block/kyber-iosched.c

   1 /*
   2  * The Kyber I/O scheduler. Controls latency by throttling queue depths using
   3  * scalable techniques.
   4  *
   5  * Copyright (C) 2017 Facebook
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public
   9  * License v2 as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program.  If not, see <https://www.gnu.org/licenses/>.
  18  */
  19
  20 #include <linux/kernel.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/blk-mq.h>
  23 #include <linux/elevator.h>
  24 #include <linux/module.h>
  25 #include <linux/sbitmap.h>
  26
  27 #include "blk.h"
  28 #include "blk-mq.h"
  29 #include "blk-mq-debugfs.h"
  30 #include "blk-mq-sched.h"
  31 #include "blk-mq-tag.h"
  32 #include "blk-stat.h"
  33
  34 /* Scheduling domains. */
  35 enum {
  36         KYBER_READ,
  37         KYBER_SYNC_WRITE,
  38         KYBER_OTHER, /* Async writes, discard, etc. */
  39         KYBER_NUM_DOMAINS,
  40 };
  41
  42 enum {
  43         KYBER_MIN_DEPTH = 256,
  44
  45         /*
  46          * In order to prevent starvation of synchronous requests by a flood of
  47          * asynchronous requests, we reserve 25% of requests for synchronous
  48          * operations.
  49          */
  50         KYBER_ASYNC_PERCENT = 75,
  51 };
  52
  53 /*
  54  * Initial device-wide depths for each scheduling domain.
  55  *
  56  * Even for fast devices with lots of tags like NVMe, you can saturate
  57  * the device with only a fraction of the maximum possible queue depth.
  58  * So, we cap these to a reasonable value.
  59  */
  60 static const unsigned int kyber_depth[] = {
  61         [KYBER_READ] = 256,
  62         [KYBER_SYNC_WRITE] = 128,
  63         [KYBER_OTHER] = 64,
  64 };
  65
  66 /*
  67  * Scheduling domain batch sizes. We favor reads.
  68  */
  69 static const unsigned int kyber_batch_size[] = {
  70         [KYBER_READ] = 16,
  71         [KYBER_SYNC_WRITE] = 8,
  72         [KYBER_OTHER] = 8,
  73 };
  74
  75 struct kyber_queue_data {
  76         struct request_queue *q;
  77
  78         struct blk_stat_callback *cb;
  79
  80         /*
  81          * The device is divided into multiple scheduling domains based on the
  82          * request type. Each domain has a fixed number of in-flight requests of
  83          * that type device-wide, limited by these tokens.
  84          */
  85         struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
  86
  87         /*
  88          * Async request percentage, converted to per-word depth for
  89          * sbitmap_get_shallow().
  90          */
  91         unsigned int async_depth;
  92
  93         /* Target latencies in nanoseconds. */
  94         u64 read_lat_nsec, write_lat_nsec;
  95 };
  96
  97 struct kyber_hctx_data {
  98         spinlock_t lock;
  99         struct list_head rqs[KYBER_NUM_DOMAINS];
 100         unsigned int cur_domain;
 101         unsigned int batching;
 102         wait_queue_t domain_wait[KYBER_NUM_DOMAINS];
 103         atomic_t wait_index[KYBER_NUM_DOMAINS];
 104 };
 105
 106 static int rq_sched_domain(const struct request *rq)
 107 {
 108         unsigned int op = rq->cmd_flags;
 109
 110         if ((op & REQ_OP_MASK) == REQ_OP_READ)
 111                 return KYBER_READ;
 112         else if ((op & REQ_OP_MASK) == REQ_OP_WRITE && op_is_sync(op))
 113                 return KYBER_SYNC_WRITE;
 114         else
 115                 return KYBER_OTHER;
 116 }
 117
 118 enum {
 119         NONE = 0,
 120         GOOD = 1,
 121         GREAT = 2,
 122         BAD = -1,
 123         AWFUL = -2,
 124 };
 125
 126 #define IS_GOOD(status) ((status) > 0)
 127 #define IS_BAD(status) ((status) < 0)
 128
 129 static int kyber_lat_status(struct blk_stat_callback *cb,
 130                             unsigned int sched_domain, u64 target)
 131 {
 132         u64 latency;
 133
 134         if (!cb->stat[sched_domain].nr_samples)
 135                 return NONE;
 136
 137         latency = cb->stat[sched_domain].mean;
 138         if (latency >= 2 * target)
 139                 return AWFUL;
 140         else if (latency > target)
 141                 return BAD;
 142         else if (latency <= target / 2)
 143                 return GREAT;
 144         else /* (latency <= target) */
 145                 return GOOD;
 146 }
 147
 148 /*
 149  * Adjust the read or synchronous write depth given the status of reads and
 150  * writes. The goal is that the latencies of the two domains are fair (i.e., if
 151  * one is good, then the other is good).
