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 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
  72 {
  73         if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
  74                 return false;
  75
  76         if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
  77                 struct request_queue *q = hctx->queue;
  78
  79                 if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
  80                         atomic_dec(&q->shared_hctx_restart);
  81         } else
  82                 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
  83
  84         if (blk_mq_hctx_has_pending(hctx)) {
  85                 blk_mq_run_hw_queue(hctx, true);
  86                 return true;
  87         }
  88
  89         return false;
  90 }
  91
  92 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
  93 {
  94         struct request_queue *q = hctx->queue;
  95         struct elevator_queue *e = q->elevator;
  96         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
  97         bool do_sched_dispatch = true;
  98         LIST_HEAD(rq_list);
  99
 100         /* RCU or SRCU read lock is needed before checking quiesced flag */
 101         if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
 102                 return;
 103
 104         hctx->run++;
 105
 106         /*
 107          * If we have previous entries on our dispatch list, grab them first for
 108          * more fair dispatch.
 109          */
 110         if (!list_empty_careful(&hctx->dispatch)) {
 111                 spin_lock(&hctx->lock);
 112                 if (!list_empty(&hctx->dispatch))
 113                         list_splice_init(&hctx->dispatch, &rq_list);
 114                 spin_unlock(&hctx->lock);
 115         }
 116
 117         /*
 118          * Only ask the scheduler for requests, if we didn't have residual
 119          * requests from the dispatch list. This is to avoid the case where
 120          * we only ever dispatch a fraction of the requests available because
 121          * of low device queue depth. Once we pull requests out of the IO
 122          * scheduler, we can no longer merge or sort them. So it's best to
 123          * leave them there for as long as we can. Mark the hw queue as
 124          * needing a restart in that case.
 125          */
 126         if (!list_empty(&rq_list)) {
 127                 blk_mq_sched_mark_restart_hctx(hctx);
 128                 do_sched_dispatch = blk_mq_dispatch_rq_list(q, &rq_list);
 129         } else if (!has_sched_dispatch) {
 130                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
 131                 blk_mq_dispatch_rq_list(q, &rq_list);
 132         }
 133
 134         /*
 135          * We want to dispatch from the scheduler if there was nothing
 136          * on the dispatch list or we were able to dispatch from the
 137          * dispatch list.
 138          */
 139         if (do_sched_dispatch && has_sched_dispatch) {
 140                 do {
 141                         struct request *rq;
 142
 143                         rq = e->type->ops.mq.dispatch_request(hctx);
 144                         if (!rq)
 145                                 break;
 146                         list_add(&rq->queuelist, &rq_list);
 147                 } while (blk_mq_dispatch_rq_list(q, &rq_list));
 148         }
 149 }
 150
 151 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 152                             struct request **merged_request)
 153 {
 154         struct request *rq;
 155
 156         switch (elv_merge(q, &rq, bio)) {
 157         case ELEVATOR_BACK_MERGE:
 158                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 159                         return false;
 160                 if (!bio_attempt_back_merge(q, rq, bio))
 161                         return false;
 162                 *merged_request = attempt_back_merge(q, rq);
 163                 if (!*merged_request)
 164                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 165                 return true;
 166         case ELEVATOR_FRONT_MERGE:
 167                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 168                         return false;
 169                 if (!bio_attempt_front_merge(q, rq, bio))
 170                         return false;
 171                 *merged_request = attempt_front_merge(q, rq);
 172                 if (!*merged_request)
 173                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 174                 return true;
 175         default:
 176                 return false;
 177         }
 178 }
 179 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 180
 181 /*
 182  * Reverse check our software queue for entries that we could potentially
 183  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 184  * too much time checking for merges.
 185  */
 186 static bool blk_mq_attempt_merge(struct request_queue *q,
 187                                  struct blk_mq_ctx *ctx, struct bio *bio)
 188 {
 189         struct request *rq;
 190         int checked = 8;
 191
 192         lockdep_assert_held(&ctx->lock);
 193
 194         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 195                 bool merged = false;
 196
 197                 if (!checked--)
 198                         break;
 199
 200                 if (!blk_rq_merge_ok(rq, bio))
 201                         continue;
 202
 203                 switch (blk_try_merge(rq, bio)) {
 204                 case ELEVATOR_BACK_MERGE:
 205                         if (blk_mq_sched_allow_merge(q, rq, bio))
 206                                 merged = bio_attempt_back_merge(q, rq, bio);
 207                         break;
 208                 case ELEVATOR_FRONT_MERGE:
 209                         if (blk_mq_sched_allow_merge(q, rq, bio))
 210                                 merged = bio_attempt_front_merge(q, rq, bio);
 211                         break;
 212                 case ELEVATOR_DISCARD_MERGE:
 213                         merged = bio_attempt_discard_merge(q, rq, bio);
 214                         break;
 215                 default:
 216                         continue;
 217                 }
 218
 219                 if (merged)
 220                         ctx->rq_merged++;
 221                 return merged;
 222         }
 223
 224         return false;
 225 }
 226
 227 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 228 {
 229         struct elevator_queue *e = q->elevator;
 230         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 231         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 232         bool ret = false;
 233
 234         if (e && e->type->ops.mq.bio_merge) {
 235                 blk_mq_put_ctx(ctx);
 236                 return e->type->ops.mq.bio_merge(hctx, bio);
 237         }
 238
 239         if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
 240                         !list_empty_careful(&ctx->rq_list)) {
 241                 /* default per sw-queue merge */
 242                 spin_lock(&ctx->lock);
 243                 ret = blk_mq_attempt_merge(q, ctx, bio);
 244                 spin_unlock(&ctx->lock);
 245         }
 246
 247         blk_mq_put_ctx(ctx);
 248         return ret;
 249 }
 250
 251 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 252 {
 253         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 254 }
 255 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 256
 257 void blk_mq_sched_request_inserted(struct request *rq)
 258 {
 259         trace_block_rq_insert(rq->q, rq);
 260 }
 261 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 262
 263 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
 264                                        struct request *rq)
 265 {
 266         if (rq->tag == -1) {
 267                 rq->rq_flags |= RQF_SORTED;
 268                 return false;
 269         }
 270
 271         /*
 272          * If we already have a real request tag, send directly to
 273          * the dispatch list.
 274          */
 275         spin_lock(&hctx->lock);
 276         list_add(&rq->queuelist, &hctx->dispatch);
 277         spin_unlock(&hctx->lock);
 278         return true;
 279 }
 280
 281 /**
 282  * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
 283  * @pos:    loop cursor.
 284  * @skip:   the list element that will not be examined. Iteration starts at
 285  *          @skip->next.
 286  * @head:   head of the list to examine. This list must have at least one
 287  *          element, namely @skip.
 288  * @member: name of the list_head structure within typeof(*pos).
 289  */
 290 #define list_for_each_entry_rcu_rr(pos, skip, head, member)             \
 291         for ((pos) = (skip);                                            \
 292              (pos = (pos)->member.next != (head) ? list_entry_rcu(      \
 293                         (pos)->member.next, typeof(*pos), member) :     \
 294               list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
 295              (pos) != (skip); )
 296
 297 /*
 298  * Called after a driver tag has been freed to check whether a hctx needs to
 299  * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
 300  * queues in a round-robin fashion if the tag set of @hctx is shared with other
 301  * hardware queues.
 302  */
 303 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
 304 {
 305         struct blk_mq_tags *const tags = hctx->tags;
 306         struct blk_mq_tag_set *const set = hctx->queue->tag_set;
 307         struct request_queue *const queue = hctx->queue, *q;
 308         struct blk_mq_hw_ctx *hctx2;
 309         unsigned int i, j;
 310
 311         if (set->flags & BLK_MQ_F_TAG_SHARED) {
 312                 /*
 313                  * If this is 0, then we know that no hardware queues
 314                  * have RESTART marked. We're done.
