Serge E. Hallyn wrote: > Quoting Oren Laadan (orenl@xxxxxxxxxxxxxxx): >> >> Sukadev Bhattiprolu wrote: >>> The help message for --freezer=CGROUP option says: >>> >>> freeze tasks in freezer group CGROUP on success >>> >>> and if this option specifies an existing cgroup, then (and only >>> then) is the RESTART_FROZEN flag to sys_restart() set. >>> >>> Does the --freezer option have to specify an existing cgroup ? >> As it is now - yes; >> >> In theory I would like use -F (--freezer) without an argument, >> to use RESTART_FROZEN without explicitly adding tasks to a >> target cgroup. >> >> The trick is that we need to put all the restarted tasks in the >> traget cgroup. The coordinator must remain outside that cgroup, >> or it will be frozen as well when restart succeeds. >> >> This means that restarted tasks (at least the root task) have >> to be explicitly added to the cgroup - which is why I require >> that the cgroup name be passed. >> >> This could be avoided if we could add 'restart' to a cgruop and >> then after spawning the entire tree, we could remove 'restart' >> from that cgroup. I don't think this is possible ? > > Why not? You can certainly move to freezer:/restart.XYZ, spawn > tasks, then move back to freezer:/. And since the tasks should > hang in sys_restart() until the coordinator calls sys_restart() > itself, there should be no race issues. Do you mean that something > else won't work? No, I just wasn't thinking right... But even in this case, we need some more wisdom: we need to know where the coordinator starts (in which cgroup) before moving it to a new cgroup, so we can put it back in the original cgroup. Is there an API for that ? > >>> Is there a way to specify that the restarted application belong >>> to a new (yet-to-be created) cgroup AND to leave the application >>> FROZEN after restart ? Or would we need a new command line option >>> and field in 'struct app_restart_args' to specify this ? >> This is exactly what --freezer does. Well, alsmot - except for >> "yet-to-be created". Do you want 'restart' to create the cgroup >> if it doesn't exist ? > > I thought that was what Suka was asking for. If he wants to add it to the 'restart' utility, then sure, with a proper argument. I'm not sure that creating a control group fits into the mandate of a library call/api. Instead, I'd expect the caller to arrange for a cgroup to be created. Oren. _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers
