dragonfly: parent process does not receive POLLHUP events

[mikdusan]

In working towards porting to dragonfly, zig.exe fails right out the gate because it cannot fork/exec and capture stdout/stderr. The defacto probe to learn about host libc_installation such as cc -E -Wp,-v -xc /dev/null hangs.

This appears to be an upstream issue. The code converted to .c works for linux, macos, freebsd, netbsd, openbsd and not for dragonfly.

• upstream bug report filed: https://bugs.dragonflybsd.org/issues/3268
• test upstream patch
• upstream patch landed
• add comptime workaround for dragonfly POLLHUP in child_process.collectOutputPosix()

EDIT:
With merge of #8730 we have a comptime-workaround for dragonfly POLLHUP bug manifesting in collectOutputPosix() and that should be adequate. In the future we can revisit and add a runtime-version-detection of dragonfly to enable a workaround for dfly version > 6.0 . But the cost of this runtime-detection seems more expensive (still cheap though) than always using a workaround, so I'm backtracking because NOT-WORTH-IT and no longer find it necessary to have a versioned workaround.

[mikdusan]

• main.c is a C reduction

#include <errno.h>
#include <poll.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>

// invoke with something like: `main /bin/ls -al /`
// cc -E -Wp,-v -xc /dev/null

int main(int argc, char *argv[], char *envp[]) {
    int i = 0;
    for (int i = 0; i < argc; i += 1) {
        printf("[%d] %s\n", i, argv[i]);
    }

    if (argc < 2) {
        fprintf(stderr, "usage: %s exe [arg]...\n", argv[0]);
        exit(1);
    }

    int stdout_pipe[2];
    if (pipe(stdout_pipe) == -1) {
        fprintf(stderr, "pipe(stdout_pipe) failed: %s\n", strerror(errno));
        exit(1);
    }

    int stderr_pipe[2];
    if (pipe(stderr_pipe) == -1) {
        fprintf(stderr, "pipe(stderr_pipe) failed: %s\n", strerror(errno));
        exit(1);
    }

    printf("pipes: %d,%d,%d,%d\n", stdout_pipe[0], stdout_pipe[1], stderr_pipe[0], stderr_pipe[1]);

    pid_t child_pid = fork();
    if (child_pid == -1) {
        fprintf(stderr, "fork failed: %s\n", strerror(errno));
        exit(1);
    }

    // child
    if (child_pid == 0) {
        if (dup2(stdout_pipe[1], STDOUT_FILENO) == -1) {
            fprintf(stderr, "dup2(STDOUT_FILENO) failed: %s\n", strerror(errno));
            exit(1);
        }
        if (dup2(stderr_pipe[1], STDERR_FILENO) == -1) {
            fprintf(stderr, "dup2(STDOUT_FILENO) failed: %s\n", strerror(errno));
            exit(1);
        }

        close(stdout_pipe[0]);
        close(stdout_pipe[1]);
        close(stderr_pipe[0]);
        close(stderr_pipe[1]);

        int rv = execve(argv[1], &argv[1], envp);
        if (rv == -1) {
            fprintf(stderr, "execve failed: %s\n", strerror(errno));
            exit(1);
        }
    }

    // parent

    close(stdout_pipe[1]);
    close(stderr_pipe[1]);

    struct pollfd fds[2];
    fds[0].fd = stdout_pipe[0];
    fds[0].events = POLLIN;
    fds[1].fd = stderr_pipe[0];
    fds[1].events = POLLIN;

    int nactive = 2;
    while (nactive > 0) {
        int rv = poll(fds, 2, -1);
        if (rv == -1) {
            fprintf(stderr, "poll failed: %s\n", strerror(errno));
            exit(1);
        }
        printf("poll rv=%d\n", rv);
        if (rv == 0) continue;

        printf("fd[0].revents=0x%04x [1].revents=0x%04x\n", fds[0].revents, fds[1].revents);
        char buf[8192];
        
        if (fds[0].revents & POLLIN) {
            memset(buf, 0, sizeof(buf));
            ssize_t nbytes = read(fds[0].fd, buf, sizeof(buf) - 1);
            printf("fd[0] read=%ld bytes\n", nbytes);
        }

        if (fds[1].revents & POLLIN) {
            memset(buf, 0, sizeof(buf));
            ssize_t nbytes = read(fds[1].fd, buf, sizeof(buf) - 1);
            printf("fd[1] read=%ld bytes\n", nbytes);
        }

        if (fds[0].revents & (POLLERR | POLLHUP | POLLNVAL)) {
            fds[0].fd = -1;
            nactive -= 1;
        }
        if (fds[1].revents & (POLLERR | POLLHUP | POLLNVAL)) {
            fds[1].fd = -1;
            nactive -= 1;
        }
    }

    int status;
    int rv = waitpid(child_pid, &status, 0);
    if (rv == -1) {
        fprintf(stderr, "waitpid failed: %s\n", strerror(errno));
        exit(1);
    }

    printf("child %d exited with status %d\n", child_pid, status);

    return 0;
}

• output from dragonfly: ./main /bin/echo hello

[0] ./main
[1] /bin/echo
[2] hello
pipes: 3,4,5,6
poll rv=1
fd[0].revents=0x0001 [1].revents=0x0000
fd[0] read=6 bytes
poll rv=2
fd[0].revents=0x0001 [1].revents=0x0001
fd[0] read=0 bytes
fd[1] read=0 bytes
poll rv=2
fd[0].revents=0x0001 [1].revents=0x0001
fd[0] read=0 bytes
fd[1] read=0 bytes
poll rv=2
.... infinite loop

• output from openbsd: ./main /bin/echo hello

[0] ./main
[1] /bin/echo
[2] hello
pipes: 3,4,5,6
poll rv=1
fd[0].revents=0x0001 [1].revents=0x0000
fd[0] read=6 bytes
poll rv=2
fd[0].revents=0x0011 [1].revents=0x0011
fd[0] read=0 bytes
fd[1] read=0 bytes
child 72887 exited with status 0

[mikdusan]

[mikdusan]

Closing. We're in a good place now. See edit in original-post.