xref: /linux/block/kyber-iosched.c (revision 1ac731c529cd4d6adbce134754b51ff7d822b145)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * The Kyber I/O scheduler. Controls latency by throttling queue depths using
4  * scalable techniques.
5  *
6  * Copyright (C) 2017 Facebook
7  */
8 
9 #include <linux/kernel.h>
10 #include <linux/blkdev.h>
11 #include <linux/module.h>
12 #include <linux/sbitmap.h>
13 
14 #include <trace/events/block.h>
15 
16 #include "elevator.h"
17 #include "blk.h"
18 #include "blk-mq.h"
19 #include "blk-mq-debugfs.h"
20 #include "blk-mq-sched.h"
21 
22 #define CREATE_TRACE_POINTS
23 #include <trace/events/kyber.h>
24 
25 /*
26  * Scheduling domains: the device is divided into multiple domains based on the
27  * request type.
28  */
29 enum {
30 	KYBER_READ,
31 	KYBER_WRITE,
32 	KYBER_DISCARD,
33 	KYBER_OTHER,
34 	KYBER_NUM_DOMAINS,
35 };
36 
37 static const char *kyber_domain_names[] = {
38 	[KYBER_READ] = "READ",
39 	[KYBER_WRITE] = "WRITE",
40 	[KYBER_DISCARD] = "DISCARD",
41 	[KYBER_OTHER] = "OTHER",
42 };
43 
44 enum {
45 	/*
46 	 * In order to prevent starvation of synchronous requests by a flood of
47 	 * asynchronous requests, we reserve 25% of requests for synchronous
48 	 * operations.
49 	 */
50 	KYBER_ASYNC_PERCENT = 75,
51 };
52 
53 /*
54  * Maximum device-wide depth for each scheduling domain.
55  *
56  * Even for fast devices with lots of tags like NVMe, you can saturate the
57  * device with only a fraction of the maximum possible queue depth. So, we cap
58  * these to a reasonable value.
59  */
60 static const unsigned int kyber_depth[] = {
61 	[KYBER_READ] = 256,
62 	[KYBER_WRITE] = 128,
63 	[KYBER_DISCARD] = 64,
64 	[KYBER_OTHER] = 16,
65 };
66 
67 /*
68  * Default latency targets for each scheduling domain.
69  */
70 static const u64 kyber_latency_targets[] = {
71 	[KYBER_READ] = 2ULL * NSEC_PER_MSEC,
72 	[KYBER_WRITE] = 10ULL * NSEC_PER_MSEC,
73 	[KYBER_DISCARD] = 5ULL * NSEC_PER_SEC,
74 };
75 
76 /*
77  * Batch size (number of requests we'll dispatch in a row) for each scheduling
78  * domain.
79  */
80 static const unsigned int kyber_batch_size[] = {
81 	[KYBER_READ] = 16,
82 	[KYBER_WRITE] = 8,
83 	[KYBER_DISCARD] = 1,
84 	[KYBER_OTHER] = 1,
85 };
86 
87 /*
88  * Requests latencies are recorded in a histogram with buckets defined relative
89  * to the target latency:
90  *
91  * <= 1/4 * target latency
92  * <= 1/2 * target latency
93  * <= 3/4 * target latency
94  * <= target latency
95  * <= 1 1/4 * target latency
96  * <= 1 1/2 * target latency
97  * <= 1 3/4 * target latency
98  * > 1 3/4 * target latency
99  */
100 enum {
101 	/*
102 	 * The width of the latency histogram buckets is
103 	 * 1 / (1 << KYBER_LATENCY_SHIFT) * target latency.
104 	 */
105 	KYBER_LATENCY_SHIFT = 2,
106 	/*
107 	 * The first (1 << KYBER_LATENCY_SHIFT) buckets are <= target latency,
108 	 * thus, "good".
109 	 */
110 	KYBER_GOOD_BUCKETS = 1 << KYBER_LATENCY_SHIFT,
111 	/* There are also (1 << KYBER_LATENCY_SHIFT) "bad" buckets. */
112 	KYBER_LATENCY_BUCKETS = 2 << KYBER_LATENCY_SHIFT,
113 };
114 
115 /*
116  * We measure both the total latency and the I/O latency (i.e., latency after
117  * submitting to the device).
118  */
119 enum {
120 	KYBER_TOTAL_LATENCY,
121 	KYBER_IO_LATENCY,
122 };
123 
124 static const char *kyber_latency_type_names[] = {
125 	[KYBER_TOTAL_LATENCY] = "total",
126 	[KYBER_IO_LATENCY] = "I/O",
127 };
128 
129 /*
130  * Per-cpu latency histograms: total latency and I/O latency for each scheduling
131  * domain except for KYBER_OTHER.
132  */
133 struct kyber_cpu_latency {
134 	atomic_t buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
135 };
136 
137 /*
138  * There is a same mapping between ctx & hctx and kcq & khd,
139  * we use request->mq_ctx->index_hw to index the kcq in khd.
