1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * DAMON api
4  *
5  * Author: SeongJae Park <sjpark@amazon.de>
6  */
7 
8 #ifndef _DAMON_H_
9 #define _DAMON_H_
10 
11 #include <linux/mutex.h>
12 #include <linux/time64.h>
13 #include <linux/types.h>
14 #include <linux/random.h>
15 
16 /* Minimal region size.  Every damon_region is aligned by this. */
17 #define DAMON_MIN_REGION	PAGE_SIZE
18 /* Max priority score for DAMON-based operation schemes */
19 #define DAMOS_MAX_SCORE		(99)
20 
21 /* Get a random number in [l, r) */
damon_rand(unsigned long l,unsigned long r)22 static inline unsigned long damon_rand(unsigned long l, unsigned long r)
23 {
24 	return l + prandom_u32_max(r - l);
25 }
26 
27 /**
28  * struct damon_addr_range - Represents an address region of [@start, @end).
29  * @start:	Start address of the region (inclusive).
30  * @end:	End address of the region (exclusive).
31  */
32 struct damon_addr_range {
33 	unsigned long start;
34 	unsigned long end;
35 };
36 
37 /**
38  * struct damon_region - Represents a monitoring target region.
39  * @ar:			The address range of the region.
40  * @sampling_addr:	Address of the sample for the next access check.
41  * @nr_accesses:	Access frequency of this region.
42  * @list:		List head for siblings.
43  * @age:		Age of this region.
44  *
45  * @age is initially zero, increased for each aggregation interval, and reset
46  * to zero again if the access frequency is significantly changed.  If two
47  * regions are merged into a new region, both @nr_accesses and @age of the new
48  * region are set as region size-weighted average of those of the two regions.
49  */
50 struct damon_region {
51 	struct damon_addr_range ar;
52 	unsigned long sampling_addr;
53 	unsigned int nr_accesses;
54 	struct list_head list;
55 
56 	unsigned int age;
57 /* private: Internal value for age calculation. */
58 	unsigned int last_nr_accesses;
59 };
60 
61 /**
62  * struct damon_target - Represents a monitoring target.
63  * @pid:		The PID of the virtual address space to monitor.
64  * @nr_regions:		Number of monitoring target regions of this target.
65  * @regions_list:	Head of the monitoring target regions of this target.
66  * @list:		List head for siblings.
67  *
68  * Each monitoring context could have multiple targets.  For example, a context
69  * for virtual memory address spaces could have multiple target processes.  The
70  * @pid should be set for appropriate &struct damon_operations including the
71  * virtual address spaces monitoring operations.
72  */
73 struct damon_target {
74 	struct pid *pid;
75 	unsigned int nr_regions;
76 	struct list_head regions_list;
77 	struct list_head list;
78 };
79 
80 /**
81  * enum damos_action - Represents an action of a Data Access Monitoring-based
82  * Operation Scheme.
83  *
84  * @DAMOS_WILLNEED:	Call ``madvise()`` for the region with MADV_WILLNEED.
85  * @DAMOS_COLD:		Call ``madvise()`` for the region with MADV_COLD.
86  * @DAMOS_PAGEOUT:	Call ``madvise()`` for the region with MADV_PAGEOUT.
87  * @DAMOS_HUGEPAGE:	Call ``madvise()`` for the region with MADV_HUGEPAGE.
88  * @DAMOS_NOHUGEPAGE:	Call ``madvise()`` for the region with MADV_NOHUGEPAGE.
89  * @DAMOS_LRU_PRIO:	Prioritize the region on its LRU lists.
90  * @DAMOS_LRU_DEPRIO:	Deprioritize the region on its LRU lists.
91  * @DAMOS_STAT:		Do nothing but count the stat.
92  * @NR_DAMOS_ACTIONS:	Total number of DAMOS actions
93  */
94 enum damos_action {
95 	DAMOS_WILLNEED,
96 	DAMOS_COLD,
97 	DAMOS_PAGEOUT,
98 	DAMOS_HUGEPAGE,
99 	DAMOS_NOHUGEPAGE,
100 	DAMOS_LRU_PRIO,
101 	DAMOS_LRU_DEPRIO,
102 	DAMOS_STAT,		/* Do nothing but only record the stat */
103 	NR_DAMOS_ACTIONS,
104 };
105 
106 /**
107  * struct damos_quota - Controls the aggressiveness of the given scheme.
