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 ctx->passed_sample_intervals = 0;
431 /* These will be set from kdamond_init_intervals_sis() */
432 ctx->next_aggregation_sis = 0;
433 ctx->next_ops_update_sis = 0;
435 mutex_init(&ctx->kdamond_lock);
437 ctx->attrs.min_nr_regions = 10;
438 ctx->attrs.max_nr_regions = 1000;
440 INIT_LIST_HEAD(&ctx->adaptive_targets);
441 INIT_LIST_HEAD(&ctx->schemes);
446 static void damon_destroy_targets(struct damon_ctx *ctx)
448 struct damon_target *t, *next_t;
450 if (ctx->ops.cleanup) {
451 ctx->ops.cleanup(ctx);
455 damon_for_each_target_safe(t, next_t, ctx)
456 damon_destroy_target(t);
459 void damon_destroy_ctx(struct damon_ctx *ctx)
461 struct damos *s, *next_s;
463 damon_destroy_targets(ctx);
465 damon_for_each_scheme_safe(s, next_s, ctx)
466 damon_destroy_scheme(s);
471 static unsigned int damon_age_for_new_attrs(unsigned int age,
472 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
474 return age * old_attrs->aggr_interval / new_attrs->aggr_interval;
477 /* convert access ratio in bp (per 10,000) to nr_accesses */
478 static unsigned int damon_accesses_bp_to_nr_accesses(
479 unsigned int accesses_bp, struct damon_attrs *attrs)
481 return accesses_bp * damon_max_nr_accesses(attrs) / 10000;
484 /* convert nr_accesses to access ratio in bp (per 10,000) */
485 static unsigned int damon_nr_accesses_to_accesses_bp(
486 unsigned int nr_accesses, struct damon_attrs *attrs)
488 return nr_accesses * 10000 / damon_max_nr_accesses(attrs);
491 static unsigned int damon_nr_accesses_for_new_attrs(unsigned int nr_accesses,
492 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
494 return damon_accesses_bp_to_nr_accesses(
495 damon_nr_accesses_to_accesses_bp(
496 nr_accesses, old_attrs),
500 static void damon_update_monitoring_result(struct damon_region *r,
501 struct damon_attrs *old_attrs, struct damon_attrs *new_attrs)
503 r->nr_accesses = damon_nr_accesses_for_new_attrs(r->nr_accesses,
504 old_attrs, new_attrs);
505 r->age = damon_age_for_new_attrs(r->age, old_attrs, new_attrs);
509 * region->nr_accesses is the number of sampling intervals in the last
510 * aggregation interval that access to the region has found, and region->age is
511 * the number of aggregation intervals that its access pattern has maintained.
512 * For the reason, the real meaning of the two fields depend on current
513 * sampling interval and aggregation interval. This function updates
514 * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.
516 static void damon_update_monitoring_results(struct damon_ctx *ctx,
517 struct damon_attrs *new_attrs)
519 struct damon_attrs *old_attrs = &ctx->attrs;
520 struct damon_target *t;
521 struct damon_region *r;
523 /* if any interval is zero, simply forgive conversion */
524 if (!old_attrs->sample_interval || !old_attrs->aggr_interval ||
525 !new_attrs->sample_interval ||
526 !new_attrs->aggr_interval)
529 damon_for_each_target(t, ctx)
530 damon_for_each_region(r, t)
531 damon_update_monitoring_result(
532 r, old_attrs, new_attrs);
536 * damon_set_attrs() - Set attributes for the monitoring.
537 * @ctx: monitoring context
538 * @attrs: monitoring attributes
540 * This function should not be called while the kdamond is running.
541 * Every time interval is in micro-seconds.
543 * Return: 0 on success, negative error code otherwise.
545 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs)
547 unsigned long sample_interval = attrs->sample_interval ?
