🚇 fix: Repair Direct Sandbox Startup And IPv6 Binding - #26

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ LeonardSEO
❌ GPTLeonard
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

@codex review

@LeonardSEO if you currently dont find the time could take over the PR and finish it as these would eliminate my custom fixes i have currently deployed.

@snapydziuba Continued the PR and addressed the remaining review finding in b777895: SANDBOX_BIND_HOST now passes through the microVM launcher allowlist, with regression coverage in the existing launcher test. The new GitHub Actions run is currently marked action_required and has not started any jobs, so it needs maintainer approval before CI can run.

Hi @LeonardSEO — thanks for splitting this into the parts that were fixable without touching the storage or egress architecture. That framing held up. Reviewing it against current main: one of the three fixes has since landed independently, and the other two are still needed and still broken, so this shouldn't just be closed.

Status of each part

#22 was fixed incidentally by #66, which replaced the whole bind sequence with a statically-linked docker/rootfs-setup.c that binds /sandbox_api and /pkgs first, then a single /usr last — the same fix you proposed, for the same reason. Your diagnosis was right; it just arrived via the bridge work, which is why nobody linked it here.

#23 is untouched — api/src/config.ts:49 is still bind_address: `0.0.0.0:${process.env.PORT ?? 2000}` and api/src/index.ts:72 still does config.bind_address.split(':').

#24's empty-mount hazard is untouched too, and #66 actually inherited it. docker/rootfs-setup.c:72 binds on existence alone:

if (access("/host-packages", F_OK) == 0 &&
    bind_mount("/host-packages", "/pkgs", 0) != 0) {

and api/Dockerfile:260, docker/Dockerfile.worker-sandbox:271 and launcher/Dockerfile:137 all mkdir -p /host-packages — so the directory always exists and an empty one always shadows $ROOTFS/pkgs. That's precisely the case your patch guards against.

What a rebase needs

I test-merged this against main: exactly one file conflicts — docker/start-direct-sandbox.sh. Everything else, including both api/src/config.ts and api/src/index.ts, auto-merges cleanly. So the IPv6 half needs no work at all.

The catch is that your conflicting block carries both the superseded /usr bind and the still-wanted /host-packages check. Resolving it by taking main's side wholesale would silently drop the #24 fix — that logic now lives in rootfs-setup.c, so it needs porting rather than deleting:

static int has_entries(const char *path) {
  DIR *dir = opendir(path);
  if (dir == NULL) return 0;
  struct dirent *entry;
  int found = 0;
  while ((entry = readdir(dir)) != NULL) {
    if (strcmp(entry->d_name, ".") != 0 && strcmp(entry->d_name, "..") != 0) {
      found = 1;
      break;
    }
  }
  closedir(dir);
  return found;
}

then gate the existing bind on has_entries("/host-packages") instead of access(..., F_OK), keeping the baked $ROOTFS/pkgs bind as the fallback.

tests/start-direct-sandbox.test.sh will need rewriting against the new entrypoint too, since the shell script it exercised is now a compiled binary.

If it's easier, this is also fine to split: the IPv6 fix is self-contained, needs no conflict resolution at all, and could land on its own straight away.

(Context: #22 is now closed as completed. This is the only open PR carrying fixes for #23 and #24, so I'd rather see it rebased than superseded.)