1 // SPDX-License-Identifier: GPL-2.0
5 * Author: SeongJae Park <sjpark@amazon.de>
8 #define pr_fmt(fmt) "damon: " fmt
10 #include <linux/damon.h>
11 #include <linux/delay.h>
12 #include <linux/kthread.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
17 #define CREATE_TRACE_POINTS
18 #include <trace/events/damon.h>
20 #ifdef CONFIG_DAMON_KUNIT_TEST
21 #undef DAMON_MIN_REGION
22 #define DAMON_MIN_REGION 1
25 static DEFINE_MUTEX(damon_lock);
26 static int nr_running_ctxs;
27 static bool running_exclusive_ctxs;
29 static DEFINE_MUTEX(damon_ops_lock);
30 static struct damon_operations damon_registered_ops[NR_DAMON_OPS];
32 static struct kmem_cache *damon_region_cache __ro_after_init;
34 /* Should be called under damon_ops_lock with id smaller than NR_DAMON_OPS */
35 static bool __damon_is_registered_ops(enum damon_ops_id id)
37 struct damon_operations empty_ops = {};
39 if (!memcmp(&empty_ops, &damon_registered_ops[id], sizeof(empty_ops)))
45 * damon_is_registered_ops() - Check if a given damon_operations is registered.
46 * @id: Id of the damon_operations to check if registered.
48 * Return: true if the ops is set, false otherwise.
50 bool damon_is_registered_ops(enum damon_ops_id id)
54 if (id >= NR_DAMON_OPS)
56 mutex_lock(&damon_ops_lock);
57 registered = __damon_is_registered_ops(id);
58 mutex_unlock(&damon_ops_lock);
63 * damon_register_ops() - Register a monitoring operations set to DAMON.
64 * @ops: monitoring operations set to register.
66 * This function registers a monitoring operations set of valid &struct
67 * damon_operations->id so that others can find and use them later.
69 * Return: 0 on success, negative error code otherwise.
71 int damon_register_ops(struct damon_operations *ops)
75 if (ops->id >= NR_DAMON_OPS)
77 mutex_lock(&damon_ops_lock);
78 /* Fail for already registered ops */
79 if (__damon_is_registered_ops(ops->id)) {
83 damon_registered_ops[ops->id] = *ops;
85 mutex_unlock(&damon_ops_lock);
90 * damon_select_ops() - Select a monitoring operations to use with the context.
91 * @ctx: monitoring context to use the operations.
92 * @id: id of the registered monitoring operations to select.
94 * This function finds registered monitoring operations set of @id and make
97 * Return: 0 on success, negative error code otherwise.
99 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id)
103 if (id >= NR_DAMON_OPS)
106 mutex_lock(&damon_ops_lock);
107 if (!__damon_is_registered_ops(id))
110 ctx->ops = damon_registered_ops[id];
111 mutex_unlock(&damon_ops_lock);
116 * Construct a damon_region struct
118 * Returns the pointer to the new struct if success, or NULL otherwise
120 struct damon_region *damon_new_region(unsigned long start, unsigned long end)
122 struct damon_region *region;
124 region = kmem_cache_alloc(damon_region_cache, GFP_KERNEL);
128 region->ar.start = start;
129 region->ar.end = end;
130 region->nr_accesses = 0;
131 INIT_LIST_HEAD(®ion->list);
134 region->last_nr_accesses = 0;
139 void damon_add_region(struct damon_region *r, struct damon_target *t)
141 list_add_tail(&r->list, &t->regions_list);
145 static void damon_del_region(struct damon_region *r, struct damon_target *t)
151 static void damon_free_region(struct damon_region *r)
153 kmem_cache_free(damon_region_cache, r);
156 void damon_destroy_region(struct damon_region *r, struct damon_target *t)
158 damon_del_region(r, t);
159 damon_free_region(r);
163 * Check whether a region is intersecting an address range
165 * Returns true if it is.
167 static bool damon_intersect(struct damon_region *r,
168 struct damon_addr_range *re)
170 return !(r->ar.end <= re->start || re->end <= r->ar.start);
174 * Fill holes in regions with new regions.
176 static int damon_fill_regions_holes(struct damon_region *first,
177 struct damon_region *last, struct damon_target *t)
179 struct damon_region *r = first;
181 damon_for_each_region_from(r, t) {
182 struct damon_region *next, *newr;
186 next = damon_next_region(r);
187 if (r->ar.end != next->ar.start) {
188 newr = damon_new_region(r->ar.end, next->ar.start);
191 damon_insert_region(newr, r, next, t);
198 * damon_set_regions() - Set regions of a target for given address ranges.
199 * @t: the given target.
