Re: [PATCH] selftests/bpf: Enforce definition order for __msg macros in BPF tests

[Cupertino Miranda <cupertino.miranda@oracle.com>]

Hi Eduard,

I might be mistaken, but as it is skip_dynamic_pfx is dropping the index 
from the string that is passed to push_msg.
This is the case since it is matching the '=' in its passed macros;

   #define TEST_TAG_EXPECT_MSG_PFX "comment:test_expect_msg="

dropping the index and the other equal in the call to strchr(msg, '=').

I know the patch is rather bigger then perhaps minimally necessary but I 
did not want to change the semantics of current implementation, except 
the array ordering.

Did not really thought about string ordering it later, however, if I 
think about it, I would personally prefer to order the array at the 
push. Ordering later during a processing stage IMHO would obfuscate the 
fact that the array is ordered by its index.

If we change skip_dynamic_pfx to also return the "<index>=" then we 
could parse the string and insert it ordered in push_msg, not really 
needing to wait for a later stage.

Anyway, up to you. If you think it is worth it, I can do that.

Cheers,
Cupertino

On 18-03-2026 11:07 PM, Eduard Zingerman wrote:
> On Thu, 2026-03-05 at 13:00 +0000, Cupertino Miranda wrote:
>> Ensure that __msg macros, and similars, used for BPF test validation are
>> processed in the order they are defined in the source code.
>>
>> Neither GCC nor the C standard defines the order in which function
>> attributes are consumed. While Clang tends to preserve definition order,
>> GCC may process them out of sequence. This inconsistency causes BPF
>> tests with multiple __msg entries to fail when compiled with GCC.
>>
>> This patch fixes the test runner to explicitly sort or process these
>> attributes to guarantee predictable matching behavior across different
>> compilers.
>>
>> Signed-off-by: Cupertino Miranda <cupertino.miranda@oracle.com>
>> Cc: David Faust  <david.faust@oracle.com>
>> Cc: Jose Marchesi  <jose.marchesi@oracle.com>
>> Cc: Elena Zannoni  <elena.zannoni@oracle.com>
>> ---
> 
> Looks like this patch fell through the cracks.
> 
> It looks a bit intrusive, given the amount of places index has to be
> passed through. An alternative would be to modify parse_test_spec()
> such that it collects decl tag ids into the array first and then sorts
> the array. It will have to use strverscmp() for comparator, though.
> This way all the changes would be localized to a single function
> collecting the array.
> 
> Up to you if that is worth the hassle.
> 
> Reviewed-by: Eduard Zingerman <eddyz87@gmail.com>
> 
> [...]