On Tue, Sep 13, 2016 at 12:53:27PM -0700, Arve Hjønnevåg wrote: > On Tue, Sep 13, 2016 at 12:32 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > On Mon, Sep 12, 2016 at 08:44:09PM -0700, Arve Hjønnevåg wrote: > > > >> A previous attempt to fix this problem, changed the lock to use > >> rt_mutex instead of mutex, but this apparently did not work as well as > >> this patch. I believe the added overhead was noticeable, and it did > >> not work when the preempted thread was in a different cgroup (I don't > >> know if this is still the case). > > > > Do you actually have RR/FIFO/DL tasks? Currently PI isn't > > defined/implemented for OTHER. > > > > Most of the tasks here are not RR/FIFO/DL tasks. I don't see anything > in the rtmutex code or documentation that indicates that they don't > work for normal tasks. From what I can tell the priority gets boosted > in every case. This may not work as well for CFS tasks as for realtime > tasks, but it should at least help when there is a large priority > difference. It does something (it used to explicitly ignore OTHER) but its not something well defined or usable. > > cgroups should be irrelevant, PI is unaware of them. > > I don't think cgroups are irrelevant. PI being unaware of them > explains the problem I described. If the task that holds the lock is > in a cgroup that has a low cpu.shares value, then boosting the task's > priority within that group does necessarily make it any more likely to > run. See, the problem is that 'priority' is a meaningless concept for OTHER/CFS. In any case, typically only RT tasks care about PI, and the traditional Priority Inheritance algorithm only really works correctly for FIFO. As is RR has issues and DL is a crude hack, CFS is really just an accident by not explicitly exempting it. We could define a meaningful something for CFS and implement that, but it isn't currently done. _______________________________________________ devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxx http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