140  */
141 struct kyber_ctx_queue {
142 	/*
143 	 * Used to ensure operations on rq_list and kcq_map to be an atmoic one.
144 	 * Also protect the rqs on rq_list when merge.
145 	 */
146 	spinlock_t lock;
147 	struct list_head rq_list[KYBER_NUM_DOMAINS];
148 } ____cacheline_aligned_in_smp;
149 
150 struct kyber_queue_data {
151 	struct request_queue *q;
152 	dev_t dev;
153 
154 	/*
155 	 * Each scheduling domain has a limited number of in-flight requests
156 	 * device-wide, limited by these tokens.
157 	 */
158 	struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
159 
160 	/*
161 	 * Async request percentage, converted to per-word depth for
162 	 * sbitmap_get_shallow().
163 	 */
164 	unsigned int async_depth;
165 
166 	struct kyber_cpu_latency __percpu *cpu_latency;
167 
168 	/* Timer for stats aggregation and adjusting domain tokens. */
169 	struct timer_list timer;
170 
171 	unsigned int latency_buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
172 
173 	unsigned long latency_timeout[KYBER_OTHER];
174 
175 	int domain_p99[KYBER_OTHER];
176 
177 	/* Target latencies in nanoseconds. */
178 	u64 latency_targets[KYBER_OTHER];
179 };
180 
181 struct kyber_hctx_data {
182 	spinlock_t lock;
183 	struct list_head rqs[KYBER_NUM_DOMAINS];
184 	unsigned int cur_domain;
185 	unsigned int batching;
186 	struct kyber_ctx_queue *kcqs;
187 	struct sbitmap kcq_map[KYBER_NUM_DOMAINS];
188 	struct sbq_wait domain_wait[KYBER_NUM_DOMAINS];
189 	struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
190 	atomic_t wait_index[KYBER_NUM_DOMAINS];
191 };
192 
193 static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
194 			     void *key);
195 
kyber_sched_domain(blk_opf_t opf)196 static unsigned int kyber_sched_domain(blk_opf_t opf)
197 {
198 	switch (opf & REQ_OP_MASK) {
199 	case REQ_OP_READ:
200 		return KYBER_READ;
201 	case REQ_OP_WRITE:
202 		return KYBER_WRITE;
203 	case REQ_OP_DISCARD:
204 		return KYBER_DISCARD;
205 	default:
206 		return KYBER_OTHER;
207 	}
208 }
209 
flush_latency_buckets(struct kyber_queue_data * kqd,struct kyber_cpu_latency * cpu_latency,unsigned int sched_domain,unsigned int type)210 static void flush_latency_buckets(struct kyber_queue_data *kqd,
211 				  struct kyber_cpu_latency *cpu_latency,
212 				  unsigned int sched_domain, unsigned int type)
213 {
214 	unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
215 	atomic_t *cpu_buckets = cpu_latency->buckets[sched_domain][type];
216 	unsigned int bucket;
217 
218 	for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
219 		buckets[bucket] += atomic_xchg(&cpu_buckets[bucket], 0);
220 }
221 
222 /*
223  * Calculate the histogram bucket with the given percentile rank, or -1 if there
224  * aren't enough samples yet.
225  */
calculate_percentile(struct kyber_queue_data * kqd,unsigned int sched_domain,unsigned int type,unsigned int percentile)226 static int calculate_percentile(struct kyber_queue_data *kqd,
227 				unsigned int sched_domain, unsigned int type,
228 				unsigned int percentile)
229 {
230 	unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
231 	unsigned int bucket, samples = 0, percentile_samples;
232 
233 	for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
234 		samples += buckets[bucket];
235 
236 	if (!samples)
237 		return -1;
238 
239 	/*
240 	 * We do the calculation once we have 500 samples or one second passes
241 	 * since the first sample was recorded, whichever comes first.