200 * @ranges: array of new monitoring target ranges.
201 * @nr_ranges: length of @ranges.
203 * This function adds new regions to, or modify existing regions of a
204 * monitoring target to fit in specific ranges.
206 * Return: 0 if success, or negative error code otherwise.
208 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
209 unsigned int nr_ranges)
211 struct damon_region *r, *next;
215 /* Remove regions which are not in the new ranges */
216 damon_for_each_region_safe(r, next, t) {
217 for (i = 0; i < nr_ranges; i++) {
218 if (damon_intersect(r, &ranges[i]))
222 damon_destroy_region(r, t);
225 r = damon_first_region(t);
226 /* Add new regions or resize existing regions to fit in the ranges */
227 for (i = 0; i < nr_ranges; i++) {
228 struct damon_region *first = NULL, *last, *newr;
229 struct damon_addr_range *range;
232 /* Get the first/last regions intersecting with the range */
233 damon_for_each_region_from(r, t) {
234 if (damon_intersect(r, range)) {
239 if (r->ar.start >= range->end)
243 /* no region intersects with this range */
244 newr = damon_new_region(
245 ALIGN_DOWN(range->start,
247 ALIGN(range->end, DAMON_MIN_REGION));
250 damon_insert_region(newr, damon_prev_region(r), r, t);
252 /* resize intersecting regions to fit in this range */
253 first->ar.start = ALIGN_DOWN(range->start,
255 last->ar.end = ALIGN(range->end, DAMON_MIN_REGION);
257 /* fill possible holes in the range */
258 err = damon_fill_regions_holes(first, last, t);
266 struct damos_filter *damos_new_filter(enum damos_filter_type type,
269 struct damos_filter *filter;
271 filter = kmalloc(sizeof(*filter), GFP_KERNEL);
275 filter->matching = matching;
276 INIT_LIST_HEAD(&filter->list);
280 void damos_add_filter(struct damos *s, struct damos_filter *f)
282 list_add_tail(&f->list, &s->filters);
285 static void damos_del_filter(struct damos_filter *f)
290 static void damos_free_filter(struct damos_filter *f)
295 void damos_destroy_filter(struct damos_filter *f)
298 damos_free_filter(f);
301 /* initialize private fields of damos_quota and return the pointer */
302 static struct damos_quota *damos_quota_init_priv(struct damos_quota *quota)
304 quota->total_charged_sz = 0;
305 quota->total_charged_ns = 0;
307 quota->charged_sz = 0;
308 quota->charged_from = 0;
309 quota->charge_target_from = NULL;
310 quota->charge_addr_from = 0;
314 struct damos *damon_new_scheme(struct damos_access_pattern *pattern,
315 enum damos_action action, struct damos_quota *quota,
316 struct damos_watermarks *wmarks)
318 struct damos *scheme;
320 scheme = kmalloc(sizeof(*scheme), GFP_KERNEL);
323 scheme->pattern = *pattern;
324 scheme->action = action;
325 INIT_LIST_HEAD(&scheme->filters);
326 scheme->stat = (struct damos_stat){};
327 INIT_LIST_HEAD(&scheme->list);
329 scheme->quota = *(damos_quota_init_priv(quota));
331 scheme->wmarks = *wmarks;
332 scheme->wmarks.activated = true;
337 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s)
339 list_add_tail(&s->list, &ctx->schemes);
342 static void damon_del_scheme(struct damos *s)
347 static void damon_free_scheme(struct damos *s)
352 void damon_destroy_scheme(struct damos *s)
354 struct damos_filter *f, *next;
356 damos_for_each_filter_safe(f, next, s)
357 damos_destroy_filter(f);
359 damon_free_scheme(s);
363 * Construct a damon_target struct
365 * Returns the pointer to the new struct if success, or NULL otherwise
367 struct damon_target *damon_new_target(void)
369 struct damon_target *t;
371 t = kmalloc(sizeof(*t), GFP_KERNEL);
377 INIT_LIST_HEAD(&t->regions_list);
378 INIT_LIST_HEAD(&t->list);
383 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)
385 list_add_tail(&t->list, &ctx->adaptive_targets);
388 bool damon_targets_empty(struct damon_ctx *ctx)
390 return list_empty(&ctx->adaptive_targets);
393 static void damon_del_target(struct damon_target *t)
398 void damon_free_target(struct damon_target *t)
400 struct damon_region *r, *next;
402 damon_for_each_region_safe(r, next, t)
403 damon_free_region(r);
407 void damon_destroy_target(struct damon_target *t)
410 damon_free_target(t);
413 unsigned int damon_nr_regions(struct damon_target *t)
415 return t->nr_regions;
418 struct damon_ctx *damon_new_ctx(void)
420 struct damon_ctx *ctx;
422 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
426 ctx->attrs.sample_interval = 5 * 1000;
