svn commit: r354672 - head/lib/libc/secure

[Kyle Evans]

On Wed, Nov 13, 2019 at 12:23 PM Warner Losh <imp at bsdimp.com> wrote:
>
>
>
> On Wed, Nov 13, 2019 at 8:52 AM Pedro Giffuni <pfg at freebsd.org> wrote:
>>
>> Hi;
>>
>> On 12/11/2019 23:44, Warner Losh wrote:
>>
>>
>>
>> On Tue, Nov 12, 2019 at 9:20 PM Kyle Evans <kevans at freebsd.org> wrote:
>>>
>>>
>>>
>>> On Tue, Nov 12, 2019, 22:04 Pedro Giffuni <pfg at freebsd.org> wrote:
>>>>
>>>>
>>>> On 12/11/2019 22:00, Kyle Evans wrote:
>>>>
>>>> Author: kevans
>>>> Date: Wed Nov 13 03:00:32 2019
>>>> New Revision: 354672
>>>> URL: https://svnweb.freebsd.org/changeset/base/354672
>>>>
>>>> Log:
>>>>   ssp: rework the logic to use priority=200 on clang builds
>>>>
>>>>   The preproc logic was added at the last minute to appease GCC 4.2, and
>>>>   kevans@ did clearly not go back and double-check that the logic worked out
>>>>   for clang builds to use the new variant.
>>>>
>>>>   It turns out that clang defines __GNUC__ == 4. Flip it around and check
>>>>   __clang__ as well, leaving a note to remove it later.
>>>>
>>>>
>>>> clang reports itself as GCC 4.2, the priority argument was introduced in GCC 4.3.
>>>>
>>>>   Reported by: cem
>>>>
>>>> Modified:
>>>>   head/lib/libc/secure/stack_protector.c
>>>>
>>>> Modified: head/lib/libc/secure/stack_protector.c
>>>> ==============================================================================
>>>> --- head/lib/libc/secure/stack_protector.c Wed Nov 13 02:22:00 2019 (r354671)
>>>> +++ head/lib/libc/secure/stack_protector.c Wed Nov 13 03:00:32 2019 (r354672)
>>>> @@ -47,13 +47,15 @@ __FBSDID("$FreeBSD$");
>>>>   * they're either not usually statically linked or they simply don't do things
>>>>   * in constructors that would be adversely affected by their positioning with
>>>>   * respect to this initialization.
>>>> + *
>>>> + * This conditional should be removed when GCC 4.2 is removed.
>>>>   */
>>>> -#if defined(__GNUC__) && __GNUC__ <= 4
>>>> -#define _GUARD_SETUP_CTOR_ATTR \
>>>> -    __attribute__((__constructor__, __used__));
>>>> -#else
>>>> +#if defined(__clang__) || (defined(__GNUC__) && __GNUC__ > 4)
>>>>  #define _GUARD_SETUP_CTOR_ATTR \
>>>>      __attribute__((__constructor__ (200), __used__));
>>>> +#else
>>>> +#define _GUARD_SETUP_CTOR_ATTR \
>>>> +    __attribute__((__constructor__, __used__));
>>>>  #endif
>>>>
>>>>  extern int __sysctl(const int *name, u_int namelen, void *oldp,
>>>>
>>>> Please fix properly. Assuming clang always supported it, something like:
>>>>
>>>> #if __GNUC_PREREQ__(4, 3) || __has_attribute(__constructor__)
>>>>
>>>> should work
>>>>
>>>> Cheers,
>>>
>>>
>>> I considered something of this sort, but searching for information on the priority argument in the first place was annoying enough that I had too much search-fatigue to figure out when GCC actually corrected this, thus assuming that GCC5 (which seemed to be an all-around good release if memory serves) and later (since I confirmed GCC6) was sufficient.
>>>
>>> I'll fix it in the morning (~8 hours) if I receive no further objections to address.
>>
>>
>> Soon enough this can be removed entirely... Getting it pedantically right in the mean time has little value. We don't really support gcc5 at the moment. gcc6 and later have good support, but anything new between 4.3 and 6.0 likely is poorly tagged...
>>
>>
>> Well, tracking attributes on GCC versions is not easy but I did spend a good amount of time getting the attributes right on cdefs.h and while I lost the battle to get support for older GCC versions deprecated, getting the attributes properly defined in the GCC 4.2 vs clang vicinity is particularly important.
>
> Not really. We only support 4.2.1 + freebsd hacks and then 6.<something>. Further refining stuff is useless. Refining 4.3 vs 6.0 buys us nothing and distracts our limited resources getting correct something we are definitely removing from the tree in a couple of months. Going back and refining it gives us no practical benefit. While I don't object to the change, per se, I don't view it as required given our future plans.
>
> We should scrub cdefs.h. We've needed to for a while...
>>
>> I particularly dislike the idea of leaving notes of stuff that can be removed when an existing compiler is gone. In this case, we can fix this without adding more lines of code, and that also helps in case the code is MFCd.
>>
>> Now ... if you want to be pedantic: this code doesn't handle the case for non-GCC based compilers, and it probably could be done more generic and clean in cdefs.h where it can be reused. But I am not asking for that ;).
>
> I guess I disagree here. The current code is adequate and can be MFC'd.  It's not as perfect as it could be, but it's not wrong enough to fuss with.... If Kyle wants to, great, I'm not standing in the way, but I want to send the clear message that we don't need to get gcc 4.2 era stuff perfect because such distinctions are currently muddy at best. We won't work with a stock 4.2 compiler, and we already use 4.2 as a proxy for our current gcc compiler, not a perfect thing today anyway. Spending time on it doesn't give good value for the time spent on it, especially if we spend that time on other things that give better ROI.
>

I went ahead and fixed it because it was already agreed that
__has_attribute is a better test for clang than defined(__clang__), so
I might as well get the version correct since I'm already touching the
line and writing a commit message for it.

I do still think it should unifdef'd when GCC 4.2 goes away, so the
note still reflects my wishes. Given that this is a libc/*.c file, we
can handle any of the edge cases for non-clang/newer GCC compilers as
they crop up, if they crop up. We don't seem to put much effort (that
I've noticed) into making sure that libc can be compiled with other
compilers than those noted, but we can always make exception if
someone comes forward with a need for it.