block/kyber-iosched.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * The Kyber I/O scheduler. Controls latency by throttling queue depths using
   4  * scalable techniques.
   5  *
   6  * Copyright (C) 2017 Facebook
   7  */
   8
   9 #include <linux/kernel.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/blk-mq.h>
  12 #include <linux/elevator.h>
  13 #include <linux/module.h>
  14 #include <linux/sbitmap.h>
  15
  16 #include <trace/events/block.h>
  17
  18 #include "blk.h"
  19 #include "blk-mq.h"
  20 #include "blk-mq-debugfs.h"
  21 #include "blk-mq-sched.h"
  22 #include "blk-mq-tag.h"
  23
  24 #define CREATE_TRACE_POINTS
  25 #include <trace/events/kyber.h>
  26
  27 /*
  28  * Scheduling domains: the device is divided into multiple domains based on the
  29  * request type.
  30  */
  31 enum {
  32         KYBER_READ,
  33         KYBER_WRITE,
  34         KYBER_DISCARD,
  35         KYBER_OTHER,
  36         KYBER_NUM_DOMAINS,
  37 };
  38
  39 static const char *kyber_domain_names[] = {
  40         [KYBER_READ] = "READ",
  41         [KYBER_WRITE] = "WRITE",
  42         [KYBER_DISCARD] = "DISCARD",
  43         [KYBER_OTHER] = "OTHER",
  44 };
  45
  46 enum {
  47         /*
  48          * In order to prevent starvation of synchronous requests by a flood of
  49          * asynchronous requests, we reserve 25% of requests for synchronous
  50          * operations.
  51          */
  52         KYBER_ASYNC_PERCENT = 75,
  53 };
  54
  55 /*
  56  * Maximum device-wide depth for each scheduling domain.
  57  *
  58  * Even for fast devices with lots of tags like NVMe, you can saturate the
  59  * device with only a fraction of the maximum possible queue depth. So, we cap
  60  * these to a reasonable value.
  61  */
  62 static const unsigned int kyber_depth[] = {
  63         [KYBER_READ] = 256,
  64         [KYBER_WRITE] = 128,
  65         [KYBER_DISCARD] = 64,
  66         [KYBER_OTHER] = 16,
  67 };
  68
  69 /*
  70  * Default latency targets for each scheduling domain.
  71  */
  72 static const u64 kyber_latency_targets[] = {
  73         [KYBER_READ] = 2ULL * NSEC_PER_MSEC,
  74         [KYBER_WRITE] = 10ULL * NSEC_PER_MSEC,
  75         [KYBER_DISCARD] = 5ULL * NSEC_PER_SEC,
  76 };
  77
  78 /*
  79  * Batch size (number of requests we'll dispatch in a row) for each scheduling
  80  * domain.
  81  */
  82 static const unsigned int kyber_batch_size[] = {
  83         [KYBER_READ] = 16,
  84         [KYBER_WRITE] = 8,
  85         [KYBER_DISCARD] = 1,
  86         [KYBER_OTHER] = 1,
  87 };
  88
  89 /*
  90  * Requests latencies are recorded in a histogram with buckets defined relative
  91  * to the target latency:
  92  *
  93  * <= 1/4 * target latency
  94  * <= 1/2 * target latency
  95  * <= 3/4 * target latency
  96  * <= target latency
  97  * <= 1 1/4 * target latency
  98  * <= 1 1/2 * target latency
  99  * <= 1 3/4 * target latency
 100  * > 1 3/4 * target latency
 101  */
 102 enum {
 103         /*
 104          * The width of the latency histogram buckets is
 105          * 1 / (1 << KYBER_LATENCY_SHIFT) * target latency.
 106          */
 107         KYBER_LATENCY_SHIFT = 2,
 108         /*
 109          * The first (1 << KYBER_LATENCY_SHIFT) buckets are <= target latency,
 110          * thus, "good".
 111          */
 112         KYBER_GOOD_BUCKETS = 1 << KYBER_LATENCY_SHIFT,
 113         /* There are also (1 << KYBER_LATENCY_SHIFT) "bad" buckets. */
 114         KYBER_LATENCY_BUCKETS = 2 << KYBER_LATENCY_SHIFT,
 115 };
 116
 117 /*
 118  * We measure both the total latency and the I/O latency (i.e., latency after
 119  * submitting to the device).
 120  */
 121 enum {
 122         KYBER_TOTAL_LATENCY,
 123         KYBER_IO_LATENCY,
 124 };
 125
 126 static const char *kyber_latency_type_names[] = {
 127         [KYBER_TOTAL_LATENCY] = "total",
 128         [KYBER_IO_LATENCY] = "I/O",
 129 };
 130
 131 /*
 132  * Per-cpu latency histograms: total latency and I/O latency for each scheduling
 133  * domain except for KYBER_OTHER.
