xref: /linux/mm/damon/core.c (revision 9052e9c95d908d6c3d7570aadc8898e1d871c8bb)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Data Access Monitor
4  *
5  * Author: SeongJae Park <sjpark@amazon.de>
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/random.h>
14 #include <linux/slab.h>
15 
16 #define CREATE_TRACE_POINTS
17 #include <trace/events/damon.h>
18 
19 #ifdef CONFIG_DAMON_KUNIT_TEST
20 #undef DAMON_MIN_REGION
21 #define DAMON_MIN_REGION 1
22 #endif
23 
24 /* Get a random number in [l, r) */
25 #define damon_rand(l, r) (l + prandom_u32_max(r - l))
26 
27 static DEFINE_MUTEX(damon_lock);
28 static int nr_running_ctxs;
29 
30 /*
31  * Construct a damon_region struct
32  *
33  * Returns the pointer to the new struct if success, or NULL otherwise
34  */
35 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
36 {
37 	struct damon_region *region;
38 
39 	region = kmalloc(sizeof(*region), GFP_KERNEL);
40 	if (!region)
41 		return NULL;
42 
43 	region->ar.start = start;
44 	region->ar.end = end;
45 	region->nr_accesses = 0;
46 	INIT_LIST_HEAD(&region->list);
47 
48 	return region;
49 }
50 
51 /*
52  * Add a region between two other regions
53  */
54 inline void damon_insert_region(struct damon_region *r,
55 		struct damon_region *prev, struct damon_region *next,
56 		struct damon_target *t)
57 {
58 	__list_add(&r->list, &prev->list, &next->list);
59 	t->nr_regions++;
60 }
61 
62 void damon_add_region(struct damon_region *r, struct damon_target *t)
63 {
64 	list_add_tail(&r->list, &t->regions_list);
65 	t->nr_regions++;
66 }
67 
68 static void damon_del_region(struct damon_region *r, struct damon_target *t)
69 {
70 	list_del(&r->list);
71 	t->nr_regions--;
72 }
73 
74 static void damon_free_region(struct damon_region *r)
75 {
76 	kfree(r);
77 }
78 
79 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
80 {
81 	damon_del_region(r, t);
82 	damon_free_region(r);
83 }
84 
85 /*
86  * Construct a damon_target struct
87  *
88  * Returns the pointer to the new struct if success, or NULL otherwise
89  */
90 struct damon_target *damon_new_target(unsigned long id)
91 {
92 	struct damon_target *t;
93 
94 	t = kmalloc(sizeof(*t), GFP_KERNEL);
95 	if (!t)
96 		return NULL;
97 
98 	t->id = id;
99 	t->nr_regions = 0;
100 	INIT_LIST_HEAD(&t->regions_list);
101 
102 	return t;
103 }
104 
105 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
106 {
107 	list_add_tail(&t->list, &ctx->adaptive_targets);
108 }
109 
110 static void damon_del_target(struct damon_target *t)
111 {
112 	list_del(&t->list);
113 }
114 
115 void damon_free_target(struct damon_target *t)
116 {
117 	struct damon_region *r, *next;
118 
119 	damon_for_each_region_safe(r, next, t)
120 		damon_free_region(r);
121 	kfree(t);
122 }
123 
124 void damon_destroy_target(struct damon_target *t)
125 {
126 	damon_del_target(t);
127 	damon_free_target(t);
128 }
129 
130 unsigned int damon_nr_regions(struct damon_target *t)
131 {
132 	return t->nr_regions;
133 }
134 
135 struct damon_ctx *damon_new_ctx(void)
136 {
137 	struct damon_ctx *ctx;
138 
139 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
140 	if (!ctx)
141 		return NULL;
142 
143 	ctx->sample_interval = 5 * 1000;
144 	ctx->aggr_interval = 100 * 1000;
145 	ctx->primitive_update_interval = 60 * 1000 * 1000;
146 
147 	ktime_get_coarse_ts64(&ctx->last_aggregation);
148 	ctx->last_primitive_update = ctx->last_aggregation;
149 
150 	mutex_init(&ctx->kdamond_lock);
151 
152 	ctx->min_nr_regions = 10;
153 	ctx->max_nr_regions = 1000;
154 
155 	INIT_LIST_HEAD(&ctx->adaptive_targets);
156 
157 	return ctx;
158 }
159 
160 static void damon_destroy_targets(struct damon_ctx *ctx)
161 {
162 	struct damon_target *t, *next_t;
163 
164 	if (ctx->primitive.cleanup) {
165 		ctx->primitive.cleanup(ctx);
166 		return;
167 	}
168 
169 	damon_for_each_target_safe(t, next_t, ctx)
170 		damon_destroy_target(t);
171 }
172 
173 void damon_destroy_ctx(struct damon_ctx *ctx)
174 {
175 	damon_destroy_targets(ctx);
176 	kfree(ctx);
177 }
178 
179 /**
180  * damon_set_targets() - Set monitoring targets.
