xref: /linux/mm/damon/core.c (revision 28e43197c4460e08b55554c05998d6de9da7d0b6)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Data Access Monitor
4  *
5  * Author: SeongJae Park <sj@kernel.org>
6  */
7 
8 #define pr_fmt(fmt) "damon: " fmt
9 
10 #include <linux/damon.h>
11 #include <linux/delay.h>
12 #include <linux/kthread.h>
13 #include <linux/mm.h>
14 #include <linux/psi.h>
15 #include <linux/slab.h>
16 #include <linux/string.h>
17 
18 #define CREATE_TRACE_POINTS
19 #include <trace/events/damon.h>
20 
21 #ifdef CONFIG_DAMON_KUNIT_TEST
22 #undef DAMON_MIN_REGION
23 #define DAMON_MIN_REGION 1
24 #endif
25 
26 static DEFINE_MUTEX(damon_lock);
27 static int nr_running_ctxs;
28 static bool running_exclusive_ctxs;
29 
30 static DEFINE_MUTEX(damon_ops_lock);
31 static struct damon_operations damon_registered_ops[NR_DAMON_OPS];
32 
33 static struct kmem_cache *damon_region_cache __ro_after_init;
34 
35 /* Should be called under damon_ops_lock with id smaller than NR_DAMON_OPS */
__damon_is_registered_ops(enum damon_ops_id id)36 static bool __damon_is_registered_ops(enum damon_ops_id id)
37 {
38 	struct damon_operations empty_ops = {};
39 
40 	if (!memcmp(&empty_ops, &damon_registered_ops[id], sizeof(empty_ops)))
41 		return false;
42 	return true;
43 }
44 
45 /**
46  * damon_is_registered_ops() - Check if a given damon_operations is registered.
47  * @id:	Id of the damon_operations to check if registered.
48  *
49  * Return: true if the ops is set, false otherwise.
50  */
damon_is_registered_ops(enum damon_ops_id id)51 bool damon_is_registered_ops(enum damon_ops_id id)
52 {
53 	bool registered;
54 
55 	if (id >= NR_DAMON_OPS)
56 		return false;
57 	mutex_lock(&damon_ops_lock);
58 	registered = __damon_is_registered_ops(id);
59 	mutex_unlock(&damon_ops_lock);
60 	return registered;
61 }
62 
63 /**
64  * damon_register_ops() - Register a monitoring operations set to DAMON.
65  * @ops:	monitoring operations set to register.
66  *
67  * This function registers a monitoring operations set of valid &struct
68  * damon_operations->id so that others can find and use them later.
69  *
70  * Return: 0 on success, negative error code otherwise.
71  */
damon_register_ops(struct damon_operations * ops)72 int damon_register_ops(struct damon_operations *ops)
73 {
74 	int err = 0;
75 
76 	if (ops->id >= NR_DAMON_OPS)
77 		return -EINVAL;
78 	mutex_lock(&damon_ops_lock);
79 	/* Fail for already registered ops */
80 	if (__damon_is_registered_ops(ops->id)) {
81 		err = -EINVAL;
82 		goto out;
83 	}
84 	damon_registered_ops[ops->id] = *ops;
85 out:
86 	mutex_unlock(&damon_ops_lock);
87 	return err;
88 }
89 
90 /**
91  * damon_select_ops() - Select a monitoring operations to use with the context.
92  * @ctx:	monitoring context to use the operations.
93  * @id:		id of the registered monitoring operations to select.
94  *
95  * This function finds registered monitoring operations set of @id and make
96  * @ctx to use it.
97  *
98  * Return: 0 on success, negative error code otherwise.
99  */
damon_select_ops(struct damon_ctx * ctx,enum damon_ops_id id)100 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id)
101 {
102 	int err = 0;
103 
104 	if (id >= NR_DAMON_OPS)
105 		return -EINVAL;
106 
107 	mutex_lock(&damon_ops_lock);
108 	if (!__damon_is_registered_ops(id))
109 		err = -EINVAL;
110 	else
111 		ctx->ops = damon_registered_ops[id];
112 	mutex_unlock(&damon_ops_lock);
113 	return err;
114 }
115 
116 /*
117  * Construct a damon_region struct
118  *
119  * Returns the pointer to the new struct if success, or NULL otherwise
120  */
damon_new_region(unsigned long start,unsigned long end)121 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
122 {
123 	struct damon_region *region;
124 
125 	region = kmem_cache_alloc(damon_region_cache, GFP_KERNEL);
126 	if (!region)
127 		return NULL;
128 
129 	region->ar.start = start;
130 	region->ar.end = end;
131 	region->nr_accesses = 0;
132 	region->nr_accesses_bp = 0;
133 	INIT_LIST_HEAD(&region->list);
134 
135 	region->age = 0;
136 	region->last_nr_accesses = 0;
137 
138 	return region;
139 }
140 
damon_add_region(struct damon_region * r,struct damon_target * t)141 void damon_add_region(struct damon_region *r, struct damon_target *t)
142 {
143 	list_add_tail(&r->list, &t->regions_list);
144 	t->nr_regions++;
145 }
146 
damon_del_region(struct damon_region * r,struct damon_target * t)147 static void damon_del_region(struct damon_region *r, struct damon_target *t)
148 {
149 	list_del(&r->list);
150 	t->nr_regions--;
151 }
152 
damon_free_region(struct damon_region * r)153 static void damon_free_region(struct damon_region *r)
154 {
155 	kmem_cache_free(damon_region_cache, r);
156 }
157 
damon_destroy_region(struct damon_region * r,struct damon_target * t)158 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
159 {
160 	damon_del_region(r, t);
161 	damon_free_region(r);
162 }
163 
164 /*
165  * Check whether a region is intersecting an address range
166  *
167  * Returns true if it is.
168  */
damon_intersect(struct damon_region * r,struct damon_addr_range * re)169 static bool damon_intersect(struct damon_region *r,
170 		struct damon_addr_range *re)
171 {
172 	return !(r->ar.end <= re->start || re->end <= r->ar.start);
173 }
174 
175 /*
176  * Fill holes in regions with new regions.
177  */
damon_fill_regions_holes(struct damon_region * first,struct damon_region * last,struct damon_target * t)178 static int damon_fill_regions_holes(struct damon_region *first,
179 		struct damon_region *last, struct damon_target *t)
180 {
181 	struct damon_region *r = first;
182 
183 	damon_for_each_region_from(r, t) {
184 		struct damon_region *next, *newr;
185 
186 		if (r == last)
187 			break;
188 		next = damon_next_region(r);
189 		if (r->ar.end != next->ar.start) {
190 			newr = damon_new_region(r->ar.end, next->ar.start);
191 			if (!newr)
192 				return -ENOMEM;
193 			damon_insert_region(newr, r, next, t);
194 		}
195 	}
196 	return 0;
197 }
198 
199 /*
200  * damon_set_regions() - Set regions of a target for given address ranges.
201  * @t:		the given target.
202  * @ranges:	array of new monitoring target ranges.
203  * @nr_ranges:	length of @ranges.
204  *
205  * This function adds new regions to, or modify existing regions of a
206  * monitoring target to fit in specific ranges.
207  *
208  * Return: 0 if success, or negative error code otherwise.
209  */
damon_set_regions(struct damon_target * t,struct damon_addr_range * ranges,unsigned int nr_ranges)210 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
211 		unsigned int nr_ranges)
212 {
213 	struct damon_region *r, *next;
214 	unsigned int i;
215 	int err;
216 
217 	/* Remove regions which are not in the new ranges */
218 	damon_for_each_region_safe(r, next, t) {
219 		for (i = 0; i < nr_ranges; i++) {
220 			if (damon_intersect(r, &ranges[i]))
221 				break;
222 		}
223 		if (i == nr_ranges)
224 			damon_destroy_region(r, t);
225 	}
226 
227 	r = damon_first_region(t);
228 	/* Add new regions or resize existing regions to fit in the ranges */
229 	for (i = 0; i < nr_ranges; i++) {
230 		struct damon_region *first = NULL, *last, *newr;
231 		struct damon_addr_range *range;
232 
233 		range = &ranges[i];
234 		/* Get the first/last regions intersecting with the range */
235 		damon_for_each_region_from(r, t) {
236 			if (damon_intersect(r, range)) {
237 				if (!first)
238 					first = r;
239 				last = r;
240 			}
241 			if (r->ar.start >= range->end)
242 				break;
243 		}
244 		if (!first) {
245 			/* no region intersects with this range */
246 			newr = damon_new_region(
247 					ALIGN_DOWN(range->start,
248 						DAMON_MIN_REGION),
249 					ALIGN(range->end, DAMON_MIN_REGION));
250 			if (!newr)
251 				return -ENOMEM;
252 			damon_insert_region(newr, damon_prev_region(r), r, t);
253 		} else {
254 			/* resize intersecting regions to fit in this range */
255 			first->ar.start = ALIGN_DOWN(range->start,
256 					DAMON_MIN_REGION);
257 			last->ar.end = ALIGN(range->end, DAMON_MIN_REGION);
258 
259 			/* fill possible holes in the range */
260 			err = damon_fill_regions_holes(first, last, t);
261 			if (err)
262 				return err;
263 		}
264 	}
265 	return 0;
266 }
267 
damos_new_filter(enum damos_filter_type type,bool matching)268 struct damos_filter *damos_new_filter(enum damos_filter_type type,
269 		bool matching)
270 {
271 	struct damos_filter *filter;
272 
273 	filter = kmalloc(sizeof(*filter), GFP_KERNEL);
274 	if (!filter)
275 		return NULL;
276 	filter->type = type;
277 	filter->matching = matching;
278 	INIT_LIST_HEAD(&filter->list);
279 	return filter;
280 }
281 
damos_add_filter(struct damos * s,struct damos_filter * f)282 void damos_add_filter(struct damos *s, struct damos_filter *f)
283 {
284 	list_add_tail(&f->list, &s->filters);
285 }
286 
damos_del_filter(struct damos_filter * f)287 static void damos_del_filter(struct damos_filter *f)
288 {
289 	list_del(&f->list);
290 }
291 
damos_free_filter(struct damos_filter * f)292 static void damos_free_filter(struct damos_filter *f)
293 {
294 	kfree(f);
295 }
296 
damos_destroy_filter(struct damos_filter * f)297 void damos_destroy_filter(struct damos_filter *f)
298 {
299 	damos_del_filter(f);
300 	damos_free_filter(f);
301 }
302 
damos_new_quota_goal(enum damos_quota_goal_metric metric,unsigned long target_value)303 struct damos_quota_goal *damos_new_quota_goal(
304 		enum damos_quota_goal_metric metric,
305 		unsigned long target_value)
306 {
307 	struct damos_quota_goal *goal;
308 
309 	goal = kmalloc(sizeof(*goal), GFP_KERNEL);
310 	if (!goal)
311 		return NULL;
312 	goal->metric = metric;
313 	goal->target_value = target_value;
314 	INIT_LIST_HEAD(&goal->list);
315 	return goal;
316 }
317 
damos_add_quota_goal(struct damos_quota * q,struct damos_quota_goal * g)318 void damos_add_quota_goal(struct damos_quota *q, struct damos_quota_goal *g)
319 {
320 	list_add_tail(&g->list, &q->goals);
321 }
322 
damos_del_quota_goal(struct damos_quota_goal * g)323 static void damos_del_quota_goal(struct damos_quota_goal *g)
324 {
325 	list_del(&g->list);
326 }
327 
damos_free_quota_goal(struct damos_quota_goal * g)328 static void damos_free_quota_goal(struct damos_quota_goal *g)
329 {
330 	kfree(g);
331 }
332 
damos_destroy_quota_goal(struct damos_quota_goal * g)333 void damos_destroy_quota_goal(struct damos_quota_goal *g)
334 {
335 	damos_del_quota_goal(g);
336 	damos_free_quota_goal(g);
337 }
338 
339 /* initialize fields of @quota that normally API users wouldn't set */
damos_quota_init(struct damos_quota * quota)340 static struct damos_quota *damos_quota_init(struct damos_quota *quota)
341 {
342 	quota->esz = 0;
343 	quota->total_charged_sz = 0;
344 	quota->total_charged_ns = 0;
345 	quota->charged_sz = 0;
346 	quota->charged_from = 0;
347 	quota->charge_target_from = NULL;
348 	quota->charge_addr_from = 0;
349 	quota->esz_bp = 0;
350 	return quota;
351 }
352 
damon_new_scheme(struct damos_access_pattern * pattern,enum damos_action action,unsigned long apply_interval_us,struct damos_quota * quota,struct damos_watermarks * wmarks,int target_nid)353 struct damos *damon_new_scheme(struct damos_access_pattern *pattern,
354 			enum damos_action action,
355 			unsigned long apply_interval_us,
356 			struct damos_quota *quota,
357 			struct damos_watermarks *wmarks,
358 			int target_nid)
359 {
360 	struct damos *scheme;
361 
362 	scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
363 	if (!scheme)
364 		return NULL;
365 	scheme->pattern = *pattern;
366 	scheme->action = action;
367 	scheme->apply_interval_us = apply_interval_us;
368 	/*
369 	 * next_apply_sis will be set when kdamond starts.  While kdamond is
370 	 * running, it will also updated when it is added to the DAMON context,
371 	 * or damon_attrs are updated.