427 ctx->attrs.aggr_interval = 100 * 1000;
428 ctx->attrs.ops_update_interval = 60 * 1000 * 1000;
430 ktime_get_coarse_ts64(&ctx->last_aggregation);
431 ctx->last_ops_update = ctx->last_aggregation;
433 mutex_init(&ctx->kdamond_lock);
435 ctx->attrs.min_nr_regions = 10;
436 ctx->attrs.max_nr_regions = 1000;
438 INIT_LIST_HEAD(&ctx->adaptive_targets);
439 INIT_LIST_HEAD(&ctx->schemes);
444 static void damon_destroy_targets(struct damon_ctx *ctx)
446 struct damon_target *t, *next_t;
448 if (ctx->ops.cleanup) {
449 ctx->ops.cleanup(ctx);
453 damon_for_each_target_safe(t, next_t, ctx)
454 damon_destroy_target(t);
457 void damon_destroy_ctx(struct damon_ctx *ctx)
459 struct damos *s, *next_s;
461 damon_destroy_targets(ctx);
463 damon_for_each_scheme_safe(s, next_s, ctx)
464 damon_destroy_scheme(s);
469 static unsigned int damon_age_for_new_attrs(unsigned int age,
470 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
472 return age * old_attrs->aggr_interval / new_attrs->aggr_interval;
475 /* convert access ratio in bp (per 10,000) to nr_accesses */
476 static unsigned int damon_accesses_bp_to_nr_accesses(
477 unsigned int accesses_bp, struct damon_attrs *attrs)
479 unsigned int max_nr_accesses =
480 attrs->aggr_interval / attrs->sample_interval;
482 return accesses_bp * max_nr_accesses / 10000;
485 /* convert nr_accesses to access ratio in bp (per 10,000) */
486 static unsigned int damon_nr_accesses_to_accesses_bp(
487 unsigned int nr_accesses, struct damon_attrs *attrs)
489 unsigned int max_nr_accesses =
490 attrs->aggr_interval / attrs->sample_interval;
492 return nr_accesses * 10000 / max_nr_accesses;
495 static unsigned int damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,
496 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
498 return damon_accesses_bp_to_nr_accesses(
499 damon_nr_accesses_to_accesses_bp(
500 nr_accesses, old_attrs),
504 static void damon_update_monitoring_result(struct damon_region *r,
505 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
507 r->nr_accesses = damon_nr_accesses_for_new_attrs(r->nr_accesses,
508 old_attrs, new_attrs);
509 r->age = damon_age_for_new_attrs(r->age, old_attrs, new_attrs);
513 * region->nr_accesses is the number of sampling intervals in the last
514 * aggregation interval that access to the region has found, and region->age is
515 * the number of aggregation intervals that its access pattern has maintained.
516 * For the reason, the real meaning of the two fields depend on current
517 * sampling interval and aggregation interval. This function updates
518 * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.
520 static void damon_update_monitoring_results(struct damon_ctx *ctx,
521 struct damon_attrs *new_attrs)
523 struct damon_attrs *old_attrs = &ctx->attrs;
524 struct damon_target *t;
525 struct damon_region *r;
527 /* if any interval is zero, simply forgive conversion */
528 if (!old_attrs->sample_interval || !old_attrs->aggr_interval ||
529 !new_attrs->sample_interval ||
530 !new_attrs->aggr_interval)
533 damon_for_each_target(t, ctx)
534 damon_for_each_region(r, t)
535 damon_update_monitoring_result(
536 r, old_attrs, new_attrs);
540 * damon_set_attrs() - Set attributes for the monitoring.
541 * @ctx: monitoring context
542 * @attrs: monitoring attributes
544 * This function should not be called while the kdamond is running.
545 * Every time interval is in micro-seconds.
547 * Return: 0 on success, negative error code otherwise.
549 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
551 if (attrs->min_nr_regions < 3)
553 if (attrs->min_nr_regions > attrs->max_nr_regions)
555 if (attrs->sample_interval > attrs->aggr_interval)
558 damon_update_monitoring_results(ctx, attrs);
564 * damon_set_schemes() - Set data access monitoring based operation schemes.
565 * @ctx: monitoring context
566 * @schemes: array of the schemes
567 * @nr_schemes: number of entries in @schemes
569 * This function should not be called while the kdamond of the context is
572 void damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
575 struct damos *s, *next;
578 damon_for_each_scheme_safe(s, next, ctx)
579 damon_destroy_scheme(s);
580 for (i = 0; i < nr_schemes; i++)
581 damon_add_scheme(ctx, schemes[i]);
585 * damon_nr_running_ctxs() - Return number of currently running contexts.
