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                 rq->rq_flags &= ~RQF_TIMED_OUT;
 675                 if (q->dma_drain_size && blk_rq_bytes(rq))
 676                         rq->nr_phys_segments--;
 677         }
 678 }
 679
 680 void blk_mq_requeue_request(struct request *rq, bool kick_requeue_list)
 681 {
 682         __blk_mq_requeue_request(rq);
 683
 684         /* this request will be re-inserted to io scheduler queue */
 685         blk_mq_sched_requeue_request(rq);
 686
 687         BUG_ON(blk_queued_rq(rq));
 688         blk_mq_add_to_requeue_list(rq, true, kick_requeue_list);
 689 }
 690 EXPORT_SYMBOL(blk_mq_requeue_request);
 691
 692 static void blk_mq_requeue_work(struct work_struct *work)
 693 {
 694         struct request_queue *q =
 695                 container_of(work, struct request_queue, requeue_work.work);
 696         LIST_HEAD(rq_list);
 697         struct request *rq, *next;
 698
 699         spin_lock_irq(&q->requeue_lock);
 700         list_splice_init(&q->requeue_list, &rq_list);
 701         spin_unlock_irq(&q->requeue_lock);
 702
 703         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 704                 if (!(rq->rq_flags & RQF_SOFTBARRIER))
 705                         continue;
 706
 707                 rq->rq_flags &= ~RQF_SOFTBARRIER;
 708                 list_del_init(&rq->queuelist);
 709                 blk_mq_sched_insert_request(rq, true, false, false);
 710         }
 711
 712         while (!list_empty(&rq_list)) {
 713                 rq = list_entry(rq_list.next, struct request, queuelist);
 714                 list_del_init(&rq->queuelist);
 715                 blk_mq_sched_insert_request(rq, false, false, false);
 716         }
 717
 718         blk_mq_run_hw_queues(q, false);
 719 }
 720
 721 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head,
 722                                 bool kick_requeue_list)
 723 {
 724         struct request_queue *q = rq->q;
 725         unsigned long flags;
 726
 727         /*
 728          * We abuse this flag that is otherwise used by the I/O scheduler to
 729          * request head insertion from the workqueue.
 730          */
 731         BUG_ON(rq->rq_flags & RQF_SOFTBARRIER);
 732
 733         spin_lock_irqsave(&q->requeue_lock, flags);
 734         if (at_head) {
 735                 rq->rq_flags |= RQF_SOFTBARRIER;
 736                 list_add(&rq->queuelist, &q->requeue_list);
 737         } else {
 738                 list_add_tail(&rq->queuelist, &q->requeue_list);
 739         }
 740         spin_unlock_irqrestore(&q->requeue_lock, flags);
 741
 742         if (kick_requeue_list)
 743                 blk_mq_kick_requeue_list(q);
 744 }
 745 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
 746
 747 void blk_mq_kick_requeue_list(struct request_queue *q)
 748 {
 749         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work, 0);
 750 }
 751 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
 752
 753 void blk_mq_delay_kick_requeue_list(struct request_queue *q,
 754                                     unsigned long msecs)
 755 {
 756         kblockd_mod_delayed_work_on(WORK_CPU_UNBOUND, &q->requeue_work,
 757                                     msecs_to_jiffies(msecs));
 758 }
 759 EXPORT_SYMBOL(blk_mq_delay_kick_requeue_list);
 760
 761 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 762 {
 763         if (tag < tags->nr_tags) {
 764                 prefetch(tags->rqs[tag]);
 765                 return tags->rqs[tag];
 766         }
 767
 768         return NULL;
 769 }
 770 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 771
 772 static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 773 {
 774         req->rq_flags |= RQF_TIMED_OUT;
 775         if (req->q->mq_ops->timeout) {
 776                 enum blk_eh_timer_return ret;
 777
 778                 ret = req->q->mq_ops->timeout(req, reserved);
 779                 if (ret == BLK_EH_DONE)
 780                         return;
 781                 WARN_ON_ONCE(ret != BLK_EH_RESET_TIMER);
 782         }
 783
 784         blk_add_timer(req);
 785 }
 786
 787 static bool blk_mq_req_expired(struct request *rq, unsigned long *next)
 788 {
 789         unsigned long deadline;
 790
 791         if (blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 792                 return false;
 793         if (rq->rq_flags & RQF_TIMED_OUT)
 794                 return false;
 795
 796         deadline = blk_rq_deadline(rq);
 797         if (time_after_eq(jiffies, deadline))
 798                 return true;
 799
 800         if (*next == 0)
 801                 *next = deadline;
 802         else if (time_after(*next, deadline))
 803                 *next = deadline;
 804         return false;
 805 }
 806
 807 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 808                 struct request *rq, void *priv, bool reserved)
 809 {
 810         unsigned long *next = priv;
 811
 812         /*
 813          * Just do a quick check if it is expired before locking the request in
 814          * so we're not unnecessarilly synchronizing across CPUs.
 815          */
 816         if (!blk_mq_req_expired(rq, next))
 817                 return;
 818
 819         /*
 820          * We have reason to believe the request may be expired. Take a
 821          * reference on the request to lock this request lifetime into its
 822          * currently allocated context to prevent it from being reallocated in
 823          * the event the completion by-passes this timeout handler.
 824          *
 825          * If the reference was already released, then the driver beat the
 826          * timeout handler to posting a natural completion.
 827          */
 828         if (!refcount_inc_not_zero(&rq->ref))
 829                 return;
 830
 831         /*
 832          * The request is now locked and cannot be reallocated underneath the
 833          * timeout handler's processing. Re-verify this exact request is truly
 834          * expired; if it is not expired, then the request was completed and
 835          * reallocated as a new request.
 836          */
 837         if (blk_mq_req_expired(rq, next))
 838                 blk_mq_rq_timed_out(rq, reserved);
 839         if (refcount_dec_and_test(&rq->ref))
 840                 __blk_mq_free_request(rq);
 841 }
 842
 843 static void blk_mq_timeout_work(struct work_struct *work)
 844 {
 845         struct request_queue *q =
 846                 container_of(work, struct request_queue, timeout_work);
 847         unsigned long next = 0;
 848         struct blk_mq_hw_ctx *hctx;
 849         int i;
 850
 851         /* A deadlock might occur if a request is stuck requiring a
 852          * timeout at the same time a queue freeze is waiting
 853          * completion, since the timeout code would not be able to
 854          * acquire the queue reference here.
 855          *
 856          * That's why we don't use blk_queue_enter here; instead, we use
 857          * percpu_ref_tryget directly, because we need to be able to
 858          * obtain a reference even in the short window between the queue
 859          * starting to freeze, by dropping the first reference in
 860          * blk_freeze_queue_start, and the moment the last request is
 861          * consumed, marked by the instant q_usage_counter reaches
 862          * zero.
 863          */
 864         if (!percpu_ref_tryget(&q->q_usage_counter))
 865                 return;
 866
 867         blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &next);
 868
 869         if (next != 0) {
 870                 mod_timer(&q->timeout, next);
 871         } else {
 872                 /*
 873                  * Request timeouts are handled as a forward rolling timer. If
 874                  * we end up here it means that no requests are pending and
 875                  * also that no request has been pending for a while. Mark
 876                  * each hctx as idle.
 877                  */
 878                 queue_for_each_hw_ctx(q, hctx, i) {
 879                         /* the hctx may be unmapped, so check it here */
 880                         if (blk_mq_hw_queue_mapped(hctx))
 881                                 blk_mq_tag_idle(hctx);
 882                 }
 883         }
 884         blk_queue_exit(q);
 885 }
 886
 887 struct flush_busy_ctx_data {
 888         struct blk_mq_hw_ctx *hctx;
 889         struct list_head *list;
 890 };
 891
 892 static bool flush_busy_ctx(struct sbitmap *sb, unsigned int bitnr, void *data)
 893 {
 894         struct flush_busy_ctx_data *flush_data = data;
 895         struct blk_mq_hw_ctx *hctx = flush_data->hctx;
 896         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 897
 898         spin_lock(&ctx->lock);
 899         list_splice_tail_init(&ctx->rq_list, flush_data->list);
 900         sbitmap_clear_bit(sb, bitnr);
 901         spin_unlock(&ctx->lock);
 902         return true;
 903 }
 904
 905 /*
 906  * Process software queues that have been marked busy, splicing them
 907  * to the for-dispatch
 908  */
 909 void blk_mq_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
 910 {
 911         struct flush_busy_ctx_data data = {
 912                 .hctx = hctx,
 913                 .list = list,
 914         };
 915
 916         sbitmap_for_each_set(&hctx->ctx_map, flush_busy_ctx, &data);
 917 }
 918 EXPORT_SYMBOL_GPL(blk_mq_flush_busy_ctxs);
 919
 920 struct dispatch_rq_data {
 921         struct blk_mq_hw_ctx *hctx;
 922         struct request *rq;
 923 };
 924
 925 static bool dispatch_rq_from_ctx(struct sbitmap *sb, unsigned int bitnr,
 926                 void *data)
 927 {
 928         struct dispatch_rq_data *dispatch_data = data;
 929         struct blk_mq_hw_ctx *hctx = dispatch_data->hctx;
 930         struct blk_mq_ctx *ctx = hctx->ctxs[bitnr];
 931
 932         spin_lock(&ctx->lock);
 933         if (!list_empty(&ctx->rq_list)) {
 934                 dispatch_data->rq = list_entry_rq(ctx->rq_list.next);
 935                 list_del_init(&dispatch_data->rq->queuelist);
 936                 if (list_empty(&ctx->rq_list))
 937                         sbitmap_clear_bit(sb, bitnr);
 938         }
 939         spin_unlock(&ctx->lock);
 940
 941         return !dispatch_data->rq;
 942 }
 943
 944 struct request *blk_mq_dequeue_from_ctx(struct blk_mq_hw_ctx *hctx,
 945                                         struct blk_mq_ctx *start)
 946 {
 947         unsigned off = start ? start->index_hw : 0;
 948         struct dispatch_rq_data data = {
 949                 .hctx = hctx,
 950                 .rq   = NULL,
 951         };
 952
 953         __sbitmap_for_each_set(&hctx->ctx_map, off,
 954                                dispatch_rq_from_ctx, &data);
 955
 956         return data.rq;
 957 }
 958
 959 static inline unsigned int queued_to_index(unsigned int queued)
 960 {
 961         if (!queued)
 962                 return 0;
 963
 964         return min(BLK_MQ_MAX_DISPATCH_ORDER - 1, ilog2(queued) + 1);
 965 }
 966
 967 bool blk_mq_get_driver_tag(struct request *rq, struct blk_mq_hw_ctx **hctx,
 968                            bool wait)
 969 {
 970         struct blk_mq_alloc_data data = {
 971                 .q = rq->q,
 972                 .hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu),
 973                 .flags = wait ? 0 : BLK_MQ_REQ_NOWAIT,
 974         };
 975
 976         might_sleep_if(wait);
 977
 978         if (rq->tag != -1)
 979                 goto done;
 980
 981         if (blk_mq_tag_is_reserved(data.hctx->sched_tags, rq->internal_tag))
 982                 data.flags |= BLK_MQ_REQ_RESERVED;
 983
 984         rq->tag = blk_mq_get_tag(&data);
 985         if (rq->tag >= 0) {
 986                 if (blk_mq_tag_busy(data.hctx)) {
 987                         rq->rq_flags |= RQF_MQ_INFLIGHT;
 988                         atomic_inc(&data.hctx->nr_active);
 989                 }
 990                 data.hctx->tags->rqs[rq->tag] = rq;
 991         }
 992
 993 done:
 994         if (hctx)
 995                 *hctx = data.hctx;
 996         return rq->tag != -1;
 997 }
 998
 999 static int blk_mq_dispatch_wake(wait_queue_entry_t *wait, unsigned mode,
1000                                 int flags, void *key)
1001 {
1002         struct blk_mq_hw_ctx *hctx;
1003
1004         hctx = container_of(wait, struct blk_mq_hw_ctx, dispatch_wait);
1005
1006         list_del_init(&wait->entry);
1007         blk_mq_run_hw_queue(hctx, true);
1008         return 1;
1009 }
1010
1011 /*
1012  * Mark us waiting for a tag. For shared tags, this involves hooking us into
1013  * the tag wakeups. For non-shared tags, we can simply mark us needing a
1014  * restart. For both cases, take care to check the condition again after
1015  * marking us as waiting.
