFP_KERNEL, set->numa_node);
3088                 if (!set->map[i].mq_map)
3089                         goto out_free_mq_map;
3090                 set->map[i].nr_queues = is_kdump_kernel() ? 1 : set->nr_hw_queues;
3091         }
3092
3093         ret = blk_mq_update_queue_map(set);
3094         if (ret)
3095                 goto out_free_mq_map;
3096
3097         ret = blk_mq_alloc_rq_maps(set);
3098         if (ret)
3099                 goto out_free_mq_map;
3100
3101         mutex_init(&set->tag_list_lock);
3102         INIT_LIST_HEAD(&set->tag_list);
3103
3104         return 0;
3105
3106 out_free_mq_map:
3107         for (i = 0; i < set->nr_maps; i++) {
3108                 kfree(set->map[i].mq_map);
3109                 set->map[i].mq_map = NULL;
3110         }
3111         kfree(set->tags);
3112         set->tags = NULL;
3113         return ret;
3114 }
3115 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
3116
3117 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
3118 {
3119         int i, j;
3120
3121         for (i = 0; i < nr_hw_queues(set); i++)
3122                 blk_mq_free_map_and_requests(set, i);
3123
3124         for (j = 0; j < set->nr_maps; j++) {
3125                 kfree(set->map[j].mq_map);
3126                 set->map[j].mq_map = NULL;
3127         }
3128
3129         kfree(set->tags);
3130         set->tags = NULL;
3131 }
3132 EXPORT_SYMBOL(blk_mq_free_tag_set);
3133
3134 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
3135 {
3136         struct blk_mq_tag_set *set = q->tag_set;
3137         struct blk_mq_hw_ctx *hctx;
3138         int i, ret;
3139
3140         if (!set)
3141                 return -EINVAL;
3142
3143         if (q->nr_requests == nr)
3144                 return 0;
3145
3146         blk_mq_freeze_queue(q);
3147         blk_mq_quiesce_queue(q);
3148
3149         ret = 0;
3150         queue_for_each_hw_ctx(q, hctx, i) {
3151                 if (!hctx->tags)
3152                         continue;
3153                 /*
3154                  * If we're using an MQ scheduler, just update the scheduler
3155                  * queue depth. This is similar to what the old code would do.
3156                  */
3157                 if (!hctx->sched_tags) {
3158                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
3159                                                         false);
3160                 } else {
3161                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
3162                                                         nr, true);
3163                 }
3164                 if (ret)
3165                         break;
3166                 if (q->elevator && q->elevator->type->ops.depth_updated)
3167                         q->elevator->type->ops.depth_updated(hctx);
3168         }
3169
3170         if (!ret)
3171                 q->nr_requests = nr;
3172
3173         blk_mq_unquiesce_queue(q);
3174         blk_mq_unfreeze_queue(q);
3175
3176         return ret;
3177 }
3178
3179 /*
3180  * request_queue and elevator_type pair.
3181  * It is just used by __blk_mq_update_nr_hw_queues to cache
3182  * the elevator_type associated with a request_queue.
3183  */
3184 struct blk_mq_qe_pair {
3185         struct list_head node;
3186         struct request_queue *q;
3187         struct elevator_type *type;
3188 };
3189
3190 /*
3191  * Cache the elevator_type in qe pair list and switch the
3192  * io scheduler to 'none'
3193  */
3194 static bool blk_mq_elv_switch_none(struct list_head *head,
3195                 struct request_queue *q)
3196 {
3197         struct blk_mq_qe_pair *qe;
3198
3199         if (!q->elevator)
3200                 return true;
3201
3202         qe = kmalloc(sizeof(*qe), GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY);
3203         if (!qe)
3204                 return false;
3205
3206         INIT_LIST_HEAD(&qe->node);
3207         qe->q = q;
3208         qe->type = q->elevator->type;
3209         list_add(&qe->node, head);
3210
3211         mutex_lock(&q->sysfs_lock);
3212         /*
3213          * After elevator_switch_mq, the previous elevator_queue will be
3214          * released by elevator_release. The reference of the io scheduler
3215          * module get by elevator_get will also be put. So we need to get
3216          * a reference of the io scheduler module here to prevent it to be
3217          * removed.
3218          */
3219         __module_get(qe->type->elevator_owner);
3220         elevator_switch_mq(q, NULL);
3221         mutex_unlock(&q->sysfs_lock);
3222
3223         return true;
3224 }
3225
3226 static void blk_mq_elv_switch_back(struct list_head *head,
3227                 struct request_queue *q)
3228 {
3229         struct blk_mq_qe_pair *qe;
3230         struct elevator_type *t = NULL;
3231
3232         list_for_each_entry(qe, head, node)
3233                 if (qe->q == q) {
3234                         t = qe->type;
3235                         break;
3236                 }
3237
3238         if (!t)
3239                 return;
3240
3241         list_del(&qe->node);
3242         kfree(qe);
3243
3244         mutex_lock(&q->sysfs_lock);
3245         elevator_switch_mq(q, t);
3246         mutex_unlock(&q->sysfs_lock);
3247 }
3248
3249 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
3250                                                         int nr_hw_queues)
3251 {
3252         struct request_queue *q;
3253         LIST_HEAD(head);
3254         int prev_nr_hw_queues;
3255
3256         lockdep_assert_held(&set->tag_list_lock);
3257
3258         if (set->nr_maps == 1 && nr_hw_queues > nr_cpu_ids)
3259                 nr_hw_queues = nr_cpu_ids;
3260         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
3261                 return;
3262
3263         list_for_each_entry(q, &set->tag_list, tag_set_list)
3264                 blk_mq_freeze_queue(q);
3265         /*
3266          * Sync with blk_mq_queue_tag_busy_iter.
3267          */
3268         synchronize_rcu();
3269         /*
3270          * Switch IO scheduler to 'none', cleaning up the data associated
3271          * with the previous scheduler. We will switch back once we are done
3272          * updating the new sw to hw queue mappings.
3273          */
3274         list_for_each_entry(q, &set->tag_list, tag_set_list)
3275                 if (!blk_mq_elv_switch_none(&head, q))
3276                         goto switch_back;
3277
3278         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3279                 blk_mq_debugfs_unregister_hctxs(q);
3280                 blk_mq_sysfs_unregister(q);
3281         }
3282
3283         prev_nr_hw_queues = set->nr_hw_queues;
3284         set->nr_hw_queues = nr_hw_queues;
3285         blk_mq_update_queue_map(set);
3286 fallback:
3287         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3288                 blk_mq_realloc_hw_ctxs(set, q);
3289                 if (q->nr_hw_queues != set->nr_hw_queues) {
3290                         pr_warn("Increasing nr_hw_queues to %d fails, fallback to %d\n",
3291                                         nr_hw_queues, prev_nr_hw_queues);
3292                         set->nr_hw_queues = prev_nr_hw_queues;
3293                         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
3294                         goto fallback;
3295                 }
3296                 blk_mq_map_swqueue(q);
3297         }
3298
3299         list_for_each_entry(q, &set->tag_list, tag_set_list) {
3300                 blk_mq_sysfs_register(q);
3301                 blk_mq_debugfs_register_hctxs(q);
3302         }
3303
3304 switch_back:
3305         list_for_each_entry(q, &set->tag_list, tag_set_list)
3306                 blk_mq_elv_switch_back(&head, q);
3307
3308         list_for_each_entry(q, &set->tag_list, tag_set_list)
3309                 blk_mq_unfreeze_queue(q);
3310 }
3311
3312 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
3313 {
3314         mutex_lock(&set->tag_list_lock);
3315         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
3316         mutex_unlock(&set->tag_list_lock);
3317 }
3318 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
3319
3320 /* Enable polling stats and return whether they were already enabled. */
3321 static bool blk_poll_stats_enable(struct request_queue *q)
3322 {
3323         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3324             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
3325                 return true;
3326         blk_stat_add_callback(q, q->poll_cb);
3327         return false;
3328 }
3329
3330 static void blk_mq_poll_stats_start(struct request_queue *q)
3331 {
3332         /*
3333          * We don't arm the callback if polling stats are not enabled or the
3334          * callback is already active.
3335          */
3336         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
3337             blk_stat_is_active(q->poll_cb))
3338                 return;
3339
3340         blk_stat_activate_msecs(q->poll_cb, 100);
3341 }
3342
3343 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
3344 {
3345         struct request_queue *q = cb->data;
3346         int bucket;
3347
3348         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
3349                 if (cb->stat[bucket].nr_samples)
3350                         q->poll_stat[bucket] = cb->stat[bucket];
3351         }
3352 }
3353
3354 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
3355                                        struct blk_mq_hw_ctx *hctx,
3356                                        struct request *rq)
3357 {
3358         unsigned long ret = 0;
3359         int bucket;
3360
3361         /*
3362          * If stats collection isn't on, don't sleep but turn it on for
3363          * future users
3364          */
3365         if (!blk_poll_stats_enable(q))
3366                 return 0;
3367
3368         /*
3369          * As an optimistic guess, use half of the mean service time
3370          * for this type of request. We can (and should) make this smarter.
3371          * For instance, if the completion latencies are tight, we can
3372          * get closer than just half the mean. This is especially
3373          * important on devices where the completion latencies are longer
3374          * than ~10 usec. We do use the stats for the relevant IO size
3375          * if available which does lead to better estimates.
