There is a functionality issue reported on some AMD platforms
http://lkml.indiana.edu/hypermail/linux/kernel/0912.2/01118.html
wherein, fdformat fails when HPET MSI based percpu timer is used.
We do not have the real root-cause for that problem. But, that
report exposed an issue with our current usage HPET MSI timers.
We use HPET MSI timers even on platforms that do not have
support for C2/C3 states. On those systems we should rather be
using LAPIC timers.
So, this series of patches does just that.
* Use LAPIC timer when there is always running APIC timer
* Use LAPIC timer on platforms that do not have support for deep C states
* Only use HPET MSI timers as percpu timers on systems that have LAPICs
that stop in deep C-states _and_ system supports deep C-states
The change turned out to be more than what I expected, due to
the current static nature of clockevent rating and unrelated
issues in acpi processor driver resume path. I also ended up
touching different subsystems to handle this.
If the patchset resolves the issue for Mark and if it looks sane
can one of the maintainers queue it up for .34
Reported-by: Mark Hounschell <[email protected]>
Signed-off-by: Venkatesh Pallipadi <[email protected]>
--
On 01/14/2010 08:39 PM, Venkatesh Pallipadi wrote:
> There is a functionality issue reported on some AMD platforms
> http://lkml.indiana.edu/hypermail/linux/kernel/0912.2/01118.html
> wherein, fdformat fails when HPET MSI based percpu timer is used.
>
> We do not have the real root-cause for that problem. But, that
> report exposed an issue with our current usage HPET MSI timers.
> We use HPET MSI timers even on platforms that do not have
> support for C2/C3 states. On those systems we should rather be
> using LAPIC timers.
>
> So, this series of patches does just that.
> * Use LAPIC timer when there is always running APIC timer
> * Use LAPIC timer on platforms that do not have support for deep C states
> * Only use HPET MSI timers as percpu timers on systems that have LAPICs
> that stop in deep C-states _and_ system supports deep C-states
>
> The change turned out to be more than what I expected, due to
> the current static nature of clockevent rating and unrelated
> issues in acpi processor driver resume path. I also ended up
> touching different subsystems to handle this.
>
> If the patchset resolves the issue for Mark and if it looks sane
> can one of the maintainers queue it up for .34
>
> Reported-by: Mark Hounschell <[email protected]>
> Signed-off-by: Venkatesh Pallipadi <[email protected]>
>
It does resolve my problem
Mark