587 int damon_nr_running_ctxs(void)
591 mutex_lock(&damon_lock);
592 nr_ctxs = nr_running_ctxs;
593 mutex_unlock(&damon_lock);
598 /* Returns the size upper limit for each monitoring region */
599 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
601 struct damon_target *t;
602 struct damon_region *r;
603 unsigned long sz = 0;
605 damon_for_each_target(t, ctx) {
606 damon_for_each_region(r, t)
607 sz += damon_sz_region(r);
610 if (ctx->attrs.min_nr_regions)
611 sz /= ctx->attrs.min_nr_regions;
612 if (sz < DAMON_MIN_REGION)
613 sz = DAMON_MIN_REGION;
618 static int kdamond_fn(void *data);
621 * __damon_start() - Starts monitoring with given context.
622 * @ctx: monitoring context
624 * This function should be called while damon_lock is hold.
626 * Return: 0 on success, negative error code otherwise.
628 static int __damon_start(struct damon_ctx *ctx)
632 mutex_lock(&ctx->kdamond_lock);
635 ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
637 if (IS_ERR(ctx->kdamond)) {
638 err = PTR_ERR(ctx->kdamond);
642 mutex_unlock(&ctx->kdamond_lock);
648 * damon_start() - Starts the monitorings for a given group of contexts.
649 * @ctxs: an array of the pointers for contexts to start monitoring
650 * @nr_ctxs: size of @ctxs
651 * @exclusive: exclusiveness of this contexts group
653 * This function starts a group of monitoring threads for a group of monitoring
654 * contexts. One thread per each context is created and run in parallel. The
655 * caller should handle synchronization between the threads by itself. If
656 * @exclusive is true and a group of threads that created by other
657 * 'damon_start()' call is currently running, this function does nothing but
660 * Return: 0 on success, negative error code otherwise.
662 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
667 mutex_lock(&damon_lock);
668 if ((exclusive && nr_running_ctxs) ||
669 (!exclusive && running_exclusive_ctxs)) {
670 mutex_unlock(&damon_lock);
674 for (i = 0; i < nr_ctxs; i++) {
675 err = __damon_start(ctxs[i]);
680 if (exclusive && nr_running_ctxs)
681 running_exclusive_ctxs = true;
682 mutex_unlock(&damon_lock);
688 * __damon_stop() - Stops monitoring of a given context.
689 * @ctx: monitoring context
691 * Return: 0 on success, negative error code otherwise.
693 static int __damon_stop(struct damon_ctx *ctx)
695 struct task_struct *tsk;
697 mutex_lock(&ctx->kdamond_lock);
700 get_task_struct(tsk);
701 mutex_unlock(&ctx->kdamond_lock);
703 put_task_struct(tsk);
706 mutex_unlock(&ctx->kdamond_lock);
712 * damon_stop() - Stops the monitorings for a given group of contexts.
713 * @ctxs: an array of the pointers for contexts to stop monitoring
714 * @nr_ctxs: size of @ctxs
716 * Return: 0 on success, negative error code otherwise.
718 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
722 for (i = 0; i < nr_ctxs; i++) {
723 /* nr_running_ctxs is decremented in kdamond_fn */
724 err = __damon_stop(ctxs[i]);
732 * damon_check_reset_time_interval() - Check if a time interval is elapsed.
733 * @baseline: the time to check whether the interval has elapsed since
734 * @interval: the time interval (microseconds)
736 * See whether the given time interval has passed since the given baseline
737 * time. If so, it also updates the baseline to current time for next check.
739 * Return: true if the time interval has passed, or false otherwise.
741 static bool damon_check_reset_time_interval(struct timespec64 *baseline,
742 unsigned long interval)
744 struct timespec64 now;
746 ktime_get_coarse_ts64(&now);
747 if ((timespec64_to_ns(&now) - timespec64_to_ns(baseline)) <
755 * Check whether it is time to flush the aggregated information
757 static bool kdamond_aggregate_interval_passed(struct damon_ctx *ctx)
759 return damon_check_reset_time_interval(&ctx->last_aggregation,
760 ctx->attrs.aggr_interval);
764 * Reset the aggregated monitoring results ('nr_accesses' of each region).