242 	 */
243 	if (!kqd->latency_timeout[sched_domain])
244 		kqd->latency_timeout[sched_domain] = max(jiffies + HZ, 1UL);
245 	if (samples < 500 &&
246 	    time_is_after_jiffies(kqd->latency_timeout[sched_domain])) {
247 		return -1;
248 	}
249 	kqd->latency_timeout[sched_domain] = 0;
250 
251 	percentile_samples = DIV_ROUND_UP(samples * percentile, 100);
252 	for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS - 1; bucket++) {
253 		if (buckets[bucket] >= percentile_samples)
254 			break;
255 		percentile_samples -= buckets[bucket];
256 	}
257 	memset(buckets, 0, sizeof(kqd->latency_buckets[sched_domain][type]));
258 
259 	trace_kyber_latency(kqd->dev, kyber_domain_names[sched_domain],
260 			    kyber_latency_type_names[type], percentile,
261 			    bucket + 1, 1 << KYBER_LATENCY_SHIFT, samples);
262 
263 	return bucket;
264 }
265 
kyber_resize_domain(struct kyber_queue_data * kqd,unsigned int sched_domain,unsigned int depth)266 static void kyber_resize_domain(struct kyber_queue_data *kqd,
267 				unsigned int sched_domain, unsigned int depth)
268 {
269 	depth = clamp(depth, 1U, kyber_depth[sched_domain]);
270 	if (depth != kqd->domain_tokens[sched_domain].sb.depth) {
271 		sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
272 		trace_kyber_adjust(kqd->dev, kyber_domain_names[sched_domain],
273 				   depth);
274 	}
275 }
276 
kyber_timer_fn(struct timer_list * t)277 static void kyber_timer_fn(struct timer_list *t)
278 {
279 	struct kyber_queue_data *kqd = from_timer(kqd, t, timer);
280 	unsigned int sched_domain;
281 	int cpu;
282 	bool bad = false;
283 
284 	/* Sum all of the per-cpu latency histograms. */
285 	for_each_online_cpu(cpu) {
286 		struct kyber_cpu_latency *cpu_latency;
287 
288 		cpu_latency = per_cpu_ptr(kqd->cpu_latency, cpu);
289 		for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
290 			flush_latency_buckets(kqd, cpu_latency, sched_domain,
291 					      KYBER_TOTAL_LATENCY);
292 			flush_latency_buckets(kqd, cpu_latency, sched_domain,
293 					      KYBER_IO_LATENCY);
294 		}
295 	}
296 
297 	/*
298 	 * Check if any domains have a high I/O latency, which might indicate
299 	 * congestion in the device. Note that we use the p90; we don't want to
300 	 * be too sensitive to outliers here.
301 	 */
302 	for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
303 		int p90;
304 
305 		p90 = calculate_percentile(kqd, sched_domain, KYBER_IO_LATENCY,
306 					   90);
307 		if (p90 >= KYBER_GOOD_BUCKETS)
308 			bad = true;
309 	}
310 
311 	/*
312 	 * Adjust the scheduling domain depths. If we determined that there was
313 	 * congestion, we throttle all domains with good latencies. Either way,
314 	 * we ease up on throttling domains with bad latencies.
315 	 */
316 	for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
317 		unsigned int orig_depth, depth;
318 		int p99;
319 
320 		p99 = calculate_percentile(kqd, sched_domain,
321 					   KYBER_TOTAL_LATENCY, 99);
322 		/*
323 		 * This is kind of subtle: different domains will not
324 		 * necessarily have enough samples to calculate the latency
325 		 * percentiles during the same window, so we have to remember
326 		 * the p99 for the next time we observe congestion; once we do,
327 		 * we don't want to throttle again until we get more data, so we
328 		 * reset it to -1.
329 		 */
330 		if (bad) {
331 			if (p99 < 0)
332 				p99 = kqd->domain_p99[sched_domain];
333 			kqd->domain_p99[sched_domain] = -1;
334 		} else if (p99 >= 0) {
335 			kqd->domain_p99[sched_domain] = p99;
336 		}
337 		if (p99 < 0)
338 			continue;
339 
340 		/*
341 		 * If this domain has bad latency, throttle less. Otherwise,
342 		 * throttle more iff we determined that there is congestion.
343 		 *
344 		 * The new depth is scaled linearly with the p99 latency vs the
345 		 * latency target. E.g., if the p99 is 3/4 of the target, then
346 		 * we throttle down to 3/4 of the current depth, and if the p99
347 		 * is 2x the target, then we double the depth.
348 		 */
349 		if (bad || p99 >= KYBER_GOOD_BUCKETS) {
350 			orig_depth = kqd->domain_tokens[sched_domain].sb.depth;
351 			depth = (orig_depth * (p99 + 1)) >> KYBER_LATENCY_SHIFT;
352 			kyber_resize_domain(kqd, sched_domain, depth);
353 		}
354 	}
355 }
356 
kyber_queue_data_alloc(struct request_queue * q)357 static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
358 {
359 	struct kyber_queue_data *kqd;
360 	int ret = -ENOMEM;
361 	int i;
362 
363 	kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
364 	if (!kqd)
365 		goto err;
366 
367 	kqd->q = q;
368 	kqd->dev = disk_devt(q->disk);
369 
370 	kqd->cpu_latency = alloc_percpu_gfp(struct kyber_cpu_latency,
371 					    GFP_KERNEL | __GFP_ZERO);
372 	if (!kqd->cpu_latency)
373 		goto err_kqd;
374 
375 	timer_setup(&kqd->timer, kyber_timer_fn, 0);
376 
377 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
378 		WARN_ON(!kyber_depth[i]);
379 		WARN_ON(!kyber_batch_size[i]);
380 		ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
381 					      kyber_depth[i], -1, false,
382 					      GFP_KERNEL, q->node);
383 		if (ret) {
