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/module.h>
13 #include <linux/sbitmap.h>
14 
15 #include <trace/events/block.h>
16 
17 #include "elevator.h"
18 #include "blk.h"
19 #include "blk-mq.h"
20 #include "blk-mq-debugfs.h"
21 #include "blk-mq-sched.h"
22 #include "blk-mq-tag.h"
23 
24 #define CREATE_TRACE_POINTS
25 #include <trace/events/kyber.h>
26 
27 /*
28  * Scheduling domains: the device is divided into multiple domains based on the
29  * request type.
30  */
31 enum {
32 	KYBER_READ,
33 	KYBER_WRITE,
34 	KYBER_DISCARD,
35 	KYBER_OTHER,
36 	KYBER_NUM_DOMAINS,
37 };
38 
39 static const char *kyber_domain_names[] = {
40 	[KYBER_READ] = "READ",
41 	[KYBER_WRITE] = "WRITE",
42 	[KYBER_DISCARD] = "DISCARD",
43 	[KYBER_OTHER] = "OTHER",
44 };
45 
46 enum {
47 	/*
48 	 * In order to prevent starvation of synchronous requests by a flood of
49 	 * asynchronous requests, we reserve 25% of requests for synchronous
50 	 * operations.
51 	 */
52 	KYBER_ASYNC_PERCENT = 75,
53 };
54 
55 /*
56  * Maximum device-wide depth for each scheduling domain.
57  *
58  * Even for fast devices with lots of tags like NVMe, you can saturate the
59  * device with only a fraction of the maximum possible queue depth. So, we cap
60  * these to a reasonable value.
61  */
62 static const unsigned int kyber_depth[] = {
63 	[KYBER_READ] = 256,
64 	[KYBER_WRITE] = 128,
65 	[KYBER_DISCARD] = 64,
66 	[KYBER_OTHER] = 16,
67 };
68 
69 /*
70  * Default latency targets for each scheduling domain.
71  */
72 static const u64 kyber_latency_targets[] = {
73 	[KYBER_READ] = 2ULL * NSEC_PER_MSEC,
74 	[KYBER_WRITE] = 10ULL * NSEC_PER_MSEC,
75 	[KYBER_DISCARD] = 5ULL * NSEC_PER_SEC,
76 };
77 
78 /*
79  * Batch size (number of requests we'll dispatch in a row) for each scheduling
80  * domain.
81  */
82 static const unsigned int kyber_batch_size[] = {
83 	[KYBER_READ] = 16,
84 	[KYBER_WRITE] = 8,
85 	[KYBER_DISCARD] = 1,
86 	[KYBER_OTHER] = 1,
87 };
88 
89 /*
90  * Requests latencies are recorded in a histogram with buckets defined relative
91  * to the target latency:
92  *
93  * <= 1/4 * target latency
94  * <= 1/2 * target latency
95  * <= 3/4 * target latency
96  * <= target latency
97  * <= 1 1/4 * target latency
98  * <= 1 1/2 * target latency
99  * <= 1 3/4 * target latency
100  * > 1 3/4 * target latency
101  */
102 enum {
103 	/*
104 	 * The width of the latency histogram buckets is
105 	 * 1 / (1 << KYBER_LATENCY_SHIFT) * target latency.
106 	 */
107 	KYBER_LATENCY_SHIFT = 2,
108 	/*
109 	 * The first (1 << KYBER_LATENCY_SHIFT) buckets are <= target latency,
110 	 * thus, "good".
111 	 */
112 	KYBER_GOOD_BUCKETS = 1 << KYBER_LATENCY_SHIFT,
113 	/* There are also (1 << KYBER_LATENCY_SHIFT) "bad" buckets. */
114 	KYBER_LATENCY_BUCKETS = 2 << KYBER_LATENCY_SHIFT,
115 };
116 
117 /*
118  * We measure both the total latency and the I/O latency (i.e., latency after
119  * submitting to the device).
120  */
121 enum {
122 	KYBER_TOTAL_LATENCY,
123 	KYBER_IO_LATENCY,
124 };
125 
126 static const char *kyber_latency_type_names[] = {
127 	[KYBER_TOTAL_LATENCY] = "total",
128 	[KYBER_IO_LATENCY] = "I/O",
129 };
130 
131 /*
132  * Per-cpu latency histograms: total latency and I/O latency for each scheduling
133  * domain except for KYBER_OTHER.
134  */
135 struct kyber_cpu_latency {
136 	atomic_t buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
137 };
138 
139 /*
140  * There is a same mapping between ctx & hctx and kcq & khd,
141  * we use request->mq_ctx->index_hw to index the kcq in khd.
142  */
143 struct kyber_ctx_queue {
144 	/*
145 	 * Used to ensure operations on rq_list and kcq_map to be an atmoic one.
146 	 * Also protect the rqs on rq_list when merge.
147 	 */
148 	spinlock_t lock;
149 	struct list_head rq_list[KYBER_NUM_DOMAINS];
150 } ____cacheline_aligned_in_smp;
151 
152 struct kyber_queue_data {
153 	struct request_queue *q;
154 	dev_t dev;
155 
156 	/*
157 	 * Each scheduling domain has a limited number of in-flight requests
158 	 * device-wide, limited by these tokens.
159 	 */
160 	struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
161 
162 	/*
163 	 * Async request percentage, converted to per-word depth for
164 	 * sbitmap_get_shallow().
165 	 */
166 	unsigned int async_depth;
167 
168 	struct kyber_cpu_latency __percpu *cpu_latency;
169 
170 	/* Timer for stats aggregation and adjusting domain tokens. */
171 	struct timer_list timer;
172 
173 	unsigned int latency_buckets[KYBER_OTHER][2][KYBER_LATENCY_BUCKETS];
174 
175 	unsigned long latency_timeout[KYBER_OTHER];
176 
177 	int domain_p99[KYBER_OTHER];
178 
179 	/* Target latencies in nanoseconds. */
180 	u64 latency_targets[KYBER_OTHER];
181 };
182 
183 struct kyber_hctx_data {
184 	spinlock_t lock;
185 	struct list_head rqs[KYBER_NUM_DOMAINS];
186 	unsigned int cur_domain;
187 	unsigned int batching;
188 	struct kyber_ctx_queue *kcqs;
189 	struct sbitmap kcq_map[KYBER_NUM_DOMAINS];
190 	struct sbq_wait domain_wait[KYBER_NUM_DOMAINS];
191 	struct sbq_wait_state *domain_ws[KYBER_NUM_DOMAINS];
192 	atomic_t wait_index[KYBER_NUM_DOMAINS];
193 };
194 
195 static int kyber_domain_wake(wait_queue_entry_t *wait, unsigned mode, int flags,
196 			     void *key);
197 
kyber_sched_domain(blk_opf_t opf)198 static unsigned int kyber_sched_domain(blk_opf_t opf)
199 {
200 	switch (opf & REQ_OP_MASK) {
201 	case REQ_OP_READ:
202 		return KYBER_READ;
203 	case REQ_OP_WRITE:
204 		return KYBER_WRITE;
205 	case REQ_OP_DISCARD:
206 		return KYBER_DISCARD;
207 	default:
208 		return KYBER_OTHER;
209 	}
210 }
211 
flush_latency_buckets(struct kyber_queue_data * kqd,struct kyber_cpu_latency * cpu_latency,unsigned int sched_domain,unsigned int type)212 static void flush_latency_buckets(struct kyber_queue_data *kqd,
213 				  struct kyber_cpu_latency *cpu_latency,
214 				  unsigned int sched_domain, unsigned int type)
215 {
216 	unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
217 	atomic_t *cpu_buckets = cpu_latency->buckets[sched_domain][type];
218 	unsigned int bucket;
219 
220 	for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
221 		buckets[bucket] += atomic_xchg(&cpu_buckets[bucket], 0);
222 }
223 
224 /*
225  * Calculate the histogram bucket with the given percentile rank, or -1 if there
226  * aren't enough samples yet.
227  */
calculate_percentile(struct kyber_queue_data * kqd,unsigned int sched_domain,unsigned int type,unsigned int percentile)228 static int calculate_percentile(struct kyber_queue_data *kqd,
229 				unsigned int sched_domain, unsigned int type,
230 				unsigned int percentile)
231 {
232 	unsigned int *buckets = kqd->latency_buckets[sched_domain][type];
233 	unsigned int bucket, samples = 0, percentile_samples;
234 
235 	for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS; bucket++)
236 		samples += buckets[bucket];
237 
238 	if (!samples)
239 		return -1;
240 
241 	/*
242 	 * We do the calculation once we have 500 samples or one second passes
243 	 * since the first sample was recorded, whichever comes first.
244 	 */
245 	if (!kqd->latency_timeout[sched_domain])
246 		kqd->latency_timeout[sched_domain] = max(jiffies + HZ, 1UL);
247 	if (samples < 500 &&
248 	    time_is_after_jiffies(kqd->latency_timeout[sched_domain])) {
249 		return -1;
250 	}
251 	kqd->latency_timeout[sched_domain] = 0;
252 
253 	percentile_samples = DIV_ROUND_UP(samples * percentile, 100);
254 	for (bucket = 0; bucket < KYBER_LATENCY_BUCKETS - 1; bucket++) {
255 		if (buckets[bucket] >= percentile_samples)
256 			break;
257 		percentile_samples -= buckets[bucket];
258 	}
259 	memset(buckets, 0, sizeof(kqd->latency_buckets[sched_domain][type]));
260 
261 	trace_kyber_latency(kqd->dev, kyber_domain_names[sched_domain],
262 			    kyber_latency_type_names[type], percentile,
263 			    bucket + 1, 1 << KYBER_LATENCY_SHIFT, samples);
264 
265 	return bucket;
266 }
267 
kyber_resize_domain(struct kyber_queue_data * kqd,unsigned int sched_domain,unsigned int depth)268 static void kyber_resize_domain(struct kyber_queue_data *kqd,
269 				unsigned int sched_domain, unsigned int depth)
270 {
271 	depth = clamp(depth, 1U, kyber_depth[sched_domain]);
272 	if (depth != kqd->domain_tokens[sched_domain].sb.depth) {
273 		sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
274 		trace_kyber_adjust(kqd->dev, kyber_domain_names[sched_domain],
275 				   depth);
276 	}
277 }
278 
kyber_timer_fn(struct timer_list * t)279 static void kyber_timer_fn(struct timer_list *t)
280 {
281 	struct kyber_queue_data *kqd = from_timer(kqd, t, timer);
282 	unsigned int sched_domain;
283 	int cpu;
284 	bool bad = false;
285 
286 	/* Sum all of the per-cpu latency histograms. */
287 	for_each_online_cpu(cpu) {
288 		struct kyber_cpu_latency *cpu_latency;
289 
290 		cpu_latency = per_cpu_ptr(kqd->cpu_latency, cpu);
291 		for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
292 			flush_latency_buckets(kqd, cpu_latency, sched_domain,
293 					      KYBER_TOTAL_LATENCY);
294 			flush_latency_buckets(kqd, cpu_latency, sched_domain,
295 					      KYBER_IO_LATENCY);
296 		}
297 	}
298 
299 	/*
300 	 * Check if any domains have a high I/O latency, which might indicate
301 	 * congestion in the device. Note that we use the p90; we don't want to
302 	 * be too sensitive to outliers here.