766 static void kdamond_reset_aggregated(struct damon_ctx *c)
768 struct damon_target *t;
769 unsigned int ti = 0; /* target's index */
771 damon_for_each_target(t, c) {
772 struct damon_region *r;
774 damon_for_each_region(r, t) {
775 trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
776 r->last_nr_accesses = r->nr_accesses;
783 static void damon_split_region_at(struct damon_target *t,
784 struct damon_region *r, unsigned long sz_r);
786 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
790 sz = damon_sz_region(r);
791 return s->pattern.min_sz_region <= sz &&
792 sz <= s->pattern.max_sz_region &&
793 s->pattern.min_nr_accesses <= r->nr_accesses &&
794 r->nr_accesses <= s->pattern.max_nr_accesses &&
795 s->pattern.min_age_region <= r->age &&
796 r->age <= s->pattern.max_age_region;
799 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
800 struct damon_region *r, struct damos *s)
802 bool ret = __damos_valid_target(r, s);
804 if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
807 return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
811 * damos_skip_charged_region() - Check if the given region or starting part of
812 * it is already charged for the DAMOS quota.
813 * @t: The target of the region.
814 * @rp: The pointer to the region.
815 * @s: The scheme to be applied.
817 * If a quota of a scheme has exceeded in a quota charge window, the scheme's
818 * action would applied to only a part of the target access pattern fulfilling
819 * regions. To avoid applying the scheme action to only already applied
820 * regions, DAMON skips applying the scheme action to the regions that charged
821 * in the previous charge window.
823 * This function checks if a given region should be skipped or not for the
824 * reason. If only the starting part of the region has previously charged,
825 * this function splits the region into two so that the second one covers the
826 * area that not charged in the previous charge widnow and saves the second
827 * region in *rp and returns false, so that the caller can apply DAMON action
830 * Return: true if the region should be entirely skipped, false otherwise.
832 static bool damos_skip_charged_region(struct damon_target *t,
833 struct damon_region **rp, struct damos *s)
835 struct damon_region *r = *rp;
836 struct damos_quota *quota = &s->quota;
837 unsigned long sz_to_skip;
839 /* Skip previously charged regions */
840 if (quota->charge_target_from) {
841 if (t != quota->charge_target_from)
843 if (r == damon_last_region(t)) {
844 quota->charge_target_from = NULL;
845 quota->charge_addr_from = 0;
848 if (quota->charge_addr_from &&
849 r->ar.end <= quota->charge_addr_from)
852 if (quota->charge_addr_from && r->ar.start <
853 quota->charge_addr_from) {
854 sz_to_skip = ALIGN_DOWN(quota->charge_addr_from -
855 r->ar.start, DAMON_MIN_REGION);
857 if (damon_sz_region(r) <= DAMON_MIN_REGION)
859 sz_to_skip = DAMON_MIN_REGION;
861 damon_split_region_at(t, r, sz_to_skip);
862 r = damon_next_region(r);
865 quota->charge_target_from = NULL;
866 quota->charge_addr_from = 0;
871 static void damos_update_stat(struct damos *s,
872 unsigned long sz_tried, unsigned long sz_applied)
875 s->stat.sz_tried += sz_tried;
877 s->stat.nr_applied++;
878 s->stat.sz_applied += sz_applied;
881 static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
882 struct damon_region *r, struct damos *s)
884 struct damos_quota *quota = &s->quota;
885 unsigned long sz = damon_sz_region(r);
886 struct timespec64 begin, end;
887 unsigned long sz_applied = 0;
890 if (c->ops.apply_scheme) {
891 if (quota->esz && quota->charged_sz + sz > quota->esz) {
892 sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
896 damon_split_region_at(t, r, sz);
898 ktime_get_coarse_ts64(&begin);
899 if (c->callback.before_damos_apply)
900 err = c->callback.before_damos_apply(c, t, r, s);
902 sz_applied = c->ops.apply_scheme(c, t, r, s);
903 ktime_get_coarse_ts64(&end);
904 quota->total_charged_ns += timespec64_to_ns(&end) -
905 timespec64_to_ns(&begin);
906 quota->charged_sz += sz;
907 if (quota->esz && quota->charged_sz >= quota->esz) {
908 quota->charge_target_from = t;
909 quota->charge_addr_from = r->ar.end + 1;
912 if (s->action != DAMOS_STAT)
916 damos_update_stat(s, sz, sz_applied);
919 static void damon_do_apply_schemes(struct damon_ctx *c,
920 struct damon_target *t,
921 struct damon_region *r)
925 damon_for_each_scheme(s, c) {
926 struct damos_quota *quota = &s->quota;
928 if (!s->wmarks.activated)
931 /* Check the quota */
932 if (quota->esz && quota->charged_sz >= quota->esz)
935 if (damos_skip_charged_region(t, &r, s))
938 if (!damos_valid_target(c, t, r, s))
941 damos_apply_scheme(c, t, r, s);
945 /* Shouldn't be called if quota->ms and quota->sz are zero */
946 static void damos_set_effective_quota(struct damos_quota *quota)
948 unsigned long throughput;
952 quota->esz = quota->sz;
956 if (quota->total_charged_ns)
957 throughput = quota->total_charged_sz * 1000000 /
958 quota->total_charged_ns;
960 throughput = PAGE_SIZE * 1024;
961 esz = throughput * quota->ms;
963 if (quota->sz && quota->sz < esz)
968 static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
970 struct damos_quota *quota = &s->quota;
971 struct damon_target *t;
972 struct damon_region *r;
973 unsigned long cumulated_sz;
974 unsigned int score, max_score = 0;
976 if (!quota->ms && !quota->sz)
979 /* New charge window starts */
980 if (time_after_eq(jiffies, quota->charged_from +
981 msecs_to_jiffies(quota->reset_interval))) {
982 if (quota->esz && quota->charged_sz >= quota->esz)
983 s->stat.qt_exceeds++;
984 quota->total_charged_sz += quota->charged_sz;
985 quota->charged_from = jiffies;
986 quota->charged_sz = 0;
987 damos_set_effective_quota(quota);
990 if (!c->ops.get_scheme_score)
993 /* Fill up the score histogram */
994 memset(quota->histogram, 0, sizeof(quota->histogram));
995 damon_for_each_target(t, c) {
996 damon_for_each_region(r, t) {
997 if (!__damos_valid_target(r, s))
999 score = c->ops.get_scheme_score(c, t, r, s);
1000 quota->histogram[score] += damon_sz_region(r);
1001 if (score > max_score)
1006 /* Set the min score limit */
1007 for (cumulated_sz = 0, score = max_score; ; score--) {
1008 cumulated_sz += quota->histogram[score];
1009 if (cumulated_sz >= quota->esz || !score)
1012 quota->min_score = score;
1015 static void kdamond_apply_schemes(struct damon_ctx *c)
1017 struct damon_target *t;
1018 struct damon_region *r, *next_r;
1021 damon_for_each_scheme(s, c) {
1022 if (!s->wmarks.activated)
1025 damos_adjust_quota(c, s);
1028 damon_for_each_target(t, c) {
1029 damon_for_each_region_safe(r, next_r, t)
1030 damon_do_apply_schemes(c, t, r);
1035 * Merge two adjacent regions into one region
1037 static void damon_merge_two_regions(struct damon_target *t,
1038 struct damon_region *l, struct damon_region *r)
1040 unsigned long sz_l = damon_sz_region(l), sz_r = damon_sz_region(r);
1042 l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
1044 l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
1045 l->ar.end = r->ar.end;
1046 damon_destroy_region(r, t);
1050 * Merge adjacent regions having similar access frequencies
1052 * t target affected by this merge operation
1053 * thres '->nr_accesses' diff threshold for the merge
1054 * sz_limit size upper limit of each region
1056 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
1057 unsigned long sz_limit)
1059 struct damon_region *r, *prev = NULL, *next;
1061 damon_for_each_region_safe(r, next, t) {
1062 if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
1067 if (prev && prev->ar.end == r->ar.start &&
1068 abs(prev->nr_accesses - r->nr_accesses) <= thres &&
1069 damon_sz_region(prev) + damon_sz_region(r) <= sz_limit)
1070 damon_merge_two_regions(t, prev, r);
1077 * Merge adjacent regions having similar access frequencies
1079 * threshold '->nr_accesses' diff threshold for the merge
1080 * sz_limit size upper limit of each region
1082 * This function merges monitoring target regions which are adjacent and their
1083 * access frequencies are similar. This is for minimizing the monitoring
1084 * overhead under the dynamically changeable access pattern. If a merge was
1085 * unnecessarily made, later 'kdamond_split_regions()' will revert it.
1087 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
1088 unsigned long sz_limit)
1090 struct damon_target *t;
1092 damon_for_each_target(t, c)
1093 damon_merge_regions_of(t, threshold, sz_limit);
1097 * Split a region in two
1099 * r the region to be split
1100 * sz_r size of the first sub-region that will be made
1102 static void damon_split_region_at(struct damon_target *t,
1103 struct damon_region *r, unsigned long sz_r)
1105 struct damon_region *new;
1107 new = damon_new_region(r->ar.start + sz_r, r->ar.end);
1111 r->ar.end = new->ar.start;
1114 new->last_nr_accesses = r->last_nr_accesses;
1116 damon_insert_region(new, r, damon_next_region(r), t);
1119 /* Split every region in the given target into 'nr_subs' regions */
1120 static void damon_split_regions_of(struct damon_target *t, int nr_subs)
1122 struct damon_region *r, *next;
1123 unsigned long sz_region, sz_sub = 0;
1126 damon_for_each_region_safe(r, next, t) {
1127 sz_region = damon_sz_region(r);
1129 for (i = 0; i < nr_subs - 1 &&
1130 sz_region > 2 * DAMON_MIN_REGION; i++) {
1132 * Randomly select size of left sub-region to be at
1133 * least 10 percent and at most 90% of original region
1135 sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
1136 sz_region / 10, DAMON_MIN_REGION);
1137 /* Do not allow blank region */
1138 if (sz_sub == 0 || sz_sub >= sz_region)
1141 damon_split_region_at(t, r, sz_sub);
1148 * Split every target region into randomly-sized small regions
1150 * This function splits every target region into random-sized small regions if
1151 * current total number of the regions is equal or smaller than half of the
1152 * user-specified maximum number of regions. This is for maximizing the
1153 * monitoring accuracy under the dynamically changeable access patterns. If a
1154 * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
1157 static void kdamond_split_regions(struct damon_ctx *ctx)
1159 struct damon_target *t;
1160 unsigned int nr_regions = 0;
1161 static unsigned int last_nr_regions;
1162 int nr_subregions = 2;
1164 damon_for_each_target(t, ctx)
1165 nr_regions += damon_nr_regions(t);
1167 if (nr_regions > ctx->attrs.max_nr_regions / 2)
1170 /* Maybe the middle of the region has different access frequency */
1171 if (last_nr_regions == nr_regions &&
1172 nr_regions < ctx->attrs.max_nr_regions / 3)
1175 damon_for_each_target(t, ctx)
1176 damon_split_regions_of(t, nr_subregions);
1178 last_nr_regions = nr_regions;
1182 * Check whether it is time to check and apply the operations-related data
1185 * Returns true if it is.
1187 static bool kdamond_need_update_operations(struct damon_ctx *ctx)
1189 return damon_check_reset_time_interval(&ctx->last_ops_update,
1190 ctx->attrs.ops_update_interval);
1194 * Check whether current monitoring should be stopped
1196 * The monitoring is stopped when either the user requested to stop, or all
1197 * monitoring targets are invalid.
1199 * Returns true if need to stop current monitoring.
1201 static bool kdamond_need_stop(struct damon_ctx *ctx)
1203 struct damon_target *t;
1205 if (kthread_should_stop())
1208 if (!ctx->ops.target_valid)
1211 damon_for_each_target(t, ctx) {
1212 if (ctx->ops.target_valid(t))
1219 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
1224 case DAMOS_WMARK_FREE_MEM_RATE:
1226 return i.freeram * 1000 / i.totalram;
1234 * Returns zero if the scheme is active. Else, returns time to wait for next
1235 * watermark check in micro-seconds.
1237 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1239 unsigned long metric;
1241 if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
1244 metric = damos_wmark_metric_value(scheme->wmarks.metric);
1245 /* higher than high watermark or lower than low watermark */
1246 if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1247 if (scheme->wmarks.activated)
1248 pr_debug("deactivate a scheme (%d) for %s wmark\n",
1250 metric > scheme->wmarks.high ?
1252 scheme->wmarks.activated = false;
1253 return scheme->wmarks.interval;
1256 /* inactive and higher than middle watermark */
1257 if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1258 !scheme->wmarks.activated)
1259 return scheme->wmarks.interval;
1261 if (!scheme->wmarks.activated)
1262 pr_debug("activate a scheme (%d)\n", scheme->action);
1263 scheme->wmarks.activated = true;
1267 static void kdamond_usleep(unsigned long usecs)
1269 /* See Documentation/timers/timers-howto.rst for the thresholds */
1270 if (usecs > 20 * USEC_PER_MSEC)
1271 schedule_timeout_idle(usecs_to_jiffies(usecs));
1273 usleep_idle_range(usecs, usecs + 1);
1276 /* Returns negative error code if it's not activated but should return */