 134  */
 135 struct kyber_cpu_latency {
 136         atomic_t buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
 137 };
 138
 139 /*
 140  * There is a same mapping between ctx & hctx and kcq & khd,
 141  * we use request->mq_ctx->index_hw to index the kcq in khd.
 142  */
 143 struct kyber_ctx_queue {
 144         /*
 145          * Used to ensure operations on rq_list and kcq_map to be an atmoic one.
 146          * Also protect the rqs on rq_list when merge.
 147          */
 148         spinlock_t lock;
 149         struct list_head rq_list[KYBER_NUM_DOMAINS];
 150 } ____cacheline_aligned_in_smp;
 151
 152 struct kyber_queue_data {
 153         struct request_queue *q;
 154
 155         /*
 156          * Each scheduling domain has a limited number of in-flight requests
 157          * device-wide, limited by these tokens.
 158          */
 159         struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
 160
 161         /*
 162          * Async request percentage, converted to per-word depth for
 163          * sbitmap_get_shallow().
 164          */
 165         unsigned int async_depth;
 166
 167         struct kyber_cpu_latency __percpu *cpu_latency;
 168
 169         /* Timer for stats aggregation and adjusting domain tokens. */
 170         struct timer_list timer;
 171
 172         unsigned int latency_buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
 173
 174         unsigned long latency_timeout[KYBER_OTHER];
 175
 176         int domain_p99[KYBER_OTHER];
 177
 178         /* Target latencies in nanoseconds. */
 179         u64 latency_targets[KYBER_OTHER];
 180 };
 181
 182 struct kyber_hctx_data {
 183         spinlock_t lock;
 184         struct list_head rqs[KYBER_NUM_DOMAINS];
 185         unsigned int cur_domain;
 186         unsigned int batching;
 187         struct kyber_ctx_queue *kcqs;
 188         struct sbitmap kcq_map[KYBER_NUM_DOMAINS];
 189         struct sbq_wait domain_wait[KYBER_NUM_DOMAINS];
 190         struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
 191         atomic_t wait_index[KYBER_NUM_DOMAINS];
 192 };
 193
 194 static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
 195                              void *key);
 196
 197 static unsigned int kyber_sched_domain(unsigned int op)
 198 {
 199         switch (op & REQ_OP_MASK) {
 200         case REQ_OP_READ:
 201                 return KYBER_READ;
 202         case REQ_OP_WRITE:
 203                 return KYBER_WRITE;
 204         case REQ_OP_DISCARD:
 205                 return KYBER_DISCARD;
 206         default:
 207                 return KYBER_OTHER;
 208         }
 209 }
 210
 211 static void flush_latency_buckets(struct kyber_queue_data *kqd,
 212                                   struct kyber_cpu_latency *cpu_latency,
 213                                   unsigned int sched_domain, unsigned int type)
 214 {
 215         unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
 216         atomic_t *cpu_buckets = cpu_latency->buckets[sched_domain][type];
 217         unsigned int bucket;
 218
 219         for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
 220                 buckets[bucket] += atomic_xchg(&cpu_buckets[bucket], 0);
 221 }
 222
 223 /*
 224  * Calculate the histogram bucket with the given percentile rank, or -1 if there
 225  * aren't enough samples yet.
 226  */
 227 static int calculate_percentile(struct kyber_queue_data *kqd,
 228                                 unsigned int sched_domain, unsigned int type,
 229                                 unsigned int percentile)
 230 {
 231         unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
 232         unsigned int bucket, samples = 0, percentile_samples;
 233
 234         for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
 235                 samples += buckets[bucket];
 236
 237         if (!samples)
 238                 return -1;
 239
 240         /*
 241          * We do the calculation once we have 500 samples or one second passes
 242          * since the first sample was recorded, whichever comes first.
 243          */
 244         if (!kqd->latency_timeout[sched_domain])
 245                 kqd->latency_timeout[sched_domain] = max(jiffies + HZ, 1UL);
 246         if (samples < 500 &&
 247             time_is_after_jiffies(kqd->latency_timeout[sched_domain])) {
 248                 return -1;
 249         }
 250         kqd->latency_timeout[sched_domain] = 0;
 251
 252         percentile_samples = DIV_ROUND_UP(samples * percentile, 100);
 253         for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS - 1; bucket++) {
 254                 if (buckets[bucket] >= percentile_samples)
 255                         break;
 256                 percentile_samples -= buckets[bucket];
 257         }
 258         memset(buckets, 0, sizeof(kqd->latency_buckets[sched_domain][type]));
 259
 260         trace_kyber_latency(kqd->q, kyber_domain_names[sched_domain],
 261                             kyber_latency_type_names[type], percentile,
 262                             bucket + 1, 1 << KYBER_LATENCY_SHIFT, samples);
 263
 264         return bucket;
 265 }
 266
 267 static void kyber_resize_domain(struct kyber_queue_data *kqd,
 268                                 unsigned int sched_domain, unsigned int depth)
 269 {
 270         depth = clamp(depth, 1U, kyber_depth[sched_domain]);
 271         if (depth != kqd->domain_tokens[sched_domain].sb.depth) {
 272                 sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
 273                 trace_kyber_adjust(kqd->q, kyber_domain_names[sched_domain],
 274                                    depth);
 275         }
 276 }
 277
 278 static void kyber_timer_fn(struct timer_list *t)
 279 {
 280         struct kyber_queue_data *kqd = from_timer(kqd, t, timer);
 281         unsigned int sched_domain;
 282         int cpu;
 283         bool bad = false;
 284
 285         /* Sum all of the per-cpu latency histograms. */
 286         for_each_online_cpu(cpu) {
 287                 struct kyber_cpu_latency *cpu_latency;
 288
 289                 cpu_latency = per_cpu_ptr(kqd->cpu_latency, cpu);
 290                 for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
 291                         flush_latency_buckets(kqd, cpu_latency, sched_domain,
 292                                               KYBER_TOTAL_LATENCY);
 293                         flush_latency_buckets(kqd, cpu_latency, sched_domain,
 294                                               KYBER_IO_LATENCY);
 295                 }
 296         }
 297
 298         /*
 299          * Check if any domains have a high I/O latency, which might indicate
 300          * congestion in the device. Note that we use the p90; we don't want to
 301          * be too sensitive to outliers here.
 302          */
 303         for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
 304                 int p90;
 305
 306                 p90 = calculate_percentile(kqd, sched_domain, KYBER_IO_LATENCY,
 307                                            90);
 308                 if (p90 >= KYBER_GOOD_BUCKETS)
 309                         bad = true;
 310         }
 311
 312         /*
 313          * Adjust the scheduling domain depths. If we determined that there was
 314          * congestion, we throttle all domains with good latencies. Either way,
 315          * we ease up on throttling domains with bad latencies.
 316          */
 317         for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
 318                 unsigned int orig_depth, depth;
 319                 int p99;
 320
 321                 p99 = calculate_percentile(kqd, sched_domain,
 322                                            KYBER_TOTAL_LATENCY, 99);
 323                 /*
 324                  * This is kind of subtle: different domains will not
 325                  * necessarily have enough samples to calculate the latency
 326                  * percentiles during the same window, so we have to remember
 327                  * the p99 for the next time we observe congestion; once we do,
 328                  * we don't want to throttle again until we get more data, so we
 329                  * reset it to -1.
 330                  */
 331                 if (bad) {
 332                         if (p99 < 0)
 333                                 p99 = kqd->domain_p99[sched_domain];
 334                         kqd->domain_p99[sched_domain] = -1;
 335                 } else if (p99 >= 0) {
 336                         kqd->domain_p99[sched_domain] = p99;
 337                 }
 338                 if (p99 < 0)
 339                         continue;
 340
 341                 /*
 342                  * If this domain has bad latency, throttle less. Otherwise,
 343                  * throttle more iff we determined that there is congestion.
 344                  *
 345                  * The new depth is scaled linearly with the p99 latency vs the
 346                  * latency target. E.g., if the p99 is 3/4 of the target, then
 347                  * we throttle down to 3/4 of the current depth, and if the p99
 348                  * is 2x the target, then we double the depth.
 349                  */
 350                 if (bad || p99 >= KYBER_GOOD_BUCKETS) {
 351                         orig_depth = kqd->domain_tokens[sched_domain].sb.depth;
 352                         depth = (orig_depth * (p99 + 1)) >> KYBER_LATENCY_SHIFT;
 353                         kyber_resize_domain(kqd, sched_domain, depth);
 354                 }
 355         }
 356 }
 357
 358 static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
 359 {
 360         struct kyber_queue_data *kqd;
 361         int ret = -ENOMEM;
 362         int i;
 363
 364         kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
 365         if (!kqd)
 366                 goto err;
 367
 368         kqd->q = q;
 369
 370         kqd->cpu_latency = alloc_percpu_gfp(struct kyber_cpu_latency,
 371                                             GFP_KERNEL | __GFP_ZERO);
 372         if (!kqd->cpu_latency)
 373                 goto err_kqd;
 374
 375         timer_setup(&kqd->timer, kyber_timer_fn, 0);
 376
 377         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 378                 WARN_ON(!kyber_depth[i]);
 379                 WARN_ON(!kyber_batch_size[i]);
 380                 ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
 381                                               kyber_depth[i], -1, false,
 382                                               GFP_KERNEL, q->node);
 383                 if (ret) {
 384                         while (--i >= 0)
 385                                 sbitmap_queue_free(&kqd->domain_tokens[i]);
 386                         goto err_buckets;
 387                 }
 388         }
 389
 390         for (i = 0; i < KYBER_OTHER; i++) {
 391                 kqd->domain_p99[i] = -1;
 392                 kqd->latency_targets[i] = kyber_latency_targets[i];
 393         }
 394
 395         return kqd;
 396
 397 err_buckets:
 398         free_percpu(kqd->cpu_latency);
 399 err_kqd:
 400         kfree(kqd);
 401 err:
 402         return ERR_PTR(ret);
 403 }
 404
 405 static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
 406 {
 407         struct kyber_queue_data *kqd;
 408         struct elevator_queue *eq;
 409
 410         eq = elevator_alloc(q, e);
 411         if (!eq)
 412                 return -ENOMEM;
 413
 414         kqd = kyber_queue_data_alloc(q);
 415         if (IS_ERR(kqd)) {
 416                 kobject_put(&eq->kobj);
 417                 return PTR_ERR(kqd);
 418         }
 419
 420         blk_stat_enable_accounting(q);
 421
 422         eq->elevator_data = kqd;
 423         q->elevator = eq;
 424
 425         return 0;
 426 }
 427
 428 static void kyber_exit_sched(struct elevator_queue *e)
 429 {
 430         struct kyber_queue_data *kqd = e->elevator_data;
 431         int i;
 432
 433         del_timer_sync(&kqd->timer);
 434
 435         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 436                 sbitmap_queue_free(&kqd->domain_tokens[i]);
 437         free_percpu(kqd->cpu_latency);
 438         kfree(kqd);
 439 }
 440
 441 static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq)
 442 {
 443         unsigned int i;
 444
 445         spin_lock_init(&kcq->lock);
 446         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 447                 INIT_LIST_HEAD(&kcq->rq_list[i]);
 448 }
 449
 450 static void kyber_depth_updated(struct blk_mq_hw_ctx *hctx)
 451 {
 452         struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
 453         struct blk_mq_tags *tags = hctx->sched_tags;
 454         unsigned int shift = tags->bitmap_tags->sb.shift;
 455
 456         kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
 457
 458         sbitmap_queue_min_shallow_depth(tags->bitmap_tags, kqd->async_depth);
 459 }
 460
 461 static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 462 {
 463         struct kyber_hctx_data *khd;
 464         int i;
 465
 466         khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
 467         if (!khd)
 468                 return -ENOMEM;
 469
 470         khd->kcqs = kmalloc_array_node(hctx->nr_ctx,
 471                                        sizeof(struct kyber_ctx_queue),
 472                                        GFP_KERNEL, hctx->numa_node);
 473         if (!khd->kcqs)
 474                 goto err_khd;
 475
 476         for (i = 0; i < hctx->nr_ctx; i++)
 477                 kyber_ctx_queue_init(&khd->kcqs[i]);
 478
 479         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 480                 if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx,
 481                                       ilog2(8), GFP_KERNEL, hctx->numa_node,
 482                                       false, false)) {
 483                         while (--i >= 0)
 484                                 sbitmap_free(&khd->kcq_map[i]);
 485                         goto err_kcqs;
 486                 }
 487         }
 488
 489         spin_lock_init(&khd->lock);
 490
 491         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 492                 INIT_LIST_HEAD(&khd->rqs[i]);
 493                 khd->domain_wait[i].sbq = NULL;
 494                 init_waitqueue_func_entry(&khd->domain_wait[i].wait,
 495                                           kyber_domain_wake);
 496                 khd->domain_wait[i].wait.private = hctx;
 497                 INIT_LIST_HEAD(&khd->domain_wait[i].wait.entry);
 498                 atomic_set(&khd->wait_index[i], 0);
 499         }
 500
 501         khd->cur_domain = 0;
 502         khd->batching = 0;
 503
 504         hctx->sched_data = khd;
 505         kyber_depth_updated(hctx);
 506
 507         return 0;
 508
 509 err_kcqs:
 510         kfree(khd->kcqs);
 511 err_khd:
 512         kfree(khd);
 513         return -ENOMEM;
 514 }
 515
 516 static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 517 {
 518         struct kyber_hctx_data *khd = hctx->sched_data;
 519         int i;
 520
 521         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 522                 sbitmap_free(&khd->kcq_map[i]);
 523         kfree(khd->kcqs);
 524         kfree(hctx->sched_data);
 525 }
 526
 527 static int rq_get_domain_token(struct request *rq)
 528 {
 529         return (long)rq->elv.priv[0];
 530 }
 531
 532 static void rq_set_domain_token(struct request *rq, int token)
 533 {
 534         rq->elv.priv[0] = (void *)(long)token;
 535 }
 536
 537 static void rq_clear_domain_token(struct kyber_queue_data *kqd,
 538                                   struct request *rq)
 539 {
 540         unsigned int sched_domain;
 541         int nr;
 542
 543         nr = rq_get_domain_token(rq);
 544         if (nr != -1) {
 545                 sched_domain = kyber_sched_domain(rq->cmd_flags);
 546                 sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
 547                                     rq->mq_ctx->cpu);
 548         }
 549 }
 550
 551 static void kyber_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
 552 {
 553         /*
 554          * We use the scheduler tags as per-hardware queue queueing tokens.
