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 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         set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
  63 }
  64 EXPORT_SYMBOL_GPL(blk_mq_sched_mark_restart_hctx);
  65
  66 void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
  67 {
  68         if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
  69                 return;
  70         clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
  71
  72         blk_mq_run_hw_queue(hctx, true);
  73 }
  74
  75 /*
  76  * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
  77  * its queue by itself in its completion handler, so we don't need to
  78  * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
  79  */
  80 static void blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
  81 {
  82         struct request_queue *q = hctx->queue;
  83         struct elevator_queue *e = q->elevator;
  84         LIST_HEAD(rq_list);
  85
  86         do {
  87                 struct request *rq;
  88
  89                 if (e->type->ops.mq.has_work &&
  90                                 !e->type->ops.mq.has_work(hctx))
  91                         break;
  92
  93                 if (!blk_mq_get_dispatch_budget(hctx))
  94                         break;
  95
  96                 rq = e->type->ops.mq.dispatch_request(hctx);
  97                 if (!rq) {
  98                         blk_mq_put_dispatch_budget(hctx);
  99                         break;
 100                 }
 101
 102                 /*
 103                  * Now this rq owns the budget which has to be released
 104                  * if this rq won't be queued to driver via .queue_rq()
 105                  * in blk_mq_dispatch_rq_list().
 106                  */
 107                 list_add(&rq->queuelist, &rq_list);
 108         } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
 109 }
 110
 111 static struct blk_mq_ctx *blk_mq_next_ctx(struct blk_mq_hw_ctx *hctx,
 112                                           struct blk_mq_ctx *ctx)
 113 {
 114         unsigned idx = ctx->index_hw;
 115
 116         if (++idx == hctx->nr_ctx)
 117                 idx = 0;
 118
 119         return hctx->ctxs[idx];
 120 }
 121
 122 /*
 123  * Only SCSI implements .get_budget and .put_budget, and SCSI restarts
 124  * its queue by itself in its completion handler, so we don't need to
 125  * restart queue if .get_budget() returns BLK_STS_NO_RESOURCE.
 126  */
 127 static void blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
 128 {
 129         struct request_queue *q = hctx->queue;
 130         LIST_HEAD(rq_list);
 131         struct blk_mq_ctx *ctx = READ_ONCE(hctx->dispatch_from);
 132
 133         do {
 134                 struct request *rq;
 135
 136                 if (!sbitmap_any_bit_set(&hctx->ctx_map))
 137                         break;
 138
 139                 if (!blk_mq_get_dispatch_budget(hctx))
 140                         break;
 141
 142                 rq = blk_mq_dequeue_from_ctx(hctx, ctx);
 143                 if (!rq) {
 144                         blk_mq_put_dispatch_budget(hctx);
 145                         break;
 146                 }
 147
 148                 /*
 149                  * Now this rq owns the budget which has to be released
 150                  * if this rq won't be queued to driver via .queue_rq()
 151                  * in blk_mq_dispatch_rq_list().
 152                  */
 153                 list_add(&rq->queuelist, &rq_list);
 154
 155                 /* round robin for fair dispatch */
 156                 ctx = blk_mq_next_ctx(hctx, rq->mq_ctx);
 157
 158         } while (blk_mq_dispatch_rq_list(q, &rq_list, true));
 159
 160         WRITE_ONCE(hctx->dispatch_from, ctx);
 161 }
 162
 163 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
 164 {
 165         struct request_queue *q = hctx->queue;
 166         struct elevator_queue *e = q->elevator;
 167         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
 168         LIST_HEAD(rq_list);
 169
 170         /* RCU or SRCU read lock is needed before checking quiesced flag */
 171         if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
 172                 return;
 173
 174         hctx->run++;
 175
 176         /*
 177          * If we have previous entries on our dispatch list, grab them first for
 178          * more fair dispatch.
