block/blk-mq-sched.c

   1 /*
   2  * blk-mq scheduling framework
   3  *
   4  * Copyright (C) 2016 Jens Axboe
   5  */
   6 #include <linux/kernel.h>
   7 #include <linux/module.h>
   8 #include <linux/blk-mq.h>
   9
  10 #include <trace/events/block.h>
  11
  12 #include "blk.h"
  13 #include "blk-mq.h"
  14 #include "blk-mq-debugfs.h"
  15 #include "blk-mq-sched.h"
  16 #include "blk-mq-tag.h"
  17 #include "blk-wbt.h"
  18
  19 void blk_mq_sched_free_hctx_data(struct request_queue *q,
  20                                  void (*exit)(struct blk_mq_hw_ctx *))
  21 {
  22         struct blk_mq_hw_ctx *hctx;
  23         int i;
  24
  25         queue_for_each_hw_ctx(q, hctx, i) {
  26                 if (exit && hctx->sched_data)
  27                         exit(hctx);
  28                 kfree(hctx->sched_data);
  29                 hctx->sched_data = NULL;
  30         }
  31 }
  32 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
  33
  34 void blk_mq_sched_assign_ioc(struct request *rq, struct bio *bio)
  35 {
  36         struct request_queue *q = rq->q;
  37         struct io_context *ioc = rq_ioc(bio);
  38         struct io_cq *icq;
  39
  40         spin_lock_irq(q->queue_lock);
  41         icq = ioc_lookup_icq(ioc, q);
  42         spin_unlock_irq(q->queue_lock);
  43
  44         if (!icq) {
  45                 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
  46                 if (!icq)
  47                         return;
  48         }
  49         get_io_context(icq->ioc);
  50         rq->elv.icq = icq;
  51 }
  52
  53 /*
  54  * Mark a hardware queue as needing a restart. For shared queues, maintain
  55  * a count of how many hardware queues are marked for restart.
  56  */
  57 static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
  58 {
  59         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
  60                 return;
  61
  62         if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
  63                 struct request_queue *q = hctx->queue;
  64
  65                 if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
  66                         atomic_inc(&q->shared_hctx_restart);
  67         } else
  68                 set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
  69 }
  70
  71 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
  72 {
  73         if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
  74                 return;
  75
  76         clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
  77
  78         if (blk_mq_hctx_has_pending(hctx)) {
  79                 blk_mq_run_hw_queue(hctx, true);
  80                 return;
  81         }
  82 }
  83
  84 /* return true if hctx need to run again */
  85 static bool blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
  86 {
  87         struct request_queue *q = hctx->queue;
  88         struct elevator_queue *e = q->elevator;
  89         LIST_HEAD(rq_list);
  90
  91         do {
  92                 struct request *rq;
  93                 blk_status_t ret;
  94
  95                 if (e->type->ops.mq.has_work &&
  96                                 !e->type->ops.mq.has_work(hctx))
  97                         break;
  98
  99                 ret = blk_mq_get_dispatch_budget(hctx);
 100                 if (ret == BLK_STS_RESOURCE)
 101                         return true;
 102
 103                 rq = e->type->ops.mq.dispatch_request(hctx);
 104                 if (!rq) {
 105                         blk_mq_put_dispatch_budget(hctx);
 106                         break;
 107                 } else if (ret != BLK_STS_OK) {
 108                         blk_mq_end_request(rq, ret);
 109                         continue;
 110                 }
 111
 112                 /*
 113                  * Now this rq owns the budget which has to be released
 114                  * if this rq won't be queued to driver via .queue_rq()
 115                  * in blk_mq_dispatch_rq_list().