 315                  */
 316                 if (!atomic_read(&queue->shared_hctx_restart))
 317                         return;
 318
 319                 rcu_read_lock();
 320                 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
 321                                            tag_set_list) {
 322                         queue_for_each_hw_ctx(q, hctx2, i)
 323                                 if (hctx2->tags == tags &&
 324                                     blk_mq_sched_restart_hctx(hctx2))
 325                                         goto done;
 326                 }
 327                 j = hctx->queue_num + 1;
 328                 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
 329                         if (j == queue->nr_hw_queues)
 330                                 j = 0;
 331                         hctx2 = queue->queue_hw_ctx[j];
 332                         if (hctx2->tags == tags &&
 333                             blk_mq_sched_restart_hctx(hctx2))
 334                                 break;
 335                 }
 336 done:
 337                 rcu_read_unlock();
 338         } else {
 339                 blk_mq_sched_restart_hctx(hctx);
 340         }
 341 }
 342
 343 /*
 344  * Add flush/fua to the queue. If we fail getting a driver tag, then
 345  * punt to the requeue list. Requeue will re-invoke us from a context
 346  * that's safe to block from.
 347  */
 348 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
 349                                       struct request *rq, bool can_block)
 350 {
 351         if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
 352                 blk_insert_flush(rq);
 353                 blk_mq_run_hw_queue(hctx, true);
 354         } else
 355                 blk_mq_add_to_requeue_list(rq, false, true);
 356 }
 357
 358 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 359                                  bool run_queue, bool async, bool can_block)
 360 {
 361         struct request_queue *q = rq->q;
 362         struct elevator_queue *e = q->elevator;
 363         struct blk_mq_ctx *ctx = rq->mq_ctx;
 364         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 365
 366         if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
 367                 blk_mq_sched_insert_flush(hctx, rq, can_block);
 368                 return;
 369         }
 370
 371         if (e && blk_mq_sched_bypass_insert(hctx, rq))
 372                 goto run;
 373
 374         if (e && e->type->ops.mq.insert_requests) {
 375                 LIST_HEAD(list);
 376
 377                 list_add(&rq->queuelist, &list);
 378                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 379         } else {
 380                 spin_lock(&ctx->lock);
 381                 __blk_mq_insert_request(hctx, rq, at_head);
 382                 spin_unlock(&ctx->lock);
 383         }
 384
 385 run:
 386         if (run_queue)
 387                 blk_mq_run_hw_queue(hctx, async);
 388 }
 389
 390 void blk_mq_sched_insert_requests(struct request_queue *q,
 391                                   struct blk_mq_ctx *ctx,
 392                                   struct list_head *list, bool run_queue_async)
 393 {
 394         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 395         struct elevator_queue *e = hctx->queue->elevator;
 396
 397         if (e) {
 398                 struct request *rq, *next;
 399
 400                 /*
 401                  * We bypass requests that already have a driver tag assigned,
 402                  * which should only be flushes. Flushes are only ever inserted
 403                  * as single requests, so we shouldn't ever hit the
 404                  * WARN_ON_ONCE() below (but let's handle it just in case).
 405                  */
 406                 list_for_each_entry_safe(rq, next, list, queuelist) {
 407                         if (WARN_ON_ONCE(rq->tag != -1)) {
 408                                 list_del_init(&rq->queuelist);
 409                                 blk_mq_sched_bypass_insert(hctx, rq);
 410                         }
 411                 }
 412         }
 413
 414         if (e && e->type->ops.mq.insert_requests)
 415                 e->type->ops.mq.insert_requests(hctx, list, false);
 416         else
 417                 blk_mq_insert_requests(hctx, ctx, list);
 418
 419         blk_mq_run_hw_queue(hctx, run_queue_async);
 420 }
 421
 422 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 423                                    struct blk_mq_hw_ctx *hctx,
 424                                    unsigned int hctx_idx)
 425 {
 426         if (hctx->sched_tags) {
 427                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 428                 blk_mq_free_rq_map(hctx->sched_tags);
 429                 hctx->sched_tags = NULL;
 430         }
 431 }
 432
 433 static int blk_mq_sched_alloc_tags(struct request_queue *q,
 434                                    struct blk_mq_hw_ctx *hctx,
