* menage@xxxxxxxxxx <menage@xxxxxxxxxx> [2009-07-13 23:49:16]: > On Mon, Jul 13, 2009 at 10:56 PM, Balbir Singh<balbir@xxxxxxxxxxxxxxxxxx> wrote: > >> > >> Waiting for the next scheduling point might be too long, since a > >> thread can block for arbitrary amounts of time and keeping the marker > >> around for arbitrary time (unless we add a new task_struct field) > >> would be tricky. Marking the cgroup or tgid as being migrated which > >> then triggers the extra synchronization in the fork path (but which > >> isn't needed at other times) is probably where we'll end up. > > > > > > Hmm... but we would not need that information till we schedule the > > tasks, adding a field to task_struct is what I had in mind. > > Waiting until schedule to move the threads would result in them still > showing up in the old "tasks" file until they next ran, which would be > confusing and misleading. > Yes, agreed, even first access to the data struture based lazy migration might not be too helpful, since it can be very context dependent. > As a first cut, we were planning to add an rwsem that gets taken for > read in cgroup_fork(), released in cgroup_post_fork(), and taken for > write when moving an entire process to a new cgroup; not ideal > performance-wise, but safe. > > If adding a field to task_struct is an option, then the rwsem could be > per thread-group leader, which would reduce contention. > We should also document that moving large processes with several threads can be expensive. -- Balbir _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers
