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 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
  54 {
  55         struct request_queue *q = hctx->queue;
  56         struct elevator_queue *e = q->elevator;
  57         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
  58         bool did_work = false;
  59         LIST_HEAD(rq_list);
  60
  61         /* RCU or SRCU read lock is needed before checking quiesced flag */
  62         if (unlikely(blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)))
  63                 return;
  64
  65         hctx->run++;
  66
  67         /*
  68          * If we have previous entries on our dispatch list, grab them first for
  69          * more fair dispatch.
  70          */
  71         if (!list_empty_careful(&hctx->dispatch)) {
  72                 spin_lock(&hctx->lock);
  73                 if (!list_empty(&hctx->dispatch))
  74                         list_splice_init(&hctx->dispatch, &rq_list);
  75                 spin_unlock(&hctx->lock);
  76         }
  77
  78         /*
  79          * Only ask the scheduler for requests, if we didn't have residual
  80          * requests from the dispatch list. This is to avoid the case where
  81          * we only ever dispatch a fraction of the requests available because
  82          * of low device queue depth. Once we pull requests out of the IO
  83          * scheduler, we can no longer merge or sort them. So it's best to
  84          * leave them there for as long as we can. Mark the hw queue as
  85          * needing a restart in that case.
  86          */
  87         if (!list_empty(&rq_list)) {
  88                 blk_mq_sched_mark_restart_hctx(hctx);
  89                 did_work = blk_mq_dispatch_rq_list(q, &rq_list);
  90         } else if (!has_sched_dispatch) {
  91                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
  92                 blk_mq_dispatch_rq_list(q, &rq_list);
  93         }
  94
  95         /*
  96          * We want to dispatch from the scheduler if we had no work left
  97          * on the dispatch list, OR if we did have work but weren't able
  98          * to make progress.
  99          */
 100         if (!did_work && has_sched_dispatch) {
 101                 do {
 102                         struct request *rq;
 103
 104                         rq = e->type->ops.mq.dispatch_request(hctx);
 105                         if (!rq)
 106                                 break;
 107                         list_add(&rq->queuelist, &rq_list);
 108                 } while (blk_mq_dispatch_rq_list(q, &rq_list));
 109         }
 110 }
 111
 112 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 113                             struct request **merged_request)
 114 {
 115         struct request *rq;
 116
 117         switch (elv_merge(q, &rq, bio)) {
 118         case ELEVATOR_BACK_MERGE:
 119                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 120                         return false;
 121                 if (!bio_attempt_back_merge(q, rq, bio))
 122                         return false;
 123                 *merged_request = attempt_back_merge(q, rq);
 124                 if (!*merged_request)
 125                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 126                 return true;
 127         case ELEVATOR_FRONT_MERGE:
 128                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 129                         return false;
 130                 if (!bio_attempt_front_merge(q, rq, bio))
 131                         return false;
 132                 *merged_request = attempt_front_merge(q, rq);
 133                 if (!*merged_request)
 134                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 135                 return true;
 136         default:
 137                 return false;
 138         }
 139 }
 140 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 141
 142 /*
 143  * Reverse check our software queue for entries that we could potentially
 144  * merge with. Currently includes a hand-wavy stop count of 8, to not spend
 145  * too much time checking for merges.