372 	 */
373 	scheme->next_apply_sis = 0;
374 	INIT_LIST_HEAD(&scheme->filters);
375 	scheme->stat = (struct damos_stat){};
376 	INIT_LIST_HEAD(&scheme->list);
377 
378 	scheme->quota = *(damos_quota_init(quota));
379 	/* quota.goals should be separately set by caller */
380 	INIT_LIST_HEAD(&scheme->quota.goals);
381 
382 	scheme->wmarks = *wmarks;
383 	scheme->wmarks.activated = true;
384 
385 	scheme->target_nid = target_nid;
386 
387 	return scheme;
388 }
389 
damos_set_next_apply_sis(struct damos * s,struct damon_ctx * ctx)390 static void damos_set_next_apply_sis(struct damos *s, struct damon_ctx *ctx)
391 {
392 	unsigned long sample_interval = ctx->attrs.sample_interval ?
393 		ctx->attrs.sample_interval : 1;
394 	unsigned long apply_interval = s->apply_interval_us ?
395 		s->apply_interval_us : ctx->attrs.aggr_interval;
396 
397 	s->next_apply_sis = ctx->passed_sample_intervals +
398 		apply_interval / sample_interval;
399 }
400 
damon_add_scheme(struct damon_ctx * ctx,struct damos * s)401 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
402 {
403 	list_add_tail(&s->list, &ctx->schemes);
404 	damos_set_next_apply_sis(s, ctx);
405 }
406 
damon_del_scheme(struct damos * s)407 static void damon_del_scheme(struct damos *s)
408 {
409 	list_del(&s->list);
410 }
411 
damon_free_scheme(struct damos * s)412 static void damon_free_scheme(struct damos *s)
413 {
414 	kfree(s);
415 }
416 
damon_destroy_scheme(struct damos * s)417 void damon_destroy_scheme(struct damos *s)
418 {
419 	struct damos_quota_goal *g, *g_next;
420 	struct damos_filter *f, *next;
421 
422 	damos_for_each_quota_goal_safe(g, g_next, &s->quota)
423 		damos_destroy_quota_goal(g);
424 
425 	damos_for_each_filter_safe(f, next, s)
426 		damos_destroy_filter(f);
427 	damon_del_scheme(s);
428 	damon_free_scheme(s);
429 }
430 
431 /*
432  * Construct a damon_target struct
433  *
434  * Returns the pointer to the new struct if success, or NULL otherwise
435  */
damon_new_target(void)436 struct damon_target *damon_new_target(void)
437 {
438 	struct damon_target *t;
439 
440 	t = kmalloc(sizeof(*t), GFP_KERNEL);
441 	if (!t)
442 		return NULL;
443 
444 	t->pid = NULL;
445 	t->nr_regions = 0;
446 	INIT_LIST_HEAD(&t->regions_list);
447 	INIT_LIST_HEAD(&t->list);
448 
449 	return t;
450 }
451 
damon_add_target(struct damon_ctx * ctx,struct damon_target * t)452 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
453 {
454 	list_add_tail(&t->list, &ctx->adaptive_targets);
455 }
456 
damon_targets_empty(struct damon_ctx * ctx)457 bool damon_targets_empty(struct damon_ctx *ctx)
458 {
459 	return list_empty(&ctx->adaptive_targets);
460 }
461 
damon_del_target(struct damon_target * t)462 static void damon_del_target(struct damon_target *t)
463 {
464 	list_del(&t->list);
465 }
466 
damon_free_target(struct damon_target * t)467 void damon_free_target(struct damon_target *t)
468 {
469 	struct damon_region *r, *next;
470 
471 	damon_for_each_region_safe(r, next, t)
472 		damon_free_region(r);
473 	kfree(t);
474 }
475 
damon_destroy_target(struct damon_target * t)476 void damon_destroy_target(struct damon_target *t)
477 {
478 	damon_del_target(t);
479 	damon_free_target(t);
480 }
481 
damon_nr_regions(struct damon_target * t)482 unsigned int damon_nr_regions(struct damon_target *t)
483 {
484 	return t->nr_regions;
485 }
486 
damon_new_ctx(void)487 struct damon_ctx *damon_new_ctx(void)
488 {
489 	struct damon_ctx *ctx;
490 
491 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
492 	if (!ctx)
493 		return NULL;
494 
495 	init_completion(&ctx->kdamond_started);
496 
497 	ctx->attrs.sample_interval = 5 * 1000;
498 	ctx->attrs.aggr_interval = 100 * 1000;
499 	ctx->attrs.ops_update_interval = 60 * 1000 * 1000;
500 
501 	ctx->passed_sample_intervals = 0;
502 	/* These will be set from kdamond_init_intervals_sis() */
503 	ctx->next_aggregation_sis = 0;
504 	ctx->next_ops_update_sis = 0;
505 
506 	mutex_init(&ctx->kdamond_lock);
507 
508 	ctx->attrs.min_nr_regions = 10;
509 	ctx->attrs.max_nr_regions = 1000;
510 
511 	INIT_LIST_HEAD(&ctx->adaptive_targets);
512 	INIT_LIST_HEAD(&ctx->schemes);
513 
514 	return ctx;
515 }
516 
damon_destroy_targets(struct damon_ctx * ctx)517 static void damon_destroy_targets(struct damon_ctx *ctx)
518 {
519 	struct damon_target *t, *next_t;
520 
521 	if (ctx->ops.cleanup) {
522 		ctx->ops.cleanup(ctx);
523 		return;
524 	}
525 
526 	damon_for_each_target_safe(t, next_t, ctx)
527 		damon_destroy_target(t);
528 }
529 
damon_destroy_ctx(struct damon_ctx * ctx)530 void damon_destroy_ctx(struct damon_ctx *ctx)
531 {
532 	struct damos *s, *next_s;
533 
534 	damon_destroy_targets(ctx);
535 
536 	damon_for_each_scheme_safe(s, next_s, ctx)
537 		damon_destroy_scheme(s);
538 
539 	kfree(ctx);
540 }
541 
damon_age_for_new_attrs(unsigned int age,struct damon_attrs * old_attrs,struct damon_attrs * new_attrs)542 static unsigned int damon_age_for_new_attrs(unsigned int age,
543 		struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
544 {
545 	return age * old_attrs->aggr_interval / new_attrs->aggr_interval;
546 }
547 
548 /* convert access ratio in bp (per 10,000) to nr_accesses */
damon_accesses_bp_to_nr_accesses(unsigned int accesses_bp,struct damon_attrs * attrs)549 static unsigned int damon_accesses_bp_to_nr_accesses(
550 		unsigned int accesses_bp, struct damon_attrs *attrs)
551 {
552 	return accesses_bp * damon_max_nr_accesses(attrs) / 10000;
553 }
554 
555 /*
556  * Convert nr_accesses to access ratio in bp (per 10,000).
557  *
558  * Callers should ensure attrs.aggr_interval is not zero, like
559  * damon_update_monitoring_results() does .  Otherwise, divide-by-zero would
560  * happen.
561  */
damon_nr_accesses_to_accesses_bp(unsigned int nr_accesses,struct damon_attrs * attrs)562 static unsigned int damon_nr_accesses_to_accesses_bp(
563 		unsigned int nr_accesses, struct damon_attrs *attrs)
564 {
565 	return nr_accesses * 10000 / damon_max_nr_accesses(attrs);
566 }
567 
damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,struct damon_attrs * old_attrs,struct damon_attrs * new_attrs)568 static unsigned int damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,
569 		struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
570 {
571 	return damon_accesses_bp_to_nr_accesses(
572 			damon_nr_accesses_to_accesses_bp(
573 				nr_accesses, old_attrs),
574 			new_attrs);
575 }
576 
damon_update_monitoring_result(struct damon_region * r,struct damon_attrs * old_attrs,struct damon_attrs * new_attrs)577 static void damon_update_monitoring_result(struct damon_region *r,
578 		struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
579 {
580 	r->nr_accesses = damon_nr_accesses_for_new_attrs(r->nr_accesses,
581 			old_attrs, new_attrs);
582 	r->nr_accesses_bp = r->nr_accesses * 10000;
583 	r->age = damon_age_for_new_attrs(r->age, old_attrs, new_attrs);
584 }
585 
586 /*
587  * region->nr_accesses is the number of sampling intervals in the last
588  * aggregation interval that access to the region has found, and region->age is
589  * the number of aggregation intervals that its access pattern has maintained.
