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/module.h>
  12 #include <linux/sbitmap.h>
  13
  14 #include <trace/events/block.h>
  15
  16 #include "elevator.h"
  17 #include "blk.h"
  18 #include "blk-mq.h"
  19 #include "blk-mq-debugfs.h"
  20 #include "blk-mq-sched.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         dev_t dev;
 153
 154         /*
 155          * Each scheduling domain has a limited number of in-flight requests
 156          * device-wide, limited by these tokens.
 157          */
 158         struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
 159
 160         /*
 161          * Async request percentage, converted to per-word depth for
 162          * sbitmap_get_shallow().
 163          */
 164         unsigned int async_depth;
 165
 166         struct kyber_cpu_latency __percpu *cpu_latency;
 167
 168         /* Timer for stats aggregation and adjusting domain tokens. */
 169         struct timer_list timer;
 170
 171         unsigned int latency_buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
 172
 173         unsigned long latency_timeout[KYBER_OTHER];
 174
 175         int domain_p99[KYBER_OTHER];
 176
 177         /* Target latencies in nanoseconds. */
 178         u64 latency_targets[KYBER_OTHER];
 179 };
 180
 181 struct kyber_hctx_data {
 182         spinlock_t lock;
 183         struct list_head rqs[KYBER_NUM_DOMAINS];
 184         unsigned int cur_domain;
 185         unsigned int batching;
 186         struct kyber_ctx_queue *kcqs;
 187         struct sbitmap kcq_map[KYBER_NUM_DOMAINS];
 188         struct sbq_wait domain_wait[KYBER_NUM_DOMAINS];
 189         struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
 190         atomic_t wait_index[KYBER_NUM_DOMAINS];
 191 };
 192
 193 static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
 194                              void *key);
 195
 196 static unsigned int kyber_sched_domain(blk_opf_t opf)
 197 {
 198         switch (opf & REQ_OP_MASK) {
 199         case REQ_OP_READ:
 200                 return KYBER_READ;
 201         case REQ_OP_WRITE:
 202                 return KYBER_WRITE;
 203         case REQ_OP_DISCARD:
 204                 return KYBER_DISCARD;
 205         default:
 206                 return KYBER_OTHER;
 207         }
 208 }
 209
 210 static void flush_latency_buckets(struct kyber_queue_data *kqd,
 211                                   struct kyber_cpu_latency *cpu_latency,
 212                                   unsigned int sched_domain, unsigned int type)
 213 {
 214         unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
 215         atomic_t *cpu_buckets = cpu_latency->buckets[sched_domain][type];
 216         unsigned int bucket;
 217
 218         for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
 219                 buckets[bucket] += atomic_xchg(&cpu_buckets[bucket], 0);
 220 }
 221
 222 /*
 223  * Calculate the histogram bucket with the given percentile rank, or -1 if there
 224  * aren't enough samples yet.
 225  */
 226 static int calculate_percentile(struct kyber_queue_data *kqd,
 227                                 unsigned int sched_domain, unsigned int type,
 228                                 unsigned int percentile)
 229 {
 230         unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
 231         unsigned int bucket, samples = 0, percentile_samples;
 232
 233         for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
 234                 samples += buckets[bucket];
 235
 236         if (!samples)
 237                 return -1;
 238
 239         /*
 240          * We do the calculation once we have 500 samples or one second passes
 241          * since the first sample was recorded, whichever comes first.