 146  */
 147 static bool blk_mq_attempt_merge(struct request_queue *q,
 148                                  struct blk_mq_ctx *ctx, struct bio *bio)
 149 {
 150         struct request *rq;
 151         int checked = 8;
 152
 153         lockdep_assert_held(&ctx->lock);
 154
 155         list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
 156                 bool merged = false;
 157
 158                 if (!checked--)
 159                         break;
 160
 161                 if (!blk_rq_merge_ok(rq, bio))
 162                         continue;
 163
 164                 switch (blk_try_merge(rq, bio)) {
 165                 case ELEVATOR_BACK_MERGE:
 166                         if (blk_mq_sched_allow_merge(q, rq, bio))
 167                                 merged = bio_attempt_back_merge(q, rq, bio);
 168                         break;
 169                 case ELEVATOR_FRONT_MERGE:
 170                         if (blk_mq_sched_allow_merge(q, rq, bio))
 171                                 merged = bio_attempt_front_merge(q, rq, bio);
 172                         break;
 173                 case ELEVATOR_DISCARD_MERGE:
 174                         merged = bio_attempt_discard_merge(q, rq, bio);
 175                         break;
 176                 default:
 177                         continue;
 178                 }
 179
 180                 if (merged)
 181                         ctx->rq_merged++;
 182                 return merged;
 183         }
 184
 185         return false;
 186 }
 187
 188 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 189 {
 190         struct elevator_queue *e = q->elevator;
 191         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 192         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 193         bool ret = false;
 194
 195         if (e && e->type->ops.mq.bio_merge) {
 196                 blk_mq_put_ctx(ctx);
 197                 return e->type->ops.mq.bio_merge(hctx, bio);
 198         }
 199
 200         if (hctx->flags & BLK_MQ_F_SHOULD_MERGE) {
 201                 /* default per sw-queue merge */
 202                 spin_lock(&ctx->lock);
 203                 ret = blk_mq_attempt_merge(q, ctx, bio);
 204                 spin_unlock(&ctx->lock);
 205         }
 206
 207         blk_mq_put_ctx(ctx);
 208         return ret;
 209 }
 210
 211 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 212 {
 213         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 214 }
 215 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 216
 217 void blk_mq_sched_request_inserted(struct request *rq)
 218 {
 219         trace_block_rq_insert(rq->q, rq);
 220 }
 221 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 222
 223 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
 224                                        struct request *rq)
 225 {
 226         if (rq->tag == -1) {
 227                 rq->rq_flags |= RQF_SORTED;
 228                 return false;
 229         }
 230
 231         /*
 232          * If we already have a real request tag, send directly to
 233          * the dispatch list.
 234          */
 235         spin_lock(&hctx->lock);
 236         list_add(&rq->queuelist, &hctx->dispatch);
 237         spin_unlock(&hctx->lock);
 238         return true;
 239 }
 240
 241 static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
 242 {
 243         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
 244                 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 245                 if (blk_mq_hctx_has_pending(hctx)) {
 246                         blk_mq_run_hw_queue(hctx, true);
 247                         return true;
 248                 }
 249         }
 250         return false;
 251 }
 252
 253 /**
 254  * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
 255  * @pos:    loop cursor.
 256  * @skip:   the list element that will not be examined. Iteration starts at
 257  *          @skip->next.
 258  * @head:   head of the list to examine. This list must have at least one
 259  *          element, namely @skip.
 260  * @member: name of the list_head structure within typeof(*pos).
 261  */
 262 #define list_for_each_entry_rcu_rr(pos, skip, head, member)             \
 263         for ((pos) = (skip);                                            \
 264              (pos = (pos)->member.next != (head) ? list_entry_rcu(      \
 265                         (pos)->member.next, typeof(*pos), member) :     \
 266               list_entry_rcu((pos)->member.next->next, typeof(*pos), member)), \
 267              (pos) != (skip); )
 268
 269 /*
 270  * Called after a driver tag has been freed to check whether a hctx needs to
 271  * be restarted. Restarts @hctx if its tag set is not shared. Restarts hardware
 272  * queues in a round-robin fashion if the tag set of @hctx is shared with other
 273  * hardware queues.
 274  */
 275 void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
 276 {
 277         struct blk_mq_tags *const tags = hctx->tags;
 278         struct blk_mq_tag_set *const set = hctx->queue->tag_set;
 279         struct request_queue *const queue = hctx->queue, *q;
 280         struct blk_mq_hw_ctx *hctx2;
 281         unsigned int i, j;
 282
 283         if (set->flags & BLK_MQ_F_TAG_SHARED) {
 284                 rcu_read_lock();
 285                 list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
 286                                            tag_set_list) {
 287                         queue_for_each_hw_ctx(q, hctx2, i)
 288                                 if (hctx2->tags == tags &&
 289                                     blk_mq_sched_restart_hctx(hctx2))
 290                                         goto done;
 291                 }
 292                 j = hctx->queue_num + 1;
 293                 for (i = 0; i < queue->nr_hw_queues; i++, j++) {
 294                         if (j == queue->nr_hw_queues)
 295                                 j = 0;
 296                         hctx2 = queue->queue_hw_ctx[j];
 297                         if (hctx2->tags == tags &&
 298                             blk_mq_sched_restart_hctx(hctx2))
 299                                 break;
 300                 }
 301 done:
 302                 rcu_read_unlock();
 303         } else {
 304                 blk_mq_sched_restart_hctx(hctx);
 305         }
 306 }
 307
 308 /*
 309  * Add flush/fua to the queue. If we fail getting a driver tag, then
 310  * punt to the requeue list. Requeue will re-invoke us from a context
 311  * that's safe to block from.
