Qemu nixpkgs vm hangs on boot sometimes, with traceback in Console that suggests FD_SETSIZE overflow in g_poll. both aarch64 and x86_64 emulated VMs are affected. lsof shows qemu using ~1024+ fds then.

Sometimes the issue happens, sometimes not (so a workaround is rebooting the VM several times until it "settles" on a relatively low number of FDs, ~500-700).

qemu has its own "g_poll" implementation for platforms that support ppoll() that doesn't go through glib loop. AFAIU macos doesn't support it. (?)

is this a known issue? is there anything we could do about it?

in glib, there's this in meson

if host_system in ['windows', 'darwin']
  # Poll doesn't work on devices on Windows, and macOS's poll() implementation is known to be broken
  glib_conf.set('BROKEN_POLL', true)
endif

/* The poll() emulation on OS/X doesn't handle fds=NULL, nfds=0,
 * so we prefer our own poll emulation.

that's on nixos

(ins)[nix-shell:/tmp]$ ./a.out
poll() returned: 0

(ins)[nix-shell:/tmp]$ cat test.c
#include <stdio.h>
#include <poll.h>
#include <errno.h>
#include <string.h>

int main() {
    int result = poll(NULL, 0, 1000);

    printf("poll() returned: %d\n", result);

    if (result == -1) {
        printf("Error: %s\n", strerror(errno));
    }

    return 0;
}


(ins)[nix-shell:/tmp]$ clang ./test.c

(ins)[nix-shell:/tmp]$ ./a.out
poll() returned: 0