> On May 14, 2014, 10:16 a.m., Jie Yu wrote: > > src/linux/cgroups.cpp, lines 1634-1645 > > <https://reviews.apache.org/r/20818/diff/4/?file=572874#file572874line1634> > > > > This sequence of operations are introduced in this commit. Probably you > > should contact the author to see if your current logic is OK. If yes, > > please add some comments to explain why it is OK. > > > > commit 77db3cb32f9d25656b86607f6f241b2303dbd982 > > Author: Brenden Matthews <brenden.matth...@airbnb.com> > > Date: Fri May 10 16:00:15 2013 -0700 > > > > Changed cgroups killTasks() sequence. > > > > The sequence is as follows: > > stop -> kill -> empty -> freeze -> kill -> thaw -> empty > > > > We also now ignore ESRCH errors from kill() in cgroups::kill(). > > > > Review: https://reviews.apache.org/r/11131
Brendan commented that this was introduced when it was observed that OOM'ed processes weren't being killed successfully. At this time the cgroups_isolator was disabling the kernel OOM killer and trying to handle the OOM in Mesos. Trying to handle this in userspace is a bad idea and could lead to processes stuck in weird states. This is no longer done in Mesos and the kernel OOM killer is used. - Ian ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/20818/#review42977 ----------------------------------------------------------- On May 5, 2014, 1:16 p.m., Ian Downes wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/20818/ > ----------------------------------------------------------- > > (Updated May 5, 2014, 1:16 p.m.) > > > Review request for mesos, Benjamin Hindman, Ben Mahler, Chi Zhang, Jie Yu, > and Vinod Kone. > > > Bugs: MESOS-759 and MESOS-976 > https://issues.apache.org/jira/browse/MESOS-759 > https://issues.apache.org/jira/browse/MESOS-976 > > > Repository: mesos-git > > > Description > ------- > > internal::Destroyer now uses a single pass of freeze, kill, thaw, reap > to kill all processes in a freezer cgroup. cgroups::destroy will not > complete until all processes have been reaped. A timeout is used for > each step so destroy will eventually fail rather than constantly > retrying. > > Added a test for destroying a freezer cgroup containing a stopped > process. > > > Diffs > ----- > > src/linux/cgroups.hpp 5a5735721fb9f051eee661edb08d1cdaa163d0f3 > src/linux/cgroups.cpp 8202c282f580d027a60ded2081962e96e4860f60 > src/slave/containerizer/isolators/cgroups/cpushare.cpp > b494a9236210245383e20fa9ab3dbac01e42f8dd > src/slave/containerizer/linux_launcher.cpp > 530e0bd64d71bad761a2eab3d6e2f2179a167b4b > src/tests/cgroups_tests.cpp 6ba9de622953e158feadaa9950618b0b13c9e832 > > Diff: https://reviews.apache.org/r/20818/diff/ > > > Testing > ------- > > make check # Linux > > > Thanks, > > Ian Downes > >
