1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * The Kyber I/O scheduler. Controls latency by throttling queue depths using
4 * scalable techniques.
5 *
6 * Copyright (C) 2017 Facebook
7 */
8
9 #include <linux/kernel.h>
10 #include <linux/blkdev.h>
11 #include <linux/module.h>
12 #include <linux/sbitmap.h>
13
14 #include <trace/events/block.h>
15
16 #include "elevator.h"
17 #include "blk.h"
18 #include "blk-mq.h"
19 #include "blk-mq-debugfs.h"
20 #include "blk-mq-sched.h"
21
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/kyber.h>
24
25 /*
26 * Scheduling domains: the device is divided into multiple domains based on the
27 * request type.
28 */
29 enum {
30 KYBER_READ,
31 KYBER_WRITE,
32 KYBER_DISCARD,
33 KYBER_OTHER,
34 KYBER_NUM_DOMAINS,
35 };
36
37 static const char *kyber_domain_names[] = {
38 [KYBER_READ] = "READ",
39 [KYBER_WRITE] = "WRITE",
40 [KYBER_DISCARD] = "DISCARD",
41 [KYBER_OTHER] = "OTHER",
42 };
43
44 enum {
45 /*
46 * In order to prevent starvation of synchronous requests by a flood of
47 * asynchronous requests, we reserve 25% of requests for synchronous
48 * operations.
49 */
50 KYBER_ASYNC_PERCENT = 75,
51 };
52
53 /*
54 * Maximum device-wide depth for each scheduling domain.
55 *
56 * Even for fast devices with lots of tags like NVMe, you can saturate the
57 * device with only a fraction of the maximum possible queue depth. So, we cap
58 * these to a reasonable value.
59 */
60 static const unsigned int kyber_depth[] = {
61 [KYBER_READ] = 256,
62 [KYBER_WRITE] = 128,
63 [KYBER_DISCARD] = 64,
64 [KYBER_OTHER] = 16,
65 };
66
67 /*
68 * Default latency targets for each scheduling domain.
69 */
70 static const u64 kyber_latency_targets[] = {
71 [KYBER_READ] = 2ULL * NSEC_PER_MSEC,
72 [KYBER_WRITE] = 10ULL * NSEC_PER_MSEC,
73 [KYBER_DISCARD] = 5ULL * NSEC_PER_SEC,
74 };
75
76 /*
77 * Batch size (number of requests we'll dispatch in a row) for each scheduling
78 * domain.
79 */
80 static const unsigned int kyber_batch_size[] = {
81 [KYBER_READ] = 16,
82 [KYBER_WRITE] = 8,
83 [KYBER_DISCARD] = 1,
84 [KYBER_OTHER] = 1,
85 };
86
87 /*
88 * Requests latencies are recorded in a histogram with buckets defined relative
89 * to the target latency:
90 *
91 * <= 1/4 * target latency
92 * <= 1/2 * target latency
93 * <= 3/4 * target latency
94 * <= target latency
95 * <= 1 1/4 * target latency
96 * <= 1 1/2 * target latency
97 * <= 1 3/4 * target latency
98 * > 1 3/4 * target latency
99 */
100 enum {
101 /*
102 * The width of the latency histogram buckets is
103 * 1 / (1 << KYBER_LATENCY_SHIFT) * target latency.
104 */
105 KYBER_LATENCY_SHIFT = 2,
106 /*
107 * The first (1 << KYBER_LATENCY_SHIFT) buckets are <= target latency,
108 * thus, "good".
109 */
110 KYBER_GOOD_BUCKETS = 1 << KYBER_LATENCY_SHIFT,
111 /* There are also (1 << KYBER_LATENCY_SHIFT) "bad" buckets. */
112 KYBER_LATENCY_BUCKETS = 2 << KYBER_LATENCY_SHIFT,
113 };
114
115 /*
116 * We measure both the total latency and the I/O latency (i.e., latency after
117 * submitting to the device).
118 */
119 enum {
120 KYBER_TOTAL_LATENCY,
121 KYBER_IO_LATENCY,
122 };
123
124 static const char *kyber_latency_type_names[] = {
125 [KYBER_TOTAL_LATENCY] = "total",
126 [KYBER_IO_LATENCY] = "I/O",
127 };
128
129 /*
130 * Per-cpu latency histograms: total latency and I/O latency for each scheduling
131 * domain except for KYBER_OTHER.
