Oddities in -current post-eventtimer

[Alexander Motin]

On 03.01.2011 12:28, Matthew D. Fuller wrote:
> OK, this has happened a couple times now.  I'm running a mid-Oct
> -CURRENT, and at around 25 days uptime (not exact but consistently in
> that vicinity), things start getting very choppy.  It's easily visible
> in playing videos; things get very jerky and slow, but all sorts of
> things start acting like they're happening in little chunks of time;
> keyboard repeats get very slow, things that often take notable time
> take much more, etc.  It's accompanied by a big spat of "calcru:
> runtime went backwards" messages (presumably just another symptom).
>
> The only fix I've found is to reboot, and then it's good for another
> 25ish days.  As a workaround, enabling kern.eventtimer.idletick sets
> things rightish.  A look at the interrupts turns up a hint; while
> vmstat says the overall average for cpu0 is just under 300/s, systat
> -vmstat shows that it's currnetly running around 20-some.  The other
> CPU's also settle at much lower levels.
>
> Another more tiring workaround is just slinging the mouse around real
> fast; that seems to hint to the system to keep checking stuff.
> Watching systat, that doesn't seem to bring the cpuX interrupt rate up
> very much, but the videos start playing smoothly.
>
> FreeBSD 9.0-CURRENT #0 r214107: Wed Oct 20 06:25:50 CDT 2010
> Quad-core running amd64.

Symptoms look very alike to ones fixed at r214597 on 2010-10-31:

Fix callout_tickstofirst() behavior after signed integer ticks overflow.
This should fix callout precision drop to 1/4s after 25 days of uptime
with HZ = 1000.

-- 
Alexander Motin