Hi Vincent, On Mon, Jun 06, 2016 at 02:32:39PM +0200, Vincent Guittot wrote: > > Unlike the switch to fair class case, if the task is on_rq, it will be > > enqueued after we move task groups, so the simplest solution is to reset > > the task's last_update_time when we do task_move_group(), but not to > > attach sched avgs in task_move_group(), and then let enqueue_task() do > > the sched avgs attachment. > > According to the review of the previous version > http://www.gossamer-threads.com/lists/linux/kernel/2450678#2450678, > only the use case switch to fair with a task that has never been > queued as a CFS task before, can have the issue and this will not > happen for other use cases described above. > So can you limit the description in the commit message to just this > use case unless you have discovered new use cases and in this case > please add the description.
I assure you that I will, :) > Then, the problem with this use case, comes that last_update_time == 0 > has a special meaning ( task has migrated ) and we initialize > last_update_time with this value. In general, the meaning of last_update_time == 0 is: not yet attached to any cfs queue. > A much more simple solution would be > to prevent last_update_time to be initialized with this special value. > We can initialize the last_update_time of a sched_entity to 1 as an > example which is easier than these changes. Then at least, we must differentiate CONFIG_FAIR_GROUP_SCHED and !CONFIG_FAIR_GROUP_SCHED. And I am not sure whether this is a simpler solution, as I haven't sorted out the whole thing. IMO, this solution seems a little too hacky and I'd prefer not if we have alternative.