590  * For the reason, the real meaning of the two fields depend on current
591  * sampling interval and aggregation interval.  This function updates
592  * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.
593  */
damon_update_monitoring_results(struct damon_ctx * ctx,struct damon_attrs * new_attrs)594 static void damon_update_monitoring_results(struct damon_ctx *ctx,
595 		struct damon_attrs *new_attrs)
596 {
597 	struct damon_attrs *old_attrs = &ctx->attrs;
598 	struct damon_target *t;
599 	struct damon_region *r;
600 
601 	/* if any interval is zero, simply forgive conversion */
602 	if (!old_attrs->sample_interval || !old_attrs->aggr_interval ||
603 			!new_attrs->sample_interval ||
604 			!new_attrs->aggr_interval)
605 		return;
606 
607 	damon_for_each_target(t, ctx)
608 		damon_for_each_region(r, t)
609 			damon_update_monitoring_result(
610 					r, old_attrs, new_attrs);
611 }
612 
613 /**
614  * damon_set_attrs() - Set attributes for the monitoring.
615  * @ctx:		monitoring context
616  * @attrs:		monitoring attributes
617  *
618  * This function should be called while the kdamond is not running, or an
619  * access check results aggregation is not ongoing (e.g., from
620  * &struct damon_callback->after_aggregation or
621  * &struct damon_callback->after_wmarks_check callbacks).
622  *
623  * Every time interval is in micro-seconds.
624  *
625  * Return: 0 on success, negative error code otherwise.
626  */
damon_set_attrs(struct damon_ctx * ctx,struct damon_attrs * attrs)627 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
628 {
629 	unsigned long sample_interval = attrs->sample_interval ?
630 		attrs->sample_interval : 1;
631 	struct damos *s;
632 
633 	if (attrs->min_nr_regions < 3)
634 		return -EINVAL;
635 	if (attrs->min_nr_regions > attrs->max_nr_regions)
636 		return -EINVAL;
637 	if (attrs->sample_interval > attrs->aggr_interval)
638 		return -EINVAL;
639 
640 	ctx->next_aggregation_sis = ctx->passed_sample_intervals +
641 		attrs->aggr_interval / sample_interval;
642 	ctx->next_ops_update_sis = ctx->passed_sample_intervals +
643 		attrs->ops_update_interval / sample_interval;
644 
645 	damon_update_monitoring_results(ctx, attrs);
646 	ctx->attrs = *attrs;
647 
648 	damon_for_each_scheme(s, ctx)
649 		damos_set_next_apply_sis(s, ctx);
650 
651 	return 0;
652 }
653 
654 /**
655  * damon_set_schemes() - Set data access monitoring based operation schemes.
656  * @ctx:	monitoring context
657  * @schemes:	array of the schemes
658  * @nr_schemes:	number of entries in @schemes
659  *
660  * This function should not be called while the kdamond of the context is
661  * running.
662  */
damon_set_schemes(struct damon_ctx * ctx,struct damos ** schemes,ssize_t nr_schemes)663 void damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
664 			ssize_t nr_schemes)
665 {
666 	struct damos *s, *next;
667 	ssize_t i;
668 
669 	damon_for_each_scheme_safe(s, next, ctx)
670 		damon_destroy_scheme(s);
671 	for (i = 0; i < nr_schemes; i++)
672 		damon_add_scheme(ctx, schemes[i]);
673 }
674 
damos_nth_quota_goal(int n,struct damos_quota * q)675 static struct damos_quota_goal *damos_nth_quota_goal(
676 		int n, struct damos_quota *q)
677 {
678 	struct damos_quota_goal *goal;
679 	int i = 0;
680 
681 	damos_for_each_quota_goal(goal, q) {
682 		if (i++ == n)
683 			return goal;
684 	}
685 	return NULL;
686 }
687 
damos_commit_quota_goal(struct damos_quota_goal * dst,struct damos_quota_goal * src)688 static void damos_commit_quota_goal(
689 		struct damos_quota_goal *dst, struct damos_quota_goal *src)
690 {
691 	dst->metric = src->metric;
692 	dst->target_value = src->target_value;
693 	if (dst->metric == DAMOS_QUOTA_USER_INPUT)
694 		dst->current_value = src->current_value;
695 	/* keep last_psi_total as is, since it will be updated in next cycle */
696 }
697 
698 /**
699  * damos_commit_quota_goals() - Commit DAMOS quota goals to another quota.
700  * @dst:	The commit destination DAMOS quota.
701  * @src:	The commit source DAMOS quota.
702  *
703  * Copies user-specified parameters for quota goals from @src to @dst.  Users
704  * should use this function for quota goals-level parameters update of running
705  * DAMON contexts, instead of manual in-place updates.
706  *
707  * This function should be called from parameters-update safe context, like
708  * DAMON callbacks.
709  */
damos_commit_quota_goals(struct damos_quota * dst,struct damos_quota * src)710 int damos_commit_quota_goals(struct damos_quota *dst, struct damos_quota *src)
711 {
712 	struct damos_quota_goal *dst_goal, *next, *src_goal, *new_goal;
713 	int i = 0, j = 0;
714 
715 	damos_for_each_quota_goal_safe(dst_goal, next, dst) {
716 		src_goal = damos_nth_quota_goal(i++, src);
717 		if (src_goal)
718 			damos_commit_quota_goal(dst_goal, src_goal);
719 		else
720 			damos_destroy_quota_goal(dst_goal);
721 	}
722 	damos_for_each_quota_goal_safe(src_goal, next, src) {
723 		if (j++ < i)
724 			continue;
725 		new_goal = damos_new_quota_goal(
726 				src_goal->metric, src_goal->target_value);
727 		if (!new_goal)
728 			return -ENOMEM;
729 		damos_add_quota_goal(dst, new_goal);
730 	}
731 	return 0;
732 }
733 
damos_commit_quota(struct damos_quota * dst,struct damos_quota * src)734 static int damos_commit_quota(struct damos_quota *dst, struct damos_quota *src)
735 {
736 	int err;
737 
738 	dst->reset_interval = src->reset_interval;
739 	dst->ms = src->ms;
740 	dst->sz = src->sz;
741 	err = damos_commit_quota_goals(dst, src);
742 	if (err)
743 		return err;
744 	dst->weight_sz = src->weight_sz;
745 	dst->weight_nr_accesses = src->weight_nr_accesses;
746 	dst->weight_age = src->weight_age;
747 	return 0;
748 }
749 
damos_nth_filter(int n,struct damos * s)750 static struct damos_filter *damos_nth_filter(int n, struct damos *s)
751 {
752 	struct damos_filter *filter;
753 	int i = 0;
754 
755 	damos_for_each_filter(filter, s) {
756 		if (i++ == n)
757 			return filter;
758 	}
759 	return NULL;
760 }
761 
damos_commit_filter_arg(struct damos_filter * dst,struct damos_filter * src)762 static void damos_commit_filter_arg(
763 		struct damos_filter *dst, struct damos_filter *src)
764 {
765 	switch (dst->type) {
766 	case DAMOS_FILTER_TYPE_MEMCG:
767 		dst->memcg_id = src->memcg_id;
768 		break;
769 	case DAMOS_FILTER_TYPE_ADDR:
770 		dst->addr_range = src->addr_range;
771 		break;
772 	case DAMOS_FILTER_TYPE_TARGET:
773 		dst->target_idx = src->target_idx;
774 		break;
775 	default:
776 		break;
777 	}
778 }
779 
damos_commit_filter(struct damos_filter * dst,struct damos_filter * src)780 static void damos_commit_filter(
781 		struct damos_filter *dst, struct damos_filter *src)
782 {
783 	dst->type = src->type;
784 	dst->matching = src->matching;
785 	damos_commit_filter_arg(dst, src);
786 }
787 
damos_commit_filters(struct damos * dst,struct damos * src)788 static int damos_commit_filters(struct damos *dst, struct damos *src)
789 {
790 	struct damos_filter *dst_filter, *next, *src_filter, *new_filter;
791 	int i = 0, j = 0;
792 
793 	damos_for_each_filter_safe(dst_filter, next, dst) {
794 		src_filter = damos_nth_filter(i++, src);
795 		if (src_filter)
796 			damos_commit_filter(dst_filter, src_filter);
797 		else
798 			damos_destroy_filter(dst_filter);
799 	}
800 
801 	damos_for_each_filter_safe(src_filter, next, src) {
802 		if (j++ < i)
803 			continue;
804 
805 		new_filter = damos_new_filter(
806 				src_filter->type, src_filter->matching);
807 		if (!new_filter)
808 			return -ENOMEM;
809 		damos_commit_filter_arg(new_filter, src_filter);
810 		damos_add_filter(dst, new_filter);
811 	}
812 	return 0;
813 }
814 
damon_nth_scheme(int n,struct damon_ctx * ctx)815 static struct damos *damon_nth_scheme(int n, struct damon_ctx *ctx)
816 {
817 	struct damos *s;
818 	int i = 0;
819 
820 	damon_for_each_scheme(s, ctx) {
821 		if (i++ == n)
822 			return s;
823 	}
824 	return NULL;
825 }
826 
damos_commit(struct damos * dst,struct damos * src)827 static int damos_commit(struct damos *dst, struct damos *src)
828 {
829 	int err;
830 
831 	dst->pattern = src->pattern;
832 	dst->action = src->action;
833 	dst->apply_interval_us = src->apply_interval_us;
834 
835 	err = damos_commit_quota(&dst->quota, &src->quota);
836 	if (err)
837 		return err;
838 
839 	dst->wmarks = src->wmarks;
840 
841 	err = damos_commit_filters(dst, src);
842 	return err;
843 }
844 
damon_commit_schemes(struct damon_ctx * dst,struct damon_ctx * src)845 static int damon_commit_schemes(struct damon_ctx *dst, struct damon_ctx *src)
846 {
847 	struct damos *dst_scheme, *next, *src_scheme, *new_scheme;
848 	int i = 0, j = 0, err;
849 
850 	damon_for_each_scheme_safe(dst_scheme, next, dst) {
851 		src_scheme = damon_nth_scheme(i++, src);
852 		if (src_scheme) {
853 			err = damos_commit(dst_scheme, src_scheme);
854 			if (err)
855 				return err;
856 		} else {
857 			damon_destroy_scheme(dst_scheme);
858 		}
859 	}
860 
861 	damon_for_each_scheme_safe(src_scheme, next, src) {
862 		if (j++ < i)
863 			continue;
864 		new_scheme = damon_new_scheme(&src_scheme->pattern,
865 				src_scheme->action,
866 				src_scheme->apply_interval_us,
867 				&src_scheme->quota, &src_scheme->wmarks,
868 				NUMA_NO_NODE);
869 		if (!new_scheme)
870 			return -ENOMEM;
871 		damon_add_scheme(dst, new_scheme);
872 	}
873 	return 0;
874 }
875 
damon_nth_target(int n,struct damon_ctx * ctx)876 static struct damon_target *damon_nth_target(int n, struct damon_ctx *ctx)
877 {
878 	struct damon_target *t;
879 	int i = 0;
880 
881 	damon_for_each_target(t, ctx) {
882 		if (i++ == n)
883 			return t;
884 	}
885 	return NULL;
886 }
887 
888 /*
889  * The caller should ensure the regions of @src are
890  * 1. valid (end >= src) and
891  * 2. sorted by starting address.
