On Thu, Sep 05, 2019 at 09:31:27 +0100, Peter Zijlstra wrote...
> On Fri, Aug 30, 2019 at 10:49:36AM -0700, subhra mazumdar wrote: >> Add Cgroup interface for latency-nice. Each CPU Cgroup adds a new file >> "latency-nice" which is shared by all the threads in that Cgroup. > > *sigh*, no. We start with a normal per task attribute, and then later, > if it is needed and makes sense, we add it to cgroups. FWIW, to add on top of what Peter says, we used this same approach for uclamp and it proved to be a very effective way to come up with a good design. General principles have been: - a system wide API [1] (under /proc/sys/kernel/sched_*) defines default values for all tasks affected by that feature. This interface has to define also upper bounds for task specific values. Thus, in the case of latency-nice, it should be set by default to the MIN value, since that's the current mainline behaviour: all tasks are latency sensitive. - a per-task API [2] (via the sched_setattr() syscall) can be used to relax the system wide setting thus implementing a "nice" policy. - a per-taskgroup API [3] (via cpu controller's attributes) can be used to relax the system-wide settings and restrict the per-task API. The above features are worth to be added in that exact order. > Also, your Changelog fails on pretty much every point. It doesn't > explain why, it doesn't describe anything and so on. On the description side, I guess it's worth to mention somewhere to which scheduling classes this feature can be useful for. It's worth to mention that it can apply only to: - CFS tasks: for example, at wakeup time a task with an high latency-nice should avoid to preempt a low latency-nice task. Maybe by mapping the latency nice value into proper vruntime normalization value? - RT tasks: for example, at wakeup time a task with an high latency-nice value could avoid to preempt a CFS task. I'm sure there will be discussion about some of these features, that's why it's important in the proposal presentation to keep a well defined distinction among the "mechanisms and API" and how we use the new concept to "bias" some scheduler policies. > From just reading the above, I would expect it to have the range > [-20,19] just like normal nice. Apparently this is not so. Regarding the range for the latency-nice values, I guess we have two options: - [-20..19], which makes it similar to priorities downside: we quite likely end up with a kernel space representation which does not match the user-space one, e.g. look at task_struct::prio. - [0..1024], which makes it more similar to a "percentage" Being latency-nice a new concept, we are not constrained by POSIX and IMHO the [0..1024] scale is a better fit. That will translate into: latency-nice=0 : default (current mainline) behaviour, all "biasing" policies are disabled and we wakeup up as fast as possible latency-nice=1024 : maximum niceness, where for example we can imaging to turn switch a CFS task to be SCHED_IDLE? Best, Patrick [1] commit e8f14172c6b1 ("sched/uclamp: Add system default clamps") [2] commit a509a7cd7974 ("sched/uclamp: Extend sched_setattr() to support utilization clamping") [3] 5 patches in today's tip/sched/core up to: commit babbe170e053 ("sched/uclamp: Update CPU's refcount on TG's clamp changes") -- #include <best/regards.h> Patrick Bellasi