108  * @ms:			Maximum milliseconds that the scheme can use.
109  * @sz:			Maximum bytes of memory that the action can be applied.
110  * @reset_interval:	Charge reset interval in milliseconds.
111  *
112  * @weight_sz:		Weight of the region's size for prioritization.
113  * @weight_nr_accesses:	Weight of the region's nr_accesses for prioritization.
114  * @weight_age:		Weight of the region's age for prioritization.
115  *
116  * To avoid consuming too much CPU time or IO resources for applying the
117  * &struct damos->action to large memory, DAMON allows users to set time and/or
118  * size quotas.  The quotas can be set by writing non-zero values to &ms and
119  * &sz, respectively.  If the time quota is set, DAMON tries to use only up to
120  * &ms milliseconds within &reset_interval for applying the action.  If the
121  * size quota is set, DAMON tries to apply the action only up to &sz bytes
122  * within &reset_interval.
123  *
124  * Internally, the time quota is transformed to a size quota using estimated
125  * throughput of the scheme's action.  DAMON then compares it against &sz and
126  * uses smaller one as the effective quota.
127  *
128  * For selecting regions within the quota, DAMON prioritizes current scheme's
129  * target memory regions using the &struct damon_operations->get_scheme_score.
130  * You could customize the prioritization logic by setting &weight_sz,
131  * &weight_nr_accesses, and &weight_age, because monitoring operations are
132  * encouraged to respect those.
133  */
134 struct damos_quota {
135 	unsigned long ms;
136 	unsigned long sz;
137 	unsigned long reset_interval;
138 
139 	unsigned int weight_sz;
140 	unsigned int weight_nr_accesses;
141 	unsigned int weight_age;
142 
143 /* private: */
144 	/* For throughput estimation */
145 	unsigned long total_charged_sz;
146 	unsigned long total_charged_ns;
147 
148 	unsigned long esz;	/* Effective size quota in bytes */
149 
150 	/* For charging the quota */
151 	unsigned long charged_sz;
152 	unsigned long charged_from;
153 	struct damon_target *charge_target_from;
154 	unsigned long charge_addr_from;
155 
156 	/* For prioritization */
157 	unsigned long histogram[DAMOS_MAX_SCORE + 1];
158 	unsigned int min_score;
159 };
160 
161 /**
162  * enum damos_wmark_metric - Represents the watermark metric.
163  *
164  * @DAMOS_WMARK_NONE:		Ignore the watermarks of the given scheme.
165  * @DAMOS_WMARK_FREE_MEM_RATE:	Free memory rate of the system in [0,1000].
166  * @NR_DAMOS_WMARK_METRICS:	Total number of DAMOS watermark metrics
167  */
168 enum damos_wmark_metric {
169 	DAMOS_WMARK_NONE,
170 	DAMOS_WMARK_FREE_MEM_RATE,
171 	NR_DAMOS_WMARK_METRICS,
172 };
173 
174 /**
175  * struct damos_watermarks - Controls when a given scheme should be activated.
176  * @metric:	Metric for the watermarks.
177  * @interval:	Watermarks check time interval in microseconds.
178  * @high:	High watermark.
179  * @mid:	Middle watermark.
180  * @low:	Low watermark.
181  *
182  * If &metric is &DAMOS_WMARK_NONE, the scheme is always active.  Being active
183  * means DAMON does monitoring and applying the action of the scheme to
184  * appropriate memory regions.  Else, DAMON checks &metric of the system for at
185  * least every &interval microseconds and works as below.
186  *
187  * If &metric is higher than &high, the scheme is inactivated.  If &metric is
188  * between &mid and &low, the scheme is activated.  If &metric is lower than
189  * &low, the scheme is inactivated.
190  */
191 struct damos_watermarks {
192 	enum damos_wmark_metric metric;
193 	unsigned long interval;
194 	unsigned long high;
195 	unsigned long mid;
196 	unsigned long low;
197 
198 /* private: */
199 	bool activated;
200 };
201 
202 /**
203  * struct damos_stat - Statistics on a given scheme.
