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 blk_mq_hw_ctx *hctx;
 114         struct blk_mq_ctx *ctx;
 115         struct request *rq;
 116
 117         blk_queue_enter_live(q);
 118         ctx = blk_mq_get_ctx(q);
 119         hctx = blk_mq_map_queue(q, ctx->cpu);
 120
 121         blk_mq_set_alloc_data(data, q, data->flags, ctx, hctx);
 122
 123         if (e) {
 124                 data->flags |= BLK_MQ_REQ_INTERNAL;
 125
 126                 /*
 127                  * Flush requests are special and go directly to the
 128                  * dispatch list.
 129                  */
 130                 if (!op_is_flush(op) && e->type->ops.mq.get_request) {
 131                         rq = e->type->ops.mq.get_request(q, op, data);
 132                         if (rq)
 133                                 rq->rq_flags |= RQF_QUEUED;
 134                 } else
 135                         rq = __blk_mq_alloc_request(data, op);
 136         } else {
 137                 rq = __blk_mq_alloc_request(data, op);
 138                 if (rq)
 139                         data->hctx->tags->rqs[rq->tag] = rq;
 140         }
 141
 142         if (rq) {
 143                 if (!op_is_flush(op)) {
 144                         rq->elv.icq = NULL;
 145                         if (e && e->type->icq_cache)
 146                                 blk_mq_sched_assign_ioc(q, rq, bio);
 147                 }
 148                 data->hctx->queued++;
 149                 return rq;
 150         }
 151
 152         blk_queue_exit(q);
 153         return NULL;
 154 }
 155
 156 void blk_mq_sched_put_request(struct request *rq)
 157 {
 158         struct request_queue *q = rq->q;
 159         struct elevator_queue *e = q->elevator;
 160
 161         if (rq->rq_flags & RQF_ELVPRIV) {
 162                 blk_mq_sched_put_rq_priv(rq->q, rq);
 163                 if (rq->elv.icq) {
 164                         put_io_context(rq->elv.icq->ioc);
 165                         rq->elv.icq = NULL;
 166                 }
 167         }
 168
 169         if ((rq->rq_flags & RQF_QUEUED) && e && e->type->ops.mq.put_request)
 170                 e->type->ops.mq.put_request(rq);
 171         else
 172                 blk_mq_finish_request(rq);
 173 }
 174
 175 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
 176 {
 177         struct elevator_queue *e = hctx->queue->elevator;
 178         const bool has_sched_dispatch = e && e->type->ops.mq.dispatch_request;
 179         bool did_work = false;
 180         LIST_HEAD(rq_list);
 181
 182         if (unlikely(blk_mq_hctx_stopped(hctx)))
 183                 return;
 184
 185         hctx->run++;
 186
 187         /*
 188          * If we have previous entries on our dispatch list, grab them first for
 189          * more fair dispatch.
 190          */
 191         if (!list_empty_careful(&hctx->dispatch)) {
 192                 spin_lock(&hctx->lock);
 193                 if (!list_empty(&hctx->dispatch))
 194                         list_splice_init(&hctx->dispatch, &rq_list);
 195                 spin_unlock(&hctx->lock);
 196         }
 197
 198         /*
 199          * Only ask the scheduler for requests, if we didn't have residual
 200          * requests from the dispatch list. This is to avoid the case where
 201          * we only ever dispatch a fraction of the requests available because
 202          * of low device queue depth. Once we pull requests out of the IO
 203          * scheduler, we can no longer merge or sort them. So it's best to
 204          * leave them there for as long as we can. Mark the hw queue as
 205          * needing a restart in that case.
 206          */
 207         if (!list_empty(&rq_list)) {
 208                 blk_mq_sched_mark_restart_hctx(hctx);
 209                 did_work = blk_mq_dispatch_rq_list(hctx, &rq_list);
 210         } else if (!has_sched_dispatch) {
 211                 blk_mq_flush_busy_ctxs(hctx, &rq_list);
 212                 blk_mq_dispatch_rq_list(hctx, &rq_list);
 213         }
 214
 215         /*
 216          * We want to dispatch from the scheduler if we had no work left
 217          * on the dispatch list, OR if we did have work but weren't able
 218          * to make progress.
