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