 555          * Async requests can be limited at this stage.
 556          */
 557         if (!op_is_sync(op)) {
 558                 struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
 559
 560                 data->shallow_depth = kqd->async_depth;
 561         }
 562 }
 563
 564 static bool kyber_bio_merge(struct request_queue *q, struct bio *bio,
 565                 unsigned int nr_segs)
 566 {
 567         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 568         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
 569         struct kyber_hctx_data *khd = hctx->sched_data;
 570         struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]];
 571         unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);
 572         struct list_head *rq_list = &kcq->rq_list[sched_domain];
 573         bool merged;
 574
 575         spin_lock(&kcq->lock);
 576         merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs);
 577         spin_unlock(&kcq->lock);
 578
 579         return merged;
 580 }
 581
 582 static void kyber_prepare_request(struct request *rq)
 583 {
 584         rq_set_domain_token(rq, -1);
 585 }
 586
 587 static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx,
 588                                   struct list_head *rq_list, bool at_head)
 589 {
 590         struct kyber_hctx_data *khd = hctx->sched_data;
 591         struct request *rq, *next;
 592
 593         list_for_each_entry_safe(rq, next, rq_list, queuelist) {
 594                 unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags);
 595                 struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]];
 596                 struct list_head *head = &kcq->rq_list[sched_domain];
 597
 598                 spin_lock(&kcq->lock);
 599                 trace_block_rq_insert(rq);
 600                 if (at_head)
 601                         list_move(&rq->queuelist, head);
 602                 else
 603                         list_move_tail(&rq->queuelist, head);
 604                 sbitmap_set_bit(&khd->kcq_map[sched_domain],
 605                                 rq->mq_ctx->index_hw[hctx->type]);
 606                 spin_unlock(&kcq->lock);
 607         }
 608 }
 609
 610 static void kyber_finish_request(struct request *rq)
 611 {
 612         struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
 613
 614         rq_clear_domain_token(kqd, rq);
 615 }
 616
 617 static void add_latency_sample(struct kyber_cpu_latency *cpu_latency,
 618                                unsigned int sched_domain, unsigned int type,
 619                                u64 target, u64 latency)
 620 {
 621         unsigned int bucket;
 622         u64 divisor;
 623
 624         if (latency > 0) {
 625                 divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1);
 626                 bucket = min_t(unsigned int, div64_u64(latency - 1, divisor),
 627                                KYBER_LATENCY_BUCKETS - 1);
 628         } else {
 629                 bucket = 0;
 630         }
 631
 632         atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]);
 633 }
 634
 635 static void kyber_completed_request(struct request *rq, u64 now)
 636 {
 637         struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
 638         struct kyber_cpu_latency *cpu_latency;
 639         unsigned int sched_domain;
 640         u64 target;
 641
 642         sched_domain = kyber_sched_domain(rq->cmd_flags);
 643         if (sched_domain == KYBER_OTHER)
 644                 return;
 645
 646         cpu_latency = get_cpu_ptr(kqd->cpu_latency);
 647         target = kqd->latency_targets[sched_domain];
 648         add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY,
 649                            target, now - rq->start_time_ns);
 650         add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target,
 651                            now - rq->io_start_time_ns);
 652         put_cpu_ptr(kqd->cpu_latency);
 653
 654         timer_reduce(&kqd->timer, jiffies + HZ / 10);
 655 }
 656
 657 struct flush_kcq_data {
 658         struct kyber_hctx_data *khd;
 659         unsigned int sched_domain;
 660         struct list_head *list;
 661 };
 662
 663 static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data)
 664 {
 665         struct flush_kcq_data *flush_data = data;
 666         struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr];
 667
 668         spin_lock(&kcq->lock);
 669         list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain],
 670                               flush_data->list);
 671         sbitmap_clear_bit(sb, bitnr);
 672         spin_unlock(&kcq->lock);
 673
 674         return true;
 675 }
 676
 677 static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd,
 678                                   unsigned int sched_domain,
 679                                   struct list_head *list)
 680 {
 681         struct flush_kcq_data data = {
 682                 .khd = khd,
 683                 .sched_domain = sched_domain,
 684                 .list = list,
 685         };
 686
 687         sbitmap_for_each_set(&khd->kcq_map[sched_domain],
 688                              flush_busy_kcq, &data);
 689 }
 690
 691 static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags,
 692                              void *key)
 693 {
 694         struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private);
 695         struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait);
 696
 697         sbitmap_del_wait_queue(wait);
 698         blk_mq_run_hw_queue(hctx, true);
 699         return 1;
 700 }
 701
 702 static int kyber_get_domain_token(struct kyber_queue_data *kqd,
 703                                   struct kyber_hctx_data *khd,
 704                                   struct blk_mq_hw_ctx *hctx)
 705 {
 706         unsigned int sched_domain = khd->cur_domain;
 707         struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
 708         struct sbq_wait *wait = &khd->domain_wait[sched_domain];
 709         struct sbq_wait_state *ws;
 710         int nr;
 711
 712         nr = __sbitmap_queue_get(domain_tokens);
 713
 714         /*
 715          * If we failed to get a domain token, make sure the hardware queue is
 716          * run when one becomes available. Note that this is serialized on
 717          * khd->lock, but we still need to be careful about the waker.
