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, struct io_context *ioc)
  72 {
  73         struct io_cq *icq;
  74
  75         spin_lock_irq(q->queue_lock);
  76         icq = ioc_lookup_icq(ioc, q);
  77         spin_unlock_irq(q->queue_lock);
  78
  79         if (!icq) {
  80                 icq = ioc_create_icq(ioc, q, GFP_ATOMIC);
  81                 if (!icq)
  82                         return;
  83         }
  84
  85         rq->elv.icq = icq;
  86         if (!blk_mq_sched_get_rq_priv(q, rq)) {
  87                 rq->rq_flags |= RQF_ELVPRIV;
  88                 get_io_context(icq->ioc);
  89                 return;
  90         }
  91
  92         rq->elv.icq = NULL;
  93 }
  94
  95 static void blk_mq_sched_assign_ioc(struct request_queue *q,
  96                                     struct request *rq, struct bio *bio)
  97 {
  98         struct io_context *ioc;
  99
 100         ioc = rq_ioc(bio);
 101         if (ioc)
 102                 __blk_mq_sched_assign_ioc(q, rq, ioc);
 103 }
 104
 105 struct request *blk_mq_sched_get_request(struct request_queue *q,
 106                                          struct bio *bio,
 107                                          unsigned int op,
 108                                          struct blk_mq_alloc_data *data)
 109 {
 110         struct elevator_queue *e = q->elevator;
 111         struct blk_mq_hw_ctx *hctx;
 112         struct blk_mq_ctx *ctx;
 113         struct request *rq;
 114         const bool is_flush = op & (REQ_PREFLUSH | REQ_FUA);
 115
 116         blk_queue_enter_live(q);
 117         ctx = blk_mq_get_ctx(q);
 118         hctx = blk_mq_map_queue(q, ctx->cpu);
 119
 120         blk_mq_set_alloc_data(data, q, data->flags, ctx, hctx);
 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 (!is_flush && 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                 if (rq)
 138                         data->hctx->tags->rqs[rq->tag] = rq;
 139         }
 140
 141         if (rq) {
 142                 if (!is_flush) {
 143                         rq->elv.icq = NULL;
 144                         if (e && e->type->icq_cache)
 145                                 blk_mq_sched_assign_ioc(q, rq, bio);
 146                 }
 147                 data->hctx->queued++;
 148                 return rq;
 149         }
 150
 151         blk_queue_exit(q);
 152         return NULL;
 153 }
 154
 155 void blk_mq_sched_put_request(struct request *rq)
 156 {
 157         struct request_queue *q = rq->q;
 158         struct elevator_queue *e = q->elevator;
 159
 160         if (rq->rq_flags & RQF_ELVPRIV) {
 161                 blk_mq_sched_put_rq_priv(rq->q, rq);
 162                 if (rq->elv.icq) {
 163                         put_io_context(rq->elv.icq->ioc);
 164                         rq->elv.icq = NULL;
 165                 }
 166         }
 167
 168         if ((rq->rq_flags & RQF_QUEUED) && e && e->type->ops.mq.put_request)
 169                 e->type->ops.mq.put_request(rq);
 170         else
 171                 blk_mq_finish_request(rq);
 172 }
 173
 174 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx)
 175 {
 176         struct elevator_queue *e = hctx->queue->elevator;
 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                 if (e && e->type->ops.mq.dispatch_requests)
 206                         e->type->ops.mq.dispatch_requests(hctx, &rq_list);
 207                 else
 208                         blk_mq_flush_busy_ctxs(hctx, &rq_list);
 209         } else
 210                 blk_mq_sched_mark_restart(hctx);
 211
 212         blk_mq_dispatch_rq_list(hctx, &rq_list);
 213 }
 214
 215 void blk_mq_sched_move_to_dispatch(struct blk_mq_hw_ctx *hctx,
 216                                    struct list_head *rq_list,
 217                                    struct request *(*get_rq)(struct blk_mq_hw_ctx *))
 218 {
 219         do {
 220                 struct request *rq;
 221
 222                 rq = get_rq(hctx);
 223                 if (!rq)
 224                         break;
 225
 226                 list_add_tail(&rq->queuelist, rq_list);
 227         } while (1);
 228 }
 229 EXPORT_SYMBOL_GPL(blk_mq_sched_move_to_dispatch);
 230
 231 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio)
 232 {
 233         struct request *rq;
 234         int ret;
 235
 236         ret = elv_merge(q, &rq, bio);
 237         if (ret == ELEVATOR_BACK_MERGE) {
 238                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 239                         return false;
 240                 if (bio_attempt_back_merge(q, rq, bio)) {
 241                         if (!attempt_back_merge(q, rq))
 242                                 elv_merged_request(q, rq, ret);
 243                         return true;
 244                 }
 245         } else if (ret == ELEVATOR_FRONT_MERGE) {
 246                 if (!blk_mq_sched_allow_merge(q, rq, bio))
 247                         return false;
 248                 if (bio_attempt_front_merge(q, rq, bio)) {
 249                         if (!attempt_front_merge(q, rq))
 250                                 elv_merged_request(q, rq, ret);
 251                         return true;
 252                 }
 253         }
 254
 255         return false;
 256 }
 257 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);
 258
 259 bool __blk_mq_sched_bio_merge(struct request_queue *q, struct bio *bio)
 260 {
 261         struct elevator_queue *e = q->elevator;
 262
 263         if (e->type->ops.mq.bio_merge) {
 264                 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 265                 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, ctx->cpu);
 266
 267                 blk_mq_put_ctx(ctx);
 268                 return e->type->ops.mq.bio_merge(hctx, bio);
 269         }
 270
 271         return false;
 272 }
 273
 274 bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq)
 275 {
 276         return rq_mergeable(rq) && elv_attempt_insert_merge(q, rq);
 277 }
 278 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 279
 280 void blk_mq_sched_request_inserted(struct request *rq)
 281 {
 282         trace_block_rq_insert(rq->q, rq);
 283 }
 284 EXPORT_SYMBOL_GPL(blk_mq_sched_request_inserted);
 285
 286 bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx, struct request *rq)
 287 {
 288         if (rq->tag == -1) {
 289                 rq->rq_flags |= RQF_SORTED;
 290                 return false;
 291         }
 292
 293         /*
 294          * If we already have a real request tag, send directly to
 295          * the dispatch list.
