block/blk-mq.c

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