Posted Apr 24, 2013 4:23 UTC (Wed) by dlang (✭ supporter ✭, #313)
Parent article: LSFMM: Soft reclaim
I can sure see cases where the soft limit of one level could be such that all children of it can still be below their soft limits.
As a result, it seems to be that both modes of operation need to be supported
1. if a group is over it's softlimit, look inside that group.
1b. if any child is over it's limit, reclaim from it
1c. if all children are within their soft limit, but the parent is still over it's softlimit, reclaim from all children.
that way the kernel is not determining the policy that the limit must be >= the sum of the limits of it's children. If people believe that's the way they want to configure things, it works. If people want to configure things differently, it still works in a predictable way.
and let's face it, people are going to misconfigure systems, even if the policy is that the parent must be >= the children, how is the kernel going to deal with these systems?