384 			while (--i >= 0)
385 				sbitmap_queue_free(&kqd->domain_tokens[i]);
386 			goto err_buckets;
387 		}
388 	}
389 
390 	for (i = 0; i < KYBER_OTHER; i++) {
391 		kqd->domain_p99[i] = -1;
392 		kqd->latency_targets[i] = kyber_latency_targets[i];
393 	}
394 
395 	return kqd;
396 
397 err_buckets:
398 	free_percpu(kqd->cpu_latency);
399 err_kqd:
400 	kfree(kqd);
401 err:
402 	return ERR_PTR(ret);
403 }
404 
kyber_init_sched(struct request_queue * q,struct elevator_type * e)405 static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
406 {
407 	struct kyber_queue_data *kqd;
408 	struct elevator_queue *eq;
409 
410 	eq = elevator_alloc(q, e);
411 	if (!eq)
412 		return -ENOMEM;
413 
414 	kqd = kyber_queue_data_alloc(q);
415 	if (IS_ERR(kqd)) {
416 		kobject_put(&eq->kobj);
417 		return PTR_ERR(kqd);
418 	}
419 
420 	blk_stat_enable_accounting(q);
421 
422 	blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);
423 
424 	eq->elevator_data = kqd;
425 	q->elevator = eq;
426 
427 	return 0;
428 }
429 
kyber_exit_sched(struct elevator_queue * e)430 static void kyber_exit_sched(struct elevator_queue *e)
431 {
432 	struct kyber_queue_data *kqd = e->elevator_data;
433 	int i;
434 
435 	timer_shutdown_sync(&kqd->timer);
436 	blk_stat_disable_accounting(kqd->q);
437 
438 	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
439 		sbitmap_queue_free(&kqd->domain_tokens[i]);
440 	free_percpu(kqd->cpu_latency);
441 	kfree(kqd);
442 }
443 
kyber_ctx_queue_init(struct kyber_ctx_queue * kcq)444 static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq)
445 {
446 	unsigned int i;
447 
448 	spin_lock_init(&kcq->lock);
449 	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
450 		INIT_LIST_HEAD(&kcq->rq_list[i]);
451 }
452 
kyber_depth_updated(struct blk_mq_hw_ctx * hctx)453 static void kyber_depth_updated(struct blk_mq_hw_ctx *hctx)
454 {
455 	struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
456 	struct blk_mq_tags *tags = hctx->sched_tags;
457 	unsigned int shift = tags->bitmap_tags.sb.shift;
458 
459 	kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
460 
461 	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, kqd->async_depth);
462 }
463 
kyber_init_hctx(struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)464 static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
465 {
466 	struct kyber_hctx_data *khd;
467 	int i;
468 
469 	khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
470 	if (!khd)
471 		return -ENOMEM;
472 
473 	khd->kcqs = kmalloc_array_node(hctx->nr_ctx,
474 				       sizeof(struct kyber_ctx_queue),
475 				       GFP_KERNEL, hctx->numa_node);
476 	if (!khd->kcqs)
477 		goto err_khd;
478 
479 	for (i = 0; i < hctx->nr_ctx; i++)
480 		kyber_ctx_queue_init(&khd->kcqs[i]);
481 
482 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
483 		if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx,
484 				      ilog2(8), GFP_KERNEL, hctx->numa_node,
485 				      false, false)) {
486 			while (--i >= 0)
487 				sbitmap_free(&khd->kcq_map[i]);
488 			goto err_kcqs;
489 		}
490 	}
491 
492 	spin_lock_init(&khd->lock);
493 
494 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
495 		INIT_LIST_HEAD(&khd->rqs[i]);
496 		khd->domain_wait[i].sbq = NULL;
497 		init_waitqueue_func_entry(&khd->domain_wait[i].wait,
498 					  kyber_domain_wake);
499 		khd->domain_wait[i].wait.private = hctx;
500 		INIT_LIST_HEAD(&khd->domain_wait[i].wait.entry);
501 		atomic_set(&khd->wait_index[i], 0);
502 	}
503 
504 	khd->cur_domain = 0;
505 	khd->batching = 0;
506 
507 	hctx->sched_data = khd;
508 	kyber_depth_updated(hctx);
509 
510 	return 0;
511 
512 err_kcqs:
513 	kfree(khd->kcqs);
514 err_khd:
515 	kfree(khd);
516 	return -ENOMEM;
517 }
518 
kyber_exit_hctx(struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)519 static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
520 {
521 	struct kyber_hctx_data *khd = hctx->sched_data;
522 	int i;
523 
524 	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
525 		sbitmap_free(&khd->kcq_map[i]);
526 	kfree(khd->kcqs);
527 	kfree(hctx->sched_data);
528 }
529 
rq_get_domain_token(struct request * rq)530 static int rq_get_domain_token(struct request *rq)
531 {
532 	return (long)rq->elv.priv[0];
533 }
534 
rq_set_domain_token(struct request * rq,int token)535 static void rq_set_domain_token(struct request *rq, int token)
536 {
537 	rq->elv.priv[0] = (void *)(long)token;
538 }
539 
rq_clear_domain_token(struct kyber_queue_data * kqd,struct request * rq)540 static void rq_clear_domain_token(struct kyber_queue_data *kqd,
541 				  struct request *rq)
542 {
543 	unsigned int sched_domain;
544 	int nr;
545 
546 	nr = rq_get_domain_token(rq);
547 	if (nr != -1) {
548 		sched_domain = kyber_sched_domain(rq->cmd_flags);
549 		sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
550 				    rq->mq_ctx->cpu);
551 	}
552 }
553 
kyber_limit_depth(blk_opf_t opf,struct blk_mq_alloc_data * data)554 static void kyber_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
555 {
556 	/*
557 	 * We use the scheduler tags as per-hardware queue queueing tokens.