892  *
893  * If @src has no region, @dst keeps current regions.
894  */
damon_commit_target_regions(struct damon_target * dst,struct damon_target * src)895 static int damon_commit_target_regions(
896 		struct damon_target *dst, struct damon_target *src)
897 {
898 	struct damon_region *src_region;
899 	struct damon_addr_range *ranges;
900 	int i = 0, err;
901 
902 	damon_for_each_region(src_region, src)
903 		i++;
904 	if (!i)
905 		return 0;
906 
907 	ranges = kmalloc_array(i, sizeof(*ranges), GFP_KERNEL | __GFP_NOWARN);
908 	if (!ranges)
909 		return -ENOMEM;
910 	i = 0;
911 	damon_for_each_region(src_region, src)
912 		ranges[i++] = src_region->ar;
913 	err = damon_set_regions(dst, ranges, i);
914 	kfree(ranges);
915 	return err;
916 }
917 
damon_commit_target(struct damon_target * dst,bool dst_has_pid,struct damon_target * src,bool src_has_pid)918 static int damon_commit_target(
919 		struct damon_target *dst, bool dst_has_pid,
920 		struct damon_target *src, bool src_has_pid)
921 {
922 	int err;
923 
924 	err = damon_commit_target_regions(dst, src);
925 	if (err)
926 		return err;
927 	if (dst_has_pid)
928 		put_pid(dst->pid);
929 	if (src_has_pid)
930 		get_pid(src->pid);
931 	dst->pid = src->pid;
932 	return 0;
933 }
934 
damon_commit_targets(struct damon_ctx * dst,struct damon_ctx * src)935 static int damon_commit_targets(
936 		struct damon_ctx *dst, struct damon_ctx *src)
937 {
938 	struct damon_target *dst_target, *next, *src_target, *new_target;
939 	int i = 0, j = 0, err;
940 
941 	damon_for_each_target_safe(dst_target, next, dst) {
942 		src_target = damon_nth_target(i++, src);
943 		if (src_target) {
944 			err = damon_commit_target(
945 					dst_target, damon_target_has_pid(dst),
946 					src_target, damon_target_has_pid(src));
947 			if (err)
948 				return err;
949 		} else {
950 			if (damon_target_has_pid(dst))
951 				put_pid(dst_target->pid);
952 			damon_destroy_target(dst_target);
953 		}
954 	}
955 
956 	damon_for_each_target_safe(src_target, next, src) {
957 		if (j++ < i)
958 			continue;
959 		new_target = damon_new_target();
960 		if (!new_target)
961 			return -ENOMEM;
962 		err = damon_commit_target(new_target, false,
963 				src_target, damon_target_has_pid(src));
964 		if (err)
965 			return err;
966 	}
967 	return 0;
968 }
969 
970 /**
971  * damon_commit_ctx() - Commit parameters of a DAMON context to another.
972  * @dst:	The commit destination DAMON context.
973  * @src:	The commit source DAMON context.
974  *
975  * This function copies user-specified parameters from @src to @dst and update
976  * the internal status and results accordingly.  Users should use this function
977  * for context-level parameters update of running context, instead of manual
978  * in-place updates.
979  *
980  * This function should be called from parameters-update safe context, like
981  * DAMON callbacks.
982  */
damon_commit_ctx(struct damon_ctx * dst,struct damon_ctx * src)983 int damon_commit_ctx(struct damon_ctx *dst, struct damon_ctx *src)
984 {
985 	int err;
986 
987 	err = damon_commit_schemes(dst, src);
988 	if (err)
989 		return err;
990 	err = damon_commit_targets(dst, src);
991 	if (err)
992 		return err;
993 	/*
994 	 * schemes and targets should be updated first, since
995 	 * 1. damon_set_attrs() updates monitoring results of targets and
996 	 * next_apply_sis of schemes, and
997 	 * 2. ops update should be done after pid handling is done (target
998 	 *    committing require putting pids).
999 	 */
1000 	err = damon_set_attrs(dst, &src->attrs);
1001 	if (err)
1002 		return err;
1003 	dst->ops = src->ops;
1004 
1005 	return 0;
1006 }
1007 
1008 /**
1009  * damon_nr_running_ctxs() - Return number of currently running contexts.
1010  */
damon_nr_running_ctxs(void)1011 int damon_nr_running_ctxs(void)
1012 {
1013 	int nr_ctxs;
1014 
1015 	mutex_lock(&damon_lock);
1016 	nr_ctxs = nr_running_ctxs;
1017 	mutex_unlock(&damon_lock);
1018 
1019 	return nr_ctxs;
1020 }
1021 
1022 /* Returns the size upper limit for each monitoring region */
damon_region_sz_limit(struct damon_ctx * ctx)1023 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
1024 {
1025 	struct damon_target *t;
1026 	struct damon_region *r;
1027 	unsigned long sz = 0;
1028 
1029 	damon_for_each_target(t, ctx) {
1030 		damon_for_each_region(r, t)
1031 			sz += damon_sz_region(r);
1032 	}
1033 
1034 	if (ctx->attrs.min_nr_regions)
1035 		sz /= ctx->attrs.min_nr_regions;
1036 	if (sz < DAMON_MIN_REGION)
1037 		sz = DAMON_MIN_REGION;
1038 
1039 	return sz;
1040 }
1041 
1042 static int kdamond_fn(void *data);
1043 
1044 /*
1045  * __damon_start() - Starts monitoring with given context.
1046  * @ctx:	monitoring context
1047  *
1048  * This function should be called while damon_lock is hold.
1049  *
1050  * Return: 0 on success, negative error code otherwise.
1051  */
__damon_start(struct damon_ctx * ctx)1052 static int __damon_start(struct damon_ctx *ctx)
1053 {
1054 	int err = -EBUSY;
1055 
1056 	mutex_lock(&ctx->kdamond_lock);
1057 	if (!ctx->kdamond) {
1058 		err = 0;
1059 		reinit_completion(&ctx->kdamond_started);
1060 		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
1061 				nr_running_ctxs);
1062 		if (IS_ERR(ctx->kdamond)) {
1063 			err = PTR_ERR(ctx->kdamond);
1064 			ctx->kdamond = NULL;
1065 		} else {
1066 			wait_for_completion(&ctx->kdamond_started);
1067 		}
1068 	}
1069 	mutex_unlock(&ctx->kdamond_lock);
1070 
1071 	return err;
1072 }
1073 
1074 /**
1075  * damon_start() - Starts the monitorings for a given group of contexts.
1076  * @ctxs:	an array of the pointers for contexts to start monitoring
1077  * @nr_ctxs:	size of @ctxs
1078  * @exclusive:	exclusiveness of this contexts group
1079  *
1080  * This function starts a group of monitoring threads for a group of monitoring
1081  * contexts.  One thread per each context is created and run in parallel.  The
1082  * caller should handle synchronization between the threads by itself.  If
1083  * @exclusive is true and a group of threads that created by other
1084  * 'damon_start()' call is currently running, this function does nothing but
1085  * returns -EBUSY.
1086  *
1087  * Return: 0 on success, negative error code otherwise.
1088  */
damon_start(struct damon_ctx ** ctxs,int nr_ctxs,bool exclusive)1089 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
1090 {
1091 	int i;
1092 	int err = 0;
1093 
1094 	mutex_lock(&damon_lock);
1095 	if ((exclusive && nr_running_ctxs) ||
1096 			(!exclusive && running_exclusive_ctxs)) {
1097 		mutex_unlock(&damon_lock);
1098 		return -EBUSY;
1099 	}
1100 
1101 	for (i = 0; i < nr_ctxs; i++) {
1102 		err = __damon_start(ctxs[i]);
1103 		if (err)
1104 			break;
1105 		nr_running_ctxs++;
1106 	}
1107 	if (exclusive && nr_running_ctxs)
1108 		running_exclusive_ctxs = true;
1109 	mutex_unlock(&damon_lock);
1110 
1111 	return err;
1112 }
1113 
1114 /*
1115  * __damon_stop() - Stops monitoring of a given context.
1116  * @ctx:	monitoring context
1117  *
1118  * Return: 0 on success, negative error code otherwise.
1119  */
__damon_stop(struct damon_ctx * ctx)1120 static int __damon_stop(struct damon_ctx *ctx)
1121 {
1122 	struct task_struct *tsk;
1123 
1124 	mutex_lock(&ctx->kdamond_lock);
1125 	tsk = ctx->kdamond;
1126 	if (tsk) {
1127 		get_task_struct(tsk);
1128 		mutex_unlock(&ctx->kdamond_lock);
1129 		kthread_stop_put(tsk);
1130 		return 0;
1131 	}
1132 	mutex_unlock(&ctx->kdamond_lock);
1133 
1134 	return -EPERM;
1135 }
1136 
1137 /**
1138  * damon_stop() - Stops the monitorings for a given group of contexts.
1139  * @ctxs:	an array of the pointers for contexts to stop monitoring
1140  * @nr_ctxs:	size of @ctxs
1141  *
1142  * Return: 0 on success, negative error code otherwise.
1143  */
damon_stop(struct damon_ctx ** ctxs,int nr_ctxs)1144 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
1145 {
1146 	int i, err = 0;
1147 
1148 	for (i = 0; i < nr_ctxs; i++) {
1149 		/* nr_running_ctxs is decremented in kdamond_fn */
1150 		err = __damon_stop(ctxs[i]);
1151 		if (err)
1152 			break;
1153 	}
1154 	return err;
1155 }
1156 
1157 /*
1158  * Reset the aggregated monitoring results ('nr_accesses' of each region).