303 	 */
304 	for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
305 		int p90;
306 
307 		p90 = calculate_percentile(kqd, sched_domain, KYBER_IO_LATENCY,
308 					   90);
309 		if (p90 >= KYBER_GOOD_BUCKETS)
310 			bad = true;
311 	}
312 
313 	/*
314 	 * Adjust the scheduling domain depths. If we determined that there was
315 	 * congestion, we throttle all domains with good latencies. Either way,
316 	 * we ease up on throttling domains with bad latencies.
317 	 */
318 	for (sched_domain = 0; sched_domain < KYBER_OTHER; sched_domain++) {
319 		unsigned int orig_depth, depth;
320 		int p99;
321 
322 		p99 = calculate_percentile(kqd, sched_domain,
323 					   KYBER_TOTAL_LATENCY, 99);
324 		/*
325 		 * This is kind of subtle: different domains will not
326 		 * necessarily have enough samples to calculate the latency
327 		 * percentiles during the same window, so we have to remember
328 		 * the p99 for the next time we observe congestion; once we do,
329 		 * we don't want to throttle again until we get more data, so we
330 		 * reset it to -1.
331 		 */
332 		if (bad) {
333 			if (p99 < 0)
334 				p99 = kqd->domain_p99[sched_domain];
335 			kqd->domain_p99[sched_domain] = -1;
336 		} else if (p99 >= 0) {
337 			kqd->domain_p99[sched_domain] = p99;
338 		}
339 		if (p99 < 0)
340 			continue;
341 
342 		/*
343 		 * If this domain has bad latency, throttle less. Otherwise,
344 		 * throttle more iff we determined that there is congestion.
345 		 *
346 		 * The new depth is scaled linearly with the p99 latency vs the
347 		 * latency target. E.g., if the p99 is 3/4 of the target, then
348 		 * we throttle down to 3/4 of the current depth, and if the p99
349 		 * is 2x the target, then we double the depth.
350 		 */
351 		if (bad || p99 >= KYBER_GOOD_BUCKETS) {
352 			orig_depth = kqd->domain_tokens[sched_domain].sb.depth;
353 			depth = (orig_depth * (p99 + 1)) >> KYBER_LATENCY_SHIFT;
354 			kyber_resize_domain(kqd, sched_domain, depth);
355 		}
356 	}
357 }
358 
kyber_queue_data_alloc(struct request_queue * q)359 static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
360 {
361 	struct kyber_queue_data *kqd;
362 	int ret = -ENOMEM;
363 	int i;
364 
365 	kqd = kzalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
366 	if (!kqd)
367 		goto err;
368 
369 	kqd->q = q;
370 	kqd->dev = disk_devt(q->disk);
371 
372 	kqd->cpu_latency = alloc_percpu_gfp(struct kyber_cpu_latency,
373 					    GFP_KERNEL | __GFP_ZERO);
374 	if (!kqd->cpu_latency)
375 		goto err_kqd;
376 
377 	timer_setup(&kqd->timer, kyber_timer_fn, 0);
378 
379 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
380 		WARN_ON(!kyber_depth[i]);
381 		WARN_ON(!kyber_batch_size[i]);
382 		ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
383 					      kyber_depth[i], -1, false,
384 					      GFP_KERNEL, q->node);
385 		if (ret) {
386 			while (--i >= 0)
387 				sbitmap_queue_free(&kqd->domain_tokens[i]);
388 			goto err_buckets;
389 		}
390 	}
391 
392 	for (i = 0; i < KYBER_OTHER; i++) {
393 		kqd->domain_p99[i] = -1;
394 		kqd->latency_targets[i] = kyber_latency_targets[i];
395 	}
396 
397 	return kqd;
398 
399 err_buckets:
400 	free_percpu(kqd->cpu_latency);
401 err_kqd:
402 	kfree(kqd);
403 err:
404 	return ERR_PTR(ret);
405 }
406 
kyber_init_sched(struct request_queue * q,struct elevator_type * e)407 static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
408 {
409 	struct kyber_queue_data *kqd;
410 	struct elevator_queue *eq;
411 
412 	eq = elevator_alloc(q, e);
413 	if (!eq)
414 		return -ENOMEM;
415 
416 	kqd = kyber_queue_data_alloc(q);
417 	if (IS_ERR(kqd)) {
418 		kobject_put(&eq->kobj);
419 		return PTR_ERR(kqd);
420 	}
421 
422 	blk_stat_enable_accounting(q);
423 
424 	blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);
425 
426 	eq->elevator_data = kqd;
427 	q->elevator = eq;
428 
429 	return 0;
430 }
431 
kyber_exit_sched(struct elevator_queue * e)432 static void kyber_exit_sched(struct elevator_queue *e)
433 {