181  * @ctx:	monitoring context
182  * @ids:	array of target ids
183  * @nr_ids:	number of entries in @ids
184  *
185  * This function should not be called while the kdamond is running.
186  *
187  * Return: 0 on success, negative error code otherwise.
188  */
189 int damon_set_targets(struct damon_ctx *ctx,
190 		      unsigned long *ids, ssize_t nr_ids)
191 {
192 	ssize_t i;
193 	struct damon_target *t, *next;
194 
195 	damon_destroy_targets(ctx);
196 
197 	for (i = 0; i < nr_ids; i++) {
198 		t = damon_new_target(ids[i]);
199 		if (!t) {
200 			pr_err("Failed to alloc damon_target\n");
201 			/* The caller should do cleanup of the ids itself */
202 			damon_for_each_target_safe(t, next, ctx)
203 				damon_destroy_target(t);
204 			return -ENOMEM;
205 		}
206 		damon_add_target(ctx, t);
207 	}
208 
209 	return 0;
210 }
211 
212 /**
213  * damon_set_attrs() - Set attributes for the monitoring.
214  * @ctx:		monitoring context
215  * @sample_int:		time interval between samplings
216  * @aggr_int:		time interval between aggregations
217  * @primitive_upd_int:	time interval between monitoring primitive updates
218  * @min_nr_reg:		minimal number of regions
219  * @max_nr_reg:		maximum number of regions
220  *
221  * This function should not be called while the kdamond is running.
222  * Every time interval is in micro-seconds.
223  *
224  * Return: 0 on success, negative error code otherwise.
225  */
226 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
227 		    unsigned long aggr_int, unsigned long primitive_upd_int,
228 		    unsigned long min_nr_reg, unsigned long max_nr_reg)
229 {
230 	if (min_nr_reg < 3) {
231 		pr_err("min_nr_regions (%lu) must be at least 3\n",
232 				min_nr_reg);
233 		return -EINVAL;
234 	}
235 	if (min_nr_reg > max_nr_reg) {
236 		pr_err("invalid nr_regions.  min (%lu) > max (%lu)\n",
237 				min_nr_reg, max_nr_reg);
238 		return -EINVAL;
239 	}
240 
241 	ctx->sample_interval = sample_int;
242 	ctx->aggr_interval = aggr_int;
243 	ctx->primitive_update_interval = primitive_upd_int;
244 	ctx->min_nr_regions = min_nr_reg;
245 	ctx->max_nr_regions = max_nr_reg;
246 
247 	return 0;
248 }
249 
250 /**
251  * damon_nr_running_ctxs() - Return number of currently running contexts.
252  */
253 int damon_nr_running_ctxs(void)
254 {
255 	int nr_ctxs;
256 
257 	mutex_lock(&damon_lock);
258 	nr_ctxs = nr_running_ctxs;
259 	mutex_unlock(&damon_lock);
260 
261 	return nr_ctxs;
262 }
263 
264 /* Returns the size upper limit for each monitoring region */
265 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
266 {
267 	struct damon_target *t;
268 	struct damon_region *r;
269 	unsigned long sz = 0;
270 
271 	damon_for_each_target(t, ctx) {
272 		damon_for_each_region(r, t)
273 			sz += r->ar.end - r->ar.start;
274 	}
275 
276 	if (ctx->min_nr_regions)
277 		sz /= ctx->min_nr_regions;
278 	if (sz < DAMON_MIN_REGION)
279 		sz = DAMON_MIN_REGION;
280 
281 	return sz;
282 }
283 
284 static bool damon_kdamond_running(struct damon_ctx *ctx)
285 {
286 	bool running;
287 
288 	mutex_lock(&ctx->kdamond_lock);
289 	running = ctx->kdamond != NULL;
290 	mutex_unlock(&ctx->kdamond_lock);
291 
292 	return running;
293 }
294 
295 static int kdamond_fn(void *data);
296 
297 /*
298  * __damon_start() - Starts monitoring with given context.
299  * @ctx:	monitoring context
300  *
301  * This function should be called while damon_lock is hold.
302  *
303  * Return: 0 on success, negative error code otherwise.
304  */
305 static int __damon_start(struct damon_ctx *ctx)
306 {
307 	int err = -EBUSY;
308 
309 	mutex_lock(&ctx->kdamond_lock);
310 	if (!ctx->kdamond) {
311 		err = 0;
312 		ctx->kdamond_stop = false;
313 		ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
314 				nr_running_ctxs);
315 		if (IS_ERR(ctx->kdamond)) {
316 			err = PTR_ERR(ctx->kdamond);
317 			ctx->kdamond = 0;
318 		}
319 	}
320 	mutex_unlock(&ctx->kdamond_lock);
321 
322 	return err;
323 }
324 
325 /**
326  * damon_start() - Starts the monitorings for a given group of contexts.
327  * @ctxs:	an array of the pointers for contexts to start monitoring
328  * @nr_ctxs:	size of @ctxs
329  *
330  * This function starts a group of monitoring threads for a group of monitoring
331  * contexts.  One thread per each context is created and run in parallel.  The
332  * caller should handle synchronization between the threads by itself.  If a
333  * group of threads that created by other 'damon_start()' call is currently
334  * running, this function does nothing but returns -EBUSY.
335  *
336  * Return: 0 on success, negative error code otherwise.
337  */
338 int damon_start(struct damon_ctx **ctxs, int nr_ctxs)
339 {
340 	int i;
341 	int err = 0;
342 
343 	mutex_lock(&damon_lock);
344 	if (nr_running_ctxs) {
345 		mutex_unlock(&damon_lock);
346 		return -EBUSY;
347 	}
348 
349 	for (i = 0; i < nr_ctxs; i++) {
350 		err = __damon_start(ctxs[i]);
351 		if (err)
352 			break;
353 		nr_running_ctxs++;
354 	}
355 	mutex_unlock(&damon_lock);
356 
357 	return err;
358 }
359 
360 /*
361  * __damon_stop() - Stops monitoring of given context.
362  * @ctx:	monitoring context
363  *
364  * Return: 0 on success, negative error code otherwise.
365  */
366 static int __damon_stop(struct damon_ctx *ctx)
367 {
368 	mutex_lock(&ctx->kdamond_lock);
369 	if (ctx->kdamond) {
370 		ctx->kdamond_stop = true;
371 		mutex_unlock(&ctx->kdamond_lock);
372 		while (damon_kdamond_running(ctx))
373 			usleep_range(ctx->sample_interval,
374 					ctx->sample_interval * 2);
375 		return 0;
376 	}
377 	mutex_unlock(&ctx->kdamond_lock);
378 
379 	return -EPERM;
380 }
381 
382 /**
383  * damon_stop() - Stops the monitorings for a given group of contexts.
