Re: [patch V4 part 4 02/24] x86/int3: Avoid atomic instrumentation

[Andy Lutomirski]

> On May 14, 2020, at 6:15 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> ïOn Thu, May 14, 2020 at 02:51:32PM +0200, Thomas Gleixner wrote:
>> Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:
>>> On Wed, May 13, 2020 at 09:57:52PM -0700, Andy Lutomirski wrote:
>>>>> On Tue, May 5, 2020 at 7:15 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>>>>>> 
>>>>>> From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>>>>>> 
>>>>>> Use arch_atomic_*() and READ_ONCE_NOCHECK() to ensure nothing untoward
>>>>>> creeps in and ruins things.
>>>>>> 
>>>>>> That is; this is the INT3 text poke handler, strictly limit the code
>>>>>> that runs in it, lest it inadvertenly hits yet another INT3.
>>>>> 
>>>>> 
>>>>> Acked-by: Andy Lutomirski <luto@xxxxxxxxxx>
>>>>> 
>>>>> Does objtool catch this error?
>>> 
>>> It does not. I'll put it on the (endless) todo list..
>> 
>> Well, at least it detects when that code calls out into something which
>> is not in the non-instrumentable section.
> 
> True, but the more specific problem is that noinstr code can use
> jump_label/static_call just fine.
> 
> So a more specific test is validating none of that happens in the INT3
> handler before poke_int3_handler(). Which is what I think Andy was
> after.

Exactly.  I admit that sleep-deprived Andy was actually thinking âtglx and/or PeterZ found this by inspection, and somewhere it escaped objtoolâs notice,â which is sort of the same thing :)