 179          */
 180         if (!list_empty_careful(&hctx->dispatch)) {
 181                 spin_lock(&hctx->lock);
 182                 if (!list_empty(&hctx->dispatch))
 183                         list_splice_init(&hctx->dispatch, &rq_list);
 184                 spin_unlock(&hctx->lock);
 185         }
 186
 187         /*
 188          * Only ask the scheduler for requests, if we didn't have residual
 189          * requests from the dispatch list. This is to avoid the case where
 190          * we only ever dispatch a fraction of the requests available because
 191          * of low device queue depth. Once we pull requests out of the IO
 192          * scheduler, we can no longer merge or sort them. So it's best to
 193          * leave them there for as long as we can. Mark the hw queue as
 194          * needing a restart in that case.
 195          *
 196          * We want to dispatch from the scheduler if there was nothing
 197          * on the dispatch list or we were able to dispatch from the
 198          * dispatch list.
 199          */
 200         if (!list_empty(&rq_list)) {
 201                 blk_mq_sched_mark_restart_hctx(hctx);
 202                 if (blk_mq_dispatch_rq_list(q, &rq_list, false)) {
 203                         if (has_sched_dispatch)
 204                                 blk_mq_do_dispatch_sched(hctx);
 205                         else
 206                                 blk_mq_do_dispatch_ctx(hctx);
 207                 }
 208         } else if (has_sched_dispatch) {
 209                 blk_mq_do_dispatch_sched(hctx);
 210         } else if (hctx->dispatch_busy) {
 211                 /* dequeue request one by one from sw queue if queue is busy */
 212                 blk_mq_do_dispatch_ctx(hctx);
 213         } else {
 214                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
 215                 blk_mq_dispatch_rq_list(q, &rq_list, false);
 216         }
 217 }
 218
 219 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 220                             struct request **merged_request)
 221 {
 222         struct request *rq;
 223
 224         switch (elv_merge(q, &rq, bio)) {
 225         case ELEVATOR_BACK_MERGE:
 226                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 227                         return false;
 228                 if (!bio_attempt_back_merge(q, rq, bio))
 229                         return false;
 230                 *merged_request = attempt_back_merge(q, rq);
 231                 if (!*merged_request)
 232                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 233                 return true;
 234         case ELEVATOR_FRONT_MERGE:
 235                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 236                         return false;
 237                 if (!bio_attempt_front_merge(q, rq, bio))
 238                         return false;
 239                 *merged_request = attempt_front_merge(q, rq);
 240                 if (!*merged_request)
 241                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 242                 return true;
 243         case ELEVATOR_DISCARD_MERGE:
 244                 return bio_attempt_discard_merge(q, rq, bio);
 245         default:
 246                 return false;
 247         }
 248 }
 249 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 250
 251 /*
 252  * Iterate list of requests and see if we can merge this bio with any
 253  * of them.
 254  */
 255 bool blk_mq_bio_list_merge(struct request_queue *q, struct list_head *list,
 256                            struct bio *bio)
 257 {
 258         struct request *rq;
 259         int checked = 8;
 260
 261         list_for_each_entry_reverse(rq, list, queuelist) {
 262                 bool merged = false;
 263
 264                 if (!checked--)
 265                         break;
 266
 267                 if (!blk_rq_merge_ok(rq, bio))
 268                         continue;
 269
 270                 switch (blk_try_merge(rq, bio)) {
 271                 case ELEVATOR_BACK_MERGE:
 272                         if (blk_mq_sched_allow_merge(q, rq, bio))
 273                                 merged = bio_attempt_back_merge(q, rq, bio);
 274                         break;
 275                 case ELEVATOR_FRONT_MERGE:
 276                         if (blk_mq_sched_allow_merge(q, rq, bio))
 277                                 merged = bio_attempt_front_merge(q, rq, bio);
 278                         break;
 279                 case ELEVATOR_DISCARD_MERGE:
 280                         merged = bio_attempt_discard_merge(q, rq, bio);
 281                         break;
 282                 default:
 283                         continue;
 284                 }
 285
 286                 return merged;
 287         }
 288
 289         return false;
 290 }
 291 EXPORT_SYMBOL_GPL(blk_mq_bio_list_merge);
 292
 293 /*
 294  * Reverse check our software queue for entries that we could potentially
 295  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 296  * too much time checking for merges.
