tree bd7777531a7bab3726058d38ddaa2f5e94bc8622
parent 03ad4a42545591863f3f01be31e15d221cf29c75
author Brian Norris <briannorris@chromium.org> 1709252991 -0800
committer Nico Huber <nico.h@gmx.de> 1709990495 +0000

udelay: Lower the sleep vs delay threshold

By default, we busy-loop (a.k.a., "delay") for most delay values, and
only allow sleeping for large delays. But busy-looping is expensive, as
it wastes CPU cycles.

In a simple program that runs a bunch of samples of [1] over 1000
samples, I find that for 0.1 s (100000 us):

 64x2 AMD CPU (CONFIG_HZ=250 / CONFIG_NO_HZ_FULL=y):
   min diff: 60 us
   max diff: 831 us
   mean diff: 135 us

 4+2 Mediatek MT8183 CPU (CONFIG_HZ=1000 / CONFIG_NO_HZ_IDLE=y /
                          sysctl kernel.timer_highres=1):
   min diff: 70 us
   max diff: 1556 us
   mean diff: 146 us

 4+2 Mediatek MT8183 CPU (CONFIG_HZ=1000 / CONFIG_NO_HZ_IDLE=y /
                          sysctl kernel.timer_highres=0):
   min diff: 94 us
   max diff: 7222 us
   mean diff: 1201 us

i.e., maximum 1.5% error, typically ~0.1% error with high resolution
timers. Max 7% error, typical 1% error with low resolution timers.

This seems reasonable.

[1] Stripped / pseudocode:

  clock_gettime(CLOCK_MONOTONIC, before);
  nanosleep({ .tv_nsec = usecs * 1000 }, NULL);
  clock_gettime(CLOCK_MONOTONIC, after);
  diff = abs((after - before) / 1000 - usecs));

Change-Id: Ifd4821c66c5564f7c975c08769a6742f645e9be0
Signed-off-by: Brian Norris <briannorris@chromium.org>
Reviewed-on: https://review.sourcearcade.org/c/flashprog/+/97
Reviewed-by: Arthur Heymans <arthur@aheymans.xyz>
Reviewed-by: Nico Huber <nico.h@gmx.de>
Tested-by: Nico Huber <nico.h@gmx.de>