132 */
133 struct kyber_cpu_latency {
134 atomic_t buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
135 };
136
137 /*
138 * There is a same mapping between ctx & hctx and kcq & khd,
139 * we use request->mq_ctx->index_hw to index the kcq in khd.
140 */
141 struct kyber_ctx_queue {
142 /*
143 * Used to ensure operations on rq_list and kcq_map to be an atmoic one.
144 * Also protect the rqs on rq_list when merge.
145 */
146 spinlock_t lock;
147 struct list_head rq_list[KYBER_NUM_DOMAINS];
148 } ____cacheline_aligned_in_smp;
149
150 struct kyber_queue_data {
151 struct request_queue *q;
152 dev_t dev;
153
154 /*
155 * Each scheduling domain has a limited number of in-flight requests
156 * device-wide, limited by these tokens.
157 */
158 struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
159
160 /* Number of allowed async requests. */
161 unsigned int async_depth;
162
163 struct kyber_cpu_latency __percpu *cpu_latency;
164
165 /* Timer for stats aggregation and adjusting domain tokens. */
166 struct timer_list timer;
167
168 unsigned int latency_buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
169
170 unsigned long latency_timeout[KYBER_OTHER];
171
172 int domain_p99[KYBER_OTHER];
173
174 /* Target latencies in nanoseconds. */
175 u64 latency_targets[KYBER_OTHER];
176 };
177
178 struct kyber_hctx_data {
179 spinlock_t lock;
180 struct list_head rqs[KYBER_NUM_DOMAINS];
181 unsigned int cur_domain;
182 unsigned int batching;
183 struct kyber_ctx_queue *kcqs;
184 struct sbitmap kcq_map[KYBER_NUM_DOMAINS];
185 struct sbq_wait domain_wait[KYBER_NUM_DOMAINS];
186 struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
187 atomic_t wait_index[KYBER_NUM_DOMAINS];
188 };
189
190 static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
191 void *key);
192
kyber_sched_domain(blk_opf_t opf)193 static unsigned int kyber_sched_domain(blk_opf_t opf)
194 {
195 switch (opf & REQ_OP_MASK) {
196 case REQ_OP_READ:
197 return KYBER_READ;
198 case REQ_OP_WRITE:
199 return KYBER_WRITE;
200 case REQ_OP_DISCARD:
201 return KYBER_DISCARD;
202 default:
203 return KYBER_OTHER;
204 }
205 }
206
flush_latency_buckets(struct kyber_queue_data * kqd,struct kyber_cpu_latency * cpu_latency,unsigned int sched_domain,unsigned int type)207 static void flush_latency_buckets(struct kyber_queue_data *kqd,
208 struct kyber_cpu_latency *cpu_latency,
209 unsigned int sched_domain, unsigned int type)
210 {
211 unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
212 atomic_t *cpu_buckets = cpu_latency->buckets[sched_domain][type];
213 unsigned int bucket;
214
215 for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
216 buckets[bucket] += atomic_xchg(&cpu_buckets[bucket], 0);
217 }
218
219 /*
220 * Calculate the histogram bucket with the given percentile rank, or -1 if there
221 * aren't enough samples yet.
222 */
calculate_percentile(struct kyber_queue_data * kqd,unsigned int sched_domain,unsigned int type,unsigned int percentile)223 static int calculate_percentile(struct kyber_queue_data *kqd,
224 unsigned int sched_domain, unsigned int type,
225 unsigned int percentile)
226 {
227 unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
228 unsigned int bucket, samples = 0, percentile_samples;
229
230 for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
231 samples += buckets[bucket];
232
233 if (!samples)
234 return -1;
235
236 /*
237 * We do the calculation once we have 500 samples or one second passes
238 * since the first sample was recorded, whichever comes first.
