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