Re: [PATCH v1 0/2] perf: Drop leaked kernel samples

["Jin, Yao" <yao.jin@linux.intel.com>]

On 6/15/2018 11:35 AM, Kyle Huey wrote:
> I strongly object to this patch as written. As I said when I
> originally reported[0] the regression introduced by the previous
> version of this patch a year ago.
> 
> "It seems like this change should, at a bare minimum, be limited to
> counters that actually perform sampling of register state when the
> interrupt fires.  In our case, with the retired conditional branches
> counter restricted to counting userspace events only, it makes no
> difference that the PMU interrupt happened to be delivered in the
> kernel."
> 
> This means identifying which values of `perf_event_attr::sample_type`
> are security concerns (presumably PERF_SAMPLE_IP is, and
> PERF_SAMPLE_TIME is not, and someone needs to go through and decide on
> all of them) and filtering on those values for this new behavior.
> 
> And because rr sets its sample_type to 0, once you do that, the sysctl
> will not be necessary.
> 
> - Kyle
> 

Since rr sets sample_type to 0, the easiest way is to add checking like:

if (event->attr.sample_type) {
	if (event->attr.exclude_kernel && !user_mode(regs))
		return false;
}

So the rr doesn't need to be changed and for other use cases the leaked 
kernel samples will be dropped.

But I don't like this is because:

1. It's too specific for rr case.

2. If we create a new sample_type, e.g. PERF_SAMPLE_ALLOW_LEAKAGE, the 
code will be:

if !(event->attr.sample_type & PERF_SAMPLE_ALLOW_LEAKAGE) {
	if (event->attr.exclude_kernel && !user_mode(regs))
		return false;
}

But rr needs to add PERF_SAMPLE_ALLOW_LEAKAGE to sample_type since by 
default the bit is not set.

3. Sysctl is a more flexible way. It provides us with an option when we 
want to see if skid is existing, we can use sysctl to turn on that.

Thanks
Jin Yao