Loading...
1/*
2 * The Kyber I/O scheduler. Controls latency by throttling queue depths using
3 * scalable techniques.
4 *
5 * Copyright (C) 2017 Facebook
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public
9 * License v2 as published by the Free Software Foundation.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program. If not, see <https://www.gnu.org/licenses/>.
18 */
19
20#include <linux/kernel.h>
21#include <linux/blkdev.h>
22#include <linux/blk-mq.h>
23#include <linux/elevator.h>
24#include <linux/module.h>
25#include <linux/sbitmap.h>
26
27#include "blk.h"
28#include "blk-mq.h"
29#include "blk-mq-debugfs.h"
30#include "blk-mq-sched.h"
31#include "blk-mq-tag.h"
32#include "blk-stat.h"
33
34/* Scheduling domains. */
35enum {
36 KYBER_READ,
37 KYBER_SYNC_WRITE,
38 KYBER_OTHER, /* Async writes, discard, etc. */
39 KYBER_NUM_DOMAINS,
40};
41
42enum {
43 KYBER_MIN_DEPTH = 256,
44
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 * Initial device-wide depths for each scheduling domain.
55 *
56 * Even for fast devices with lots of tags like NVMe, you can saturate
57 * the device with only a fraction of the maximum possible queue depth.
58 * So, we cap these to a reasonable value.
59 */
60static const unsigned int kyber_depth[] = {
61 [KYBER_READ] = 256,
62 [KYBER_SYNC_WRITE] = 128,
63 [KYBER_OTHER] = 64,
64};
65
66/*
67 * Scheduling domain batch sizes. We favor reads.
68 */
69static const unsigned int kyber_batch_size[] = {
70 [KYBER_READ] = 16,
71 [KYBER_SYNC_WRITE] = 8,
72 [KYBER_OTHER] = 8,
73};
74
75struct kyber_queue_data {
76 struct request_queue *q;
77
78 struct blk_stat_callback *cb;
79
80 /*
81 * The device is divided into multiple scheduling domains based on the
82 * request type. Each domain has a fixed number of in-flight requests of
83 * that type device-wide, limited by these tokens.
84 */
85 struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
86
87 /*
88 * Async request percentage, converted to per-word depth for
89 * sbitmap_get_shallow().
90 */
91 unsigned int async_depth;
92
93 /* Target latencies in nanoseconds. */
94 u64 read_lat_nsec, write_lat_nsec;
95};
96
97struct kyber_hctx_data {
98 spinlock_t lock;
99 struct list_head rqs[KYBER_NUM_DOMAINS];
100 unsigned int cur_domain;
101 unsigned int batching;
102 wait_queue_entry_t domain_wait[KYBER_NUM_DOMAINS];
103 struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
104 atomic_t wait_index[KYBER_NUM_DOMAINS];
105};
106
107static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
108 void *key);
109
110static int rq_sched_domain(const struct request *rq)
111{
112 unsigned int op = rq->cmd_flags;
113
114 if ((op & REQ_OP_MASK) == REQ_OP_READ)
115 return KYBER_READ;
116 else if ((op & REQ_OP_MASK) == REQ_OP_WRITE && op_is_sync(op))
117 return KYBER_SYNC_WRITE;
118 else
119 return KYBER_OTHER;
120}
121
122enum {
123 NONE = 0,
124 GOOD = 1,
125 GREAT = 2,
126 BAD = -1,
127 AWFUL = -2,
128};
129
130#define IS_GOOD(status) ((status) > 0)
131#define IS_BAD(status) ((status) < 0)
132
133static int kyber_lat_status(struct blk_stat_callback *cb,
134 unsigned int sched_domain, u64 target)
135{
136 u64 latency;
137
138 if (!cb->stat[sched_domain].nr_samples)
139 return NONE;
140
141 latency = cb->stat[sched_domain].mean;
142 if (latency >= 2 * target)
143 return AWFUL;
144 else if (latency > target)
145 return BAD;
146 else if (latency <= target / 2)
147 return GREAT;
148 else /* (latency <= target) */
149 return GOOD;
150}
151
152/*
153 * Adjust the read or synchronous write depth given the status of reads and
154 * writes. The goal is that the latencies of the two domains are fair (i.e., if
155 * one is good, then the other is good).
156 */
157static void kyber_adjust_rw_depth(struct kyber_queue_data *kqd,
158 unsigned int sched_domain, int this_status,
159 int other_status)
160{
161 unsigned int orig_depth, depth;
162
163 /*
164 * If this domain had no samples, or reads and writes are both good or
165 * both bad, don't adjust the depth.
166 */
167 if (this_status == NONE ||
168 (IS_GOOD(this_status) && IS_GOOD(other_status)) ||
169 (IS_BAD(this_status) && IS_BAD(other_status)))
170 return;
171
172 orig_depth = depth = kqd->domain_tokens[sched_domain].sb.depth;
173
174 if (other_status == NONE) {
175 depth++;
176 } else {
177 switch (this_status) {
178 case GOOD:
179 if (other_status == AWFUL)
180 depth -= max(depth / 4, 1U);
181 else
182 depth -= max(depth / 8, 1U);
183 break;
184 case GREAT:
185 if (other_status == AWFUL)
186 depth /= 2;
187 else
188 depth -= max(depth / 4, 1U);
189 break;
190 case BAD:
191 depth++;
192 break;
193 case AWFUL:
194 if (other_status == GREAT)
195 depth += 2;
196 else
197 depth++;
198 break;
199 }
200 }
201
202 depth = clamp(depth, 1U, kyber_depth[sched_domain]);
203 if (depth != orig_depth)
204 sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
205}
206
207/*
208 * Adjust the depth of other requests given the status of reads and synchronous
209 * writes. As long as either domain is doing fine, we don't throttle, but if
210 * both domains are doing badly, we throttle heavily.
211 */
212static void kyber_adjust_other_depth(struct kyber_queue_data *kqd,
213 int read_status, int write_status,
214 bool have_samples)
215{
216 unsigned int orig_depth, depth;
217 int status;
218
219 orig_depth = depth = kqd->domain_tokens[KYBER_OTHER].sb.depth;
220
221 if (read_status == NONE && write_status == NONE) {
222 depth += 2;
223 } else if (have_samples) {
224 if (read_status == NONE)
225 status = write_status;
226 else if (write_status == NONE)
227 status = read_status;
228 else
229 status = max(read_status, write_status);
230 switch (status) {
231 case GREAT:
232 depth += 2;
233 break;
234 case GOOD:
235 depth++;
236 break;
237 case BAD:
238 depth -= max(depth / 4, 1U);
239 break;
240 case AWFUL:
241 depth /= 2;
242 break;
243 }
244 }
245
246 depth = clamp(depth, 1U, kyber_depth[KYBER_OTHER]);
247 if (depth != orig_depth)
248 sbitmap_queue_resize(&kqd->domain_tokens[KYBER_OTHER], depth);
249}
250
251/*
252 * Apply heuristics for limiting queue depths based on gathered latency
253 * statistics.
254 */
255static void kyber_stat_timer_fn(struct blk_stat_callback *cb)
256{
257 struct kyber_queue_data *kqd = cb->data;
258 int read_status, write_status;
259
260 read_status = kyber_lat_status(cb, KYBER_READ, kqd->read_lat_nsec);
261 write_status = kyber_lat_status(cb, KYBER_SYNC_WRITE, kqd->write_lat_nsec);
262
263 kyber_adjust_rw_depth(kqd, KYBER_READ, read_status, write_status);
264 kyber_adjust_rw_depth(kqd, KYBER_SYNC_WRITE, write_status, read_status);
265 kyber_adjust_other_depth(kqd, read_status, write_status,
266 cb->stat[KYBER_OTHER].nr_samples != 0);
267
268 /*
269 * Continue monitoring latencies if we aren't hitting the targets or
270 * we're still throttling other requests.
271 */
272 if (!blk_stat_is_active(kqd->cb) &&
273 ((IS_BAD(read_status) || IS_BAD(write_status) ||
274 kqd->domain_tokens[KYBER_OTHER].sb.depth < kyber_depth[KYBER_OTHER])))
275 blk_stat_activate_msecs(kqd->cb, 100);
276}
277
278static unsigned int kyber_sched_tags_shift(struct kyber_queue_data *kqd)
279{
280 /*
281 * All of the hardware queues have the same depth, so we can just grab
282 * the shift of the first one.
283 */
284 return kqd->q->queue_hw_ctx[0]->sched_tags->bitmap_tags.sb.shift;
285}
286
287static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
288{
289 struct kyber_queue_data *kqd;
290 unsigned int max_tokens;
291 unsigned int shift;
292 int ret = -ENOMEM;
293 int i;
294
295 kqd = kmalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
296 if (!kqd)
297 goto err;
298 kqd->q = q;
299
300 kqd->cb = blk_stat_alloc_callback(kyber_stat_timer_fn, rq_sched_domain,
301 KYBER_NUM_DOMAINS, kqd);
302 if (!kqd->cb)
303 goto err_kqd;
304
305 /*
306 * The maximum number of tokens for any scheduling domain is at least
307 * the queue depth of a single hardware queue. If the hardware doesn't
308 * have many tags, still provide a reasonable number.
309 */
310 max_tokens = max_t(unsigned int, q->tag_set->queue_depth,
311 KYBER_MIN_DEPTH);
312 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
313 WARN_ON(!kyber_depth[i]);
314 WARN_ON(!kyber_batch_size[i]);
315 ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
316 max_tokens, -1, false, GFP_KERNEL,
317 q->node);
318 if (ret) {
319 while (--i >= 0)
320 sbitmap_queue_free(&kqd->domain_tokens[i]);
321 goto err_cb;
322 }
323 sbitmap_queue_resize(&kqd->domain_tokens[i], kyber_depth[i]);
324 }
325
326 shift = kyber_sched_tags_shift(kqd);
327 kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
328
329 kqd->read_lat_nsec = 2000000ULL;
330 kqd->write_lat_nsec = 10000000ULL;
331
332 return kqd;
333
334err_cb:
335 blk_stat_free_callback(kqd->cb);
336err_kqd:
337 kfree(kqd);
338err:
339 return ERR_PTR(ret);
340}
341
342static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
343{
344 struct kyber_queue_data *kqd;
345 struct elevator_queue *eq;
346
347 eq = elevator_alloc(q, e);
348 if (!eq)
349 return -ENOMEM;
350
351 kqd = kyber_queue_data_alloc(q);
352 if (IS_ERR(kqd)) {
353 kobject_put(&eq->kobj);
354 return PTR_ERR(kqd);
355 }
356
357 eq->elevator_data = kqd;
358 q->elevator = eq;
359
360 blk_stat_add_callback(q, kqd->cb);
361
362 return 0;
363}
364
365static void kyber_exit_sched(struct elevator_queue *e)
366{
367 struct kyber_queue_data *kqd = e->elevator_data;
368 struct request_queue *q = kqd->q;
369 int i;
370
371 blk_stat_remove_callback(q, kqd->cb);
372
373 for (i = 0; i < KYBER_NUM_DOMAINS; i++)
374 sbitmap_queue_free(&kqd->domain_tokens[i]);
375 blk_stat_free_callback(kqd->cb);
376 kfree(kqd);
377}
378
379static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
380{
381 struct kyber_hctx_data *khd;
382 int i;
383
384 khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
385 if (!khd)
386 return -ENOMEM;
387
388 spin_lock_init(&khd->lock);
389
390 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
391 INIT_LIST_HEAD(&khd->rqs[i]);
392 init_waitqueue_func_entry(&khd->domain_wait[i],
393 kyber_domain_wake);
394 khd->domain_wait[i].private = hctx;
395 INIT_LIST_HEAD(&khd->domain_wait[i].entry);
396 atomic_set(&khd->wait_index[i], 0);
397 }
398
399 khd->cur_domain = 0;
400 khd->batching = 0;
401
402 hctx->sched_data = khd;
403
404 return 0;
405}
406
407static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
408{
409 kfree(hctx->sched_data);
410}
411
412static int rq_get_domain_token(struct request *rq)
413{
414 return (long)rq->elv.priv[0];
415}
416
417static void rq_set_domain_token(struct request *rq, int token)
418{
419 rq->elv.priv[0] = (void *)(long)token;
420}
421
422static void rq_clear_domain_token(struct kyber_queue_data *kqd,
423 struct request *rq)
424{
425 unsigned int sched_domain;
426 int nr;
427
428 nr = rq_get_domain_token(rq);
429 if (nr != -1) {
430 sched_domain = rq_sched_domain(rq);
431 sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
432 rq->mq_ctx->cpu);
433 }
434}
435
436static void kyber_limit_depth(unsigned int op, struct blk_mq_alloc_data *data)
437{
438 /*
439 * We use the scheduler tags as per-hardware queue queueing tokens.
440 * Async requests can be limited at this stage.
441 */
442 if (!op_is_sync(op)) {
443 struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
444
445 data->shallow_depth = kqd->async_depth;
446 }
447}
448
449static void kyber_prepare_request(struct request *rq, struct bio *bio)
450{
451 rq_set_domain_token(rq, -1);
452}
453
454static void kyber_finish_request(struct request *rq)
455{
456 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
457
458 rq_clear_domain_token(kqd, rq);
459}
460
461static void kyber_completed_request(struct request *rq)
462{
463 struct request_queue *q = rq->q;
464 struct kyber_queue_data *kqd = q->elevator->elevator_data;
465 unsigned int sched_domain;
466 u64 now, latency, target;
467
468 /*
469 * Check if this request met our latency goal. If not, quickly gather
470 * some statistics and start throttling.
471 */
472 sched_domain = rq_sched_domain(rq);
473 switch (sched_domain) {
474 case KYBER_READ:
475 target = kqd->read_lat_nsec;
476 break;
477 case KYBER_SYNC_WRITE:
478 target = kqd->write_lat_nsec;
479 break;
480 default:
481 return;
482 }
483
484 /* If we are already monitoring latencies, don't check again. */
485 if (blk_stat_is_active(kqd->cb))
486 return;
487
488 now = __blk_stat_time(ktime_to_ns(ktime_get()));
489 if (now < blk_stat_time(&rq->issue_stat))
490 return;
491
492 latency = now - blk_stat_time(&rq->issue_stat);
493
494 if (latency > target)
495 blk_stat_activate_msecs(kqd->cb, 10);
496}
497
498static void kyber_flush_busy_ctxs(struct kyber_hctx_data *khd,
499 struct blk_mq_hw_ctx *hctx)
500{
501 LIST_HEAD(rq_list);
502 struct request *rq, *next;
503
504 blk_mq_flush_busy_ctxs(hctx, &rq_list);
505 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
506 unsigned int sched_domain;
507
508 sched_domain = rq_sched_domain(rq);
509 list_move_tail(&rq->queuelist, &khd->rqs[sched_domain]);
510 }
511}
512
513static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
514 void *key)
515{
516 struct blk_mq_hw_ctx *hctx = READ_ONCE(wait->private);
517
518 list_del_init(&wait->entry);
519 blk_mq_run_hw_queue(hctx, true);
520 return 1;
521}
522
523static int kyber_get_domain_token(struct kyber_queue_data *kqd,
524 struct kyber_hctx_data *khd,
525 struct blk_mq_hw_ctx *hctx)
526{
527 unsigned int sched_domain = khd->cur_domain;
528 struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
529 wait_queue_entry_t *wait = &khd->domain_wait[sched_domain];
530 struct sbq_wait_state *ws;
531 int nr;
532
533 nr = __sbitmap_queue_get(domain_tokens);
534
535 /*
536 * If we failed to get a domain token, make sure the hardware queue is
537 * run when one becomes available. Note that this is serialized on
538 * khd->lock, but we still need to be careful about the waker.
539 */
540 if (nr < 0 && list_empty_careful(&wait->entry)) {
541 ws = sbq_wait_ptr(domain_tokens,
542 &khd->wait_index[sched_domain]);
543 khd->domain_ws[sched_domain] = ws;
544 add_wait_queue(&ws->wait, wait);
545
546 /*
547 * Try again in case a token was freed before we got on the wait
548 * queue.
549 */
550 nr = __sbitmap_queue_get(domain_tokens);
551 }
552
553 /*
554 * If we got a token while we were on the wait queue, remove ourselves
555 * from the wait queue to ensure that all wake ups make forward
556 * progress. It's possible that the waker already deleted the entry
557 * between the !list_empty_careful() check and us grabbing the lock, but
558 * list_del_init() is okay with that.
559 */
560 if (nr >= 0 && !list_empty_careful(&wait->entry)) {
561 ws = khd->domain_ws[sched_domain];
562 spin_lock_irq(&ws->wait.lock);
563 list_del_init(&wait->entry);
564 spin_unlock_irq(&ws->wait.lock);
565 }
566
567 return nr;
568}
569
570static struct request *
571kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
572 struct kyber_hctx_data *khd,
573 struct blk_mq_hw_ctx *hctx,
574 bool *flushed)
575{
576 struct list_head *rqs;
577 struct request *rq;
578 int nr;
579
580 rqs = &khd->rqs[khd->cur_domain];
581 rq = list_first_entry_or_null(rqs, struct request, queuelist);
582
583 /*
584 * If there wasn't already a pending request and we haven't flushed the
585 * software queues yet, flush the software queues and check again.
586 */
587 if (!rq && !*flushed) {
588 kyber_flush_busy_ctxs(khd, hctx);
589 *flushed = true;
590 rq = list_first_entry_or_null(rqs, struct request, queuelist);
591 }
592
593 if (rq) {
594 nr = kyber_get_domain_token(kqd, khd, hctx);
595 if (nr >= 0) {
596 khd->batching++;
597 rq_set_domain_token(rq, nr);
598 list_del_init(&rq->queuelist);
599 return rq;
600 }
601 }
602
603 /* There were either no pending requests or no tokens. */
604 return NULL;
605}
606
607static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
608{
609 struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
610 struct kyber_hctx_data *khd = hctx->sched_data;
611 bool flushed = false;
612 struct request *rq;
613 int i;
614
615 spin_lock(&khd->lock);
616
617 /*
618 * First, if we are still entitled to batch, try to dispatch a request
619 * from the batch.
620 */
621 if (khd->batching < kyber_batch_size[khd->cur_domain]) {
622 rq = kyber_dispatch_cur_domain(kqd, khd, hctx, &flushed);
623 if (rq)
624 goto out;
625 }
626
627 /*
628 * Either,
629 * 1. We were no longer entitled to a batch.
630 * 2. The domain we were batching didn't have any requests.
631 * 3. The domain we were batching was out of tokens.
632 *
633 * Start another batch. Note that this wraps back around to the original
634 * domain if no other domains have requests or tokens.
635 */
636 khd->batching = 0;
637 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
638 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
639 khd->cur_domain = 0;
640 else
641 khd->cur_domain++;
642
643 rq = kyber_dispatch_cur_domain(kqd, khd, hctx, &flushed);
644 if (rq)
645 goto out;
646 }
647
648 rq = NULL;
649out:
650 spin_unlock(&khd->lock);
651 return rq;
652}
653
654static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
655{
656 struct kyber_hctx_data *khd = hctx->sched_data;
657 int i;
658
659 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
660 if (!list_empty_careful(&khd->rqs[i]))
661 return true;
662 }
663 return sbitmap_any_bit_set(&hctx->ctx_map);
664}
665
666#define KYBER_LAT_SHOW_STORE(op) \
667static ssize_t kyber_##op##_lat_show(struct elevator_queue *e, \
668 char *page) \
669{ \
670 struct kyber_queue_data *kqd = e->elevator_data; \
671 \
672 return sprintf(page, "%llu\n", kqd->op##_lat_nsec); \
673} \
674 \
675static ssize_t kyber_##op##_lat_store(struct elevator_queue *e, \
676 const char *page, size_t count) \
677{ \
678 struct kyber_queue_data *kqd = e->elevator_data; \
679 unsigned long long nsec; \
680 int ret; \
681 \
682 ret = kstrtoull(page, 10, &nsec); \
683 if (ret) \
684 return ret; \
685 \
686 kqd->op##_lat_nsec = nsec; \
687 \
688 return count; \
689}
690KYBER_LAT_SHOW_STORE(read);
691KYBER_LAT_SHOW_STORE(write);
692#undef KYBER_LAT_SHOW_STORE
693
694#define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
695static struct elv_fs_entry kyber_sched_attrs[] = {
696 KYBER_LAT_ATTR(read),
697 KYBER_LAT_ATTR(write),
698 __ATTR_NULL
699};
700#undef KYBER_LAT_ATTR
701
702#ifdef CONFIG_BLK_DEBUG_FS
703#define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name) \
704static int kyber_##name##_tokens_show(void *data, struct seq_file *m) \
705{ \
706 struct request_queue *q = data; \
707 struct kyber_queue_data *kqd = q->elevator->elevator_data; \
708 \
709 sbitmap_queue_show(&kqd->domain_tokens[domain], m); \
710 return 0; \
711} \
712 \
713static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos) \
714 __acquires(&khd->lock) \
715{ \
716 struct blk_mq_hw_ctx *hctx = m->private; \
717 struct kyber_hctx_data *khd = hctx->sched_data; \
718 \
719 spin_lock(&khd->lock); \
720 return seq_list_start(&khd->rqs[domain], *pos); \
721} \
722 \
723static void *kyber_##name##_rqs_next(struct seq_file *m, void *v, \
724 loff_t *pos) \
725{ \
726 struct blk_mq_hw_ctx *hctx = m->private; \
727 struct kyber_hctx_data *khd = hctx->sched_data; \
728 \
729 return seq_list_next(v, &khd->rqs[domain], pos); \
730} \
731 \
732static void kyber_##name##_rqs_stop(struct seq_file *m, void *v) \
733 __releases(&khd->lock) \
734{ \
735 struct blk_mq_hw_ctx *hctx = m->private; \
736 struct kyber_hctx_data *khd = hctx->sched_data; \
737 \
738 spin_unlock(&khd->lock); \
739} \
740 \
741static const struct seq_operations kyber_##name##_rqs_seq_ops = { \
742 .start = kyber_##name##_rqs_start, \
743 .next = kyber_##name##_rqs_next, \
744 .stop = kyber_##name##_rqs_stop, \
745 .show = blk_mq_debugfs_rq_show, \
746}; \
747 \
748static int kyber_##name##_waiting_show(void *data, struct seq_file *m) \
749{ \
750 struct blk_mq_hw_ctx *hctx = data; \
751 struct kyber_hctx_data *khd = hctx->sched_data; \
752 wait_queue_entry_t *wait = &khd->domain_wait[domain]; \
753 \
754 seq_printf(m, "%d\n", !list_empty_careful(&wait->entry)); \
755 return 0; \
756}
757KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
758KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_SYNC_WRITE, sync_write)
759KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
760#undef KYBER_DEBUGFS_DOMAIN_ATTRS
761
762static int kyber_async_depth_show(void *data, struct seq_file *m)
763{
764 struct request_queue *q = data;
765 struct kyber_queue_data *kqd = q->elevator->elevator_data;
766
767 seq_printf(m, "%u\n", kqd->async_depth);
768 return 0;
769}
770
771static int kyber_cur_domain_show(void *data, struct seq_file *m)
772{
773 struct blk_mq_hw_ctx *hctx = data;
774 struct kyber_hctx_data *khd = hctx->sched_data;
775
776 switch (khd->cur_domain) {
777 case KYBER_READ:
778 seq_puts(m, "READ\n");
779 break;
780 case KYBER_SYNC_WRITE:
781 seq_puts(m, "SYNC_WRITE\n");
782 break;
783 case KYBER_OTHER:
784 seq_puts(m, "OTHER\n");
785 break;
786 default:
787 seq_printf(m, "%u\n", khd->cur_domain);
788 break;
789 }
790 return 0;
791}
792
793static int kyber_batching_show(void *data, struct seq_file *m)
794{
795 struct blk_mq_hw_ctx *hctx = data;
796 struct kyber_hctx_data *khd = hctx->sched_data;
797
798 seq_printf(m, "%u\n", khd->batching);
799 return 0;
800}
801
802#define KYBER_QUEUE_DOMAIN_ATTRS(name) \
803 {#name "_tokens", 0400, kyber_##name##_tokens_show}
804static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
805 KYBER_QUEUE_DOMAIN_ATTRS(read),
806 KYBER_QUEUE_DOMAIN_ATTRS(sync_write),
807 KYBER_QUEUE_DOMAIN_ATTRS(other),
808 {"async_depth", 0400, kyber_async_depth_show},
809 {},
810};
811#undef KYBER_QUEUE_DOMAIN_ATTRS
812
813#define KYBER_HCTX_DOMAIN_ATTRS(name) \
814 {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops}, \
815 {#name "_waiting", 0400, kyber_##name##_waiting_show}
816static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
817 KYBER_HCTX_DOMAIN_ATTRS(read),
818 KYBER_HCTX_DOMAIN_ATTRS(sync_write),
819 KYBER_HCTX_DOMAIN_ATTRS(other),
820 {"cur_domain", 0400, kyber_cur_domain_show},
821 {"batching", 0400, kyber_batching_show},
822 {},
823};
824#undef KYBER_HCTX_DOMAIN_ATTRS
825#endif
826
827static struct elevator_type kyber_sched = {
828 .ops.mq = {
829 .init_sched = kyber_init_sched,
830 .exit_sched = kyber_exit_sched,
831 .init_hctx = kyber_init_hctx,
832 .exit_hctx = kyber_exit_hctx,
833 .limit_depth = kyber_limit_depth,
834 .prepare_request = kyber_prepare_request,
835 .finish_request = kyber_finish_request,
836 .requeue_request = kyber_finish_request,
837 .completed_request = kyber_completed_request,
838 .dispatch_request = kyber_dispatch_request,
839 .has_work = kyber_has_work,
840 },
841 .uses_mq = true,
842#ifdef CONFIG_BLK_DEBUG_FS
843 .queue_debugfs_attrs = kyber_queue_debugfs_attrs,
844 .hctx_debugfs_attrs = kyber_hctx_debugfs_attrs,
845#endif
846 .elevator_attrs = kyber_sched_attrs,
847 .elevator_name = "kyber",
848 .elevator_owner = THIS_MODULE,
849};
850
