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-sched.h"
  15 #include "blk-mq-tag.h"
  16 #include "blk-wbt.h"
  17
  18 void blk_mq_sched_free_hctx_data(struct request_queue *q,
  19                                  void (*exit)(struct blk_mq_hw_ctx *))
  20 {
  21         struct blk_mq_hw_ctx *hctx;
  22         int i;
  23
  24         queue_for_each_hw_ctx(q, hctx, i) {
  25                 if (exit && hctx->sched_data)
  26                         exit(hctx);
  27                 kfree(hctx->sched_data);
  28                 hctx->sched_data = NULL;
  29         }
  30 }
  31 EXPORT_SYMBOL_GPL(blk_mq_sched_free_hctx_data);
  32
  33 int blk_mq_sched_init_hctx_data(struct request_queue *q, size_t size,
  34                                 int (*init)(struct blk_mq_hw_ctx *),
  35                                 void (*exit)(struct blk_mq_hw_ctx *))
  36 {
  37         struct blk_mq_hw_ctx *hctx;
  38         int ret;
  39         int i;
  40
  41         queue_for_each_hw_ctx(q, hctx, i) {
  42                 hctx->sched_data = kmalloc_node(size, GFP_KERNEL, hctx->numa_node);
  43                 if (!hctx->sched_data) {
  44                         ret = -ENOMEM;
  45                         goto error;
  46                 }
  47
  48                 if (init) {
  49                         ret = init(hctx);
  50                         if (ret) {
  51                                 /*
  52                                  * We don't want to give exit() a partially
  53                                  * initialized sched_data. init() must clean up
  54                                  * if it fails.
  55                                  */
  56                                 kfree(hctx->sched_data);
  57                                 hctx->sched_data = NULL;
  58                                 goto error;
  59                         }
  60                 }
  61         }
  62
  63         return 0;
  64 error:
  65         blk_mq_sched_free_hctx_data(q, exit);
  66         return ret;
  67 }
  68 EXPORT_SYMBOL_GPL(blk_mq_sched_init_hctx_data);
  69
  70 static void __blk_mq_sched_assign_ioc(struct request_queue *q,
  71                                       struct request *rq,
  72                                       struct bio *bio,
  73                                       struct io_context *ioc)
  74 {
  75         struct io_cq *icq;
  76
  77         spin_lock_irq(q->queue_lock);
  78         icq = ioc_lookup_icq(ioc, q);
  79         spin_unlock_irq(q->queue_lock);
  80
  81         if (!icq) {
  82                 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
  83                 if (!icq)
  84                         return;
  85         }
  86
  87         rq->elv.icq = icq;
  88         if (!blk_mq_sched_get_rq_priv(q, rq, bio)) {
  89                 rq->rq_flags |= RQF_ELVPRIV;
  90                 get_io_context(icq->ioc);
  91                 return;
  92         }
  93
  94         rq->elv.icq = NULL;
  95 }
  96
  97 static void blk_mq_sched_assign_ioc(struct request_queue *q,
  98                                     struct request *rq, struct bio *bio)
  99 {
 100         struct io_context *ioc;
 101
 102         ioc = rq_ioc(bio);
 103         if (ioc)
 104                 __blk_mq_sched_assign_ioc(q, rq, bio, ioc);
 105 }
 106
 107 struct request *blk_mq_sched_get_request(struct request_queue *q,
 108                                          struct bio *bio,
 109                                          unsigned int op,
 110                                          struct blk_mq_alloc_data *data)
 111 {
 112         struct elevator_queue *e = q->elevator;
 113         struct request *rq;
 114
 115         blk_queue_enter_live(q);
 116         data->q = q;
 117         if (likely(!data->ctx))
 118                 data->ctx = blk_mq_get_ctx(q);
 119         if (likely(!data->hctx))
 120                 data->hctx = blk_mq_map_queue(q, data->ctx->cpu);
 121
 122         if (e) {
 123                 data->flags |= BLK_MQ_REQ_INTERNAL;
 124
 125                 /*
 126                  * Flush requests are special and go directly to the
 127                  * dispatch list.
 128                  */
 129                 if (!op_is_flush(op) && e->type->ops.mq.get_request) {
 130                         rq = e->type->ops.mq.get_request(q, op, data);
 131                         if (rq)
 132                                 rq->rq_flags |= RQF_QUEUED;
 133                 } else
 134                         rq = __blk_mq_alloc_request(data, op);
 135         } else {
 136                 rq = __blk_mq_alloc_request(data, op);
 137         }
 138
 139         if (rq) {
 140                 if (!op_is_flush(op)) {
 141                         rq->elv.icq = NULL;
 142                         if (e && e->type->icq_cache)
 143                                 blk_mq_sched_assign_ioc(q, rq, bio);
 144                 }
 145                 data->hctx->queued++;
 146                 return rq;
 147         }
 148
 149         blk_queue_exit(q);
 150         return NULL;
 151 }
 152
 153 void blk_mq_sched_put_request(struct request *rq)
 154 {
 155         struct request_queue *q = rq->q;
 156         struct elevator_queue *e = q->elevator;
 157
 158         if (rq->rq_flags & RQF_ELVPRIV) {
 159                 blk_mq_sched_put_rq_priv(rq->q, rq);
 160                 if (rq->elv.icq) {
 161                         put_io_context(rq->elv.icq->ioc);
 162                         rq->elv.icq = NULL;
 163                 }
 164         }
 165
 166         if ((rq->rq_flags & RQF_QUEUED) && e && e->type->ops.mq.put_request)
 167                 e->type->ops.mq.put_request(rq);
 168         else
 169                 blk_mq_finish_request(rq);
 170 }
 171
 172 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
 173 {
 174         struct request_queue *q = hctx->queue;
 175         struct elevator_queue *e = q->elevator;
 176         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
 177         bool did_work = false;
 178         LIST_HEAD(rq_list);
 179
 180         if (unlikely(blk_mq_hctx_stopped(hctx)))
 181                 return;
 182
 183         hctx->run++;
 184
 185         /*
 186          * If we have previous entries on our dispatch list, grab them first for
 187          * more fair dispatch.
 188          */
 189         if (!list_empty_careful(&hctx->dispatch)) {
 190                 spin_lock(&hctx->lock);
 191                 if (!list_empty(&hctx->dispatch))
 192                         list_splice_init(&hctx->dispatch, &rq_list);
 193                 spin_unlock(&hctx->lock);
 194         }
 195
 196         /*
 197          * Only ask the scheduler for requests, if we didn't have residual
 198          * requests from the dispatch list. This is to avoid the case where
 199          * we only ever dispatch a fraction of the requests available because
 200          * of low device queue depth. Once we pull requests out of the IO
 201          * scheduler, we can no longer merge or sort them. So it's best to
 202          * leave them there for as long as we can. Mark the hw queue as
 203          * needing a restart in that case.
 204          */
 205         if (!list_empty(&rq_list)) {
 206                 blk_mq_sched_mark_restart_hctx(hctx);
 207                 did_work = blk_mq_dispatch_rq_list(q, &rq_list);
 208         } else if (!has_sched_dispatch) {
 209                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
 210                 blk_mq_dispatch_rq_list(q, &rq_list);
 211         }
 212
 213         /*
 214          * We want to dispatch from the scheduler if we had no work left
 215          * on the dispatch list, OR if we did have work but weren't able
 216          * to make progress.
 217          */
 218         if (!did_work && has_sched_dispatch) {
 219                 do {
 220                         struct request *rq;
 221
 222                         rq = e->type->ops.mq.dispatch_request(hctx);
 223                         if (!rq)
 224                                 break;
 225                         list_add(&rq->queuelist, &rq_list);
 226                 } while (blk_mq_dispatch_rq_list(q, &rq_list));
 227         }
 228 }
 229
 230 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx *hctx,
 231                                    struct list_head *rq_list,
 232                                    struct request *(*get_rq)(struct blk_mq_hw_ctx *))
 233 {
 234         do {
 235                 struct request *rq;
 236
 237                 rq = get_rq(hctx);
 238                 if (!rq)
 239                         break;
 240
 241                 list_add_tail(&rq->queuelist, rq_list);
 242         } while (1);
 243 }
 244 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch);
 245
 246 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 247                             struct request **merged_request)
 248 {
 249         struct request *rq;
 250
 251         switch (elv_merge(q, &rq, bio)) {
 252         case ELEVATOR_BACK_MERGE:
 253                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 254                         return false;
 255                 if (!bio_attempt_back_merge(q, rq, bio))
 256                         return false;
 257                 *merged_request = attempt_back_merge(q, rq);
 258                 if (!*merged_request)
 259                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 260                 return true;
 261         case ELEVATOR_FRONT_MERGE:
 262                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 263                         return false;
 264                 if (!bio_attempt_front_merge(q, rq, bio))
 265                         return false;
 266                 *merged_request = attempt_front_merge(q, rq);
 267                 if (!*merged_request)
 268                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 269                 return true;
 270         default:
 271                 return false;
 272         }
 273 }
 274 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 275
 276 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 277 {
 278         struct elevator_queue *e = q->elevator;
 279
 280         if (e->type->ops.mq.bio_merge) {
 281                 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 282                 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 283
 284                 blk_mq_put_ctx(ctx);
 285                 return e->type->ops.mq.bio_merge(hctx, bio);
 286         }
 287
 288         return false;
 289 }
 290
 291 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 292 {
 293         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 294 }
 295 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 296
 297 void blk_mq_sched_request_inserted(struct request *rq)
 298 {
 299         trace_block_rq_insert(rq->q, rq);
 300 }
 301 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 302
 303 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
 304                                        struct request *rq)
 305 {
 306         if (rq->tag == -1) {
 307                 rq->rq_flags |= RQF_SORTED;
 308                 return false;
 309         }
 310
 311         /*
 312          * If we already have a real request tag, send directly to
 313          * the dispatch list.