 242          */
 243         if (!kqd->latency_timeout[sched_domain])
 244                 kqd->latency_timeout[sched_domain] = max(jiffies + HZ, 1UL);
 245         if (samples < 500 &&
 246             time_is_after_jiffies(kqd->latency_timeout[sched_domain])) {
 247                 return -1;
 248         }
 249         kqd->latency_timeout[sched_domain] = 0;
 250
 251         percentile_samples = DIV_ROUND_UP(samples * percentile, 100);
 252         for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS - 1; bucket++) {
 253                 if (buckets[bucket] >= percentile_samples)
 254                         break;
 255                 percentile_samples -= buckets[bucket];
 256         }
 257         memset(buckets, 0, sizeof(kqd->latency_buckets[sched_domain][type]));
 258
 259         trace_kyber_latency(kqd->dev, kyber_domain_names[sched_domain],
 260                             kyber_latency_type_names[type], percentile,
 261                             bucket + 1, 1 << KYBER_LATENCY_SHIFT, samples);
 262
 263         return bucket;
 264 }
 265
 266 static void kyber_resize_domain(struct kyber_queue_data *kqd,
 267                                 unsigned int sched_domain, unsigned int depth)
 268 {
 269         depth = clamp(depth, 1U, kyber_depth[sched_domain]);
 270         if (depth != kqd->domain_tokens[sched_domain].sb.depth) {
 271                 sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
 272                 trace_kyber_adjust(kqd->dev, kyber_domain_names[sched_domain],
 273                                    depth);
 274         }
 275 }
 276
 277 static void kyber_timer_fn(struct timer_list *t)
 278 {
 279         struct kyber_queue_data *kqd = from_timer(kqd, t, timer);
 280         unsigned int sched_domain;
 281         int cpu;
 282         bool bad = false;
 283
 284         /* Sum all of the per-cpu latency histograms. */
 285         for_each_online_cpu(cpu) {
 286                 struct kyber_cpu_latency *cpu_latency;
 287
 288                 cpu_latency = per_cpu_ptr(kqd->cpu_latency, cpu);
 289                 for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
 290                         flush_latency_buckets(kqd, cpu_latency, sched_domain,
 291                                               KYBER_TOTAL_LATENCY);
 292                         flush_latency_buckets(kqd, cpu_latency, sched_domain,
 293                                               KYBER_IO_LATENCY);
 294                 }
 295         }
 296
 297         /*
 298          * Check if any domains have a high I/O latency, which might indicate
 299          * congestion in the device. Note that we use the p90; we don't want to
 300          * be too sensitive to outliers here.
 301          */
 302         for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
 303                 int p90;
 304
 305                 p90 = calculate_percentile(kqd, sched_domain, KYBER_IO_LATENCY,
 306                                            90);
 307                 if (p90 >= KYBER_GOOD_BUCKETS)
 308                         bad = true;
 309         }
 310
 311         /*
 312          * Adjust the scheduling domain depths. If we determined that there was
 313          * congestion, we throttle all domains with good latencies. Either way,
 314          * we ease up on throttling domains with bad latencies.
 315          */
 316         for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
 317                 unsigned int orig_depth, depth;
 318                 int p99;
 319
 320                 p99 = calculate_percentile(kqd, sched_domain,
 321                                            KYBER_TOTAL_LATENCY, 99);
 322                 /*
 323                  * This is kind of subtle: different domains will not
 324                  * necessarily have enough samples to calculate the latency
 325                  * percentiles during the same window, so we have to remember
 326                  * the p99 for the next time we observe congestion; once we do,
 327                  * we don't want to throttle again until we get more data, so we
 328                  * reset it to -1.
 329                  */
 330                 if (bad) {
 331                         if (p99 < 0)
 332                                 p99 = kqd->domain_p99[sched_domain];
 333                         kqd->domain_p99[sched_domain] = -1;
 334                 } else if (p99 >= 0) {
 335                         kqd->domain_p99[sched_domain] = p99;
 336                 }
 337                 if (p99 < 0)
 338                         continue;
 339
 340                 /*
 341                  * If this domain has bad latency, throttle less. Otherwise,
 342                  * throttle more iff we determined that there is congestion.
 343                  *
 344                  * The new depth is scaled linearly with the p99 latency vs the
 345                  * latency target. E.g., if the p99 is 3/4 of the target, then
 346                  * we throttle down to 3/4 of the current depth, and if the p99
 347                  * is 2x the target, then we double the depth.
