2237 lines · c
1// SPDX-License-Identifier: GPL-2.02/*3 * Data Access Monitor4 *5 * Author: SeongJae Park <sj@kernel.org>6 */7 8#define pr_fmt(fmt) "damon: " fmt9 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_POINTS19#include <trace/events/damon.h>20 21#ifdef CONFIG_DAMON_KUNIT_TEST22#undef DAMON_MIN_REGION23#define DAMON_MIN_REGION 124#endif25 26static DEFINE_MUTEX(damon_lock);27static int nr_running_ctxs;28static bool running_exclusive_ctxs;29 30static DEFINE_MUTEX(damon_ops_lock);31static struct damon_operations damon_registered_ops[NR_DAMON_OPS];32 33static 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 */36static 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 */51bool 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 &struct68 * damon_operations->id so that others can find and use them later.69 *70 * Return: 0 on success, negative error code otherwise.71 */72int 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;85out: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 make96 * @ctx to use it.97 *98 * Return: 0 on success, negative error code otherwise.99 */100int 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 else111 ctx->ops = damon_registered_ops[id];112 mutex_unlock(&damon_ops_lock);113 return err;114}115 116/*117 * Construct a damon_region struct118 *119 * Returns the pointer to the new struct if success, or NULL otherwise120 */121struct 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(®ion->list);134 135 region->age = 0;136 region->last_nr_accesses = 0;137 138 return region;139}140 141void 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 147static void damon_del_region(struct damon_region *r, struct damon_target *t)148{149 list_del(&r->list);150 t->nr_regions--;151}152 153static void damon_free_region(struct damon_region *r)154{155 kmem_cache_free(damon_region_cache, r);156}157 158void 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 range166 *167 * Returns true if it is.168 */169static 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 */178static 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 a206 * monitoring target to fit in specific ranges.207 *208 * Return: 0 if success, or negative error code otherwise.209 */210int 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 268struct 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 282void damos_add_filter(struct damos *s, struct damos_filter *f)283{284 list_add_tail(&f->list, &s->filters);285}286 287static void damos_del_filter(struct damos_filter *f)288{289 list_del(&f->list);290}291 292static void damos_free_filter(struct damos_filter *f)293{294 kfree(f);295}296 297void damos_destroy_filter(struct damos_filter *f)298{299 damos_del_filter(f);300 damos_free_filter(f);301}302 303struct 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 318void damos_add_quota_goal(struct damos_quota *q, struct damos_quota_goal *g)319{320 list_add_tail(&g->list, &q->goals);321}322 323static void damos_del_quota_goal(struct damos_quota_goal *g)324{325 list_del(&g->list);326}327 328static void damos_free_quota_goal(struct damos_quota_goal *g)329{330 kfree(g);331}332 333void 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 */340static 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 353struct 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 is370 * 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 390static 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 401void 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 407static void damon_del_scheme(struct damos *s)408{409 list_del(&s->list);410}411 412static void damon_free_scheme(struct damos *s)413{414 kfree(s);415}416 417void 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 struct433 *434 * Returns the pointer to the new struct if success, or NULL otherwise435 */436struct 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 452void damon_add_target(struct damon_ctx *ctx, struct damon_target *t)453{454 list_add_tail(&t->list, &ctx->adaptive_targets);455}456 457bool damon_targets_empty(struct damon_ctx *ctx)458{459 return list_empty(&ctx->adaptive_targets);460}461 462static void damon_del_target(struct damon_target *t)463{464 list_del(&t->list);465}466 467void 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 476void damon_destroy_target(struct damon_target *t)477{478 damon_del_target(t);479 damon_free_target(t);480}481 482unsigned int damon_nr_regions(struct damon_target *t)483{484 return t->nr_regions;485}486 487struct 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 517static 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 530void 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 542static 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 */549static 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, like559 * damon_update_monitoring_results() does . Otherwise, divide-by-zero would560 * happen.561 */562static 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 568static 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 577static 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 last588 * aggregation interval that access to the region has found, and region->age is589 * the number of aggregation intervals that its access pattern has maintained.590 * For the reason, the real meaning of the two fields depend on current591 * sampling interval and aggregation interval. This function updates592 * ->nr_accesses and ->age of given damon_ctx's regions for new damon_attrs.593 */594static 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 context616 * @attrs: monitoring attributes617 *618 * This function should be called while the kdamond is not running, or an619 * access check results aggregation is not ongoing (e.g., from620 * &struct damon_callback->after_aggregation or621 * &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 */627int 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 context657 * @schemes: array of the schemes658 * @nr_schemes: number of entries in @schemes659 *660 * This function should not be called while the kdamond of the context is661 * running.662 */663void 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 675static 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 688static 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. Users704 * should use this function for quota goals-level parameters update of running705 * DAMON contexts, instead of manual in-place updates.706 *707 * This function should be called from parameters-update safe context, like708 * DAMON callbacks.709 */710int 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 else720 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 734static 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 750static 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 762static 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 780static 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 788static 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 else798 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 815static 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 827static 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 845static 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 876static 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 are890 * 1. valid (end >= src) and891 * 2. sorted by starting address.892 *893 * If @src has no region, @dst keeps current regions.894 */895static 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 918static 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 935static 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 update976 * the internal status and results accordingly. Users should use this function977 * for context-level parameters update of running context, instead of manual978 * in-place updates.979 *980 * This function should be called from parameters-update safe context, like981 * DAMON callbacks.982 */983int 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, since995 * 1. damon_set_attrs() updates monitoring results of targets and996 * next_apply_sis of schemes, and997 * 2. ops update should be done after pid handling is done (target998 * 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 */1011int 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 */1023static 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 1042static int kdamond_fn(void *data);1043 1044/*1045 * __damon_start() - Starts monitoring with given context.1046 * @ctx: monitoring context1047 *1048 * This function should be called while damon_lock is hold.1049 *1050 * Return: 0 on success, negative error code otherwise.1051 */1052static 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 monitoring1077 * @nr_ctxs: size of @ctxs1078 * @exclusive: exclusiveness of this contexts group1079 *1080 * This function starts a group of monitoring threads for a group of monitoring1081 * contexts. One thread per each context is created and run in parallel. The1082 * caller should handle synchronization between the threads by itself. If1083 * @exclusive is true and a group of threads that created by other1084 * 'damon_start()' call is currently running, this function does nothing but1085 * returns -EBUSY.1086 *1087 * Return: 0 on success, negative error code otherwise.1088 */1089int 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 context1117 *1118 * Return: 0 on success, negative error code otherwise.1119 */1120static 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 monitoring1140 * @nr_ctxs: size of @ctxs1141 *1142 * Return: 0 on success, negative error code otherwise.1143 */1144int 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 */1160static 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 1177static void damon_split_region_at(struct damon_target *t,1178 struct damon_region *r, unsigned long sz_r);1179 1180static 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 1194static 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 of1207 * 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's1213 * action would applied to only a part of the target access pattern fulfilling1214 * regions. To avoid applying the scheme action to only already applied1215 * regions, DAMON skips applying the scheme action to the regions that charged1216 * in the previous charge window.1217 *1218 * This function checks if a given region should be skipped or not for the1219 * 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 the1221 * area that not charged in the previous charge widnow and saves the second1222 * region in *rp and returns false, so that the caller can apply DAMON action1223 * to the second one.1224 *1225 * Return: true if the region should be entirely skipped, false otherwise.1226 */1227static 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 1266static 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 1276static 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 1324static 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 1336static 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 sysfs1346 * implies with 'nr_contexts' file. Nevertheless, only single context1347 * per kdamond is supported for now. So, we can simply use '0' context1348 * 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 1402update_stat:1403 damos_update_stat(s, sz, sz_applied);1404}1405 1406static 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 input1441 * and current score. Assuming the input and the score are positively1442 * proportional, calculate how much compensation should be added to or1443 * subtracted from the last input as a proportion of the last input. Avoid1444 * next input always being zero by setting it non-zero always. In short form1445 * (assuming support of float and signed calculations), the algorithm is as1446 * 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. The1451 * caller should adjust @score for this.1452 *1453 * Returns next input that assumed to achieve the target score.1454 */1455static 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 else1472 score_goal_diff = goal - score;1473 1474 if (last_input < ULONG_MAX / score_goal_diff)1475 compensation = last_input * score_goal_diff / goal;1476 else1477 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_PSI1487 1488static 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 1498static inline u64 damos_get_some_mem_psi_total(void)1499{1500 return 0;1501};1502 1503#endif /* CONFIG_PSI */1504 1505static 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 */1524static 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 empty1541 */1542static 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("a->goals)) {1548 quota->esz = quota->sz;1549 return;1550 }1551 1552 if (!list_empty("a->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 else1566 throughput = PAGE_SIZE * 1024;1567 if (!list_empty("a->goals))1568 esz = min(throughput * quota->ms, esz);1569 else1570 esz = throughput * quota->ms;1571 }1572 1573 if (quota->sz && quota->sz < esz)1574 esz = quota->sz;1575 1576 quota->esz = esz;1577}1578 1579static 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("a->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 1629static 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 region1669 */1670static 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 frequencies1685 *1686 * t target affected by this merge operation1687 * thres '->nr_accesses' diff threshold for the merge1688 * sz_limit size upper limit of each region1689 */1690static 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 else1699 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 else1706 prev = r;1707 }1708}1709 1710/*1711 * Merge adjacent regions having similar access frequencies1712 *1713 * threshold '->nr_accesses' diff threshold for the merge1714 * sz_limit size upper limit of each region1715 *1716 * This function merges monitoring target regions which are adjacent and their1717 * access frequencies are similar. This is for minimizing the monitoring1718 * overhead under the dynamically changeable access pattern. If a merge was1719 * 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 update1723 * max_nr_regions to a number that lower than the current number of regions1724 * while DAMON is running. For such a case, repeat merging until the limit is1725 * met while increasing @threshold up to possible maximum level.1726 */1727static 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 two1749 *1750 * r the region to be split1751 * sz_r size of the first sub-region that will be made1752 */1753static 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 */1773static 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 at1786 * least 10 percent and at most 90% of original region1787 */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 regions1802 *1803 * This function splits every target region into random-sized small regions if1804 * current total number of the regions is equal or smaller than half of the1805 * user-specified maximum number of regions. This is for maximizing the1806 * monitoring accuracy under the dynamically changeable access patterns. If a1807 * split was unnecessarily made, later 'kdamond_merge_regions()' will revert1808 * it.1809 */1810static 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 stopped1836 *1837 * The monitoring is stopped when either the user requested to stop, or all1838 * monitoring targets are invalid.1839 *1840 * Returns true if need to stop current monitoring.1841 */1842static 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 1860static 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 next1876 * watermark check in micro-seconds.1877 */1878static 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 1907static 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 else1913 usleep_idle_range(usecs, usecs + 1);1914}1915 1916/* Returns negative error code if it's not activated but should return */1917static 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 1944static 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 thread1965 */1966static 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 could1993 * be changed from after_wmarks_check() or after_aggregation()1994 * callbacks. Read the values here, and use those for this1995 * iteration. That is, damon_set_attrs() updated new values1996 * 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() if2028 * possible, to reduce overhead2029 */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 }2054done: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 of2082 * [@start, @end).2083 * @start: Start address of the region (inclusive).2084 * @end: End address of the region (exclusive).2085 */2086struct damon_system_ram_region {2087 unsigned long start;2088 unsigned long end;2089};2090 2091static 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 in2104 * @start and @end, respectively. If no System RAM is found, returns false.2105 */2106static 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 given2123 * 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 the2129 * values of @start and @end are zero, however, this function finds the biggest2130 * 'System RAM' resource and sets the region to cover the resource. In the2131 * latter case, this function saves the start and end addresses of the resource2132 * in @start and @end, respectively.2133 *2134 * Return: 0 on success, negative error code otherwise.2135 */2136int 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, but2161 * the cost of keeping past values can be high for arbitrary window size. This2162 * function implements a lightweight pseudo moving sum function that doesn't2163 * keep the past window values.2164 *2165 * It simply assumes there was no noise in the past, and get the no-noise2166 * assumed past value to drop from @nomvsum and @len_window. @nomvsum is a2167 * non-moving sum of the last window. For example, if @len_window is 10 and we2168 * have 25 values, @nomvsum is the sum of the 11th to 20th values of the 252169 * values. Hence, this function simply drops @nomvsum / @len_window from2170 * given @mvsum and add @new_value.2171 *2172 * For example, if @len_window is 10 and @nomvsum is 50, the last 10 values for2173 * the last window could be vary, e.g., 0, 10, 0, 10, 0, 10, 0, 0, 0, 20. For2174 * calculating next moving sum with a new value, we should drop 0 from 50 and2175 * add the new value. However, this function assumes it got value 5 for each2176 * of the last ten times. Based on the assumption, when the next value is2177 * measured, it drops the assumed past value, 5 from the current sum, and add2178 * the new value to get the updated pseduo-moving average.2179 *2180 * This means the value could have errors, but the errors will be disappeared2181 * for every @len_window aligned calls. For example, if @len_window is 10, the2182 * pseudo moving sum with 11th value to 19th value would have an error. But2183 * the sum with 20th value will not have the error.2184 *2185 * Return: Pseudo-moving average after getting the @new_value.2186 */2187static 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 interval2200 * access check result.2201 *2202 * Usually this will be called by &damon_operations->check_accesses callback.2203 */2204void 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 than2211 * 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 2223static 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 2234subsys_initcall(damon_init);2235 2236#include "tests/core-kunit.h"2237