434 	struct kyber_queue_data *kqd = e->elevator_data;
435 	int i;
436 
437 	timer_shutdown_sync(&kqd->timer);
438 	blk_stat_disable_accounting(kqd->q);
439 
440 	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
441 		sbitmap_queue_free(&kqd->domain_tokens[i]);
442 	free_percpu(kqd->cpu_latency);
443 	kfree(kqd);
444 }
445 
kyber_ctx_queue_init(struct kyber_ctx_queue * kcq)446 static void kyber_ctx_queue_init(struct kyber_ctx_queue *kcq)
447 {
448 	unsigned int i;
449 
450 	spin_lock_init(&kcq->lock);
451 	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
452 		INIT_LIST_HEAD(&kcq->rq_list[i]);
453 }
454 
kyber_depth_updated(struct blk_mq_hw_ctx * hctx)455 static void kyber_depth_updated(struct blk_mq_hw_ctx *hctx)
456 {
457 	struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
458 	struct blk_mq_tags *tags = hctx->sched_tags;
459 	unsigned int shift = tags->bitmap_tags.sb.shift;
460 
461 	kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
462 
463 	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, kqd->async_depth);
464 }
465 
kyber_init_hctx(struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)466 static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
467 {
468 	struct kyber_hctx_data *khd;
469 	int i;
470 
471 	khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
472 	if (!khd)
473 		return -ENOMEM;
474 
475 	khd->kcqs = kmalloc_array_node(hctx->nr_ctx,
476 				       sizeof(struct kyber_ctx_queue),
477 				       GFP_KERNEL, hctx->numa_node);
478 	if (!khd->kcqs)
479 		goto err_khd;
480 
481 	for (i = 0; i < hctx->nr_ctx; i++)
482 		kyber_ctx_queue_init(&khd->kcqs[i]);
483 
484 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
485 		if (sbitmap_init_node(&khd->kcq_map[i], hctx->nr_ctx,
486 				      ilog2(8), GFP_KERNEL, hctx->numa_node,
487 				      false, false)) {
488 			while (--i >= 0)
489 				sbitmap_free(&khd->kcq_map[i]);
490 			goto err_kcqs;
491 		}
492 	}
493 
494 	spin_lock_init(&khd->lock);
495 
496 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
497 		INIT_LIST_HEAD(&khd->rqs[i]);
498 		khd->domain_wait[i].sbq = NULL;
499 		init_waitqueue_func_entry(&khd->domain_wait[i].wait,
500 					  kyber_domain_wake);
501 		khd->domain_wait[i].wait.private = hctx;
502 		INIT_LIST_HEAD(&khd->domain_wait[i].wait.entry);
503 		atomic_set(&khd->wait_index[i], 0);
504 	}
505 
506 	khd->cur_domain = 0;
507 	khd->batching = 0;
508 
509 	hctx->sched_data = khd;
510 	kyber_depth_updated(hctx);
511 
512 	return 0;
513 
514 err_kcqs:
515 	kfree(khd->kcqs);
516 err_khd:
517 	kfree(khd);
518 	return -ENOMEM;
519 }
520 
kyber_exit_hctx(struct blk_mq_hw_ctx * hctx,unsigned int hctx_idx)521 static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
522 {
523 	struct kyber_hctx_data *khd = hctx->sched_data;
524 	int i;
525 
526 	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
527 		sbitmap_free(&khd->kcq_map[i]);
528 	kfree(khd->kcqs);
529 	kfree(hctx->sched_data);
530 }
531 
rq_get_domain_token(struct request * rq)532 static int rq_get_domain_token(struct request *rq)
533 {
534 	return (long)rq->elv.priv[0];
535 }
536 
rq_set_domain_token(struct request * rq,int token)537 static void rq_set_domain_token(struct request *rq, int token)
538 {
539 	rq->elv.priv[0] = (void *)(long)token;
540 }
541 
rq_clear_domain_token(struct kyber_queue_data * kqd,struct request * rq)542 static void rq_clear_domain_token(struct kyber_queue_data *kqd,
543 				  struct request *rq)
544 {
545 	unsigned int sched_domain;
546 	int nr;
547 
548 	nr = rq_get_domain_token(rq);
549 	if (nr != -1) {
550 		sched_domain = kyber_sched_domain(rq->cmd_flags);
551 		sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
552 				    rq->mq_ctx->cpu);
553 	}
554 }
555 
kyber_limit_depth(blk_opf_t opf,struct blk_mq_alloc_data * data)556 static void kyber_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
557 {
558 	/*
559 	 * We use the scheduler tags as per-hardware queue queueing tokens.
560 	 * Async requests can be limited at this stage.
561 	 */
562 	if (!op_is_sync(opf)) {
563 		struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
564 
565 		data->shallow_depth = kqd->async_depth;
566 	}
567 }
568 
kyber_bio_merge(struct request_queue * q,struct bio * bio,unsigned int nr_segs)569 static bool kyber_bio_merge(struct request_queue *q, struct bio *bio,
570 		unsigned int nr_segs)
571 {
572 	struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
573 	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, bio->bi_opf, ctx);
574 	struct kyber_hctx_data *khd = hctx->sched_data;
575 	struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]];
576 	unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);
577 	struct list_head *rq_list = &kcq->rq_list[sched_domain];
578 	bool merged;
579 
580 	spin_lock(&kcq->lock);
581 	merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs);
582 	spin_unlock(&kcq->lock);
583 
584 	return merged;
585 }
586 
kyber_prepare_request(struct request * rq)587 static void kyber_prepare_request(struct request *rq)
588 {
589 	rq_set_domain_token(rq, -1);
590 }
591 
kyber_insert_requests(struct blk_mq_hw_ctx * hctx,struct list_head * rq_list,bool at_head)592 static void kyber_insert_requests(struct blk_mq_hw_ctx *hctx,
593 				  struct list_head *rq_list, bool at_head)
594 {
595 	struct kyber_hctx_data *khd = hctx->sched_data;
596 	struct request *rq, *next;
597 
598 	list_for_each_entry_safe(rq, next, rq_list, queuelist) {
599 		unsigned int sched_domain = kyber_sched_domain(rq->cmd_flags);
600 		struct kyber_ctx_queue *kcq = &khd->kcqs[rq->mq_ctx->index_hw[hctx->type]];
601 		struct list_head *head = &kcq->rq_list[sched_domain];
602 
603 		spin_lock(&kcq->lock);
604 		trace_block_rq_insert(rq);
605 		if (at_head)
606 			list_move(&rq->queuelist, head);
607 		else
608 			list_move_tail(&rq->queuelist, head);
609 		sbitmap_set_bit(&khd->kcq_map[sched_domain],
610 				rq->mq_ctx->index_hw[hctx->type]);
611 		spin_unlock(&kcq->lock);
612 	}
613 }
614 
kyber_finish_request(struct request * rq)615 static void kyber_finish_request(struct request *rq)
616 {
617 	struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
618 
619 	rq_clear_domain_token(kqd, rq);
620 }
621 
add_latency_sample(struct kyber_cpu_latency * cpu_latency,unsigned int sched_domain,unsigned int type,u64 target,u64 latency)622 static void add_latency_sample(struct kyber_cpu_latency *cpu_latency,
623 			       unsigned int sched_domain, unsigned int type,
624 			       u64 target, u64 latency)
625 {
626 	unsigned int bucket;
627 	u64 divisor;
628 
629 	if (latency > 0) {
630 		divisor = max_t(u64, target >> KYBER_LATENCY_SHIFT, 1);
631 		bucket = min_t(unsigned int, div64_u64(latency - 1, divisor),
632 			       KYBER_LATENCY_BUCKETS - 1);
633 	} else {
634 		bucket = 0;
635 	}
636 
637 	atomic_inc(&cpu_latency->buckets[sched_domain][type][bucket]);
638 }
639 
kyber_completed_request(struct request * rq,u64 now)640 static void kyber_completed_request(struct request *rq, u64 now)
641 {
642 	struct kyber_queue_data *kqd = rq->q->elevator->elevator_data;
643 	struct kyber_cpu_latency *cpu_latency;
644 	unsigned int sched_domain;
645 	u64 target;
646 
647 	sched_domain = kyber_sched_domain(rq->cmd_flags);
648 	if (sched_domain == KYBER_OTHER)
649 		return;
650 
651 	cpu_latency = get_cpu_ptr(kqd->cpu_latency);
652 	target = kqd->latency_targets[sched_domain];
653 	add_latency_sample(cpu_latency, sched_domain, KYBER_TOTAL_LATENCY,
654 			   target, now - rq->start_time_ns);
655 	add_latency_sample(cpu_latency, sched_domain, KYBER_IO_LATENCY, target,
656 			   now - rq->io_start_time_ns);
657 	put_cpu_ptr(kqd->cpu_latency);
658 
659 	timer_reduce(&kqd->timer, jiffies + HZ / 10);
660 }
661 
662 struct flush_kcq_data {
663 	struct kyber_hctx_data *khd;
664 	unsigned int sched_domain;
665 	struct list_head *list;
666 };
667 
flush_busy_kcq(struct sbitmap * sb,unsigned int bitnr,void * data)668 static bool flush_busy_kcq(struct sbitmap *sb, unsigned int bitnr, void *data)
669 {
670 	struct flush_kcq_data *flush_data = data;
671 	struct kyber_ctx_queue *kcq = &flush_data->khd->kcqs[bitnr];
672 
673 	spin_lock(&kcq->lock);
674 	list_splice_tail_init(&kcq->rq_list[flush_data->sched_domain],
675 			      flush_data->list);
676 	sbitmap_clear_bit(sb, bitnr);
677 	spin_unlock(&kcq->lock);
678 
679 	return true;
680 }
681 
kyber_flush_busy_kcqs(struct kyber_hctx_data * khd,unsigned int sched_domain,struct list_head * list)682 static void kyber_flush_busy_kcqs(struct kyber_hctx_data *khd,
683 				  unsigned int sched_domain,
684 				  struct list_head *list)
685 {
686 	struct flush_kcq_data data = {
687 		.khd = khd,
