xref: /openbmc/linux/include/linux/damon.h (revision e7f127b2)
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) */
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  * @id:			Unique identifier for this target.
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  * @id of each target should be unique among the targets of the context.  For
71  * example, in the virtual address monitoring context, it could be a pidfd or
72  * an address of an mm_struct.
73  */
74 struct damon_target {
75 	unsigned long id;
76 	unsigned int nr_regions;
77 	struct list_head regions_list;
78 	struct list_head list;
79 };
80 
81 /**
82  * enum damos_action - Represents an action of a Data Access Monitoring-based
83  * Operation Scheme.
84  *
85  * @DAMOS_WILLNEED:	Call ``madvise()`` for the region with MADV_WILLNEED.
86  * @DAMOS_COLD:		Call ``madvise()`` for the region with MADV_COLD.
87  * @DAMOS_PAGEOUT:	Call ``madvise()`` for the region with MADV_PAGEOUT.
88  * @DAMOS_HUGEPAGE:	Call ``madvise()`` for the region with MADV_HUGEPAGE.
89  * @DAMOS_NOHUGEPAGE:	Call ``madvise()`` for the region with MADV_NOHUGEPAGE.
90  * @DAMOS_STAT:		Do nothing but count the stat.
91  */
92 enum damos_action {
93 	DAMOS_WILLNEED,
94 	DAMOS_COLD,
95 	DAMOS_PAGEOUT,
96 	DAMOS_HUGEPAGE,
97 	DAMOS_NOHUGEPAGE,
98 	DAMOS_STAT,		/* Do nothing but only record the stat */
99 };
100 
101 /**
102  * struct damos_quota - Controls the aggressiveness of the given scheme.
103  * @ms:			Maximum milliseconds that the scheme can use.
104  * @sz:			Maximum bytes of memory that the action can be applied.
105  * @reset_interval:	Charge reset interval in milliseconds.
106  *
107  * @weight_sz:		Weight of the region's size for prioritization.
108  * @weight_nr_accesses:	Weight of the region's nr_accesses for prioritization.
109  * @weight_age:		Weight of the region's age for prioritization.
110  *
111  * To avoid consuming too much CPU time or IO resources for applying the
112  * &struct damos->action to large memory, DAMON allows users to set time and/or
113  * size quotas.  The quotas can be set by writing non-zero values to &ms and
114  * &sz, respectively.  If the time quota is set, DAMON tries to use only up to
115  * &ms milliseconds within &reset_interval for applying the action.  If the
116  * size quota is set, DAMON tries to apply the action only up to &sz bytes
117  * within &reset_interval.
118  *
119  * Internally, the time quota is transformed to a size quota using estimated
120  * throughput of the scheme's action.  DAMON then compares it against &sz and
121  * uses smaller one as the effective quota.
122  *
123  * For selecting regions within the quota, DAMON prioritizes current scheme's
124  * target memory regions using the &struct damon_primitive->get_scheme_score.
125  * You could customize the prioritization logic by setting &weight_sz,
126  * &weight_nr_accesses, and &weight_age, because monitoring primitives are
127  * encouraged to respect those.
128  */
129 struct damos_quota {
130 	unsigned long ms;
131 	unsigned long sz;
132 	unsigned long reset_interval;
133 
134 	unsigned int weight_sz;
135 	unsigned int weight_nr_accesses;
136 	unsigned int weight_age;
137 
138 /* private: */
139 	/* For throughput estimation */
140 	unsigned long total_charged_sz;
141 	unsigned long total_charged_ns;
142 
143 	unsigned long esz;	/* Effective size quota in bytes */
144 
145 	/* For charging the quota */
146 	unsigned long charged_sz;
147 	unsigned long charged_from;
148 	struct damon_target *charge_target_from;
149 	unsigned long charge_addr_from;
150 
151 	/* For prioritization */
152 	unsigned long histogram[DAMOS_MAX_SCORE + 1];
153 	unsigned int min_score;
154 };
155 
156 /**
157  * enum damos_wmark_metric - Represents the watermark metric.
158  *
159  * @DAMOS_WMARK_NONE:		Ignore the watermarks of the given scheme.
160  * @DAMOS_WMARK_FREE_MEM_RATE:	Free memory rate of the system in [0,1000].
161  */
162 enum damos_wmark_metric {
163 	DAMOS_WMARK_NONE,
164 	DAMOS_WMARK_FREE_MEM_RATE,
165 };
166 
167 /**
168  * struct damos_watermarks - Controls when a given scheme should be activated.
169  * @metric:	Metric for the watermarks.
170  * @interval:	Watermarks check time interval in microseconds.
171  * @high:	High watermark.
172  * @mid:	Middle watermark.
173  * @low:	Low watermark.
174  *
175  * If &metric is &DAMOS_WMARK_NONE, the scheme is always active.  Being active
176  * means DAMON does monitoring and applying the action of the scheme to
177  * appropriate memory regions.  Else, DAMON checks &metric of the system for at
178  * least every &interval microseconds and works as below.
179  *
180  * If &metric is higher than &high, the scheme is inactivated.  If &metric is
181  * between &mid and &low, the scheme is activated.  If &metric is lower than
182  * &low, the scheme is inactivated.
183  */
184 struct damos_watermarks {
185 	enum damos_wmark_metric metric;
186 	unsigned long interval;
187 	unsigned long high;
188 	unsigned long mid;
189 	unsigned long low;
190 
191 /* private: */
192 	bool activated;
193 };
194 
195 /**
196  * struct damos_stat - Statistics on a given scheme.
197  * @nr_tried:	Total number of regions that the scheme is tried to be applied.
198  * @sz_tried:	Total size of regions that the scheme is tried to be applied.
199  * @nr_applied:	Total number of regions that the scheme is applied.
200  * @sz_applied:	Total size of regions that the scheme is applied.
201  * @qt_exceeds: Total number of times the quota of the scheme has exceeded.
202  */
203 struct damos_stat {
204 	unsigned long nr_tried;
205 	unsigned long sz_tried;
206 	unsigned long nr_applied;
207 	unsigned long sz_applied;
208 	unsigned long qt_exceeds;
209 };
210 
211 /**
212  * struct damos - Represents a Data Access Monitoring-based Operation Scheme.
213  * @min_sz_region:	Minimum size of target regions.
214  * @max_sz_region:	Maximum size of target regions.
215  * @min_nr_accesses:	Minimum ``->nr_accesses`` of target regions.
216  * @max_nr_accesses:	Maximum ``->nr_accesses`` of target regions.
217  * @min_age_region:	Minimum age of target regions.
218  * @max_age_region:	Maximum age of target regions.
219  * @action:		&damo_action to be applied to the target regions.
220  * @quota:		Control the aggressiveness of this scheme.
221  * @wmarks:		Watermarks for automated (in)activation of this scheme.
222  * @stat:		Statistics of this scheme.
223  * @list:		List head for siblings.
224  *
225  * For each aggregation interval, DAMON finds regions which fit in the
226  * condition (&min_sz_region, &max_sz_region, &min_nr_accesses,
227  * &max_nr_accesses, &min_age_region, &max_age_region) and applies &action to
228  * those.  To avoid consuming too much CPU time or IO resources for the
229  * &action, &quota is used.
230  *
231  * To do the work only when needed, schemes can be activated for specific
232  * system situations using &wmarks.  If all schemes that registered to the
233  * monitoring context are inactive, DAMON stops monitoring either, and just
234  * repeatedly checks the watermarks.
235  *
236  * If all schemes that registered to a &struct damon_ctx are inactive, DAMON
237  * stops monitoring and just repeatedly checks the watermarks.
238  *
239  * After applying the &action to each region, &stat_count and &stat_sz is
240  * updated to reflect the number of regions and total size of regions that the
241  * &action is applied.
242  */
243 struct damos {
244 	unsigned long min_sz_region;
245 	unsigned long max_sz_region;
246 	unsigned int min_nr_accesses;
247 	unsigned int max_nr_accesses;
248 	unsigned int min_age_region;
249 	unsigned int max_age_region;
250 	enum damos_action action;
251 	struct damos_quota quota;
252 	struct damos_watermarks wmarks;
253 	struct damos_stat stat;
254 	struct list_head list;
255 };
256 
257 struct damon_ctx;
258 
259 /**
260  * struct damon_primitive - Monitoring primitives for given use cases.
261  *
262  * @init:			Initialize primitive-internal data structures.
263  * @update:			Update primitive-internal data structures.
264  * @prepare_access_checks:	Prepare next access check of target regions.
265  * @check_accesses:		Check the accesses to target regions.
266  * @reset_aggregated:		Reset aggregated accesses monitoring results.
267  * @get_scheme_score:		Get the score of a region for a scheme.
268  * @apply_scheme:		Apply a DAMON-based operation scheme.
269  * @target_valid:		Determine if the target is valid.
270  * @cleanup:			Clean up the context.
271  *
272  * DAMON can be extended for various address spaces and usages.  For this,
273  * users should register the low level primitives for their target address
274  * space and usecase via the &damon_ctx.primitive.  Then, the monitoring thread
275  * (&damon_ctx.kdamond) calls @init and @prepare_access_checks before starting
276  * the monitoring, @update after each &damon_ctx.primitive_update_interval, and
277  * @check_accesses, @target_valid and @prepare_access_checks after each
278  * &damon_ctx.sample_interval.  Finally, @reset_aggregated is called after each
279  * &damon_ctx.aggr_interval.
280  *
281  * @init should initialize primitive-internal data structures.  For example,
282  * this could be used to construct proper monitoring target regions and link
283  * those to @damon_ctx.adaptive_targets.
284  * @update should update the primitive-internal data structures.  For example,
285  * this could be used to update monitoring target regions for current status.
286  * @prepare_access_checks should manipulate the monitoring regions to be
287  * prepared for the next access check.
288  * @check_accesses should check the accesses to each region that made after the
289  * last preparation and update the number of observed accesses of each region.
290  * It should also return max number of observed accesses that made as a result
291  * of its update.  The value will be used for regions adjustment threshold.
292  * @reset_aggregated should reset the access monitoring results that aggregated
293  * by @check_accesses.
294  * @get_scheme_score should return the priority score of a region for a scheme
295  * as an integer in [0, &DAMOS_MAX_SCORE].
296  * @apply_scheme is called from @kdamond when a region for user provided
297  * DAMON-based operation scheme is found.  It should apply the scheme's action
298  * to the region and return bytes of the region that the action is successfully
299  * applied.
300  * @target_valid should check whether the target is still valid for the
301  * monitoring.
302  * @cleanup is called from @kdamond just before its termination.
303  */
304 struct damon_primitive {
305 	void (*init)(struct damon_ctx *context);
306 	void (*update)(struct damon_ctx *context);
307 	void (*prepare_access_checks)(struct damon_ctx *context);
308 	unsigned int (*check_accesses)(struct damon_ctx *context);
309 	void (*reset_aggregated)(struct damon_ctx *context);
310 	int (*get_scheme_score)(struct damon_ctx *context,
311 			struct damon_target *t, struct damon_region *r,
312 			struct damos *scheme);
313 	unsigned long (*apply_scheme)(struct damon_ctx *context,
314 			struct damon_target *t, struct damon_region *r,
315 			struct damos *scheme);
316 	bool (*target_valid)(void *target);
317 	void (*cleanup)(struct damon_ctx *context);
318 };
319 
320 /**
321  * struct damon_callback - Monitoring events notification callbacks.
322  *
323  * @before_start:	Called before starting the monitoring.
324  * @after_sampling:	Called after each sampling.
325  * @after_aggregation:	Called after each aggregation.
326  * @before_terminate:	Called before terminating the monitoring.
327  * @private:		User private data.
328  *
329  * The monitoring thread (&damon_ctx.kdamond) calls @before_start and
330  * @before_terminate just before starting and finishing the monitoring,
331  * respectively.  Therefore, those are good places for installing and cleaning
332  * @private.