204  * @nr_tried:	Total number of regions that the scheme is tried to be applied.
205  * @sz_tried:	Total size of regions that the scheme is tried to be applied.
206  * @nr_applied:	Total number of regions that the scheme is applied.
207  * @sz_applied:	Total size of regions that the scheme is applied.
208  * @qt_exceeds: Total number of times the quota of the scheme has exceeded.
209  */
210 struct damos_stat {
211 	unsigned long nr_tried;
212 	unsigned long sz_tried;
213 	unsigned long nr_applied;
214 	unsigned long sz_applied;
215 	unsigned long qt_exceeds;
216 };
217 
218 /**
219  * struct damos_access_pattern - Target access pattern of the given scheme.
220  * @min_sz_region:	Minimum size of target regions.
221  * @max_sz_region:	Maximum size of target regions.
222  * @min_nr_accesses:	Minimum ``->nr_accesses`` of target regions.
223  * @max_nr_accesses:	Maximum ``->nr_accesses`` of target regions.
224  * @min_age_region:	Minimum age of target regions.
225  * @max_age_region:	Maximum age of target regions.
226  */
227 struct damos_access_pattern {
228 	unsigned long min_sz_region;
229 	unsigned long max_sz_region;
230 	unsigned int min_nr_accesses;
231 	unsigned int max_nr_accesses;
232 	unsigned int min_age_region;
233 	unsigned int max_age_region;
234 };
235 
236 /**
237  * struct damos - Represents a Data Access Monitoring-based Operation Scheme.
238  * @pattern:		Access pattern of target regions.
239  * @action:		&damo_action to be applied to the target regions.
240  * @quota:		Control the aggressiveness of this scheme.
241  * @wmarks:		Watermarks for automated (in)activation of this scheme.
242  * @stat:		Statistics of this scheme.
243  * @list:		List head for siblings.
244  *
245  * For each aggregation interval, DAMON finds regions which fit in the
246  * &pattern and applies &action to those. To avoid consuming too much
247  * CPU time or IO resources for the &action, &quota is used.
248  *
249  * To do the work only when needed, schemes can be activated for specific
250  * system situations using &wmarks.  If all schemes that registered to the
251  * monitoring context are inactive, DAMON stops monitoring either, and just
252  * repeatedly checks the watermarks.
253  *
254  * If all schemes that registered to a &struct damon_ctx are inactive, DAMON
255  * stops monitoring and just repeatedly checks the watermarks.
256  *
257  * After applying the &action to each region, &stat_count and &stat_sz is
258  * updated to reflect the number of regions and total size of regions that the
259  * &action is applied.
260  */
261 struct damos {
262 	struct damos_access_pattern pattern;
263 	enum damos_action action;
264 	struct damos_quota quota;
265 	struct damos_watermarks wmarks;
266 	struct damos_stat stat;
267 	struct list_head list;
268 };
269 
270 /**
271  * enum damon_ops_id - Identifier for each monitoring operations implementation
272  *
273  * @DAMON_OPS_VADDR:	Monitoring operations for virtual address spaces
274  * @DAMON_OPS_FVADDR:	Monitoring operations for only fixed ranges of virtual
275  *			address spaces
276  * @DAMON_OPS_PADDR:	Monitoring operations for the physical address space
277  * @NR_DAMON_OPS:	Number of monitoring operations implementations
278  */
279 enum damon_ops_id {
280 	DAMON_OPS_VADDR,
281 	DAMON_OPS_FVADDR,
282 	DAMON_OPS_PADDR,
283 	NR_DAMON_OPS,
284 };
285 
286 struct damon_ctx;
287 
288 /**
289  * struct damon_operations - Monitoring operations for given use cases.
290  *
291  * @id:				Identifier of this operations set.
292  * @init:			Initialize operations-related data structures.
293  * @update:			Update operations-related data structures.
294  * @prepare_access_checks:	Prepare next access check of target regions.
295  * @check_accesses:		Check the accesses to target regions.
296  * @reset_aggregated:		Reset aggregated accesses monitoring results.
