time: use precise callout for clock_nanosleep(2) and nanosleep(2)

[glebius]

Summary

Don't apply tc_precexp (a derivative of kern.timecounter.alloweddeviation)
to sleep callout when processing the default and precise clocks. The
default timer deviation of 5% is our internal optimization in the kernel,
and we shouldn't leak that into the POSIX APIs. Note that application
doesn't have any control to cancel the deviation, only a superuser can
change the global tunable [with side effects].

Leave the deviation for CLOCK_*_FAST and CLOCK_SECOND that are documented
as imprecise.

Note: although this change cancels the gargantuan 5% deviation, some sleep
precision tests would still provide better results on a system with
kern.timecounter.alloweddeviation=0. This could be explained by other
events aggregated into buckets of our precise events, producing some
effect.

[imp]

This looks good to me.

sys/kern/kern_time.c

523

need a fallthrough marking?

538

here too?

[kib]

I would add a knob to allow user to restore current behavior. Might be, *_PRECISE must be always precise indeed, and CLOCK_REALTIME/MONOTONIC/UPTIME managed by this knob.

[glebius]

Also eliminate TIMESEL() for precise APIs. Now that becomes truly precise.

From mav@.

[glebius]

• Initialize sbt properly (fixup of last update).
• Add sysctl to switch to the old behavior.
• Improve documentation.

[ziaee]

Apropos is showing me all the clocks throughout the manual are unmarked or Dvs (including other places on this page), so we should keep them together I think.

lib/libsys/nanosleep.2

105

It should probably be another commit, and I'm happy to do it if you want, but these are Dvs.

129

131

137

140

142

148

248

These need to be on their own lines to register correctly. But they aren't Va, in the rest of the manual they're Dv. You can check with apropos Va=CLOCK_MONOTONIC and try different mdoc macros.

[glebius]

Use .Dv for clocks in the manual page.

[ziaee]

lib/libsys/nanosleep.2

140

This still registers as a single defined variable in the whatis database called "CLOCK_MONTONIC, CLOCK_REALTIME". Nit, oxford comma

[glebius]

One more place where to values were on same line.

[ziaee]

Accept as manpages, sorry for the nits.