Oleg Nesterov <oleg@xxxxxxxxxx> writes: > On 06/18, Oleg Nesterov wrote: >> >> I only try to discuss the idea to break the circular reference. > > I don't know what I have missed, but this looks really right to me. > Besides, we have yet another problem: proc_flush_task()->mntput() > is just wrong. Consider the multithreaded execing init. > > I am going to simplify, test, and send the fix which moves mntput() > into free_pid_ns() paths. free_pid_ns is comparatively late, to release the kern_mount. > But first of all I think we should cleanup the pid_ns_prepare_proc() > logic. Imho, this code is really ugly. Please see the patches. Since I have a patchset that makes it possible to unshare the pid namespace about ready to send I figure we should combine the two efforts. This patchset is a prerequisite to my patches for giving namespaces file descriptors and allowing you to join and existing namespace. When I look over my old notes it appears there Daniel managed to hit this proc_mnt reference counting in that context. So that is definitely interesting. Oleg take a look I think I have combined the best of our two patchsets. Eric _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers
