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 blk_mq_hw_ctx *hctx, struct bio *bio, 566 unsigned int nr_segs) 567 { 568 struct kyber_hctx_data *khd = hctx->sched_data; 569 struct blk_mq_ctx *ctx = blk_mq_get_ctx(hctx->queue); 570 struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]]; 571 unsigned int sched_domain = kyber_sched_domain(bio->bi_opf); 572 struct list_head *rq_list = &kcq->rq_list[sched_domain]; 573 bool merged; 574 575 spin_lock(&kcq->lock); 576 merged = blk_mq_bio_list_merge(hctx->queue, rq_list, bio, nr_segs); 577 spin_unlock(&kcq->lock); 578 579 return merged; 580 } 581 582 static void kyber_prepare_request(struct request *rq) 583 { 584 rq_set_domain_token(rq, -1); 585 } 586 587 static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx, 588 struct list_head *rq_list, bool at_head) 589 { 590 struct kyber_hctx_data *khd = hctx->sched_data; 591 struct request *rq, *next; 592 593 list_for_each_entry_safe(rq, next, rq_list, queuelist) { 594 unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags); 595 struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]]; 596 struct list_head *head = &kcq->rq_list[sched_domain]; 597 598 spin_lock(&kcq->lock); 599 if (at_head) 600 list_move(&rq->queuelist, head); 601 else 602 list_move_tail(&rq->queuelist, head); 603 sbitmap_set_bit(&khd->kcq_map[sched_domain], 604 rq->mq_ctx->index_hw[hctx->type]); 605 blk_mq_sched_request_inserted(rq); 606 spin_unlock(&kcq->lock); 607 } 608 } 609 610 static void kyber_finish_request(struct request *rq) 611 { 612 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; 613 614 rq_clear_domain_token(kqd, rq); 615 } 616 617 static void add_latency_sample(struct kyber_cpu_latency *cpu_latency, 618 unsigned int sched_domain, unsigned int type, 619 u64 target, u64 latency) 620 { 621 unsigned int bucket; 622 u64 divisor; 623 624 if (latency > 0) { 625 divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1); 626 bucket = min_t(unsigned int, div64_u64(latency - 1, divisor), 627 KYBER_LATENCY_BUCKETS - 1); 628 } else { 629 bucket = 0; 630 } 631 632 atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]); 633 } 634 635 static void kyber_completed_request(struct request *rq, u64 now) 636 { 637 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data; 638 struct kyber_cpu_latency *cpu_latency; 639 unsigned int sched_domain; 640 u64 target; 641 642 sched_domain = kyber_sched_domain(rq->cmd_flags); 643 if (sched_domain == KYBER_OTHER) 644 return; 645 646 cpu_latency = get_cpu_ptr(kqd->cpu_latency); 647 target = kqd->latency_targets[sched_domain]; 648 add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY, 649 target, now - rq->start_time_ns); 650 add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target, 651 now - rq->io_start_time_ns); 652 put_cpu_ptr(kqd->cpu_latency); 653 654 timer_reduce(&kqd->timer, jiffies + HZ / 10); 655 } 656 657 struct flush_kcq_data { 658 struct kyber_hctx_data *khd; 659 unsigned int sched_domain; 660 struct list_head *list; 661 }; 662 663 static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data) 664 { 665 struct flush_kcq_data *flush_data = data; 666 struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr]; 667 668 spin_lock(&kcq->lock); 669 list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain], 670 flush_data->list); 671 sbitmap_clear_bit(sb, bitnr); 672 spin_unlock(&kcq->lock); 673 674 return true; 675 } 676 677 static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd, 678 unsigned int sched_domain, 679 struct list_head *list) 680 { 681 struct flush_kcq_data data = { 682 .khd = khd, 683 .sched_domain = sched_domain, 684 .list = list, 685 }; 686 687 sbitmap_for_each_set(&khd->kcq_map[sched_domain], 688 flush_busy_kcq, &data); 689 } 690 691 static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags, 692 void *key) 693 { 694 struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private); 695 struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait); 696 697 sbitmap_del_wait_queue(wait); 698 blk_mq_run_hw_queue(hctx, true); 699 return 1; 700 } 701 702 static int kyber_get_domain_token(struct kyber_queue_data *kqd, 703 struct kyber_hctx_data *khd, 704 struct blk_mq_hw_ctx *hctx) 705 { 706 unsigned int sched_domain = khd->cur_domain; 707 struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain]; 708 struct sbq_wait *wait = &khd->domain_wait[sched_domain]; 709 struct sbq_wait_state *ws; 710 int nr; 711 712 nr = __sbitmap_queue_get(domain_tokens); 713 714 /* 715 * If we failed to get a domain token, make sure the hardware queue is 716 * run when one becomes available. Note that this is serialized on 717 * khd->lock, but we still need to be careful about the waker. 718 */ 719 if (nr < 0 && list_empty_careful(&wait->wait.entry)) { 720 ws = sbq_wait_ptr(domain_tokens, 721 &khd->wait_index[sched_domain]); 722 khd->domain_ws[sched_domain] = ws; 723 sbitmap_add_wait_queue(domain_tokens, ws, wait); 724 725 /* 726 * Try again in case a token was freed before we got on the wait 727 * queue. 728 */ 729 nr = __sbitmap_queue_get(domain_tokens); 730 } 731 732 /* 733 * If we got a token while we were on the wait queue, remove ourselves 734 * from the wait queue to ensure that all wake ups make forward 735 * progress. It's possible that the waker already deleted the entry 736 * between the !list_empty_careful() check and us grabbing the lock, but 737 * list_del_init() is okay with that. 738 */ 739 if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) { 740 ws = khd->domain_ws[sched_domain]; 741 spin_lock_irq(&ws->wait.lock); 742 sbitmap_del_wait_queue(wait); 743 spin_unlock_irq(&ws->wait.lock); 744 } 745 746 return nr; 747 } 748 749 static struct request * 750 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd, 751 struct kyber_hctx_data *khd, 752 struct blk_mq_hw_ctx *hctx) 753 { 754 struct list_head *rqs; 755 struct request *rq; 756 int nr; 757 758 rqs = &khd->rqs[khd->cur_domain]; 759 760 /* 761 * If we already have a flushed request, then we just need to get a 762 * token for it. Otherwise, if there are pending requests in the kcqs, 763 * flush the kcqs, but only if we can get a token. If not, we should 764 * leave the requests in the kcqs so that they can be merged. Note that 765 * khd->lock serializes the flushes, so if we observed any bit set in 766 * the kcq_map, we will always get a request. 767 */ 768 rq = list_first_entry_or_null(rqs, struct request, queuelist); 769 if (rq) { 770 nr = kyber_get_domain_token(kqd, khd, hctx); 771 if (nr >= 0) { 772 khd->batching++; 773 rq_set_domain_token(rq, nr); 774 list_del_init(&rq->queuelist); 775 return rq; 776 } else { 777 trace_kyber_throttled(kqd->q, 778 kyber_domain_names[khd->cur_domain]); 779 } 780 } else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) { 781 nr = kyber_get_domain_token(kqd, khd, hctx); 782 if (nr >= 0) { 783 kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs); 784 rq = list_first_entry(rqs, struct request, queuelist); 785 khd->batching++; 786 rq_set_domain_token(rq, nr); 787 list_del_init(&rq->queuelist); 788 return rq; 789 } else { 790 trace_kyber_throttled(kqd->q, 791 kyber_domain_names[khd->cur_domain]); 792 } 793 } 794 795 /* There were either no pending requests or no tokens. */ 796 return NULL; 797 } 798 799 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx) 800 { 801 struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data; 802 struct kyber_hctx_data *khd = hctx->sched_data; 803 struct request *rq; 804 int i; 805 806 spin_lock(&khd->lock); 807 808 /* 809 * First, if we are still entitled to batch, try to dispatch a request 810 * from the batch. 