297  * @get_scheme_score:		Get the score of a region for a scheme.
298  * @apply_scheme:		Apply a DAMON-based operation scheme.
299  * @target_valid:		Determine if the target is valid.
300  * @cleanup:			Clean up the context.
301  *
302  * DAMON can be extended for various address spaces and usages.  For this,
303  * users should register the low level operations for their target address
304  * space and usecase via the &damon_ctx.ops.  Then, the monitoring thread
305  * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting
306  * the monitoring, @update after each &damon_ctx.ops_update_interval, and
307  * @check_accesses, @target_valid and @prepare_access_checks after each
308  * &damon_ctx.sample_interval.  Finally, @reset_aggregated is called after each
309  * &damon_ctx.aggr_interval.
310  *
311  * Each &struct damon_operations instance having valid @id can be registered
312  * via damon_register_ops() and selected by damon_select_ops() later.
313  * @init should initialize operations-related data structures.  For example,
314  * this could be used to construct proper monitoring target regions and link
315  * those to @damon_ctx.adaptive_targets.
316  * @update should update the operations-related data structures.  For example,
317  * this could be used to update monitoring target regions for current status.
318  * @prepare_access_checks should manipulate the monitoring regions to be
319  * prepared for the next access check.
320  * @check_accesses should check the accesses to each region that made after the
321  * last preparation and update the number of observed accesses of each region.
322  * It should also return max number of observed accesses that made as a result
323  * of its update.  The value will be used for regions adjustment threshold.
324  * @reset_aggregated should reset the access monitoring results that aggregated
325  * by @check_accesses.
326  * @get_scheme_score should return the priority score of a region for a scheme
327  * as an integer in [0, &DAMOS_MAX_SCORE].
328  * @apply_scheme is called from @kdamond when a region for user provided
329  * DAMON-based operation scheme is found.  It should apply the scheme's action
330  * to the region and return bytes of the region that the action is successfully
331  * applied.
332  * @target_valid should check whether the target is still valid for the
333  * monitoring.
334  * @cleanup is called from @kdamond just before its termination.
335  */
336 struct damon_operations {
337 	enum damon_ops_id id;
338 	void (*init)(struct damon_ctx *context);
339 	void (*update)(struct damon_ctx *context);
340 	void (*prepare_access_checks)(struct damon_ctx *context);
341 	unsigned int (*check_accesses)(struct damon_ctx *context);
342 	void (*reset_aggregated)(struct damon_ctx *context);
343 	int (*get_scheme_score)(struct damon_ctx *context,
344 			struct damon_target *t, struct damon_region *r,
345 			struct damos *scheme);
346 	unsigned long (*apply_scheme)(struct damon_ctx *context,
347 			struct damon_target *t, struct damon_region *r,
348 			struct damos *scheme);
349 	bool (*target_valid)(struct damon_target *t);
350 	void (*cleanup)(struct damon_ctx *context);
351 };
352 
353 /**
354  * struct damon_callback - Monitoring events notification callbacks.
355  *
356  * @before_start:	Called before starting the monitoring.
357  * @after_wmarks_check:	Called after each schemes' watermarks check.
358  * @after_sampling:	Called after each sampling.
359  * @after_aggregation:	Called after each aggregation.
360  * @before_terminate:	Called before terminating the monitoring.
361  * @private:		User private data.
362  *
363  * The monitoring thread (&damon_ctx.kdamond) calls @before_start and
364  * @before_terminate just before starting and finishing the monitoring,
365  * respectively.  Therefore, those are good places for installing and cleaning
366  * @private.
367  *
368  * The monitoring thread calls @after_wmarks_check after each DAMON-based
369  * operation schemes' watermarks check.  If users need to make changes to the
370  * attributes of the monitoring context while it's deactivated due to the
371  * watermarks, this is the good place to do.
372  *
373  * The monitoring thread calls @after_sampling and @after_aggregation for each
374  * of the sampling intervals and aggregation intervals, respectively.
375  * Therefore, users can safely access the monitoring results without additional
376  * protection.  For the reason, users are recommended to use these callback for
377  * the accesses to the results.
378  *
379  * If any callback returns non-zero, monitoring stops.