688 		.sched_domain = sched_domain,
689 		.list = list,
690 	};
691 
692 	sbitmap_for_each_set(&khd->kcq_map[sched_domain],
693 			     flush_busy_kcq, &data);
694 }
695 
kyber_domain_wake(wait_queue_entry_t * wqe,unsigned mode,int flags,void * key)696 static int kyber_domain_wake(wait_queue_entry_t *wqe, unsigned mode, int flags,
697 			     void *key)
698 {
699 	struct blk_mq_hw_ctx *hctx = READ_ONCE(wqe->private);
700 	struct sbq_wait *wait = container_of(wqe, struct sbq_wait, wait);
701 
702 	sbitmap_del_wait_queue(wait);
703 	blk_mq_run_hw_queue(hctx, true);
704 	return 1;
705 }
706 
kyber_get_domain_token(struct kyber_queue_data * kqd,struct kyber_hctx_data * khd,struct blk_mq_hw_ctx * hctx)707 static int kyber_get_domain_token(struct kyber_queue_data *kqd,
708 				  struct kyber_hctx_data *khd,
709 				  struct blk_mq_hw_ctx *hctx)
710 {
711 	unsigned int sched_domain = khd->cur_domain;
712 	struct sbitmap_queue *domain_tokens = &kqd->domain_tokens[sched_domain];
713 	struct sbq_wait *wait = &khd->domain_wait[sched_domain];
714 	struct sbq_wait_state *ws;
715 	int nr;
716 
717 	nr = __sbitmap_queue_get(domain_tokens);
718 
719 	/*
720 	 * If we failed to get a domain token, make sure the hardware queue is
721 	 * run when one becomes available. Note that this is serialized on
722 	 * khd->lock, but we still need to be careful about the waker.
723 	 */
724 	if (nr < 0 && list_empty_careful(&wait->wait.entry)) {
725 		ws = sbq_wait_ptr(domain_tokens,
726 				  &khd->wait_index[sched_domain]);
727 		khd->domain_ws[sched_domain] = ws;
728 		sbitmap_add_wait_queue(domain_tokens, ws, wait);
729 
730 		/*
731 		 * Try again in case a token was freed before we got on the wait
732 		 * queue.
733 		 */
734 		nr = __sbitmap_queue_get(domain_tokens);
735 	}
736 
737 	/*
738 	 * If we got a token while we were on the wait queue, remove ourselves
739 	 * from the wait queue to ensure that all wake ups make forward
740 	 * progress. It's possible that the waker already deleted the entry
741 	 * between the !list_empty_careful() check and us grabbing the lock, but
742 	 * list_del_init() is okay with that.
743 	 */
744 	if (nr >= 0 && !list_empty_careful(&wait->wait.entry)) {
745 		ws = khd->domain_ws[sched_domain];
746 		spin_lock_irq(&ws->wait.lock);
747 		sbitmap_del_wait_queue(wait);
748 		spin_unlock_irq(&ws->wait.lock);
749 	}
750 
751 	return nr;
752 }
753 
754 static struct request *
kyber_dispatch_cur_domain(struct kyber_queue_data * kqd,struct kyber_hctx_data * khd,struct blk_mq_hw_ctx * hctx)755 kyber_dispatch_cur_domain(struct kyber_queue_data *kqd,
756 			  struct kyber_hctx_data *khd,
757 			  struct blk_mq_hw_ctx *hctx)
758 {
759 	struct list_head *rqs;
760 	struct request *rq;
761 	int nr;
762 
763 	rqs = &khd->rqs[khd->cur_domain];
764 
765 	/*
766 	 * If we already have a flushed request, then we just need to get a
767 	 * token for it. Otherwise, if there are pending requests in the kcqs,
768 	 * flush the kcqs, but only if we can get a token. If not, we should
769 	 * leave the requests in the kcqs so that they can be merged. Note that
770 	 * khd->lock serializes the flushes, so if we observed any bit set in
771 	 * the kcq_map, we will always get a request.
772 	 */
773 	rq = list_first_entry_or_null(rqs, struct request, queuelist);
774 	if (rq) {
775 		nr = kyber_get_domain_token(kqd, khd, hctx);
776 		if (nr >= 0) {
777 			khd->batching++;
778 			rq_set_domain_token(rq, nr);
779 			list_del_init(&rq->queuelist);
780 			return rq;
781 		} else {
782 			trace_kyber_throttled(kqd->dev,
783 					      kyber_domain_names[khd->cur_domain]);
784 		}
785 	} else if (sbitmap_any_bit_set(&khd->kcq_map[khd->cur_domain])) {
786 		nr = kyber_get_domain_token(kqd, khd, hctx);
787 		if (nr >= 0) {
788 			kyber_flush_busy_kcqs(khd, khd->cur_domain, rqs);
789 			rq = list_first_entry(rqs, struct request, queuelist);
790 			khd->batching++;
791 			rq_set_domain_token(rq, nr);
792 			list_del_init(&rq->queuelist);
793 			return rq;
794 		} else {
795 			trace_kyber_throttled(kqd->dev,
796 					      kyber_domain_names[khd->cur_domain]);
797 		}
798 	}
799 
800 	/* There were either no pending requests or no tokens. */
801 	return NULL;
802 }
803 
kyber_dispatch_request(struct blk_mq_hw_ctx * hctx)804 static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
805 {
806 	struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
807 	struct kyber_hctx_data *khd = hctx->sched_data;
808 	struct request *rq;
809 	int i;
810 
811 	spin_lock(&khd->lock);
812 
813 	/*
814 	 * First, if we are still entitled to batch, try to dispatch a request
815 	 * from the batch.
