mm/damon/core: avoid time-quota permanently disabling scheme

RFC PATCH In review SeongJae Park 1 participant 2 messages Apr 5 – Apr 6

Subsystems (3)

Stable branch Memory management DAMON

lore

Share & tools

lore

SeongJae Park

author maintainer

2026-04-05 22:54 UTC #0 lore raw

RFC PATCH 1/1 mm/damon/core: avoid time-quota permanently disabling scheme

When the throughput of a DAMOS scheme is very slow, DAMOS time quota can make the effective size quota smaller than damon_ctx->min_region_sz. In the case, damos_apply_scheme() will skip applying the action, because the action is tried at region level, which requires >=min_region_sz size. That is, the quota is effectively exceeded for the quota charge window.

Because no action will be applied, the total_charged_sz and total_charged_ns are also not updated. damos_set_effective_quota() will try to update the effective size quota before starting the next charge window. However, because the total_charged_sz and total_charged_ns have not updated, the throughput and effective size quota are also not changed. Since effective size quota can only be decreased, other effective size quota update factors including DAMOS quota goals and size quota cannot make any change, either.

As a result, the scheme is unexpectedly deactivated until the user notices and mitigates the situation. The users can mitigate this situation by changing the time quota online or re-install the scheme. While the mitigation is somewhat straightforward, finding the situation would be challenging, because DAMON is not providing good observabilities for that. Even if such observability is provided, doing the additional monitoring and the mitigation is somewhat cumbersome and not aligned to the intention of the time quota. The time quota was intended to help reduce the user's administration overhead.

Fix the problem by setting time quota-modified effective size quota be at least min_region_sz always.

The issue was discovered [1] by sashiko.

[1] https://lore.kernel.org/[email protected]

Fixes 1cd243030059 ("mm/damon/schemes: implement time quota")

Cc <[email protected]> # 5.16.x

Signed-off-by SeongJae Park <[email protected]>

---
 mm/damon/core.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

mm/damon/core.c +5 -3

index 3bc7a2bbfe7de..12544c60531d3 100644
@@ -2384,7 +2384,8 @@ static void damos_goal_tune_esz_bp_temporal(struct damos_quota *quota)
 /*
  * Called only if quota->ms, or quota->sz are set, or quota->goals is not empty
  */
-static void damos_set_effective_quota(struct damos_quota *quota)
+static void damos_set_effective_quota(struct damos_quota *quota,
+		struct damon_ctx *ctx)
 {
 	unsigned long throughput;
 	unsigned long esz = ULONG_MAX;

@@ -2409,6 +2410,7 @@ static void damos_set_effective_quota(struct damos_quota *quota)
 		else
 			throughput = PAGE_SIZE * 1024;
 		esz = min(throughput * quota->ms, esz);
+		esz = max(ctx->min_region_sz, esz);
 	}
 
 	if (quota->sz && quota->sz < esz)

@@ -2445,7 +2447,7 @@ static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
 	/* First charge window */
 	if (!quota->total_charged_sz && !quota->charged_from) {
 		quota->charged_from = jiffies;
-		damos_set_effective_quota(quota);
+		damos_set_effective_quota(quota, c);
 	}
 
 	/* New charge window starts */

@@ -2460,7 +2462,7 @@ static void damos_adjust_quota(struct damon_ctx *c, struct damos *s)
 		quota->charged_sz = 0;
 		if (trace_damos_esz_enabled())
 			cached_esz = quota->esz;
-		damos_set_effective_quota(quota);
+		damos_set_effective_quota(quota, c);
 		if (trace_damos_esz_enabled() && quota->esz != cached_esz)
 			damos_trace_esz(c, s, quota);
 	}

base-commit: 8c08ea3625d4fa8c1c74b208c0c630fce76e14a9

Signature · 1 line

2.47.3

SeongJae Park

author maintainer

2026-04-06 15:57 UTC #1 lore raw

Re: (sashiko review) [RFC PATCH] mm/damon/core: avoid time-quota permanently disabling scheme

TL; DR: sashiko found no issue in this patch. I will repost this patch after dropping RFC, soon.

Forwarding sashiko review in a reply format with my inline comments below, for details of my view and doing discussions via mails if needed.

Quoted 35 lines

# review url: https://sashiko.dev/#/patchset/[email protected]
# start of sashiko.dev inline review
commit 93dc55170095d9a174cb626df2cde4fe5e49b9a6
Author: SeongJae Park <[email protected]>
mm/damon/core: avoid time-quota permanently disabling scheme
When a DAMOS scheme throughput is very slow, the time quota can make the effective size quota smaller than the minimum region size. This causes the action to be skipped, freezing the effective size quota updates. This patch sets the time quota-modified effective size quota to be at least min_region_sz.
Link https://lore.kernel.org/[email protected] open here
Quoted 4 lines
mm/damon/core.c +0 -0
index 3bc7a2bbfe7de..12544c60531d3 100644
[ ... ]
mm/damon/core.c static void damos_set_effective_quota(struct damos_quota *quota) · line 2413
Quoted 5 lines
mm/damon/core.c +1 -0
@@ -2409,6 +2410,7 @@ static void damos_set_effective_quota(struct damos_quota *quota)
 		else
 			throughput = PAGE_SIZE * 1024;
 		esz = min(throughput * quota->ms, esz);
+		esz = max(ctx->min_region_sz, esz);
Does this minimum region size bound incorrectly override DAMOS quota goals?
In damos_set_effective_quota(), esz is first initialized based on quota->goals. The goal tuner can legitimately reduce esz below ctx->min_region_sz (for example, to 0) to completely pause the scheme if the system is over-achieving its target limits.

Quoted 8 lines

Because the max() operates on the combined esz instead of just the time quota component, it artificially inflates the quota back up to ctx->min_region_sz. This forces the scheme to process at least one region per charge window, overriding the goal tuner's request to pause the scheme.
If quota->ms is 0, this block is bypassed and the goal is properly respected, leading to inconsistent behavior depending on whether a time quota is configured.

That's correct, and it is an intended behavior.

It's true that it is bit inconsistent. But I don't think that will cause any real issue, since the difference is quite minimal (min_region_sz quota).

Quoted 8 lines

Could the lower bound be applied exclusively to the calculated time quota before it is combined with the goal-based quota?
For example:
unsigned long time_esz = max(throughput * quota->ms, ctx->min_region_sz); esz = min(time_esz, esz);

Maybe we can later improve the behavior by doing the time quota based esz setup before goals based setup. But that's not good for stable kernels hotfix in my opinion, so it is out of the scope of this patch.

Quoted 4 lines


# end of sashiko.dev inline review
# review url: https://sashiko.dev/#/patchset/[email protected]

Thanks,
SJ

# hkml [1] generated a draft of this mail.  You can regenerate
# this using below command:
#
#     hkml patch sashiko_dev --for_forwarding \
#             [email protected]
#
# [1] https://github.com/sjp38/hackermail