 718          */
 719         if (nr < 0 && list_empty_careful(&wait->wait.entry)) {
 720                 ws = sbq_wait_ptr(domain_tokens,
 721                                   &khd->wait_index[sched_domain]);
 722                 khd->domain_ws[sched_domain] = ws;
 723                 sbitmap_add_wait_queue(domain_tokens, ws, wait);
 724
 725                 /*
 726                  * Try again in case a token was freed before we got on the wait
 727                  * queue.
 728                  */
 729                 nr = __sbitmap_queue_get(domain_tokens);
 730         }
 731
 732         /*
 733          * If we got a token while we were on the wait queue, remove ourselves
 734          * from the wait queue to ensure that all wake ups make forward
 735          * progress. It's possible that the waker already deleted the entry
 736          * between the !list_empty_careful() check and us grabbing the lock, but
 737          * list_del_init() is okay with that.
 738          */
 739         if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) {
 740                 ws = khd->domain_ws[sched_domain];
 741                 spin_lock_irq(&ws->wait.lock);
 742                 sbitmap_del_wait_queue(wait);
 743                 spin_unlock_irq(&ws->wait.lock);
 744         }
 745
 746         return nr;
 747 }
 748
 749 static struct request *
 750 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
 751                           struct kyber_hctx_data *khd,
 752                           struct blk_mq_hw_ctx *hctx)
 753 {
 754         struct list_head *rqs;
 755         struct request *rq;
 756         int nr;
 757
 758         rqs = &khd->rqs[khd->cur_domain];
 759
 760         /*
 761          * If we already have a flushed request, then we just need to get a
 762          * token for it. Otherwise, if there are pending requests in the kcqs,
 763          * flush the kcqs, but only if we can get a token. If not, we should
 764          * leave the requests in the kcqs so that they can be merged. Note that
 765          * khd->lock serializes the flushes, so if we observed any bit set in
 766          * the kcq_map, we will always get a request.
 767          */
 768         rq = list_first_entry_or_null(rqs, struct request, queuelist);
 769         if (rq) {
 770                 nr = kyber_get_domain_token(kqd, khd, hctx);
 771                 if (nr >= 0) {
 772                         khd->batching++;
 773                         rq_set_domain_token(rq, nr);
 774                         list_del_init(&rq->queuelist);
 775                         return rq;
 776                 } else {
 777                         trace_kyber_throttled(kqd->q,
 778                                               kyber_domain_names[khd->cur_domain]);
 779                 }
 780         } else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) {
 781                 nr = kyber_get_domain_token(kqd, khd, hctx);
 782                 if (nr >= 0) {
 783                         kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs);
 784                         rq = list_first_entry(rqs, struct request, queuelist);
 785                         khd->batching++;
 786                         rq_set_domain_token(rq, nr);
 787                         list_del_init(&rq->queuelist);
 788                         return rq;
 789                 } else {
 790                         trace_kyber_throttled(kqd->q,
 791                                               kyber_domain_names[khd->cur_domain]);
 792                 }
 793         }
 794
 795         /* There were either no pending requests or no tokens. */
 796         return NULL;
 797 }
 798
 799 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
 800 {
 801         struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
 802         struct kyber_hctx_data *khd = hctx->sched_data;
 803         struct request *rq;
 804         int i;
 805
 806         spin_lock(&khd->lock);
 807
 808         /*
 809          * First, if we are still entitled to batch, try to dispatch a request
 810          * from the batch.