 348                  */
 349                 if (bad || p99 >= KYBER_GOOD_BUCKETS) {
 350                         orig_depth = kqd->domain_tokens[sched_domain].sb.depth;
 351                         depth = (orig_depth * (p99 + 1)) >> KYBER_LATENCY_SHIFT;
 352                         kyber_resize_domain(kqd, sched_domain, depth);
 353                 }
 354         }
 355 }
 356
 357 static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
 358 {
 359         struct kyber_queue_data *kqd;
 360         int ret = -ENOMEM;
 361         int i;
 362
 363         kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
 364         if (!kqd)
 365                 goto err;
 366
 367         kqd->q = q;
 368         kqd->dev = disk_devt(q->disk);
 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         blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);
 423
 424         eq->elevator_data = kqd;
 425         q->elevator = eq;
 426
 427         return 0;
 428 }
 429
 430 static void kyber_exit_sched(struct elevator_queue *e)
 431 {
 432         struct kyber_queue_data *kqd = e->elevator_data;
 433         int i;
 434
 435         timer_shutdown_sync(&kqd->timer);
 436         blk_stat_disable_accounting(kqd->q);
 437
 438         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 439                 sbitmap_queue_free(&kqd->domain_tokens[i]);
 440         free_percpu(kqd->cpu_latency);
 441         kfree(kqd);
 442 }
 443
 444 static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq)
 445 {
 446         unsigned int i;
 447
 448         spin_lock_init(&kcq->lock);
 449         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 450                 INIT_LIST_HEAD(&kcq->rq_list[i]);
 451 }
 452
 453 static void kyber_depth_updated(struct blk_mq_hw_ctx *hctx)
 454 {
 455         struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
 456         struct blk_mq_tags *tags = hctx->sched_tags;
 457         unsigned int shift = tags->bitmap_tags.sb.shift;
 458
 459         kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
 460
 461         sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, kqd->async_depth);
 462 }
 463
 464 static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 465 {
 466         struct kyber_hctx_data *khd;
 467         int i;
 468
 469         khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
 470         if (!khd)
 471                 return -ENOMEM;
 472
 473         khd->kcqs = kmalloc_array_node(hctx->nr_ctx,
 474                                        sizeof(struct kyber_ctx_queue),
 475                                        GFP_KERNEL, hctx->numa_node);
 476         if (!khd->kcqs)
 477                 goto err_khd;
 478
 479         for (i = 0; i < hctx->nr_ctx; i++)
 480                 kyber_ctx_queue_init(&khd->kcqs[i]);
 481
 482         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 483                 if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx,
 484                                       ilog2(8), GFP_KERNEL, hctx->numa_node,
 485                                       false, false)) {
 486                         while (--i >= 0)
 487                                 sbitmap_free(&khd->kcq_map[i]);
 488                         goto err_kcqs;
 489                 }
 490         }
 491
 492         spin_lock_init(&khd->lock);
 493
 494         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 495                 INIT_LIST_HEAD(&khd->rqs[i]);
 496                 khd->domain_wait[i].sbq = NULL;
 497                 init_waitqueue_func_entry(&khd->domain_wait[i].wait,
 498                                           kyber_domain_wake);
 499                 khd->domain_wait[i].wait.private = hctx;
 500                 INIT_LIST_HEAD(&khd->domain_wait[i].wait.entry);
 501                 atomic_set(&khd->wait_index[i], 0);
 502         }
 503
 504         khd->cur_domain = 0;
 505         khd->batching = 0;
 506
 507         hctx->sched_data = khd;
 508         kyber_depth_updated(hctx);
 509
 510         return 0;
 511
 512 err_kcqs:
 513         kfree(khd->kcqs);
 514 err_khd:
 515         kfree(khd);
 516         return -ENOMEM;
 517 }
 518
 519 static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 520 {
 521         struct kyber_hctx_data *khd = hctx->sched_data;
 522         int i;
 523
 524         for (i = 0; i < KYBER_NUM_DOMAINS; i++)
 525                 sbitmap_free(&khd->kcq_map[i]);
 526         kfree(khd->kcqs);
 527         kfree(hctx->sched_data);
 528 }
 529
 530 static int rq_get_domain_token(struct request *rq)
 531 {
 532         return (long)rq->elv.priv[0];
 533 }
 534
 535 static void rq_set_domain_token(struct request *rq, int token)
 536 {
 537         rq->elv.priv[0] = (void *)(long)token;
 538 }
 539
 540 static void rq_clear_domain_token(struct kyber_queue_data *kqd,
 541                                   struct request *rq)
 542 {
 543         unsigned int sched_domain;
 544         int nr;
 545
 546         nr = rq_get_domain_token(rq);
 547         if (nr != -1) {
 548                 sched_domain = kyber_sched_domain(rq->cmd_flags);
 549                 sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
 550                                     rq->mq_ctx->cpu);
 551         }
 552 }
 553
 554 static void kyber_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 555 {
 556         /*
 557          * We use the scheduler tags as per-hardware queue queueing tokens.
 558          * Async requests can be limited at this stage.
 559          */
 560         if (!op_is_sync(opf)) {
 561                 struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
 562
 563                 data->shallow_depth = kqd->async_depth;
 564         }
 565 }
 566
 567 static bool kyber_bio_merge(struct request_queue *q, struct bio *bio,
 568                 unsigned int nr_segs)
 569 {
 570         struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 571         struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
 572         struct kyber_hctx_data *khd = hctx->sched_data;
 573         struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]];
 574         unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);
 575         struct list_head *rq_list = &kcq->rq_list[sched_domain];
 576         bool merged;
 577
 578         spin_lock(&kcq->lock);
 579         merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs);
 580         spin_unlock(&kcq->lock);
 581
 582         return merged;
 583 }
 584
 585 static void kyber_prepare_request(struct request *rq)
 586 {
 587         rq_set_domain_token(rq, -1);
 588 }
 589
 590 static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx,
 591                                   struct list_head *rq_list,
 592                                   blk_insert_t flags)
 593 {
 594         struct kyber_hctx_data *khd = hctx->sched_data;
 595         struct request *rq, *next;
 596
 597         list_for_each_entry_safe(rq, next, rq_list, queuelist) {
 598                 unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags);
 599                 struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]];
 600                 struct list_head *head = &kcq->rq_list[sched_domain];
 601
 602                 spin_lock(&kcq->lock);
 603                 trace_block_rq_insert(rq);
 604                 if (flags & BLK_MQ_INSERT_AT_HEAD)
 605                         list_move(&rq->queuelist, head);
 606                 else
 607                         list_move_tail(&rq->queuelist, head);
 608                 sbitmap_set_bit(&khd->kcq_map[sched_domain],
 609                                 rq->mq_ctx->index_hw[hctx->type]);
 610                 spin_unlock(&kcq->lock);
 611         }
 612 }
 613
 614 static void kyber_finish_request(struct request *rq)
 615 {
 616         struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
 617
 618         rq_clear_domain_token(kqd, rq);
 619 }
 620
 621 static void add_latency_sample(struct kyber_cpu_latency *cpu_latency,
 622                                unsigned int sched_domain, unsigned int type,
 623                                u64 target, u64 latency)
 624 {
 625         unsigned int bucket;
 626         u64 divisor;
 627
 628         if (latency > 0) {
 629                 divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1);
 630                 bucket = min_t(unsigned int, div64_u64(latency - 1, divisor),
 631                                KYBER_LATENCY_BUCKETS - 1);
 632         } else {
 633                 bucket = 0;
 634         }
 635
 636         atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]);