558 	 * Async requests can be limited at this stage.
559 	 */
560 	if (!op_is_sync(opf)) {
561 		struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
562 
563 		data->shallow_depth = kqd->async_depth;
564 	}
565 }
566 
kyber_bio_merge(struct request_queue * q,struct bio * bio,unsigned int nr_segs)567 static bool kyber_bio_merge(struct request_queue *q, struct bio *bio,
568 		unsigned int nr_segs)
569 {
570 	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
571 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
572 	struct kyber_hctx_data *khd = hctx->sched_data;
573 	struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]];
574 	unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);
575 	struct list_head *rq_list = &kcq->rq_list[sched_domain];
576 	bool merged;
577 
578 	spin_lock(&kcq->lock);
579 	merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs);
580 	spin_unlock(&kcq->lock);
581 
582 	return merged;
583 }
584 
kyber_prepare_request(struct request * rq)585 static void kyber_prepare_request(struct request *rq)
586 {
587 	rq_set_domain_token(rq, -1);
588 }
589 
kyber_insert_requests(struct blk_mq_hw_ctx * hctx,struct list_head * rq_list,blk_insert_t flags)590 static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx,
591 				  struct list_head *rq_list,
592 				  blk_insert_t flags)
593 {
594 	struct kyber_hctx_data *khd = hctx->sched_data;
595 	struct request *rq, *next;
596 
597 	list_for_each_entry_safe(rq, next, rq_list, queuelist) {
598 		unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags);
599 		struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]];
600 		struct list_head *head = &kcq->rq_list[sched_domain];
601 
602 		spin_lock(&kcq->lock);
603 		trace_block_rq_insert(rq);
604 		if (flags & BLK_MQ_INSERT_AT_HEAD)
605 			list_move(&rq->queuelist, head);
606 		else
607 			list_move_tail(&rq->queuelist, head);
608 		sbitmap_set_bit(&khd->kcq_map[sched_domain],
609 				rq->mq_ctx->index_hw[hctx->type]);
610 		spin_unlock(&kcq->lock);
611 	}
612 }
613 
kyber_finish_request(struct request * rq)614 static void kyber_finish_request(struct request *rq)
615 {
616 	struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
617 
618 	rq_clear_domain_token(kqd, rq);
619 }
620 
add_latency_sample(struct kyber_cpu_latency * cpu_latency,unsigned int sched_domain,unsigned int type,u64 target,u64 latency)621 static void add_latency_sample(struct kyber_cpu_latency *cpu_latency,
622 			       unsigned int sched_domain, unsigned int type,
623 			       u64 target, u64 latency)
624 {
625 	unsigned int bucket;
626 	u64 divisor;
627 
628 	if (latency > 0) {
629 		divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1);
630 		bucket = min_t(unsigned int, div64_u64(latency - 1, divisor),
631 			       KYBER_LATENCY_BUCKETS - 1);
632 	} else {
633 		bucket = 0;
634 	}
635 
636 	atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]);
637 }
638 
kyber_completed_request(struct request * rq,u64 now)639 static void kyber_completed_request(struct request *rq, u64 now)
640 {
641 	struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
642 	struct kyber_cpu_latency *cpu_latency;
643 	unsigned int sched_domain;
644 	u64 target;
645 
646 	sched_domain = kyber_sched_domain(rq->cmd_flags);
647 	if (sched_domain == KYBER_OTHER)
648 		return;
649 
650 	cpu_latency = get_cpu_ptr(kqd->cpu_latency);
651 	target = kqd->latency_targets[sched_domain];
652 	add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY,
653 			   target, now - rq->start_time_ns);
654 	add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target,
655 			   now - rq->io_start_time_ns);
656 	put_cpu_ptr(kqd->cpu_latency);
657 
658 	timer_reduce(&kqd->timer, jiffies + HZ / 10);
659 }
660 
661 struct flush_kcq_data {
662 	struct kyber_hctx_data *khd;
663 	unsigned int sched_domain;
664 	struct list_head *list;
665 };
666 
flush_busy_kcq(struct sbitmap * sb,unsigned int bitnr,void * data)667 static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data)
668 {
669 	struct flush_kcq_data *flush_data = data;
670 	struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr];
671 
672 	spin_lock(&kcq->lock);
673 	list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain],
674 			      flush_data->list);
675 	sbitmap_clear_bit(sb, bitnr);
676 	spin_unlock(&kcq->lock);
677 
678 	return true;
679 }
680 
kyber_flush_busy_kcqs(struct kyber_hctx_data * khd,unsigned int sched_domain,struct list_head * list)681 static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd,
682 				  unsigned int sched_domain,
683 				  struct list_head *list)
684 {
685 	struct flush_kcq_data data = {