333  *
334  * The monitoring thread calls @after_sampling and @after_aggregation for each
335  * of the sampling intervals and aggregation intervals, respectively.
336  * Therefore, users can safely access the monitoring results without additional
337  * protection.  For the reason, users are recommended to use these callback for
338  * the accesses to the results.
339  *
340  * If any callback returns non-zero, monitoring stops.
341  */
342 struct damon_callback {
343 	void *private;
344 
345 	int (*before_start)(struct damon_ctx *context);
346 	int (*after_sampling)(struct damon_ctx *context);
347 	int (*after_aggregation)(struct damon_ctx *context);
348 	void (*before_terminate)(struct damon_ctx *context);
349 };
350 
351 /**
352  * struct damon_ctx - Represents a context for each monitoring.  This is the
353  * main interface that allows users to set the attributes and get the results
354  * of the monitoring.
355  *
356  * @sample_interval:		The time between access samplings.
357  * @aggr_interval:		The time between monitor results aggregations.
358  * @primitive_update_interval:	The time between monitoring primitive updates.
359  *
360  * For each @sample_interval, DAMON checks whether each region is accessed or
361  * not.  It aggregates and keeps the access information (number of accesses to
362  * each region) for @aggr_interval time.  DAMON also checks whether the target
363  * memory regions need update (e.g., by ``mmap()`` calls from the application,
364  * in case of virtual memory monitoring) and applies the changes for each
365  * @primitive_update_interval.  All time intervals are in micro-seconds.
366  * Please refer to &struct damon_primitive and &struct damon_callback for more
367  * detail.
368  *
369  * @kdamond:		Kernel thread who does the monitoring.
370  * @kdamond_stop:	Notifies whether kdamond should stop.
371  * @kdamond_lock:	Mutex for the synchronizations with @kdamond.
372  *
373  * For each monitoring context, one kernel thread for the monitoring is
374  * created.  The pointer to the thread is stored in @kdamond.
375  *
376  * Once started, the monitoring thread runs until explicitly required to be
377  * terminated or every monitoring target is invalid.  The validity of the
378  * targets is checked via the &damon_primitive.target_valid of @primitive.  The
379  * termination can also be explicitly requested by writing non-zero to
380  * @kdamond_stop.  The thread sets @kdamond to NULL when it terminates.
381  * Therefore, users can know whether the monitoring is ongoing or terminated by
382  * reading @kdamond.  Reads and writes to @kdamond and @kdamond_stop from
383  * outside of the monitoring thread must be protected by @kdamond_lock.
384  *
385  * Note that the monitoring thread protects only @kdamond and @kdamond_stop via
386  * @kdamond_lock.  Accesses to other fields must be protected by themselves.
387  *
388  * @primitive:	Set of monitoring primitives for given use cases.
389  * @callback:	Set of callbacks for monitoring events notifications.
390  *
391  * @min_nr_regions:	The minimum number of adaptive monitoring regions.
392  * @max_nr_regions:	The maximum number of adaptive monitoring regions.
393  * @adaptive_targets:	Head of monitoring targets (&damon_target) list.
394  * @schemes:		Head of schemes (&damos) list.