1016  */
1017 static bool blk_mq_mark_tag_wait(struct blk_mq_hw_ctx **hctx,
1018                                  struct request *rq)
1019 {
1020         struct blk_mq_hw_ctx *this_hctx = *hctx;
1021         struct sbq_wait_state *ws;
1022         wait_queue_entry_t *wait;
1023         bool ret;
1024
1025         if (!(this_hctx->flags & BLK_MQ_F_TAG_SHARED)) {
1026                 if (!test_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state))
1027                         set_bit(BLK_MQ_S_SCHED_RESTART, &this_hctx->state);
1028
1029                 /*
1030                  * It's possible that a tag was freed in the window between the
1031                  * allocation failure and adding the hardware queue to the wait
1032                  * queue.
1033                  *
1034                  * Don't clear RESTART here, someone else could have set it.
1035                  * At most this will cost an extra queue run.
1036                  */
1037                 return blk_mq_get_driver_tag(rq, hctx, false);
1038         }
1039
1040         wait = &this_hctx->dispatch_wait;
1041         if (!list_empty_careful(&wait->entry))
1042                 return false;
1043
1044         spin_lock(&this_hctx->lock);
1045         if (!list_empty(&wait->entry)) {
1046                 spin_unlock(&this_hctx->lock);
1047                 return false;
1048         }
1049
1050         ws = bt_wait_ptr(&this_hctx->tags->bitmap_tags, this_hctx);
1051         add_wait_queue(&ws->wait, wait);
1052
1053         /*
1054          * It's possible that a tag was freed in the window between the
1055          * allocation failure and adding the hardware queue to the wait
1056          * queue.
1057          */
1058         ret = blk_mq_get_driver_tag(rq, hctx, false);
1059         if (!ret) {
1060                 spin_unlock(&this_hctx->lock);
1061                 return false;
1062         }
1063
1064         /*
1065          * We got a tag, remove ourselves from the wait queue to ensure
1066          * someone else gets the wakeup.
1067          */
1068         spin_lock_irq(&ws->wait.lock);
1069         list_del_init(&wait->entry);
1070         spin_unlock_irq(&ws->wait.lock);
1071         spin_unlock(&this_hctx->lock);
1072
1073         return true;
1074 }
1075
1076 #define BLK_MQ_RESOURCE_DELAY   3               /* ms units */
1077
1078 /*
1079  * Returns true if we did some work AND can potentially do more.
1080  */
1081 bool blk_mq_dispatch_rq_list(struct request_queue *q, struct list_head *list,
1082                              bool got_budget)
1083 {
1084         struct blk_mq_hw_ctx *hctx;
1085         struct request *rq, *nxt;
1086         bool no_tag = false;
1087         int errors, queued;
1088         blk_status_t ret = BLK_STS_OK;
1089
1090         if (list_empty(list))
1091                 return false;
1092
1093         WARN_ON(!list_is_singular(list) && got_budget);
1094
1095         /*
1096          * Now process all the entries, sending them to the driver.
1097          */
1098         errors = queued = 0;
1099         do {
1100                 struct blk_mq_queue_data bd;
1101
1102                 rq = list_first_entry(list, struct request, queuelist);
1103
1104                 hctx = blk_mq_map_queue(rq->q, rq->mq_ctx->cpu);
1105                 if (!got_budget && !blk_mq_get_dispatch_budget(hctx))
1106                         break;
1107
1108                 if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1109                         /*
1110                          * The initial allocation attempt failed, so we need to
1111                          * rerun the hardware queue when a tag is freed. The
1112                          * waitqueue takes care of that. If the queue is run
1113                          * before we add this entry back on the dispatch list,
1114                          * we'll re-run it below.
1115                          */
1116                         if (!blk_mq_mark_tag_wait(&hctx, rq)) {
1117                                 blk_mq_put_dispatch_budget(hctx);
1118                                 /*
1119                                  * For non-shared tags, the RESTART check
1120                                  * will suffice.
1121                                  */
1122                                 if (hctx->flags & BLK_MQ_F_TAG_SHARED)
1123                                         no_tag = true;
1124                                 break;
1125                         }
1126                 }
1127
1128                 list_del_init(&rq->queuelist);
1129
1130                 bd.rq = rq;
1131
1132                 /*
1133                  * Flag last if we have no more requests, or if we have more
1134                  * but can't assign a driver tag to it.
1135                  */
1136                 if (list_empty(list))
1137                         bd.last = true;
1138                 else {
1139                         nxt = list_first_entry(list, struct request, queuelist);
1140                         bd.last = !blk_mq_get_driver_tag(nxt, NULL, false);
1141                 }
1142
1143                 ret = q->mq_ops->queue_rq(hctx, &bd);
1144                 if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE) {
1145                         /*
1146                          * If an I/O scheduler has been configured and we got a
1147                          * driver tag for the next request already, free it
1148                          * again.
1149                          */
1150                         if (!list_empty(list)) {
1151                                 nxt = list_first_entry(list, struct request, queuelist);
1152                                 blk_mq_put_driver_tag(nxt);
1153                         }
1154                         list_add(&rq->queuelist, list);
1155                         __blk_mq_requeue_request(rq);
1156                         break;
1157                 }
1158
1159                 if (unlikely(ret != BLK_STS_OK)) {
1160                         errors++;
1161                         blk_mq_end_request(rq, BLK_STS_IOERR);
1162                         continue;
1163                 }
1164
1165                 queued++;
1166         } while (!list_empty(list));
1167
1168         hctx->dispatched[queued_to_index(queued)]++;
1169
1170         /*
1171          * Any items that need requeuing? Stuff them into hctx->dispatch,
1172          * that is where we will continue on next queue run.
1173          */
1174         if (!list_empty(list)) {
1175                 bool needs_restart;
1176
1177                 spin_lock(&hctx->lock);
1178                 list_splice_init(list, &hctx->dispatch);
1179                 spin_unlock(&hctx->lock);
1180
1181                 /*
1182                  * If SCHED_RESTART was set by the caller of this function and
1183                  * it is no longer set that means that it was cleared by another
1184                  * thread and hence that a queue rerun is needed.
1185                  *
1186                  * If 'no_tag' is set, that means that we failed getting
1187                  * a driver tag with an I/O scheduler attached. If our dispatch
1188                  * waitqueue is no longer active, ensure that we run the queue
1189                  * AFTER adding our entries back to the list.
1190                  *
1191                  * If no I/O scheduler has been configured it is possible that
1192                  * the hardware queue got stopped and restarted before requests
1193                  * were pushed back onto the dispatch list. Rerun the queue to
1194                  * avoid starvation. Notes:
1195                  * - blk_mq_run_hw_queue() checks whether or not a queue has
1196                  *   been stopped before rerunning a queue.
1197                  * - Some but not all block drivers stop a queue before
1198                  *   returning BLK_STS_RESOURCE. Two exceptions are scsi-mq
1199                  *   and dm-rq.
1200                  *
1201                  * If driver returns BLK_STS_RESOURCE and SCHED_RESTART
1202                  * bit is set, run queue after a delay to avoid IO stalls
1203                  * that could otherwise occur if the queue is idle.
1204                  */
1205                 needs_restart = blk_mq_sched_needs_restart(hctx);
1206                 if (!needs_restart ||
1207                     (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
1208                         blk_mq_run_hw_queue(hctx, true);
1209                 else if (needs_restart && (ret == BLK_STS_RESOURCE))
1210                         blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
1211
1212                 return false;
1213         }
1214
1215         /*
1216          * If the host/device is unable to accept more work, inform the
1217          * caller of that.
1218          */
1219         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1220                 return false;
1221
1222         return (queued + errors) != 0;
1223 }
1224
1225 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
1226 {
1227         int srcu_idx;
1228
1229         /*
1230          * We should be running this queue from one of the CPUs that
1231          * are mapped to it.
1232          *
1233          * There are at least two related races now between setting
1234          * hctx->next_cpu from blk_mq_hctx_next_cpu() and running
1235          * __blk_mq_run_hw_queue():
1236          *
1237          * - hctx->next_cpu is found offline in blk_mq_hctx_next_cpu(),
1238          *   but later it becomes online, then this warning is harmless
1239          *   at all
1240          *
1241          * - hctx->next_cpu is found online in blk_mq_hctx_next_cpu(),
1242          *   but later it becomes offline, then the warning can't be
1243          *   triggered, and we depend on blk-mq timeout handler to
1244          *   handle dispatched requests to this hctx
1245          */
1246         if (!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask) &&
1247                 cpu_online(hctx->next_cpu)) {
1248                 printk(KERN_WARNING "run queue from wrong CPU %d, hctx %s\n",
1249                         raw_smp_processor_id(),
1250                         cpumask_empty(hctx->cpumask) ? "inactive": "active");
1251                 dump_stack();
1252         }
1253
1254         /*
1255          * We can't run the queue inline with ints disabled. Ensure that
1256          * we catch bad users of this early.
1257          */
1258         WARN_ON_ONCE(in_interrupt());
1259
1260         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1261
1262         hctx_lock(hctx, &srcu_idx);
1263         blk_mq_sched_dispatch_requests(hctx);
1264         hctx_unlock(hctx, srcu_idx);
1265 }
1266
1267 static inline int blk_mq_first_mapped_cpu(struct blk_mq_hw_ctx *hctx)
1268 {
1269         int cpu = cpumask_first_and(hctx->cpumask, cpu_online_mask);
1270
1271         if (cpu >= nr_cpu_ids)
1272                 cpu = cpumask_first(hctx->cpumask);
1273         return cpu;
1274 }
1275
1276 /*
1277  * It'd be great if the workqueue API had a way to pass
1278  * in a mask and had some smarts for more clever placement.
1279  * For now we just round-robin here, switching for every
1280  * BLK_MQ_CPU_WORK_BATCH queued items.
1281  */
1282 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
1283 {
1284         bool tried = false;
1285         int next_cpu = hctx->next_cpu;
1286
1287         if (hctx->queue->nr_hw_queues == 1)
1288                 return WORK_CPU_UNBOUND;
1289
1290         if (--hctx->next_cpu_batch <= 0) {
1291 select_cpu:
1292                 next_cpu = cpumask_next_and(next_cpu, hctx->cpumask,
1293                                 cpu_online_mask);
1294                 if (next_cpu >= nr_cpu_ids)
1295                         next_cpu = blk_mq_first_mapped_cpu(hctx);
1296                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1297         }
1298
1299         /*
1300          * Do unbound schedule if we can't find a online CPU for this hctx,
1301          * and it should only happen in the path of handling CPU DEAD.
1302          */
1303         if (!cpu_online(next_cpu)) {
1304                 if (!tried) {
1305                         tried = true;
1306                         goto select_cpu;
1307                 }
1308
1309                 /*
1310                  * Make sure to re-select CPU next time once after CPUs
1311                  * in hctx->cpumask become online again.
1312                  */
1313                 hctx->next_cpu = next_cpu;
1314                 hctx->next_cpu_batch = 1;
1315                 return WORK_CPU_UNBOUND;
1316         }
1317
1318         hctx->next_cpu = next_cpu;
1319         return next_cpu;
1320 }
1321
1322 static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
1323                                         unsigned long msecs)
1324 {
1325         if (unlikely(blk_mq_hctx_stopped(hctx)))
1326                 return;
1327
1328         if (!async && !(hctx->flags & BLK_MQ_F_BLOCKING)) {
1329                 int cpu = get_cpu();
1330                 if (cpumask_test_cpu(cpu, hctx->cpumask)) {
1331                         __blk_mq_run_hw_queue(hctx);
1332                         put_cpu();
1333                         return;
1334                 }
1335
1336                 put_cpu();
1337         }
1338
1339         kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
1340                                     msecs_to_jiffies(msecs));
1341 }
1342
1343 void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
1344 {
1345         __blk_mq_delay_run_hw_queue(hctx, true, msecs);
1346 }
1347 EXPORT_SYMBOL(blk_mq_delay_run_hw_queue);
1348
1349 bool blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1350 {
1351         int srcu_idx;
1352         bool need_run;
1353
1354         /*
1355          * When queue is quiesced, we may be switching io scheduler, or
1356          * updating nr_hw_queues, or other things, and we can't run queue
1357          * any more, even __blk_mq_hctx_has_pending() can't be called safely.
1358          *
1359          * And queue will be rerun in blk_mq_unquiesce_queue() if it is
1360          * quiesced.
1361          */
1362         hctx_lock(hctx, &srcu_idx);
1363         need_run = !blk_queue_quiesced(hctx->queue) &&
1364                 blk_mq_hctx_has_pending(hctx);
1365         hctx_unlock(hctx, srcu_idx);
1366
1367         if (need_run) {
1368                 __blk_mq_delay_run_hw_queue(hctx, async, 0);
1369                 return true;
1370         }
1371
1372         return false;
1373 }
1374 EXPORT_SYMBOL(blk_mq_run_hw_queue);
1375
1376 void blk_mq_run_hw_queues(struct request_queue *q, bool async)
1377 {
1378         struct blk_mq_hw_ctx *hctx;
1379         int i;
1380
1381         queue_for_each_hw_ctx(q, hctx, i) {
1382                 if (blk_mq_hctx_stopped(hctx))
1383                         continue;
1384
1385                 blk_mq_run_hw_queue(hctx, async);
1386         }
1387 }
1388 EXPORT_SYMBOL(blk_mq_run_hw_queues);
1389
1390 /**
1391  * blk_mq_queue_stopped() - check whether one or more hctxs have been stopped
1392  * @q: request queue.
1393  *
1394  * The caller is responsible for serializing this function against
1395  * blk_mq_{start,stop}_hw_queue().
1396  */
1397 bool blk_mq_queue_stopped(struct request_queue *q)
1398 {
1399         struct blk_mq_hw_ctx *hctx;
1400         int i;
1401
1402         queue_for_each_hw_ctx(q, hctx, i)
1403                 if (blk_mq_hctx_stopped(hctx))
1404                         return true;
1405
1406         return false;
1407 }
1408 EXPORT_SYMBOL(blk_mq_queue_stopped);
1409
1410 /*
1411  * This function is often used for pausing .queue_rq() by driver when
1412  * there isn't enough resource or some conditions aren't satisfied, and
1413  * BLK_STS_RESOURCE is usually returned.
1414  *
1415  * We do not guarantee that dispatch can be drained or blocked
1416  * after blk_mq_stop_hw_queue() returns. Please use
1417  * blk_mq_quiesce_queue() for that requirement.
1418  */
1419 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
1420 {
1421         cancel_delayed_work(&hctx->run_work);
1422
1423         set_bit(BLK_MQ_S_STOPPED, &hctx->state);
1424 }
1425 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
1426
1427 /*
1428  * This function is often used for pausing .queue_rq() by driver when
1429  * there isn't enough resource or some conditions aren't satisfied, and
1430  * BLK_STS_RESOURCE is usually returned.
1431  *
1432  * We do not guarantee that dispatch can be drained or blocked
1433  * after blk_mq_stop_hw_queues() returns. Please use
1434  * blk_mq_quiesce_queue() for that requirement.
1435  */
1436 void blk_mq_stop_hw_queues(struct request_queue *q)
1437 {
1438         struct blk_mq_hw_ctx *hctx;
1439         int i;
1440
1441         queue_for_each_hw_ctx(q, hctx, i)
1442                 blk_mq_stop_hw_queue(hctx);
1443 }
1444 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
1445
1446 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
1447 {
1448         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1449
1450         blk_mq_run_hw_queue(hctx, false);
1451 }
1452 EXPORT_SYMBOL(blk_mq_start_hw_queue);
1453
1454 void blk_mq_start_hw_queues(struct request_queue *q)
1455 {
1456         struct blk_mq_hw_ctx *hctx;
1457         int i;
1458
1459         queue_for_each_hw_ctx(q, hctx, i)
1460                 blk_mq_start_hw_queue(hctx);
1461 }
1462 EXPORT_SYMBOL(blk_mq_start_hw_queues);
1463
1464 void blk_mq_start_stopped_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
1465 {
1466         if (!blk_mq_hctx_stopped(hctx))
1467                 return;
1468
1469         clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
1470         blk_mq_run_hw_queue(hctx, async);
1471 }
1472 EXPORT_SYMBOL_GPL(blk_mq_start_stopped_hw_queue);
1473
1474 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
1475 {
1476         struct blk_mq_hw_ctx *hctx;
1477         int i;
1478
1479         queue_for_each_hw_ctx(q, hctx, i)
1480                 blk_mq_start_stopped_hw_queue(hctx, async);
1481 }
1482 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
1483
1484 static void blk_mq_run_work_fn(struct work_struct *work)
1485 {
1486         struct blk_mq_hw_ctx *hctx;
1487
1488         hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
1489
1490         /*
1491          * If we are stopped, don't run the queue.
1492          */
1493         if (test_bit(BLK_MQ_S_STOPPED, &hctx->state))
1494                 return;
1495
1496         __blk_mq_run_hw_queue(hctx);
1497 }
1498
1499 static inline void __blk_mq_insert_req_list(struct blk_mq_hw_ctx *hctx,
1500                                             struct request *rq,
1501                                             bool at_head)
1502 {
1503         struct blk_mq_ctx *ctx = rq->mq_ctx;
1504
1505         lockdep_assert_held(&ctx->lock);
1506
1507         trace_block_rq_insert(hctx->queue, rq);
1508
1509         if (at_head)
1510                 list_add(&rq->queuelist, &ctx->rq_list);
1511         else
1512                 list_add_tail(&rq->queuelist, &ctx->rq_list);
1513 }
1514
1515 void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
1516                              bool at_head)
1517 {
1518         struct blk_mq_ctx *ctx = rq->mq_ctx;
1519
1520         lockdep_assert_held(&ctx->lock);
1521
1522         __blk_mq_insert_req_list(hctx, rq, at_head);
1523         blk_mq_hctx_mark_pending(hctx, ctx);
1524 }
1525
1526 /*
1527  * Should only be used carefully, when the caller knows we want to
1528  * bypass a potential IO scheduler on the target device.
1529  */
1530 void blk_mq_request_bypass_insert(struct request *rq, bool run_queue)
1531 {
1532         struct blk_mq_ctx *ctx = rq->mq_ctx;
1533         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1534
1535         spin_lock(&hctx->lock);
1536         list_add_tail(&rq->queuelist, &hctx->dispatch);
1537         spin_unlock(&hctx->lock);
1538
1539         if (run_queue)
1540                 blk_mq_run_hw_queue(hctx, false);
1541 }
1542
1543 void blk_mq_insert_requests(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
1544                             struct list_head *list)
1545
1546 {
1547         /*
1548          * preemption doesn't flush plug list, so it's possible ctx->cpu is
1549          * offline now
1550          */
1551         spin_lock(&ctx->lock);
1552         while (!list_empty(list)) {
1553                 struct request *rq;
1554
1555                 rq = list_first_entry(list, struct request, queuelist);
1556                 BUG_ON(rq->mq_ctx != ctx);
1557                 list_del_init(&rq->queuelist);
1558                 __blk_mq_insert_req_list(hctx, rq, false);
1559         }
1560         blk_mq_hctx_mark_pending(hctx, ctx);
1561         spin_unlock(&ctx->lock);
1562 }
1563
1564 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
1565 {
1566         struct request *rqa = container_of(a, struct request, queuelist);
1567         struct request *rqb = container_of(b, struct request, queuelist);
1568
1569         return !(rqa->mq_ctx < rqb->mq_ctx ||
1570                  (rqa->mq_ctx == rqb->mq_ctx &&
1571                   blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1572 }
1573
1574 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1575 {
1576         struct blk_mq_ctx *this_ctx;
1577         struct request_queue *this_q;
1578         struct request *rq;
1579         LIST_HEAD(list);
1580         LIST_HEAD(ctx_list);
1581         unsigned int depth;
1582
1583         list_splice_init(&plug->mq_list, &list);
1584
1585         list_sort(NULL, &list, plug_ctx_cmp);
1586
1587         this_q = NULL;
1588         this_ctx = NULL;
1589         depth = 0;
1590
1591         while (!list_empty(&list)) {
1592                 rq = list_entry_rq(list.next);
1593                 list_del_init(&rq->queuelist);
1594                 BUG_ON(!rq->q);
1595                 if (rq->mq_ctx != this_ctx) {
1596                         if (this_ctx) {
1597                                 trace_block_unplug(this_q, depth, from_schedule);
1598                                 blk_mq_sched_insert_requests(this_q, this_ctx,
1599                                                                 &ctx_list,
1600                                                                 from_schedule);
1601                         }
1602
1603                         this_ctx = rq->mq_ctx;
1604                         this_q = rq->q;
1605                         depth = 0;
1606                 }
1607
1608                 depth++;
1609                 list_add_tail(&rq->queuelist, &ctx_list);
1610         }
1611
1612         /*
1613          * If 'this_ctx' is set, we know we have entries to complete
1614          * on 'ctx_list'. Do those.
1615          */
1616         if (this_ctx) {
1617                 trace_block_unplug(this_q, depth, from_schedule);
1618                 blk_mq_sched_insert_requests(this_q, this_ctx, &ctx_list,
1619                                                 from_schedule);
1620         }
1621 }
1622
1623 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1624 {
1625         blk_init_request_from_bio(rq, bio);
1626
1627         blk_rq_set_rl(rq, blk_get_rl(rq->q, bio));
1628
1629         blk_account_io_start(rq, true);
1630 }
1631
1632 static blk_qc_t request_to_qc_t(struct blk_mq_hw_ctx *hctx, struct request *rq)
1633 {
1634         if (rq->tag != -1)
1635                 return blk_tag_to_qc_t(rq->tag, hctx->queue_num, false);
1636
1637         return blk_tag_to_qc_t(rq->internal_tag, hctx->queue_num, true);
1638 }
1639
1640 static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
1641                                             struct request *rq,
1642                                             blk_qc_t *cookie)
1643 {
1644         struct request_queue *q = rq->q;
1645         struct blk_mq_queue_data bd = {
1646                 .rq = rq,
1647                 .last = true,
1648         };
1649         blk_qc_t new_cookie;
1650         blk_status_t ret;
1651
1652         new_cookie = request_to_qc_t(hctx, rq);
1653
1654         /*
1655          * For OK queue, we are done. For error, caller may kill it.
1656          * Any other error (busy), just add it to our list as we
1657          * previously would have done.
1658          */
1659         ret = q->mq_ops->queue_rq(hctx, &bd);
1660         switch (ret) {
1661         case BLK_STS_OK:
1662                 *cookie = new_cookie;
1663                 break;
1664         case BLK_STS_RESOURCE:
1665         case BLK_STS_DEV_RESOURCE:
1666                 __blk_mq_requeue_request(rq);
1667                 break;
1668         default:
1669                 *cookie = BLK_QC_T_NONE;
1670                 break;
1671         }
1672
1673         return ret;
1674 }
1675
1676 static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1677                                                 struct request *rq,
1678                                                 blk_qc_t *cookie,
1679                                                 bool bypass_insert)
1680 {
1681         struct request_queue *q = rq->q;
1682         bool run_queue = true;
1683
1684         /*
1685          * RCU or SRCU read lock is needed before checking quiesced flag.
1686          *
1687          * When queue is stopped or quiesced, ignore 'bypass_insert' from
1688          * blk_mq_request_issue_directly(), and return BLK_STS_OK to caller,
1689          * and avoid driver to try to dispatch again.
1690          */
1691         if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
1692                 run_queue = false;
1693                 bypass_insert = false;
1694                 goto insert;
1695         }
1696
1697         if (q->elevator && !bypass_insert)
1698                 goto insert;
1699
1700         if (!blk_mq_get_dispatch_budget(hctx))
1701                 goto insert;
1702
1703         if (!blk_mq_get_driver_tag(rq, NULL, false)) {
1704                 blk_mq_put_dispatch_budget(hctx);
1705                 goto insert;
1706         }
1707
1708         return __blk_mq_issue_directly(hctx, rq, cookie);
1709 insert:
1710         if (bypass_insert)
1711                 return BLK_STS_RESOURCE;
1712
1713         blk_mq_sched_insert_request(rq, false, run_queue, false);
1714         return BLK_STS_OK;
1715 }
1716
1717 static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
1718                 struct request *rq, blk_qc_t *cookie)
1719 {
1720         blk_status_t ret;
1721         int srcu_idx;
1722
1723         might_sleep_if(hctx->flags & BLK_MQ_F_BLOCKING);
1724
1725         hctx_lock(hctx, &srcu_idx);
1726
1727         ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
1728         if (ret == BLK_STS_RESOURCE || ret == BLK_STS_DEV_RESOURCE)
1729                 blk_mq_sched_insert_request(rq, false, true, false);
1730         else if (ret != BLK_STS_OK)
1731                 blk_mq_end_request(rq, ret);
1732
1733         hctx_unlock(hctx, srcu_idx);
1734 }
1735
1736 blk_status_t blk_mq_request_issue_directly(struct request *rq)
1737 {
1738         blk_status_t ret;
1739         int srcu_idx;
1740         blk_qc_t unused_cookie;
1741         struct blk_mq_ctx *ctx = rq->mq_ctx;
1742         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
1743
1744         hctx_lock(hctx, &srcu_idx);
1745         ret = __blk_mq_try_issue_directly(hctx, rq, &unused_cookie, true);
1746         hctx_unlock(hctx, srcu_idx);
1747
1748         return ret;
1749 }
1750
1751 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
1752 {
1753         const int is_sync = op_is_sync(bio->bi_opf);
1754         const int is_flush_fua = op_is_flush(bio->bi_opf);
1755         struct blk_mq_alloc_data data = { .flags = 0 };
1756         struct request *rq;
1757         unsigned int request_count = 0;
1758         struct blk_plug *plug;
1759         struct request *same_queue_rq = NULL;
1760         blk_qc_t cookie;
1761         unsigned int wb_acct;
1762
1763         blk_queue_bounce(q, &bio);
1764
1765         blk_queue_split(q, &bio);
1766
1767         if (!bio_integrity_prep(bio))
1768                 return BLK_QC_T_NONE;
1769
1770         if (!is_flush_fua && !blk_queue_nomerges(q) &&
1771             blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
1772                 return BLK_QC_T_NONE;
1773
1774         if (blk_mq_sched_bio_merge(q, bio))
1775                 return BLK_QC_T_NONE;
1776
1777         wb_acct = wbt_wait(q->rq_wb, bio, NULL);
1778
1779         trace_block_getrq(q, bio, bio->bi_opf);
1780
1781         rq = blk_mq_get_request(q, bio, bio->bi_opf, &data);
1782         if (unlikely(!rq)) {
1783                 __wbt_done(q->rq_wb, wb_acct);
1784                 if (bio->bi_opf & REQ_NOWAIT)
1785                         bio_wouldblock_error(bio);
1786                 return BLK_QC_T_NONE;
1787         }
1788
1789         wbt_track(rq, wb_acct);
1790
1791         cookie = request_to_qc_t(data.hctx, rq);
1792
1793         plug = current->plug;
1794         if (unlikely(is_flush_fua)) {
1795                 blk_mq_put_ctx(data.ctx);
1796                 blk_mq_bio_to_request(rq, bio);
1797
1798                 /* bypass scheduler for flush rq */
1799                 blk_insert_flush(rq);
1800                 blk_mq_run_hw_queue(data.hctx, true);
1801         } else if (plug && q->nr_hw_queues == 1) {
1802                 struct request *last = NULL;
1803
1804                 blk_mq_put_ctx(data.ctx);
1805                 blk_mq_bio_to_request(rq, bio);
1806
1807                 /*
1808                  * @request_count may become stale because of schedule
1809                  * out, so check the list again.
1810                  */
1811                 if (list_empty(&plug->mq_list))
1812                         request_count = 0;
1813                 else if (blk_queue_nomerges(q))
1814                         request_count = blk_plug_queued_count(q);
1815
1816                 if (!request_count)
1817                         trace_block_plug(q);
1818                 else
1819                         last = list_entry_rq(plug->mq_list.prev);
1820
1821                 if (request_count >= BLK_MAX_REQUEST_COUNT || (last &&
1822                     blk_rq_bytes(last) >= BLK_PLUG_FLUSH_SIZE)) {
1823                         blk_flush_plug_list(plug, false);
1824                         trace_block_plug(q);
1825                 }
1826
1827                 list_add_tail(&rq->queuelist, &plug->mq_list);
1828         } else if (plug && !blk_queue_nomerges(q)) {
1829                 blk_mq_bio_to_request(rq, bio);
1830
1831                 /*
1832                  * We do limited plugging. If the bio can be merged, do that.
1833                  * Otherwise the existing request in the plug list will be
1834                  * issued. So the plug list will have one request at most
1835                  * The plug list might get flushed before this. If that happens,
1836                  * the plug list is empty, and same_queue_rq is invalid.
