dhcpcd-discuss

On 02/10/2017 20:13, M. Buecher wrote:
> On 2017-10-02 12:12, Roy Marples wrote:
>> On 02/10/2017 10:39, M. Buecher wrote:
>>> On 2017-10-01 23:26, Roy Marples wrote:
>>>> On 01/10/2017 20:48, M. Buecher wrote:
>>>>> It builds except for one error:
>>>>> ../../src/eloop.o: In function `eloop_event_add_rw':
>>>>> ./src/eloop.c:353: undefined reference to `reallocarray'
>>>>> collect2: error: ld returned 1 exit status
>>>>> Makefile:39: recipe for target 'eloop-bench' failed
>>>>>
>>>>> Don't know if this a source or Raspbian/Debian issue.
>>>>> Anybody got a clue?
>>>>
>>>> I can replicate it on my Sparky Linux image.
>>>> But it's tempermental, I can't always replicate it.
>>>>
>>>> I *think* it's a bug with GNU make - I cannot replicate it on any BSD
>>>> platform but then reallocarray is also present so that particular
>>>> issue would never be hit either.
>>>>
>>>> When it happens, I was able to build simply by touching src/eloop.c.
>>>> This changes the datetime, which make does use, implying it's a make
>>>> issue, or maybe an issue with my Makefile, but I don't understand what
>>>> or how.
>>>>
>>>> The eloop_bench isn't really a test, it just measures performance of
>>>> the chosen polling mechanism. I wrote it to highlight an issue with
>>>> the kqueue implementation on NetBSD compared to other BSDs. Linux of
>>>> course uses epoll.
>>>>
>>>> Roy
>>>
>>> Just touching src/eloop.c didn't work for Raspbian/Debian and debuild.
>>>
>>> From what I have seen so far my best guess is that
>>> "compat/reallocarray.o" is not available for linking. Just guessing as
>>> in the related call I cannot see that object referenced.
>>>
>>> cc  -Wl,-z,relro -Wl,-z,now -o eloop-bench eloop-bench.o
>>> ../../src/eloop.o
>>
>> It's not used.
>> CFLAGS sets -DNO_CONFIG_H so eloop will use it's internal version.
>>
>> Is something else setting HAVE_REALLOCARRAY maybe?
>>
>> Roy
> 
> On Raspbian `configure` doesn't find reallocarray():
> Configuring dhcpcd for ... linux
> Testing for reallocarray ... no
> 
> Therefore the other build and link commands for dhcpcd and the crypt
> test use "config.h" and "compat/reallocarray.o".
> This is why "src/eloop.o" references an external reallocarray() function
> (from "compat/reallocarray.o").
> 
> The CFLAG -DNO_CONFIG_H has only effect when compiling
> "tests/eloop-bench/eloop-bench.c".
> When linking with "src/eloop.o" it still needs an external
> reallocarray() implementation.
> 
> Either "src/eloop.c" is build again into another directory with CFLAG
> -DNO_CONFIG_H and then this object is linked with
> "tests/eloop-bench/eloop-bench.o".
> Or "tests/eloop-bench/eloop-bench.o" provides a reallocarray()
> implementation for "src/eloop.o".
> Haven't checked the source yet, only the build log.

The latter is the easiest fix.
https://roy.marples.name/git/dhcpcd.git/commit/?id=9f23d8c75a9cf46e0c5ce46f0c4a5828e5312404

To have a seperate build directory (which would be better) I'm not sure
is possible while being agnostic to GNU make and BSD make.

Roy