384  * @ctxs:	an array of the pointers for contexts to stop monitoring
385  * @nr_ctxs:	size of @ctxs
386  *
387  * Return: 0 on success, negative error code otherwise.
388  */
389 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
390 {
391 	int i, err = 0;
392 
393 	for (i = 0; i < nr_ctxs; i++) {
394 		/* nr_running_ctxs is decremented in kdamond_fn */
395 		err = __damon_stop(ctxs[i]);
396 		if (err)
397 			return err;
398 	}
399 
400 	return err;
401 }
402 
403 /*
404  * damon_check_reset_time_interval() - Check if a time interval is elapsed.
405  * @baseline:	the time to check whether the interval has elapsed since
406  * @interval:	the time interval (microseconds)
407  *
408  * See whether the given time interval has passed since the given baseline
409  * time.  If so, it also updates the baseline to current time for next check.
410  *
411  * Return:	true if the time interval has passed, or false otherwise.
412  */
413 static bool damon_check_reset_time_interval(struct timespec64 *baseline,
414 		unsigned long interval)
415 {
416 	struct timespec64 now;
417 
418 	ktime_get_coarse_ts64(&now);
419 	if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
420 			interval * 1000)
421 		return false;
422 	*baseline = now;
423 	return true;
424 }
425 
426 /*
427  * Check whether it is time to flush the aggregated information
428  */
429 static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
430 {
431 	return damon_check_reset_time_interval(&ctx->last_aggregation,
432 			ctx->aggr_interval);
433 }
434 
435 /*
436  * Reset the aggregated monitoring results ('nr_accesses' of each region).
437  */
438 static void kdamond_reset_aggregated(struct damon_ctx *c)
439 {
440 	struct damon_target *t;
441 
442 	damon_for_each_target(t, c) {
443 		struct damon_region *r;
444 
445 		damon_for_each_region(r, t) {
446 			trace_damon_aggregated(t, r, damon_nr_regions(t));
447 			r->nr_accesses = 0;
448 		}
449 	}
450 }
451 
452 #define sz_damon_region(r) (r->ar.end - r->ar.start)
453 
454 /*
455  * Merge two adjacent regions into one region
456  */
457 static void damon_merge_two_regions(struct damon_target *t,
458 		struct damon_region *l, struct damon_region *r)
459 {
460 	unsigned long sz_l = sz_damon_region(l), sz_r = sz_damon_region(r);
461 
462 	l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
463 			(sz_l + sz_r);
464 	l->ar.end = r->ar.end;
465 	damon_destroy_region(r, t);
466 }
467 
468 #define diff_of(a, b) (a > b ? a - b : b - a)
469 
470 /*
471  * Merge adjacent regions having similar access frequencies
472  *
473  * t		target affected by this merge operation
474  * thres	'->nr_accesses' diff threshold for the merge
475  * sz_limit	size upper limit of each region
476  */
477 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
478 				   unsigned long sz_limit)
479 {
480 	struct damon_region *r, *prev = NULL, *next;
481 
482 	damon_for_each_region_safe(r, next, t) {
483 		if (prev && prev->ar.end == r->ar.start &&
484 		    diff_of(prev->nr_accesses, r->nr_accesses) <= thres &&
485 		    sz_damon_region(prev) + sz_damon_region(r) <= sz_limit)
486 			damon_merge_two_regions(t, prev, r);
487 		else
488 			prev = r;
489 	}
490 }
491 
492 /*
493  * Merge adjacent regions having similar access frequencies
494  *
495  * threshold	'->nr_accesses' diff threshold for the merge
496  * sz_limit	size upper limit of each region
497  *
498  * This function merges monitoring target regions which are adjacent and their
499  * access frequencies are similar.  This is for minimizing the monitoring
500  * overhead under the dynamically changeable access pattern.  If a merge was
501  * unnecessarily made, later 'kdamond_split_regions()' will revert it.