 297  */
 298 static bool blk_mq_attempt_merge(struct request_queue *q,
 299                                  struct blk_mq_ctx *ctx, struct bio *bio)
 300 {
 301         lockdep_assert_held(&ctx->lock);
 302
 303         if (blk_mq_bio_list_merge(q, &ctx->rq_list, bio)) {
 304                 ctx->rq_merged++;
 305                 return true;
 306         }
 307
 308         return false;
 309 }
 310
 311 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 312 {
 313         struct elevator_queue *e = q->elevator;
 314         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 315         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 316         bool ret = false;
 317
 318         if (e && e->type->ops.mq.bio_merge) {
 319                 blk_mq_put_ctx(ctx);
 320                 return e->type->ops.mq.bio_merge(hctx, bio);
 321         }
 322
 323         if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
 324                         !list_empty_careful(&ctx->rq_list)) {
 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                                        bool has_sched,
 349                                        struct request *rq)
 350 {
 351         /* dispatch flush rq directly */
 352         if (rq->rq_flags & RQF_FLUSH_SEQ) {
 353                 spin_lock(&hctx->lock);
 354                 list_add(&rq->queuelist, &hctx->dispatch);
 355                 spin_unlock(&hctx->lock);
 356                 return true;
 357         }
 358
 359         if (has_sched)
 360                 rq->rq_flags |= RQF_SORTED;
 361
 362         return false;
 363 }
 364
 365 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 366                                  bool run_queue, bool async)
 367 {
 368         struct request_queue *q = rq->q;
 369         struct elevator_queue *e = q->elevator;
 370         struct blk_mq_ctx *ctx = rq->mq_ctx;
 371         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 372
 373         /* flush rq in flush machinery need to be dispatched directly */
 374         if (!(rq->rq_flags & RQF_FLUSH_SEQ) && op_is_flush(rq->cmd_flags)) {
 375                 blk_insert_flush(rq);
 376                 goto run;
 377         }
 378
 379         WARN_ON(e && (rq->tag != -1));
 380
 381         if (blk_mq_sched_bypass_insert(hctx, !!e, rq))
 382                 goto run;
 383
 384         if (e && e->type->ops.mq.insert_requests) {
 385                 LIST_HEAD(list);
 386
 387                 list_add(&rq->queuelist, &list);
 388                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 389         } else {
 390                 spin_lock(&ctx->lock);
 391                 __blk_mq_insert_request(hctx, rq, at_head);
 392                 spin_unlock(&ctx->lock);
 393         }
 394
 395 run:
 396         if (run_queue)
 397                 blk_mq_run_hw_queue(hctx, async);
 398 }
 399
 400 void blk_mq_sched_insert_requests(struct request_queue *q,
 401                                   struct blk_mq_ctx *ctx,
 402                                   struct list_head *list, bool run_queue_async)
 403 {
 404         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 405         struct elevator_queue *e = hctx->queue->elevator;
 406
 407         if (e && e->type->ops.mq.insert_requests)
 408                 e->type->ops.mq.insert_requests(hctx, list, false);
 409         else {
 410                 /*
 411                  * try to issue requests directly if the hw queue isn't
 412                  * busy in case of 'none' scheduler, and this way may save
 413                  * us one extra enqueue & dequeue to sw queue.