239 */
240 if (!kqd->latency_timeout[sched_domain])
241 kqd->latency_timeout[sched_domain] = max(jiffies + HZ, 1UL);
242 if (samples < 500 &&
243 time_is_after_jiffies(kqd->latency_timeout[sched_domain])) {
244 return -1;
245 }
246 kqd->latency_timeout[sched_domain] = 0;
247
248 percentile_samples = DIV_ROUND_UP(samples * percentile, 100);
249 for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS - 1; bucket++) {
250 if (buckets[bucket] >= percentile_samples)
251 break;
252 percentile_samples -= buckets[bucket];
253 }
254 memset(buckets, 0, sizeof(kqd->latency_buckets[sched_domain][type]));
255
256 trace_kyber_latency(kqd->dev, kyber_domain_names[sched_domain],
257 kyber_latency_type_names[type], percentile,
258 bucket + 1, 1 << KYBER_LATENCY_SHIFT, samples);
259
260 return bucket;
261 }
262
kyber_resize_domain(struct kyber_queue_data * kqd,unsigned int sched_domain,unsigned int depth)263 static void kyber_resize_domain(struct kyber_queue_data *kqd,
264 unsigned int sched_domain, unsigned int depth)
265 {
266 depth = clamp(depth, 1U, kyber_depth[sched_domain]);
267 if (depth != kqd->domain_tokens[sched_domain].sb.depth) {
268 sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
269 trace_kyber_adjust(kqd->dev, kyber_domain_names[sched_domain],
270 depth);
271 }
272 }
273
kyber_timer_fn(struct timer_list * t)274 static void kyber_timer_fn(struct timer_list *t)
275 {
276 struct kyber_queue_data *kqd = from_timer(kqd, t, timer);
277 unsigned int sched_domain;
278 int cpu;
279 bool bad = false;
280
281 /* Sum all of the per-cpu latency histograms. */
282 for_each_online_cpu(cpu) {
283 struct kyber_cpu_latency *cpu_latency;
284
285 cpu_latency = per_cpu_ptr(kqd->cpu_latency, cpu);
286 for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
287 flush_latency_buckets(kqd, cpu_latency, sched_domain,
288 KYBER_TOTAL_LATENCY);
289 flush_latency_buckets(kqd, cpu_latency, sched_domain,
290 KYBER_IO_LATENCY);
291 }
292 }
293
294 /*
295 * Check if any domains have a high I/O latency, which might indicate
296 * congestion in the device. Note that we use the p90; we don't want to
297 * be too sensitive to outliers here.
298 */
299 for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
300 int p90;
301
302 p90 = calculate_percentile(kqd, sched_domain, KYBER_IO_LATENCY,
303 90);
304 if (p90 >= KYBER_GOOD_BUCKETS)
305 bad = true;
306 }
307
308 /*
309 * Adjust the scheduling domain depths. If we determined that there was
310 * congestion, we throttle all domains with good latencies. Either way,
311 * we ease up on throttling domains with bad latencies.
312 */
313 for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
314 unsigned int orig_depth, depth;
315 int p99;
316
317 p99 = calculate_percentile(kqd, sched_domain,
318 KYBER_TOTAL_LATENCY, 99);
319 /*
320 * This is kind of subtle: different domains will not
321 * necessarily have enough samples to calculate the latency
322 * percentiles during the same window, so we have to remember
323 * the p99 for the next time we observe congestion; once we do,
324 * we don't want to throttle again until we get more data, so we
325 * reset it to -1.
326 */
327 if (bad) {
328 if (p99 < 0)
329 p99 = kqd->domain_p99[sched_domain];
330 kqd->domain_p99[sched_domain] = -1;
331 } else if (p99 >= 0) {
332 kqd->domain_p99[sched_domain] = p99;
333 }
334 if (p99 < 0)
335 continue;
336
337 /*
338 * If this domain has bad latency, throttle less. Otherwise,
339 * throttle more iff we determined that there is congestion.
340 *
341 * The new depth is scaled linearly with the p99 latency vs the
342 * latency target. E.g., if the p99 is 3/4 of the target, then
343 * we throttle down to 3/4 of the current depth, and if the p99
344 * is 2x the target, then we double the depth.
