On Wed, Jun 15, 2011 at 03:01:59PM -0400, Konrad Rzeszutek Wilk wrote:
> Before this patch if we failed the transition (either p-state or
> voltage) we would still try to submit the "new" frequencies to cpufreq.
> That is incorrect - also we could submit a non-existing frequency value
> which would cause cpufreq to crash. The ultimate fix is in cpufreq
> to deal with incorrect values, but this patch improves the error
> recovery in the AMD powernowk8 driver.
>
> The failure that was reported was as follow:
>
> powernow-k8: Found 1 AMD Athlon(tm) 64 Processor 3700+ (1 cpu cores) (version 2.20.00)
> powernow-k8: fid 0x2 (1000 MHz), vid 0x12
> powernow-k8: fid 0xa (1800 MHz), vid 0xa
> powernow-k8: fid 0xc (2000 MHz), vid 0x8
> powernow-k8: fid 0xe (2200 MHz), vid 0x8
> Marking TSC unstable due to cpufreq changes
> powernow-k8: fid trans failed, fid 0x2, curr 0x0
> BUG: unable to handle kernel paging request at ffff880807e07b78
> IP: [<ffffffff81479163>] cpufreq_stats_update+0x46/0x5b
> ...
>
> And transition fails and data->currfid ends up with 0. Since
> the machine does not support 800Mhz value when the calculation is
> done ('find_khz_freq_from_fid(data->currfid);') it reports the
> new frequency as 800000 which is bogus. This patch fixes
> the issue during target setting.
>
> The patch however does not fix the issue in 'powernowk8_cpu_init'
> where the pol->cur can also be set with the 800000 value:
>
> pol->cur = find_khz_freq_from_fid(data->currfid);
> dprintk("policy current frequency %d kHz\n", pol->cur);
>
> /* min/max the cpu is capable of */
> if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
>
> The fix for that looks to update cpufreq_frequency_table_cpuinfo to
> check pol->cur.... but that would cause an regression in how the
> acpi-cpufreq driver works (it sets cpu->cur after calling
> cpufreq_frequency_table_cpuinfo). Instead the fix will be to let
> cpufreq gracefully handle bogus data.
>
> Reported-by: Tobias Diedrich <ranma+xen@xxxxxxxxxxxx>
> Tested-by: Tobias Diedrich <ranma+xen@xxxxxxxxxxxx>
> Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>
> ---
> arch/x86/kernel/cpu/cpufreq/powernow-k8.c | 7 ++++++-
> 1 files changed, 6 insertions(+), 1 deletions(-)
>
> diff --git a/arch/x86/kernel/cpu/cpufreq/powernow-k8.c b/arch/x86/kernel/cpu/cpufreq/powernow-k8.c
> index 2368e38..7e4a664 100644
> --- a/arch/x86/kernel/cpu/cpufreq/powernow-k8.c
> +++ b/arch/x86/kernel/cpu/cpufreq/powernow-k8.c
> @@ -1079,8 +1079,10 @@ static int transition_frequency_fidvid(struct powernow_k8_data *data,
> }
>
> res = transition_fid_vid(data, fid, vid);
> - freqs.new = find_khz_freq_from_fid(data->currfid);
> + if (res)
> + return res;
>
> + freqs.new = find_khz_freq_from_fid(data->currfid);
> for_each_cpu(i, data->available_cores) {
> freqs.cpu = i;
> cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
> @@ -1112,6 +1114,9 @@ static int transition_frequency_pstate(struct powernow_k8_data *data,
> }
>
> res = transition_pstate(data, pstate);
> + if (res)
> + return res;
That's wrong because transition_pstate() returns 0 unconditionally
(at least it does so on 3.0-rc3). But this change accidentally fixes
a different bug because res is used uninitialized, containing stack
garbage otherwise.
A proper fix should be to check against data->max_hw_pstate and
check whether the entry is not CPUFREQ_ENTRY_INVALID (look at
fill_powernow_table_pstate() for example).
I'm guessing this oops happens when powernow-k8 is loaded in the guest
and that the actual power management is done in the hypervisor?
Adding some more people to CC.
--
Regards/Gruss,
Boris.
--
To unsubscribe from this list: send the line "unsubscribe cpufreq" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