1837                  */
1838                 if (list_empty(&plug->mq_list))
1839                         same_queue_rq = NULL;
1840                 if (same_queue_rq)
1841                         list_del_init(&same_queue_rq->queuelist);
1842                 list_add_tail(&rq->queuelist, &plug->mq_list);
1843
1844                 blk_mq_put_ctx(data.ctx);
1845
1846                 if (same_queue_rq) {
1847                         data.hctx = blk_mq_map_queue(q,
1848                                         same_queue_rq->mq_ctx->cpu);
1849                         blk_mq_try_issue_directly(data.hctx, same_queue_rq,
1850                                         &cookie);
1851                 }
1852         } else if (q->nr_hw_queues > 1 && is_sync) {
1853                 blk_mq_put_ctx(data.ctx);
1854                 blk_mq_bio_to_request(rq, bio);
1855                 blk_mq_try_issue_directly(data.hctx, rq, &cookie);
1856         } else {
1857                 blk_mq_put_ctx(data.ctx);
1858                 blk_mq_bio_to_request(rq, bio);
1859                 blk_mq_sched_insert_request(rq, false, true, true);
1860         }
1861
1862         return cookie;
1863 }
1864
1865 void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1866                      unsigned int hctx_idx)
1867 {
1868         struct page *page;
1869
1870         if (tags->rqs && set->ops->exit_request) {
1871                 int i;
1872
1873                 for (i = 0; i < tags->nr_tags; i++) {
1874                         struct request *rq = tags->static_rqs[i];
1875
1876                         if (!rq)
1877                                 continue;
1878                         set->ops->exit_request(set, rq, hctx_idx);
1879                         tags->static_rqs[i] = NULL;
1880                 }
1881         }
1882
1883         while (!list_empty(&tags->page_list)) {
1884                 page = list_first_entry(&tags->page_list, struct page, lru);
1885                 list_del_init(&page->lru);
1886                 /*
1887                  * Remove kmemleak object previously allocated in
1888                  * blk_mq_init_rq_map().
1889                  */
1890                 kmemleak_free(page_address(page));
1891                 __free_pages(page, page->private);
1892         }
1893 }
1894
1895 void blk_mq_free_rq_map(struct blk_mq_tags *tags)
1896 {
1897         kfree(tags->rqs);
1898         tags->rqs = NULL;
1899         kfree(tags->static_rqs);
1900         tags->static_rqs = NULL;
1901
1902         blk_mq_free_tags(tags);
1903 }
1904
1905 struct blk_mq_tags *blk_mq_alloc_rq_map(struct blk_mq_tag_set *set,
1906                                         unsigned int hctx_idx,
1907                                         unsigned int nr_tags,
1908                                         unsigned int reserved_tags)
1909 {
1910         struct blk_mq_tags *tags;
1911         int node;
1912
1913         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1914         if (node == NUMA_NO_NODE)
1915                 node = set->numa_node;
1916
1917         tags = blk_mq_init_tags(nr_tags, reserved_tags, node,
1918                                 BLK_MQ_FLAG_TO_ALLOC_POLICY(set->flags));
1919         if (!tags)
1920                 return NULL;
1921
1922         tags->rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1923                                  GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1924                                  node);
1925         if (!tags->rqs) {
1926                 blk_mq_free_tags(tags);
1927                 return NULL;
1928         }
1929
1930         tags->static_rqs = kcalloc_node(nr_tags, sizeof(struct request *),
1931                                         GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY,
1932                                         node);
1933         if (!tags->static_rqs) {
1934                 kfree(tags->rqs);
1935                 blk_mq_free_tags(tags);
1936                 return NULL;
1937         }
1938
1939         return tags;
1940 }
1941
1942 static size_t order_to_size(unsigned int order)
1943 {
1944         return (size_t)PAGE_SIZE << order;
1945 }
1946
1947 static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1948                                unsigned int hctx_idx, int node)
1949 {
1950         int ret;
1951
1952         if (set->ops->init_request) {
1953                 ret = set->ops->init_request(set, rq, hctx_idx, node);
1954                 if (ret)
1955                         return ret;
1956         }
1957
1958         WRITE_ONCE(rq->state, MQ_RQ_IDLE);
1959         return 0;
1960 }
1961
1962 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
1963                      unsigned int hctx_idx, unsigned int depth)
1964 {
1965         unsigned int i, j, entries_per_page, max_order = 4;
1966         size_t rq_size, left;
1967         int node;
1968
1969         node = blk_mq_hw_queue_to_node(set->mq_map, hctx_idx);
1970         if (node == NUMA_NO_NODE)
1971                 node = set->numa_node;
1972
1973         INIT_LIST_HEAD(&tags->page_list);
1974
1975         /*
1976          * rq_size is the size of the request plus driver payload, rounded
1977          * to the cacheline size
1978          */
1979         rq_size = round_up(sizeof(struct request) + set->cmd_size,
1980                                 cache_line_size());
1981         left = rq_size * depth;
1982
1983         for (i = 0; i < depth; ) {
1984                 int this_order = max_order;
1985                 struct page *page;
1986                 int to_do;
1987                 void *p;
1988
1989                 while (this_order && left < order_to_size(this_order - 1))
1990                         this_order--;
1991
1992                 do {
1993                         page = alloc_pages_node(node,
1994                                 GFP_NOIO | __GFP_NOWARN | __GFP_NORETRY | __GFP_ZERO,
1995                                 this_order);
1996                         if (page)
1997                                 break;
1998                         if (!this_order--)
1999                                 break;
2000                         if (order_to_size(this_order) < rq_size)
2001                                 break;
2002                 } while (1);
2003
2004                 if (!page)
2005                         goto fail;
2006
2007                 page->private = this_order;
2008                 list_add_tail(&page->lru, &tags->page_list);
2009
2010                 p = page_address(page);
2011                 /*
2012                  * Allow kmemleak to scan these pages as they contain pointers
2013                  * to additional allocations like via ops->init_request().
2014                  */
2015                 kmemleak_alloc(p, order_to_size(this_order), 1, GFP_NOIO);
2016                 entries_per_page = order_to_size(this_order) / rq_size;
2017                 to_do = min(entries_per_page, depth - i);
2018                 left -= to_do * rq_size;
2019                 for (j = 0; j < to_do; j++) {
2020                         struct request *rq = p;
2021
2022                         tags->static_rqs[i] = rq;
2023                         if (blk_mq_init_request(set, rq, hctx_idx, node)) {
2024                                 tags->static_rqs[i] = NULL;
2025                                 goto fail;
2026                         }
2027
2028                         p += rq_size;
2029                         i++;
2030                 }
2031         }
2032         return 0;
2033
2034 fail:
2035         blk_mq_free_rqs(set, tags, hctx_idx);
2036         return -ENOMEM;
2037 }
2038
2039 /*
2040  * 'cpu' is going away. splice any existing rq_list entries from this
2041  * software queue to the hw queue dispatch list, and ensure that it
2042  * gets run.
2043  */
2044 static int blk_mq_hctx_notify_dead(unsigned int cpu, struct hlist_node *node)
2045 {
2046         struct blk_mq_hw_ctx *hctx;
2047         struct blk_mq_ctx *ctx;
2048         LIST_HEAD(tmp);
2049
2050         hctx = hlist_entry_safe(node, struct blk_mq_hw_ctx, cpuhp_dead);
2051         ctx = __blk_mq_get_ctx(hctx->queue, cpu);
2052
2053         spin_lock(&ctx->lock);
2054         if (!list_empty(&ctx->rq_list)) {
2055                 list_splice_init(&ctx->rq_list, &tmp);
2056                 blk_mq_hctx_clear_pending(hctx, ctx);
2057         }
2058         spin_unlock(&ctx->lock);
2059
2060         if (list_empty(&tmp))
2061                 return 0;
2062
2063         spin_lock(&hctx->lock);
2064         list_splice_tail_init(&tmp, &hctx->dispatch);
2065         spin_unlock(&hctx->lock);
2066
2067         blk_mq_run_hw_queue(hctx, true);
2068         return 0;
2069 }
2070
2071 static void blk_mq_remove_cpuhp(struct blk_mq_hw_ctx *hctx)
2072 {
2073         cpuhp_state_remove_instance_nocalls(CPUHP_BLK_MQ_DEAD,
2074                                             &hctx->cpuhp_dead);
2075 }
2076
2077 /* hctx->ctxs will be freed in queue's release handler */
2078 static void blk_mq_exit_hctx(struct request_queue *q,
2079                 struct blk_mq_tag_set *set,
2080                 struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
2081 {
2082         blk_mq_debugfs_unregister_hctx(hctx);
2083
2084         if (blk_mq_hw_queue_mapped(hctx))
2085                 blk_mq_tag_idle(hctx);
2086
2087         if (set->ops->exit_request)
2088                 set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
2089
2090         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2091
2092         if (set->ops->exit_hctx)
2093                 set->ops->exit_hctx(hctx, hctx_idx);
2094
2095         if (hctx->flags & BLK_MQ_F_BLOCKING)
2096                 cleanup_srcu_struct(hctx->srcu);
2097
2098         blk_mq_remove_cpuhp(hctx);
2099         blk_free_flush_queue(hctx->fq);
2100         sbitmap_free(&hctx->ctx_map);
2101 }
2102
2103 static void blk_mq_exit_hw_queues(struct request_queue *q,
2104                 struct blk_mq_tag_set *set, int nr_queue)
2105 {
2106         struct blk_mq_hw_ctx *hctx;
2107         unsigned int i;
2108
2109         queue_for_each_hw_ctx(q, hctx, i) {
2110                 if (i == nr_queue)
2111                         break;
2112                 blk_mq_exit_hctx(q, set, hctx, i);
2113         }
2114 }
2115
2116 static int blk_mq_init_hctx(struct request_queue *q,
2117                 struct blk_mq_tag_set *set,
2118                 struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
2119 {
2120         int node;
2121
2122         node = hctx->numa_node;
2123         if (node == NUMA_NO_NODE)
2124                 node = hctx->numa_node = set->numa_node;
2125
2126         INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
2127         spin_lock_init(&hctx->lock);
2128         INIT_LIST_HEAD(&hctx->dispatch);
2129         hctx->queue = q;
2130         hctx->flags = set->flags & ~BLK_MQ_F_TAG_SHARED;
2131
2132         cpuhp_state_add_instance_nocalls(CPUHP_BLK_MQ_DEAD, &hctx->cpuhp_dead);