 116                  */
 117                 list_add(&rq->queuelist, &rq_list);
 118         } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
 119
 120         return false;
 121 }
 122
 123 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
 124                                           struct blk_mq_ctx *ctx)
 125 {
 126         unsigned idx = ctx->index_hw;
 127
 128         if (++idx == hctx->nr_ctx)
 129                 idx = 0;
 130
 131         return hctx->ctxs[idx];
 132 }
 133
 134 /* return true if hctx need to run again */
 135 static bool blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
 136 {
 137         struct request_queue *q = hctx->queue;
 138         LIST_HEAD(rq_list);
 139         struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
 140
 141         do {
 142                 struct request *rq;
 143                 blk_status_t ret;
 144
 145                 if (!sbitmap_any_bit_set(&hctx->ctx_map))
 146                         break;
 147
 148                 ret = blk_mq_get_dispatch_budget(hctx);
 149                 if (ret == BLK_STS_RESOURCE)
 150                         return true;
 151
 152                 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
 153                 if (!rq) {
 154                         blk_mq_put_dispatch_budget(hctx);
 155                         break;
 156                 } else if (ret != BLK_STS_OK) {
 157                         blk_mq_end_request(rq, ret);
 158                         continue;
 159                 }
 160
 161                 /*
 162                  * Now this rq owns the budget which has to be released
 163                  * if this rq won't be queued to driver via .queue_rq()
 164                  * in blk_mq_dispatch_rq_list().
 165                  */
 166                 list_add(&rq->queuelist, &rq_list);
 167
 168                 /* round robin for fair dispatch */
 169                 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
 170
 171         } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
 172
 173         WRITE_ONCE(hctx->dispatch_from, ctx);
 174
 175         return false;
 176 }
 177
 178 /* return true if hw queue need to be run again */
 179 bool blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
 180 {
 181         struct request_queue *q = hctx->queue;
 182         struct elevator_queue *e = q->elevator;
 183         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
 184         LIST_HEAD(rq_list);
 185         bool run_queue = false;
 186
 187         /* RCU or SRCU read lock is needed before checking quiesced flag */
 188         if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
 189                 return false;
 190
 191         hctx->run++;
 192
 193         /*
 194          * If we have previous entries on our dispatch list, grab them first for
 195          * more fair dispatch.
 196          */
 197         if (!list_empty_careful(&hctx->dispatch)) {
 198                 spin_lock(&hctx->lock);
 199                 if (!list_empty(&hctx->dispatch))
 200                         list_splice_init(&hctx->dispatch, &rq_list);
 201                 spin_unlock(&hctx->lock);
 202         }
 203
 204         /*
 205          * Only ask the scheduler for requests, if we didn't have residual
 206          * requests from the dispatch list. This is to avoid the case where
 207          * we only ever dispatch a fraction of the requests available because
 208          * of low device queue depth. Once we pull requests out of the IO
 209          * scheduler, we can no longer merge or sort them. So it's best to
 210          * leave them there for as long as we can. Mark the hw queue as
 211          * needing a restart in that case.
 212          *
 213          * We want to dispatch from the scheduler if there was nothing
 214          * on the dispatch list or we were able to dispatch from the
 215          * dispatch list.
 216          */
 217         if (!list_empty(&rq_list)) {
 218                 blk_mq_sched_mark_restart_hctx(hctx);
 219                 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
 220                         if (has_sched_dispatch)
 221                                 run_queue = blk_mq_do_dispatch_sched(hctx);
 222                         else
 223                                 run_queue = blk_mq_do_dispatch_ctx(hctx);
 224                 }
 225         } else if (has_sched_dispatch) {
 226                 run_queue = blk_mq_do_dispatch_sched(hctx);
 227         } else if (q->mq_ops->get_budget) {
 228                 /*
 229                  * If we need to get budget before queuing request, we
 230                  * dequeue request one by one from sw queue for avoiding
 231                  * to mess up I/O merge when dispatch runs out of resource.
 232                  *
 233                  * TODO: get more budgets, and dequeue more requests in
 234                  * one time.
