Re: [RFC PATCH v3 09/14] mm/memcontrol: Make memory.high tier-aware

[Joshua Hahn]

On Tue, 18 Aug 2026 09:21:15 +0800 Song Hu <husong@xxxxxxxxxx> wrote:

Hi Song, thank you for your reply.

> (pre-existing, not from this series - mentioning it here because
> it reworks memory.high behaviour)
> 
> The penalty computation in calculate_high_delay() can overflow:
> 
> 	penalty_jiffies = max_overage * max_overage * HZ;
> 
> max_overage is a fixed-point relative overage:
> calculate_overage() returns (usage - high) << 20 / high, so the
> square times HZ wraps once (usage - high) / high reaches
> 4096 / sqrt(HZ).  That is roughly 130x at HZ=1000 and 259x at
> HZ=250 - reachable when memory.high is tightened far below the
> current usage of a large working set.
> 
> Most wrapped values are still caught by the 2s clamp in
> mem_cgroup_handle_over_high(), but residues below the clamp pass
> through, and some of them produce far less throttling than
> intended - down to no sleep at all.
> 
> check_mul_overflow() with a clamp would saturate the penalty
> instead of wrapping it.  Could be folded in while memory.high is
> being reworked, or a standalone fix.

I think this could be a real concern, but I'm not sure that my series
is the right place to fold this change in. Namely because I don't
touch the penalty_jiffies / throttling part at all.

Maybe a standalone fix could be worth it? : -) I have no plans at the
moment so please feel free to take a crack at it if you think it's
a good idea!
Joshua