Re: x86_32 tsc/pit and hrtimers

[Chris Snook <csnook@redhat.com>]

Jeff Hansen wrote:
> Linus, Ingo, All,
> 
> I've been struggling with hrtimer support in 2.6.26.5 on an older 
> x86_32/i386 system, and I'm wondering if there are any easy fixes that you
> (or anyone else) would suggest.
> 
> Basically, this system does not print out the message:
> 
> "Switched to high resolution mode on CPU 0"
> 
> indicating that one-shot, hrtimers, etc. won't work, since high resolution
> mode has not been enabled.  I've verified that hrtimers started with
> hrtimer_start do not have the expected resolution further than 1/HZ.
> 
> This system does not have LAPIC, ACPI, or HPET, so really the only
> clocksources I can use are TSC and PIT.  This should be fine (in theory,
> unless it wasn't designed like that), but apparently the clocksource flags
> are not initialized in such a way that one of them ever gets marked as
> CLOCK_SOURCE_VALID_FOR_HRES.
> 
> The flow of the flags on each of these clocksources is as follows:
> 
> 1) The flags on the TSC clocksource are CLOCK_SOURCE_IS_CONTINUOUS |
>     CLOCK_SOURCE_MUST_VERIFY, which causes PIT to be used as the watchdog
>     clocksource. (see kernel/time/clocksource.c:~171)
> 2) Around line 122 in kernel/time/clocksource.c, where most clocksources'
>     flags usually get ORed with CLOCK_SOURCE_VALID_FOR_HRES, the PIT's do
>     not because it is not CLOCK_SOURCE_IS_CONTINUOUS, and the TSC's do not
>     also because the PIT (as the watchdog) is not
>     CLOCK_SOURCE_IS_CONTINUOUS.
> 
> I get the same results on a new laptop booting into 32-bit Linux with hpet
> and acpi disabled.
> 
> Can you please tell me if this is supposed to work, and I just have a
> poorly configured kernel; or if TSC/PIT drivers were not designed to work
> this way in the first place.  If it wasn't designed to do this, do you
> have any tips on implementing this, since I'll be needing to do that?
> 
> -Jeff Hansen

This is not supposed to work, but it might be worthwhile to add a boot option to 
force the kernel to trust the TSC, as hardware that lacks any high-res timers 
also tends to be primitive enough that the TSC can be trusted, if it exists.  If 
you patch out the CLOCK_SOURCE_MUST_VERIFY flag on the TSC, do you get 
correctly-functioning high-res timers on this system?

-- Chris