1159  */
kdamond_reset_aggregated(struct damon_ctx * c)1160 static void kdamond_reset_aggregated(struct damon_ctx *c)
1161 {
1162 	struct damon_target *t;
1163 	unsigned int ti = 0;	/* target's index */
1164 
1165 	damon_for_each_target(t, c) {
1166 		struct damon_region *r;
1167 
1168 		damon_for_each_region(r, t) {
1169 			trace_damon_aggregated(ti, r, damon_nr_regions(t));
1170 			r->last_nr_accesses = r->nr_accesses;
1171 			r->nr_accesses = 0;
1172 		}
1173 		ti++;
1174 	}
1175 }
1176 
1177 static void damon_split_region_at(struct damon_target *t,
1178 				  struct damon_region *r, unsigned long sz_r);
1179 
__damos_valid_target(struct damon_region * r,struct damos * s)1180 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
1181 {
1182 	unsigned long sz;
1183 	unsigned int nr_accesses = r->nr_accesses_bp / 10000;
1184 
1185 	sz = damon_sz_region(r);
1186 	return s->pattern.min_sz_region <= sz &&
1187 		sz <= s->pattern.max_sz_region &&
1188 		s->pattern.min_nr_accesses <= nr_accesses &&
1189 		nr_accesses <= s->pattern.max_nr_accesses &&
1190 		s->pattern.min_age_region <= r->age &&
1191 		r->age <= s->pattern.max_age_region;
1192 }
1193 
damos_valid_target(struct damon_ctx * c,struct damon_target * t,struct damon_region * r,struct damos * s)1194 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
1195 		struct damon_region *r, struct damos *s)
1196 {
1197 	bool ret = __damos_valid_target(r, s);
1198 
1199 	if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
1200 		return ret;
1201 
1202 	return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
1203 }
1204 
1205 /*
1206  * damos_skip_charged_region() - Check if the given region or starting part of
1207  * it is already charged for the DAMOS quota.
1208  * @t:	The target of the region.
1209  * @rp:	The pointer to the region.
1210  * @s:	The scheme to be applied.
1211  *
1212  * If a quota of a scheme has exceeded in a quota charge window, the scheme's
1213  * action would applied to only a part of the target access pattern fulfilling
1214  * regions.  To avoid applying the scheme action to only already applied
1215  * regions, DAMON skips applying the scheme action to the regions that charged
1216  * in the previous charge window.
1217  *
1218  * This function checks if a given region should be skipped or not for the
1219  * reason.  If only the starting part of the region has previously charged,
1220  * this function splits the region into two so that the second one covers the
1221  * area that not charged in the previous charge widnow and saves the second
1222  * region in *rp and returns false, so that the caller can apply DAMON action
1223  * to the second one.
1224  *
1225  * Return: true if the region should be entirely skipped, false otherwise.
1226  */
damos_skip_charged_region(struct damon_target * t,struct damon_region ** rp,struct damos * s)1227 static bool damos_skip_charged_region(struct damon_target *t,
1228 		struct damon_region **rp, struct damos *s)
1229 {
1230 	struct damon_region *r = *rp;
1231 	struct damos_quota *quota = &s->quota;
1232 	unsigned long sz_to_skip;
1233 
1234 	/* Skip previously charged regions */
1235 	if (quota->charge_target_from) {
1236 		if (t != quota->charge_target_from)
1237 			return true;
1238 		if (r == damon_last_region(t)) {
1239 			quota->charge_target_from = NULL;
1240 			quota->charge_addr_from = 0;
1241 			return true;
1242 		}
1243 		if (quota->charge_addr_from &&
1244 				r->ar.end <= quota->charge_addr_from)
1245 			return true;
1246 
1247 		if (quota->charge_addr_from && r->ar.start <
1248 				quota->charge_addr_from) {
1249 			sz_to_skip = ALIGN_DOWN(quota->charge_addr_from -
1250 					r->ar.start, DAMON_MIN_REGION);
1251 			if (!sz_to_skip) {
1252 				if (damon_sz_region(r) <= DAMON_MIN_REGION)
1253 					return true;
1254 				sz_to_skip = DAMON_MIN_REGION;
1255 			}
1256 			damon_split_region_at(t, r, sz_to_skip);
1257 			r = damon_next_region(r);
1258 			*rp = r;
1259 		}
1260 		quota->charge_target_from = NULL;
1261 		quota->charge_addr_from = 0;
1262 	}
1263 	return false;
1264 }
1265 
damos_update_stat(struct damos * s,unsigned long sz_tried,unsigned long sz_applied)1266 static void damos_update_stat(struct damos *s,
1267 		unsigned long sz_tried, unsigned long sz_applied)
1268 {
1269 	s->stat.nr_tried++;
1270 	s->stat.sz_tried += sz_tried;
1271 	if (sz_applied)
1272 		s->stat.nr_applied++;
1273 	s->stat.sz_applied += sz_applied;
1274 }
1275 
__damos_filter_out(struct damon_ctx * ctx,struct damon_target * t,struct damon_region * r,struct damos_filter * filter)1276 static bool __damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
1277 		struct damon_region *r, struct damos_filter *filter)
1278 {
1279 	bool matched = false;
1280 	struct damon_target *ti;
1281 	int target_idx = 0;
1282 	unsigned long start, end;
1283 
1284 	switch (filter->type) {
1285 	case DAMOS_FILTER_TYPE_TARGET:
1286 		damon_for_each_target(ti, ctx) {
1287 			if (ti == t)
1288 				break;
1289 			target_idx++;
1290 		}
1291 		matched = target_idx == filter->target_idx;
1292 		break;
1293 	case DAMOS_FILTER_TYPE_ADDR:
1294 		start = ALIGN_DOWN(filter->addr_range.start, DAMON_MIN_REGION);
1295 		end = ALIGN_DOWN(filter->addr_range.end, DAMON_MIN_REGION);
1296 
1297 		/* inside the range */
1298 		if (start <= r->ar.start && r->ar.end <= end) {
1299 			matched = true;
1300 			break;
1301 		}
1302 		/* outside of the range */
1303 		if (r->ar.end <= start || end <= r->ar.start) {
1304 			matched = false;
1305 			break;
1306 		}
1307 		/* start before the range and overlap */
1308 		if (r->ar.start < start) {
1309 			damon_split_region_at(t, r, start - r->ar.start);
1310 			matched = false;
1311 			break;
1312 		}
1313 		/* start inside the range */
1314 		damon_split_region_at(t, r, end - r->ar.start);
1315 		matched = true;
1316 		break;
1317 	default:
1318 		return false;
1319 	}
1320 
1321 	return matched == filter->matching;
1322 }
1323 
damos_filter_out(struct damon_ctx * ctx,struct damon_target * t,struct damon_region * r,struct damos * s)1324 static bool damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
1325 		struct damon_region *r, struct damos *s)
1326 {
1327 	struct damos_filter *filter;
1328 
1329 	damos_for_each_filter(filter, s) {
1330 		if (__damos_filter_out(ctx, t, r, filter))
1331 			return true;
1332 	}
1333 	return false;
1334 }
1335 
damos_apply_scheme(struct damon_ctx * c,struct damon_target * t,struct damon_region * r,struct damos * s)1336 static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
1337 		struct damon_region *r, struct damos *s)
1338 {
1339 	struct damos_quota *quota = &s->quota;
1340 	unsigned long sz = damon_sz_region(r);
1341 	struct timespec64 begin, end;
1342 	unsigned long sz_applied = 0;
1343 	int err = 0;
1344 	/*
1345 	 * We plan to support multiple context per kdamond, as DAMON sysfs
1346 	 * implies with 'nr_contexts' file.  Nevertheless, only single context
1347 	 * per kdamond is supported for now.  So, we can simply use '0' context
1348 	 * index here.
1349 	 */
1350 	unsigned int cidx = 0;
1351 	struct damos *siter;		/* schemes iterator */
1352 	unsigned int sidx = 0;
1353 	struct damon_target *titer;	/* targets iterator */
1354 	unsigned int tidx = 0;
1355 	bool do_trace = false;
1356 
1357 	/* get indices for trace_damos_before_apply() */
1358 	if (trace_damos_before_apply_enabled()) {
1359 		damon_for_each_scheme(siter, c) {
1360 			if (siter == s)
1361 				break;
1362 			sidx++;
1363 		}
1364 		damon_for_each_target(titer, c) {
1365 			if (titer == t)
1366 				break;
1367 			tidx++;
1368 		}
1369 		do_trace = true;
1370 	}
1371 
1372 	if (c->ops.apply_scheme) {
1373 		if (quota->esz && quota->charged_sz + sz > quota->esz) {
1374 			sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
1375 					DAMON_MIN_REGION);
1376 			if (!sz)
1377 				goto update_stat;
1378 			damon_split_region_at(t, r, sz);
1379 		}
1380 		if (damos_filter_out(c, t, r, s))
1381 			return;
1382 		ktime_get_coarse_ts64(&begin);
1383 		if (c->callback.before_damos_apply)
1384 			err = c->callback.before_damos_apply(c, t, r, s);
1385 		if (!err) {
1386 			trace_damos_before_apply(cidx, sidx, tidx, r,
1387 					damon_nr_regions(t), do_trace);
1388 			sz_applied = c->ops.apply_scheme(c, t, r, s);
1389 		}
1390 		ktime_get_coarse_ts64(&end);
1391 		quota->total_charged_ns += timespec64_to_ns(&end) -
1392 			timespec64_to_ns(&begin);
1393 		quota->charged_sz += sz;
1394 		if (quota->esz && quota->charged_sz >= quota->esz) {
1395 			quota->charge_target_from = t;
1396 			quota->charge_addr_from = r->ar.end + 1;
1397 		}
1398 	}
1399 	if (s->action != DAMOS_STAT)
1400 		r->age = 0;
1401 
1402 update_stat:
1403 	damos_update_stat(s, sz, sz_applied);
1404 }
1405 
damon_do_apply_schemes(struct damon_ctx * c,struct damon_target * t,struct damon_region * r)1406 static void damon_do_apply_schemes(struct damon_ctx *c,
1407 				   struct damon_target *t,
1408 				   struct damon_region *r)
1409 {
1410 	struct damos *s;
1411 
1412 	damon_for_each_scheme(s, c) {
1413 		struct damos_quota *quota = &s->quota;
1414 
1415 		if (c->passed_sample_intervals < s->next_apply_sis)
1416 			continue;
1417 
1418 		if (!s->wmarks.activated)
1419 			continue;
1420 
1421 		/* Check the quota */
1422 		if (quota->esz && quota->charged_sz >= quota->esz)
1423 			continue;
1424 
1425 		if (damos_skip_charged_region(t, &r, s))
1426 			continue;
1427 
1428 		if (!damos_valid_target(c, t, r, s))
1429 			continue;
1430 
1431 		damos_apply_scheme(c, t, r, s);
1432 	}
1433 }
1434 
1435 /*
1436  * damon_feed_loop_next_input() - get next input to achieve a target score.
1437  * @last_input	The last input.