 637 }
 638
 639 static void kyber_completed_request(struct request *rq, u64 now)
 640 {
 641         struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
 642         struct kyber_cpu_latency *cpu_latency;
 643         unsigned int sched_domain;
 644         u64 target;
 645
 646         sched_domain = kyber_sched_domain(rq->cmd_flags);
 647         if (sched_domain == KYBER_OTHER)
 648                 return;
 649
 650         cpu_latency = get_cpu_ptr(kqd->cpu_latency);
 651         target = kqd->latency_targets[sched_domain];
 652         add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY,
 653                            target, now - rq->start_time_ns);
 654         add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target,
 655                            now - rq->io_start_time_ns);
 656         put_cpu_ptr(kqd->cpu_latency);
 657
 658         timer_reduce(&kqd->timer, jiffies + HZ / 10);
 659 }
 660
 661 struct flush_kcq_data {
 662         struct kyber_hctx_data *khd;
 663         unsigned int sched_domain;
 664         struct list_head *list;
 665 };
 666
 667 static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data)
 668 {
 669         struct flush_kcq_data *flush_data = data;
 670         struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr];
 671
 672         spin_lock(&kcq->lock);
 673         list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain],
 674                               flush_data->list);
 675         sbitmap_clear_bit(sb, bitnr);
 676         spin_unlock(&kcq->lock);
 677
 678         return true;
 679 }
 680
 681 static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd,
 682                                   unsigned int sched_domain,
 683                                   struct list_head *list)
 684 {
 685         struct flush_kcq_data data = {
 686                 .khd = khd,
 687                 .sched_domain = sched_domain,
 688                 .list = list,
 689         };
 690
 691         sbitmap_for_each_set(&khd->kcq_map[sched_domain],
 692                              flush_busy_kcq, &data);
 693 }
 694
 695 static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags,
 696                              void *key)
 697 {
 698         struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private);
 699         struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait);
 700
 701         sbitmap_del_wait_queue(wait);
 702         blk_mq_run_hw_queue(hctx, true);
 703         return 1;
 704 }
 705
 706 static int kyber_get_domain_token(struct kyber_queue_data *kqd,
 707                                   struct kyber_hctx_data *khd,
 708                                   struct blk_mq_hw_ctx *hctx)
 709 {
 710         unsigned int sched_domain = khd->cur_domain;
 711         struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
 712         struct sbq_wait *wait = &khd->domain_wait[sched_domain];
 713         struct sbq_wait_state *ws;
 714         int nr;
 715
 716         nr = __sbitmap_queue_get(domain_tokens);
 717
 718         /*
 719          * If we failed to get a domain token, make sure the hardware queue is
 720          * run when one becomes available. Note that this is serialized on
 721          * khd->lock, but we still need to be careful about the waker.
 722          */
 723         if (nr < 0 && list_empty_careful(&wait->wait.entry)) {
 724                 ws = sbq_wait_ptr(domain_tokens,
 725                                   &khd->wait_index[sched_domain]);
 726                 khd->domain_ws[sched_domain] = ws;
 727                 sbitmap_add_wait_queue(domain_tokens, ws, wait);
 728
 729                 /*
 730                  * Try again in case a token was freed before we got on the wait
 731                  * queue.
 732                  */
 733                 nr = __sbitmap_queue_get(domain_tokens);
 734         }
 735
 736         /*
 737          * If we got a token while we were on the wait queue, remove ourselves
 738          * from the wait queue to ensure that all wake ups make forward
 739          * progress. It's possible that the waker already deleted the entry
 740          * between the !list_empty_careful() check and us grabbing the lock, but
 741          * list_del_init() is okay with that.
