Re: [PATCH] perf: Check all MSRs before passing hw check

[George Dunlap <george.dunlap@eu.citrix.com>]

On 18/03/13 08:42, Ingo Molnar wrote:
> * George Dunlap <george.dunlap@eu.citrix.com> wrote:
>
>> check_hw_exists has a number of checks which go to two exit paths:
>> msr_fail and bios_fail.  Checks classified as msr_fail will cause
>> check_hw_exists() to return false, causing the PMU not to be used;
>> bios_fail checks will only cause a warning to be printed, but will
>> return true.
>>
>> The problem is that if there are both msr failures and bios failures,
>> and the routine hits a bios_fail check first, it will exit early and
>> return true, not finishing the rest of the msr checks.  If those msrs
>> are in fact broken, it will cause them to be used erroneously.
>>
>> This changset causes check_hw_exists() to go through all of the msr
>> checks, failing and returning false if any of them fail.
>>
>> This problem affects kernels as far back as 3.2, and should thus be
>> considered for backport.
>>
>> Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
>> CC: Konrad Wilk <konrad.wilk@oracle.com>
>> CC: Thomas Gleixner <tglx@linutronix.de>
>> CC: "H. Peter Anvin" <hpa@zytor.com>
>> CC: x86@kernel.org
>> ---
>>   arch/x86/kernel/cpu/perf_event.c |   20 ++++++++++----------
>>   1 file changed, 10 insertions(+), 10 deletions(-)
> What is missing is a description of what specific platform this gets
> triggered on and exactly why. Is some hw feature emulation missing that
> causes the check to fail?

Remember, there are two checks failing: the second one is supposed to 
fail and disable the PMU entirely, but it's not getting there because 
when the first one fails, it skips the rest but returns "success" anyway.

The warning on the first check is as follows:

[    8.131985] Performance Events: Broken BIOS detected, complain to 
your hardware vendor.^M
[    8.139997] [Firmware Bug]: the BIOS has corrupted hw-PMU resources 
(MSR c0010000 is 530076)^M

c0010000 is the AMD  MSR_K7_EVNTSEL0; the check it's failing is:
   if (val & ARCH_PERFMON_EVENTSEL_ENABLE)

So it discovers that one of the performance counters is already enabled 
-- worth a warning, but by itself not worth disabling the PMU.  This is 
most likely to be exactly what the warning message says: a buggy BIOS 
that enables perfcounters enabled for some reason.

The second check is supposed to detect that the PMU is actually not 
usable -- in my case because it's running virtualized (under Xen).

  -George