345 */
346 if (bad || p99 >= KYBER_GOOD_BUCKETS) {
347 orig_depth = kqd->domain_tokens[sched_domain].sb.depth;
348 depth = (orig_depth * (p99 + 1)) >> KYBER_LATENCY_SHIFT;
349 kyber_resize_domain(kqd, sched_domain, depth);
350 }
351 }
352 }
353
kyber_queue_data_alloc(struct request_queue * q)354 static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
355 {
356 struct kyber_queue_data *kqd;
357 int ret = -ENOMEM;
358 int i;
359
360 kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
361 if (!kqd)
362 goto err;
363
364 kqd->q = q;
365 kqd->dev = disk_devt(q->disk);
366
367 kqd->cpu_latency = alloc_percpu_gfp(struct kyber_cpu_latency,
368 GFP_KERNEL | __GFP_ZERO);
369 if (!kqd->cpu_latency)
370 goto err_kqd;
371
372 timer_setup(&kqd->timer, kyber_timer_fn, 0);
373
374 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
375 WARN_ON(!kyber_depth[i]);
376 WARN_ON(!kyber_batch_size[i]);
377 ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
378 kyber_depth[i], -1, false,
379 GFP_KERNEL, q->node);
380 if (ret) {
381 while (--i >= 0)
382 sbitmap_queue_free(&kqd->domain_tokens[i]);
383 goto err_buckets;
384 }
385 }
386
387 for (i = 0; i < KYBER_OTHER; i++) {
388 kqd->domain_p99[i] = -1;
389 kqd->latency_targets[i] = kyber_latency_targets[i];
390 }
391
392 return kqd;
393
394 err_buckets:
395 free_percpu(kqd->cpu_latency);
396 err_kqd:
397 kfree(kqd);
398 err:
399 return ERR_PTR(ret);
400 }
401
kyber_init_sched(struct request_queue * q,struct elevator_type * e)402 static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
403 {
404 struct kyber_queue_data *kqd;
405 struct elevator_queue *eq;
406
407 eq = elevator_alloc(q, e);
408 if (!eq)
409 return -ENOMEM;
410
411 kqd = kyber_queue_data_alloc(q);
412 if (IS_ERR(kqd)) {
413 kobject_put(&eq->kobj);
414 return PTR_ERR(kqd);
415 }
416
417 blk_stat_enable_accounting(q);
418
419 blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);
420
421 eq->elevator_data = kqd;
422 q->elevator = eq;
423
424 return 0;
425 }
426
kyber_exit_sched(struct elevator_queue * e)427 static void kyber_exit_sched(struct elevator_queue *e)
428 {
429 struct kyber_queue_data *kqd = e->elevator_data;
430 int i;
431
432 timer_shutdown_sync(&kqd->timer);
433 blk_stat_disable_accounting(kqd->q);
434
435 for (i = 0; i < KYBER_NUM_DOMAINS; i++)
436 sbitmap_queue_free(&kqd->domain_tokens[i]);
437 free_percpu(kqd->cpu_latency);
438 kfree(kqd);
439 }
440
kyber_ctx_queue_init(struct kyber_ctx_queue * kcq)441 static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq)
442 {
443 unsigned int i;
444
445 spin_lock_init(&kcq->lock);
446 for (i = 0; i < KYBER_NUM_DOMAINS; i++)
447 INIT_LIST_HEAD(&kcq->rq_list[i]);
448 }
449
kyber_depth_updated(struct blk_mq_hw_ctx * hctx)450 static void kyber_depth_updated(struct blk_mq_hw_ctx *hctx)
451 {
452 struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
453 struct blk_mq_tags *tags = hctx->sched_tags;
454
455 kqd->async_depth = hctx->queue->nr_requests * KYBER_ASYNC_PERCENT / 100U;
456 sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, kqd->async_depth);
457 }
458
kyber_init_hctx(struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)459 static 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 false, false)) {
481 while (--i >= 0)
482 sbitmap_free(&khd->kcq_map[i]);
483 goto err_kcqs;
484 }
485 }
486
487 spin_lock_init(&khd->lock);
488
489 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
490 INIT_LIST_HEAD(&khd->rqs[i]);
491 khd->domain_wait[i].sbq = NULL;
492 init_waitqueue_func_entry(&khd->domain_wait[i].wait,
493 kyber_domain_wake);
494 khd->domain_wait[i].wait.private = hctx;
495 INIT_LIST_HEAD(&khd->domain_wait[i].wait.entry);
496 atomic_set(&khd->wait_index[i], 0);
497 }
498
499 khd->cur_domain = 0;
500 khd->batching = 0;
501
502 hctx->sched_data = khd;
503 kyber_depth_updated(hctx);
504
505 return 0;
506
507 err_kcqs:
508 kfree(khd->kcqs);
509 err_khd:
510 kfree(khd);
511 return -ENOMEM;
512 }
513
kyber_exit_hctx(struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)514 static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
515 {
516 struct kyber_hctx_data *khd = hctx->sched_data;
517 int i;
518
519 for (i = 0; i < KYBER_NUM_DOMAINS; i++)
520 sbitmap_free(&khd->kcq_map[i]);
521 kfree(khd->kcqs);
522 kfree(hctx->sched_data);
523 }
524
rq_get_domain_token(struct request * rq)525 static int rq_get_domain_token(struct request *rq)
526 {
527 return (long)rq->elv.priv[0];
528 }
529
rq_set_domain_token(struct request * rq,int token)530 static void rq_set_domain_token(struct request *rq, int token)
531 {
532 rq->elv.priv[0] = (void *)(long)token;
533 }
534
rq_clear_domain_token(struct kyber_queue_data * kqd,struct request * rq)535 static void rq_clear_domain_token(struct kyber_queue_data *kqd,
536 struct request *rq)
537 {
538 unsigned int sched_domain;
539 int nr;
540
541 nr = rq_get_domain_token(rq);
542 if (nr != -1) {
543 sched_domain = kyber_sched_domain(rq->cmd_flags);
544 sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
545 rq->mq_ctx->cpu);
546 }
547 }
548
kyber_limit_depth(blk_opf_t opf,struct blk_mq_alloc_data * data)549 static void kyber_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
550 {
551 /*
552 * We use the scheduler tags as per-hardware queue queueing tokens.
553 * Async requests can be limited at this stage.
554 */
555 if (!op_is_sync(opf)) {
556 struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
557
558 data->shallow_depth = kqd->async_depth;
559 }
560 }
561
kyber_bio_merge(struct request_queue * q,struct bio * bio,unsigned int nr_segs)562 static bool kyber_bio_merge(struct request_queue *q, struct bio *bio,
563 unsigned int nr_segs)
564 {
565 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
566 struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
567 struct kyber_hctx_data *khd = hctx->sched_data;
568 struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]];
569 unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);
570 struct list_head *rq_list = &kcq->rq_list[sched_domain];
571 bool merged;
572
573 spin_lock(&kcq->lock);
574 merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs);
575 spin_unlock(&kcq->lock);
576
577 return merged;
578 }
579
kyber_prepare_request(struct request * rq)580 static void kyber_prepare_request(struct request *rq)
581 {
582 rq_set_domain_token(rq, -1);
583 }
584
kyber_insert_requests(struct blk_mq_hw_ctx * hctx,struct list_head * rq_list,blk_insert_t flags)585 static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx,
586 struct list_head *rq_list,
587 blk_insert_t flags)
588 {
589 struct kyber_hctx_data *khd = hctx->sched_data;
590 struct request *rq, *next;
591
592 list_for_each_entry_safe(rq, next, rq_list, queuelist) {
593 unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags);
594 struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]];
595 struct list_head *head = &kcq->rq_list[sched_domain];
596
597 spin_lock(&kcq->lock);
598 trace_block_rq_insert(rq);
599 if (flags & BLK_MQ_INSERT_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 spin_unlock(&kcq->lock);
606 }
607 }
608
kyber_finish_request(struct request * rq)609 static void kyber_finish_request(struct request *rq)
610 {
611 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
612
613 rq_clear_domain_token(kqd, rq);
614 }
615
add_latency_sample(struct kyber_cpu_latency * cpu_latency,unsigned int sched_domain,unsigned int type,u64 target,u64 latency)616 static void add_latency_sample(struct kyber_cpu_latency *cpu_latency,
617 unsigned int sched_domain, unsigned int type,
618 u64 target, u64 latency)
619 {
620 unsigned int bucket;
621 u64 divisor;
622
623 if (latency > 0) {
624 divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1);
625 bucket = min_t(unsigned int, div64_u64(latency - 1, divisor),
626 KYBER_LATENCY_BUCKETS - 1);
627 } else {
628 bucket = 0;
629 }
630
631 atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]);
632 }
633
kyber_completed_request(struct request * rq,u64 now)634 static void kyber_completed_request(struct request *rq, u64 now)
635 {
636 struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
637 struct kyber_cpu_latency *cpu_latency;
638 unsigned int sched_domain;
639 u64 target;
640
641 sched_domain = kyber_sched_domain(rq->cmd_flags);
642 if (sched_domain == KYBER_OTHER)
643 return;
644
645 cpu_latency = get_cpu_ptr(kqd->cpu_latency);
646 target = kqd->latency_targets[sched_domain];
647 add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY,
648 target, now - rq->start_time_ns);
649 add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target,
650 now - rq->io_start_time_ns);
651 put_cpu_ptr(kqd->cpu_latency);
652
653 timer_reduce(&kqd->timer, jiffies + HZ / 10);
654 }
655
656 struct flush_kcq_data {
657 struct kyber_hctx_data *khd;
658 unsigned int sched_domain;
659 struct list_head *list;
660 };
661
flush_busy_kcq(struct sbitmap * sb,unsigned int bitnr,void * data)662 static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data)