395  */
396 struct damon_ctx {
397 	unsigned long sample_interval;
398 	unsigned long aggr_interval;
399 	unsigned long primitive_update_interval;
400 
401 /* private: internal use only */
402 	struct timespec64 last_aggregation;
403 	struct timespec64 last_primitive_update;
404 
405 /* public: */
406 	struct task_struct *kdamond;
407 	struct mutex kdamond_lock;
408 
409 	struct damon_primitive primitive;
410 	struct damon_callback callback;
411 
412 	unsigned long min_nr_regions;
413 	unsigned long max_nr_regions;
414 	struct list_head adaptive_targets;
415 	struct list_head schemes;
416 };
417 
418 static inline struct damon_region *damon_next_region(struct damon_region *r)
419 {
420 	return container_of(r->list.next, struct damon_region, list);
421 }
422 
423 static inline struct damon_region *damon_prev_region(struct damon_region *r)
424 {
425 	return container_of(r->list.prev, struct damon_region, list);
426 }
427 
428 static inline struct damon_region *damon_last_region(struct damon_target *t)
429 {
430 	return list_last_entry(&t->regions_list, struct damon_region, list);
431 }
432 
433 #define damon_for_each_region(r, t) \
434 	list_for_each_entry(r, &t->regions_list, list)
435 
436 #define damon_for_each_region_safe(r, next, t) \
437 	list_for_each_entry_safe(r, next, &t->regions_list, list)
438 
439 #define damon_for_each_target(t, ctx) \
440 	list_for_each_entry(t, &(ctx)->adaptive_targets, list)
441 
442 #define damon_for_each_target_safe(t, next, ctx)	\
443 	list_for_each_entry_safe(t, next, &(ctx)->adaptive_targets, list)
444 
445 #define damon_for_each_scheme(s, ctx) \
446 	list_for_each_entry(s, &(ctx)->schemes, list)
447 
448 #define damon_for_each_scheme_safe(s, next, ctx) \
449 	list_for_each_entry_safe(s, next, &(ctx)->schemes, list)
450 
451 #ifdef CONFIG_DAMON
452 
453 struct damon_region *damon_new_region(unsigned long start, unsigned long end);
454 
455 /*
456  * Add a region between two other regions
457  */
458 static inline void damon_insert_region(struct damon_region *r,
459 		struct damon_region *prev, struct damon_region *next,
460 		struct damon_target *t)
461 {
462 	__list_add(&r->list, &prev->list, &next->list);
463 	t->nr_regions++;
464 }
465 
466 void damon_add_region(struct damon_region *r, struct damon_target *t);
467 void damon_destroy_region(struct damon_region *r, struct damon_target *t);
468 
469 struct damos *damon_new_scheme(
470 		unsigned long min_sz_region, unsigned long max_sz_region,
471 		unsigned int min_nr_accesses, unsigned int max_nr_accesses,
472 		unsigned int min_age_region, unsigned int max_age_region,
473 		enum damos_action action, struct damos_quota *quota,
474 		struct damos_watermarks *wmarks);
475 void damon_add_scheme(struct damon_ctx *ctx, struct damos *s);
476 void damon_destroy_scheme(struct damos *s);
477 
478 struct damon_target *damon_new_target(unsigned long id);
479 void damon_add_target(struct damon_ctx *ctx, struct damon_target *t);
480 bool damon_targets_empty(struct damon_ctx *ctx);
481 void damon_free_target(struct damon_target *t);
482 void damon_destroy_target(struct damon_target *t);
483 unsigned int damon_nr_regions(struct damon_target *t);
484 
485 struct damon_ctx *damon_new_ctx(void);
486 void damon_destroy_ctx(struct damon_ctx *ctx);
487 int damon_set_targets(struct damon_ctx *ctx,
488 		unsigned long *ids, ssize_t nr_ids);
489 int damon_set_attrs(struct damon_ctx *ctx, unsigned long sample_int,
490 		unsigned long aggr_int, unsigned long primitive_upd_int,
491 		unsigned long min_nr_reg, unsigned long max_nr_reg);
492 int damon_set_schemes(struct damon_ctx *ctx,
493 			struct damos **schemes, ssize_t nr_schemes);
494 int damon_nr_running_ctxs(void);
495 
496 int damon_start(struct damon_ctx **ctxs, int nr_ctxs);
497 int damon_stop(struct damon_ctx **ctxs, int nr_ctxs);
498 
499 #endif	/* CONFIG_DAMON */
500 
501 #ifdef CONFIG_DAMON_VADDR
502 bool damon_va_target_valid(void *t);
503 void damon_va_set_primitives(struct damon_ctx *ctx);
504 #endif	/* CONFIG_DAMON_VADDR */
505 
506 #ifdef CONFIG_DAMON_PADDR
507 bool damon_pa_target_valid(void *t);
508 void damon_pa_set_primitives(struct damon_ctx *ctx);
509 #endif	/* CONFIG_DAMON_PADDR */
510 
511 #endif	/* _DAMON_H */
512