1438  * @score	Current score that made with @last_input.
1439  *
1440  * Calculate next input to achieve the target score, based on the last input
1441  * and current score.  Assuming the input and the score are positively
1442  * proportional, calculate how much compensation should be added to or
1443  * subtracted from the last input as a proportion of the last input.  Avoid
1444  * next input always being zero by setting it non-zero always.  In short form
1445  * (assuming support of float and signed calculations), the algorithm is as
1446  * below.
1447  *
1448  * next_input = max(last_input * ((goal - current) / goal + 1), 1)
1449  *
1450  * For simple implementation, we assume the target score is always 10,000.  The
1451  * caller should adjust @score for this.
1452  *
1453  * Returns next input that assumed to achieve the target score.
1454  */
damon_feed_loop_next_input(unsigned long last_input,unsigned long score)1455 static unsigned long damon_feed_loop_next_input(unsigned long last_input,
1456 		unsigned long score)
1457 {
1458 	const unsigned long goal = 10000;
1459 	/* Set minimum input as 10000 to avoid compensation be zero */
1460 	const unsigned long min_input = 10000;
1461 	unsigned long score_goal_diff, compensation;
1462 	bool over_achieving = score > goal;
1463 
1464 	if (score == goal)
1465 		return last_input;
1466 	if (score >= goal * 2)
1467 		return min_input;
1468 
1469 	if (over_achieving)
1470 		score_goal_diff = score - goal;
1471 	else
1472 		score_goal_diff = goal - score;
1473 
1474 	if (last_input < ULONG_MAX / score_goal_diff)
1475 		compensation = last_input * score_goal_diff / goal;
1476 	else
1477 		compensation = last_input / goal * score_goal_diff;
1478 
1479 	if (over_achieving)
1480 		return max(last_input - compensation, min_input);
1481 	if (last_input < ULONG_MAX - compensation)
1482 		return last_input + compensation;
1483 	return ULONG_MAX;
1484 }
1485 
1486 #ifdef CONFIG_PSI
1487 
damos_get_some_mem_psi_total(void)1488 static u64 damos_get_some_mem_psi_total(void)
1489 {
1490 	if (static_branch_likely(&psi_disabled))
1491 		return 0;
1492 	return div_u64(psi_system.total[PSI_AVGS][PSI_MEM * 2],
1493 			NSEC_PER_USEC);
1494 }
1495 
1496 #else	/* CONFIG_PSI */
1497 
damos_get_some_mem_psi_total(void)1498 static inline u64 damos_get_some_mem_psi_total(void)
1499 {
1500 	return 0;
1501 };
1502 
1503 #endif	/* CONFIG_PSI */
1504 
damos_set_quota_goal_current_value(struct damos_quota_goal * goal)1505 static void damos_set_quota_goal_current_value(struct damos_quota_goal *goal)
1506 {
1507 	u64 now_psi_total;
1508 
1509 	switch (goal->metric) {
1510 	case DAMOS_QUOTA_USER_INPUT:
1511 		/* User should already set goal->current_value */
1512 		break;
1513 	case DAMOS_QUOTA_SOME_MEM_PSI_US:
1514 		now_psi_total = damos_get_some_mem_psi_total();
1515 		goal->current_value = now_psi_total - goal->last_psi_total;
1516 		goal->last_psi_total = now_psi_total;
1517 		break;
1518 	default:
1519 		break;
1520 	}
1521 }
1522 
1523 /* Return the highest score since it makes schemes least aggressive */
damos_quota_score(struct damos_quota * quota)1524 static unsigned long damos_quota_score(struct damos_quota *quota)
1525 {
1526 	struct damos_quota_goal *goal;
1527 	unsigned long highest_score = 0;
1528 
1529 	damos_for_each_quota_goal(goal, quota) {
1530 		damos_set_quota_goal_current_value(goal);
1531 		highest_score = max(highest_score,
1532 				goal->current_value * 10000 /
1533 				goal->target_value);
1534 	}
1535 
1536 	return highest_score;
1537 }
1538 
1539 /*
1540  * Called only if quota->ms, or quota->sz are set, or quota->goals is not empty
1541  */
damos_set_effective_quota(struct damos_quota * quota)1542 static void damos_set_effective_quota(struct damos_quota *quota)
1543 {
1544 	unsigned long throughput;
1545 	unsigned long esz;
1546 
1547 	if (!quota->ms && list_empty(&quota->goals)) {
1548 		quota->esz = quota->sz;
1549 		return;
1550 	}
1551 
1552 	if (!list_empty(&quota->goals)) {
1553 		unsigned long score = damos_quota_score(quota);
1554 
1555 		quota->esz_bp = damon_feed_loop_next_input(
1556 				max(quota->esz_bp, 10000UL),
1557 				score);
1558 		esz = quota->esz_bp / 10000;
1559 	}
1560 
1561 	if (quota->ms) {
1562 		if (quota->total_charged_ns)
1563 			throughput = quota->total_charged_sz * 1000000 /
1564 				quota->total_charged_ns;
1565 		else
1566 			throughput = PAGE_SIZE * 1024;
1567 		if (!list_empty(&quota->goals))
1568 			esz = min(throughput * quota->ms, esz);
1569 		else
1570 			esz = throughput * quota->ms;
1571 	}
1572 
1573 	if (quota->sz && quota->sz < esz)
1574 		esz = quota->sz;
1575 
1576 	quota->esz = esz;
1577 }
1578 
damos_adjust_quota(struct damon_ctx * c,struct damos * s)1579 static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
1580 {
1581 	struct damos_quota *quota = &s->quota;
1582 	struct damon_target *t;
1583 	struct damon_region *r;
1584 	unsigned long cumulated_sz;
1585 	unsigned int score, max_score = 0;
1586 
1587 	if (!quota->ms && !quota->sz && list_empty(&quota->goals))
1588 		return;
1589 
1590 	/* New charge window starts */
1591 	if (time_after_eq(jiffies, quota->charged_from +
1592 				msecs_to_jiffies(quota->reset_interval))) {
1593 		if (quota->esz && quota->charged_sz >= quota->esz)
1594 			s->stat.qt_exceeds++;
1595 		quota->total_charged_sz += quota->charged_sz;
1596 		quota->charged_from = jiffies;
1597 		quota->charged_sz = 0;
1598 		damos_set_effective_quota(quota);
1599 	}
1600 
1601 	if (!c->ops.get_scheme_score)
1602 		return;
1603 
1604 	/* Fill up the score histogram */
1605 	memset(c->regions_score_histogram, 0,
1606 			sizeof(*c->regions_score_histogram) *
1607 			(DAMOS_MAX_SCORE + 1));
1608 	damon_for_each_target(t, c) {
1609 		damon_for_each_region(r, t) {
1610 			if (!__damos_valid_target(r, s))
1611 				continue;
1612 			score = c->ops.get_scheme_score(c, t, r, s);
1613 			c->regions_score_histogram[score] +=
1614 				damon_sz_region(r);
1615 			if (score > max_score)
1616 				max_score = score;
1617 		}
1618 	}
1619 
1620 	/* Set the min score limit */
1621 	for (cumulated_sz = 0, score = max_score; ; score--) {
1622 		cumulated_sz += c->regions_score_histogram[score];
1623 		if (cumulated_sz >= quota->esz || !score)
1624 			break;
1625 	}
1626 	quota->min_score = score;
1627 }
1628 
kdamond_apply_schemes(struct damon_ctx * c)1629 static void kdamond_apply_schemes(struct damon_ctx *c)
1630 {
1631 	struct damon_target *t;
1632 	struct damon_region *r, *next_r;
1633 	struct damos *s;
1634 	unsigned long sample_interval = c->attrs.sample_interval ?
1635 		c->attrs.sample_interval : 1;
1636 	bool has_schemes_to_apply = false;
1637 
1638 	damon_for_each_scheme(s, c) {
1639 		if (c->passed_sample_intervals < s->next_apply_sis)
1640 			continue;
1641 
1642 		if (!s->wmarks.activated)
1643 			continue;
1644 
1645 		has_schemes_to_apply = true;
1646 
1647 		damos_adjust_quota(c, s);
1648 	}
1649 
1650 	if (!has_schemes_to_apply)
1651 		return;
1652 
1653 	damon_for_each_target(t, c) {
1654 		damon_for_each_region_safe(r, next_r, t)
1655 			damon_do_apply_schemes(c, t, r);
1656 	}
1657 
1658 	damon_for_each_scheme(s, c) {
1659 		if (c->passed_sample_intervals < s->next_apply_sis)
1660 			continue;
1661 		s->next_apply_sis = c->passed_sample_intervals +
1662 			(s->apply_interval_us ? s->apply_interval_us :
1663 			 c->attrs.aggr_interval) / sample_interval;
1664 	}
1665 }
1666 
1667 /*
1668  * Merge two adjacent regions into one region
1669  */
damon_merge_two_regions(struct damon_target * t,struct damon_region * l,struct damon_region * r)1670 static void damon_merge_two_regions(struct damon_target *t,
1671 		struct damon_region *l, struct damon_region *r)
1672 {
1673 	unsigned long sz_l = damon_sz_region(l), sz_r = damon_sz_region(r);
1674 
1675 	l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
1676 			(sz_l + sz_r);
1677 	l->nr_accesses_bp = l->nr_accesses * 10000;
1678 	l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
1679 	l->ar.end = r->ar.end;
1680 	damon_destroy_region(r, t);
1681 }
1682 
1683 /*
1684  * Merge adjacent regions having similar access frequencies
1685  *
1686  * t		target affected by this merge operation
1687  * thres	'->nr_accesses' diff threshold for the merge
1688  * sz_limit	size upper limit of each region
1689  */
damon_merge_regions_of(struct damon_target * t,unsigned int thres,unsigned long sz_limit)1690 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
1691 				   unsigned long sz_limit)
1692 {
1693 	struct damon_region *r, *prev = NULL, *next;
1694 
1695 	damon_for_each_region_safe(r, next, t) {
1696 		if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
1697 			r->age = 0;
1698 		else
1699 			r->age++;
1700 
1701 		if (prev && prev->ar.end == r->ar.start &&
1702 		    abs(prev->nr_accesses - r->nr_accesses) <= thres &&
1703 		    damon_sz_region(prev) + damon_sz_region(r) <= sz_limit)
1704 			damon_merge_two_regions(t, prev, r);
1705 		else
1706 			prev = r;
1707 	}
1708 }
1709 
1710 /*
1711  * Merge adjacent regions having similar access frequencies
1712  *
1713  * threshold	'->nr_accesses' diff threshold for the merge
1714  * sz_limit	size upper limit of each region
1715  *
1716  * This function merges monitoring target regions which are adjacent and their
1717  * access frequencies are similar.  This is for minimizing the monitoring
1718  * overhead under the dynamically changeable access pattern.  If a merge was
1719  * unnecessarily made, later 'kdamond_split_regions()' will revert it.