 742          */
 743         if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) {
 744                 ws = khd->domain_ws[sched_domain];
 745                 spin_lock_irq(&ws->wait.lock);
 746                 sbitmap_del_wait_queue(wait);
 747                 spin_unlock_irq(&ws->wait.lock);
 748         }
 749
 750         return nr;
 751 }
 752
 753 static struct request *
 754 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
 755                           struct kyber_hctx_data *khd,
 756                           struct blk_mq_hw_ctx *hctx)
 757 {
 758         struct list_head *rqs;
 759         struct request *rq;
 760         int nr;
 761
 762         rqs = &khd->rqs[khd->cur_domain];
 763
 764         /*
 765          * If we already have a flushed request, then we just need to get a
 766          * token for it. Otherwise, if there are pending requests in the kcqs,
 767          * flush the kcqs, but only if we can get a token. If not, we should
 768          * leave the requests in the kcqs so that they can be merged. Note that
 769          * khd->lock serializes the flushes, so if we observed any bit set in
 770          * the kcq_map, we will always get a request.
 771          */
 772         rq = list_first_entry_or_null(rqs, struct request, queuelist);
 773         if (rq) {
 774                 nr = kyber_get_domain_token(kqd, khd, hctx);
 775                 if (nr >= 0) {
 776                         khd->batching++;
 777                         rq_set_domain_token(rq, nr);
 778                         list_del_init(&rq->queuelist);
 779                         return rq;
 780                 } else {
 781                         trace_kyber_throttled(kqd->dev,
 782                                               kyber_domain_names[khd->cur_domain]);
 783                 }
 784         } else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) {
 785                 nr = kyber_get_domain_token(kqd, khd, hctx);
 786                 if (nr >= 0) {
 787                         kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs);
 788                         rq = list_first_entry(rqs, struct request, queuelist);
 789                         khd->batching++;
 790                         rq_set_domain_token(rq, nr);
 791                         list_del_init(&rq->queuelist);
 792                         return rq;
 793                 } else {
 794                         trace_kyber_throttled(kqd->dev,
 795                                               kyber_domain_names[khd->cur_domain]);
 796                 }
 797         }
 798
 799         /* There were either no pending requests or no tokens. */
 800         return NULL;
 801 }
 802
 803 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
 804 {
 805         struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
 806         struct kyber_hctx_data *khd = hctx->sched_data;
 807         struct request *rq;
 808         int i;
 809
 810         spin_lock(&khd->lock);
 811
 812         /*
 813          * First, if we are still entitled to batch, try to dispatch a request
 814          * from the batch.
 815          */
 816         if (khd->batching < kyber_batch_size[khd->cur_domain]) {
 817                 rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
 818                 if (rq)
 819                         goto out;
 820         }
 821
 822         /*
 823          * Either,
 824          * 1. We were no longer entitled to a batch.
 825          * 2. The domain we were batching didn't have any requests.
 826          * 3. The domain we were batching was out of tokens.
 827          *
 828          * Start another batch. Note that this wraps back around to the original
 829          * domain if no other domains have requests or tokens.