502  */
503 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
504 				  unsigned long sz_limit)
505 {
506 	struct damon_target *t;
507 
508 	damon_for_each_target(t, c)
509 		damon_merge_regions_of(t, threshold, sz_limit);
510 }
511 
512 /*
513  * Split a region in two
514  *
515  * r		the region to be split
516  * sz_r		size of the first sub-region that will be made
517  */
518 static void damon_split_region_at(struct damon_ctx *ctx,
519 		struct damon_target *t, struct damon_region *r,
520 		unsigned long sz_r)
521 {
522 	struct damon_region *new;
523 
524 	new = damon_new_region(r->ar.start + sz_r, r->ar.end);
525 	if (!new)
526 		return;
527 
528 	r->ar.end = new->ar.start;
529 
530 	damon_insert_region(new, r, damon_next_region(r), t);
531 }
532 
533 /* Split every region in the given target into 'nr_subs' regions */
534 static void damon_split_regions_of(struct damon_ctx *ctx,
535 				     struct damon_target *t, int nr_subs)
536 {
537 	struct damon_region *r, *next;
538 	unsigned long sz_region, sz_sub = 0;
539 	int i;
540 
541 	damon_for_each_region_safe(r, next, t) {
542 		sz_region = r->ar.end - r->ar.start;
543 
544 		for (i = 0; i < nr_subs - 1 &&
545 				sz_region > 2 * DAMON_MIN_REGION; i++) {
546 			/*
547 			 * Randomly select size of left sub-region to be at
548 			 * least 10 percent and at most 90% of original region
549 			 */
550 			sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
551 					sz_region / 10, DAMON_MIN_REGION);
552 			/* Do not allow blank region */
553 			if (sz_sub == 0 || sz_sub >= sz_region)
554 				continue;
555 
556 			damon_split_region_at(ctx, t, r, sz_sub);
557 			sz_region = sz_sub;
558 		}
559 	}
560 }
561 
562 /*
563  * Split every target region into randomly-sized small regions
564  *
565  * This function splits every target region into random-sized small regions if
566  * current total number of the regions is equal or smaller than half of the
567  * user-specified maximum number of regions.  This is for maximizing the
568  * monitoring accuracy under the dynamically changeable access patterns.  If a
569  * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
570  * it.
571  */
572 static void kdamond_split_regions(struct damon_ctx *ctx)
573 {
574 	struct damon_target *t;
575 	unsigned int nr_regions = 0;
576 	static unsigned int last_nr_regions;
577 	int nr_subregions = 2;
578 
579 	damon_for_each_target(t, ctx)
580 		nr_regions += damon_nr_regions(t);
581 
582 	if (nr_regions > ctx->max_nr_regions / 2)
583 		return;
584 
585 	/* Maybe the middle of the region has different access frequency */
586 	if (last_nr_regions == nr_regions &&
587 			nr_regions < ctx->max_nr_regions / 3)
588 		nr_subregions = 3;
589 
590 	damon_for_each_target(t, ctx)
591 		damon_split_regions_of(ctx, t, nr_subregions);
592 
593 	last_nr_regions = nr_regions;
594 }
595 
596 /*
597  * Check whether it is time to check and apply the target monitoring regions
598  *
599  * Returns true if it is.
600  */
601 static bool kdamond_need_update_primitive(struct damon_ctx *ctx)
602 {
603 	return damon_check_reset_time_interval(&ctx->last_primitive_update,
604 			ctx->primitive_update_interval);
605 }
606 
607 /*
608  * Check whether current monitoring should be stopped
609  *
610  * The monitoring is stopped when either the user requested to stop, or all
611  * monitoring targets are invalid.
612  *
613  * Returns true if need to stop current monitoring.
