On Thu, Nov 21, 2024 at 8:15 AM Rafael J. Wysocki <rjw@xxxxxxxxxxxxx> wrote: > > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx> > > As stated by Len in [1], the extra delay added by msleep() to the > sleep time value passed to it can be significant, roughly between > 1.5 ns on systems with HZ = 1000 and as much as 15 ms on systems with > HZ = 100, which is hardly acceptable, at least for small sleep time > values. Maybe the problem statement is more clear with a concrete example: msleep(5) on the default HZ=250 on a modern PC takes about 11.9 ms. This results in over 800 ms of spurious system resume delay on systems such as the Dell XPS-13-9300, which use ASL Sleep(5ms) in a tight loop. (yes, this additional cost used to be over 1200 ms before the v6.12 msleep rounding fix) > - msleep(ms); > + u64 usec = ms * USEC_PER_MSEC, delta_us = 50; > + if (ms > 5) > + delta_us = (USEC_PER_MSEC / 100) * ms I measured 100 resume cycles on the Dell XPS 13 9300 on 4 kernels. Here is the measured fastest kernel resume time in msec for each: 1. 1921.292 v6.12 msleep (baseline) 2. 1115.579 v6.12 delta_us = (USEC_PER_MSEC / 100) * ms (this patch) 3. 1113.396 v6.12 delta_us = 50 4. 1107.835 v6.12 delta_us = 0 (I didn't average the 100 runs, because random very long device hiccups throw off the average) So any of #2, #3 and #4 are a huge step forward from what is shipping today! So considering #2 vs #3 vs #4.... I agree that it is a problem for the timer sub-system to work to maintain a 1ns granularity that it can't actually deliver. I think it is fine for the timer sub-system to allow calls to opt into timer slack -- some callers may actually know what number to use. However, I don't think that the timer sub-system should force callers to guess how much slack is appropriate. I think that a caller with 0 slack should be internally rounded up by the timer sub-system to the granularity that it can actually deliver with the timer that is currently in use on that system. Note also that slack of 0 doesn't mean that no coalescing can happen. A slack=0 timer can land within the slack another timer, and the other timer will be pulled forward to coalesce. The 50 usec default for user timer slack is certainly a magic number born of tests of interesting workloads on interesting systems on a certain date. It may not be the right number for other workloads, or other systems with other timers on other dates. My opinion... I don't see a justification for increasing timer slack with increasing duration. User-space timers don't pay this additional delay, why should the ASL programmer? Also, the graduated increasing slack with duration is a guess no more valid than the guess of a flat 50 usec. A flat 50 or a flat 0 have the virtue of being simple -- they will be simpler to understand and maintain in the future. But I can live with any of these options, since they are all a big step forward. thanks, Len Brown, Intel Open Source Technology Center