1720  *
1721  * The total number of regions could be higher than the user-defined limit,
1722  * max_nr_regions for some cases.  For example, the user can update
1723  * max_nr_regions to a number that lower than the current number of regions
1724  * while DAMON is running.  For such a case, repeat merging until the limit is
1725  * met while increasing @threshold up to possible maximum level.
1726  */
kdamond_merge_regions(struct damon_ctx * c,unsigned int threshold,unsigned long sz_limit)1727 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
1728 				  unsigned long sz_limit)
1729 {
1730 	struct damon_target *t;
1731 	unsigned int nr_regions;
1732 	unsigned int max_thres;
1733 
1734 	max_thres = c->attrs.aggr_interval /
1735 		(c->attrs.sample_interval ?  c->attrs.sample_interval : 1);
1736 	do {
1737 		nr_regions = 0;
1738 		damon_for_each_target(t, c) {
1739 			damon_merge_regions_of(t, threshold, sz_limit);
1740 			nr_regions += damon_nr_regions(t);
1741 		}
1742 		threshold = max(1, threshold * 2);
1743 	} while (nr_regions > c->attrs.max_nr_regions &&
1744 			threshold / 2 < max_thres);
1745 }
1746 
1747 /*
1748  * Split a region in two
1749  *
1750  * r		the region to be split
1751  * sz_r		size of the first sub-region that will be made
1752  */
damon_split_region_at(struct damon_target * t,struct damon_region * r,unsigned long sz_r)1753 static void damon_split_region_at(struct damon_target *t,
1754 				  struct damon_region *r, unsigned long sz_r)
1755 {
1756 	struct damon_region *new;
1757 
1758 	new = damon_new_region(r->ar.start + sz_r, r->ar.end);
1759 	if (!new)
1760 		return;
1761 
1762 	r->ar.end = new->ar.start;
1763 
1764 	new->age = r->age;
1765 	new->last_nr_accesses = r->last_nr_accesses;
1766 	new->nr_accesses_bp = r->nr_accesses_bp;
1767 	new->nr_accesses = r->nr_accesses;
1768 
1769 	damon_insert_region(new, r, damon_next_region(r), t);
1770 }
1771 
1772 /* Split every region in the given target into 'nr_subs' regions */
damon_split_regions_of(struct damon_target * t,int nr_subs)1773 static void damon_split_regions_of(struct damon_target *t, int nr_subs)
1774 {
1775 	struct damon_region *r, *next;
1776 	unsigned long sz_region, sz_sub = 0;
1777 	int i;
1778 
1779 	damon_for_each_region_safe(r, next, t) {
1780 		sz_region = damon_sz_region(r);
1781 
1782 		for (i = 0; i < nr_subs - 1 &&
1783 				sz_region > 2 * DAMON_MIN_REGION; i++) {
1784 			/*
1785 			 * Randomly select size of left sub-region to be at
1786 			 * least 10 percent and at most 90% of original region
1787 			 */
1788 			sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
1789 					sz_region / 10, DAMON_MIN_REGION);
1790 			/* Do not allow blank region */
1791 			if (sz_sub == 0 || sz_sub >= sz_region)
1792 				continue;
1793 
1794 			damon_split_region_at(t, r, sz_sub);
1795 			sz_region = sz_sub;
1796 		}
1797 	}
1798 }
1799 
1800 /*
1801  * Split every target region into randomly-sized small regions
1802  *
1803  * This function splits every target region into random-sized small regions if
1804  * current total number of the regions is equal or smaller than half of the
1805  * user-specified maximum number of regions.  This is for maximizing the
1806  * monitoring accuracy under the dynamically changeable access patterns.  If a
1807  * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
1808  * it.
1809  */
kdamond_split_regions(struct damon_ctx * ctx)1810 static void kdamond_split_regions(struct damon_ctx *ctx)
1811 {
1812 	struct damon_target *t;
1813 	unsigned int nr_regions = 0;
1814 	static unsigned int last_nr_regions;
1815 	int nr_subregions = 2;
1816 
1817 	damon_for_each_target(t, ctx)
1818 		nr_regions += damon_nr_regions(t);
1819 
1820 	if (nr_regions > ctx->attrs.max_nr_regions / 2)
1821 		return;
1822 
1823 	/* Maybe the middle of the region has different access frequency */
1824 	if (last_nr_regions == nr_regions &&
1825 			nr_regions < ctx->attrs.max_nr_regions / 3)
1826 		nr_subregions = 3;
1827 
1828 	damon_for_each_target(t, ctx)
1829 		damon_split_regions_of(t, nr_subregions);
1830 
1831 	last_nr_regions = nr_regions;
1832 }
1833 
1834 /*
1835  * Check whether current monitoring should be stopped
1836  *
1837  * The monitoring is stopped when either the user requested to stop, or all
1838  * monitoring targets are invalid.
1839  *
1840  * Returns true if need to stop current monitoring.
1841  */
kdamond_need_stop(struct damon_ctx * ctx)1842 static bool kdamond_need_stop(struct damon_ctx *ctx)
1843 {
1844 	struct damon_target *t;
1845 
1846 	if (kthread_should_stop())
1847 		return true;
1848 
1849 	if (!ctx->ops.target_valid)
1850 		return false;
1851 
1852 	damon_for_each_target(t, ctx) {
1853 		if (ctx->ops.target_valid(t))
1854 			return false;
1855 	}
1856 
1857 	return true;
1858 }
1859 
damos_get_wmark_metric_value(enum damos_wmark_metric metric,unsigned long * metric_value)1860 static int damos_get_wmark_metric_value(enum damos_wmark_metric metric,
1861 					unsigned long *metric_value)
1862 {
1863 	switch (metric) {
1864 	case DAMOS_WMARK_FREE_MEM_RATE:
1865 		*metric_value = global_zone_page_state(NR_FREE_PAGES) * 1000 /
1866 		       totalram_pages();
1867 		return 0;
1868 	default:
1869 		break;
1870 	}
1871 	return -EINVAL;
1872 }
1873 
1874 /*
1875  * Returns zero if the scheme is active.  Else, returns time to wait for next
1876  * watermark check in micro-seconds.
1877  */
damos_wmark_wait_us(struct damos * scheme)1878 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1879 {
1880 	unsigned long metric;
1881 
1882 	if (damos_get_wmark_metric_value(scheme->wmarks.metric, &metric))
1883 		return 0;
1884 
1885 	/* higher than high watermark or lower than low watermark */
1886 	if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1887 		if (scheme->wmarks.activated)
1888 			pr_debug("deactivate a scheme (%d) for %s wmark\n",
1889 					scheme->action,
1890 					metric > scheme->wmarks.high ?
1891 					"high" : "low");
1892 		scheme->wmarks.activated = false;
1893 		return scheme->wmarks.interval;
1894 	}
1895 
1896 	/* inactive and higher than middle watermark */
1897 	if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1898 			!scheme->wmarks.activated)
1899 		return scheme->wmarks.interval;
1900 
1901 	if (!scheme->wmarks.activated)
1902 		pr_debug("activate a scheme (%d)\n", scheme->action);
1903 	scheme->wmarks.activated = true;
1904 	return 0;
1905 }
1906 
kdamond_usleep(unsigned long usecs)1907 static void kdamond_usleep(unsigned long usecs)
1908 {
1909 	/* See Documentation/timers/timers-howto.rst for the thresholds */
1910 	if (usecs > 20 * USEC_PER_MSEC)
1911 		schedule_timeout_idle(usecs_to_jiffies(usecs));
1912 	else
1913 		usleep_idle_range(usecs, usecs + 1);
1914 }
1915 
1916 /* Returns negative error code if it's not activated but should return */
kdamond_wait_activation(struct damon_ctx * ctx)1917 static int kdamond_wait_activation(struct damon_ctx *ctx)
1918 {
1919 	struct damos *s;
1920 	unsigned long wait_time;
1921 	unsigned long min_wait_time = 0;
1922 	bool init_wait_time = false;
1923 
1924 	while (!kdamond_need_stop(ctx)) {
1925 		damon_for_each_scheme(s, ctx) {
1926 			wait_time = damos_wmark_wait_us(s);
1927 			if (!init_wait_time || wait_time < min_wait_time) {
1928 				init_wait_time = true;
1929 				min_wait_time = wait_time;
1930 			}
1931 		}
1932 		if (!min_wait_time)
1933 			return 0;
1934 
1935 		kdamond_usleep(min_wait_time);
1936 
1937 		if (ctx->callback.after_wmarks_check &&
1938 				ctx->callback.after_wmarks_check(ctx))
1939 			break;
1940 	}
1941 	return -EBUSY;
1942 }
1943 
kdamond_init_intervals_sis(struct damon_ctx * ctx)1944 static void kdamond_init_intervals_sis(struct damon_ctx *ctx)
1945 {
1946 	unsigned long sample_interval = ctx->attrs.sample_interval ?
1947 		ctx->attrs.sample_interval : 1;
1948 	unsigned long apply_interval;
1949 	struct damos *scheme;
1950 
1951 	ctx->passed_sample_intervals = 0;
1952 	ctx->next_aggregation_sis = ctx->attrs.aggr_interval / sample_interval;
1953 	ctx->next_ops_update_sis = ctx->attrs.ops_update_interval /
1954 		sample_interval;
1955 
1956 	damon_for_each_scheme(scheme, ctx) {
1957 		apply_interval = scheme->apply_interval_us ?
1958 			scheme->apply_interval_us : ctx->attrs.aggr_interval;
1959 		scheme->next_apply_sis = apply_interval / sample_interval;
1960 	}
1961 }
1962 
1963 /*
1964  * The monitoring daemon that runs as a kernel thread
1965  */
kdamond_fn(void * data)1966 static int kdamond_fn(void *data)
1967 {
1968 	struct damon_ctx *ctx = data;
1969 	struct damon_target *t;
1970 	struct damon_region *r, *next;
1971 	unsigned int max_nr_accesses = 0;
1972 	unsigned long sz_limit = 0;
1973 
1974 	pr_debug("kdamond (%d) starts\n", current->pid);
1975 
1976 	complete(&ctx->kdamond_started);
1977 	kdamond_init_intervals_sis(ctx);
1978 
1979 	if (ctx->ops.init)
1980 		ctx->ops.init(ctx);
1981 	if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1982 		goto done;
1983 	ctx->regions_score_histogram = kmalloc_array(DAMOS_MAX_SCORE + 1,
1984 			sizeof(*ctx->regions_score_histogram), GFP_KERNEL);
1985 	if (!ctx->regions_score_histogram)
1986 		goto done;
1987 
1988 	sz_limit = damon_region_sz_limit(ctx);
1989 
1990 	while (!kdamond_need_stop(ctx)) {
1991 		/*
1992 		 * ctx->attrs and ctx->next_{aggregation,ops_update}_sis could
1993 		 * be changed from after_wmarks_check() or after_aggregation()
1994 		 * callbacks.  Read the values here, and use those for this
1995 		 * iteration.  That is, damon_set_attrs() updated new values
1996 		 * are respected from next iteration.