 830          */
 831         khd->batching = 0;
 832         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 833                 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
 834                         khd->cur_domain = 0;
 835                 else
 836                         khd->cur_domain++;
 837
 838                 rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
 839                 if (rq)
 840                         goto out;
 841         }
 842
 843         rq = NULL;
 844 out:
 845         spin_unlock(&khd->lock);
 846         return rq;
 847 }
 848
 849 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
 850 {
 851         struct kyber_hctx_data *khd = hctx->sched_data;
 852         int i;
 853
 854         for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
 855                 if (!list_empty_careful(&khd->rqs[i]) ||
 856                     sbitmap_any_bit_set(&khd->kcq_map[i]))
 857                         return true;
 858         }
 859
 860         return false;
 861 }
 862
 863 #define KYBER_LAT_SHOW_STORE(domain, name)                              \
 864 static ssize_t kyber_##name##_lat_show(struct elevator_queue *e,        \
 865                                        char *page)                      \
 866 {                                                                       \
 867         struct kyber_queue_data *kqd = e->elevator_data;                \
 868                                                                         \
 869         return sprintf(page, "%llu\n", kqd->latency_targets[domain]);   \
 870 }                                                                       \
 871                                                                         \
 872 static ssize_t kyber_##name##_lat_store(struct elevator_queue *e,       \
 873                                         const char *page, size_t count) \
 874 {                                                                       \
 875         struct kyber_queue_data *kqd = e->elevator_data;                \
 876         unsigned long long nsec;                                        \
 877         int ret;                                                        \
 878                                                                         \
 879         ret = kstrtoull(page, 10, &nsec);                               \
 880         if (ret)                                                        \
 881                 return ret;                                             \
 882                                                                         \
 883         kqd->latency_targets[domain] = nsec;                            \
 884                                                                         \
 885         return count;                                                   \
 886 }
 887 KYBER_LAT_SHOW_STORE(KYBER_READ, read);
 888 KYBER_LAT_SHOW_STORE(KYBER_WRITE, write);
 889 #undef KYBER_LAT_SHOW_STORE
 890
 891 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
 892 static struct elv_fs_entry kyber_sched_attrs[] = {
 893         KYBER_LAT_ATTR(read),
 894         KYBER_LAT_ATTR(write),
 895         __ATTR_NULL
 896 };
 897 #undef KYBER_LAT_ATTR
 898
 899 #ifdef CONFIG_BLK_DEBUG_FS
 900 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name)                        \
 901 static int kyber_##name##_tokens_show(void *data, struct seq_file *m)   \
 902 {                                                                       \
 903         struct request_queue *q = data;                                 \
 904         struct kyber_queue_data *kqd = q->elevator->elevator_data;      \
 905                                                                         \
 906         sbitmap_queue_show(&kqd->domain_tokens[domain], m);             \
 907         return 0;                                                       \
 908 }                                                                       \
 909                                                                         \
 910 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos)  \
 911         __acquires(&khd->lock)                                          \
 912 {                                                                       \
 913         struct blk_mq_hw_ctx *hctx = m->private;                        \
 914         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 915                                                                         \
 916         spin_lock(&khd->lock);                                          \
 917         return seq_list_start(&khd->rqs[domain], *pos);                 \
 918 }                                                                       \
 919                                                                         \
 920 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v,       \
 921                                      loff_t *pos)                       \
 922 {                                                                       \
 923         struct blk_mq_hw_ctx *hctx = m->private;                        \
 924         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 925                                                                         \
 926         return seq_list_next(v, &khd->rqs[domain], pos);                \
 927 }                                                                       \
 928                                                                         \
 929 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v)        \
 930         __releases(&khd->lock)                                          \