 235                  */
 236                 run_queue = blk_mq_do_dispatch_ctx(hctx);
 237         } else {
 238                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
 239                 blk_mq_dispatch_rq_list(q, &rq_list, false);
 240         }
 241
 242         if (run_queue && !blk_mq_sched_needs_restart(hctx) &&
 243                         !test_bit(BLK_MQ_S_TAG_WAITING, &hctx->state)) {
 244                 blk_mq_sched_mark_restart_hctx(hctx);
 245                 return true;
 246         }
 247
 248         return false;
 249 }
 250
 251 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 252                             struct request **merged_request)
 253 {
 254         struct request *rq;
 255
 256         switch (elv_merge(q, &rq, bio)) {
 257         case ELEVATOR_BACK_MERGE:
 258                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 259                         return false;
 260                 if (!bio_attempt_back_merge(q, rq, bio))
 261                         return false;
 262                 *merged_request = attempt_back_merge(q, rq);
 263                 if (!*merged_request)
 264                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 265                 return true;
 266         case ELEVATOR_FRONT_MERGE:
 267                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 268                         return false;
 269                 if (!bio_attempt_front_merge(q, rq, bio))
 270                         return false;
 271                 *merged_request = attempt_front_merge(q, rq);
 272                 if (!*merged_request)
 273                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 274                 return true;
 275         default:
 276                 return false;
 277         }
 278 }
 279 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 280
 281 /*
 282  * Reverse check our software queue for entries that we could potentially
 283  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 284  * too much time checking for merges.
 285  */
 286 static bool blk_mq_attempt_merge(struct request_queue *q,
 287                                  struct blk_mq_ctx *ctx, struct bio *bio)
 288 {
 289         struct request *rq;
 290         int checked = 8;
 291
 292         lockdep_assert_held(&ctx->lock);
 293
 294         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 295                 bool merged = false;
 296
 297                 if (!checked--)
 298                         break;
 299
 300                 if (!blk_rq_merge_ok(rq, bio))
 301                         continue;
 302
 303                 switch (blk_try_merge(rq, bio)) {
 304                 case ELEVATOR_BACK_MERGE:
 305                         if (blk_mq_sched_allow_merge(q, rq, bio))
 306                                 merged = bio_attempt_back_merge(q, rq, bio);
 307                         break;
 308                 case ELEVATOR_FRONT_MERGE:
 309                         if (blk_mq_sched_allow_merge(q, rq, bio))
 310                                 merged = bio_attempt_front_merge(q, rq, bio);
 311                         break;
 312                 case ELEVATOR_DISCARD_MERGE:
 313                         merged = bio_attempt_discard_merge(q, rq, bio);
 314                         break;
 315                 default:
 316                         continue;
 317                 }
 318
 319                 if (merged)
 320                         ctx->rq_merged++;
 321                 return merged;
 322         }
 323
 324         return false;
 325 }
 326
 327 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 328 {
 329         struct elevator_queue *e = q->elevator;
 330         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 331         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 332         bool ret = false;
 333
 334         if (e && e->type->ops.mq.bio_merge) {
 335                 blk_mq_put_ctx(ctx);
 336                 return e->type->ops.mq.bio_merge(hctx, bio);
 337         }
 338
 339         if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
 340                 /* default per sw-queue merge */
 341                 spin_lock(&ctx->lock);
 342                 ret = blk_mq_attempt_merge(q, ctx, bio);
 343                 spin_unlock(&ctx->lock);
 344         }
 345
 346         blk_mq_put_ctx(ctx);
 347         return ret;
 348 }
 349
 350 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 351 {
 352         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 353 }
 354 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 355
 356 void blk_mq_sched_request_inserted(struct request *rq)
 357 {
 358         trace_block_rq_insert(rq->q, rq);
 359 }
 360 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 361
 362 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
 363                                        struct request *rq)
 364 {
 365         if (rq->tag == -1) {
 366                 rq->rq_flags |= RQF_SORTED;
 367                 return false;
 368         }
 369
 370         /*
 371          * If we already have a real request tag, send directly to
 372          * the dispatch list.