 152  */
 153 static void kyber_adjust_rw_depth(struct kyber_queue_data *kqd,
 154                                   unsigned int sched_domain, int this_status,
 155                                   int other_status)
 156 {
 157         unsigned int orig_depth, depth;
 158
 159         /*
 160          * If this domain had no samples, or reads and writes are both good or
 161          * both bad, don't adjust the depth.
 162          */
 163         if (this_status == NONE ||
 164             (IS_GOOD(this_status) && IS_GOOD(other_status)) ||
 165             (IS_BAD(this_status) && IS_BAD(other_status)))
 166                 return;
 167
 168         orig_depth = depth = kqd->domain_tokens[sched_domain].sb.depth;
 169
 170         if (other_status == NONE) {
 171                 depth++;
 172         } else {
 173                 switch (this_status) {
 174                 case GOOD:
 175                         if (other_status == AWFUL)
 176                                 depth -= max(depth / 4, 1U);
 177                         else
 178                                 depth -= max(depth / 8, 1U);
 179                         break;
 180                 case GREAT:
 181                         if (other_status == AWFUL)
 182                                 depth /= 2;
 183                         else
 184                                 depth -= max(depth / 4, 1U);
 185                         break;
 186                 case BAD:
 187                         depth++;
 188                         break;
 189                 case AWFUL:
 190                         if (other_status == GREAT)
 191                                 depth += 2;
 192                         else
 193                                 depth++;
 194                         break;
 195                 }
 196         }
 197
 198         depth = clamp(depth, 1U, kyber_depth[sched_domain]);
 199         if (depth != orig_depth)
 200                 sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
 201 }
 202
 203 /*
 204  * Adjust the depth of other requests given the status of reads and synchronous
 205  * writes. As long as either domain is doing fine, we don't throttle, but if
 206  * both domains are doing badly, we throttle heavily.
 207  */
 208 static void kyber_adjust_other_depth(struct kyber_queue_data *kqd,
 209                                      int read_status, int write_status,
 210                                      bool have_samples)
 211 {
 212         unsigned int orig_depth, depth;
 213         int status;
 214
 215         orig_depth = depth = kqd->domain_tokens[KYBER_OTHER].sb.depth;
 216
 217         if (read_status == NONE && write_status == NONE) {
 218                 depth += 2;
 219         } else if (have_samples) {
 220                 if (read_status == NONE)
 221                         status = write_status;
 222                 else if (write_status == NONE)
 223                         status = read_status;
 224                 else
 225                         status = max(read_status, write_status);
 226                 switch (status) {
 227                 case GREAT:
 228                         depth += 2;
 229                         break;
 230                 case GOOD:
 231                         depth++;
 232                         break;
 233                 case BAD:
 234                         depth -= max(depth / 4, 1U);
 235                         break;
 236                 case AWFUL:
 237                         depth /= 2;
 238                         break;
 239                 }
 240         }
 241
 242         depth = clamp(depth, 1U, kyber_depth[KYBER_OTHER]);
 243         if (depth != orig_depth)
 244                 sbitmap_queue_resize(&kqd->domain_tokens[KYBER_OTHER], depth);
 245 }
 246
 247 /*
 248  * Apply heuristics for limiting queue depths based on gathered latency
 249  * statistics.
 250  */
 251 static void kyber_stat_timer_fn(struct blk_stat_callback *cb)
 252 {
 253         struct kyber_queue_data *kqd = cb->data;
 254         int read_status, write_status;
 255
 256         read_status = kyber_lat_status(cb, KYBER_READ, kqd->read_lat_nsec);
 257         write_status = kyber_lat_status(cb, KYBER_SYNC_WRITE, kqd->write_lat_nsec);
 258
 259         kyber_adjust_rw_depth(kqd, KYBER_READ, read_status, write_status);
 260         kyber_adjust_rw_depth(kqd, KYBER_SYNC_WRITE, write_status, read_status);
 261         kyber_adjust_other_depth(kqd, read_status, write_status,
 262                                  cb->stat[KYBER_OTHER].nr_samples != 0);
 263
 264         /*
 265          * Continue monitoring latencies if we aren't hitting the targets or
 266          * we're still throttling other requests.
 267          */
 268         if (!blk_stat_is_active(kqd->cb) &&
 269             ((IS_BAD(read_status) || IS_BAD(write_status) ||
 270               kqd->domain_tokens[KYBER_OTHER].sb.depth < kyber_depth[KYBER_OTHER])))
 271                 blk_stat_activate_msecs(kqd->cb, 100);
 272 }
 273
 274 static unsigned int kyber_sched_tags_shift(struct kyber_queue_data *kqd)
 275 {
 276         /*
 277          * All of the hardware queues have the same depth, so we can just grab
 278          * the shift of the first one.
 279          */
 280         return kqd->q->queue_hw_ctx[0]->sched_tags->bitmap_tags.sb.shift;
 281 }
 282
 283 static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
 284 {
 285         struct kyber_queue_data *kqd;
 286         unsigned int max_tokens;
 287         unsigned int shift;
 288         int ret = -ENOMEM;
 289         int i;
 290
 291         kqd = kmalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
 292         if (!kqd)
 293                 goto err;
 294         kqd->q = q;
 295
 296         kqd->cb = blk_stat_alloc_callback(kyber_stat_timer_fn, rq_sched_domain,
 297                                           KYBER_NUM_DOMAINS, kqd);
 298         if (!kqd->cb)
 299                 goto err_kqd;
 300
 301         /*
 302          * The maximum number of tokens for any scheduling domain is at least
 303          * the queue depth of a single hardware queue. If the hardware doesn't
 304          * have many tags, still provide a reasonable number.
 305          */
 306         max_tokens = max_t(unsigned int, q->tag_set->queue_depth,
 307                            KYBER_MIN_DEPTH);
 308         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 309                 WARN_ON(!kyber_depth[i]);
 310                 WARN_ON(!kyber_batch_size[i]);
 311                 ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
 312                                               max_tokens, -1, false, GFP_KERNEL,
 313                                               q->node);
 314                 if (ret) {
 315                         while (--i >= 0)
 316                                 sbitmap_queue_free(&kqd->domain_tokens[i]);
 317                         goto err_cb;
 318                 }
 319                 sbitmap_queue_resize(&kqd->domain_tokens[i], kyber_depth[i]);
 320         }
 321
 322         shift = kyber_sched_tags_shift(kqd);
 323         kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
 324
 325         kqd->read_lat_nsec = 2000000ULL;
 326         kqd->write_lat_nsec = 10000000ULL;
 327
 328         return kqd;
 329
 330 err_cb:
 331         blk_stat_free_callback(kqd->cb);
 332 err_kqd:
 333         kfree(kqd);
 334 err:
 335         return ERR_PTR(ret);
 336 }
 337
 338 static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
 339 {
 340         struct kyber_queue_data *kqd;
 341         struct elevator_queue *eq;
 342
 343         eq = elevator_alloc(q, e);
 344         if (!eq)
 345                 return -ENOMEM;
 346
 347         kqd = kyber_queue_data_alloc(q);
 348         if (IS_ERR(kqd)) {
 349                 kobject_put(&eq->kobj);
 350                 return PTR_ERR(kqd);
 351         }
 352
 353         eq->elevator_data = kqd;
 354         q->elevator = eq;
 355
 356         blk_stat_add_callback(q, kqd->cb);
 357
 358         return 0;
 359 }
 360
 361 static void kyber_exit_sched(struct elevator_queue *e)
 362 {
 363         struct kyber_queue_data *kqd = e->elevator_data;
 364         struct request_queue *q = kqd->q;
 365         int i;
 366
 367         blk_stat_remove_callback(q, kqd->cb);
 368
 369         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 370                 sbitmap_queue_free(&kqd->domain_tokens[i]);
 371         blk_stat_free_callback(kqd->cb);
 372         kfree(kqd);
 373 }
 374
 375 static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 376 {
 377         struct kyber_hctx_data *khd;
 378         int i;
 379
 380         khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
 381         if (!khd)
 382                 return -ENOMEM;
 383
 384         spin_lock_init(&khd->lock);
 385
 386         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 387                 INIT_LIST_HEAD(&khd->rqs[i]);