 931 {                                                                       \
 932         struct blk_mq_hw_ctx *hctx = m->private;                        \
 933         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 934                                                                         \
 935         spin_unlock(&khd->lock);                                        \
 936 }                                                                       \
 937                                                                         \
 938 static const struct seq_operations kyber_##name##_rqs_seq_ops = {       \
 939         .start  = kyber_##name##_rqs_start,                             \
 940         .next   = kyber_##name##_rqs_next,                              \
 941         .stop   = kyber_##name##_rqs_stop,                              \
 942         .show   = blk_mq_debugfs_rq_show,                               \
 943 };                                                                      \
 944                                                                         \
 945 static int kyber_##name##_waiting_show(void *data, struct seq_file *m)  \
 946 {                                                                       \
 947         struct blk_mq_hw_ctx *hctx = data;                              \
 948         struct kyber_hctx_data *khd = hctx->sched_data;                 \
 949         wait_queue_entry_t *wait = &khd->domain_wait[domain].wait;      \
 950                                                                         \
 951         seq_printf(m, "%d\n", !list_empty_careful(&wait->entry));       \
 952         return 0;                                                       \
 953 }
 954 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
 955 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write)
 956 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
 957 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
 958 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
 959
 960 static int kyber_async_depth_show(void *data, struct seq_file *m)
 961 {
 962         struct request_queue *q = data;
 963         struct kyber_queue_data *kqd = q->elevator->elevator_data;
 964
 965         seq_printf(m, "%u\n", kqd->async_depth);
 966         return 0;
 967 }
 968
 969 static int kyber_cur_domain_show(void *data, struct seq_file *m)
 970 {
 971         struct blk_mq_hw_ctx *hctx = data;
 972         struct kyber_hctx_data *khd = hctx->sched_data;
 973
 974         seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]);
 975         return 0;
 976 }
 977
 978 static int kyber_batching_show(void *data, struct seq_file *m)
 979 {
 980         struct blk_mq_hw_ctx *hctx = data;
 981         struct kyber_hctx_data *khd = hctx->sched_data;
 982
 983         seq_printf(m, "%u\n", khd->batching);
 984         return 0;
 985 }
 986
 987 #define KYBER_QUEUE_DOMAIN_ATTRS(name)  \
 988         {#name "_tokens", 0400, kyber_##name##_tokens_show}
 989 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
 990         KYBER_QUEUE_DOMAIN_ATTRS(read),
 991         KYBER_QUEUE_DOMAIN_ATTRS(write),
 992         KYBER_QUEUE_DOMAIN_ATTRS(discard),
 993         KYBER_QUEUE_DOMAIN_ATTRS(other),
 994         {"async_depth", 0400, kyber_async_depth_show},
 995         {},
 996 };
 997 #undef KYBER_QUEUE_DOMAIN_ATTRS
 998
 999 #define KYBER_HCTX_DOMAIN_ATTRS(name)                                   \
1000         {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops},   \
1001         {#name "_waiting", 0400, kyber_##name##_waiting_show}
1002 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
1003         KYBER_HCTX_DOMAIN_ATTRS(read),
1004         KYBER_HCTX_DOMAIN_ATTRS(write),
1005         KYBER_HCTX_DOMAIN_ATTRS(discard),
1006         KYBER_HCTX_DOMAIN_ATTRS(other),
1007         {"cur_domain", 0400, kyber_cur_domain_show},
1008         {"batching", 0400, kyber_batching_show},
1009         {},
1010 };
1011 #undef KYBER_HCTX_DOMAIN_ATTRS
1012 #endif
1013
1014 static struct elevator_type kyber_sched = {
1015         .ops = {
1016                 .init_sched = kyber_init_sched,
1017                 .exit_sched = kyber_exit_sched,
1018                 .init_hctx = kyber_init_hctx,
1019                 .exit_hctx = kyber_exit_hctx,
1020                 .limit_depth = kyber_limit_depth,
1021                 .bio_merge = kyber_bio_merge,
1022                 .prepare_request = kyber_prepare_request,
1023                 .insert_requests = kyber_insert_requests,
1024                 .finish_request = kyber_finish_request,
1025                 .requeue_request = kyber_finish_request,
1026                 .completed_request = kyber_completed_request,
1027                 .dispatch_request = kyber_dispatch_request,
1028                 .has_work = kyber_has_work,
1029                 .depth_updated = kyber_depth_updated,
1030         },
1031 #ifdef CONFIG_BLK_DEBUG_FS
1032         .queue_debugfs_attrs = kyber_queue_debugfs_attrs,
1033         .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs,
1034 #endif
1035         .elevator_attrs = kyber_sched_attrs,
1036         .elevator_name = "kyber",
1037         .elevator_owner = THIS_MODULE,
1038 };
1039
1040 static int __init kyber_init(void)
1041 {
1042         return elv_register(&kyber_sched);
1043 }
1044
1045 static void __exit kyber_exit(void)
1046 {
1047         elv_unregister(&kyber_sched);
1048 }
1049
1050 module_init(kyber_init);
1051 module_exit(kyber_exit);
1052
1053 MODULE_AUTHOR("Omar Sandoval");
1054 MODULE_LICENSE("GPL");
1055 MODULE_DESCRIPTION("Kyber I/O scheduler");