On Mon, Feb 06, 2023 at 02:39:17PM -0800, Yosry Ahmed wrote: > On Mon, Feb 6, 2023 at 2:36 PM Tejun Heo <tj@xxxxxxxxxx> wrote: > > > > On Mon, Feb 06, 2023 at 02:32:10PM -0800, Yosry Ahmed wrote: > > > I guess it boils down to which we want: > > > (a) Limit the amount of memory processes in a cgroup can be pinned/locked. > > > (b) Limit the amount of memory charged to a cgroup that can be pinned/locked. > > > > > > The proposal is doing (a), I suppose if this was part of memcg it > > > would be (b), right? > > > > > > I am not saying it should be one or the other, I am just making sure > > > my understanding is clear. > > > > I don't quite understand what the distinction would mean in practice. It's > > just odd to put locked memory in a separate controller from interface POV. > > Assume we have 2 cgroups, A and B. A process in cgroup A creates a > tmpfs file and writes to it, so the memory is now charged to cgroup A. > Now imagine a process in cgroup B tries to lock this memory. > - With (a) the amount of locked memory will count toward's cgroup A's > limit, because cgroup A is charged for the memory. > - With (b) the amount of locked memory will count toward's cgroup B's > limit, because a process in cgroup B is locking the memory. > > I agree that it is confusing from an interface POV. Oh yeah, that's confusing. I'd go with (a) for consistency with the rest of memcg - locked memory should fit inside e.g. memory.max. The problem with shared memory accounting exists for non-locked memory as well and prolly best to handle the same way rather than handling differently. Thanks. -- tejun
