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 elevator_queue *e = hctx->queue->elevator;
 175         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
 176         bool did_work = false;
 177         LIST_HEAD(rq_list);
 178
 179         if (unlikely(blk_mq_hctx_stopped(hctx)))
 180                 return;
 181
 182         hctx->run++;
 183
 184         /*
 185          * If we have previous entries on our dispatch list, grab them first for
 186          * more fair dispatch.
 187          */
 188         if (!list_empty_careful(&hctx->dispatch)) {
 189                 spin_lock(&hctx->lock);
 190                 if (!list_empty(&hctx->dispatch))
 191                         list_splice_init(&hctx->dispatch, &rq_list);
 192                 spin_unlock(&hctx->lock);
 193         }
 194
 195         /*
 196          * Only ask the scheduler for requests, if we didn't have residual
 197          * requests from the dispatch list. This is to avoid the case where
 198          * we only ever dispatch a fraction of the requests available because
 199          * of low device queue depth. Once we pull requests out of the IO
 200          * scheduler, we can no longer merge or sort them. So it's best to
 201          * leave them there for as long as we can. Mark the hw queue as
 202          * needing a restart in that case.
 203          */
 204         if (!list_empty(&rq_list)) {
 205                 blk_mq_sched_mark_restart_hctx(hctx);
 206                 did_work = blk_mq_dispatch_rq_list(hctx, &rq_list);
 207         } else if (!has_sched_dispatch) {
 208                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
 209                 blk_mq_dispatch_rq_list(hctx, &rq_list);
 210         }
 211
 212         /*
 213          * We want to dispatch from the scheduler if we had no work left
 214          * on the dispatch list, OR if we did have work but weren't able
 215          * to make progress.
 216          */
 217         if (!did_work && has_sched_dispatch) {
 218                 do {
 219                         struct request *rq;
 220
 221                         rq = e->type->ops.mq.dispatch_request(hctx);
 222                         if (!rq)
 223                                 break;
 224                         list_add(&rq->queuelist, &rq_list);
 225                 } while (blk_mq_dispatch_rq_list(hctx, &rq_list));
 226         }
 227 }
 228
 229 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx *hctx,
 230                                    struct list_head *rq_list,
 231                                    struct request *(*get_rq)(struct blk_mq_hw_ctx *))
 232 {
 233         do {
 234                 struct request *rq;
 235
 236                 rq = get_rq(hctx);
 237                 if (!rq)
 238                         break;
 239
 240                 list_add_tail(&rq->queuelist, rq_list);
 241         } while (1);
 242 }
 243 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch);
 244
 245 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 246                             struct request **merged_request)
 247 {
 248         struct request *rq;
 249
 250         switch (elv_merge(q, &rq, bio)) {
 251         case ELEVATOR_BACK_MERGE:
 252                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 253                         return false;
 254                 if (!bio_attempt_back_merge(q, rq, bio))
 255                         return false;
 256                 *merged_request = attempt_back_merge(q, rq);
 257                 if (!*merged_request)
 258                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 259                 return true;
 260         case ELEVATOR_FRONT_MERGE:
 261                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 262                         return false;
 263                 if (!bio_attempt_front_merge(q, rq, bio))
 264                         return false;
 265                 *merged_request = attempt_front_merge(q, rq);
 266                 if (!*merged_request)
 267                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 268                 return true;
 269         default:
 270                 return false;
 271         }
 272 }
 273 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 274
 275 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 276 {
 277         struct elevator_queue *e = q->elevator;
 278
 279         if (e->type->ops.mq.bio_merge) {
 280                 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 281                 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 282
 283                 blk_mq_put_ctx(ctx);
 284                 return e->type->ops.mq.bio_merge(hctx, bio);
 285         }
 286
 287         return false;
 288 }
 289
 290 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 291 {
 292         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 293 }
 294 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 295
 296 void blk_mq_sched_request_inserted(struct request *rq)
 297 {
 298         trace_block_rq_insert(rq->q, rq);
 299 }
 300 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 301
 302 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
 303                                        struct request *rq)
 304 {
 305         if (rq->tag == -1) {
 306                 rq->rq_flags |= RQF_SORTED;
 307                 return false;
 308         }
 309
 310         /*
 311          * If we already have a real request tag, send directly to
 312          * the dispatch list.
 313          */
 314         spin_lock(&hctx->lock);
 315         list_add(&rq->queuelist, &hctx->dispatch);
 316         spin_unlock(&hctx->lock);
 317         return true;
 318 }
 319
 320 static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
 321 {
 322         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
 323                 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 324                 if (blk_mq_hctx_has_pending(hctx))
 325                         blk_mq_run_hw_queue(hctx, true);
 326         }
 327 }
 328
 329 void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx *hctx)
 330 {
 331         struct request_queue *q = hctx->queue;
 332         unsigned int i;
 333
 334         if (test_bit(QUEUE_FLAG_RESTART, &q->queue_flags)) {
 335                 if (test_and_clear_bit(QUEUE_FLAG_RESTART, &q->queue_flags)) {
 336                         queue_for_each_hw_ctx(q, hctx, i)
 337                                 blk_mq_sched_restart_hctx(hctx);
 338                 }
 339         } else {
 340                 blk_mq_sched_restart_hctx(hctx);
 341         }
 342 }
 343
 344 /*
 345  * Add flush/fua to the queue. If we fail getting a driver tag, then
 346  * punt to the requeue list. Requeue will re-invoke us from a context
 347  * that's safe to block from.
 348  */
 349 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
 350                                       struct request *rq, bool can_block)
 351 {
 352         if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
 353                 blk_insert_flush(rq);
 354                 blk_mq_run_hw_queue(hctx, true);
 355         } else
 356                 blk_mq_add_to_requeue_list(rq, false, true);
 357 }
 358
 359 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 360                                  bool run_queue, bool async, bool can_block)
 361 {
 362         struct request_queue *q = rq->q;
 363         struct elevator_queue *e = q->elevator;
 364         struct blk_mq_ctx *ctx = rq->mq_ctx;
 365         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 366
 367         if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
 368                 blk_mq_sched_insert_flush(hctx, rq, can_block);
 369                 return;
 370         }
 371
 372         if (e && blk_mq_sched_bypass_insert(hctx, rq))
 373                 goto run;
 374
 375         if (e && e->type->ops.mq.insert_requests) {
 376                 LIST_HEAD(list);
 377
 378                 list_add(&rq->queuelist, &list);
 379                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 380         } else {
 381                 spin_lock(&ctx->lock);
 382                 __blk_mq_insert_request(hctx, rq, at_head);
 383                 spin_unlock(&ctx->lock);
 384         }
 385
 386 run:
 387         if (run_queue)
 388                 blk_mq_run_hw_queue(hctx, async);
 389 }
 390
 391 void blk_mq_sched_insert_requests(struct request_queue *q,
 392                                   struct blk_mq_ctx *ctx,
 393                                   struct list_head *list, bool run_queue_async)
 394 {
 395         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 396         struct elevator_queue *e = hctx->queue->elevator;
 397
 398         if (e) {
 399                 struct request *rq, *next;
 400
 401                 /*
 402                  * We bypass requests that already have a driver tag assigned,
 403                  * which should only be flushes. Flushes are only ever inserted
 404                  * as single requests, so we shouldn't ever hit the
 405                  * WARN_ON_ONCE() below (but let's handle it just in case).
 406                  */
 407                 list_for_each_entry_safe(rq, next, list, queuelist) {
 408                         if (WARN_ON_ONCE(rq->tag != -1)) {
 409                                 list_del_init(&rq->queuelist);
 410                                 blk_mq_sched_bypass_insert(hctx, rq);
 411                         }
 412                 }
 413         }
 414
 415         if (e && e->type->ops.mq.insert_requests)
 416                 e->type->ops.mq.insert_requests(hctx, list, false);
 417         else
 418                 blk_mq_insert_requests(hctx, ctx, list);
 419
 420         blk_mq_run_hw_queue(hctx, run_queue_async);
 421 }
 422
 423 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 424                                    struct blk_mq_hw_ctx *hctx,
 425                                    unsigned int hctx_idx)
 426 {
 427         if (hctx->sched_tags) {
 428                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 429                 blk_mq_free_rq_map(hctx->sched_tags);
 430                 hctx->sched_tags = NULL;
 431         }
 432 }
 433
 434 int blk_mq_sched_setup(struct request_queue *q)
 435 {
 436         struct blk_mq_tag_set *set = q->tag_set;
 437         struct blk_mq_hw_ctx *hctx;
 438         int ret, i;
 439
 440         /*
 441          * Default to 256, since we don't split into sync/async like the
 442          * old code did. Additionally, this is a per-hw queue depth.
 443          */
 444         q->nr_requests = 2 * BLKDEV_MAX_RQ;
 445
 446         /*
 447          * We're switching to using an IO scheduler, so setup the hctx
 448          * scheduler tags and switch the request map from the regular
 449          * tags to scheduler tags. First allocate what we need, so we
 450          * can safely fail and fallback, if needed.
 451          */
 452         ret = 0;
 453         queue_for_each_hw_ctx(q, hctx, i) {
 454                 hctx->sched_tags = blk_mq_alloc_rq_map(set, i,
 455                                 q->nr_requests, set->reserved_tags);
 456                 if (!hctx->sched_tags) {
 457                         ret = -ENOMEM;
 458                         break;
 459                 }
 460                 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, i, q->nr_requests);
 461                 if (ret)
 462                         break;
 463         }
 464
 465         /*
 466          * If we failed, free what we did allocate
 467          */
 468         if (ret) {
 469                 queue_for_each_hw_ctx(q, hctx, i) {
 470                         if (!hctx->sched_tags)
 471                                 continue;
 472                         blk_mq_sched_free_tags(set, hctx, i);
 473                 }
 474
 475                 return ret;
 476         }
 477
 478         return 0;
 479 }
 480
 481 void blk_mq_sched_teardown(struct request_queue *q)
 482 {
 483         struct blk_mq_tag_set *set = q->tag_set;
 484         struct blk_mq_hw_ctx *hctx;
 485         int i;
 486
 487         queue_for_each_hw_ctx(q, hctx, i)
 488                 blk_mq_sched_free_tags(set, hctx, i);
 489 }
 490
 491 int blk_mq_sched_init(struct request_queue *q)
 492 {
 493         int ret;
 494
 495         mutex_lock(&q->sysfs_lock);
 496         ret = elevator_init(q, NULL);
 497         mutex_unlock(&q->sysfs_lock);
 498
 499         return ret;
 500 }