 314          */
 315         spin_lock(&hctx->lock);
 316         list_add(&rq->queuelist, &hctx->dispatch);
 317         spin_unlock(&hctx->lock);
 318         return true;
 319 }
 320
 321 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
 322 {
 323         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
 324                 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 325                 if (blk_mq_hctx_has_pending(hctx)) {
 326                         blk_mq_run_hw_queue(hctx, true);
 327                         return true;
 328                 }
 329         }
 330         return false;
 331 }
 332
 333 /**
 334  * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
 335  * @pos:    loop cursor.
 336  * @skip:   the list element that will not be examined. Iteration starts at
 337  *          @skip->next.
 338  * @head:   head of the list to examine. This list must have at least one
 339  *          element, namely @skip.
 340  * @member: name of the list_head structure within typeof(*pos).
 341  */
 342 #define list_for_each_entry_rcu_rr(pos, skip, head, member)             \
 343         for ((pos) = (skip);                                            \
 344              (pos = (pos)->member.next != (head) ? list_entry_rcu(      \
 345                         (pos)->member.next, typeof(*pos), member) :     \
 346               list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
 347              (pos) != (skip); )
 348
 349 /*
 350  * Called after a driver tag has been freed to check whether a hctx needs to
 351  * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
 352  * queues in a round-robin fashion if the tag set of @hctx is shared with other
 353  * hardware queues.
 354  */
 355 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
 356 {
 357         struct blk_mq_tags *const tags = hctx->tags;
 358         struct blk_mq_tag_set *const set = hctx->queue->tag_set;
 359         struct request_queue *const queue = hctx->queue, *q;
 360         struct blk_mq_hw_ctx *hctx2;
 361         unsigned int i, j;
 362
 363         if (set->flags & BLK_MQ_F_TAG_SHARED) {
 364                 rcu_read_lock();
 365                 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
 366                                            tag_set_list) {
 367                         queue_for_each_hw_ctx(q, hctx2, i)
 368                                 if (hctx2->tags == tags &&
 369                                     blk_mq_sched_restart_hctx(hctx2))
 370                                         goto done;
 371                 }
 372                 j = hctx->queue_num + 1;
 373                 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
 374                         if (j == queue->nr_hw_queues)
 375                                 j = 0;
 376                         hctx2 = queue->queue_hw_ctx[j];
 377                         if (hctx2->tags == tags &&
 378                             blk_mq_sched_restart_hctx(hctx2))
 379                                 break;
 380                 }
 381 done:
 382                 rcu_read_unlock();
 383         } else {
 384                 blk_mq_sched_restart_hctx(hctx);
 385         }
 386 }
 387
 388 /*
 389  * Add flush/fua to the queue. If we fail getting a driver tag, then
 390  * punt to the requeue list. Requeue will re-invoke us from a context
 391  * that's safe to block from.
 392  */
 393 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
 394                                       struct request *rq, bool can_block)
 395 {
 396         if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
 397                 blk_insert_flush(rq);
 398                 blk_mq_run_hw_queue(hctx, true);
 399         } else
 400                 blk_mq_add_to_requeue_list(rq, false, true);
 401 }
 402
 403 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 404                                  bool run_queue, bool async, bool can_block)
 405 {
 406         struct request_queue *q = rq->q;
 407         struct elevator_queue *e = q->elevator;
 408         struct blk_mq_ctx *ctx = rq->mq_ctx;
 409         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 410
 411         if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
 412                 blk_mq_sched_insert_flush(hctx, rq, can_block);
 413                 return;
 414         }
 415
 416         if (e && blk_mq_sched_bypass_insert(hctx, rq))
 417                 goto run;
 418
 419         if (e && e->type->ops.mq.insert_requests) {
 420                 LIST_HEAD(list);
 421
 422                 list_add(&rq->queuelist, &list);
 423                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 424         } else {
 425                 spin_lock(&ctx->lock);
 426                 __blk_mq_insert_request(hctx, rq, at_head);
 427                 spin_unlock(&ctx->lock);
 428         }
 429
 430 run:
 431         if (run_queue)
 432                 blk_mq_run_hw_queue(hctx, async);
 433 }
 434
 435 void blk_mq_sched_insert_requests(struct request_queue *q,
 436                                   struct blk_mq_ctx *ctx,
 437                                   struct list_head *list, bool run_queue_async)
 438 {
 439         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 440         struct elevator_queue *e = hctx->queue->elevator;
 441
 442         if (e) {
 443                 struct request *rq, *next;
 444
 445                 /*
 446                  * We bypass requests that already have a driver tag assigned,
 447                  * which should only be flushes. Flushes are only ever inserted
 448                  * as single requests, so we shouldn't ever hit the
 449                  * WARN_ON_ONCE() below (but let's handle it just in case).