 219          */
 220         if (!did_work && has_sched_dispatch) {
 221                 do {
 222                         struct request *rq;
 223
 224                         rq = e->type->ops.mq.dispatch_request(hctx);
 225                         if (!rq)
 226                                 break;
 227                         list_add(&rq->queuelist, &rq_list);
 228                 } while (blk_mq_dispatch_rq_list(hctx, &rq_list));
 229         }
 230 }
 231
 232 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx *hctx,
 233                                    struct list_head *rq_list,
 234                                    struct request *(*get_rq)(struct blk_mq_hw_ctx *))
 235 {
 236         do {
 237                 struct request *rq;
 238
 239                 rq = get_rq(hctx);
 240                 if (!rq)
 241                         break;
 242
 243                 list_add_tail(&rq->queuelist, rq_list);
 244         } while (1);
 245 }
 246 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch);
 247
 248 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
 249                             struct request **merged_request)
 250 {
 251         struct request *rq;
 252
 253         switch (elv_merge(q, &rq, bio)) {
 254         case ELEVATOR_BACK_MERGE:
 255                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 256                         return false;
 257                 if (!bio_attempt_back_merge(q, rq, bio))
 258                         return false;
 259                 *merged_request = attempt_back_merge(q, rq);
 260                 if (!*merged_request)
 261                         elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
 262                 return true;
 263         case ELEVATOR_FRONT_MERGE:
 264                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 265                         return false;
 266                 if (!bio_attempt_front_merge(q, rq, bio))
 267                         return false;
 268                 *merged_request = attempt_front_merge(q, rq);
 269                 if (!*merged_request)
 270                         elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
 271                 return true;
 272         default:
 273                 return false;
 274         }
 275 }
 276 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 277
 278 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 279 {
 280         struct elevator_queue *e = q->elevator;
 281
 282         if (e->type->ops.mq.bio_merge) {
 283                 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 284                 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 285
 286                 blk_mq_put_ctx(ctx);
 287                 return e->type->ops.mq.bio_merge(hctx, bio);
 288         }
 289
 290         return false;
 291 }
 292
 293 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 294 {
 295         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 296 }
 297 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 298
 299 void blk_mq_sched_request_inserted(struct request *rq)
 300 {
 301         trace_block_rq_insert(rq->q, rq);
 302 }
 303 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 304
 305 static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
 306                                        struct request *rq)
 307 {
 308         if (rq->tag == -1) {
 309                 rq->rq_flags |= RQF_SORTED;
 310                 return false;
 311         }
 312
 313         /*
 314          * If we already have a real request tag, send directly to
 315          * the dispatch list.
 316          */
 317         spin_lock(&hctx->lock);
 318         list_add(&rq->queuelist, &hctx->dispatch);
 319         spin_unlock(&hctx->lock);
 320         return true;
 321 }
 322
 323 static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
 324 {
 325         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
 326                 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 327                 if (blk_mq_hctx_has_pending(hctx))
 328                         blk_mq_run_hw_queue(hctx, true);
 329         }
 330 }
 331
 332 void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx *hctx)
 333 {
 334         struct request_queue *q = hctx->queue;
 335         unsigned int i;
 336
 337         if (test_bit(QUEUE_FLAG_RESTART, &q->queue_flags)) {
 338                 if (test_and_clear_bit(QUEUE_FLAG_RESTART, &q->queue_flags)) {
 339                         queue_for_each_hw_ctx(q, hctx, i)
 340                                 blk_mq_sched_restart_hctx(hctx);
 341                 }
 342         } else {
 343                 blk_mq_sched_restart_hctx(hctx);
 344         }
 345 }
 346
 347 /*
 348  * Add flush/fua to the queue. If we fail getting a driver tag, then
 349  * punt to the requeue list. Requeue will re-invoke us from a context
 350  * that's safe to block from.