380  */
381 struct damon_callback {
382 	void *private;
383 
384 	int (*before_start)(struct damon_ctx *context);
385 	int (*after_wmarks_check)(struct damon_ctx *context);
386 	int (*after_sampling)(struct damon_ctx *context);
387 	int (*after_aggregation)(struct damon_ctx *context);
388 	void (*before_terminate)(struct damon_ctx *context);
389 };
390 
391 /**
392  * struct damon_attrs - Monitoring attributes for accuracy/overhead control.
393  *
394  * @sample_interval:		The time between access samplings.
395  * @aggr_interval:		The time between monitor results aggregations.
396  * @ops_update_interval:	The time between monitoring operations updates.
397  * @min_nr_regions:		The minimum number of adaptive monitoring
398  *				regions.
399  * @max_nr_regions:		The maximum number of adaptive monitoring
400  *				regions.
401  *
402  * For each @sample_interval, DAMON checks whether each region is accessed or
403  * not.  It aggregates and keeps the access information (number of accesses to
404  * each region) for @aggr_interval time.  DAMON also checks whether the target
405  * memory regions need update (e.g., by ``mmap()`` calls from the application,
406  * in case of virtual memory monitoring) and applies the changes for each
407  * @ops_update_interval.  All time intervals are in micro-seconds.
408  * Please refer to &struct damon_operations and &struct damon_callback for more
409  * detail.
410  */
411 struct damon_attrs {
412 	unsigned long sample_interval;
413 	unsigned long aggr_interval;
414 	unsigned long ops_update_interval;
415 	unsigned long min_nr_regions;
416 	unsigned long max_nr_regions;
417 };
418 
419 /**
420  * struct damon_ctx - Represents a context for each monitoring.  This is the
421  * main interface that allows users to set the attributes and get the results
422  * of the monitoring.
423  *
424  * @attrs:		Monitoring attributes for accuracy/overhead control.
425  * @kdamond:		Kernel thread who does the monitoring.
426  * @kdamond_lock:	Mutex for the synchronizations with @kdamond.
427  *
428  * For each monitoring context, one kernel thread for the monitoring is
429  * created.  The pointer to the thread is stored in @kdamond.
430  *
431  * Once started, the monitoring thread runs until explicitly required to be
432  * terminated or every monitoring target is invalid.  The validity of the
433  * targets is checked via the &damon_operations.target_valid of @ops.  The
434  * termination can also be explicitly requested by calling damon_stop().
435  * The thread sets @kdamond to NULL when it terminates. Therefore, users can
436  * know whether the monitoring is ongoing or terminated by reading @kdamond.
437  * Reads and writes to @kdamond from outside of the monitoring thread must
438  * be protected by @kdamond_lock.
439  *
440  * Note that the monitoring thread protects only @kdamond via @kdamond_lock.
441  * Accesses to other fields must be protected by themselves.
442  *
443  * @ops:	Set of monitoring operations for given use cases.
444  * @callback:	Set of callbacks for monitoring events notifications.
445  *
446  * @adaptive_targets:	Head of monitoring targets (&damon_target) list.
447  * @schemes:		Head of schemes (&damos) list.