663 {
664 struct flush_kcq_data *flush_data = data;
665 struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr];
666
667 spin_lock(&kcq->lock);
668 list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain],
669 flush_data->list);
670 sbitmap_clear_bit(sb, bitnr);
671 spin_unlock(&kcq->lock);
672
673 return true;
674 }
675
kyber_flush_busy_kcqs(struct kyber_hctx_data * khd,unsigned int sched_domain,struct list_head * list)676 static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd,
677 unsigned int sched_domain,
678 struct list_head *list)
679 {
680 struct flush_kcq_data data = {
681 .khd = khd,
682 .sched_domain = sched_domain,
683 .list = list,
684 };
685
686 sbitmap_for_each_set(&khd->kcq_map[sched_domain],
687 flush_busy_kcq, &data);
688 }
689
kyber_domain_wake(wait_queue_entry_t * wqe,unsigned mode,int flags,void * key)690 static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags,
691 void *key)
692 {
693 struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private);
694 struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait);
695
696 sbitmap_del_wait_queue(wait);
697 blk_mq_run_hw_queue(hctx, true);
698 return 1;
699 }
700
kyber_get_domain_token(struct kyber_queue_data * kqd,struct kyber_hctx_data * khd,struct blk_mq_hw_ctx * hctx)701 static int kyber_get_domain_token(struct kyber_queue_data *kqd,
702 struct kyber_hctx_data *khd,
703 struct blk_mq_hw_ctx *hctx)
704 {
705 unsigned int sched_domain = khd->cur_domain;
706 struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
707 struct sbq_wait *wait = &khd->domain_wait[sched_domain];
708 struct sbq_wait_state *ws;
709 int nr;
710
711 nr = __sbitmap_queue_get(domain_tokens);
712
713 /*
714 * If we failed to get a domain token, make sure the hardware queue is
715 * run when one becomes available. Note that this is serialized on
716 * khd->lock, but we still need to be careful about the waker.
717 */
718 if (nr < 0 && list_empty_careful(&wait->wait.entry)) {
719 ws = sbq_wait_ptr(domain_tokens,
720 &khd->wait_index[sched_domain]);
721 khd->domain_ws[sched_domain] = ws;
722 sbitmap_add_wait_queue(domain_tokens, ws, wait);
723
724 /*
725 * Try again in case a token was freed before we got on the wait
726 * queue.
727 */
728 nr = __sbitmap_queue_get(domain_tokens);
729 }
730
731 /*
732 * If we got a token while we were on the wait queue, remove ourselves
733 * from the wait queue to ensure that all wake ups make forward
734 * progress. It's possible that the waker already deleted the entry
735 * between the !list_empty_careful() check and us grabbing the lock, but
736 * list_del_init() is okay with that.
737 */
738 if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) {
739 ws = khd->domain_ws[sched_domain];
740 spin_lock_irq(&ws->wait.lock);
741 sbitmap_del_wait_queue(wait);
742 spin_unlock_irq(&ws->wait.lock);
743 }
744
745 return nr;
746 }
747
748 static struct request *
kyber_dispatch_cur_domain(struct kyber_queue_data * kqd,struct kyber_hctx_data * khd,struct blk_mq_hw_ctx * hctx)749 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
750 struct kyber_hctx_data *khd,
751 struct blk_mq_hw_ctx *hctx)
752 {
753 struct list_head *rqs;
754 struct request *rq;
755 int nr;
756
757 rqs = &khd->rqs[khd->cur_domain];
758
759 /*
760 * If we already have a flushed request, then we just need to get a
761 * token for it. Otherwise, if there are pending requests in the kcqs,
762 * flush the kcqs, but only if we can get a token. If not, we should
763 * leave the requests in the kcqs so that they can be merged. Note that
764 * khd->lock serializes the flushes, so if we observed any bit set in
765 * the kcq_map, we will always get a request.
