boost::process:: child::wait() on Linux ignores a zombie process

[theroyn]

Reproducible in the following example:

#include <boost/process.hpp>

int main()
{
        namespace bp = boost::process;

        std::string appCommand = "sleep 50";
        bp::child childProcess(appCommand);

        int pid = childProcess.id();

        printf("childProcess.running()=%d\n", childProcess.running());

        childProcess.terminate();
        childProcess.wait();

        printf("childProcess.running()=%d\n", childProcess.running());

        // kill 0 returns 0 if process is running
        bool isRunning = (::kill(pid, 0) == 0);

        printf("isRunning=%d\n", isRunning);
}

output on my machine is:

childProcess.running()=1
childProcess.running()=0
(kill 0 23764): isRunning=1

The program will print isRunning=1 at its end, meaning that the process is still running. Pausing at that point with debugger and checking ps -ax will also show that the process is a zombie.

After some investigation, faulty code seem to be in include/boost/process/detail/child_decl.hpp:

    bool _exited()
    {
        return _terminated || !::boost::process::detail::api::is_running(_exit_status->load());
    };

    void terminate(std::error_code & ec) noexcept
    {
        if (valid() && running(ec) && !ec)
            boost::process::detail::api::terminate(_child_handle, ec);

        if (!ec)
            _terminated = true;
    }

    void wait(std::error_code & ec) noexcept
    {
        if (!_exited() && valid())
        {
            int exit_code = 0;
            boost::process::detail::api::wait(_child_handle, exit_code, ec);
            if (!ec)
                _exit_status->store(exit_code);
        }
    }

terminate() marks _terminated = true, and wait() never calls its main logic because _exited() evaluates true.

A possible solution would be to perform a blocking wait at the end of boost::process::detail::api::terminate(), because ATM it waits with WNOHANG.

[SilverPlate3]

@klemens-morgenstern
Do you want me to create a pull request and fix this issue?

[klemens-morgenstern]

I think I fixed this, maybe add a test for it?