 811          */
 812         if (khd->batching < kyber_batch_size[khd->cur_domain]) {
 813                 rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
 814                 if (rq)
 815                         goto out;
 816         }
 817
 818         /*
 819          * Either,
 820          * 1. We were no longer entitled to a batch.
 821          * 2. The domain we were batching didn't have any requests.
 822          * 3. The domain we were batching was out of tokens.
 823          *
 824          * Start another batch. Note that this wraps back around to the original
 825          * domain if no other domains have requests or tokens.
 826          */
 827         khd->batching = 0;
 828         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 829                 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
 830                         khd->cur_domain = 0;
 831                 else
 832                         khd->cur_domain++;
 833
 834                 rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
 835                 if (rq)
 836                         goto out;
 837         }
 838
 839         rq = NULL;
 840 out:
 841         spin_unlock(&khd->lock);
 842         return rq;
 843 }
 844
 845 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
 846 {
 847         struct kyber_hctx_data *khd = hctx->sched_data;
 848         int i;
 849
 850         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 851                 if (!list_empty_careful(&khd->rqs[i]) ||
 852                     sbitmap_any_bit_set(&khd->kcq_map[i]))
 853                         return true;
 854         }
 855
 856         return false;
 857 }
 858
 859 #define KYBER_LAT_SHOW_STORE(domain, name)                              \
 860 static ssize_t kyber_##name##_lat_show(struct elevator_queue *e,        \
 861                                        char *page)                      \
 862 {                                                                       \
 863         struct kyber_queue_data *kqd = e->elevator_data;                \
 864                                                                         \
 865         return sprintf(page, "%llu\n", kqd->latency_targets[domain]);   \
 866 }                                                                       \
 867                                                                         \
 868 static ssize_t kyber_##name##_lat_store(struct elevator_queue *e,       \
 869                                         const char *page, size_t count) \
 870 {                                                                       \
 871         struct kyber_queue_data *kqd = e->elevator_data;                \
 872         unsigned long long nsec;                                        \
 873         int ret;                                                        \
 874                                                                         \
 875         ret = kstrtoull(page, 10, &nsec);                               \
 876         if (ret)                                                        \
 877                 return ret;                                             \
 878                                                                         \
 879         kqd->latency_targets[domain] = nsec;                            \
 880                                                                         \
 881         return count;                                                   \
 882 }
 883 KYBER_LAT_SHOW_STORE(KYBER_READ, read);
 884 KYBER_LAT_SHOW_STORE(KYBER_WRITE, write);
 885 #undef KYBER_LAT_SHOW_STORE
 886
 887 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
 888 static struct elv_fs_entry kyber_sched_attrs[] = {
 889         KYBER_LAT_ATTR(read),
 890         KYBER_LAT_ATTR(write),
 891         __ATTR_NULL
 892 };
 893 #undef KYBER_LAT_ATTR
 894
 895 #ifdef CONFIG_BLK_DEBUG_FS
 896 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name)                        \
 897 static int kyber_##name##_tokens_show(void *data, struct seq_file *m)   \
 898 {                                                                       \
 899         struct request_queue *q = data;                                 \
 900         struct kyber_queue_data *kqd = q->elevator->elevator_data;      \
 901                                                                         \
 902         sbitmap_queue_show(&kqd->domain_tokens[domain], m);             \
 903         return 0;                                                       \
 904 }                                                                       \
 905                                                                         \
 906 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos)  \
 907         __acquires(&khd->lock)                                          \
 908 {                                                                       \
 909         struct blk_mq_hw_ctx *hctx = m->private;                        \
 910         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 911                                                                         \
 912         spin_lock(&khd->lock);                                          \
 913         return seq_list_start(&khd->rqs[domain], *pos);                 \
 914 }                                                                       \
 915                                                                         \
 916 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v,       \
 917                                      loff_t *pos)                       \
 918 {                                                                       \
 919         struct blk_mq_hw_ctx *hctx = m->private;                        \
 920         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 921                                                                         \
 922         return seq_list_next(v, &khd->rqs[domain], pos);                \
