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 */ 29enum { 30 KYBER_READ, 31 KYBER_WRITE, 32 KYBER_DISCARD, 33 KYBER_OTHER, 34 KYBER_NUM_DOMAINS, 35}; 36 37static const char *kyber_domain_names[] = { 38 [KYBER_READ] = "READ", 39 [KYBER_WRITE] = "WRITE", 40 [KYBER_DISCARD] = "DISCARD", 41 [KYBER_OTHER] = "OTHER", 42}; 43 44enum { 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 */ 60static 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 */ 70static 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 */ 80static 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 */ 100enum { 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 */ 119enum { 120 KYBER_TOTAL_LATENCY, 121 KYBER_IO_LATENCY, 122}; 123 124static 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 */ 133struct 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 */ 141struct 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 150struct 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 180struct 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 192static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags, 193 void *key); 194 195static 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 209static 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 */ 225static 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 265static 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 276static 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 356static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q) 357{ 358 struct kyber_queue_data *kqd; 359 int ret = -ENOMEM; 360 int i; 361 362 kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node); 363 if (!kqd) 364 goto err; 365 366 kqd->q = q; 367 368 kqd->cpu_latency = alloc_percpu_gfp(struct kyber_cpu_latency, 369 GFP_KERNEL | __GFP_ZERO); 370 if (!kqd->cpu_latency) 371 goto err_kqd; 372 373 timer_setup(&kqd->timer, kyber_timer_fn, 0); 374 375 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 376 WARN_ON(!kyber_depth[i]); 377 WARN_ON(!kyber_batch_size[i]); 378 ret = sbitmap_queue_init_node(&kqd->domain_tokens[i], 379 kyber_depth[i], -1, false, 380 GFP_KERNEL, q->node); 381 if (ret) { 382 while (--i >= 0) 383 sbitmap_queue_free(&kqd->domain_tokens[i]); 384 goto err_buckets; 385 } 386 } 387 388 for (i = 0; i < KYBER_OTHER; i++) { 389 kqd->domain_p99[i] = -1; 390 kqd->latency_targets[i] = kyber_latency_targets[i]; 391 } 392 393 return kqd; 394 395err_buckets: 396 free_percpu(kqd->cpu_latency); 397err_kqd: 398 kfree(kqd); 399err: 400 return ERR_PTR(ret); 401} 402 403static int kyber_init_sched(struct request_queue *q, struct elevator_type *e) 404{ 405 struct kyber_queue_data *kqd; 406 struct elevator_queue *eq; 407 408 eq = elevator_alloc(q, e); 409 if (!eq) 410 return -ENOMEM; 411 412 kqd = kyber_queue_data_alloc(q); 413 if (IS_ERR(kqd)) { 414 kobject_put(&eq->kobj); 415 return PTR_ERR(kqd); 416 } 417 418 blk_stat_enable_accounting(q); 419 420 eq->elevator_data = kqd; 421 q->elevator = eq; 422 423 return 0; 424} 425 426static void kyber_exit_sched(struct elevator_queue *e) 427{ 428 struct kyber_queue_data *kqd = e->elevator_data; 429 int i; 430 431 del_timer_sync(&kqd->timer); 432 433 for (i = 0; i < KYBER_NUM_DOMAINS; i++) 434 sbitmap_queue_free(&kqd->domain_tokens[i]); 435 free_percpu(kqd->cpu_latency); 436 kfree(kqd); 437} 438 439static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq) 440{ 441 unsigned int i; 442 443 spin_lock_init(&kcq->lock); 444 for (i = 0; i < KYBER_NUM_DOMAINS; i++) 445 INIT_LIST_HEAD(&kcq->rq_list[i]); 446} 447 448static void kyber_depth_updated(struct blk_mq_hw_ctx *hctx) 449{ 450 struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; 451 struct blk_mq_tags *tags = hctx->sched_tags; 452 unsigned int shift = tags->bitmap_tags->sb.shift; 453 454 kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U; 455 456 sbitmap_queue_min_shallow_depth(tags->bitmap_tags, kqd->async_depth); 457} 458 459static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) 460{ 461 struct kyber_hctx_data *khd; 462 int i; 463 464 khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node); 465 if (!khd) 466 return -ENOMEM; 467 468 khd->kcqs = kmalloc_array_node(hctx->nr_ctx, 469 sizeof(struct kyber_ctx_queue), 470 GFP_KERNEL, hctx->numa_node); 471 if (!khd->kcqs) 472 goto err_khd; 473 474 for (i = 0; i < hctx->nr_ctx; i++) 475 kyber_ctx_queue_init(&khd->kcqs[i]); 476 477 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 478 if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx, 479 ilog2(8), GFP_KERNEL, hctx->numa_node)) { 480 while (--i >= 0) 481 sbitmap_free(&khd->kcq_map[i]); 482 goto err_kcqs; 483 } 484 } 485 486 spin_lock_init(&khd->lock); 487 488 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 489 INIT_LIST_HEAD(&khd->rqs[i]); 490 khd->domain_wait[i].sbq = NULL; 491 init_waitqueue_func_entry(&khd->domain_wait[i].wait, 492 kyber_domain_wake); 493 khd->domain_wait[i].wait.private = hctx; 494 INIT_LIST_HEAD(&khd->domain_wait[i].wait.entry); 495 atomic_set(&khd->wait_index[i], 0); 496 } 497 498 khd->cur_domain = 0; 499 khd->batching = 0; 500 501 hctx->sched_data = khd; 502 kyber_depth_updated(hctx); 503 504 return 0; 505 506err_kcqs: 507 kfree(khd->kcqs); 508err_khd: 509 kfree(khd); 510 return -ENOMEM; 511} 512 513static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx) 514{ 515 struct kyber_hctx_data *khd = hctx->sched_data; 516 int i; 517 518 for (i = 0; i < KYBER_NUM_DOMAINS; i++) 519 sbitmap_free(&khd->kcq_map[i]); 520 kfree(khd->kcqs); 521 kfree(hctx->sched_data); 522} 523 524static int rq_get_domain_token(struct request *rq) 525{ 526 return (long)rq->elv.priv[0]; 527} 528 529static void rq_set_domain_token(struct request *rq, int token) 530{ 531 rq->elv.priv[0] = (void *)(long)token; 532} 533 534static void rq_clear_domain_token(struct kyber_queue_data *kqd, 535 struct request *rq) 536{ 537 unsigned int sched_domain; 538 int nr; 539 540 nr = rq_get_domain_token(rq); 541 if (nr != -1) { 542 sched_domain = kyber_sched_domain(rq->cmd_flags); 543 sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr, 544 rq->mq_ctx->cpu); 545 } 546} 547 548static void kyber_limit_depth(unsigned int op, struct blk_mq_alloc_data *data) 549{ 550 /* 551 * We use the scheduler tags as per-hardware queue queueing tokens. 552 * Async requests can be limited at this stage. 553 */ 554 if (!op_is_sync(op)) { 555 struct kyber_queue_data *kqd = data->q->elevator->elevator_data; 556 557 data->shallow_depth = kqd->async_depth; 558 } 559} 560 561static bool kyber_bio_merge(struct request_queue *q, struct bio *bio, 562 unsigned int nr_segs) 563{ 564 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q); 565 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx); 566 struct kyber_hctx_data *khd = hctx->sched_data; 567 struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]]; 568 unsigned int sched_domain = kyber_sched_domain(bio->bi_opf); 569 struct list_head *rq_list = &kcq->rq_list[sched_domain]; 570 bool merged; 571 572 spin_lock(&kcq->lock); 573 merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs); 574 spin_unlock(&kcq->lock); 575 576 return merged; 577} 578 579static void kyber_prepare_request(struct request *rq) 580{ 581 rq_set_domain_token(rq, -1); 582} 583 584static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx, 585 struct list_head *rq_list, bool at_head) 586{ 587 struct kyber_hctx_data *khd = hctx->sched_data; 588 struct request *rq, *next; 589 590 list_for_each_entry_safe(rq, next, rq_list, queuelist) { 591 unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags); 592 struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]]; 593 struct list_head *head = &kcq->rq_list[sched_domain]; 594 595 spin_lock(&kcq->lock); 596 if (at_head) 597 list_move(&rq->queuelist, head); 598 else 599 list_move_tail(&rq->queuelist, head); 600 sbitmap_set_bit(&khd->kcq_map[sched_domain], 601 rq->mq_ctx->index_hw[hctx->type]); 602 blk_mq_sched_request_inserted(rq); 603 spin_unlock(&kcq->lock); 604 } 605} 606 607static void kyber_finish_request(struct request *rq) 608{ 609 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; 610 611 rq_clear_domain_token(kqd, rq); 612} 613 614static void add_latency_sample(struct kyber_cpu_latency *cpu_latency, 615 unsigned int sched_domain, unsigned int type, 616 u64 target, u64 latency) 617{ 618 unsigned int bucket; 619 u64 divisor; 620 621 if (latency > 0) { 622 divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1); 623 bucket = min_t(unsigned int, div64_u64(latency - 1, divisor), 624 KYBER_LATENCY_BUCKETS - 1); 625 } else { 626 bucket = 0; 627 } 628 629 atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]); 630} 631 632static void kyber_completed_request(struct request *rq, u64 now) 633{ 634 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; 635 struct kyber_cpu_latency *cpu_latency; 636 unsigned int sched_domain; 637 u64 target; 638 639 sched_domain = kyber_sched_domain(rq->cmd_flags); 640 if (sched_domain == KYBER_OTHER) 641 return; 642 643 cpu_latency = get_cpu_ptr(kqd->cpu_latency); 644 target = kqd->latency_targets[sched_domain]; 645 add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY, 646 target, now - rq->start_time_ns); 647 add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target, 648 now - rq->io_start_time_ns); 649 put_cpu_ptr(kqd->cpu_latency); 650 651 timer_reduce(&kqd->timer, jiffies + HZ / 10); 652} 653 654struct flush_kcq_data { 655 struct kyber_hctx_data *khd; 656 unsigned int sched_domain; 657 struct list_head *list; 658}; 659 660static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data) 661{ 662 struct flush_kcq_data *flush_data = data; 663 struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr]; 664 665 spin_lock(&kcq->lock); 666 list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain], 667 flush_data->list); 668 sbitmap_clear_bit(sb, bitnr); 669 spin_unlock(&kcq->lock); 670 671 return true; 672} 673 674static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd, 675 unsigned int sched_domain, 676 struct list_head *list) 677{ 678 struct flush_kcq_data data = { 679 .khd = khd, 680 .sched_domain = sched_domain, 681 .list = list, 682 }; 683 684 sbitmap_for_each_set(&khd->kcq_map[sched_domain], 685 flush_busy_kcq, &data); 686} 687 688static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags, 689 void *key) 690{ 691 struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private); 692 struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait); 693 694 sbitmap_del_wait_queue(wait); 695 blk_mq_run_hw_queue(hctx, true); 696 return 1; 697} 698 699static int kyber_get_domain_token(struct kyber_queue_data *kqd, 700 struct kyber_hctx_data *khd, 701 struct blk_mq_hw_ctx *hctx) 702{ 703 unsigned int sched_domain = khd->cur_domain; 704 struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain]; 705 struct sbq_wait *wait = &khd->domain_wait[sched_domain]; 706 struct sbq_wait_state *ws; 707 int nr; 708 709 nr = __sbitmap_queue_get(domain_tokens); 710 711 /* 712 * If we failed to get a domain token, make sure the hardware queue is 713 * run when one becomes available. Note that this is serialized on 714 * khd->lock, but we still need to be careful about the waker. 715 */ 716 if (nr < 0 && list_empty_careful(&wait->wait.entry)) { 717 ws = sbq_wait_ptr(domain_tokens, 718 &khd->wait_index[sched_domain]); 719 khd->domain_ws[sched_domain] = ws; 720 sbitmap_add_wait_queue(domain_tokens, ws, wait); 721 722 /* 723 * Try again in case a token was freed before we got on the wait 724 * queue. 725 */ 726 nr = __sbitmap_queue_get(domain_tokens); 727 } 728 729 /* 730 * If we got a token while we were on the wait queue, remove ourselves 731 * from the wait queue to ensure that all wake ups make forward 732 * progress. It's possible that the waker already deleted the entry 733 * between the !list_empty_careful() check and us grabbing the lock, but 734 * list_del_init() is okay with that. 735 */ 736 if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) { 737 ws = khd->domain_ws[sched_domain]; 738 spin_lock_irq(&ws->wait.lock); 739 sbitmap_del_wait_queue(wait); 740 spin_unlock_irq(&ws->wait.lock); 741 } 742 743 return nr; 744} 745 746static struct request * 747kyber_dispatch_cur_domain(struct kyber_queue_data *kqd, 748 struct kyber_hctx_data *khd, 749 struct blk_mq_hw_ctx *hctx) 750{ 751 struct list_head *rqs; 752 struct request *rq; 753 int nr; 754 755 rqs = &khd->rqs[khd->cur_domain]; 756 757 /* 758 * If we already have a flushed request, then we just need to get a 759 * token for it. Otherwise, if there are pending requests in the kcqs, 760 * flush the kcqs, but only if we can get a token. If not, we should 761 * leave the requests in the kcqs so that they can be merged. Note that 762 * khd->lock serializes the flushes, so if we observed any bit set in 763 * the kcq_map, we will always get a request. 764 */ 765 rq = list_first_entry_or_null(rqs, struct request, queuelist); 766 if (rq) { 767 nr = kyber_get_domain_token(kqd, khd, hctx); 768 if (nr >= 0) { 769 khd->batching++; 770 rq_set_domain_token(rq, nr); 771 list_del_init(&rq->queuelist); 772 return rq; 773 } else { 774 trace_kyber_throttled(kqd->q, 775 kyber_domain_names[khd->cur_domain]); 776 } 777 } else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) { 778 nr = kyber_get_domain_token(kqd, khd, hctx); 779 if (nr >= 0) { 780 kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs); 781 rq = list_first_entry(rqs, struct request, queuelist); 782 khd->batching++; 783 rq_set_domain_token(rq, nr); 784 list_del_init(&rq->queuelist); 785 return rq; 786 } else { 787 trace_kyber_throttled(kqd->q, 788 kyber_domain_names[khd->cur_domain]); 789 } 790 } 791 792 /* There were either no pending requests or no tokens. */ 793 return NULL; 794} 795 796static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx) 797{ 798 struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; 799 struct kyber_hctx_data *khd = hctx->sched_data; 800 struct request *rq; 801 int i; 802 803 spin_lock(&khd->lock); 804 805 /* 806 * First, if we are still entitled to batch, try to dispatch a request 807 * from the batch. 808 */ 809 if (khd->batching < kyber_batch_size[khd->cur_domain]) { 810 rq = kyber_dispatch_cur_domain(kqd, khd, hctx); 811 if (rq) 812 goto out; 813 } 814 815 /* 816 * Either, 817 * 1. We were no longer entitled to a batch. 818 * 2. The domain we were batching didn't have any requests. 819 * 3. The domain we were batching was out of tokens. 820 * 821 * Start another batch. Note that this wraps back around to the original 822 * domain if no other domains have requests or tokens. 823 */ 824 khd->batching = 0; 825 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 826 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1) 827 khd->cur_domain = 0; 828 else 829 khd->cur_domain++; 830 831 rq = kyber_dispatch_cur_domain(kqd, khd, hctx); 832 if (rq) 833 goto out; 834 } 835 836 rq = NULL; 837out: 838 spin_unlock(&khd->lock); 839 return rq; 840} 841 842static bool kyber_has_work(struct blk_mq_hw_ctx *hctx) 843{ 844 struct kyber_hctx_data *khd = hctx->sched_data; 845 int i; 846 847 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 848 if (!list_empty_careful(&khd->rqs[i]) || 849 sbitmap_any_bit_set(&khd->kcq_map[i])) 850 return true; 851 } 852 853 return false; 854} 855 856#define KYBER_LAT_SHOW_STORE(domain, name) \ 857static ssize_t kyber_##name##_lat_show(struct elevator_queue *e, \ 858 char *page) \ 859{ \ 860 struct kyber_queue_data *kqd = e->elevator_data; \ 861 \ 862 return sprintf(page, "%llu\n", kqd->latency_targets[domain]); \ 863} \ 864 \ 865static ssize_t kyber_##name##_lat_store(struct elevator_queue *e, \ 866 const char *page, size_t count) \ 867{ \ 868 struct kyber_queue_data *kqd = e->elevator_data; \ 869 unsigned long long nsec; \ 870 int ret; \ 871 \ 872 ret = kstrtoull(page, 10, &nsec); \ 873 if (ret) \ 874 return ret; \ 875 \ 876 kqd->latency_targets[domain] = nsec; \ 877 \ 878 return count; \ 879} 880KYBER_LAT_SHOW_STORE(KYBER_READ, read); 881KYBER_LAT_SHOW_STORE(KYBER_WRITE, write); 882#undef KYBER_LAT_SHOW_STORE 883 884#define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store) 885static struct elv_fs_entry kyber_sched_attrs[] = { 886 KYBER_LAT_ATTR(read), 887 KYBER_LAT_ATTR(write), 888 __ATTR_NULL 889}; 890#undef KYBER_LAT_ATTR 891 892#ifdef CONFIG_BLK_DEBUG_FS 893#define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name) \ 894static int kyber_##name##_tokens_show(void *data, struct seq_file *m) \ 895{ \ 896 struct request_queue *q = data; \ 897 struct kyber_queue_data *kqd = q->elevator->elevator_data; \ 898 \ 899 sbitmap_queue_show(&kqd->domain_tokens[domain], m); \ 900 return 0; \ 901} \ 902 \ 903static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos) \ 904 __acquires(&khd->lock) \ 905{ \ 906 struct blk_mq_hw_ctx *hctx = m->private; \ 907 struct kyber_hctx_data *khd = hctx->sched_data; \ 908 \ 909 spin_lock(&khd->lock); \ 910 return seq_list_start(&khd->rqs[domain], *pos); \ 911} \ 912 \ 913static void *kyber_##name##_rqs_next(struct seq_file *m, void *v, \ 914 loff_t *pos) \ 915{ \ 916 struct blk_mq_hw_ctx *hctx = m->private; \ 917 struct kyber_hctx_data *khd = hctx->sched_data; \ 918 \ 919 return seq_list_next(v, &khd->rqs[domain], pos); \ 920} \ 921 \ 922static void kyber_##name##_rqs_stop(struct seq_file *m, void *v) \ 923 __releases(&khd->lock) \ 924{ \ 925 struct blk_mq_hw_ctx *hctx = m->private; \ 926 struct kyber_hctx_data *khd = hctx->sched_data; \ 927 \ 928 spin_unlock(&khd->lock); \ 929} \ 930 \ 931static const struct seq_operations kyber_##name##_rqs_seq_ops = { \ 932 .start = kyber_##name##_rqs_start, \ 933 .next = kyber_##name##_rqs_next, \ 934 .stop = kyber_##name##_rqs_stop, \ 935 .show = blk_mq_debugfs_rq_show, \ 936}; \ 937 \ 938static int kyber_##name##_waiting_show(void *data, struct seq_file *m) \ 939{ \ 940 struct blk_mq_hw_ctx *hctx = data; \ 941 struct kyber_hctx_data *khd = hctx->sched_data; \ 942 wait_queue_entry_t *wait = &khd->domain_wait[domain].wait; \ 943 \ 944 seq_printf(m, "%d\n", !list_empty_careful(&wait->entry)); \ 945 return 0; \ 946} 947KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read) 948KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write) 949KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard) 950KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other) 951#undef KYBER_DEBUGFS_DOMAIN_ATTRS 952 953static int kyber_async_depth_show(void *data, struct seq_file *m) 954{ 955 struct request_queue *q = data; 956 struct kyber_queue_data *kqd = q->elevator->elevator_data; 957 958 seq_printf(m, "%u\n", kqd->async_depth); 959 return 0; 960} 961 962static int kyber_cur_domain_show(void *data, struct seq_file *m) 963{ 964 struct blk_mq_hw_ctx *hctx = data; 965 struct kyber_hctx_data *khd = hctx->sched_data; 966 967 seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]); 968 return 0; 969} 970 971static int kyber_batching_show(void *data, struct seq_file *m) 972{ 973 struct blk_mq_hw_ctx *hctx = data; 974 struct kyber_hctx_data *khd = hctx->sched_data; 975 976 seq_printf(m, "%u\n", khd->batching); 977 return 0; 978} 979 980#define KYBER_QUEUE_DOMAIN_ATTRS(name) \ 981 {#name "_tokens", 0400, kyber_##name##_tokens_show} 982static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = { 983 KYBER_QUEUE_DOMAIN_ATTRS(read), 984 KYBER_QUEUE_DOMAIN_ATTRS(write), 985 KYBER_QUEUE_DOMAIN_ATTRS(discard), 986 KYBER_QUEUE_DOMAIN_ATTRS(other), 987 {"async_depth", 0400, kyber_async_depth_show}, 988 {}, 989}; 990#undef KYBER_QUEUE_DOMAIN_ATTRS 991 992#define KYBER_HCTX_DOMAIN_ATTRS(name) \ 993 {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops}, \ 994 {#name "_waiting", 0400, kyber_##name##_waiting_show} 995static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = { 996 KYBER_HCTX_DOMAIN_ATTRS(read), 997 KYBER_HCTX_DOMAIN_ATTRS(write), 998 KYBER_HCTX_DOMAIN_ATTRS(discard), 999 KYBER_HCTX_DOMAIN_ATTRS(other), 1000 {"cur_domain", 0400, kyber_cur_domain_show}, 1001 {"batching", 0400, kyber_batching_show}, 1002 {}, 1003}; 1004#undef KYBER_HCTX_DOMAIN_ATTRS 1005#endif 1006 1007static struct elevator_type kyber_sched = { 1008 .ops = { 1009 .init_sched = kyber_init_sched, 1010 .exit_sched = kyber_exit_sched, 1011 .init_hctx = kyber_init_hctx, 1012 .exit_hctx = kyber_exit_hctx, 1013 .limit_depth = kyber_limit_depth, 1014 .bio_merge = kyber_bio_merge, 1015 .prepare_request = kyber_prepare_request, 1016 .insert_requests = kyber_insert_requests, 1017 .finish_request = kyber_finish_request, 1018 .requeue_request = kyber_finish_request, 1019 .completed_request = kyber_completed_request, 1020 .dispatch_request = kyber_dispatch_request, 1021 .has_work = kyber_has_work, 1022 .depth_updated = kyber_depth_updated, 1023 }, 1024#ifdef CONFIG_BLK_DEBUG_FS 1025 .queue_debugfs_attrs = kyber_queue_debugfs_attrs, 1026 .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs, 1027#endif 1028 .elevator_attrs = kyber_sched_attrs, 1029 .elevator_name = "kyber", 1030 .elevator_owner = THIS_MODULE, 1031}; 1032 1033static int __init kyber_init(void) 1034{ 1035 return elv_register(&kyber_sched); 1036} 1037 1038static void __exit kyber_exit(void) 1039{ 1040 elv_unregister(&kyber_sched); 1041} 1042 1043module_init(kyber_init); 1044module_exit(kyber_exit); 1045 1046MODULE_AUTHOR("Omar Sandoval"); 1047MODULE_LICENSE("GPL"); 1048MODULE_DESCRIPTION("Kyber I/O scheduler"); 1049