-lsf-pc (I can add them back, but I did not want to spam the group) +Vladimir On Fri, 23 Feb 2018 08:42:01 +0100 Michal Hocko <mhocko@xxxxxxxxxx> wrote: > On Fri 23-02-18 09:01:23, Balbir Singh wrote: > > Changed the subject to reflect the discussion > > > > On Thu, 22 Feb 2018 14:34:25 +0100 > > Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > > > > On Fri 23-02-18 00:23:53, Balbir Singh wrote: > > > > On Fri, Feb 23, 2018 at 12:03 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > > > > On Thu 22-02-18 13:54:46, Balbir Singh wrote: > > > > > [...] > > > > >> 2. Memory cgroups - I don't see a pressing need for many new features, > > > > >> but I'd like to see if we can revive some old proposals around virtual > > > > >> memory limits > > > > > > > > > > Could you be more specific about usecase(s)? > > > > > > > > I had for a long time a virtual memory limit controller in -mm tree. > > > > The use case was to fail allocations as opposed to OOM'ing in the > > > > worst case as we do with the cgroup memory limits (actual page usage > > > > control). I did not push for it then since I got side-tracked. I'd > > > > like to pursue a use case for being able to fail allocations as > > > > opposed to OOM'ing on a per cgroup basis. I'd like to start the > > > > discussion again. > > > > > > So you basically want the strict no overcommit on the per memcg level? > > > > I don't think it implies strict no overcommit, the value sets the > > overcommit ratio (independent of the global vm.overcommit_ratio, which > > we can discuss on the side, since I don't want it to impact the use > > case). > > > > The goal of the controller was (and its optional, may not work well > > for sparse address spaces) > > > > 1. set the vm limit > > 2. If the limit is exceeded, fail at malloc()/mmap() as opposed to > > OOM'ing at page fault time > > this is basically strict no-overcommit I look at it more as Committed_AS accounting and controls not controlled or driven by CommitLimit, but something the administrator can derive, but your right the defaults would be CommitLimit > > > 3. Application handles the fault and decide not to proceed with the > > new task that needed more memory > > So you do not return ENOMEM but rather raise a signal? What that would > be? Nope, it will return ENOMEM > > > I think this leads to applications being able to deal with failures > > better. OOM is a big hammer > > Do you have any _specific_ usecase in mind? It's mostly my frustration with OOM kills I see, granted a lot of it is about sizing the memory cgroup correctly, but that is not an easy job. > > > > I am really skeptical, to be completely honest. The global behavior is > > > not very usable in most cases already. Making it per-memcg will just > > > amplify all the issues (application tend to overcommit their virtual > > > address space). Not to mention that you cannot really prevent from the > > > OOM killer because there are allocations outside of the address space. > > > > > > > Could you clarify on the outside address space -- as in shared > > allocations outside the cgroup? kernel allocations as a side-effect? > > basically anything that can be triggered from userspace and doesn't map > into the address space - page cache, fs metadata, drm buffers etc... Yep, the virtual memory limits controller is more about the Committed_AS. I also noticed that Vladimir tried something similar at https://lkml.org/lkml/2014/7/3/405 Balbir Singh. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