 388                 INIT_LIST_HEAD(&khd->domain_wait[i].task_list);
 389                 atomic_set(&khd->wait_index[i], 0);
 390         }
 391
 392         khd->cur_domain = 0;
 393         khd->batching = 0;
 394
 395         hctx->sched_data = khd;
 396
 397         return 0;
 398 }
 399
 400 static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 401 {
 402         kfree(hctx->sched_data);
 403 }
 404
 405 static int rq_get_domain_token(struct request *rq)
 406 {
 407         return (long)rq->elv.priv[0];
 408 }
 409
 410 static void rq_set_domain_token(struct request *rq, int token)
 411 {
 412         rq->elv.priv[0] = (void *)(long)token;
 413 }
 414
 415 static void rq_clear_domain_token(struct kyber_queue_data *kqd,
 416                                   struct request *rq)
 417 {
 418         unsigned int sched_domain;
 419         int nr;
 420
 421         nr = rq_get_domain_token(rq);
 422         if (nr != -1) {
 423                 sched_domain = rq_sched_domain(rq);
 424                 sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
 425                                     rq->mq_ctx->cpu);
 426         }
 427 }
 428
 429 static struct request *kyber_get_request(struct request_queue *q,
 430                                          unsigned int op,
 431                                          struct blk_mq_alloc_data *data)
 432 {
 433         struct kyber_queue_data *kqd = q->elevator->elevator_data;
 434         struct request *rq;
 435
 436         /*
 437          * We use the scheduler tags as per-hardware queue queueing tokens.
 438          * Async requests can be limited at this stage.
 439          */
 440         if (!op_is_sync(op))
 441                 data->shallow_depth = kqd->async_depth;
 442
 443         rq = __blk_mq_alloc_request(data, op);
 444         if (rq)
 445                 rq_set_domain_token(rq, -1);
 446         return rq;
 447 }
 448
 449 static void kyber_finish_request(struct request *rq)
 450 {
 451         struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
 452
 453         rq_clear_domain_token(kqd, rq);
 454 }
 455
 456 static void kyber_completed_request(struct request *rq)
 457 {
 458         struct request_queue *q = rq->q;
 459         struct kyber_queue_data *kqd = q->elevator->elevator_data;
 460         unsigned int sched_domain;
 461         u64 now, latency, target;
 462
 463         /*
 464          * Check if this request met our latency goal. If not, quickly gather
 465          * some statistics and start throttling.
 466          */
 467         sched_domain = rq_sched_domain(rq);
 468         switch (sched_domain) {
 469         case KYBER_READ:
 470                 target = kqd->read_lat_nsec;
 471                 break;
 472         case KYBER_SYNC_WRITE:
 473                 target = kqd->write_lat_nsec;
 474                 break;
 475         default:
 476                 return;
 477         }
 478
 479         /* If we are already monitoring latencies, don't check again. */
 480         if (blk_stat_is_active(kqd->cb))
 481                 return;
 482
 483         now = __blk_stat_time(ktime_to_ns(ktime_get()));
 484         if (now < blk_stat_time(&rq->issue_stat))
 485                 return;
 486
 487         latency = now - blk_stat_time(&rq->issue_stat);
 488
 489         if (latency > target)
 490                 blk_stat_activate_msecs(kqd->cb, 10);
 491 }
 492
 493 static void kyber_flush_busy_ctxs(struct kyber_hctx_data *khd,
 494                                   struct blk_mq_hw_ctx *hctx)
 495 {
 496         LIST_HEAD(rq_list);
 497         struct request *rq, *next;
 498
 499         blk_mq_flush_busy_ctxs(hctx, &rq_list);
 500         list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
 501                 unsigned int sched_domain;
 502
 503                 sched_domain = rq_sched_domain(rq);
 504                 list_move_tail(&rq->queuelist, &khd->rqs[sched_domain]);
 505         }
 506 }
 507
 508 static int kyber_domain_wake(wait_queue_t *wait, unsigned mode, int flags,
 509                              void *key)
 510 {
 511         struct blk_mq_hw_ctx *hctx = READ_ONCE(wait->private);
 512
 513         list_del_init(&wait->task_list);
 514         blk_mq_run_hw_queue(hctx, true);
 515         return 1;
 516 }
 517
 518 static int kyber_get_domain_token(struct kyber_queue_data *kqd,
 519                                   struct kyber_hctx_data *khd,
 520                                   struct blk_mq_hw_ctx *hctx)
 521 {
 522         unsigned int sched_domain = khd->cur_domain;
 523         struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
 524         wait_queue_t *wait = &khd->domain_wait[sched_domain];
 525         struct sbq_wait_state *ws;
 526         int nr;
 527
 528         nr = __sbitmap_queue_get(domain_tokens);
 529         if (nr >= 0)
 530                 return nr;
 531
 532         /*
 533          * If we failed to get a domain token, make sure the hardware queue is
 534          * run when one becomes available. Note that this is serialized on
 535          * khd->lock, but we still need to be careful about the waker.