 373          */
 374         spin_lock(&hctx->lock);
 375         list_add(&rq->queuelist, &hctx->dispatch);
 376         spin_unlock(&hctx->lock);
 377         return true;
 378 }
 379
 380 /*
 381  * Add flush/fua to the queue. If we fail getting a driver tag, then
 382  * punt to the requeue list. Requeue will re-invoke us from a context
 383  * that's safe to block from.
 384  */
 385 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
 386                                       struct request *rq, bool can_block)
 387 {
 388         if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
 389                 blk_insert_flush(rq);
 390                 blk_mq_run_hw_queue(hctx, true);
 391         } else
 392                 blk_mq_add_to_requeue_list(rq, false, true);
 393 }
 394
 395 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 396                                  bool run_queue, bool async, bool can_block)
 397 {
 398         struct request_queue *q = rq->q;
 399         struct elevator_queue *e = q->elevator;
 400         struct blk_mq_ctx *ctx = rq->mq_ctx;
 401         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 402
 403         if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
 404                 blk_mq_sched_insert_flush(hctx, rq, can_block);
 405                 return;
 406         }
 407
 408         if (e && blk_mq_sched_bypass_insert(hctx, rq))
 409                 goto run;
 410
 411         if (e && e->type->ops.mq.insert_requests) {
 412                 LIST_HEAD(list);
 413
 414                 list_add(&rq->queuelist, &list);
 415                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 416         } else {
 417                 spin_lock(&ctx->lock);
 418                 __blk_mq_insert_request(hctx, rq, at_head);
 419                 spin_unlock(&ctx->lock);
 420         }
 421
 422 run:
 423         if (run_queue)
 424                 blk_mq_run_hw_queue(hctx, async);
 425 }
 426
 427 void blk_mq_sched_insert_requests(struct request_queue *q,
 428                                   struct blk_mq_ctx *ctx,
 429                                   struct list_head *list, bool run_queue_async)
 430 {
 431         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 432         struct elevator_queue *e = hctx->queue->elevator;
 433
 434         if (e) {
 435                 struct request *rq, *next;
 436
 437                 /*
 438                  * We bypass requests that already have a driver tag assigned,
 439                  * which should only be flushes. Flushes are only ever inserted
 440                  * as single requests, so we shouldn't ever hit the
 441                  * WARN_ON_ONCE() below (but let's handle it just in case).
 442                  */
 443                 list_for_each_entry_safe(rq, next, list, queuelist) {
 444                         if (WARN_ON_ONCE(rq->tag != -1)) {
 445                                 list_del_init(&rq->queuelist);
 446                                 blk_mq_sched_bypass_insert(hctx, rq);
 447                         }
 448                 }
 449         }
 450
 451         if (e && e->type->ops.mq.insert_requests)
 452                 e->type->ops.mq.insert_requests(hctx, list, false);
 453         else
 454                 blk_mq_insert_requests(hctx, ctx, list);
 455
 456         blk_mq_run_hw_queue(hctx, run_queue_async);
 457 }
 458
 459 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 460                                    struct blk_mq_hw_ctx *hctx,
 461                                    unsigned int hctx_idx)
 462 {
 463         if (hctx->sched_tags) {
 464                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 465                 blk_mq_free_rq_map(hctx->sched_tags);
 466                 hctx->sched_tags = NULL;
 467         }
 468 }
 469
 470 static int blk_mq_sched_alloc_tags(struct request_queue *q,
 471                                    struct blk_mq_hw_ctx *hctx,
 472                                    unsigned int hctx_idx)
 473 {
 474         struct blk_mq_tag_set *set = q->tag_set;
 475         int ret;
 476
 477         hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
 478                                                set->reserved_tags);
 479         if (!hctx->sched_tags)
 480                 return -ENOMEM;
 481
 482         ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
 483         if (ret)
 484                 blk_mq_sched_free_tags(set, hctx, hctx_idx);