 312  */
 313 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
 314                                       struct request *rq, bool can_block)
 315 {
 316         if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
 317                 blk_insert_flush(rq);
 318                 blk_mq_run_hw_queue(hctx, true);
 319         } else
 320                 blk_mq_add_to_requeue_list(rq, false, true);
 321 }
 322
 323 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 324                                  bool run_queue, bool async, bool can_block)
 325 {
 326         struct request_queue *q = rq->q;
 327         struct elevator_queue *e = q->elevator;
 328         struct blk_mq_ctx *ctx = rq->mq_ctx;
 329         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 330
 331         if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
 332                 blk_mq_sched_insert_flush(hctx, rq, can_block);
 333                 return;
 334         }
 335
 336         if (e && blk_mq_sched_bypass_insert(hctx, rq))
 337                 goto run;
 338
 339         if (e && e->type->ops.mq.insert_requests) {
 340                 LIST_HEAD(list);
 341
 342                 list_add(&rq->queuelist, &list);
 343                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 344         } else {
 345                 spin_lock(&ctx->lock);
 346                 __blk_mq_insert_request(hctx, rq, at_head);
 347                 spin_unlock(&ctx->lock);
 348         }
 349
 350 run:
 351         if (run_queue)
 352                 blk_mq_run_hw_queue(hctx, async);
 353 }
 354
 355 void blk_mq_sched_insert_requests(struct request_queue *q,
 356                                   struct blk_mq_ctx *ctx,
 357                                   struct list_head *list, bool run_queue_async)
 358 {
 359         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 360         struct elevator_queue *e = hctx->queue->elevator;
 361
 362         if (e) {
 363                 struct request *rq, *next;
 364
 365                 /*
 366                  * We bypass requests that already have a driver tag assigned,
 367                  * which should only be flushes. Flushes are only ever inserted
 368                  * as single requests, so we shouldn't ever hit the
 369                  * WARN_ON_ONCE() below (but let's handle it just in case).
 370                  */
 371                 list_for_each_entry_safe(rq, next, list, queuelist) {
 372                         if (WARN_ON_ONCE(rq->tag != -1)) {
 373                                 list_del_init(&rq->queuelist);
 374                                 blk_mq_sched_bypass_insert(hctx, rq);
 375                         }
 376                 }
 377         }
 378
 379         if (e && e->type->ops.mq.insert_requests)
 380                 e->type->ops.mq.insert_requests(hctx, list, false);
 381         else
 382                 blk_mq_insert_requests(hctx, ctx, list);
 383
 384         blk_mq_run_hw_queue(hctx, run_queue_async);
 385 }
 386
 387 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 388                                    struct blk_mq_hw_ctx *hctx,
 389                                    unsigned int hctx_idx)
 390 {
 391         if (hctx->sched_tags) {
 392                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 393                 blk_mq_free_rq_map(hctx->sched_tags);
 394                 hctx->sched_tags = NULL;
 395         }
 396 }
 397
 398 static int blk_mq_sched_alloc_tags(struct request_queue *q,
 399                                    struct blk_mq_hw_ctx *hctx,
 400                                    unsigned int hctx_idx)
 401 {
 402         struct blk_mq_tag_set *set = q->tag_set;
 403         int ret;
 404
 405         hctx->sched_tags = blk_mq_alloc_rq_map(set, hctx_idx, q->nr_requests,
 406                                                set->reserved_tags);
 407         if (!hctx->sched_tags)
 408                 return -ENOMEM;
 409
 410         ret = blk_mq_alloc_rqs(set, hctx->sched_tags, hctx_idx, q->nr_requests);
 411         if (ret)
 412                 blk_mq_sched_free_tags(set, hctx, hctx_idx);
 413
 414         return ret;
 415 }
 416
 417 static void blk_mq_sched_tags_teardown(struct request_queue *q)