 923 }                                                                       \
 924                                                                         \
 925 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v)        \
 926         __releases(&khd->lock)                                          \
 927 {                                                                       \
 928         struct blk_mq_hw_ctx *hctx = m->private;                        \
 929         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 930                                                                         \
 931         spin_unlock(&khd->lock);                                        \
 932 }                                                                       \
 933                                                                         \
 934 static const struct seq_operations kyber_##name##_rqs_seq_ops = {       \
 935         .start  = kyber_##name##_rqs_start,                             \
 936         .next   = kyber_##name##_rqs_next,                              \
 937         .stop   = kyber_##name##_rqs_stop,                              \
 938         .show   = blk_mq_debugfs_rq_show,                               \
 939 };                                                                      \
 940                                                                         \
 941 static int kyber_##name##_waiting_show(void *data, struct seq_file *m)  \
 942 {                                                                       \
 943         struct blk_mq_hw_ctx *hctx = data;                              \
 944         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 945         wait_queue_entry_t *wait = &khd->domain_wait[domain].wait;      \
 946                                                                         \
 947         seq_printf(m, "%d\n", !list_empty_careful(&wait->entry));       \
 948         return 0;                                                       \
 949 }
 950 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
 951 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write)
 952 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
 953 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
 954 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
 955
 956 static int kyber_async_depth_show(void *data, struct seq_file *m)
 957 {
 958         struct request_queue *q = data;
 959         struct kyber_queue_data *kqd = q->elevator->elevator_data;
 960
 961         seq_printf(m, "%u\n", kqd->async_depth);
 962         return 0;
 963 }
 964
 965 static int kyber_cur_domain_show(void *data, struct seq_file *m)
 966 {
 967         struct blk_mq_hw_ctx *hctx = data;
 968         struct kyber_hctx_data *khd = hctx->sched_data;
 969
 970         seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]);
 971         return 0;
 972 }
 973
 974 static int kyber_batching_show(void *data, struct seq_file *m)
 975 {
 976         struct blk_mq_hw_ctx *hctx = data;
 977         struct kyber_hctx_data *khd = hctx->sched_data;
 978
 979         seq_printf(m, "%u\n", khd->batching);
 980         return 0;
 981 }
 982
 983 #define KYBER_QUEUE_DOMAIN_ATTRS(name)  \
 984         {#name "_tokens", 0400, kyber_##name##_tokens_show}
 985 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
 986         KYBER_QUEUE_DOMAIN_ATTRS(read),
 987         KYBER_QUEUE_DOMAIN_ATTRS(write),
 988         KYBER_QUEUE_DOMAIN_ATTRS(discard),
 989         KYBER_QUEUE_DOMAIN_ATTRS(other),
 990         {"async_depth", 0400, kyber_async_depth_show},
 991         {},
 992 };
 993 #undef KYBER_QUEUE_DOMAIN_ATTRS
 994
 995 #define KYBER_HCTX_DOMAIN_ATTRS(name)                                   \
 996         {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops},   \
 997         {#name "_waiting", 0400, kyber_##name##_waiting_show}
 998 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
 999         KYBER_HCTX_DOMAIN_ATTRS(read),
1000         KYBER_HCTX_DOMAIN_ATTRS(write),
1001         KYBER_HCTX_DOMAIN_ATTRS(discard),
1002         KYBER_HCTX_DOMAIN_ATTRS(other),
1003         {"cur_domain", 0400, kyber_cur_domain_show},
1004         {"batching", 0400, kyber_batching_show},
1005         {},
1006 };
1007 #undef KYBER_HCTX_DOMAIN_ATTRS
1008 #endif
1009
1010 static struct elevator_type kyber_sched = {
1011         .ops = {
1012                 .init_sched = kyber_init_sched,
1013                 .exit_sched = kyber_exit_sched,
1014                 .init_hctx = kyber_init_hctx,
1015                 .exit_hctx = kyber_exit_hctx,
1016                 .limit_depth = kyber_limit_depth,
1017                 .bio_merge = kyber_bio_merge,
1018                 .prepare_request = kyber_prepare_request,
1019                 .insert_requests = kyber_insert_requests,
1020                 .finish_request = kyber_finish_request,
1021                 .requeue_request = kyber_finish_request,
1022                 .completed_request = kyber_completed_request,
1023                 .dispatch_request = kyber_dispatch_request,
1024                 .has_work = kyber_has_work,
1025                 .depth_updated = kyber_depth_updated,
1026         },
1027 #ifdef CONFIG_BLK_DEBUG_FS
1028         .queue_debugfs_attrs = kyber_queue_debugfs_attrs,
1029         .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs,
1030 #endif
1031         .elevator_attrs = kyber_sched_attrs,
1032         .elevator_name = "kyber",
1033         .elevator_features = ELEVATOR_F_MQ_AWARE,
1034         .elevator_owner = THIS_MODULE,
1035 };
1036
1037 static int __init kyber_init(void)
1038 {
1039         return elv_register(&kyber_sched);
1040 }
1041
1042 static void __exit kyber_exit(void)
1043 {
1044         elv_unregister(&kyber_sched);
1045 }
1046
1047 module_init(kyber_init);
1048 module_exit(kyber_exit);
1049
1050 MODULE_AUTHOR("Omar Sandoval");
1051 MODULE_LICENSE("GPL");
1052 MODULE_DESCRIPTION("Kyber I/O scheduler");