766 */
767 rq = list_first_entry_or_null(rqs, struct request, queuelist);
768 if (rq) {
769 nr = kyber_get_domain_token(kqd, khd, hctx);
770 if (nr >= 0) {
771 khd->batching++;
772 rq_set_domain_token(rq, nr);
773 list_del_init(&rq->queuelist);
774 return rq;
775 } else {
776 trace_kyber_throttled(kqd->dev,
777 kyber_domain_names[khd->cur_domain]);
778 }
779 } else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) {
780 nr = kyber_get_domain_token(kqd, khd, hctx);
781 if (nr >= 0) {
782 kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs);
783 rq = list_first_entry(rqs, struct request, queuelist);
784 khd->batching++;
785 rq_set_domain_token(rq, nr);
786 list_del_init(&rq->queuelist);
787 return rq;
788 } else {
789 trace_kyber_throttled(kqd->dev,
790 kyber_domain_names[khd->cur_domain]);
791 }
792 }
793
794 /* There were either no pending requests or no tokens. */
795 return NULL;
796 }
797
kyber_dispatch_request(struct blk_mq_hw_ctx * hctx)798 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
799 {
800 struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
801 struct kyber_hctx_data *khd = hctx->sched_data;
802 struct request *rq;
803 int i;
804
805 spin_lock(&khd->lock);
806
807 /*
808 * First, if we are still entitled to batch, try to dispatch a request
809 * from the batch.
810 */
811 if (khd->batching < kyber_batch_size[khd->cur_domain]) {
812 rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
813 if (rq)
814 goto out;
815 }
816
817 /*
818 * Either,
819 * 1. We were no longer entitled to a batch.
820 * 2. The domain we were batching didn't have any requests.
821 * 3. The domain we were batching was out of tokens.
822 *
823 * Start another batch. Note that this wraps back around to the original
824 * domain if no other domains have requests or tokens.
825 */
826 khd->batching = 0;
827 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
828 if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
829 khd->cur_domain = 0;
830 else
831 khd->cur_domain++;
832
833 rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
834 if (rq)
835 goto out;
836 }
837
838 rq = NULL;
839 out:
840 spin_unlock(&khd->lock);
841 return rq;
842 }
843
kyber_has_work(struct blk_mq_hw_ctx * hctx)844 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
845 {
846 struct kyber_hctx_data *khd = hctx->sched_data;
847 int i;
848
849 for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
850 if (!list_empty_careful(&khd->rqs[i]) ||
851 sbitmap_any_bit_set(&khd->kcq_map[i]))
852 return true;
853 }
854
855 return false;
856 }
857
858 #define KYBER_LAT_SHOW_STORE(domain, name) \
859 static ssize_t kyber_##name##_lat_show(struct elevator_queue *e, \
860 char *page) \
861 { \
862 struct kyber_queue_data *kqd = e->elevator_data; \
863 \
864 return sprintf(page, "%llu\n", kqd->latency_targets[domain]); \
865 } \
866 \
867 static ssize_t kyber_##name##_lat_store(struct elevator_queue *e, \
868 const char *page, size_t count) \
869 { \
870 struct kyber_queue_data *kqd = e->elevator_data; \
871 unsigned long long nsec; \
872 int ret; \
873 \
874 ret = kstrtoull(page, 10, &nsec); \
875 if (ret) \
876 return ret; \
877 \
878 kqd->latency_targets[domain] = nsec; \
879 \
880 return count; \
881 }
882 KYBER_LAT_SHOW_STORE(KYBER_READ, read);
883 KYBER_LAT_SHOW_STORE(KYBER_WRITE, write);
884 #undef KYBER_LAT_SHOW_STORE
885
886 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
887 static struct elv_fs_entry kyber_sched_attrs[] = {
888 KYBER_LAT_ATTR(read),
889 KYBER_LAT_ATTR(write),
890 __ATTR_NULL
891 };
892 #undef KYBER_LAT_ATTR
893
894 #ifdef CONFIG_BLK_DEBUG_FS
895 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name) \
896 static int kyber_##name##_tokens_show(void *data, struct seq_file *m) \
897 { \
898 struct request_queue *q = data; \
899 struct kyber_queue_data *kqd = q->elevator->elevator_data; \
900 \
901 sbitmap_queue_show(&kqd->domain_tokens[domain], m); \
902 return 0; \
903 } \
904 \
905 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos) \
906 __acquires(&khd->lock) \
907 { \
908 struct blk_mq_hw_ctx *hctx = m->private; \
909 struct kyber_hctx_data *khd = hctx->sched_data; \
910 \
911 spin_lock(&khd->lock); \
912 return seq_list_start(&khd->rqs[domain], *pos); \
913 } \
914 \
915 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v, \
916 loff_t *pos) \
917 { \
918 struct blk_mq_hw_ctx *hctx = m->private; \
919 struct kyber_hctx_data *khd = hctx->sched_data; \
920 \
921 return seq_list_next(v, &khd->rqs[domain], pos); \
922 } \
923 \
924 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v) \
925 __releases(&khd->lock) \
926 { \
927 struct blk_mq_hw_ctx *hctx = m->private; \
928 struct kyber_hctx_data *khd = hctx->sched_data; \
929 \
930 spin_unlock(&khd->lock); \
931 } \
932 \
933 static const struct seq_operations kyber_##name##_rqs_seq_ops = { \
934 .start = kyber_##name##_rqs_start, \
935 .next = kyber_##name##_rqs_next, \
936 .stop = kyber_##name##_rqs_stop, \
937 .show = blk_mq_debugfs_rq_show, \
938 }; \
939 \
940 static int kyber_##name##_waiting_show(void *data, struct seq_file *m) \
941 { \
942 struct blk_mq_hw_ctx *hctx = data; \
943 struct kyber_hctx_data *khd = hctx->sched_data; \
944 wait_queue_entry_t *wait = &khd->domain_wait[domain].wait; \
945 \
946 seq_printf(m, "%d\n", !list_empty_careful(&wait->entry)); \
947 return 0; \
948 }
KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ,read)949 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
950 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write)
951 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
952 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
953 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
954
955 static int kyber_async_depth_show(void *data, struct seq_file *m)
956 {
957 struct request_queue *q = data;
958 struct kyber_queue_data *kqd = q->elevator->elevator_data;
959
960 seq_printf(m, "%u\n", kqd->async_depth);
961 return 0;
962 }
963
kyber_cur_domain_show(void * data,struct seq_file * m)964 static int kyber_cur_domain_show(void *data, struct seq_file *m)
965 {
966 struct blk_mq_hw_ctx *hctx = data;
967 struct kyber_hctx_data *khd = hctx->sched_data;
968
969 seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]);
970 return 0;
971 }
972
kyber_batching_show(void * data,struct seq_file * m)973 static int kyber_batching_show(void *data, struct seq_file *m)
974 {
975 struct blk_mq_hw_ctx *hctx = data;
976 struct kyber_hctx_data *khd = hctx->sched_data;
977
978 seq_printf(m, "%u\n", khd->batching);
979 return 0;
980 }
981
982 #define KYBER_QUEUE_DOMAIN_ATTRS(name) \
983 {#name "_tokens", 0400, kyber_##name##_tokens_show}
984 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
985 KYBER_QUEUE_DOMAIN_ATTRS(read),
986 KYBER_QUEUE_DOMAIN_ATTRS(write),
987 KYBER_QUEUE_DOMAIN_ATTRS(discard),
988 KYBER_QUEUE_DOMAIN_ATTRS(other),
989 {"async_depth", 0400, kyber_async_depth_show},
990 {},
991 };
992 #undef KYBER_QUEUE_DOMAIN_ATTRS
993
994 #define KYBER_HCTX_DOMAIN_ATTRS(name) \
995 {#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops}, \
996 {#name "_waiting", 0400, kyber_##name##_waiting_show}
997 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
998 KYBER_HCTX_DOMAIN_ATTRS(read),
999 KYBER_HCTX_DOMAIN_ATTRS(write),
1000 KYBER_HCTX_DOMAIN_ATTRS(discard),
1001 KYBER_HCTX_DOMAIN_ATTRS(other),
1002 {"cur_domain", 0400, kyber_cur_domain_show},
1003 {"batching", 0400, kyber_batching_show},
1004 {},
1005 };
1006 #undef KYBER_HCTX_DOMAIN_ATTRS
1007 #endif
1008
1009 static struct elevator_type kyber_sched = {
1010 .ops = {
1011 .init_sched = kyber_init_sched,
1012 .exit_sched = kyber_exit_sched,
1013 .init_hctx = kyber_init_hctx,
1014 .exit_hctx = kyber_exit_hctx,
1015 .limit_depth = kyber_limit_depth,
1016 .bio_merge = kyber_bio_merge,
1017 .prepare_request = kyber_prepare_request,
1018 .insert_requests = kyber_insert_requests,
1019 .finish_request = kyber_finish_request,
1020 .requeue_request = kyber_finish_request,
1021 .completed_request = kyber_completed_request,
1022 .dispatch_request = kyber_dispatch_request,
1023 .has_work = kyber_has_work,
1024 .depth_updated = kyber_depth_updated,
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
kyber_init(void)1035 static int __init kyber_init(void)
1036 {
1037 return elv_register(&kyber_sched);
1038 }
1039
kyber_exit(void)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