 414                  */
 415                 if (!hctx->dispatch_busy && !e && !run_queue_async) {
 416                         blk_mq_try_issue_list_directly(hctx, list);
 417                         if (list_empty(list))
 418                                 return;
 419                 }
 420                 blk_mq_insert_requests(hctx, ctx, list);
 421         }
 422
 423         blk_mq_run_hw_queue(hctx, run_queue_async);
 424 }
 425
 426 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 427                                    struct blk_mq_hw_ctx *hctx,
 428                                    unsigned int hctx_idx)
 429 {
 430         if (hctx->sched_tags) {
 431                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 432                 blk_mq_free_rq_map(hctx->sched_tags);
 433                 hctx->sched_tags = NULL;
 434         }
 435 }
 436
 437 static int blk_mq_sched_alloc_tags(struct request_queue *q,
 438                                    struct blk_mq_hw_ctx *hctx,
 439                                    unsigned int hctx_idx)
 440 {
 441         struct blk_mq_tag_set *set = q->tag_set;
 442         int ret;
 443
 444         hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
 445                                                set->reserved_tags);
 446         if (!hctx->sched_tags)
 447                 return -ENOMEM;
 448
 449         ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
 450         if (ret)
 451                 blk_mq_sched_free_tags(set, hctx, hctx_idx);
 452
 453         return ret;
 454 }
 455
 456 static void blk_mq_sched_tags_teardown(struct request_queue *q)
 457 {
 458         struct blk_mq_tag_set *set = q->tag_set;
 459         struct blk_mq_hw_ctx *hctx;
 460         int i;
 461
 462         queue_for_each_hw_ctx(q, hctx, i)
 463                 blk_mq_sched_free_tags(set, hctx, i);
 464 }
 465
 466 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 467 {
 468         struct blk_mq_hw_ctx *hctx;
 469         struct elevator_queue *eq;
 470         unsigned int i;
 471         int ret;
 472
 473         if (!e) {
 474                 q->elevator = NULL;
 475                 q->nr_requests = q->tag_set->queue_depth;
 476                 return 0;
 477         }
 478
 479         /*
 480          * Default to double of smaller one between hw queue_depth and 128,
 481          * since we don't split into sync/async like the old code did.
 482          * Additionally, this is a per-hw queue depth.
 483          */
 484         q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
 485                                    BLKDEV_MAX_RQ);
 486
 487         queue_for_each_hw_ctx(q, hctx, i) {
 488                 ret = blk_mq_sched_alloc_tags(q, hctx, i);
 489                 if (ret)
 490                         goto err;
 491         }
 492
 493         ret = e->ops.mq.init_sched(q, e);
 494         if (ret)
 495                 goto err;
 496
 497         blk_mq_debugfs_register_sched(q);
 498
 499         queue_for_each_hw_ctx(q, hctx, i) {
 500                 if (e->ops.mq.init_hctx) {
 501                         ret = e->ops.mq.init_hctx(hctx, i);
 502                         if (ret) {
 503                                 eq = q->elevator;
 504                                 blk_mq_exit_sched(q, eq);
 505                                 kobject_put(&eq->kobj);
 506                                 return ret;
 507                         }
 508                 }
 509                 blk_mq_debugfs_register_sched_hctx(q, hctx);
 510         }
 511
 512         return 0;
 513
 514 err:
 515         blk_mq_sched_tags_teardown(q);
 516         q->elevator = NULL;
 517         return ret;
 518 }
 519
 520 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
 521 {
 522         struct blk_mq_hw_ctx *hctx;
 523         unsigned int i;
 524
 525         queue_for_each_hw_ctx(q, hctx, i) {
 526                 blk_mq_debugfs_unregister_sched_hctx(hctx);
 527                 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 528                         e->type->ops.mq.exit_hctx(hctx, i);
 529                         hctx->sched_data = NULL;
 530                 }
 531         }
 532         blk_mq_debugfs_unregister_sched(q);
 533         if (e->type->ops.mq.exit_sched)
 534                 e->type->ops.mq.exit_sched(e);
 535         blk_mq_sched_tags_teardown(q);
 536         q->elevator = NULL;
 537 }