851static int __init kyber_init(void)
852{
853 return elv_register(&kyber_sched);
854}
855
856static void __exit kyber_exit(void)
857{
858 elv_unregister(&kyber_sched);
859}
860
861module_init(kyber_init);
862module_exit(kyber_exit);
863
864MODULE_AUTHOR("Omar Sandoval");
865MODULE_LICENSE("GPL");
866MODULE_DESCRIPTION("Kyber I/O scheduler");
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 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
365static 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
408err_buckets:
409 free_percpu(kqd->cpu_latency);
410err_kqd:
411 kfree(kqd);
412err:
413 return ERR_PTR(ret);
414}
415
416static 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
439static 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
452static 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
461static 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
510err_kcqs:
511 kfree(khd->kcqs);
512err_khd:
513 kfree(khd);
514 return -ENOMEM;
515}
516
517static 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
528static int rq_get_domain_token(struct request *rq)
529{
530 return (long)rq->elv.priv[0];
531}
532
533static void rq_set_domain_token(struct request *rq, int token)
534{
535 rq->elv.priv[0] = (void *)(long)token;
536}
537
538static 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
552static 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
565static 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
582static void kyber_prepare_request(struct request *rq)
583{
584 rq_set_domain_token(rq, -1);
585}
586
587static 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
610static 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
617static 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
635static 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
657struct flush_kcq_data {
658 struct kyber_hctx_data *khd;
659 unsigned int sched_domain;
660 struct list_head *list;
661};
662
663static 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
677static 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
691static 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
702static 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
749static struct request *
750kyber_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
799static 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;
840out:
841 spin_unlock(&khd->lock);
842 return rq;
843}
844
845static 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) \
860static 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 \
868static 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}
883KYBER_LAT_SHOW_STORE(KYBER_READ, read);
884KYBER_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)
888static 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) \
897static 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 \
906static 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 \
916static 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 \
925static 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 \
934static 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 \
941static 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}
950KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
951KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write)
952KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
953KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
954#undef KYBER_DEBUGFS_DOMAIN_ATTRS
955
956static 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
965static 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
974static 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}
985static 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}
998static 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
1010static 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
1035static int __init kyber_init(void)
1036{
1037 return elv_register(&kyber_sched);
1038}
1039
1040static void __exit kyber_exit(void)
1041{
1042 elv_unregister(&kyber_sched);
1043}
1044
1045module_init(kyber_init);
1046module_exit(kyber_exit);
1047
1048MODULE_AUTHOR("Omar Sandoval");
1049MODULE_LICENSE("GPL");
1050MODULE_DESCRIPTION("Kyber I/O scheduler");