Created attachment 598585 [details] reproducer Description of problem: When used with WNOWAIT, waitpid returns EINVAL error. This breaks ksh buildtime tests. Version-Release number of selected component (if applicable): glibc-2.15-48.fc17 How reproducible: always Steps to Reproduce: 1. gcc reproducer.c -o reproducer 2. ./reproducer 3. Actual results: ./reproducer WNOWAIT waitpid returns -1, status=32767, errno=22 : Invalid argument WNOHANG waitpid returns <PID>, status=0, errno=0 : Success 0 waitpid returns -1, status=0, errno=10 : No child processes Expected results: ./reproducer WNOWAIT waitpid returns <PID>, status=0, errno=0 : Success WNOHANG waitpid returns <PID>, status=0, errno=0 : Success 0 waitpid returns -1, status=0, errno=10 : No child processes Additional info:
This is reall a kernel issue -- glibc merely provides an interface into the kernel for system calls such as wait, waitpid, etc. It doesn't munge, interpret or otherwise care about the return value or the options for this family of system calls.
curious. As the vfork won't return until after the child has exited, I would have expected the waitpid's to be returning ECHILD in all cases, as the pid wouldn't be valid at that point. I recommend bringing this up on linux-kernel.
As far as I can see, waitpid (and wait4) don't allow WNOWAIT. That is probably why it is returning -ENIVAL, because that is not a valid option for that system call. The options that are allowed for waitpid are WNOHANG, WUNTRACED, and WCONTINUED as documented in the man page. There are Linux specific options that can be used if the children were created with clone, and those are __WCLONE, __WALL, and __WNOTHREAD. The waitid call does accept WNOWAIT, but that is not the same thing as waitpid. I'm not seeing how this is a bug.