548 attrs->sample_interval : 1;
550 if (attrs->min_nr_regions < 3)
552 if (attrs->min_nr_regions > attrs->max_nr_regions)
554 if (attrs->sample_interval > attrs->aggr_interval)
557 ctx->next_aggregation_sis = ctx->passed_sample_intervals +
558 attrs->aggr_interval / sample_interval;
559 ctx->next_ops_update_sis = ctx->passed_sample_intervals +
560 attrs->ops_update_interval / sample_interval;
562 damon_update_monitoring_results(ctx, attrs);
568 * damon_set_schemes() - Set data access monitoring based operation schemes.
569 * @ctx: monitoring context
570 * @schemes: array of the schemes
571 * @nr_schemes: number of entries in @schemes
573 * This function should not be called while the kdamond of the context is
576 void damon_set_schemes(struct damon_ctx *ctx, struct damos **schemes,
579 struct damos *s, *next;
582 damon_for_each_scheme_safe(s, next, ctx)
583 damon_destroy_scheme(s);
584 for (i = 0; i < nr_schemes; i++)
585 damon_add_scheme(ctx, schemes[i]);
589 * damon_nr_running_ctxs() - Return number of currently running contexts.
591 int damon_nr_running_ctxs(void)
595 mutex_lock(&damon_lock);
596 nr_ctxs = nr_running_ctxs;
597 mutex_unlock(&damon_lock);
602 /* Returns the size upper limit for each monitoring region */
603 static unsigned long damon_region_sz_limit(struct damon_ctx *ctx)
605 struct damon_target *t;
606 struct damon_region *r;
607 unsigned long sz = 0;
609 damon_for_each_target(t, ctx) {
610 damon_for_each_region(r, t)
611 sz += damon_sz_region(r);
614 if (ctx->attrs.min_nr_regions)
615 sz /= ctx->attrs.min_nr_regions;
616 if (sz < DAMON_MIN_REGION)
617 sz = DAMON_MIN_REGION;
622 static int kdamond_fn(void *data);
625 * __damon_start() - Starts monitoring with given context.
626 * @ctx: monitoring context
628 * This function should be called while damon_lock is hold.
630 * Return: 0 on success, negative error code otherwise.
632 static int __damon_start(struct damon_ctx *ctx)
636 mutex_lock(&ctx->kdamond_lock);
639 ctx->kdamond = kthread_run(kdamond_fn, ctx, "kdamond.%d",
641 if (IS_ERR(ctx->kdamond)) {
642 err = PTR_ERR(ctx->kdamond);
646 mutex_unlock(&ctx->kdamond_lock);
652 * damon_start() - Starts the monitorings for a given group of contexts.
653 * @ctxs: an array of the pointers for contexts to start monitoring
654 * @nr_ctxs: size of @ctxs
655 * @exclusive: exclusiveness of this contexts group
657 * This function starts a group of monitoring threads for a group of monitoring
658 * contexts. One thread per each context is created and run in parallel. The
659 * caller should handle synchronization between the threads by itself. If
660 * @exclusive is true and a group of threads that created by other
661 * 'damon_start()' call is currently running, this function does nothing but
664 * Return: 0 on success, negative error code otherwise.
666 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive)
671 mutex_lock(&damon_lock);
672 if ((exclusive && nr_running_ctxs) ||
673 (!exclusive && running_exclusive_ctxs)) {
674 mutex_unlock(&damon_lock);
678 for (i = 0; i < nr_ctxs; i++) {
679 err = __damon_start(ctxs[i]);
684 if (exclusive && nr_running_ctxs)
685 running_exclusive_ctxs = true;
686 mutex_unlock(&damon_lock);
692 * __damon_stop() - Stops monitoring of a given context.
693 * @ctx: monitoring context
695 * Return: 0 on success, negative error code otherwise.
697 static int __damon_stop(struct damon_ctx *ctx)
699 struct task_struct *tsk;
701 mutex_lock(&ctx->kdamond_lock);
704 get_task_struct(tsk);
705 mutex_unlock(&ctx->kdamond_lock);
707 put_task_struct(tsk);
710 mutex_unlock(&ctx->kdamond_lock);
716 * damon_stop() - Stops the monitorings for a given group of contexts.
717 * @ctxs: an array of the pointers for contexts to stop monitoring
718 * @nr_ctxs: size of @ctxs
720 * Return: 0 on success, negative error code otherwise.
722 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs)
726 for (i = 0; i < nr_ctxs; i++) {
727 /* nr_running_ctxs is decremented in kdamond_fn */
728 err = __damon_stop(ctxs[i]);
736 * Reset the aggregated monitoring results ('nr_accesses' of each region).
738 static void kdamond_reset_aggregated(struct damon_ctx *c)
740 struct damon_target *t;
741 unsigned int ti = 0; /* target's index */
743 damon_for_each_target(t, c) {
744 struct damon_region *r;
746 damon_for_each_region(r, t) {
747 trace_damon_aggregated(t, ti, r, damon_nr_regions(t));
748 r->last_nr_accesses = r->nr_accesses;
755 static void damon_split_region_at(struct damon_target *t,
756 struct damon_region *r, unsigned long sz_r);
758 static bool __damos_valid_target(struct damon_region *r, struct damos *s)
762 sz = damon_sz_region(r);
763 return s->pattern.min_sz_region <= sz &&
764 sz <= s->pattern.max_sz_region &&
765 s->pattern.min_nr_accesses <= r->nr_accesses &&
766 r->nr_accesses <= s->pattern.max_nr_accesses &&
767 s->pattern.min_age_region <= r->age &&
768 r->age <= s->pattern.max_age_region;
771 static bool damos_valid_target(struct damon_ctx *c, struct damon_target *t,
772 struct damon_region *r, struct damos *s)
774 bool ret = __damos_valid_target(r, s);
776 if (!ret || !s->quota.esz || !c->ops.get_scheme_score)
779 return c->ops.get_scheme_score(c, t, r, s) >= s->quota.min_score;
783 * damos_skip_charged_region() - Check if the given region or starting part of
784 * it is already charged for the DAMOS quota.
785 * @t: The target of the region.
786 * @rp: The pointer to the region.
787 * @s: The scheme to be applied.
789 * If a quota of a scheme has exceeded in a quota charge window, the scheme's
790 * action would applied to only a part of the target access pattern fulfilling
791 * regions. To avoid applying the scheme action to only already applied
792 * regions, DAMON skips applying the scheme action to the regions that charged
793 * in the previous charge window.
795 * This function checks if a given region should be skipped or not for the
796 * reason. If only the starting part of the region has previously charged,
797 * this function splits the region into two so that the second one covers the
798 * area that not charged in the previous charge widnow and saves the second
799 * region in *rp and returns false, so that the caller can apply DAMON action
802 * Return: true if the region should be entirely skipped, false otherwise.
804 static bool damos_skip_charged_region(struct damon_target *t,
805 struct damon_region **rp, struct damos *s)
807 struct damon_region *r = *rp;
808 struct damos_quota *quota = &s->quota;
809 unsigned long sz_to_skip;
811 /* Skip previously charged regions */
812 if (quota->charge_target_from) {
813 if (t != quota->charge_target_from)
815 if (r == damon_last_region(t)) {
816 quota->charge_target_from = NULL;
817 quota->charge_addr_from = 0;
820 if (quota->charge_addr_from &&
821 r->ar.end <= quota->charge_addr_from)
824 if (quota->charge_addr_from && r->ar.start <
825 quota->charge_addr_from) {
826 sz_to_skip = ALIGN_DOWN(quota->charge_addr_from -
827 r->ar.start, DAMON_MIN_REGION);
829 if (damon_sz_region(r) <= DAMON_MIN_REGION)
831 sz_to_skip = DAMON_MIN_REGION;
833 damon_split_region_at(t, r, sz_to_skip);
834 r = damon_next_region(r);
837 quota->charge_target_from = NULL;
838 quota->charge_addr_from = 0;
843 static void damos_update_stat(struct damos *s,
844 unsigned long sz_tried, unsigned long sz_applied)
847 s->stat.sz_tried += sz_tried;
849 s->stat.nr_applied++;
850 s->stat.sz_applied += sz_applied;
853 static bool __damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
854 struct damon_region *r, struct damos_filter *filter)
856 bool matched = false;
857 struct damon_target *ti;
859 unsigned long start, end;
861 switch (filter->type) {
862 case DAMOS_FILTER_TYPE_TARGET:
863 damon_for_each_target(ti, ctx) {
868 matched = target_idx == filter->target_idx;
870 case DAMOS_FILTER_TYPE_ADDR:
871 start = ALIGN_DOWN(filter->addr_range.start, DAMON_MIN_REGION);
872 end = ALIGN_DOWN(filter->addr_range.end, DAMON_MIN_REGION);
874 /* inside the range */
875 if (start <= r->ar.start && r->ar.end <= end) {
879 /* outside of the range */
880 if (r->ar.end <= start || end <= r->ar.start) {
884 /* start before the range and overlap */
885 if (r->ar.start < start) {
886 damon_split_region_at(t, r, start - r->ar.start);
890 /* start inside the range */
891 damon_split_region_at(t, r, end - r->ar.start);
898 return matched == filter->matching;
901 static bool damos_filter_out(struct damon_ctx *ctx, struct damon_target *t,
902 struct damon_region *r, struct damos *s)
904 struct damos_filter *filter;
906 damos_for_each_filter(filter, s) {
907 if (__damos_filter_out(ctx, t, r, filter))
913 static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
914 struct damon_region *r, struct damos *s)
916 struct damos_quota *quota = &s->quota;
917 unsigned long sz = damon_sz_region(r);
918 struct timespec64 begin, end;
919 unsigned long sz_applied = 0;
922 if (c->ops.apply_scheme) {
923 if (quota->esz && quota->charged_sz + sz > quota->esz) {
924 sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
928 damon_split_region_at(t, r, sz);
930 if (damos_filter_out(c, t, r, s))
932 ktime_get_coarse_ts64(&begin);
933 if (c->callback.before_damos_apply)
934 err = c->callback.before_damos_apply(c, t, r, s);
936 sz_applied = c->ops.apply_scheme(c, t, r, s);
937 ktime_get_coarse_ts64(&end);
938 quota->total_charged_ns += timespec64_to_ns(&end) -
939 timespec64_to_ns(&begin);
940 quota->charged_sz += sz;
941 if (quota->esz && quota->charged_sz >= quota->esz) {
942 quota->charge_target_from = t;
943 quota->charge_addr_from = r->ar.end + 1;
946 if (s->action != DAMOS_STAT)
950 damos_update_stat(s, sz, sz_applied);
953 static void damon_do_apply_schemes(struct damon_ctx *c,
954 struct damon_target *t,
955 struct damon_region *r)
959 damon_for_each_scheme(s, c) {
960 struct damos_quota *quota = &s->quota;
962 if (!s->wmarks.activated)
965 /* Check the quota */
966 if (quota->esz && quota->charged_sz >= quota->esz)
969 if (damos_skip_charged_region(t, &r, s))
972 if (!damos_valid_target(c, t, r, s))
975 damos_apply_scheme(c, t, r, s);
979 /* Shouldn't be called if quota->ms and quota->sz are zero */
980 static void damos_set_effective_quota(struct damos_quota *quota)
982 unsigned long throughput;
986 quota->esz = quota->sz;
990 if (quota->total_charged_ns)
991 throughput = quota->total_charged_sz * 1000000 /
992 quota->total_charged_ns;
994 throughput = PAGE_SIZE * 1024;
995 esz = throughput * quota->ms;
997 if (quota->sz && quota->sz < esz)
1002 static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
1004 struct damos_quota *quota = &s->quota;
1005 struct damon_target *t;
1006 struct damon_region *r;
1007 unsigned long cumulated_sz;
1008 unsigned int score, max_score = 0;
1010 if (!quota->ms && !quota->sz)
1013 /* New charge window starts */
1014 if (time_after_eq(jiffies, quota->charged_from +
1015 msecs_to_jiffies(quota->reset_interval))) {
1016 if (quota->esz && quota->charged_sz >= quota->esz)
1017 s->stat.qt_exceeds++;
1018 quota->total_charged_sz += quota->charged_sz;
1019 quota->charged_from = jiffies;
1020 quota->charged_sz = 0;
1021 damos_set_effective_quota(quota);
1024 if (!c->ops.get_scheme_score)
1027 /* Fill up the score histogram */
1028 memset(quota->histogram, 0, sizeof(quota->histogram));
1029 damon_for_each_target(t, c) {
1030 damon_for_each_region(r, t) {
1031 if (!__damos_valid_target(r, s))
1033 score = c->ops.get_scheme_score(c, t, r, s);
1034 quota->histogram[score] += damon_sz_region(r);
1035 if (score > max_score)
1040 /* Set the min score limit */
1041 for (cumulated_sz = 0, score = max_score; ; score--) {
1042 cumulated_sz += quota->histogram[score];
1043 if (cumulated_sz >= quota->esz || !score)
1046 quota->min_score = score;
1049 static void kdamond_apply_schemes(struct damon_ctx *c)
1051 struct damon_target *t;
1052 struct damon_region *r, *next_r;
1055 damon_for_each_scheme(s, c) {
1056 if (!s->wmarks.activated)
1059 damos_adjust_quota(c, s);
1062 damon_for_each_target(t, c) {
1063 damon_for_each_region_safe(r, next_r, t)
1064 damon_do_apply_schemes(c, t, r);
1069 * Merge two adjacent regions into one region
1071 static void damon_merge_two_regions(struct damon_target *t,
1072 struct damon_region *l, struct damon_region *r)
1074 unsigned long sz_l = damon_sz_region(l), sz_r = damon_sz_region(r);
1076 l->nr_accesses = (l->nr_accesses * sz_l + r->nr_accesses * sz_r) /
1078 l->age = (l->age * sz_l + r->age * sz_r) / (sz_l + sz_r);
1079 l->ar.end = r->ar.end;
1080 damon_destroy_region(r, t);
1084 * Merge adjacent regions having similar access frequencies
1086 * t target affected by this merge operation
1087 * thres '->nr_accesses' diff threshold for the merge
1088 * sz_limit size upper limit of each region
1090 static void damon_merge_regions_of(struct damon_target *t, unsigned int thres,
1091 unsigned long sz_limit)
1093 struct damon_region *r, *prev = NULL, *next;
1095 damon_for_each_region_safe(r, next, t) {
1096 if (abs(r->nr_accesses - r->last_nr_accesses) > thres)
1101 if (prev && prev->ar.end == r->ar.start &&
1102 abs(prev->nr_accesses - r->nr_accesses) <= thres &&
1103 damon_sz_region(prev) + damon_sz_region(r) <= sz_limit)
1104 damon_merge_two_regions(t, prev, r);
1111 * Merge adjacent regions having similar access frequencies
1113 * threshold '->nr_accesses' diff threshold for the merge
1114 * sz_limit size upper limit of each region
1116 * This function merges monitoring target regions which are adjacent and their
1117 * access frequencies are similar. This is for minimizing the monitoring
1118 * overhead under the dynamically changeable access pattern. If a merge was
1119 * unnecessarily made, later 'kdamond_split_regions()' will revert it.