614  */
615 static bool kdamond_need_stop(struct damon_ctx *ctx)
616 {
617 	struct damon_target *t;
618 	bool stop;
619 
620 	mutex_lock(&ctx->kdamond_lock);
621 	stop = ctx->kdamond_stop;
622 	mutex_unlock(&ctx->kdamond_lock);
623 	if (stop)
624 		return true;
625 
626 	if (!ctx->primitive.target_valid)
627 		return false;
628 
629 	damon_for_each_target(t, ctx) {
630 		if (ctx->primitive.target_valid(t))
631 			return false;
632 	}
633 
634 	return true;
635 }
636 
637 static void set_kdamond_stop(struct damon_ctx *ctx)
638 {
639 	mutex_lock(&ctx->kdamond_lock);
640 	ctx->kdamond_stop = true;
641 	mutex_unlock(&ctx->kdamond_lock);
642 }
643 
644 /*
645  * The monitoring daemon that runs as a kernel thread
646  */
647 static int kdamond_fn(void *data)
648 {
649 	struct damon_ctx *ctx = (struct damon_ctx *)data;
650 	struct damon_target *t;
651 	struct damon_region *r, *next;
652 	unsigned int max_nr_accesses = 0;
653 	unsigned long sz_limit = 0;
654 
655 	mutex_lock(&ctx->kdamond_lock);
656 	pr_info("kdamond (%d) starts\n", ctx->kdamond->pid);
657 	mutex_unlock(&ctx->kdamond_lock);
658 
659 	if (ctx->primitive.init)
660 		ctx->primitive.init(ctx);
661 	if (ctx->callback.before_start && ctx->callback.before_start(ctx))
662 		set_kdamond_stop(ctx);
663 
664 	sz_limit = damon_region_sz_limit(ctx);
665 
666 	while (!kdamond_need_stop(ctx)) {
667 		if (ctx->primitive.prepare_access_checks)
668 			ctx->primitive.prepare_access_checks(ctx);
669 		if (ctx->callback.after_sampling &&
670 				ctx->callback.after_sampling(ctx))
671 			set_kdamond_stop(ctx);
672 
673 		usleep_range(ctx->sample_interval, ctx->sample_interval + 1);
674 
675 		if (ctx->primitive.check_accesses)
676 			max_nr_accesses = ctx->primitive.check_accesses(ctx);
677 
678 		if (kdamond_aggregate_interval_passed(ctx)) {
679 			kdamond_merge_regions(ctx,
680 					max_nr_accesses / 10,
681 					sz_limit);
682 			if (ctx->callback.after_aggregation &&
683 					ctx->callback.after_aggregation(ctx))
684 				set_kdamond_stop(ctx);
685 			kdamond_reset_aggregated(ctx);
686 			kdamond_split_regions(ctx);
687 			if (ctx->primitive.reset_aggregated)
688 				ctx->primitive.reset_aggregated(ctx);
689 		}
690 
691 		if (kdamond_need_update_primitive(ctx)) {
692 			if (ctx->primitive.update)
693 				ctx->primitive.update(ctx);
694 			sz_limit = damon_region_sz_limit(ctx);
695 		}
696 	}
697 	damon_for_each_target(t, ctx) {
698 		damon_for_each_region_safe(r, next, t)
699 			damon_destroy_region(r, t);
700 	}
701 
702 	if (ctx->callback.before_terminate &&
703 			ctx->callback.before_terminate(ctx))
704 		set_kdamond_stop(ctx);
705 	if (ctx->primitive.cleanup)
706 		ctx->primitive.cleanup(ctx);
707 
708 	pr_debug("kdamond (%d) finishes\n", ctx->kdamond->pid);
709 	mutex_lock(&ctx->kdamond_lock);
710 	ctx->kdamond = NULL;
711 	mutex_unlock(&ctx->kdamond_lock);
712 
713 	mutex_lock(&damon_lock);
714 	nr_running_ctxs--;
715 	mutex_unlock(&damon_lock);
716 
717 	do_exit(0);
718 }
719 
720 #include "core-test.h"
721