[uclibc-ng-devel] Re: Revert request and discussion about "librt: avoid compilation error"

Hello uClibc-ng hackers, Damien, Rich, I am sending this email to discuss the possibility to revert https://cgit.uclibc-ng.org/cgi/cgit/uclibc-ng.git/commit/?id=08d46f1ce21e4e… I think this change is wrong, I have been tracking an issue for a few days and I am pretty sure the issue comes from this commit. Basically what happens is that with this commit, tst-cancel18 test is failing. The test is available over there: https://github.com/wbx-github/uclibc-ng-test/blob/master/test/nptl/tst-canc… The reason for this fail is that clock_nanosleep code from libc is badly compiled: https://elixir.bootlin.com/uclibc-ng/latest/source/librt/clock_nanosleep.c#… Because SINGLE_THREAD_P is actually defined, in case your libc build supports NPTL threads (which is, I think, the majority of configs, except very old systems using LINUXTHREADS or NOMMU maybe). In my case (kvx arch): #define SINGLE_THREAD_P __builtin_expect (THREAD_GETMEM (THREAD_SELF, header.multiple_threads) == 0, 1) See: https://elixir.bootlin.com/uclibc-ng/latest/source/libpthread/nptl/sysdeps/… This results to clock_nanosleep code being compiled to a very simple version that directly calls the syscall and does not change at all the thread cancellability to asynchronous before calling the syscall. Therefore, when the test main thread calls pthread_cancel(), no signal is sent (see https://elixir.bootlin.com/uclibc-ng/latest/source/libpthread/nptl/pthread_…) and therefore the nanosleep syscall is never interrupted and the test fails. clock_nanosleep is defined to be a mandatory cancellation point by POSIX, it must be interruptible in all cases by pthread_cancel, whether in synchronous or asynchronous mode. I think the fix for the NOMMU case should be different, maybe checking for if LIBC_CANCEL_ASYNC is defined ? Damien do you have time to work on a new fix for the NOMMU case? Are you OK with the revert?