 435                                    unsigned int hctx_idx)
 436 {
 437         struct blk_mq_tag_set *set = q->tag_set;
 438         int ret;
 439
 440         hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
 441                                                set->reserved_tags);
 442         if (!hctx->sched_tags)
 443                 return -ENOMEM;
 444
 445         ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
 446         if (ret)
 447                 blk_mq_sched_free_tags(set, hctx, hctx_idx);
 448
 449         return ret;
 450 }
 451
 452 static void blk_mq_sched_tags_teardown(struct request_queue *q)
 453 {
 454         struct blk_mq_tag_set *set = q->tag_set;
 455         struct blk_mq_hw_ctx *hctx;
 456         int i;
 457
 458         queue_for_each_hw_ctx(q, hctx, i)
 459                 blk_mq_sched_free_tags(set, hctx, i);
 460 }
 461
 462 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 463                            unsigned int hctx_idx)
 464 {
 465         struct elevator_queue *e = q->elevator;
 466         int ret;
 467
 468         if (!e)
 469                 return 0;
 470
 471         ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
 472         if (ret)
 473                 return ret;
 474
 475         if (e->type->ops.mq.init_hctx) {
 476                 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
 477                 if (ret) {
 478                         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 479                         return ret;
 480                 }
 481         }
 482
 483         blk_mq_debugfs_register_sched_hctx(q, hctx);
 484
 485         return 0;
 486 }
 487
 488 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 489                             unsigned int hctx_idx)
 490 {
 491         struct elevator_queue *e = q->elevator;
 492
 493         if (!e)
 494                 return;
 495
 496         blk_mq_debugfs_unregister_sched_hctx(hctx);
 497
 498         if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 499                 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
 500                 hctx->sched_data = NULL;
 501         }
 502
 503         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 504 }
 505
 506 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 507 {
 508         struct blk_mq_hw_ctx *hctx;
 509         struct elevator_queue *eq;
 510         unsigned int i;
 511         int ret;
 512
 513         if (!e) {
 514                 q->elevator = NULL;
 515                 return 0;
 516         }
 517
 518         /*
 519          * Default to double of smaller one between hw queue_depth and 128,
 520          * since we don't split into sync/async like the old code did.
 521          * Additionally, this is a per-hw queue depth.
 522          */
 523         q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
 524                                    BLKDEV_MAX_RQ);
 525
 526         queue_for_each_hw_ctx(q, hctx, i) {
 527                 ret = blk_mq_sched_alloc_tags(q, hctx, i);
 528                 if (ret)
 529                         goto err;
 530         }
 531
 532         ret = e->ops.mq.init_sched(q, e);
 533         if (ret)
 534                 goto err;
 535
 536         blk_mq_debugfs_register_sched(q);
 537
 538         queue_for_each_hw_ctx(q, hctx, i) {
 539                 if (e->ops.mq.init_hctx) {
 540                         ret = e->ops.mq.init_hctx(hctx, i);
 541                         if (ret) {
 542                                 eq = q->elevator;
 543                                 blk_mq_exit_sched(q, eq);
 544                                 kobject_put(&eq->kobj);
 545                                 return ret;
 546                         }
 547                 }
 548                 blk_mq_debugfs_register_sched_hctx(q, hctx);
 549         }
 550
 551         return 0;
 552
 553 err:
 554         blk_mq_sched_tags_teardown(q);
 555         q->elevator = NULL;
 556         return ret;
 557 }
 558
 559 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
 560 {
 561         struct blk_mq_hw_ctx *hctx;
 562         unsigned int i;
 563
 564         queue_for_each_hw_ctx(q, hctx, i) {
 565                 blk_mq_debugfs_unregister_sched_hctx(hctx);
 566                 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 567                         e->type->ops.mq.exit_hctx(hctx, i);
 568                         hctx->sched_data = NULL;
 569                 }
 570         }
 571         blk_mq_debugfs_unregister_sched(q);
 572         if (e->type->ops.mq.exit_sched)
 573                 e->type->ops.mq.exit_sched(e);
 574         blk_mq_sched_tags_teardown(q);
 575         q->elevator = NULL;
 576 }
 577
 578 int blk_mq_sched_init(struct request_queue *q)
 579 {
 580         int ret;
 581
 582         mutex_lock(&q->sysfs_lock);
 583         ret = elevator_init(q, NULL);
 584         mutex_unlock(&q->sysfs_lock);
 585
 586         return ret;
 587 }