1997 		 */
1998 		unsigned long next_aggregation_sis = ctx->next_aggregation_sis;
1999 		unsigned long next_ops_update_sis = ctx->next_ops_update_sis;
2000 		unsigned long sample_interval = ctx->attrs.sample_interval;
2001 
2002 		if (kdamond_wait_activation(ctx))
2003 			break;
2004 
2005 		if (ctx->ops.prepare_access_checks)
2006 			ctx->ops.prepare_access_checks(ctx);
2007 		if (ctx->callback.after_sampling &&
2008 				ctx->callback.after_sampling(ctx))
2009 			break;
2010 
2011 		kdamond_usleep(sample_interval);
2012 		ctx->passed_sample_intervals++;
2013 
2014 		if (ctx->ops.check_accesses)
2015 			max_nr_accesses = ctx->ops.check_accesses(ctx);
2016 
2017 		if (ctx->passed_sample_intervals >= next_aggregation_sis) {
2018 			kdamond_merge_regions(ctx,
2019 					max_nr_accesses / 10,
2020 					sz_limit);
2021 			if (ctx->callback.after_aggregation &&
2022 					ctx->callback.after_aggregation(ctx))
2023 				break;
2024 		}
2025 
2026 		/*
2027 		 * do kdamond_apply_schemes() after kdamond_merge_regions() if
2028 		 * possible, to reduce overhead
2029 		 */
2030 		if (!list_empty(&ctx->schemes))
2031 			kdamond_apply_schemes(ctx);
2032 
2033 		sample_interval = ctx->attrs.sample_interval ?
2034 			ctx->attrs.sample_interval : 1;
2035 		if (ctx->passed_sample_intervals >= next_aggregation_sis) {
2036 			ctx->next_aggregation_sis = next_aggregation_sis +
2037 				ctx->attrs.aggr_interval / sample_interval;
2038 
2039 			kdamond_reset_aggregated(ctx);
2040 			kdamond_split_regions(ctx);
2041 			if (ctx->ops.reset_aggregated)
2042 				ctx->ops.reset_aggregated(ctx);
2043 		}
2044 
2045 		if (ctx->passed_sample_intervals >= next_ops_update_sis) {
2046 			ctx->next_ops_update_sis = next_ops_update_sis +
2047 				ctx->attrs.ops_update_interval /
2048 				sample_interval;
2049 			if (ctx->ops.update)
2050 				ctx->ops.update(ctx);
2051 			sz_limit = damon_region_sz_limit(ctx);
2052 		}
2053 	}
2054 done:
2055 	damon_for_each_target(t, ctx) {
2056 		damon_for_each_region_safe(r, next, t)
2057 			damon_destroy_region(r, t);
2058 	}
2059 
2060 	if (ctx->callback.before_terminate)
2061 		ctx->callback.before_terminate(ctx);
2062 	if (ctx->ops.cleanup)
2063 		ctx->ops.cleanup(ctx);
2064 	kfree(ctx->regions_score_histogram);
2065 
2066 	pr_debug("kdamond (%d) finishes\n", current->pid);
2067 	mutex_lock(&ctx->kdamond_lock);
2068 	ctx->kdamond = NULL;
2069 	mutex_unlock(&ctx->kdamond_lock);
2070 
2071 	mutex_lock(&damon_lock);
2072 	nr_running_ctxs--;
2073 	if (!nr_running_ctxs && running_exclusive_ctxs)
2074 		running_exclusive_ctxs = false;
2075 	mutex_unlock(&damon_lock);
2076 
2077 	return 0;
2078 }
2079 
2080 /*
2081  * struct damon_system_ram_region - System RAM resource address region of
2082  *				    [@start, @end).
2083  * @start:	Start address of the region (inclusive).
2084  * @end:	End address of the region (exclusive).
2085  */
2086 struct damon_system_ram_region {
2087 	unsigned long start;
2088 	unsigned long end;
2089 };
2090 
walk_system_ram(struct resource * res,void * arg)2091 static int walk_system_ram(struct resource *res, void *arg)
2092 {
2093 	struct damon_system_ram_region *a = arg;
2094 
2095 	if (a->end - a->start < resource_size(res)) {
2096 		a->start = res->start;
2097 		a->end = res->end;
2098 	}
2099 	return 0;
2100 }
2101 
2102 /*
2103  * Find biggest 'System RAM' resource and store its start and end address in
2104  * @start and @end, respectively.  If no System RAM is found, returns false.
2105  */
damon_find_biggest_system_ram(unsigned long * start,unsigned long * end)2106 static bool damon_find_biggest_system_ram(unsigned long *start,
2107 						unsigned long *end)
2108 
2109 {
2110 	struct damon_system_ram_region arg = {};
2111 
2112 	walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram);
2113 	if (arg.end <= arg.start)
2114 		return false;
2115 
2116 	*start = arg.start;
2117 	*end = arg.end;
2118 	return true;
2119 }
2120 
2121 /**
2122  * damon_set_region_biggest_system_ram_default() - Set the region of the given
2123  * monitoring target as requested, or biggest 'System RAM'.
2124  * @t:		The monitoring target to set the region.
2125  * @start:	The pointer to the start address of the region.
2126  * @end:	The pointer to the end address of the region.
2127  *
2128  * This function sets the region of @t as requested by @start and @end.  If the
2129  * values of @start and @end are zero, however, this function finds the biggest
2130  * 'System RAM' resource and sets the region to cover the resource.  In the
2131  * latter case, this function saves the start and end addresses of the resource
2132  * in @start and @end, respectively.
2133  *
2134  * Return: 0 on success, negative error code otherwise.
2135  */
damon_set_region_biggest_system_ram_default(struct damon_target * t,unsigned long * start,unsigned long * end)2136 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
2137 			unsigned long *start, unsigned long *end)
2138 {
2139 	struct damon_addr_range addr_range;
2140 
2141 	if (*start > *end)
2142 		return -EINVAL;
2143 
2144 	if (!*start && !*end &&
2145 		!damon_find_biggest_system_ram(start, end))
2146 		return -EINVAL;
2147 
2148 	addr_range.start = *start;
2149 	addr_range.end = *end;
2150 	return damon_set_regions(t, &addr_range, 1);
2151 }
2152 
2153 /*
2154  * damon_moving_sum() - Calculate an inferred moving sum value.
2155  * @mvsum:	Inferred sum of the last @len_window values.
2156  * @nomvsum:	Non-moving sum of the last discrete @len_window window values.
2157  * @len_window:	The number of last values to take care of.
2158  * @new_value:	New value that will be added to the pseudo moving sum.
2159  *
2160  * Moving sum (moving average * window size) is good for handling noise, but
2161  * the cost of keeping past values can be high for arbitrary window size.  This
2162  * function implements a lightweight pseudo moving sum function that doesn't
2163  * keep the past window values.
2164  *
2165  * It simply assumes there was no noise in the past, and get the no-noise
2166  * assumed past value to drop from @nomvsum and @len_window.  @nomvsum is a
2167  * non-moving sum of the last window.  For example, if @len_window is 10 and we
2168  * have 25 values, @nomvsum is the sum of the 11th to 20th values of the 25
2169  * values.  Hence, this function simply drops @nomvsum / @len_window from
2170  * given @mvsum and add @new_value.
2171  *
2172  * For example, if @len_window is 10 and @nomvsum is 50, the last 10 values for
2173  * the last window could be vary, e.g., 0, 10, 0, 10, 0, 10, 0, 0, 0, 20.  For
2174  * calculating next moving sum with a new value, we should drop 0 from 50 and
2175  * add the new value.  However, this function assumes it got value 5 for each
2176  * of the last ten times.  Based on the assumption, when the next value is
2177  * measured, it drops the assumed past value, 5 from the current sum, and add
2178  * the new value to get the updated pseduo-moving average.
2179  *
2180  * This means the value could have errors, but the errors will be disappeared
2181  * for every @len_window aligned calls.  For example, if @len_window is 10, the
2182  * pseudo moving sum with 11th value to 19th value would have an error.  But
2183  * the sum with 20th value will not have the error.
2184  *
2185  * Return: Pseudo-moving average after getting the @new_value.
2186  */
damon_moving_sum(unsigned int mvsum,unsigned int nomvsum,unsigned int len_window,unsigned int new_value)2187 static unsigned int damon_moving_sum(unsigned int mvsum, unsigned int nomvsum,
2188 		unsigned int len_window, unsigned int new_value)
2189 {
2190 	return mvsum - nomvsum / len_window + new_value;
2191 }
2192 
2193 /**
2194  * damon_update_region_access_rate() - Update the access rate of a region.
2195  * @r:		The DAMON region to update for its access check result.
2196  * @accessed:	Whether the region has accessed during last sampling interval.
2197  * @attrs:	The damon_attrs of the DAMON context.
2198  *
2199  * Update the access rate of a region with the region's last sampling interval
2200  * access check result.
2201  *
2202  * Usually this will be called by &damon_operations->check_accesses callback.
2203  */
damon_update_region_access_rate(struct damon_region * r,bool accessed,struct damon_attrs * attrs)2204 void damon_update_region_access_rate(struct damon_region *r, bool accessed,
2205 		struct damon_attrs *attrs)
2206 {
2207 	unsigned int len_window = 1;
2208 
2209 	/*
2210 	 * sample_interval can be zero, but cannot be larger than
2211 	 * aggr_interval, owing to validation of damon_set_attrs().
2212 	 */
2213 	if (attrs->sample_interval)
2214 		len_window = damon_max_nr_accesses(attrs);
2215 	r->nr_accesses_bp = damon_moving_sum(r->nr_accesses_bp,
2216 			r->last_nr_accesses * 10000, len_window,
2217 			accessed ? 10000 : 0);
2218 
2219 	if (accessed)
2220 		r->nr_accesses++;
2221 }
2222 
damon_init(void)2223 static int __init damon_init(void)
2224 {
2225 	damon_region_cache = KMEM_CACHE(damon_region, 0);
2226 	if (unlikely(!damon_region_cache)) {
2227 		pr_err("creating damon_region_cache fails\n");
2228 		return -ENOMEM;
2229 	}
2230 
2231 	return 0;
2232 }
2233 
2234 subsys_initcall(damon_init);
2235 
2236 #include "tests/core-kunit.h"
2237