Re: [PATCH V3 5/9] objtool: Add return address unwind hints

[Alexandre Chartre <alexandre.chartre@oracle.com>]

On 4/14/20 8:42 PM, Peter Zijlstra wrote:
> On Tue, Apr 14, 2020 at 08:31:23PM +0200, Alexandre Chartre wrote:
>> On 4/14/20 7:56 PM, Peter Zijlstra wrote:
> 
>>> So what actual problem is it solving?
>>>
>>
>> The return stack stuff is here to correctly handle intra-function call so that
>> we can figure out where the ret of an intra-function call should return. We
>> don't have this challenge with regular functions because we know that a ret
>> inside such function just indicates the end of the function.
>>
>> But when there's an intra-function call, a ret instruction can either:
>>    - continue after the intra-function call (if the stack was unchanged)
>>    - jump somewhere else (if the return address was changed) and eventually
>>      return to the next return address
>>    - indicate the end of the function (if the return address was removed).
>>
>> So, all this is needed to correctly follow the flow of the code and properly
>> record stack changes.
> 
> But which intra-function calls are you worried about here? The RSB
> stuffing ones we have to explicitly forget and the retpoline ones we
> can't follow because they're indirect calls.
> 
> So again, who cares about that stack?
> 

This provides a generic code to handle any intra-function call. Currently we have
the RSB stuffing ones which are forgotten with the UNWIND_HINT_TYPE_RADDR_DELETE
directive. And for retpoline, they will not return if we have an indirect jump
(JMP_NOSPEC) but they will return if we have an indirect call (CALL_NOSPEC). The
code can handle both cases. For example, if we were to have a CALL_NOSPEC invocation
which is not in an alternative then objtool can now correctly handle it.

alex.