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