On Thu, Sep 26, 2019 at 2:23 PM Mike Kravetz mike.kravetz@oracle.com wrote:
On 9/26/19 12:28 PM, David Rientjes wrote:
On Tue, 24 Sep 2019, Mina Almasry wrote:
I personally prefer the one counter approach only for the reason that it exposes less information about hugetlb reservations. I was not around for the introduction of hugetlb reservations, but I have fixed several issues having to do with reservations. IMO, reservations should be hidden from users as much as possible. Others may disagree.
I really hope that Aneesh will comment. He added the existing hugetlb cgroup code. I was not involved in that effort, but it looks like there might have been some thought given to reservations in early versions of that code. It would be interesting to get his perspective.
Changes included in patch 4 (disable region_add file_region coalescing) would be needed in a one counter approach as well, so I do plan to review those changes.
OK, FWIW, the 1 counter approach should be sufficient for us, so I'm not really opposed. David, maybe chime in if you see a problem here? From the perspective of hiding reservations from the user as much as possible, it is an improvement.
I'm only wary about changing the behavior of the current and having that regress applications. I'm hoping you and Aneesh can shed light on this.
I think neither Aneesh nor myself are going to be able to provide a complete answer on the use of hugetlb cgroup today, anybody could be using it without our knowledge and that opens up the possibility that combining the limits would adversely affect a real system configuration.
I agree that nobody can provide complete information on hugetlb cgroup usage today. My interest was in anything Aneesh could remember about development of the current cgroup code. It 'appears' that the idea of including reservations or mmap ranges was considered or at least discussed. But, those discussions happened more than 7 years old and my searches are not providing a complete picture. My hope was that Aneesh may remember those discussions.
If that is a possibility, I think we need to do some due diligence and try to deprecate allocation limits if possible. One of the benefits to separate limits is that we can make reasonable steps to deprecating the actual allocation limits, if possible: we could add warnings about the deprecation of allocation limits and see if anybody complains.
That could take the form of two separate limits or a tunable in the root hugetlb cgroup that defines whether the limits are for allocation or reservation.
Combining them in the first pass seems to be very risky and could cause pain for users that will not detect this during an rc cycle and will report the issue only when their distro gets it. Then we are left with no alternative other than stable backports and the separation of the limits anyway.
I agree that changing behavior of the existing controller is too risky. Such a change is likely to break someone.
I'm glad we're converging on keeping the existing behavior unchanged.
The more I think about it, the best way forward will be to retain the existing controller and create a new controller that satisfies the new use cases.
My guess is that a new controller needs to support cgroups-v2, which is fine. But can a new controller also support v1? Or is there a requirement that new controllers support *only* v2? I need whatever solution here to work on v1. Added Tejun to hopefully comment on this.
The question remains as to what that new controller will be. Does it control reservations only? Is it a combination of reservations and allocations? If a combined controller will work for new use cases, that would be my preference. Of course, I have not prototyped such a controller so there may be issues when we get into the details. For a reservation only or combined controller, the region_* changes proposed by Mina would be used.
Provided we keep the existing controller untouched, should the new controller track:
1. only reservations, or 2. both reservations and allocations for which no reservations exist (such as the MAP_NORESERVE case)?
I like the 'both' approach. Seems to me a counter like that would work automatically regardless of whether the application is allocating hugetlb memory with NORESERVE or not. NORESERVE allocations cannot cut into reserved hugetlb pages, correct? If so, then applications that allocate with NORESERVE will get sigbused when they hit their limit, and applications that allocate without NORESERVE may get an error at mmap time but will always be within their limits while they access the mmap'd memory, correct? So the 'both' counter seems like a one size fits all.
I think the only sticking point left is whether an added controller can support both cgroup-v2 and cgroup-v1. If I could get confirmation on that I'll provide a patchset.
-- Mike Kravetz