[PATCH v4] linux: Add openat2 (BZ 31664)

[Paul Eggert]

On 8/26/25 07:44, Florian Weimer wrote:
> * Paul Eggert:
>> Can you give a sample call that would break?
> 
> Starting with GCC 14, the last two lines are reported as containing
> errors:
> 
> void f(char *);
> void fc(const char *);
> 
> void (*pf)(const char *) = f;
> void (*pfc)(char *) = fc;

Sure, but that's not a sample call as I requested. It's a contrived 
example that is not likely to occur in real code. (And if there is true 
concern about this instead of language-lawyering, I'm sure there are 
workarounds.)

> I don't think sigaction is a good example because it is not extensible.

It is extensible, just perhaps not in the way one might like because one 
cannot anticipate all extensions. openat2 is similar.

> POSIX examples for extensible interfaces are posix_spawn and
> pthread_attr_t, where each attribute ends up with two functions.  That
> approach isn't great, either.

Agreed that there is no perfection. However, there are real advantages 
to having two functions, one for getting and one for setting. That is 
why the getter/setter approach has so many instances already. openat2 
should follow existing practice and not try to invent a new one (for no 
particularly strong reason).

> If we don't pass through the system call interface unchanged, past
> experience tells us that this causes problems in the future.

For this particular case there will be problems no matter what we do. 
That's clear from the Linux man pages, which give two disagreeing APIs 
for openat2. In cases like these, it's often better to do the right 
thing even if it's a bit more work, than to do the wrong thing.