[PATCH v2 3/3] aarch64: GCS: add clone3 tests for shadow stack

[Adhemerval Zanella Netto]

On 15/09/25 11:15, Yury Khrustalev wrote:
> * Adhemerval
> 
> On Mon, Sep 15, 2025 at 10:31:10AM -0300, Adhemerval Zanella Netto wrote:
>> On 15/09/25 08:37, Yury Khrustalev wrote:
>>> On Tue, Sep 09, 2025 at 11:31:53AM -0300, Adhemerval Zanella Netto wrote:
>>>>
>>>>> Parsing /proc/self/smaps requires much more code, I'll add it in the
>>>>> next version of the patch. After the thread started we'll have 2
>>>>> mappings with with the ss flag: parent and child. After cancellation we
>>>>> expect 1 mapping (the parent one).
>>>>
>>>> It depends if the kernel supports MADV_GUARD_INSTALL and if the pthread
>>>> requires a guard page.  I think to simplify you can disable the guard page
>>>> to always expect one mapping.
>>>
>>> Unless I'm missing something, I don't think guard page is involved here
>>> at all. We have parent and child shadow stacks after the thread started
>>> and only parent's shadow stack remains when the thread is cancelled.
>>
>> I meant to count the number of VMA as presented in /proc/self/smaps, assuming
>> the idea is to take a snapshot before and after pthread_create/pthread_join
>> to check the difference.
>>
>> If the underlying kernel supports MADV_GUARD_INSTALL we will not have the guard
>> page.
> 
> I've been trying to write something sensible to obtain information from
> the smaps file, however the fundamental issue is that this file is big
> and it changes frequently, so reading it line by line (or via any other
> chunk size), may still be unreliable, especially if we want to use it in
> a multi-threading context.

I think for testing it should not be a generic interface that cover all
user-cases, but one works reliable for the test you aim to use and with 
some comments about its limitation if someone wants to use on a different 
one.

> 
> To save time and for the purposes of this test, I'll just save the value
> of the shadow stack pointer in the thread before it's cancelled and then
> use to check that the relevant memory was indeed unmapped.