448  */
449 struct damon_ctx {
450 	struct damon_attrs attrs;
451 
452 /* private: internal use only */
453 	struct timespec64 last_aggregation;
454 	struct timespec64 last_ops_update;
455 
456 /* public: */
457 	struct task_struct *kdamond;
458 	struct mutex kdamond_lock;
459 
460 	struct damon_operations ops;
461 	struct damon_callback callback;
462 
463 	struct list_head adaptive_targets;
464 	struct list_head schemes;
465 };
466 
damon_next_region(struct damon_region * r)467 static inline struct damon_region *damon_next_region(struct damon_region *r)
468 {
469 	return container_of(r->list.next, struct damon_region, list);
470 }
471 
damon_prev_region(struct damon_region * r)472 static inline struct damon_region *damon_prev_region(struct damon_region *r)
473 {
474 	return container_of(r->list.prev, struct damon_region, list);
475 }
476 
damon_last_region(struct damon_target * t)477 static inline struct damon_region *damon_last_region(struct damon_target *t)
478 {
479 	return list_last_entry(&t->regions_list, struct damon_region, list);
480 }
481 
damon_first_region(struct damon_target * t)482 static inline struct damon_region *damon_first_region(struct damon_target *t)
483 {
484 	return list_first_entry(&t->regions_list, struct damon_region, list);
485 }
486 
damon_sz_region(struct damon_region * r)487 static inline unsigned long damon_sz_region(struct damon_region *r)
488 {
489 	return r->ar.end - r->ar.start;
490 }
491 
492 
493 #define damon_for_each_region(r, t) \
494 	list_for_each_entry(r, &t->regions_list, list)
495 
496 #define damon_for_each_region_from(r, t) \
497 	list_for_each_entry_from(r, &t->regions_list, list)
498 
499 #define damon_for_each_region_safe(r, next, t) \
500 	list_for_each_entry_safe(r, next, &t->regions_list, list)
501 
502 #define damon_for_each_target(t, ctx) \
503 	list_for_each_entry(t, &(ctx)->adaptive_targets, list)
504 
505 #define damon_for_each_target_safe(t, next, ctx)	\
506 	list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
507 
508 #define damon_for_each_scheme(s, ctx) \
509 	list_for_each_entry(s, &(ctx)->schemes, list)
510 
511 #define damon_for_each_scheme_safe(s, next, ctx) \
512 	list_for_each_entry_safe(s, next, &(ctx)->schemes, list)
513 
514 #ifdef CONFIG_DAMON
515 
516 struct damon_region *damon_new_region(unsigned long start, unsigned long end);
517 
518 /*
519  * Add a region between two other regions
520  */
damon_insert_region(struct damon_region * r,struct damon_region * prev,struct damon_region * next,struct damon_target * t)521 static inline void damon_insert_region(struct damon_region *r,
522 		struct damon_region *prev, struct damon_region *next,
523 		struct damon_target *t)
524 {
525 	__list_add(&r->list, &prev->list, &next->list);
526 	t->nr_regions++;
527 }
528 
529 void damon_add_region(struct damon_region *r, struct damon_target *t);
530 void damon_destroy_region(struct damon_region *r, struct damon_target *t);
531 int damon_set_regions(struct damon_target *t, struct damon_addr_range *ranges,
532 		unsigned int nr_ranges);
533 
534 struct damos *damon_new_scheme(struct damos_access_pattern *pattern,
535 			enum damos_action action, struct damos_quota *quota,
536 			struct damos_watermarks *wmarks);
537 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s);
538 void damon_destroy_scheme(struct damos *s);
539 
540 struct damon_target *damon_new_target(void);
541 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
542 bool damon_targets_empty(struct damon_ctx *ctx);
543 void damon_free_target(struct damon_target *t);
544 void damon_destroy_target(struct damon_target *t);
545 unsigned int damon_nr_regions(struct damon_target *t);
546 
547 struct damon_ctx *damon_new_ctx(void);
548 void damon_destroy_ctx(struct damon_ctx *ctx);
549 int damon_set_attrs(struct damon_ctx *ctx, struct damon_attrs *attrs);
550 void damon_set_schemes(struct damon_ctx *ctx,
551 			struct damos **schemes, ssize_t nr_schemes);
552 int damon_nr_running_ctxs(void);
553 bool damon_is_registered_ops(enum damon_ops_id id);
554 int damon_register_ops(struct damon_operations *ops);
555 int damon_select_ops(struct damon_ctx *ctx, enum damon_ops_id id);
556 
damon_target_has_pid(const struct damon_ctx * ctx)557 static inline bool damon_target_has_pid(const struct damon_ctx *ctx)
558 {
559 	return ctx->ops.id == DAMON_OPS_VADDR || ctx->ops.id == DAMON_OPS_FVADDR;
560 }
561 
562 
563 int damon_start(struct damon_ctx **ctxs, int nr_ctxs, bool exclusive);
564 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
565 
566 int damon_set_region_biggest_system_ram_default(struct damon_target *t,
567 				unsigned long *start, unsigned long *end);
568 
569 #endif	/* CONFIG_DAMON */
570 
571 #endif	/* _DAMON_H */
572