Skip to contentSkip to conversation

mm/damon/core: make passed_sample_intervals comparisons overflow-safe

PATCH Applied SeongJae Park 1 participant 4 messages Mar 7

Subsystems (2)
lore
Share & tools
lore

SeongJae Park

author maintainer
#0 lore raw

cover mm/damon/core: make passed_sample_intervals comparisons overflow-safe

DAMON accounts time using its own jiffies-like time counter, namely damon_ctx->passed_sample_intervals. The counter is incremented on each iteration of kdamond_fn() main loop, which sleeps at least one sample interval. Hence the name is like that.

DAMON has time-periodic operations including monitoring results aggregation and DAMOS action application. DAMON sets the next time to do each of such operations in the passed_sample_intervals unit. And it does the operation when the counter becomes the same to or larger than the pre-set values, and update the next time for the operation. Note that the operation is done not only when the values exactly match but also when the time is passed, because the values can be updated for online-committed DAMON parameters.

The counter is 'unsigned long' type, and the comparison is done using normal comparison operators. It is not safe from overflows. This can cause rare and limited but odd situations.

Let's suppose there is an operation that should be executed every 20 sampling intervals, and the passed_sample_intervals value for next execution of the operation is ULONG_MAX - 3. Once the passed_sample_intervals reaches ULONG_MAX - 3, the operation will be executed, and the next time value for doing the operation becomes 17 (ULONG_MAX - 3 + 20), since overflow happens. In the next iteration of the kdamond_fn() main loop, passed_sample_intervals is larger than the next operation time value, so the operation will be executed again. It will continue executing the operation for each iteration, until the passed_sample_intervals also overflows.

Note that this will not be common and problematic in the real world. The sampling interval, which takes for each passed_sample_intervals increment, is 5 ms by default. And it is usually [auto-]tuned for hundreds of milliseconds. That means it takes about 248 days or 4,971 days to have the overflow on 32 bit machines when the sampling interval is 5 ms and 100 ms, respectively (1<<32 * sampling_interval_in_seconds / 3600 / 24). On 64 bit machines, the numbers become 2924712086.77536 and 58494241735.5072 years. So the real user impact is negligible. But still this is better to be fixed as long as the fix is simple and efficient.

Fix this by simply replacing the overflow-unsafe native comparison operators with the existing overflow-safe time comparison helpers.

The first patch only cleans up the next DAMOS action application time setup for consistency and reduced code. The second and the third patches update DAMOS action application time setup and rest, respectively.

SeongJae Park (3):
mm/damon/core: remove damos_set_next_apply_sis() duplicates mm/damon/core: use time_before() for next_apply_sis mm/damon/core: use time_after_eq() in kdamond_fn()

 mm/damon/core.c | 31 ++++++++++++++-----------------
 1 file changed, 14 insertions(+), 17 deletions(-)

base-commit: 4ef41fd88e37fcdc1751edd9bc57403dba2742a3

Signature · 1 line
2.47.3
author maintainer
#1 lore raw

PATCH 1/3 mm/damon/core: remove damos_set_next_apply_sis() duplicates

There is a function for damos->next_apply_sis setup. But some places are open-coding it. Consistently use the helper.

---
 mm/damon/core.c | 11 ++---------
 1 file changed, 2 insertions(+), 9 deletions(-)
mm/damon/core.c +2 -9
index e8c44541754f7..c4c30824465ee 100644
@@ -2450,8 +2450,6 @@ static void kdamond_apply_schemes(struct damon_ctx *c)
 	struct damon_target *t;
 	struct damon_region *r;
 	struct damos *s;
-	unsigned long sample_interval = c->attrs.sample_interval ?
-		c->attrs.sample_interval : 1;
 	bool has_schemes_to_apply = false;
 
 	damon_for_each_scheme(s, c) {
@@ -2482,9 +2480,7 @@ static void kdamond_apply_schemes(struct damon_ctx *c)
 		if (c->passed_sample_intervals < s->next_apply_sis)
 			continue;
 		damos_walk_complete(c, s);
-		s->next_apply_sis = c->passed_sample_intervals +
-			(s->apply_interval_us ? s->apply_interval_us :
-			 c->attrs.aggr_interval) / sample_interval;
+		damos_set_next_apply_sis(s, c);
 		s->last_applied = NULL;
 		damos_trace_stat(c, s);
 	}
@@ -2858,7 +2854,6 @@ static void kdamond_init_ctx(struct damon_ctx *ctx)
 {
 	unsigned long sample_interval = ctx->attrs.sample_interval ?
 		ctx->attrs.sample_interval : 1;
-	unsigned long apply_interval;
 	struct damos *scheme;
 
 	ctx->passed_sample_intervals = 0;
@@ -2869,9 +2864,7 @@ static void kdamond_init_ctx(struct damon_ctx *ctx)
 		ctx->attrs.intervals_goal.aggrs;
 
 	damon_for_each_scheme(scheme, ctx) {
-		apply_interval = scheme->apply_interval_us ?
-			scheme->apply_interval_us : ctx->attrs.aggr_interval;
-		scheme->next_apply_sis = apply_interval / sample_interval;
+		damos_set_next_apply_sis(scheme, ctx);
 		damos_set_filters_default_reject(scheme);
 	}
 }
Signature · 1 line
2.47.3
author maintainer
#2 lore raw

PATCH 2/3 mm/damon/core: use time_before() for next_apply_sis

In reply to SeongJae Park

damon_ctx->passed_sample_intervals and damos->next_apply_sis are unsigned long, and compared via normal comparison operators. It is unsafe from overflow. Use time_before(), which is safe from overflow when correctly used, instead.

---
 mm/damon/core.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
mm/damon/core.c +3 -3
index c4c30824465ee..4ede733a94cc0 100644
@@ -2095,7 +2095,7 @@ static void damon_do_apply_schemes(struct damon_ctx *c,
 	damon_for_each_scheme(s, c) {
 		struct damos_quota *quota = &s->quota;
 
-		if (c->passed_sample_intervals < s->next_apply_sis)
+		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
 			continue;
 
 		if (!s->wmarks.activated)
@@ -2453,7 +2453,7 @@ static void kdamond_apply_schemes(struct damon_ctx *c)
 	bool has_schemes_to_apply = false;
 
 	damon_for_each_scheme(s, c) {
-		if (c->passed_sample_intervals < s->next_apply_sis)
+		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
 			continue;
 
 		if (!s->wmarks.activated)
@@ -2477,7 +2477,7 @@ static void kdamond_apply_schemes(struct damon_ctx *c)
 	}
 
 	damon_for_each_scheme(s, c) {
-		if (c->passed_sample_intervals < s->next_apply_sis)
+		if (time_before(c->passed_sample_intervals, s->next_apply_sis))
 			continue;
 		damos_walk_complete(c, s);
 		damos_set_next_apply_sis(s, c);
Signature · 1 line
2.47.3
author maintainer
#3 lore raw

PATCH 3/3 mm/damon/core: use time_after_eq() in kdamond_fn()

In reply to SeongJae Park

damon_ctx->passed_sample_intervals and damon_ctx->next_*_sis are unsigned long. Those are compared in kdamond_fn() using normal comparison operators. It is unsafe from overflow. Use time_after_eq(), which is safe from overflows when correctly used, instead.

---
 mm/damon/core.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
mm/damon/core.c +9 -5
index 4ede733a94cc0..cd2d7a8e3fe92 100644
@@ -2915,7 +2915,8 @@ static int kdamond_fn(void *data)
 		if (ctx->ops.check_accesses)
 			max_nr_accesses = ctx->ops.check_accesses(ctx);
 
-		if (ctx->passed_sample_intervals >= next_aggregation_sis) {
+		if (time_after_eq(ctx->passed_sample_intervals,
+					next_aggregation_sis)) {
 			kdamond_merge_regions(ctx,
 					max_nr_accesses / 10,
 					sz_limit);
@@ -2935,10 +2936,12 @@ static int kdamond_fn(void *data)
 
 		sample_interval = ctx->attrs.sample_interval ?
 			ctx->attrs.sample_interval : 1;
-		if (ctx->passed_sample_intervals >= next_aggregation_sis) {
+		if (time_after_eq(ctx->passed_sample_intervals,
+					next_aggregation_sis)) {
 			if (ctx->attrs.intervals_goal.aggrs &&
-					ctx->passed_sample_intervals >=
-					ctx->next_intervals_tune_sis) {
+					time_after_eq(
+						ctx->passed_sample_intervals,
+						ctx->next_intervals_tune_sis)) {
 				/*
 				 * ctx->next_aggregation_sis might be updated
 				 * from kdamond_call().  In the case,
@@ -2972,7 +2975,8 @@ static int kdamond_fn(void *data)
 			kdamond_split_regions(ctx);
 		}
 
-		if (ctx->passed_sample_intervals >= next_ops_update_sis) {
+		if (time_after_eq(ctx->passed_sample_intervals,
+					next_ops_update_sis)) {
 			ctx->next_ops_update_sis = next_ops_update_sis +
 				ctx->attrs.ops_update_interval /
 				sample_interval;
Signature · 1 line
2.47.3