 485
 486         return ret;
 487 }
 488
 489 static void blk_mq_sched_tags_teardown(struct request_queue *q)
 490 {
 491         struct blk_mq_tag_set *set = q->tag_set;
 492         struct blk_mq_hw_ctx *hctx;
 493         int i;
 494
 495         queue_for_each_hw_ctx(q, hctx, i)
 496                 blk_mq_sched_free_tags(set, hctx, i);
 497 }
 498
 499 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 500                            unsigned int hctx_idx)
 501 {
 502         struct elevator_queue *e = q->elevator;
 503         int ret;
 504
 505         if (!e)
 506                 return 0;
 507
 508         ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
 509         if (ret)
 510                 return ret;
 511
 512         if (e->type->ops.mq.init_hctx) {
 513                 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
 514                 if (ret) {
 515                         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 516                         return ret;
 517                 }
 518         }
 519
 520         blk_mq_debugfs_register_sched_hctx(q, hctx);
 521
 522         return 0;
 523 }
 524
 525 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 526                             unsigned int hctx_idx)
 527 {
 528         struct elevator_queue *e = q->elevator;
 529
 530         if (!e)
 531                 return;
 532
 533         blk_mq_debugfs_unregister_sched_hctx(hctx);
 534
 535         if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 536                 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
 537                 hctx->sched_data = NULL;
 538         }
 539
 540         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 541 }
 542
 543 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 544 {
 545         struct blk_mq_hw_ctx *hctx;
 546         struct elevator_queue *eq;
 547         unsigned int i;
 548         int ret;
 549
 550         if (!e) {
 551                 q->elevator = NULL;
 552                 return 0;
 553         }
 554
 555         /*
 556          * Default to double of smaller one between hw queue_depth and 128,
 557          * since we don't split into sync/async like the old code did.
 558          * Additionally, this is a per-hw queue depth.
 559          */
 560         q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
 561                                    BLKDEV_MAX_RQ);
 562
 563         queue_for_each_hw_ctx(q, hctx, i) {
 564                 ret = blk_mq_sched_alloc_tags(q, hctx, i);
 565                 if (ret)
 566                         goto err;
 567         }
 568
 569         ret = e->ops.mq.init_sched(q, e);
 570         if (ret)
 571                 goto err;
 572
 573         blk_mq_debugfs_register_sched(q);
 574
 575         queue_for_each_hw_ctx(q, hctx, i) {
 576                 if (e->ops.mq.init_hctx) {
 577                         ret = e->ops.mq.init_hctx(hctx, i);
 578                         if (ret) {
 579                                 eq = q->elevator;
 580                                 blk_mq_exit_sched(q, eq);
 581                                 kobject_put(&eq->kobj);
 582                                 return ret;
 583                         }
 584                 }
 585                 blk_mq_debugfs_register_sched_hctx(q, hctx);
 586         }
 587
 588         return 0;
 589
 590 err:
 591         blk_mq_sched_tags_teardown(q);
 592         q->elevator = NULL;
 593         return ret;
 594 }
 595
 596 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
 597 {
 598         struct blk_mq_hw_ctx *hctx;
 599         unsigned int i;
 600
 601         queue_for_each_hw_ctx(q, hctx, i) {
 602                 blk_mq_debugfs_unregister_sched_hctx(hctx);
 603                 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 604                         e->type->ops.mq.exit_hctx(hctx, i);
 605                         hctx->sched_data = NULL;
 606                 }
 607         }
 608         blk_mq_debugfs_unregister_sched(q);
 609         if (e->type->ops.mq.exit_sched)
 610                 e->type->ops.mq.exit_sched(e);
 611         blk_mq_sched_tags_teardown(q);
 612         q->elevator = NULL;
 613 }
 614
 615 int blk_mq_sched_init(struct request_queue *q)
 616 {
 617         int ret;
 618
 619         mutex_lock(&q->sysfs_lock);
 620         ret = elevator_init(q, NULL);
 621         mutex_unlock(&q->sysfs_lock);
 622
 623         return ret;
 624 }