 418 {
 419         struct blk_mq_tag_set *set = q->tag_set;
 420         struct blk_mq_hw_ctx *hctx;
 421         int i;
 422
 423         queue_for_each_hw_ctx(q, hctx, i)
 424                 blk_mq_sched_free_tags(set, hctx, i);
 425 }
 426
 427 int blk_mq_sched_init_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 428                            unsigned int hctx_idx)
 429 {
 430         struct elevator_queue *e = q->elevator;
 431         int ret;
 432
 433         if (!e)
 434                 return 0;
 435
 436         ret = blk_mq_sched_alloc_tags(q, hctx, hctx_idx);
 437         if (ret)
 438                 return ret;
 439
 440         if (e->type->ops.mq.init_hctx) {
 441                 ret = e->type->ops.mq.init_hctx(hctx, hctx_idx);
 442                 if (ret) {
 443                         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 444                         return ret;
 445                 }
 446         }
 447
 448         blk_mq_debugfs_register_sched_hctx(q, hctx);
 449
 450         return 0;
 451 }
 452
 453 void blk_mq_sched_exit_hctx(struct request_queue *q, struct blk_mq_hw_ctx *hctx,
 454                             unsigned int hctx_idx)
 455 {
 456         struct elevator_queue *e = q->elevator;
 457
 458         if (!e)
 459                 return;
 460
 461         blk_mq_debugfs_unregister_sched_hctx(hctx);
 462
 463         if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 464                 e->type->ops.mq.exit_hctx(hctx, hctx_idx);
 465                 hctx->sched_data = NULL;
 466         }
 467
 468         blk_mq_sched_free_tags(q->tag_set, hctx, hctx_idx);
 469 }
 470
 471 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 472 {
 473         struct blk_mq_hw_ctx *hctx;
 474         struct elevator_queue *eq;
 475         unsigned int i;
 476         int ret;
 477
 478         if (!e) {
 479                 q->elevator = NULL;
 480                 return 0;
 481         }
 482
 483         /*
 484          * Default to 256, since we don't split into sync/async like the
 485          * old code did. Additionally, this is a per-hw queue depth.
 486          */
 487         q->nr_requests = 2 * BLKDEV_MAX_RQ;
 488
 489         queue_for_each_hw_ctx(q, hctx, i) {
 490                 ret = blk_mq_sched_alloc_tags(q, hctx, i);
 491                 if (ret)
 492                         goto err;
 493         }
 494
 495         ret = e->ops.mq.init_sched(q, e);
 496         if (ret)
 497                 goto err;
 498
 499         blk_mq_debugfs_register_sched(q);
 500
 501         queue_for_each_hw_ctx(q, hctx, i) {
 502                 if (e->ops.mq.init_hctx) {
 503                         ret = e->ops.mq.init_hctx(hctx, i);
 504                         if (ret) {
 505                                 eq = q->elevator;
 506                                 blk_mq_exit_sched(q, eq);
 507                                 kobject_put(&eq->kobj);
 508                                 return ret;
 509                         }
 510                 }
 511                 blk_mq_debugfs_register_sched_hctx(q, hctx);
 512         }
 513
 514         return 0;
 515
 516 err:
 517         blk_mq_sched_tags_teardown(q);
 518         q->elevator = NULL;
 519         return ret;
 520 }
 521
 522 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e)
 523 {
 524         struct blk_mq_hw_ctx *hctx;
 525         unsigned int i;
 526
 527         queue_for_each_hw_ctx(q, hctx, i) {
 528                 blk_mq_debugfs_unregister_sched_hctx(hctx);
 529                 if (e->type->ops.mq.exit_hctx && hctx->sched_data) {
 530                         e->type->ops.mq.exit_hctx(hctx, i);
 531                         hctx->sched_data = NULL;
 532                 }
 533         }
 534         blk_mq_debugfs_unregister_sched(q);
 535         if (e->type->ops.mq.exit_sched)
 536                 e->type->ops.mq.exit_sched(e);
 537         blk_mq_sched_tags_teardown(q);
 538         q->elevator = NULL;
 539 }
 540
 541 int blk_mq_sched_init(struct request_queue *q)
 542 {
 543         int ret;
 544
 545         mutex_lock(&q->sysfs_lock);
 546         ret = elevator_init(q, NULL);
 547         mutex_unlock(&q->sysfs_lock);
 548
 549         return ret;
 550 }