 450                  */
 451                 list_for_each_entry_safe(rq, next, list, queuelist) {
 452                         if (WARN_ON_ONCE(rq->tag != -1)) {
 453                                 list_del_init(&rq->queuelist);
 454                                 blk_mq_sched_bypass_insert(hctx, rq);
 455                         }
 456                 }
 457         }
 458
 459         if (e && e->type->ops.mq.insert_requests)
 460                 e->type->ops.mq.insert_requests(hctx, list, false);
 461         else
 462                 blk_mq_insert_requests(hctx, ctx, list);
 463
 464         blk_mq_run_hw_queue(hctx, run_queue_async);
 465 }
 466
 467 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 468                                    struct blk_mq_hw_ctx *hctx,
 469                                    unsigned int hctx_idx)
 470 {
 471         if (hctx->sched_tags) {
 472                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 473                 blk_mq_free_rq_map(hctx->sched_tags);
 474                 hctx->sched_tags = NULL;
 475         }
 476 }
 477
 478 static int blk_mq_sched_alloc_tags(struct request_queue *q,
 479                                    struct blk_mq_hw_ctx *hctx,
 480                                    unsigned int hctx_idx)
 481 {
 482         struct blk_mq_tag_set *set = q->tag_set;
 483         int ret;
 484
 485         hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
 486                                                set->reserved_tags);
 487         if (!hctx->sched_tags)
 488                 return -ENOMEM;
 489
 490         ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
 491         if (ret)
 492                 blk_mq_sched_free_tags(set, hctx, hctx_idx);
 493
 494         return ret;
 495 }
 496
 497 static void blk_mq_sched_tags_teardown(struct request_queue *q)
 498 {
 499         struct blk_mq_tag_set *set = q->tag_set;
 500         struct blk_mq_hw_ctx *hctx;
 501         int i;
 502
 503         queue_for_each_hw_ctx(q, hctx, i)
 504                 blk_mq_sched_free_tags(set, hctx, i);
 505 }
 506
 507 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 508                            unsigned int hctx_idx)
 509 {
 510         struct elevator_queue *e = q->elevator;
 511
 512         if (!e)
 513                 return 0;
 514
 515         return blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
 516 }
 517
 518 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 519                             unsigned int hctx_idx)
 520 {
 521         struct elevator_queue *e = q->elevator;
 522
 523         if (!e)
 524                 return;
 525
 526         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 527 }
 528
 529 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 530 {
 531         struct blk_mq_hw_ctx *hctx;
 532         unsigned int i;
 533         int ret;
 534
 535         if (!e) {
 536                 q->elevator = NULL;
 537                 return 0;
 538         }
 539
 540         /*
 541          * Default to 256, since we don't split into sync/async like the
 542          * old code did. Additionally, this is a per-hw queue depth.
 543          */
 544         q->nr_requests = 2 * BLKDEV_MAX_RQ;
 545
 546         queue_for_each_hw_ctx(q, hctx, i) {
 547                 ret = blk_mq_sched_alloc_tags(q, hctx, i);
 548                 if (ret)
 549                         goto err;
 550         }
 551
 552         ret = e->ops.mq.init_sched(q, e);
 553         if (ret)
 554                 goto err;
 555
 556         return 0;
 557
 558 err:
 559         blk_mq_sched_tags_teardown(q);
 560         q->elevator = NULL;
 561         return ret;
 562 }
 563
 564 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
 565 {
 566         if (e->type->ops.mq.exit_sched)
 567                 e->type->ops.mq.exit_sched(e);
 568         blk_mq_sched_tags_teardown(q);
 569         q->elevator = NULL;
 570 }
 571
 572 int blk_mq_sched_init(struct request_queue *q)
 573 {
 574         int ret;
 575
 576         mutex_lock(&q->sysfs_lock);
 577         ret = elevator_init(q, NULL);
 578         mutex_unlock(&q->sysfs_lock);
 579
 580         return ret;
 581 }