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