816 	 */
817 	if (khd->batching < kyber_batch_size[khd->cur_domain]) {
818 		rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
819 		if (rq)
820 			goto out;
821 	}
822 
823 	/*
824 	 * Either,
825 	 * 1. We were no longer entitled to a batch.
826 	 * 2. The domain we were batching didn't have any requests.
827 	 * 3. The domain we were batching was out of tokens.
828 	 *
829 	 * Start another batch. Note that this wraps back around to the original
830 	 * domain if no other domains have requests or tokens.
831 	 */
832 	khd->batching = 0;
833 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
834 		if (khd->cur_domain == KYBER_NUM_DOMAINS - 1)
835 			khd->cur_domain = 0;
836 		else
837 			khd->cur_domain++;
838 
839 		rq = kyber_dispatch_cur_domain(kqd, khd, hctx);
840 		if (rq)
841 			goto out;
842 	}
843 
844 	rq = NULL;
845 out:
846 	spin_unlock(&khd->lock);
847 	return rq;
848 }
849 
kyber_has_work(struct blk_mq_hw_ctx * hctx)850 static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
851 {
852 	struct kyber_hctx_data *khd = hctx->sched_data;
853 	int i;
854 
855 	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
856 		if (!list_empty_careful(&khd->rqs[i]) ||
857 		    sbitmap_any_bit_set(&khd->kcq_map[i]))
858 			return true;
859 	}
860 
861 	return false;
862 }
863 
864 #define KYBER_LAT_SHOW_STORE(domain, name)				\
865 static ssize_t kyber_##name##_lat_show(struct elevator_queue *e,	\
866 				       char *page)			\
867 {									\
868 	struct kyber_queue_data *kqd = e->elevator_data;		\
869 									\
870 	return sprintf(page, "%llu\n", kqd->latency_targets[domain]);	\
871 }									\
872 									\
873 static ssize_t kyber_##name##_lat_store(struct elevator_queue *e,	\
874 					const char *page, size_t count)	\
875 {									\
876 	struct kyber_queue_data *kqd = e->elevator_data;		\
877 	unsigned long long nsec;					\
878 	int ret;							\
879 									\
880 	ret = kstrtoull(page, 10, &nsec);				\
881 	if (ret)							\
882 		return ret;						\
883 									\
884 	kqd->latency_targets[domain] = nsec;				\
885 									\
886 	return count;							\
887 }
888 KYBER_LAT_SHOW_STORE(KYBER_READ, read);
889 KYBER_LAT_SHOW_STORE(KYBER_WRITE, write);
890 #undef KYBER_LAT_SHOW_STORE
891 
892 #define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
893 static struct elv_fs_entry kyber_sched_attrs[] = {
894 	KYBER_LAT_ATTR(read),
895 	KYBER_LAT_ATTR(write),
896 	__ATTR_NULL
897 };
898 #undef KYBER_LAT_ATTR
899 
900 #ifdef CONFIG_BLK_DEBUG_FS
901 #define KYBER_DEBUGFS_DOMAIN_ATTRS(domain, name)			\
902 static int kyber_##name##_tokens_show(void *data, struct seq_file *m)	\
903 {									\
904 	struct request_queue *q = data;					\
905 	struct kyber_queue_data *kqd = q->elevator->elevator_data;	\
906 									\
907 	sbitmap_queue_show(&kqd->domain_tokens[domain], m);		\
908 	return 0;							\
909 }									\
910 									\
911 static void *kyber_##name##_rqs_start(struct seq_file *m, loff_t *pos)	\
912 	__acquires(&khd->lock)						\
913 {									\
914 	struct blk_mq_hw_ctx *hctx = m->private;			\
915 	struct kyber_hctx_data *khd = hctx->sched_data;			\
916 									\
917 	spin_lock(&khd->lock);						\
918 	return seq_list_start(&khd->rqs[domain], *pos);			\
919 }									\
920 									\
921 static void *kyber_##name##_rqs_next(struct seq_file *m, void *v,	\
922 				     loff_t *pos)			\
923 {									\
924 	struct blk_mq_hw_ctx *hctx = m->private;			\
925 	struct kyber_hctx_data *khd = hctx->sched_data;			\
926 									\
927 	return seq_list_next(v, &khd->rqs[domain], pos);		\
928 }									\
929 									\
930 static void kyber_##name##_rqs_stop(struct seq_file *m, void *v)	\
931 	__releases(&khd->lock)						\
932 {									\
933 	struct blk_mq_hw_ctx *hctx = m->private;			\
934 	struct kyber_hctx_data *khd = hctx->sched_data;			\
935 									\
936 	spin_unlock(&khd->lock);					\
937 }									\
938 									\
939 static const struct seq_operations kyber_##name##_rqs_seq_ops = {	\
940 	.start	= kyber_##name##_rqs_start,				\
941 	.next	= kyber_##name##_rqs_next,				\
942 	.stop	= kyber_##name##_rqs_stop,				\