 536          */
 537         if (list_empty_careful(&wait->task_list)) {
 538                 init_waitqueue_func_entry(wait, kyber_domain_wake);
 539                 wait->private = hctx;
 540                 ws = sbq_wait_ptr(domain_tokens,
 541                                   &khd->wait_index[sched_domain]);
 542                 add_wait_queue(&ws->wait, wait);
 543
 544                 /*
 545                  * Try again in case a token was freed before we got on the wait
 546                  * queue.
 547                  */
 548                 nr = __sbitmap_queue_get(domain_tokens);
 549         }
 550         return nr;
 551 }
 552
 553 static struct request *
 554 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
 555                           struct kyber_hctx_data *khd,
 556                           struct blk_mq_hw_ctx *hctx,
 557                           bool *flushed)
 558 {
 559         struct list_head *rqs;
 560         struct request *rq;
 561         int nr;
 562
 563         rqs = &khd->rqs[khd->cur_domain];
 564         rq = list_first_entry_or_null(rqs, struct request, queuelist);
 565
 566         /*
 567          * If there wasn't already a pending request and we haven't flushed the
 568          * software queues yet, flush the software queues and check again.
 569          */
 570         if (!rq && !*flushed) {
 571                 kyber_flush_busy_ctxs(khd, hctx);
 572                 *flushed = true;
 573                 rq = list_first_entry_or_null(rqs, struct request, queuelist);
 574         }
 575
 576         if (rq) {
 577                 nr = kyber_get_domain_token(kqd, khd, hctx);
 578                 if (nr >= 0) {
 579                         khd->batching++;
 580                         rq_set_domain_token(rq, nr);
 581                         list_del_init(&rq->queuelist);
 582                         return rq;
 583                 }
 584         }
 585
 586         /* There were either no pending requests or no tokens. */
 587         return NULL;
 588 }
 589
 590 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
 591 {
 592         struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
 593         struct kyber_hctx_data *khd = hctx->sched_data;
 594         bool flushed = false;
 595         struct request *rq;
 596         int i;
 597
 598         spin_lock(&khd->lock);
 599
 600         /*
 601          * First, if we are still entitled to batch, try to dispatch a request
 602          * from the batch.
 603          */
 604         if (khd->batching < kyber_batch_size[khd->cur_domain]) {
 605                 rq = kyber_dispatch_cur_domain(kqd, khd, hctx, &flushed);
 606                 if (rq)
 607                         goto out;
 608         }
 609
 610         /*
 611          * Either,
 612          * 1. We were no longer entitled to a batch.
 613          * 2. The domain we were batching didn't have any requests.
 614          * 3. The domain we were batching was out of tokens.
 615          *
 616          * Start another batch. Note that this wraps back around to the original
 617          * domain if no other domains have requests or tokens.
 618          */
 619         khd->batching = 0;
 620         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 621                 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
 622                         khd->cur_domain = 0;
 623                 else
 624                         khd->cur_domain++;
 625
 626                 rq = kyber_dispatch_cur_domain(kqd, khd, hctx, &flushed);
 627                 if (rq)
 628                         goto out;
 629         }
 630
 631         rq = NULL;
 632 out:
 633         spin_unlock(&khd->lock);
 634         return rq;
 635 }
 636
 637 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
 638 {
 639         struct kyber_hctx_data *khd = hctx->sched_data;
 640         int i;
 641
 642         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 643                 if (!list_empty_careful(&khd->rqs[i]))
 644                         return true;
 645         }
 646         return false;
 647 }
 648
 649 #define KYBER_LAT_SHOW_STORE(op)                                        \
 650 static ssize_t kyber_##op##_lat_show(struct elevator_queue *e,          \
 651                                      char *page)                        \
 652 {                                                                       \
 653         struct kyber_queue_data *kqd = e->elevator_data;                \
 654                                                                         \
 655         return sprintf(page, "%llu\n", kqd->op##_lat_nsec);             \
 656 }                                                                       \
 657                                                                         \
 658 static ssize_t kyber_##op##_lat_store(struct elevator_queue *e,         \
 659                                       const char *page, size_t count)   \
 660 {                                                                       \
 661         struct kyber_queue_data *kqd = e->elevator_data;                \
 662         unsigned long long nsec;                                        \
 663         int ret;                                                        \
 664                                                                         \
 665         ret = kstrtoull(page, 10, &nsec);                               \
 666         if (ret)                                                        \
 667                 return ret;                                             \
 668                                                                         \
 669         kqd->op##_lat_nsec = nsec;                                      \
 670                                                                         \
 671         return count;                                                   \
 672 }
 673 KYBER_LAT_SHOW_STORE(read);
 674 KYBER_LAT_SHOW_STORE(write);
 675 #undef KYBER_LAT_SHOW_STORE
 676
 677 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
 678 static struct elv_fs_entry kyber_sched_attrs[] = {
 679         KYBER_LAT_ATTR(read),
 680         KYBER_LAT_ATTR(write),
 681         __ATTR_NULL
 682 };
 683 #undef KYBER_LAT_ATTR
 684
 685 #ifdef CONFIG_BLK_DEBUG_FS
 686 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name)                        \
 687 static int kyber_##name##_tokens_show(void *data, struct seq_file *m)   \
 688 {                                                                       \
 689         struct request_queue *q = data;                                 \
 690         struct kyber_queue_data *kqd = q->elevator->elevator_data;      \
 691                                                                         \
 692         sbitmap_queue_show(&kqd->domain_tokens[domain], m);             \
 693         return 0;                                                       \
 694 }                                                                       \
 695                                                                         \
 696 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos)  \
 697         __acquires(&khd->lock)                                          \
 698 {                                                                       \
 699         struct blk_mq_hw_ctx *hctx = m->private;                        \
 700         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 701                                                                         \
 702         spin_lock(&khd->lock);                                          \
 703         return seq_list_start(&khd->rqs[domain], *pos);                 \
 704 }                                                                       \
 705                                                                         \
 706 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v,       \
 707                                      loff_t *pos)                       \
 708 {                                                                       \
 709         struct blk_mq_hw_ctx *hctx = m->private;                        \
 710         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 711                                                                         \
 712         return seq_list_next(v, &khd->rqs[domain], pos);                \
 713 }                                                                       \
 714                                                                         \
 715 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v)        \
 716         __releases(&khd->lock)                                          \