 296          */
 297         spin_lock(&hctx->lock);
 298         list_add(&rq->queuelist, &hctx->dispatch);
 299         spin_unlock(&hctx->lock);
 300         return true;
 301 }
 302 EXPORT_SYMBOL_GPL(blk_mq_sched_bypass_insert);
 303
 304 static void blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
 305 {
 306         if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
 307                 clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
 308                 if (blk_mq_hctx_has_pending(hctx))
 309                         blk_mq_run_hw_queue(hctx, true);
 310         }
 311 }
 312
 313 void blk_mq_sched_restart_queues(struct blk_mq_hw_ctx *hctx)
 314 {
 315         unsigned int i;
 316
 317         if (!(hctx->flags & BLK_MQ_F_TAG_SHARED))
 318                 blk_mq_sched_restart_hctx(hctx);
 319         else {
 320                 struct request_queue *q = hctx->queue;
 321
 322                 if (!test_bit(QUEUE_FLAG_RESTART, &q->queue_flags))
 323                         return;
 324
 325                 clear_bit(QUEUE_FLAG_RESTART, &q->queue_flags);
 326
 327                 queue_for_each_hw_ctx(q, hctx, i)
 328                         blk_mq_sched_restart_hctx(hctx);
 329         }
 330 }
 331
 332 static void blk_mq_sched_free_tags(struct blk_mq_tag_set *set,
 333                                    struct blk_mq_hw_ctx *hctx,
 334                                    unsigned int hctx_idx)
 335 {
 336         if (hctx->sched_tags) {
 337                 blk_mq_free_rqs(set, hctx->sched_tags, hctx_idx);
 338                 blk_mq_free_rq_map(hctx->sched_tags);
 339                 hctx->sched_tags = NULL;
 340         }
 341 }
 342
 343 int blk_mq_sched_setup(struct request_queue *q)
 344 {
 345         struct blk_mq_tag_set *set = q->tag_set;
 346         struct blk_mq_hw_ctx *hctx;
 347         int ret, i;
 348
 349         /*
 350          * Default to 256, since we don't split into sync/async like the
 351          * old code did. Additionally, this is a per-hw queue depth.
 352          */
 353         q->nr_requests = 2 * BLKDEV_MAX_RQ;
 354
 355         /*
 356          * We're switching to using an IO scheduler, so setup the hctx
 357          * scheduler tags and switch the request map from the regular
 358          * tags to scheduler tags. First allocate what we need, so we
 359          * can safely fail and fallback, if needed.
 360          */
 361         ret = 0;
 362         queue_for_each_hw_ctx(q, hctx, i) {
 363                 hctx->sched_tags = blk_mq_alloc_rq_map(set, i, q->nr_requests, 0);
 364                 if (!hctx->sched_tags) {
 365                         ret = -ENOMEM;
 366                         break;
 367                 }
 368                 ret = blk_mq_alloc_rqs(set, hctx->sched_tags, i, q->nr_requests);
 369                 if (ret)
 370                         break;
 371         }
 372
 373         /*
 374          * If we failed, free what we did allocate
 375          */
 376         if (ret) {
 377                 queue_for_each_hw_ctx(q, hctx, i) {
 378                         if (!hctx->sched_tags)
 379                                 continue;
 380                         blk_mq_sched_free_tags(set, hctx, i);
 381                 }
 382
 383                 return ret;
 384         }
 385
 386         return 0;
 387 }
 388
 389 void blk_mq_sched_teardown(struct request_queue *q)
 390 {
 391         struct blk_mq_tag_set *set = q->tag_set;
 392         struct blk_mq_hw_ctx *hctx;
 393         int i;
 394
 395         queue_for_each_hw_ctx(q, hctx, i)
 396                 blk_mq_sched_free_tags(set, hctx, i);
 397 }
 398
 399 int blk_mq_sched_init(struct request_queue *q)
 400 {
 401         int ret;
 402
 403 #if defined(CONFIG_DEFAULT_SQ_NONE)
 404         if (q->nr_hw_queues == 1)
 405                 return 0;
 406 #endif
 407 #if defined(CONFIG_DEFAULT_MQ_NONE)
 408         if (q->nr_hw_queues > 1)
 409                 return 0;
 410 #endif
 411
 412         mutex_lock(&q->sysfs_lock);
 413         ret = elevator_init(q, NULL);
 414         mutex_unlock(&q->sysfs_lock);
 415
 416         return ret;
 417 }