Skip to contentSkip to conversation

mm/damon/stat: avoid 32-bit overflow in idle percentiles

PATCH In review Aveline Noir 3 participants 6 messages Oct 15 – Oct 16

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm/damon/stat: avoid 32-bit overflow in idle percentiles

On 32-bit, accounted_bytes * 100 can wrap before division by total_sz. Some entries in memory_idle_ms_percentiles then keep their old values instead of being updated.

For example, a 256 MiB i386 QEMU guest reported p0=19200 ms, but p50 and p100 were both zero.

Cast accounted_bytes to u64 before multiplying and use div64_ul() for the division. Compute the percentage once per region.

---
This is against mm-new ad9f2d463666. With the fix, the original test reports p0=p50=p100=19200 ms. The i386 kernel build, x86_64 W=1 stat.o build and external arithmetic checks passed.

I also ran 38 before/after VM tests and 20 same-image controls. Throughput was close overall, but median kdamond CPU rose by 0.0053 and 0.0221 percentage points of one guest CPU for stream and sparse, respectively. Some pairs had 5.91% lower throughput or 15.49% higher scan P99. So I can't yet tell whether these differences come from the patch.

 mm/damon/stat.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/damon/stat.c +3 -2
index b05b68f73e10..eccd372810d0 100644
@@ -114,7 +114,7 @@ static void damon_stat_set_idletime_percentiles(struct damon_ctx *c)
 {
 	struct damon_region **sorted_regions, *region;
 	int nr_regions;
-	unsigned long total_sz, accounted_bytes = 0;
+	unsigned long total_sz, accounted_bytes = 0, percentile;
 	int err, i, next_percentile = 0;
 
 	err = damon_stat_sort_regions(c, &sorted_regions, &nr_regions,
@@ -124,7 +124,8 @@ static void damon_stat_set_idletime_percentiles(struct damon_ctx *c)
 	for (i = 0; i < nr_regions; i++) {
 		region = sorted_regions[i];
 		accounted_bytes += region->ar.end - region->ar.start;
-		while (next_percentile <= accounted_bytes * 100 / total_sz)
+		percentile = div64_ul((u64)accounted_bytes * 100, total_sz);
+		while (next_percentile <= percentile)
 			memory_idle_ms_percentiles[next_percentile++] =
 				damon_stat_idletime(region) *
 				(long)c->attrs.aggr_interval / USEC_PER_MSEC;
#1 lore raw
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 4 lines

On 32-bit, accounted_bytes * 100 can wrap before division by total_sz. Some entries in memory_idle_ms_percentiles then keep their old values instead of being updated.

Quoted 8 lines

For example, a 256 MiB i386 QEMU guest reported p0=19200 ms, but p50 and p100 were both zero.

Cast accounted_bytes to u64 before multiplying and use div64_ul() for the division. Compute the percentage once per region.

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
maintainer
#2 lore raw
In reply to Aveline Noir

Hi Aveline,

This patch is sent as a followup of your previous bug report, right? Unless it is very urgent, please make sure the previous discussion is completed before making a followup action like sending this patch. As a rule of thumb for DAMON, I'd suggest waiting for ~10 hours after the last comment on the thread.

Quoted 10 lines · On Thu, 15 Oct 2026 22:52:04 +0800 Aveline Noir <[email protected]> wrote:

On 32-bit, accounted_bytes * 100 can wrap before division by total_sz. Some entries in memory_idle_ms_percentiles then keep their old values instead of being updated.

For example, a 256 MiB i386 QEMU guest reported p0=19200 ms, but p50 and p100 were both zero.

Cast accounted_bytes to u64 before multiplying and use div64_ul() for the division. Compute the percentage once per region.

Quoted 8 lines

---
This is against mm-new ad9f2d463666. With the fix, the original test reports p0=p50=p100=19200 ms. The i386 kernel build, x86_64 W=1 stat.o build and external arithmetic checks passed.

I also ran 38 before/after VM tests and 20 same-image controls. Throughput

What is 38?

Quoted 4 lines

was close overall, but median kdamond CPU rose by 0.0053 and 0.0221 percentage points of one guest CPU for stream and sparse, respectively. Some pairs had 5.91% lower throughput or 15.49% higher scan P99. So I can't yet tell whether these differences come from the patch.

Do you have a plan to run more tests to be able to tell that?

Quoted 19 lines

 mm/damon/stat.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/damon/stat.c +1 -1
index b05b68f73e10..eccd372810d0 100644
@@ -114,7 +114,7 @@ static void damon_stat_set_idletime_percentiles(struct damon_ctx *c)
 {
 	struct damon_region **sorted_regions, *region;
 	int nr_regions;
-	unsigned long total_sz, accounted_bytes = 0;
+	unsigned long total_sz, accounted_bytes = 0, percentile;
 	int err, i, next_percentile = 0;
 
 	err = damon_stat_sort_regions(c, &sorted_regions, &nr_regions,

mm/damon/stat.c static void damon_stat_set_idletime_percentiles(struct damon_ctx *c) · line 128

Quoted 6 lines
mm/damon/stat.c +2 -1
@@ -124,7 +124,8 @@ static void damon_stat_set_idletime_percentiles(struct damon_ctx *c)
 	for (i = 0; i < nr_regions; i++) {
 		region = sorted_regions[i];
 		accounted_bytes += region->ar.end - region->ar.start;
-		while (next_percentile <= accounted_bytes * 100 / total_sz)
+		percentile = div64_ul((u64)accounted_bytes * 100, total_sz);
+		while (next_percentile <= percentile)

Any reason to use div64_ul() instead of div_u64()?

Thanks,
SJ

[...]

#3 lore raw

Hi SJ,

Quoted 5 lines · SJ Park

This patch is sent as a followup of your previous bug report, right? Unless it is very urgent, please make sure the previous discussion is completed before making a followup action like sending this patch. As a rule of thumb for DAMON, I'd suggest waiting for ~10 hours after the last comment on the thread.

The patch was already prepared and additional tests were underway when your reply to the bug report arrived, so I went ahead and sent it. I'll leave more time for discussion before sending follow-up patches.

Quoted 1 line

What is 38?

38 individual VM runs, forming 19 before/after pairs. The 20 same-image runs formed 10 control pairs.

Quoted 1 line

Do you have a plan to run more tests to be able to tell that?

I've run longer VM tests, but even same-kernel runs vary substantially, so the overall patch effect remains unclear. In a fixed-input 32-bit userspace test with 10 regions over 256 MiB, the median increase was 227 ns/call (+42%). The old code updated 15 entries, versus 101 with the fix. I haven't separated the extra updates from arithmetic cost. Maybe we could record region counts and sizes alongside VM timings to check whether both versions are doing comparable work.

Quoted 1 line

Any reason to use div64_ul() instead of div_u64()?

total_sz is unsigned long. div_u64() would truncate its divisor on 64-bit systems if it exceeds U32_MAX. div64_ul() preserves that width and maps to div_u64() on 32-bit.

Thanks,
Aveline Noir

maintainer
#4 lore raw
Quoted 28 lines · On Fri, 16 Oct 2026 17:22:58 +0800 Aveline Noir <[email protected]> wrote:

Hi SJ,

Quoted 5 lines · SJ Park

This patch is sent as a followup of your previous bug report, right? Unless it is very urgent, please make sure the previous discussion is completed before making a followup action like sending this patch. As a rule of thumb for DAMON, I'd suggest waiting for ~10 hours after the last comment on the thread.

The patch was already prepared and additional tests were underway when your reply to the bug report arrived, so I went ahead and sent it. I'll leave more time for discussion before sending follow-up patches.

Quoted 1 line

What is 38?

38 individual VM runs, forming 19 before/after pairs. The 20 same-image runs formed 10 control pairs.

Quoted 1 line

Do you have a plan to run more tests to be able to tell that?

I've run longer VM tests, but even same-kernel runs vary substantially, so the overall patch effect remains unclear. In a fixed-input 32-bit userspace test with 10 regions over 256 MiB, the median increase was 227 ns/call (+42%). The old code updated 15 entries, versus 101 with the fix. I haven't separated the extra updates from arithmetic cost. Maybe we could record region counts and sizes alongside VM timings to check whether both versions are doing comparable work.

Quoted 5 lines
Quoted 1 line

Any reason to use div64_ul() instead of div_u64()?

total_sz is unsigned long. div_u64() would truncate its divisor on 64-bit systems if it exceeds U32_MAX. div64_ul() preserves that width and maps to div_u64() on 32-bit.

All answers make sense. Thank you.

Thanks,
SJ

[...]

maintainer Reviewed-by on v1 1/1
#5 lore raw
In reply to Aveline Noir
Quoted 11 lines · On Thu, 15 Oct 2026 22:52:04 +0800 Aveline Noir <[email protected]> wrote:

On 32-bit, accounted_bytes * 100 can wrap before division by total_sz. Some entries in memory_idle_ms_percentiles then keep their old values instead of being updated.

For example, a 256 MiB i386 QEMU guest reported p0=19200 ms, but p50 and p100 were both zero.

Cast accounted_bytes to u64 before multiplying and use div64_ul() for the division. Compute the percentage once per region.

From the next time, please give context about what the link is.

Quoted 1 line

I applied this patch to damon/next [1] tree, after slightly modifying the commit message for the abovely mentioned small thing. We are now quite close to next merge window. It may better to focus on making mm.git more stabilized and therefore ready for the next merge window, rather than adding more changes that are not really urgent. I understand this series is not really urgent. Hence, Andrew might not add this patch until next -rc1 release. In the case, I will request adding this to mm.git after next -rc1 release. So, no action from your side is needed for now. Let me know if you think this is really urgent.

mm/damon/stat.c

Quoted 16 lines

---
This is against mm-new ad9f2d463666. With the fix, the original test reports p0=p50=p100=19200 ms. The i386 kernel build, x86_64 W=1 stat.o build and external arithmetic checks passed.

I also ran 38 before/after VM tests and 20 same-image controls. Throughput was close overall, but median kdamond CPU rose by 0.0053 and 0.0221 percentage points of one guest CPU for stream and sparse, respectively. Some pairs had 5.91% lower throughput or 15.49% higher scan P99. So I can't yet tell whether these differences come from the patch.

 mm/damon/stat.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
mm/damon/stat.c +0 -0

The above empty space before 'diff --git' line was making 'git am' of this patch fails. I manually fixed it. Please do not add such empty space or something that could make 'git am' fails, from the next time.

[1] https://origin.kernel.org/doc/html/latest/mm/damon/maintainer-profile.html#scm-trees

Thanks,
SJ

[...]