Re: [PATCH] RFC: selinux avc trace

[peter enderborg]

On 7/30/20 9:29 PM, Steven Rostedt wrote:
> On Thu, 30 Jul 2020 21:12:39 +0200
> peter enderborg <peter.enderborg@xxxxxxxx> wrote:
>
>>>> avc: denied { find } for interface=vendor.qti.hardware.perf::IPerf sid=u:r:permissioncontroller_app:s0:c230,c256,c512,c768 pid=9164 scontext=u:r:permissioncontroller_app:s0:c230,c256,c512,c768 tcontext=u:object_r:vendor_hal_perf_hwservice:s0 tclass=hwservice_manager permissive=0
>>>> Âavc: denied { execute } for pid=13914 comm="ScionFrontendAp" path="/data/user_de/0/com.google.android.gms/app_chimera/m/00000002/oat/arm64/DynamiteLoader.odex" dev="sda77" ino=204967 scontext=u:r:platform_app:s0:c512,c768 tcontext=u:object_r:privapp_data_file:s0:c512,c768 tclass=file permissive=0 ppid=788 pcomm="main" pgid=13914 pgcomm="on.updatecenter"
>>>>
>>>> It omit the fields that are not used. Some parts are common some are not. So a correct format specification for trace will be problematic if there is no "optional" field indicator.  
>>> That's all quite noisy. What is the object of these changes? What
>>> exactly are you trying to trace and why?  
>> It is noisy, and it have to be. it covers a lot of different areas. One common problem is
>> to debug userspace applications regarding violations. You get the violation from the logs
>> and try to figure out what you did to cause it. With a trace point you can do much better
>> when combine with other traces. Having a the userspace stack is a very good way,
>> unfortunately it does not work on that many architectures within trace.
>>
>> What exactly are you doing with any trace? You collect data to analyse what's
>> going on. This is not different. Selinux do a specific thing, but is has lots of parameters.
> Have you thought of adding multiple trace events with if statements
> around them to decode each specific type of event?

Yes. And I think class is good split point. But I think it will require
a few layers, but a is mostly data driven so I think it might be hard
to do it compile time. I think a hybrid might be possible,
but it then we need some ugly part with a other separator than =,
or some escape seq to separate.

sort of "generc1=X generic2=Y variable1^x variable2^y" or

"generc1=X generic2=Y leftover=[variable1=x variable2=y]"

If there was a formal parameter tree we could maybe do some
generated printer. I don't think there are one, maybe Paul Moore or Stephen Smalley
can verify that.

Â

> Note, you can have a generic event that gets enabled by all the other
> events via the "reg" and "unreg" part of TRACE_EVENT_FN(). Say its
> called trace_avc, make a dummy trace_avc() call hat doesn't even need
> to be called anywhere, it just needs to exist to get to the other trace
> events.
>
> Then have:
>
> 	if (trace_avc_enabled()) {
> 		if (event1)
> 			trace_avc_req_event1();
> 		if (event2)
> 			trace_avc_req_event2();
> 		[..]
> 	}
>
> The reason for the trace_avc_enabled() is because that's a static
> branch, which is a nop when not enabled. When enabled, it is a jump to
> the out of band if condition block that has all the other trace events.
>
> -- Steve