On Tue, Nov 19, 2024 at 4:08 PM Arjan van de Ven <arjan@xxxxxxxxxxxxxxx> wrote: > > On 11/19/2024 5:42 AM, Rafael J. Wysocki wrote: > > On Mon, Nov 18, 2024 at 3:35 PM Arjan van de Ven <arjan@xxxxxxxxxxxxxxx> wrote: > >> > >>> And the argument seems to be that it is better to always use more > >>> resources in a given path (ACPI sleep in this particular case) than to > >>> be somewhat inaccurate which is visible in some cases. > >>> > >>> This would mean that hrtimers should always be used everywhere, but they aren't. > >> > >> more or less rule of thumb is that regular timers are optimized for not firing case > >> (e.g. timeouts that get deleted when the actual event happens) while hrtimers > >> are optimized for the case where the timer is expected to fire. > > > > I've heard that, which makes me wonder why msleep() is still there. > > > > One thing that's rarely mentioned is that programming a timer in HW > > actually takes time, so if it is done too often, it hurts performance > > through latency (even if this is the TSC deadline timer). > > yup and this is why you want to group events together "somewhat", and which is why > we have slack, to allow that to happen So what do you think would be the minimum slack to use in this case? I thought about something on the order of 199 us, but now I'm thinking that 50 us would work too. Less than this - I'm not sure.