2133
2134         hctx->tags = set->tags[hctx_idx];
2135
2136         /*
2137          * Allocate space for all possible cpus to avoid allocation at
2138          * runtime
2139          */
2140         hctx->ctxs = kmalloc_array_node(nr_cpu_ids, sizeof(void *),
2141                                         GFP_KERNEL, node);
2142         if (!hctx->ctxs)
2143                 goto unregister_cpu_notifier;
2144
2145         if (sbitmap_init_node(&hctx->ctx_map, nr_cpu_ids, ilog2(8), GFP_KERNEL,
2146                               node))
2147                 goto free_ctxs;
2148
2149         hctx->nr_ctx = 0;
2150
2151         init_waitqueue_func_entry(&hctx->dispatch_wait, blk_mq_dispatch_wake);
2152         INIT_LIST_HEAD(&hctx->dispatch_wait.entry);
2153
2154         if (set->ops->init_hctx &&
2155             set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
2156                 goto free_bitmap;
2157
2158         if (blk_mq_sched_init_hctx(q, hctx, hctx_idx))
2159                 goto exit_hctx;
2160
2161         hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
2162         if (!hctx->fq)
2163                 goto sched_exit_hctx;
2164
2165         if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
2166                 goto free_fq;
2167
2168         if (hctx->flags & BLK_MQ_F_BLOCKING)
2169                 init_srcu_struct(hctx->srcu);
2170
2171         blk_mq_debugfs_register_hctx(q, hctx);
2172
2173         return 0;
2174
2175  free_fq:
2176         kfree(hctx->fq);
2177  sched_exit_hctx:
2178         blk_mq_sched_exit_hctx(q, hctx, hctx_idx);
2179  exit_hctx:
2180         if (set->ops->exit_hctx)
2181                 set->ops->exit_hctx(hctx, hctx_idx);
2182  free_bitmap:
2183         sbitmap_free(&hctx->ctx_map);
2184  free_ctxs:
2185         kfree(hctx->ctxs);
2186  unregister_cpu_notifier:
2187         blk_mq_remove_cpuhp(hctx);
2188         return -1;
2189 }
2190
2191 static void blk_mq_init_cpu_queues(struct request_queue *q,
2192                                    unsigned int nr_hw_queues)
2193 {
2194         unsigned int i;
2195
2196         for_each_possible_cpu(i) {
2197                 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
2198                 struct blk_mq_hw_ctx *hctx;
2199
2200                 __ctx->cpu = i;
2201                 spin_lock_init(&__ctx->lock);
2202                 INIT_LIST_HEAD(&__ctx->rq_list);
2203                 __ctx->queue = q;
2204
2205                 /*
2206                  * Set local node, IFF we have more than one hw queue. If
2207                  * not, we remain on the home node of the device
2208                  */
2209                 hctx = blk_mq_map_queue(q, i);
2210                 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
2211                         hctx->numa_node = local_memory_node(cpu_to_node(i));
2212         }
2213 }
2214
2215 static bool __blk_mq_alloc_rq_map(struct blk_mq_tag_set *set, int hctx_idx)
2216 {
2217         int ret = 0;
2218
2219         set->tags[hctx_idx] = blk_mq_alloc_rq_map(set, hctx_idx,
2220                                         set->queue_depth, set->reserved_tags);
2221         if (!set->tags[hctx_idx])
2222                 return false;
2223
2224         ret = blk_mq_alloc_rqs(set, set->tags[hctx_idx], hctx_idx,
2225                                 set->queue_depth);
2226         if (!ret)
2227                 return true;
2228
2229         blk_mq_free_rq_map(set->tags[hctx_idx]);
2230         set->tags[hctx_idx] = NULL;
2231         return false;
2232 }
2233
2234 static void blk_mq_free_map_and_requests(struct blk_mq_tag_set *set,
2235                                          unsigned int hctx_idx)
2236 {
2237         if (set->tags[hctx_idx]) {
2238                 blk_mq_free_rqs(set, set->tags[hctx_idx], hctx_idx);
2239                 blk_mq_free_rq_map(set->tags[hctx_idx]);
2240                 set->tags[hctx_idx] = NULL;
2241         }
2242 }
2243
2244 static void blk_mq_map_swqueue(struct request_queue *q)
2245 {
2246         unsigned int i, hctx_idx;
2247         struct blk_mq_hw_ctx *hctx;
2248         struct blk_mq_ctx *ctx;
2249         struct blk_mq_tag_set *set = q->tag_set;
2250
2251         /*
2252          * Avoid others reading imcomplete hctx->cpumask through sysfs
2253          */
2254         mutex_lock(&q->sysfs_lock);
2255
2256         queue_for_each_hw_ctx(q, hctx, i) {
2257                 cpumask_clear(hctx->cpumask);
2258                 hctx->nr_ctx = 0;
2259                 hctx->dispatch_from = NULL;
2260         }
2261
2262         /*
2263          * Map software to hardware queues.
2264          *
2265          * If the cpu isn't present, the cpu is mapped to first hctx.
2266          */
2267         for_each_possible_cpu(i) {
2268                 hctx_idx = q->mq_map[i];
2269                 /* unmapped hw queue can be remapped after CPU topo changed */
2270                 if (!set->tags[hctx_idx] &&
2271                     !__blk_mq_alloc_rq_map(set, hctx_idx)) {
2272                         /*
2273                          * If tags initialization fail for some hctx,
2274                          * that hctx won't be brought online.  In this
2275                          * case, remap the current ctx to hctx[0] which
2276                          * is guaranteed to always have tags allocated
2277                          */
2278                         q->mq_map[i] = 0;
2279                 }
2280
2281                 ctx = per_cpu_ptr(q->queue_ctx, i);
2282                 hctx = blk_mq_map_queue(q, i);
2283
2284                 cpumask_set_cpu(i, hctx->cpumask);
2285                 ctx->index_hw = hctx->nr_ctx;
2286                 hctx->ctxs[hctx->nr_ctx++] = ctx;
2287         }
2288
2289         mutex_unlock(&q->sysfs_lock);
2290
2291         queue_for_each_hw_ctx(q, hctx, i) {
2292                 /*
2293                  * If no software queues are mapped to this hardware queue,
2294                  * disable it and free the request entries.
2295                  */
2296                 if (!hctx->nr_ctx) {
2297                         /* Never unmap queue 0.  We need it as a
2298                          * fallback in case of a new remap fails
2299                          * allocation
2300                          */
2301                         if (i && set->tags[i])
2302                                 blk_mq_free_map_and_requests(set, i);
2303
2304                         hctx->tags = NULL;
2305                         continue;
2306                 }
2307
2308                 hctx->tags = set->tags[i];
2309                 WARN_ON(!hctx->tags);
2310
2311                 /*
2312                  * Set the map size to the number of mapped software queues.
2313                  * This is more accurate and more efficient than looping
2314                  * over all possibly mapped software queues.
2315                  */
2316                 sbitmap_resize(&hctx->ctx_map, hctx->nr_ctx);
2317
2318                 /*
2319                  * Initialize batch roundrobin counts
2320                  */
2321                 hctx->next_cpu = blk_mq_first_mapped_cpu(hctx);
2322                 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
2323         }
2324 }
2325
2326 /*
2327  * Caller needs to ensure that we're either frozen/quiesced, or that
2328  * the queue isn't live yet.
2329  */
2330 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
2331 {
2332         struct blk_mq_hw_ctx *hctx;
2333         int i;
2334
2335         queue_for_each_hw_ctx(q, hctx, i) {
2336                 if (shared) {
2337                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2338                                 atomic_inc(&q->shared_hctx_restart);
2339                         hctx->flags |= BLK_MQ_F_TAG_SHARED;
2340                 } else {
2341                         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
2342                                 atomic_dec(&q->shared_hctx_restart);
2343                         hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
2344                 }
2345         }
2346 }
2347
2348 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
2349                                         bool shared)
2350 {
2351         struct request_queue *q;
2352
2353         lockdep_assert_held(&set->tag_list_lock);
2354
2355         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2356                 blk_mq_freeze_queue(q);
2357                 queue_set_hctx_shared(q, shared);
2358                 blk_mq_unfreeze_queue(q);
2359         }
2360 }
2361
2362 static void blk_mq_del_queue_tag_set(struct request_queue *q)
2363 {
2364         struct blk_mq_tag_set *set = q->tag_set;
2365
2366         mutex_lock(&set->tag_list_lock);
2367         list_del_rcu(&q->tag_set_list);
2368         if (list_is_singular(&set->tag_list)) {
2369                 /* just transitioned to unshared */
2370                 set->flags &= ~BLK_MQ_F_TAG_SHARED;
2371                 /* update existing queue */
2372                 blk_mq_update_tag_set_depth(set, false);
2373         }
2374         mutex_unlock(&set->tag_list_lock);
2375         synchronize_rcu();
2376         INIT_LIST_HEAD(&q->tag_set_list);
2377 }
2378
2379 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
2380                                      struct request_queue *q)
2381 {
2382         q->tag_set = set;
2383
2384         mutex_lock(&set->tag_list_lock);
2385
2386         /*
2387          * Check to see if we're transitioning to shared (from 1 to 2 queues).
2388          */
2389         if (!list_empty(&set->tag_list) &&
2390             !(set->flags & BLK_MQ_F_TAG_SHARED)) {
2391                 set->flags |= BLK_MQ_F_TAG_SHARED;
2392                 /* update existing queue */
2393                 blk_mq_update_tag_set_depth(set, true);
2394         }
2395         if (set->flags & BLK_MQ_F_TAG_SHARED)
2396                 queue_set_hctx_shared(q, true);
2397         list_add_tail_rcu(&q->tag_set_list, &set->tag_list);
2398
2399         mutex_unlock(&set->tag_list_lock);
2400 }
2401
2402 /*
2403  * It is the actual release handler for mq, but we do it from
2404  * request queue's release handler for avoiding use-after-free
2405  * and headache because q->mq_kobj shouldn't have been introduced,
2406  * but we can't group ctx/kctx kobj without it.
2407  */
2408 void blk_mq_release(struct request_queue *q)
2409 {
2410         struct blk_mq_hw_ctx *hctx;
2411         unsigned int i;
2412
2413         /* hctx kobj stays in hctx */
2414         queue_for_each_hw_ctx(q, hctx, i) {
2415                 if (!hctx)
2416                         continue;
2417                 kobject_put(&hctx->kobj);
2418         }
2419
2420         q->mq_map = NULL;
2421
2422         kfree(q->queue_hw_ctx);
2423
2424         /*
2425          * release .mq_kobj and sw queue's kobject now because
2426          * both share lifetime with request queue.
2427          */
2428         blk_mq_sysfs_deinit(q);
2429
2430         free_percpu(q->queue_ctx);
2431 }
2432
2433 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
2434 {
2435         struct request_queue *uninit_q, *q;
2436
2437         uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
2438         if (!uninit_q)
2439                 return ERR_PTR(-ENOMEM);
2440
2441         q = blk_mq_init_allocated_queue(set, uninit_q);
2442         if (IS_ERR(q))
2443                 blk_cleanup_queue(uninit_q);
2444
2445         return q;
2446 }
2447 EXPORT_SYMBOL(blk_mq_init_queue);
2448
2449 static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
2450 {
2451         int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
2452
2453         BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
2454                            __alignof__(struct blk_mq_hw_ctx)) !=
2455                      sizeof(struct blk_mq_hw_ctx));
2456
2457         if (tag_set->flags & BLK_MQ_F_BLOCKING)
2458                 hw_ctx_size += sizeof(struct srcu_struct);
2459
2460         return hw_ctx_size;
2461 }
2462
2463 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
2464                                                 struct request_queue *q)
2465 {
2466         int i, j;
2467         struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
2468
2469         blk_mq_sysfs_unregister(q);
2470
2471         /* protect against switching io scheduler  */
2472         mutex_lock(&q->sysfs_lock);
2473         for (i = 0; i < set->nr_hw_queues; i++) {
2474                 int node;
2475
2476                 if (hctxs[i])
2477                         continue;
2478
2479                 node = blk_mq_hw_queue_to_node(q->mq_map, i);
2480                 hctxs[i] = kzalloc_node(blk_mq_hw_ctx_size(set),
2481                                         GFP_KERNEL, node);
2482                 if (!hctxs[i])
2483                         break;
2484
2485                 if (!zalloc_cpumask_var_node(&hctxs[i]->cpumask, GFP_KERNEL,
2486                                                 node)) {
2487                         kfree(hctxs[i]);
2488                         hctxs[i] = NULL;
2489                         break;
2490                 }
2491
2492                 atomic_set(&hctxs[i]->nr_active, 0);
2493                 hctxs[i]->numa_node = node;
2494                 hctxs[i]->queue_num = i;
2495
2496                 if (blk_mq_init_hctx(q, set, hctxs[i], i)) {
2497                         free_cpumask_var(hctxs[i]->cpumask);
2498                         kfree(hctxs[i]);
2499                         hctxs[i] = NULL;
2500                         break;
2501                 }
2502                 blk_mq_hctx_kobj_init(hctxs[i]);
2503         }
2504         for (j = i; j < q->nr_hw_queues; j++) {
2505                 struct blk_mq_hw_ctx *hctx = hctxs[j];
2506
2507                 if (hctx) {
2508                         if (hctx->tags)
2509                                 blk_mq_free_map_and_requests(set, j);
2510                         blk_mq_exit_hctx(q, set, hctx, j);
2511                         kobject_put(&hctx->kobj);
2512                         hctxs[j] = NULL;
2513
2514                 }
2515         }
2516         q->nr_hw_queues = i;
2517         mutex_unlock(&q->sysfs_lock);
2518         blk_mq_sysfs_register(q);
2519 }
2520
2521 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
2522                                                   struct request_queue *q)
2523 {
2524         /* mark the queue as mq asap */
2525         q->mq_ops = set->ops;
2526
2527         q->poll_cb = blk_stat_alloc_callback(blk_mq_poll_stats_fn,
2528                                              blk_mq_poll_stats_bkt,
2529                                              BLK_MQ_POLL_STATS_BKTS, q);
2530         if (!q->poll_cb)
2531                 goto err_exit;
2532
2533         q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
2534         if (!q->queue_ctx)
2535                 goto err_exit;
2536
2537         /* init q->mq_kobj and sw queues' kobjects */
2538         blk_mq_sysfs_init(q);
2539
2540         q->queue_hw_ctx = kcalloc_node(nr_cpu_ids, sizeof(*(q->queue_hw_ctx)),
2541                                                 GFP_KERNEL, set->numa_node);
2542         if (!q->queue_hw_ctx)
2543                 goto err_percpu;
2544
2545         q->mq_map = set->mq_map;
2546
2547         blk_mq_realloc_hw_ctxs(set, q);
2548         if (!q->nr_hw_queues)
2549                 goto err_hctxs;
2550
2551         INIT_WORK(&q->timeout_work, blk_mq_timeout_work);
2552         blk_queue_rq_timeout(q, set->timeout ? set->timeout : 30 * HZ);
2553
2554         q->nr_queues = nr_cpu_ids;
2555
2556         q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
2557
2558         if (!(set->flags & BLK_MQ_F_SG_MERGE))
2559                 queue_flag_set_unlocked(QUEUE_FLAG_NO_SG_MERGE, q);
2560
2561         q->sg_reserved_size = INT_MAX;
2562
2563         INIT_DELAYED_WORK(&q->requeue_work, blk_mq_requeue_work);
2564         INIT_LIST_HEAD(&q->requeue_list);
2565         spin_lock_init(&q->requeue_lock);
2566
2567         blk_queue_make_request(q, blk_mq_make_request);
2568         if (q->mq_ops->poll)
2569                 q->poll_fn = blk_mq_poll;
2570
2571         /*
2572          * Do this after blk_queue_make_request() overrides it...
2573          */
2574         q->nr_requests = set->queue_depth;
2575
2576         /*
2577          * Default to classic polling
2578          */
2579         q->poll_nsec = -1;
2580
2581         if (set->ops->complete)
2582                 blk_queue_softirq_done(q, set->ops->complete);
2583
2584         blk_mq_init_cpu_queues(q, set->nr_hw_queues);
2585         blk_mq_add_queue_tag_set(set, q);
2586         blk_mq_map_swqueue(q);
2587
2588         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
2589                 int ret;
2590
2591                 ret = elevator_init_mq(q);
2592                 if (ret)
2593                         return ERR_PTR(ret);
2594         }
2595
2596         return q;
2597
2598 err_hctxs:
2599         kfree(q->queue_hw_ctx);
2600 err_percpu:
2601         free_percpu(q->queue_ctx);
2602 err_exit:
2603         q->mq_ops = NULL;
2604         return ERR_PTR(-ENOMEM);
2605 }
2606 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
2607
2608 void blk_mq_free_queue(struct request_queue *q)
2609 {
2610         struct blk_mq_tag_set   *set = q->tag_set;
2611
2612         blk_mq_del_queue_tag_set(q);
2613         blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
2614 }
2615
2616 /* Basically redo blk_mq_init_queue with queue frozen */
2617 static void blk_mq_queue_reinit(struct request_queue *q)
2618 {
2619         WARN_ON_ONCE(!atomic_read(&q->mq_freeze_depth));
2620
2621         blk_mq_debugfs_unregister_hctxs(q);
2622         blk_mq_sysfs_unregister(q);
2623
2624         /*
2625          * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
2626          * we should change hctx numa_node according to the new topology (this
2627          * involves freeing and re-allocating memory, worth doing?)
2628          */
2629         blk_mq_map_swqueue(q);
2630
2631         blk_mq_sysfs_register(q);
2632         blk_mq_debugfs_register_hctxs(q);
2633 }
2634
2635 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2636 {
2637         int i;
2638
2639         for (i = 0; i < set->nr_hw_queues; i++)
2640                 if (!__blk_mq_alloc_rq_map(set, i))
2641                         goto out_unwind;
2642
2643         return 0;
2644
2645 out_unwind:
2646         while (--i >= 0)
2647                 blk_mq_free_rq_map(set->tags[i]);
2648
2649         return -ENOMEM;
2650 }
2651
2652 /*
2653  * Allocate the request maps associated with this tag_set. Note that this
2654  * may reduce the depth asked for, if memory is tight. set->queue_depth
2655  * will be updated to reflect the allocated depth.
2656  */
2657 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
2658 {
2659         unsigned int depth;
2660         int err;
2661
2662         depth = set->queue_depth;
2663         do {
2664                 err = __blk_mq_alloc_rq_maps(set);
2665                 if (!err)
2666                         break;
2667
2668                 set->queue_depth >>= 1;
2669                 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
2670                         err = -ENOMEM;
2671                         break;
2672                 }
2673         } while (set->queue_depth);
2674
2675         if (!set->queue_depth || err) {
2676                 pr_err("blk-mq: failed to allocate request map\n");
2677                 return -ENOMEM;
2678         }
2679
2680         if (depth != set->queue_depth)
2681                 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
2682                                                 depth, set->queue_depth);
2683
2684         return 0;
2685 }
2686
2687 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
2688 {
2689         if (set->ops->map_queues) {
2690                 int cpu;
2691                 /*
2692                  * transport .map_queues is usually done in the following
2693                  * way:
2694                  *
2695                  * for (queue = 0; queue < set->nr_hw_queues; queue++) {
2696                  *      mask = get_cpu_mask(queue)
2697                  *      for_each_cpu(cpu, mask)
2698                  *              set->mq_map[cpu] = queue;
2699                  * }
2700                  *
2701                  * When we need to remap, the table has to be cleared for
2702                  * killing stale mapping since one CPU may not be mapped
2703                  * to any hw queue.
2704                  */
2705                 for_each_possible_cpu(cpu)
2706                         set->mq_map[cpu] = 0;
2707
2708                 return set->ops->map_queues(set);
2709         } else
2710                 return blk_mq_map_queues(set);
2711 }
2712
2713 /*
2714  * Alloc a tag set to be associated with one or more request queues.
2715  * May fail with EINVAL for various error conditions. May adjust the
2716  * requested depth down, if if it too large. In that case, the set
2717  * value will be stored in set->queue_depth.
2718  */
2719 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
2720 {
2721         int ret;
2722
2723         BUILD_BUG_ON(BLK_MQ_MAX_DEPTH > 1 << BLK_MQ_UNIQUE_TAG_BITS);
2724
2725         if (!set->nr_hw_queues)
2726                 return -EINVAL;
2727         if (!set->queue_depth)
2728                 return -EINVAL;
2729         if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
2730                 return -EINVAL;
2731
2732         if (!set->ops->queue_rq)
2733                 return -EINVAL;
2734
2735         if (!set->ops->get_budget ^ !set->ops->put_budget)
2736                 return -EINVAL;
2737
2738         if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2739                 pr_info("blk-mq: reduced tag depth to %u\n",
2740                         BLK_MQ_MAX_DEPTH);
2741                 set->queue_depth = BLK_MQ_MAX_DEPTH;
2742         }
2743
2744         /*
2745          * If a crashdump is active, then we are potentially in a very
2746          * memory constrained environment. Limit us to 1 queue and
2747          * 64 tags to prevent using too much memory.
2748          */
2749         if (is_kdump_kernel()) {
2750                 set->nr_hw_queues = 1;
2751                 set->queue_depth = min(64U, set->queue_depth);
2752         }
2753         /*
2754          * There is no use for more h/w queues than cpus.
2755          */
2756         if (set->nr_hw_queues > nr_cpu_ids)
2757                 set->nr_hw_queues = nr_cpu_ids;
2758
2759         set->tags = kcalloc_node(nr_cpu_ids, sizeof(struct blk_mq_tags *),
2760                                  GFP_KERNEL, set->numa_node);
2761         if (!set->tags)
2762                 return -ENOMEM;
2763
2764         ret = -ENOMEM;
2765         set->mq_map = kcalloc_node(nr_cpu_ids, sizeof(*set->mq_map),
2766                                    GFP_KERNEL, set->numa_node);
2767         if (!set->mq_map)
2768                 goto out_free_tags;
2769
2770         ret = blk_mq_update_queue_map(set);
2771         if (ret)
2772                 goto out_free_mq_map;
2773
2774         ret = blk_mq_alloc_rq_maps(set);
2775         if (ret)
2776                 goto out_free_mq_map;
2777
2778         mutex_init(&set->tag_list_lock);
2779         INIT_LIST_HEAD(&set->tag_list);
2780
2781         return 0;
2782
2783 out_free_mq_map:
2784         kfree(set->mq_map);
2785         set->mq_map = NULL;
2786 out_free_tags:
2787         kfree(set->tags);
2788         set->tags = NULL;
2789         return ret;
2790 }
2791 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2792
2793 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2794 {
2795         int i;
2796
2797         for (i = 0; i < nr_cpu_ids; i++)
2798                 blk_mq_free_map_and_requests(set, i);
2799
2800         kfree(set->mq_map);
2801         set->mq_map = NULL;
2802
2803         kfree(set->tags);
2804         set->tags = NULL;
2805 }
2806 EXPORT_SYMBOL(blk_mq_free_tag_set);
2807
2808 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2809 {
2810         struct blk_mq_tag_set *set = q->tag_set;
2811         struct blk_mq_hw_ctx *hctx;
2812         int i, ret;
2813
2814         if (!set)
2815                 return -EINVAL;
2816
2817         blk_mq_freeze_queue(q);
2818         blk_mq_quiesce_queue(q);
2819
2820         ret = 0;
2821         queue_for_each_hw_ctx(q, hctx, i) {
2822                 if (!hctx->tags)
2823                         continue;
2824                 /*
2825                  * If we're using an MQ scheduler, just update the scheduler
2826                  * queue depth. This is similar to what the old code would do.
2827                  */
2828                 if (!hctx->sched_tags) {
2829                         ret = blk_mq_tag_update_depth(hctx, &hctx->tags, nr,
2830                                                         false);
2831                 } else {
2832                         ret = blk_mq_tag_update_depth(hctx, &hctx->sched_tags,
2833                                                         nr, true);
2834                 }
2835                 if (ret)
2836                         break;
2837         }
2838
2839         if (!ret)
2840                 q->nr_requests = nr;
2841
2842         blk_mq_unquiesce_queue(q);
2843         blk_mq_unfreeze_queue(q);
2844
2845         return ret;
2846 }
2847
2848 static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
2849                                                         int nr_hw_queues)
2850 {
2851         struct request_queue *q;
2852
2853         lockdep_assert_held(&set->tag_list_lock);
2854
2855         if (nr_hw_queues > nr_cpu_ids)
2856                 nr_hw_queues = nr_cpu_ids;
2857         if (nr_hw_queues < 1 || nr_hw_queues == set->nr_hw_queues)
2858                 return;
2859
2860         list_for_each_entry(q, &set->tag_list, tag_set_list)
2861                 blk_mq_freeze_queue(q);
2862
2863         set->nr_hw_queues = nr_hw_queues;
2864         blk_mq_update_queue_map(set);
2865         list_for_each_entry(q, &set->tag_list, tag_set_list) {
2866                 blk_mq_realloc_hw_ctxs(set, q);
2867                 blk_mq_queue_reinit(q);
2868         }
2869
2870         list_for_each_entry(q, &set->tag_list, tag_set_list)
2871                 blk_mq_unfreeze_queue(q);
2872 }
2873
2874 void blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set, int nr_hw_queues)
2875 {
2876         mutex_lock(&set->tag_list_lock);
2877         __blk_mq_update_nr_hw_queues(set, nr_hw_queues);
2878         mutex_unlock(&set->tag_list_lock);
2879 }
2880 EXPORT_SYMBOL_GPL(blk_mq_update_nr_hw_queues);
2881
2882 /* Enable polling stats and return whether they were already enabled. */
2883 static bool blk_poll_stats_enable(struct request_queue *q)
2884 {
2885         if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2886             blk_queue_flag_test_and_set(QUEUE_FLAG_POLL_STATS, q))
2887                 return true;
2888         blk_stat_add_callback(q, q->poll_cb);
2889         return false;
2890 }
2891
2892 static void blk_mq_poll_stats_start(struct request_queue *q)
2893 {
2894         /*
2895          * We don't arm the callback if polling stats are not enabled or the
2896          * callback is already active.
2897          */
2898         if (!test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags) ||
2899             blk_stat_is_active(q->poll_cb))
2900                 return;
2901
2902         blk_stat_activate_msecs(q->poll_cb, 100);
2903 }
2904
2905 static void blk_mq_poll_stats_fn(struct blk_stat_callback *cb)
2906 {
2907         struct request_queue *q = cb->data;
2908         int bucket;
2909
2910         for (bucket = 0; bucket < BLK_MQ_POLL_STATS_BKTS; bucket++) {
2911                 if (cb->stat[bucket].nr_samples)
2912                         q->poll_stat[bucket] = cb->stat[bucket];
2913         }
2914 }
2915
2916 static unsigned long blk_mq_poll_nsecs(struct request_queue *q,
2917                                        struct blk_mq_hw_ctx *hctx,
2918                                        struct request *rq)
2919 {
2920         unsigned long ret = 0;
2921         int bucket;
2922
2923         /*
2924          * If stats collection isn't on, don't sleep but turn it on for
2925          * future users
2926          */
2927         if (!blk_poll_stats_enable(q))
2928                 return 0;
2929
2930         /*
2931          * As an optimistic guess, use half of the mean service time
2932          * for this type of request. We can (and should) make this smarter.
2933          * For instance, if the completion latencies are tight, we can
2934          * get closer than just half the mean. This is especially
2935          * important on devices where the completion latencies are longer
2936          * than ~10 usec. We do use the stats for the relevant IO size
2937          * if available which does lead to better estimates.
2938          */
2939         bucket = blk_mq_poll_stats_bkt(rq);
2940         if (bucket < 0)
2941                 return ret;
2942
2943         if (q->poll_stat[bucket].nr_samples)
2944                 ret = (q->poll_stat[bucket].mean + 1) / 2;
2945
2946         return ret;
2947 }
2948
2949 static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
2950                                      struct blk_mq_hw_ctx *hctx,
2951                                      struct request *rq)
2952 {
2953         struct hrtimer_sleeper hs;
2954         enum hrtimer_mode mode;
2955         unsigned int nsecs;
2956         ktime_t kt;
2957
2958         if (rq->rq_flags & RQF_MQ_POLL_SLEPT)
2959                 return false;
2960
2961         /*
2962          * poll_nsec can be:
2963          *
2964          * -1:  don't ever hybrid sleep
2965          *  0:  use half of prev avg
2966          * >0:  use this specific value
2967          */
2968         if (q->poll_nsec == -1)
2969                 return false;
2970         else if (q->poll_nsec > 0)
2971                 nsecs = q->poll_nsec;
2972         else
2973                 nsecs = blk_mq_poll_nsecs(q, hctx, rq);
2974
2975         if (!nsecs)
2976                 return false;
2977
2978         rq->rq_flags |= RQF_MQ_POLL_SLEPT;
2979
2980         /*
2981          * This will be replaced with the stats tracking code, using
2982          * 'avg_completion_time / 2' as the pre-sleep target.
2983          */
2984         kt = nsecs;
2985
2986         mode = HRTIMER_MODE_REL;
2987         hrtimer_init_on_stack(&hs.timer, CLOCK_MONOTONIC, mode);
2988         hrtimer_set_expires(&hs.timer, kt);
2989
2990         hrtimer_init_sleeper(&hs, current);
2991         do {
2992                 if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
2993                         break;
2994                 set_current_state(TASK_UNINTERRUPTIBLE);
2995                 hrtimer_start_expires(&hs.timer, mode);
2996                 if (hs.task)
2997                         io_schedule();
2998                 hrtimer_cancel(&hs.timer);
2999                 mode = HRTIMER_MODE_ABS;
3000         } while (hs.task && !signal_pending(current));
3001
3002         __set_current_state(TASK_RUNNING);
3003         destroy_hrtimer_on_stack(&hs.timer);
3004         return true;
3005 }
3006
3007 static bool __blk_mq_poll(struct blk_mq_hw_ctx *hctx, struct request *rq)
3008 {
3009         struct request_queue *q = hctx->queue;
3010         long state;
3011
3012         /*
3013          * If we sleep, have the caller restart the poll loop to reset
3014          * the state. Like for the other success return cases, the
3015          * caller is responsible for checking if the IO completed. If
3016          * the IO isn't complete, we'll get called again and will go
3017          * straight to the busy poll loop.
3018          */
3019         if (blk_mq_poll_hybrid_sleep(q, hctx, rq))
3020                 return true;
3021
3022         hctx->poll_considered++;
3023
3024         state = current->state;
3025         while (!need_resched()) {
3026                 int ret;
3027
3028                 hctx->poll_invoked++;
3029
3030                 ret = q->mq_ops->poll(hctx, rq->tag);
3031                 if (ret > 0) {
3032                         hctx->poll_success++;
3033                         set_current_state(TASK_RUNNING);
3034                         return true;
3035                 }
3036
3037                 if (signal_pending_state(state, current))
3038                         set_current_state(TASK_RUNNING);
3039
3040                 if (current->state == TASK_RUNNING)
3041                         return true;
3042                 if (ret < 0)
3043                         break;
3044                 cpu_relax();
3045         }
3046
3047         __set_current_state(TASK_RUNNING);
3048         return false;
3049 }
3050
3051 static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
3052 {
3053         struct blk_mq_hw_ctx *hctx;
3054         struct request *rq;
3055
3056         if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
3057                 return false;
3058
3059         hctx = q->queue_hw_ctx[blk_qc_t_to_queue_num(cookie)];
3060         if (!blk_qc_t_is_internal(cookie))
3061                 rq = blk_mq_tag_to_rq(hctx->tags, blk_qc_t_to_tag(cookie));
3062         else {
3063                 rq = blk_mq_tag_to_rq(hctx->sched_tags, blk_qc_t_to_tag(cookie));
3064                 /*
3065                  * With scheduling, if the request has completed, we'll
3066                  * get a NULL return here, as we clear the sched tag when
3067                  * that happens. The request still remains valid, like always,
3068                  * so we should be safe with just the NULL check.
3069                  */
3070                 if (!rq)
3071                         return false;
3072         }
3073
3074         return __blk_mq_poll(hctx, rq);
3075 }
3076
3077 static int __init blk_mq_init(void)
3078 {
3079         cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
3080                                 blk_mq_hctx_notify_dead);
3081         return 0;
3082 }
3083 subsys_initcall(blk_mq_init);