686 		.khd = khd,
687 		.sched_domain = sched_domain,
688 		.list = list,
689 	};
690 
691 	sbitmap_for_each_set(&khd->kcq_map[sched_domain],
692 			     flush_busy_kcq, &data);
693 }
694 
kyber_domain_wake(wait_queue_entry_t * wqe,unsigned mode,int flags,void * key)695 static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags,
696 			     void *key)
697 {
698 	struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private);
699 	struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait);
700 
701 	sbitmap_del_wait_queue(wait);
702 	blk_mq_run_hw_queue(hctx, true);
703 	return 1;
704 }
705 
kyber_get_domain_token(struct kyber_queue_data * kqd,struct kyber_hctx_data * khd,struct blk_mq_hw_ctx * hctx)706 static int kyber_get_domain_token(struct kyber_queue_data *kqd,
707 				  struct kyber_hctx_data *khd,
708 				  struct blk_mq_hw_ctx *hctx)
709 {
710 	unsigned int sched_domain = khd->cur_domain;
711 	struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
712 	struct sbq_wait *wait = &khd->domain_wait[sched_domain];
713 	struct sbq_wait_state *ws;
714 	int nr;
715 
716 	nr = __sbitmap_queue_get(domain_tokens);
717 
718 	/*
719 	 * If we failed to get a domain token, make sure the hardware queue is
720 	 * run when one becomes available. Note that this is serialized on
721 	 * khd->lock, but we still need to be careful about the waker.
722 	 */
723 	if (nr < 0 && list_empty_careful(&wait->wait.entry)) {
724 		ws = sbq_wait_ptr(domain_tokens,
725 				  &khd->wait_index[sched_domain]);
726 		khd->domain_ws[sched_domain] = ws;
727 		sbitmap_add_wait_queue(domain_tokens, ws, wait);
728 
729 		/*
730 		 * Try again in case a token was freed before we got on the wait
731 		 * queue.
732 		 */
733 		nr = __sbitmap_queue_get(domain_tokens);
734 	}
735 
736 	/*
737 	 * If we got a token while we were on the wait queue, remove ourselves
738 	 * from the wait queue to ensure that all wake ups make forward
739 	 * progress. It's possible that the waker already deleted the entry
740 	 * between the !list_empty_careful() check and us grabbing the lock, but
741 	 * list_del_init() is okay with that.
742 	 */
743 	if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) {
744 		ws = khd->domain_ws[sched_domain];
745 		spin_lock_irq(&ws->wait.lock);
746 		sbitmap_del_wait_queue(wait);
747 		spin_unlock_irq(&ws->wait.lock);
748 	}
749 
750 	return nr;
751 }
752 
753 static struct request *
kyber_dispatch_cur_domain(struct kyber_queue_data * kqd,struct kyber_hctx_data * khd,struct blk_mq_hw_ctx * hctx)754 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
755 			  struct kyber_hctx_data *khd,
756 			  struct blk_mq_hw_ctx *hctx)
757 {
758 	struct list_head *rqs;
759 	struct request *rq;
760 	int nr;
761 
762 	rqs = &khd->rqs[khd->cur_domain];
763 
764 	/*
765 	 * If we already have a flushed request, then we just need to get a
766 	 * token for it. Otherwise, if there are pending requests in the kcqs,
767 	 * flush the kcqs, but only if we can get a token. If not, we should
768 	 * leave the requests in the kcqs so that they can be merged. Note that
769 	 * khd->lock serializes the flushes, so if we observed any bit set in
770 	 * the kcq_map, we will always get a request.
771 	 */
772 	rq = list_first_entry_or_null(rqs, struct request, queuelist);
773 	if (rq) {
774 		nr = kyber_get_domain_token(kqd, khd, hctx);
775 		if (nr >= 0) {
776 			khd->batching++;
777 			rq_set_domain_token(rq, nr);
778 			list_del_init(&rq->queuelist);
779 			return rq;
780 		} else {
781 			trace_kyber_throttled(kqd->dev,
782 					      kyber_domain_names[khd->cur_domain]);
783 		}
784 	} else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) {
785 		nr = kyber_get_domain_token(kqd, khd, hctx);
786 		if (nr >= 0) {
787 			kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs);
788 			rq = list_first_entry(rqs, struct request, queuelist);
789 			khd->batching++;
790 			rq_set_domain_token(rq, nr);
791 			list_del_init(&rq->queuelist);
792 			return rq;
793 		} else {
794 			trace_kyber_throttled(kqd->dev,
795 					      kyber_domain_names[khd->cur_domain]);
796 		}
797 	}
798 
799 	/* There were either no pending requests or no tokens. */
800 	return NULL;
801 }
802 
kyber_dispatch_request(struct blk_mq_hw_ctx * hctx)803 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
804 {
805 	struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
806 	struct kyber_hctx_data *khd = hctx->sched_data;
807 	struct request *rq;
808 	int i;
809 
810 	spin_lock(&khd->lock);
811 
812 	/*
813 	 * First, if we are still entitled to batch, try to dispatch a request
814 	 * from the batch.