811 */ 812 if (khd->batching < kyber_batch_size[khd->cur_domain]) { 813 rq = kyber_dispatch_cur_domain(kqd, khd, hctx); 814 if (rq) 815 goto out; 816 } 817 818 /* 819 * Either, 820 * 1. We were no longer entitled to a batch. 821 * 2. The domain we were batching didn't have any requests. 822 * 3. The domain we were batching was out of tokens. 823 * 824 * Start another batch. Note that this wraps back around to the original 825 * domain if no other domains have requests or tokens. 826 */ 827 khd->batching = 0; 828 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 829 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1) 830 khd->cur_domain = 0; 831 else 832 khd->cur_domain++; 833 834 rq = kyber_dispatch_cur_domain(kqd, khd, hctx); 835 if (rq) 836 goto out; 837 } 838 839 rq = NULL; 840 out: 841 spin_unlock(&khd->lock); 842 return rq; 843 } 844 845 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx) 846 { 847 struct kyber_hctx_data *khd = hctx->sched_data; 848 int i; 849 850 for (i = 0; i < KYBER_NUM_DOMAINS; i++) { 851 if (!list_empty_careful(&khd->rqs[i]) || 852 sbitmap_any_bit_set(&khd->kcq_map[i])) 853 return true; 854 } 855 856 return false; 857 } 858 859 #define KYBER_LAT_SHOW_STORE(domain, name) \ 860 static ssize_t kyber_##name##_lat_show(struct elevator_queue *e, \ 861 char *page) \ 862 { \ 863 struct kyber_queue_data *kqd = e->elevator_data; \ 864 \ 865 return sprintf(page, "%llu\n", kqd->latency_targets[domain]); \ 866 } \ 867 \ 868 static ssize_t kyber_##name##_lat_store(struct elevator_queue *e, \ 869 const char *page, size_t count) \ 870 { \ 871 struct kyber_queue_data *kqd = e->elevator_data; \ 872 unsigned long long nsec; \ 873 int ret; \ 874 \ 875 ret = kstrtoull(page, 10, &nsec); \ 876 if (ret) \ 877 return ret; \ 878 \ 879 kqd->latency_targets[domain] = nsec; \ 880 \ 881 return count; \ 882 } 883 KYBER_LAT_SHOW_STORE(KYBER_READ, read); 884 KYBER_LAT_SHOW_STORE(KYBER_WRITE, write); 885 #undef KYBER_LAT_SHOW_STORE 886 887 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store) 888 static struct elv_fs_entry kyber_sched_attrs[] = { 889 KYBER_LAT_ATTR(read), 890 KYBER_LAT_ATTR(write), 891 __ATTR_NULL 892 }; 893 #undef KYBER_LAT_ATTR 894 895 #ifdef CONFIG_BLK_DEBUG_FS 896 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name) \ 897 static int kyber_##name##_tokens_show(void *data, struct seq_file *m) \ 898 { \ 899 struct request_queue *q = data; \ 900 struct kyber_queue_data *kqd = q->elevator->elevator_data; \ 901 \ 902 sbitmap_queue_show(&kqd->domain_tokens[domain], m); \ 903 return 0; \ 904 } \ 905 \ 906 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos) \ 907 __acquires(&khd->lock) \ 908 { \ 909 struct blk_mq_hw_ctx *hctx = m->private; \ 910 struct kyber_hctx_data *khd = hctx->sched_data; \ 911 \ 912 spin_lock(&khd->lock); \ 913 return seq_list_start(&khd->rqs[domain], *pos); \ 914 } \ 915 \ 916 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v, \ 917 loff_t *pos) \ 918 { \ 919 struct blk_mq_hw_ctx *hctx = m->private; \ 920 struct kyber_hctx_data *khd = hctx->sched_data; \ 921 \ 922 return seq_list_next(v, &khd->rqs[domain], pos); \ 923 } \ 924 \ 925 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v) \ 926 __releases(&khd->lock) \ 927 { \ 928 struct blk_mq_hw_ctx *hctx = m->private; \ 929 struct kyber_hctx_data *khd = hctx->sched_data; \ 930 \ 931 spin_unlock(&khd->lock); \ 932 } \ 933 \ 934 static const struct seq_operations kyber_##name##_rqs_seq_ops = { \ 935 .start = kyber_##name##_rqs_start, \ 936 .next = kyber_##name##_rqs_next, \ 937 .stop = kyber_##name##_rqs_stop, \ 938 .show = blk_mq_debugfs_rq_show, \ 939 }; \ 940 \ 941 static int kyber_##name##_waiting_show(void *data, struct seq_file *m) \ 942 { \ 943 struct blk_mq_hw_ctx *hctx = data; \ 944 struct kyber_hctx_data *khd = hctx->sched_data; \ 945 wait_queue_entry_t *wait = &khd->domain_wait[domain].wait; \ 946 \ 947 seq_printf(m, "%d\n", !list_empty_careful(&wait->entry)); \ 948 return 0; \ 949 } 950 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read) 951 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write) 952 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard) 953 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other) 954 #undef KYBER_DEBUGFS_DOMAIN_ATTRS 955 956 static int kyber_async_depth_show(void *data, struct seq_file *m) 957 { 958 struct request_queue *q = data; 959 struct kyber_queue_data *kqd = q->elevator->elevator_data; 960 961 seq_printf(m, "%u\n", kqd->async_depth); 962 return 0; 963 } 964 965 static int kyber_cur_domain_show(void *data, struct seq_file *m) 966 { 967 struct blk_mq_hw_ctx *hctx = data; 968 struct kyber_hctx_data *khd = hctx->sched_data; 969 970 seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]); 971 return 0; 972 } 973 974 static int kyber_batching_show(void *data, struct seq_file *m) 975 { 976 struct blk_mq_hw_ctx *hctx = data; 977 struct kyber_hctx_data *khd = hctx->sched_data; 978 979 seq_printf(m, "%u\n", khd->batching); 980 return 0; 981 } 982 983 #define KYBER_QUEUE_DOMAIN_ATTRS(name) \ 984 {#name "_tokens", 0400, kyber_##name##_tokens_show} 985 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = { 986 KYBER_QUEUE_DOMAIN_ATTRS(read), 987 KYBER_QUEUE_DOMAIN_ATTRS(write), 988 KYBER_QUEUE_DOMAIN_ATTRS(discard), 989 KYBER_QUEUE_DOMAIN_ATTRS(other), 990 {"async_depth", 0400, kyber_async_depth_show}, 991 {}, 992 }; 993 #undef KYBER_QUEUE_DOMAIN_ATTRS 994 995 #define KYBER_HCTX_DOMAIN_ATTRS(name) \ 996 {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops}, \ 997 {#name "_waiting", 0400, kyber_##name##_waiting_show} 998 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = { 999 KYBER_HCTX_DOMAIN_ATTRS(read), 1000 KYBER_HCTX_DOMAIN_ATTRS(write), 1001 KYBER_HCTX_DOMAIN_ATTRS(discard), 1002 KYBER_HCTX_DOMAIN_ATTRS(other), 1003 {"cur_domain", 0400, kyber_cur_domain_show}, 1004 {"batching", 0400, kyber_batching_show}, 1005 {}, 1006 }; 1007 #undef KYBER_HCTX_DOMAIN_ATTRS 1008 #endif 1009 1010 static struct elevator_type kyber_sched = { 1011 .ops = { 1012 .init_sched = kyber_init_sched, 1013 .exit_sched = kyber_exit_sched, 1014 .init_hctx = kyber_init_hctx, 1015 .exit_hctx = kyber_exit_hctx, 1016 .limit_depth = kyber_limit_depth, 1017 .bio_merge = kyber_bio_merge, 1018 .prepare_request = kyber_prepare_request, 1019 .insert_requests = kyber_insert_requests, 1020 .finish_request = kyber_finish_request, 1021 .requeue_request = kyber_finish_request, 1022 .completed_request = kyber_completed_request, 1023 .dispatch_request = kyber_dispatch_request, 1024 .has_work = kyber_has_work, 1025 }, 1026 #ifdef CONFIG_BLK_DEBUG_FS 1027 .queue_debugfs_attrs = kyber_queue_debugfs_attrs, 1028 .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs, 1029 #endif 1030 .elevator_attrs = kyber_sched_attrs, 1031 .elevator_name = "kyber", 1032 .elevator_owner = THIS_MODULE, 1033 }; 1034 1035 static int __init kyber_init(void) 1036 { 1037 return elv_register(&kyber_sched); 1038 } 1039 1040 static void __exit kyber_exit(void) 1041 { 1042 elv_unregister(&kyber_sched); 1043 } 1044 1045 module_init(kyber_init); 1046 module_exit(kyber_exit); 1047 1048 MODULE_AUTHOR("Omar Sandoval"); 1049 MODULE_LICENSE("GPL"); 1050 MODULE_DESCRIPTION("Kyber I/O scheduler"); 1051