stable-8 regression: time stands still

[Tom Evans]

On Wed, Mar 3, 2010 at 11:09 AM, Jeremy Chadwick
<freebsd at jdc.parodius.com> wrote:
> On Wed, Mar 03, 2010 at 09:09:56AM +0100, Ruben de Groot wrote:
>> > > I'm going to get this kern.timecounter sysctl from the system booted with a 7.x livecd
>> > > this evening. But really, I think this is a regression. Even if this system is the only
>> > > one known to be affected (which I sincerely doubt).
>> >
>> > You can force a timecounter choice by setting it in /boot/loader.conf.
>> > I would recommend choosing ACPI-safe on your system to see if that
>> > improves things:
>> >
>> > kern.timecounter.hardware=ACPI-safe
>>
>> Thanks, that solved my issue. Allthough for correctness I must add that this line
>> must be put in /etc/sysctl.conf; it does nothing in loader.conf.
>> Also the sysctl can be run in multiuser and immediately fixed the problem.
>
> Ah, I thought it was a loader tunable, not a sysctl tunable.  My
> apologies!  Yes, sysctl.conf is the correct place for this.
>
>> One question remains: why did this suddenly become a problem in FreeBSD 8 and not
>> in FreeBSD 7.
>
> That would be a question for the people who decided upon the timecounter
> priorities in FreeBSD, since they're hard-coded (meaning: why HPET is
> chosen above ACPI-safe).  This may have changed between RELENG_7 and
> RELENG_8; I simply don't know.  It would be easy to verify on your
> system of course (just boot a FreeBSD 7 LiveCD and provide the output
> from sysctl kern.timecounter).
>

Another possibility could be that RELENG_7 was finding ACPI-fast,
which is rated above HPET. I'm not sure why some machines get
ACPI-fast and some get ACPI-safe - I think ACPI-fast is tested at boot
up, and if the measurements are not consistent, it disables ACPI-fast
and uses ACPI-safe instead, which is then ignored when HPET is
available.

Cheers

Tom