815 	 */
816 	if (khd->batching < kyber_batch_size[khd->cur_domain]) {
817 		rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
818 		if (rq)
819 			goto out;
820 	}
821 
822 	/*
823 	 * Either,
824 	 * 1. We were no longer entitled to a batch.
825 	 * 2. The domain we were batching didn't have any requests.
826 	 * 3. The domain we were batching was out of tokens.
827 	 *
828 	 * Start another batch. Note that this wraps back around to the original
829 	 * domain if no other domains have requests or tokens.
830 	 */
831 	khd->batching = 0;
832 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
833 		if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
834 			khd->cur_domain = 0;
835 		else
836 			khd->cur_domain++;
837 
838 		rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
839 		if (rq)
840 			goto out;
841 	}
842 
843 	rq = NULL;
844 out:
845 	spin_unlock(&khd->lock);
846 	return rq;
847 }
848 
kyber_has_work(struct blk_mq_hw_ctx * hctx)849 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
850 {
851 	struct kyber_hctx_data *khd = hctx->sched_data;
852 	int i;
853 
854 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
855 		if (!list_empty_careful(&khd->rqs[i]) ||
856 		    sbitmap_any_bit_set(&khd->kcq_map[i]))
857 			return true;
858 	}
859 
860 	return false;
861 }
862 
863 #define KYBER_LAT_SHOW_STORE(domain, name)				\
864 static ssize_t kyber_##name##_lat_show(struct elevator_queue *e,	\
865 				       char *page)			\
866 {									\
867 	struct kyber_queue_data *kqd = e->elevator_data;		\
868 									\
869 	return sprintf(page, "%llu\n", kqd->latency_targets[domain]);	\
870 }									\
871 									\
872 static ssize_t kyber_##name##_lat_store(struct elevator_queue *e,	\
873 					const char *page, size_t count)	\
874 {									\
875 	struct kyber_queue_data *kqd = e->elevator_data;		\
876 	unsigned long long nsec;					\
877 	int ret;							\
878 									\
879 	ret = kstrtoull(page, 10, &nsec);				\
880 	if (ret)							\
881 		return ret;						\
882 									\
883 	kqd->latency_targets[domain] = nsec;				\
884 									\
885 	return count;							\
886 }
887 KYBER_LAT_SHOW_STORE(KYBER_READ, read);
888 KYBER_LAT_SHOW_STORE(KYBER_WRITE, write);
889 #undef KYBER_LAT_SHOW_STORE
890 
891 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
892 static struct elv_fs_entry kyber_sched_attrs[] = {
893 	KYBER_LAT_ATTR(read),
894 	KYBER_LAT_ATTR(write),
895 	__ATTR_NULL
896 };
897 #undef KYBER_LAT_ATTR
898 
899 #ifdef CONFIG_BLK_DEBUG_FS
900 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name)			\
901 static int kyber_##name##_tokens_show(void *data, struct seq_file *m)	\
902 {									\
903 	struct request_queue *q = data;					\
904 	struct kyber_queue_data *kqd = q->elevator->elevator_data;	\
905 									\
906 	sbitmap_queue_show(&kqd->domain_tokens[domain], m);		\
907 	return 0;							\
908 }									\
909 									\
910 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos)	\
911 	__acquires(&khd->lock)						\
912 {									\
913 	struct blk_mq_hw_ctx *hctx = m->private;			\
914 	struct kyber_hctx_data *khd = hctx->sched_data;			\
915 									\
916 	spin_lock(&khd->lock);						\
917 	return seq_list_start(&khd->rqs[domain], *pos);			\
918 }									\
919 									\
920 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v,	\
921 				     loff_t *pos)			\
922 {									\
923 	struct blk_mq_hw_ctx *hctx = m->private;			\
924 	struct kyber_hctx_data *khd = hctx->sched_data;			\
925 									\
926 	return seq_list_next(v, &khd->rqs[domain], pos);		\
927 }									\
928 									\
929 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v)	\
930 	__releases(&khd->lock)						\
931 {									\
932 	struct blk_mq_hw_ctx *hctx = m->private;			\
933 	struct kyber_hctx_data *khd = hctx->sched_data;			\
934 									\
935 	spin_unlock(&khd->lock);					\
936 }									\
937 									\
938 static const struct seq_operations kyber_##name##_rqs_seq_ops = {	\
939 	.start	= kyber_##name##_rqs_start,				\
940 	.next	= kyber_##name##_rqs_next,				\
941 	.stop	= kyber_##name##_rqs_stop,				\
942 	.show	= blk_mq_debugfs_rq_show,				\
943 };									\
944 									\
945 static int kyber_##name##_waiting_show(void *data, struct seq_file *m)	\
946 {									\
947 	struct blk_mq_hw_ctx *hctx = data;				\
948 	struct kyber_hctx_data *khd = hctx->sched_data;			\
949 	wait_queue_entry_t *wait = &khd->domain_wait[domain].wait;	\
950 									\
951 	seq_printf(m, "%d\n", !list_empty_careful(&wait->entry));	\
952 	return 0;							\
953 }
KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ,read)954 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
955 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write)
956 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
957 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
958 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
959 
960 static int kyber_async_depth_show(void *data, struct seq_file *m)
961 {
962 	struct request_queue *q = data;
963 	struct kyber_queue_data *kqd = q->elevator->elevator_data;
964 
965 	seq_printf(m, "%u\n", kqd->async_depth);
966 	return 0;
967 }
968 
kyber_cur_domain_show(void * data,struct seq_file * m)969 static int kyber_cur_domain_show(void *data, struct seq_file *m)
970 {
971 	struct blk_mq_hw_ctx *hctx = data;
972 	struct kyber_hctx_data *khd = hctx->sched_data;
973 
974 	seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]);
975 	return 0;
976 }
977 
kyber_batching_show(void * data,struct seq_file * m)978 static int kyber_batching_show(void *data, struct seq_file *m)
979 {
980 	struct blk_mq_hw_ctx *hctx = data;
981 	struct kyber_hctx_data *khd = hctx->sched_data;
982 
983 	seq_printf(m, "%u\n", khd->batching);
984 	return 0;
985 }
986 
987 #define KYBER_QUEUE_DOMAIN_ATTRS(name)	\
988 	{#name "_tokens", 0400, kyber_##name##_tokens_show}
989 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
990 	KYBER_QUEUE_DOMAIN_ATTRS(read),
991 	KYBER_QUEUE_DOMAIN_ATTRS(write),
992 	KYBER_QUEUE_DOMAIN_ATTRS(discard),
993 	KYBER_QUEUE_DOMAIN_ATTRS(other),
994 	{"async_depth", 0400, kyber_async_depth_show},
995 	{},
996 };
997 #undef KYBER_QUEUE_DOMAIN_ATTRS
998 
999 #define KYBER_HCTX_DOMAIN_ATTRS(name)					\
1000 	{#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops},	\
1001 	{#name "_waiting", 0400, kyber_##name##_waiting_show}
1002 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
1003 	KYBER_HCTX_DOMAIN_ATTRS(read),
1004 	KYBER_HCTX_DOMAIN_ATTRS(write),
1005 	KYBER_HCTX_DOMAIN_ATTRS(discard),
1006 	KYBER_HCTX_DOMAIN_ATTRS(other),
1007 	{"cur_domain", 0400, kyber_cur_domain_show},
1008 	{"batching", 0400, kyber_batching_show},
1009 	{},
1010 };
1011 #undef KYBER_HCTX_DOMAIN_ATTRS
1012 #endif
1013 
1014 static struct elevator_type kyber_sched = {
1015 	.ops = {
1016 		.init_sched = kyber_init_sched,
1017 		.exit_sched = kyber_exit_sched,
1018 		.init_hctx = kyber_init_hctx,
1019 		.exit_hctx = kyber_exit_hctx,
1020 		.limit_depth = kyber_limit_depth,
1021 		.bio_merge = kyber_bio_merge,
1022 		.prepare_request = kyber_prepare_request,
1023 		.insert_requests = kyber_insert_requests,
1024 		.finish_request = kyber_finish_request,
1025 		.requeue_request = kyber_finish_request,
1026 		.completed_request = kyber_completed_request,
1027 		.dispatch_request = kyber_dispatch_request,
1028 		.has_work = kyber_has_work,
1029 		.depth_updated = kyber_depth_updated,
1030 	},
1031 #ifdef CONFIG_BLK_DEBUG_FS
1032 	.queue_debugfs_attrs = kyber_queue_debugfs_attrs,
1033 	.hctx_debugfs_attrs = kyber_hctx_debugfs_attrs,
1034 #endif
1035 	.elevator_attrs = kyber_sched_attrs,
1036 	.elevator_name = "kyber",
1037 	.elevator_owner = THIS_MODULE,
1038 };
1039 
kyber_init(void)1040 static int __init kyber_init(void)
1041 {
1042 	return elv_register(&kyber_sched);
1043 }
1044 
kyber_exit(void)1045 static void __exit kyber_exit(void)
1046 {
1047 	elv_unregister(&kyber_sched);
1048 }
1049 
1050 module_init(kyber_init);
1051 module_exit(kyber_exit);
1052 
1053 MODULE_AUTHOR("Omar Sandoval");
1054 MODULE_LICENSE("GPL");
1055 MODULE_DESCRIPTION("Kyber I/O scheduler");
1056