 717 {                                                                       \
 718         struct blk_mq_hw_ctx *hctx = m->private;                        \
 719         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 720                                                                         \
 721         spin_unlock(&khd->lock);                                        \
 722 }                                                                       \
 723                                                                         \
 724 static const struct seq_operations kyber_##name##_rqs_seq_ops = {       \
 725         .start  = kyber_##name##_rqs_start,                             \
 726         .next   = kyber_##name##_rqs_next,                              \
 727         .stop   = kyber_##name##_rqs_stop,                              \
 728         .show   = blk_mq_debugfs_rq_show,                               \
 729 };                                                                      \
 730                                                                         \
 731 static int kyber_##name##_waiting_show(void *data, struct seq_file *m)  \
 732 {                                                                       \
 733         struct blk_mq_hw_ctx *hctx = data;                              \
 734         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 735         wait_queue_t *wait = &khd->domain_wait[domain];                 \
 736                                                                         \
 737         seq_printf(m, "%d\n", !list_empty_careful(&wait->task_list));   \
 738         return 0;                                                       \
 739 }
 740 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
 741 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_SYNC_WRITE, sync_write)
 742 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
 743 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
 744
 745 static int kyber_async_depth_show(void *data, struct seq_file *m)
 746 {
 747         struct request_queue *q = data;
 748         struct kyber_queue_data *kqd = q->elevator->elevator_data;
 749
 750         seq_printf(m, "%u\n", kqd->async_depth);
 751         return 0;
 752 }
 753
 754 static int kyber_cur_domain_show(void *data, struct seq_file *m)
 755 {
 756         struct blk_mq_hw_ctx *hctx = data;
 757         struct kyber_hctx_data *khd = hctx->sched_data;
 758
 759         switch (khd->cur_domain) {
 760         case KYBER_READ:
 761                 seq_puts(m, "READ\n");
 762                 break;
 763         case KYBER_SYNC_WRITE:
 764                 seq_puts(m, "SYNC_WRITE\n");
 765                 break;
 766         case KYBER_OTHER:
 767                 seq_puts(m, "OTHER\n");
 768                 break;
 769         default:
 770                 seq_printf(m, "%u\n", khd->cur_domain);
 771                 break;
 772         }
 773         return 0;
 774 }
 775
 776 static int kyber_batching_show(void *data, struct seq_file *m)
 777 {
 778         struct blk_mq_hw_ctx *hctx = data;
 779         struct kyber_hctx_data *khd = hctx->sched_data;
 780
 781         seq_printf(m, "%u\n", khd->batching);
 782         return 0;
 783 }
 784
 785 #define KYBER_QUEUE_DOMAIN_ATTRS(name)  \
 786         {#name "_tokens", 0400, kyber_##name##_tokens_show}
 787 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
 788         KYBER_QUEUE_DOMAIN_ATTRS(read),
 789         KYBER_QUEUE_DOMAIN_ATTRS(sync_write),
 790         KYBER_QUEUE_DOMAIN_ATTRS(other),
 791         {"async_depth", 0400, kyber_async_depth_show},
 792         {},
 793 };
 794 #undef KYBER_QUEUE_DOMAIN_ATTRS
 795
 796 #define KYBER_HCTX_DOMAIN_ATTRS(name)                                   \
 797         {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops},   \
 798         {#name "_waiting", 0400, kyber_##name##_waiting_show}
 799 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
 800         KYBER_HCTX_DOMAIN_ATTRS(read),
 801         KYBER_HCTX_DOMAIN_ATTRS(sync_write),
 802         KYBER_HCTX_DOMAIN_ATTRS(other),
 803         {"cur_domain", 0400, kyber_cur_domain_show},
 804         {"batching", 0400, kyber_batching_show},
 805         {},
 806 };
 807 #undef KYBER_HCTX_DOMAIN_ATTRS
 808 #endif
 809
 810 static struct elevator_type kyber_sched = {
 811         .ops.mq = {
 812                 .init_sched = kyber_init_sched,
 813                 .exit_sched = kyber_exit_sched,
 814                 .init_hctx = kyber_init_hctx,
 815                 .exit_hctx = kyber_exit_hctx,
 816                 .get_request = kyber_get_request,
 817                 .finish_request = kyber_finish_request,
 818                 .completed_request = kyber_completed_request,
 819                 .dispatch_request = kyber_dispatch_request,
 820                 .has_work = kyber_has_work,
 821         },
 822         .uses_mq = true,
 823 #ifdef CONFIG_BLK_DEBUG_FS
 824         .queue_debugfs_attrs = kyber_queue_debugfs_attrs,
 825         .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs,
 826 #endif
 827         .elevator_attrs = kyber_sched_attrs,
 828         .elevator_name = "kyber",
 829         .elevator_owner = THIS_MODULE,
 830 };
 831
 832 static int __init kyber_init(void)
 833 {
 834         return elv_register(&kyber_sched);
 835 }
 836
 837 static void __exit kyber_exit(void)
 838 {
 839         elv_unregister(&kyber_sched);
 840 }
 841
 842 module_init(kyber_init);
 843 module_exit(kyber_exit);
 844
 845 MODULE_AUTHOR("Omar Sandoval");
 846 MODULE_LICENSE("GPL");
 847 MODULE_DESCRIPTION("Kyber I/O scheduler");