3376          */
3377         bucket = blk_mq_poll_stats_bkt(rq);
3378         if (bucket < 0)
3379                 return ret;
3380
3381         if (q->poll_stat[bucket].nr_samples)
3382                 ret = (q->poll_stat[bucket].mean + 1) / 2;
3383
3384         return ret;
3385 }
3386
3387 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
3388                                      struct blk_mq_hw_ctx *hctx,
3389                                      struct request *rq)
3390 {
3391         struct hrtimer_sleeper hs;
3392         enum hrtimer_mode mode;
3393         unsigned int nsecs;
3394         ktime_t kt;
3395
3396         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
3397                 return false;
3398
3399         /*
3400          * If we get here, hybrid polling is enabled. Hence poll_nsec can be:
3401          *
3402          *  0:  use half of prev avg
3403          * >0:  use this specific value
3404          */
3405         if (q->poll_nsec > 0)
3406                 nsecs = q->poll_nsec;
3407         else
3408                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
3409
3410         if (!nsecs)
3411                 return false;
3412
3413         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
3414
3415         /*
3416          * This will be replaced with the stats tracking code, using
3417          * 'avg_completion_time / 2' as the pre-sleep target.
3418          */
3419         kt = nsecs;
3420
3421         mode = HRTIMER_MODE_REL;
3422         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
3423         hrtimer_set_expires(&hs.timer, kt);
3424
3425         hrtimer_init_sleeper(&hs, current);
3426         do {
3427                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
3428                         break;
3429                 set_current_state(TASK_UNINTERRUPTIBLE);
3430                 hrtimer_start_expires(&hs.timer, mode);
3431                 if (hs.task)
3432                         io_schedule();
3433                 hrtimer_cancel(&hs.timer);
3434                 mode = HRTIMER_MODE_ABS;
3435         } while (hs.task && !signal_pending(current));
3436
3437         __set_current_state(TASK_RUNNING);
3438         destroy_hrtimer_on_stack(&hs.timer);
3439         return true;
3440 }
3441
3442 static bool blk_mq_poll_hybrid(struct request_queue *q,
3443                                struct blk_mq_hw_ctx *hctx, blk_qc_t cookie)
3444 {
3445         struct request *rq;
3446
3447         if (q->poll_nsec == BLK_MQ_POLL_CLASSIC)
3448                 return false;
3449
3450         if (!blk_qc_t_is_internal(cookie))
3451                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3452         else {
3453                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3454                 /*
3455                  * With scheduling, if the request has completed, we'll
3456                  * get a NULL return here, as we clear the sched tag when
3457                  * that happens. The request still remains valid, like always,
3458                  * so we should be safe with just the NULL check.
3459                  */
3460                 if (!rq)
3461                         return false;
3462         }
3463
3464         return blk_mq_poll_hybrid_sleep(q, hctx, rq);
3465 }
3466
3467 /**
3468  * blk_poll - poll for IO completions
3469  * @q:  the queue
3470  * @cookie: cookie passed back at IO submission time
3471  * @spin: whether to spin for completions
3472  *
3473  * Description:
3474  *    Poll for completions on the passed in queue. Returns number of
3475  *    completed entries found. If @spin is true, then blk_poll will continue
3476  *    looping until at least one completion is found, unless the task is
3477  *    otherwise marked running (or we need to reschedule).
3478  */
3479 int blk_poll(struct request_queue *q, blk_qc_t cookie, bool spin)
3480 {
3481         struct blk_mq_hw_ctx *hctx;
3482         long state;
3483
3484         if (!blk_qc_t_valid(cookie) ||
3485             !test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3486                 return 0;
3487
3488         if (current->plug)
3489                 blk_flush_plug_list(current->plug, false);
3490
3491         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3492
3493         /*
3494          * If we sleep, have the caller restart the poll loop to reset
3495          * the state. Like for the other success return cases, the
3496          * caller is responsible for checking if the IO completed. If
3497          * the IO isn't complete, we'll get called again and will go
3498          * straight to the busy poll loop.
3499          */
3500         if (blk_mq_poll_hybrid(q, hctx, cookie))
3501                 return 1;
3502
3503         hctx->poll_considered++;
3504
3505         state = current->state;
3506         do {
3507                 int ret;
3508
3509                 hctx->poll_invoked++;
3510
3511                 ret = q->mq_ops->poll(hctx);
3512                 if (ret > 0) {
3513                         hctx->poll_success++;
3514                         __set_current_state(TASK_RUNNING);
3515                         return ret;
3516                 }
3517
3518                 if (signal_pending_state(state, current))
3519                         __set_current_state(TASK_RUNNING);
3520
3521                 if (current->state == TASK_RUNNING)
3522                         return 1;
3523                 if (ret < 0 || !spin)
3524                         break;
3525                 cpu_relax();
3526         } while (!need_resched());
3527
3528         __set_current_state(TASK_RUNNING);
3529         return 0;
3530 }
3531 EXPORT_SYMBOL_GPL(blk_poll);
3532
3533 unsigned int blk_mq_rq_cpu(struct request *rq)
3534 {
3535         return rq->mq_ctx->cpu;
3536 }
3537 EXPORT_SYMBOL(blk_mq_rq_cpu);
3538
3539 static int __init blk_mq_init(void)
3540 {
3541         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3542                                 blk_mq_hctx_notify_dead);
3543         return 0;
3544 }
3545 subsys_initcall(blk_mq_init);