1277 static int kdamond_wait_activation(struct damon_ctx *ctx)
1280 unsigned long wait_time;
1281 unsigned long min_wait_time = 0;
1282 bool init_wait_time = false;
1284 while (!kdamond_need_stop(ctx)) {
1285 damon_for_each_scheme(s, ctx) {
1286 wait_time = damos_wmark_wait_us(s);
1287 if (!init_wait_time || wait_time < min_wait_time) {
1288 init_wait_time = true;
1289 min_wait_time = wait_time;
1295 kdamond_usleep(min_wait_time);
1297 if (ctx->callback.after_wmarks_check &&
1298 ctx->callback.after_wmarks_check(ctx))
1305 * The monitoring daemon that runs as a kernel thread
1307 static int kdamond_fn(void *data)
1309 struct damon_ctx *ctx = data;
1310 struct damon_target *t;
1311 struct damon_region *r, *next;
1312 unsigned int max_nr_accesses = 0;
1313 unsigned long sz_limit = 0;
1315 pr_debug("kdamond (%d) starts\n", current->pid);
1319 if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1322 sz_limit = damon_region_sz_limit(ctx);
1324 while (!kdamond_need_stop(ctx)) {
1325 if (kdamond_wait_activation(ctx))
1328 if (ctx->ops.prepare_access_checks)
1329 ctx->ops.prepare_access_checks(ctx);
1330 if (ctx->callback.after_sampling &&
1331 ctx->callback.after_sampling(ctx))
1334 kdamond_usleep(ctx->attrs.sample_interval);
1336 if (ctx->ops.check_accesses)
1337 max_nr_accesses = ctx->ops.check_accesses(ctx);
1339 if (kdamond_aggregate_interval_passed(ctx)) {
1340 kdamond_merge_regions(ctx,
1341 max_nr_accesses / 10,
1343 if (ctx->callback.after_aggregation &&
1344 ctx->callback.after_aggregation(ctx))
1346 if (!list_empty(&ctx->schemes))
1347 kdamond_apply_schemes(ctx);
1348 kdamond_reset_aggregated(ctx);
1349 kdamond_split_regions(ctx);
1350 if (ctx->ops.reset_aggregated)
1351 ctx->ops.reset_aggregated(ctx);
1354 if (kdamond_need_update_operations(ctx)) {
1355 if (ctx->ops.update)
1356 ctx->ops.update(ctx);
1357 sz_limit = damon_region_sz_limit(ctx);
1361 damon_for_each_target(t, ctx) {
1362 damon_for_each_region_safe(r, next, t)
1363 damon_destroy_region(r, t);
1366 if (ctx->callback.before_terminate)
1367 ctx->callback.before_terminate(ctx);
1368 if (ctx->ops.cleanup)
1369 ctx->ops.cleanup(ctx);
1371 pr_debug("kdamond (%d) finishes\n", current->pid);
1372 mutex_lock(&ctx->kdamond_lock);
1373 ctx->kdamond = NULL;
1374 mutex_unlock(&ctx->kdamond_lock);
1376 mutex_lock(&damon_lock);
1378 if (!nr_running_ctxs && running_exclusive_ctxs)
1379 running_exclusive_ctxs = false;
1380 mutex_unlock(&damon_lock);
1386 * struct damon_system_ram_region - System RAM resource address region of
1388 * @start: Start address of the region (inclusive).
1389 * @end: End address of the region (exclusive).
1391 struct damon_system_ram_region {
1392 unsigned long start;
1396 static int walk_system_ram(struct resource *res, void *arg)
1398 struct damon_system_ram_region *a = arg;
1400 if (a->end - a->start < resource_size(res)) {
1401 a->start = res->start;
1408 * Find biggest 'System RAM' resource and store its start and end address in
1409 * @start and @end, respectively. If no System RAM is found, returns false.
1411 static bool damon_find_biggest_system_ram(unsigned long *start,
1415 struct damon_system_ram_region arg = {};
1417 walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram);
1418 if (arg.end <= arg.start)
1427 * damon_set_region_biggest_system_ram_default() - Set the region of the given
1428 * monitoring target as requested, or biggest 'System RAM'.
1429 * @t: The monitoring target to set the region.
1430 * @start: The pointer to the start address of the region.
1431 * @end: The pointer to the end address of the region.
1433 * This function sets the region of @t as requested by @start and @end. If the
1434 * values of @start and @end are zero, however, this function finds the biggest
1435 * 'System RAM' resource and sets the region to cover the resource. In the
1436 * latter case, this function saves the start and end addresses of the resource
1437 * in @start and @end, respectively.
1439 * Return: 0 on success, negative error code otherwise.
1441 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
1442 unsigned long *start, unsigned long *end)
1444 struct damon_addr_range addr_range;
1449 if (!*start && !*end &&
1450 !damon_find_biggest_system_ram(start, end))
1453 addr_range.start = *start;
1454 addr_range.end = *end;
1455 return damon_set_regions(t, &addr_range, 1);
1458 static int __init damon_init(void)
1460 damon_region_cache = KMEM_CACHE(damon_region, 0);
1461 if (unlikely(!damon_region_cache)) {
1462 pr_err("creating damon_region_cache fails\n");
1469 subsys_initcall(damon_init);
1471 #include "core-test.h"