HAVE_POLL_FINE is always false when cross-compiling with CMake

[Kartatz]

I did this

I was having a weird issue when using the multi interface where the curl_multi_poll() call occasionally returns CURLM_UNRECOVERABLE_POLL.

The thing is, this issue was only reproducible in cross-compiled binaries. After a bit of investigation around Curl_poll(), I found that this is caused by the lack of the HAVE_POLL_FINE define in these builds.

Looking at CMake/OtherTests.cmake I see that there is indeed a check for HAVE_POLL_FINE inside a if (NOT CMAKE_CROSSCOMPILING) condition.

Removing the NOT CMAKE_CROSSCOMPILING check and replacing check_c_source_runs() with check_c_source_compiles() fixes this.

I expected the following

Properly detection of HAVE_POLL_FINE.

curl/libcurl version

libcurl/8.10.0-DEV BearSSL nghttp2/1.61.0

operating system

Linux rhel 3.10.0-1160.119.1.el7.tuxcare.els2.x86_64 #1 SMP Mon Jul 15 12:09:18 UTC 2024 x86_64 x86_64 x86_64 GNU/Linux

[bagder]

The check for this actually builds and runs code to verify poll() and that cannot be done for cross-compiles. Thus, it guesses.

Maybe this should rather guess on it working, if not using Apple OS.

[vszakats]

One option is to hard-wire known results for (cross-)compilation targets (like done for Apple).
One other is to explicitly pass -DHAVE_POLL_FINE=ON/OFF to CMake override the default for specific builds.

[vszakats]

In ./configure this detection is reserved for cross-builds as well.

[bagder]

Classically, you'd double-check the config.h file after configure has run so that it figured out things correctly. For cross-compiles especially, as it has more limited checks due to not being able to run code for checks.

So cmake/configure needs to use a use a default for such checks. Probably going with the safe one.

[Kartatz]

Maybe this should rather guess on it working, if not using Apple OS.

That seem a reasonable solution to me. I would say that there are more platforms with a working poll() than platforms with a broken one, but I think I should not speculate about this.

Either way, I think it would be cool to have an entry about this behavior in the Known Bugs section; something like "poll() cannot be detected properly in some environments" sounds fine.

[bagder]

it would be cool to have an entry about this behavior in the Known Bugs section

It's not really a bug, it's a known limitation. And this is not unique for curl, this is how cross-compiling works for just about everyone. There several more checks that cannot be done for cross-builds.

Ideally we should have these documented, but it might be tricky.

[vszakats]

Would it help to fall back to a non-runtime detection for cross-builds, like?:

#if <UNIX> && defined(_POSIX_C_SOURCE) && _POSIX_C_SOURCE >= 200112L
/* has native poll() */
#endif

[bagder]

Yeah, that's probably a decent guess.

[vszakats]

@Kartatz Does #14734 make poll() detected in your cross-build?

[Kartatz]

Does #14734 make poll() detected in your cross-build?

Yes, it's correctly detected. Thanks!