Re: [RFC PATCH v2 1/2] x86/perf/amd: Resolve race condition when disabling PMC

["Lendacky, Thomas" <Thomas.Lendacky@amd.com>]

On 3/18/19 4:37 AM, Peter Zijlstra wrote:
> On Fri, Mar 15, 2019 at 08:40:53PM +0000, Lendacky, Thomas wrote:
>> +void amd_pmu_disable_all(void)
>> +{
>> +	unsigned long overflow_check[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
>> +	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
>> +	int idx;
>> +
>> +	bitmap_zero(overflow_check, X86_PMC_IDX_MAX);
>> +
>> +	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
>> +		u64 val;
>> +
>> +		if (!test_bit(idx, cpuc->active_mask))
>> +			continue;
>> +
>> +		rdmsrl(x86_pmu_config_addr(idx), val);
>> +		if (!(val & ARCH_PERFMON_EVENTSEL_ENABLE))
>> +			continue;
>> +
>> +		val &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
>> +		wrmsrl(x86_pmu_config_addr(idx), val);
>> +
>> +		/*
>> +		 * If the interrupt is enabled, this counter must be checked
>> +		 * for an overflow condition to avoid possibly changing the
>> +		 * counter value before the NMI handler runs.
>> +		 */
>> +		if (val & ARCH_PERFMON_EVENTSEL_INT)
>> +			__set_bit(idx, overflow_check);
>> +	}
> 
> I think you can ditch overflow_check and directly call
> x86_pmu_disable_all() here.
> 
>> +
>> +	/*
>> +	 * This shouldn't be called from NMI context, but add a safeguard here
>> +	 * to return, since if we're in NMI context we can't wait for an NMI
>> +	 * to reset an overflowed counter value.
>> +	 */
>> +	if (in_nmi())
>> +		return;
>> +
>> +	/*
>> +	 * Check each counter for overflow and wait for it to be reset by the
>> +	 * NMI if it has overflowed.
>> +	 */
>> +	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
>> +		if (!test_bit(idx, overflow_check))
> 
> And simply iterate cpuc->active_mask again here.
> 
>> +			continue;
>> +
>> +		amd_pmu_wait_on_overflow(idx);
>> +	}
>> +}
> 
> Because, per x86_pmu_hw_config() we _always_ have EVENTSEL_INT set, even
> for !sampling events -- such that we can deal with overflow, and we
> should 'always' have EVENTSEL_ENABLE set when 'active', see
> x86_pmu_enable_all().

That would make it much simpler. I will have to rely on those assumptions,
though, which I'm fine with.

I'm traveling this week so it might be a few days before I can get the
next version out.

Thanks,
Tom

> 
>