943 	.show	= blk_mq_debugfs_rq_show,				\
944 };									\
945 									\
946 static int kyber_##name##_waiting_show(void *data, struct seq_file *m)	\
947 {									\
948 	struct blk_mq_hw_ctx *hctx = data;				\
949 	struct kyber_hctx_data *khd = hctx->sched_data;			\
950 	wait_queue_entry_t *wait = &khd->domain_wait[domain].wait;	\
951 									\
952 	seq_printf(m, "%d\n", !list_empty_careful(&wait->entry));	\
953 	return 0;							\
954 }
KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ,read)955 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_READ, read)
956 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_WRITE, write)
957 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
958 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
959 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
960 
961 static int kyber_async_depth_show(void *data, struct seq_file *m)
962 {
963 	struct request_queue *q = data;
964 	struct kyber_queue_data *kqd = q->elevator->elevator_data;
965 
966 	seq_printf(m, "%u\n", kqd->async_depth);
967 	return 0;
968 }
969 
kyber_cur_domain_show(void * data,struct seq_file * m)970 static int kyber_cur_domain_show(void *data, struct seq_file *m)
971 {
972 	struct blk_mq_hw_ctx *hctx = data;
973 	struct kyber_hctx_data *khd = hctx->sched_data;
974 
975 	seq_printf(m, "%s\n", kyber_domain_names[khd->cur_domain]);
976 	return 0;
977 }
978 
kyber_batching_show(void * data,struct seq_file * m)979 static int kyber_batching_show(void *data, struct seq_file *m)
980 {
981 	struct blk_mq_hw_ctx *hctx = data;
982 	struct kyber_hctx_data *khd = hctx->sched_data;
983 
984 	seq_printf(m, "%u\n", khd->batching);
985 	return 0;
986 }
987 
988 #define KYBER_QUEUE_DOMAIN_ATTRS(name)	\
989 	{#name "_tokens", 0400, kyber_##name##_tokens_show}
990 static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
991 	KYBER_QUEUE_DOMAIN_ATTRS(read),
992 	KYBER_QUEUE_DOMAIN_ATTRS(write),
993 	KYBER_QUEUE_DOMAIN_ATTRS(discard),
994 	KYBER_QUEUE_DOMAIN_ATTRS(other),
995 	{"async_depth", 0400, kyber_async_depth_show},
996 	{},
997 };
998 #undef KYBER_QUEUE_DOMAIN_ATTRS
999 
1000 #define KYBER_HCTX_DOMAIN_ATTRS(name)					\
1001 	{#name "_rqs", 0400, .seq_ops = &kyber_##name##_rqs_seq_ops},	\
1002 	{#name "_waiting", 0400, kyber_##name##_waiting_show}
1003 static const struct blk_mq_debugfs_attr kyber_hctx_debugfs_attrs[] = {
1004 	KYBER_HCTX_DOMAIN_ATTRS(read),
1005 	KYBER_HCTX_DOMAIN_ATTRS(write),
1006 	KYBER_HCTX_DOMAIN_ATTRS(discard),
1007 	KYBER_HCTX_DOMAIN_ATTRS(other),
1008 	{"cur_domain", 0400, kyber_cur_domain_show},
1009 	{"batching", 0400, kyber_batching_show},
1010 	{},
1011 };
1012 #undef KYBER_HCTX_DOMAIN_ATTRS
1013 #endif
1014 
1015 static struct elevator_type kyber_sched = {
1016 	.ops = {
1017 		.init_sched = kyber_init_sched,
1018 		.exit_sched = kyber_exit_sched,
1019 		.init_hctx = kyber_init_hctx,
1020 		.exit_hctx = kyber_exit_hctx,
1021 		.limit_depth = kyber_limit_depth,
1022 		.bio_merge = kyber_bio_merge,
1023 		.prepare_request = kyber_prepare_request,
1024 		.insert_requests = kyber_insert_requests,
1025 		.finish_request = kyber_finish_request,
1026 		.requeue_request = kyber_finish_request,
1027 		.completed_request = kyber_completed_request,
1028 		.dispatch_request = kyber_dispatch_request,
1029 		.has_work = kyber_has_work,
1030 		.depth_updated = kyber_depth_updated,
1031 	},
1032 #ifdef CONFIG_BLK_DEBUG_FS
1033 	.queue_debugfs_attrs = kyber_queue_debugfs_attrs,
1034 	.hctx_debugfs_attrs = kyber_hctx_debugfs_attrs,
1035 #endif
1036 	.elevator_attrs = kyber_sched_attrs,
1037 	.elevator_name = "kyber",
1038 	.elevator_owner = THIS_MODULE,
1039 };
1040 
kyber_init(void)1041 static int __init kyber_init(void)
1042 {
1043 	return elv_register(&kyber_sched);
1044 }
1045 
kyber_exit(void)1046 static void __exit kyber_exit(void)
1047 {
1048 	elv_unregister(&kyber_sched);
1049 }
1050 
1051 module_init(kyber_init);
1052 module_exit(kyber_exit);
1053 
1054 MODULE_AUTHOR("Omar Sandoval");
1055 MODULE_LICENSE("GPL");
1056 MODULE_DESCRIPTION("Kyber I/O scheduler");
1057