1121 static void kdamond_merge_regions(struct damon_ctx *c, unsigned int threshold,
1122 unsigned long sz_limit)
1124 struct damon_target *t;
1126 damon_for_each_target(t, c)
1127 damon_merge_regions_of(t, threshold, sz_limit);
1131 * Split a region in two
1133 * r the region to be split
1134 * sz_r size of the first sub-region that will be made
1136 static void damon_split_region_at(struct damon_target *t,
1137 struct damon_region *r, unsigned long sz_r)
1139 struct damon_region *new;
1141 new = damon_new_region(r->ar.start + sz_r, r->ar.end);
1145 r->ar.end = new->ar.start;
1148 new->last_nr_accesses = r->last_nr_accesses;
1150 damon_insert_region(new, r, damon_next_region(r), t);
1153 /* Split every region in the given target into 'nr_subs' regions */
1154 static void damon_split_regions_of(struct damon_target *t, int nr_subs)
1156 struct damon_region *r, *next;
1157 unsigned long sz_region, sz_sub = 0;
1160 damon_for_each_region_safe(r, next, t) {
1161 sz_region = damon_sz_region(r);
1163 for (i = 0; i < nr_subs - 1 &&
1164 sz_region > 2 * DAMON_MIN_REGION; i++) {
1166 * Randomly select size of left sub-region to be at
1167 * least 10 percent and at most 90% of original region
1169 sz_sub = ALIGN_DOWN(damon_rand(1, 10) *
1170 sz_region / 10, DAMON_MIN_REGION);
1171 /* Do not allow blank region */
1172 if (sz_sub == 0 || sz_sub >= sz_region)
1175 damon_split_region_at(t, r, sz_sub);
1182 * Split every target region into randomly-sized small regions
1184 * This function splits every target region into random-sized small regions if
1185 * current total number of the regions is equal or smaller than half of the
1186 * user-specified maximum number of regions. This is for maximizing the
1187 * monitoring accuracy under the dynamically changeable access patterns. If a
1188 * split was unnecessarily made, later 'kdamond_merge_regions()' will revert
1191 static void kdamond_split_regions(struct damon_ctx *ctx)
1193 struct damon_target *t;
1194 unsigned int nr_regions = 0;
1195 static unsigned int last_nr_regions;
1196 int nr_subregions = 2;
1198 damon_for_each_target(t, ctx)
1199 nr_regions += damon_nr_regions(t);
1201 if (nr_regions > ctx->attrs.max_nr_regions / 2)
1204 /* Maybe the middle of the region has different access frequency */
1205 if (last_nr_regions == nr_regions &&
1206 nr_regions < ctx->attrs.max_nr_regions / 3)
1209 damon_for_each_target(t, ctx)
1210 damon_split_regions_of(t, nr_subregions);
1212 last_nr_regions = nr_regions;
1216 * Check whether current monitoring should be stopped
1218 * The monitoring is stopped when either the user requested to stop, or all
1219 * monitoring targets are invalid.
1221 * Returns true if need to stop current monitoring.
1223 static bool kdamond_need_stop(struct damon_ctx *ctx)
1225 struct damon_target *t;
1227 if (kthread_should_stop())
1230 if (!ctx->ops.target_valid)
1233 damon_for_each_target(t, ctx) {
1234 if (ctx->ops.target_valid(t))
1241 static unsigned long damos_wmark_metric_value(enum damos_wmark_metric metric)
1246 case DAMOS_WMARK_FREE_MEM_RATE:
1248 return i.freeram * 1000 / i.totalram;
1256 * Returns zero if the scheme is active. Else, returns time to wait for next
1257 * watermark check in micro-seconds.
1259 static unsigned long damos_wmark_wait_us(struct damos *scheme)
1261 unsigned long metric;
1263 if (scheme->wmarks.metric == DAMOS_WMARK_NONE)
1266 metric = damos_wmark_metric_value(scheme->wmarks.metric);
1267 /* higher than high watermark or lower than low watermark */
1268 if (metric > scheme->wmarks.high || scheme->wmarks.low > metric) {
1269 if (scheme->wmarks.activated)
1270 pr_debug("deactivate a scheme (%d) for %s wmark\n",
1272 metric > scheme->wmarks.high ?