 351  */
 352 static void blk_mq_sched_insert_flush(struct blk_mq_hw_ctx *hctx,
 353                                       struct request *rq, bool can_block)
 354 {
 355         if (blk_mq_get_driver_tag(rq, &hctx, can_block)) {
 356                 blk_insert_flush(rq);
 357                 blk_mq_run_hw_queue(hctx, true);
 358         } else
 359                 blk_mq_add_to_requeue_list(rq, false, true);
 360 }
 361
 362 void blk_mq_sched_insert_request(struct request *rq, bool at_head,
 363                                  bool run_queue, bool async, bool can_block)
 364 {
 365         struct request_queue *q = rq->q;
 366         struct elevator_queue *e = q->elevator;
 367         struct blk_mq_ctx *ctx = rq->mq_ctx;
 368         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 369
 370         if (rq->tag == -1 && op_is_flush(rq->cmd_flags)) {
 371                 blk_mq_sched_insert_flush(hctx, rq, can_block);
 372                 return;
 373         }
 374
 375         if (e && blk_mq_sched_bypass_insert(hctx, rq))
 376                 goto run;
 377
 378         if (e && e->type->ops.mq.insert_requests) {
 379                 LIST_HEAD(list);
 380
 381                 list_add(&rq->queuelist, &list);
 382                 e->type->ops.mq.insert_requests(hctx, &list, at_head);
 383         } else {
 384                 spin_lock(&ctx->lock);
 385                 __blk_mq_insert_request(hctx, rq, at_head);
 386                 spin_unlock(&ctx->lock);
 387         }
 388
 389 run:
 390         if (run_queue)
 391                 blk_mq_run_hw_queue(hctx, async);
 392 }
 393
 394 void blk_mq_sched_insert_requests(struct request_queue *q,
 395                                   struct blk_mq_ctx *ctx,
 396                                   struct list_head *list, bool run_queue_async)
 397 {
 398         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 399         struct elevator_queue *e = hctx->queue->elevator;
 400
 401         if (e) {
 402                 struct request *rq, *next;
 403
 404                 /*
 405                  * We bypass requests that already have a driver tag assigned,
 406                  * which should only be flushes. Flushes are only ever inserted
 407                  * as single requests, so we shouldn't ever hit the
 408                  * WARN_ON_ONCE() below (but let's handle it just in case).
 409                  */
 410                 list_for_each_entry_safe(rq, next, list, queuelist) {
 411                         if (WARN_ON_ONCE(rq->tag != -1)) {
 412                                 list_del_init(&rq->queuelist);
 413                                 blk_mq_sched_bypass_insert(hctx, rq);
 414                         }
 415                 }
 416         }
 417
 418         if (e && e->type->ops.mq.insert_requests)
 419                 e->type->ops.mq.insert_requests(hctx, list, false);
 420         else
 421                 blk_mq_insert_requests(hctx, ctx, list);
 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 int blk_mq_sched_setup(struct request_queue *q)
 438 {
 439         struct blk_mq_tag_set *set = q->tag_set;
 440         struct blk_mq_hw_ctx *hctx;
 441         int ret, i;
 442
 443         /*
 444          * Default to 256, since we don't split into sync/async like the
 445          * old code did. Additionally, this is a per-hw queue depth.
 446          */
 447         q->nr_requests = 2 * BLKDEV_MAX_RQ;
 448
 449         /*
 450          * We're switching to using an IO scheduler, so setup the hctx
 451          * scheduler tags and switch the request map from the regular
 452          * tags to scheduler tags. First allocate what we need, so we
 453          * can safely fail and fallback, if needed.
 454          */
 455         ret = 0;
 456         queue_for_each_hw_ctx(q, hctx, i) {
 457                 hctx->sched_tags = blk_mq_alloc_rq_map(set, i, q->nr_requests, 0);
 458                 if (!hctx->sched_tags) {
 459                         ret = -ENOMEM;
 460                         break;
 461                 }
 462                 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, i, q->nr_requests);
 463                 if (ret)
 464                         break;
 465         }
 466
 467         /*
 468          * If we failed, free what we did allocate
 469          */
 470         if (ret) {
 471                 queue_for_each_hw_ctx(q, hctx, i) {
 472                         if (!hctx->sched_tags)
 473                                 continue;
 474                         blk_mq_sched_free_tags(set, hctx, i);
 475                 }
 476
 477                 return ret;
 478         }
 479
 480         return 0;
 481 }
 482
 483 void blk_mq_sched_teardown(struct request_queue *q)
 484 {
 485         struct blk_mq_tag_set *set = q->tag_set;
 486         struct blk_mq_hw_ctx *hctx;
 487         int i;
 488
 489         queue_for_each_hw_ctx(q, hctx, i)
 490                 blk_mq_sched_free_tags(set, hctx, i);
 491 }
 492
 493 int blk_mq_sched_init(struct request_queue *q)
 494 {
 495         int ret;
 496
 497         mutex_lock(&q->sysfs_lock);
 498         ret = elevator_init(q, NULL);
 499         mutex_unlock(&q->sysfs_lock);
 500
 501         return ret;
 502 }