Re: kern/53016: Clock not stable

[Tero Kivinen <kivinen%iki.fi@localhost>]

The following reply was made to PR kern/53016; it has been noted by GNATS.

From: Tero Kivinen <kivinen%iki.fi@localhost>
To: gnats-bugs%NetBSD.org@localhost
Cc: kern-bug-people%netbsd.org@localhost, gnats-admin%netbsd.org@localhost, netbsd-bugs%netbsd.org@localhost
Subject: Re: kern/53016: Clock not stable
Date: Fri, 6 Apr 2018 04:15:37 +0300

 Joerg Sonnenberger writes:
 >  Can you also run some background load, main() { for (;;) ; } or so in
 >  the background to rule out whether it is deep sleep? Enough instances,
 >  even with nice 20, to just keep all cores busy.
 
 When running some tests earlier, I noticed that if I run
 
 dnssec-signzone -N INCREMENT iki.fi
 
 in a loop, I can get the clock drift to happen more often. It is also
 easy to see that the drift happened, as the time command normally
 reports:
 
 u:11.19s s:0.01s r:3.37s 331%
 
 but in this case the user cpu time was still same (11 seconds), but
 real time was only 1.6 seconds instead of 3.4 seconds and cpu usage %
 was around 700%, even though the machine only has 4 cores...
 
 So I do not think this is actually going to the any kind of sleep
 state, as dnssec-signzone actually uses all cores with full 100% cpu.
 
 I tried to see if running busy loops (like you suggested with just
 for(;;); in them) could cause issues, but that didn't affect anything.
 Next I tried to busy loop doing fpu operations (for(d=0.0; d <
 9.0e20;d += 0.00001);) but no difference there, so it might that
 dnssec-signzone does some crypto operations or more complicated fpu
 operations that cause this issue to happen. 
 -- 
 kivinen%iki.fi@localhost