1274 scheme->wmarks.activated = false;
1275 return scheme->wmarks.interval;
1278 /* inactive and higher than middle watermark */
1279 if ((scheme->wmarks.high >= metric && metric >= scheme->wmarks.mid) &&
1280 !scheme->wmarks.activated)
1281 return scheme->wmarks.interval;
1283 if (!scheme->wmarks.activated)
1284 pr_debug("activate a scheme (%d)\n", scheme->action);
1285 scheme->wmarks.activated = true;
1289 static void kdamond_usleep(unsigned long usecs)
1291 /* See Documentation/timers/timers-howto.rst for the thresholds */
1292 if (usecs > 20 * USEC_PER_MSEC)
1293 schedule_timeout_idle(usecs_to_jiffies(usecs));
1295 usleep_idle_range(usecs, usecs + 1);
1298 /* Returns negative error code if it's not activated but should return */
1299 static int kdamond_wait_activation(struct damon_ctx *ctx)
1302 unsigned long wait_time;
1303 unsigned long min_wait_time = 0;
1304 bool init_wait_time = false;
1306 while (!kdamond_need_stop(ctx)) {
1307 damon_for_each_scheme(s, ctx) {
1308 wait_time = damos_wmark_wait_us(s);
1309 if (!init_wait_time || wait_time < min_wait_time) {
1310 init_wait_time = true;
1311 min_wait_time = wait_time;
1317 kdamond_usleep(min_wait_time);
1319 if (ctx->callback.after_wmarks_check &&
1320 ctx->callback.after_wmarks_check(ctx))
1326 static void kdamond_init_intervals_sis(struct damon_ctx *ctx)
1328 unsigned long sample_interval = ctx->attrs.sample_interval ?
1329 ctx->attrs.sample_interval : 1;
1331 ctx->passed_sample_intervals = 0;
1332 ctx->next_aggregation_sis = ctx->attrs.aggr_interval / sample_interval;
1333 ctx->next_ops_update_sis = ctx->attrs.ops_update_interval /
1338 * The monitoring daemon that runs as a kernel thread
1340 static int kdamond_fn(void *data)
1342 struct damon_ctx *ctx = data;
1343 struct damon_target *t;
1344 struct damon_region *r, *next;
1345 unsigned int max_nr_accesses = 0;
1346 unsigned long sz_limit = 0;
1348 pr_debug("kdamond (%d) starts\n", current->pid);
1350 kdamond_init_intervals_sis(ctx);
1354 if (ctx->callback.before_start && ctx->callback.before_start(ctx))
1357 sz_limit = damon_region_sz_limit(ctx);
1359 while (!kdamond_need_stop(ctx)) {
1361 * ctx->attrs and ctx->next_{aggregation,ops_update}_sis could
1362 * be changed from after_wmarks_check() or after_aggregation()
1363 * callbacks. Read the values here, and use those for this
1364 * iteration. That is, damon_set_attrs() updated new values
1365 * are respected from next iteration.
1367 unsigned long next_aggregation_sis = ctx->next_aggregation_sis;
1368 unsigned long next_ops_update_sis = ctx->next_ops_update_sis;
1369 unsigned long sample_interval = ctx->attrs.sample_interval;
1371 if (kdamond_wait_activation(ctx))
1374 if (ctx->ops.prepare_access_checks)
1375 ctx->ops.prepare_access_checks(ctx);
1376 if (ctx->callback.after_sampling &&
1377 ctx->callback.after_sampling(ctx))
1380 kdamond_usleep(sample_interval);
1381 ctx->passed_sample_intervals++;
1383 if (ctx->ops.check_accesses)
1384 max_nr_accesses = ctx->ops.check_accesses(ctx);
1386 sample_interval = ctx->attrs.sample_interval ?
1387 ctx->attrs.sample_interval : 1;
1388 if (ctx->passed_sample_intervals == next_aggregation_sis) {
1389 ctx->next_aggregation_sis = next_aggregation_sis +
1390 ctx->attrs.aggr_interval / sample_interval;
1391 kdamond_merge_regions(ctx,
1392 max_nr_accesses / 10,
1394 if (ctx->callback.after_aggregation &&
1395 ctx->callback.after_aggregation(ctx))
1397 if (!list_empty(&ctx->schemes))
1398 kdamond_apply_schemes(ctx);
1399 kdamond_reset_aggregated(ctx);
1400 kdamond_split_regions(ctx);
1401 if (ctx->ops.reset_aggregated)
1402 ctx->ops.reset_aggregated(ctx);
1405 if (ctx->passed_sample_intervals == next_ops_update_sis) {
1406 ctx->next_ops_update_sis = next_ops_update_sis +
1407 ctx->attrs.ops_update_interval /
1409 if (ctx->ops.update)
1410 ctx->ops.update(ctx);
1411 sz_limit = damon_region_sz_limit(ctx);
1415 damon_for_each_target(t, ctx) {
1416 damon_for_each_region_safe(r, next, t)
1417 damon_destroy_region(r, t);
1420 if (ctx->callback.before_terminate)
1421 ctx->callback.before_terminate(ctx);
1422 if (ctx->ops.cleanup)
1423 ctx->ops.cleanup(ctx);
1425 pr_debug("kdamond (%d) finishes\n", current->pid);
1426 mutex_lock(&ctx->kdamond_lock);
1427 ctx->kdamond = NULL;
1428 mutex_unlock(&ctx->kdamond_lock);
1430 mutex_lock(&damon_lock);
1432 if (!nr_running_ctxs && running_exclusive_ctxs)
1433 running_exclusive_ctxs = false;
1434 mutex_unlock(&damon_lock);
1440 * struct damon_system_ram_region - System RAM resource address region of
1442 * @start: Start address of the region (inclusive).
1443 * @end: End address of the region (exclusive).
1445 struct damon_system_ram_region {
1446 unsigned long start;
1450 static int walk_system_ram(struct resource *res, void *arg)
1452 struct damon_system_ram_region *a = arg;
1454 if (a->end - a->start < resource_size(res)) {
1455 a->start = res->start;
1462 * Find biggest 'System RAM' resource and store its start and end address in
1463 * @start and @end, respectively. If no System RAM is found, returns false.
1465 static bool damon_find_biggest_system_ram(unsigned long *start,
1469 struct damon_system_ram_region arg = {};
1471 walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram);
1472 if (arg.end <= arg.start)
1481 * damon_set_region_biggest_system_ram_default() - Set the region of the given
1482 * monitoring target as requested, or biggest 'System RAM'.
1483 * @t: The monitoring target to set the region.
1484 * @start: The pointer to the start address of the region.
1485 * @end: The pointer to the end address of the region.
1487 * This function sets the region of @t as requested by @start and @end. If the
1488 * values of @start and @end are zero, however, this function finds the biggest
1489 * 'System RAM' resource and sets the region to cover the resource. In the
1490 * latter case, this function saves the start and end addresses of the resource
1491 * in @start and @end, respectively.
1493 * Return: 0 on success, negative error code otherwise.
1495 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
1496 unsigned long *start, unsigned long *end)
1498 struct damon_addr_range addr_range;
1503 if (!*start && !*end &&
1504 !damon_find_biggest_system_ram(start, end))
1507 addr_range.start = *start;
1508 addr_range.end = *end;
1509 return damon_set_regions(t, &addr_range, 1);
1512 static int __init damon_init(void)
1514 damon_region_cache = KMEM_CACHE(damon_region, 0);
1515 if (unlikely(!damon_region_cache)) {
1516 pr_err("creating damon_region_cache fails\n");
1523 subsys_initcall(damon_init);
1525 #include "core-test.h"