Proper sandbox mechanism

[andreypopp]

Nix uses sandbox-exec on macOS and something on Linux (cgroups?) to catch sandbox violations. We should do the same.

What we need is to allow writes only to $cur__target_dir and $cur__install and reads from some whitelisted locations + dependencies' installation locations.

[andreypopp]

Got support for that on macOS in #17. If anyone works on Linux and has knowledge of Linux namespaces — they are invited to help me with that!

[yunxing]

One thing we can try on linux is to create two mount points for the subprocess -- one for the entire filesystem and one for $cur__root. Then use chroot before executing the command, so something like:

mount --bind / /tmp/chroot_for_esy
mount -o remount,ro /tmp/chroot_for_esy 
mount --bind $cur_root /tmp/chroot_for_esy/$cur_root  # make cur_root writable

I've never tried anything like this so not sure if this approach would work.

[andreypopp]

@yunxing does mount --bind require root? That would be a downside if so.

I was thinking of namespaces on Linux to get the same level of isolation as #17 implements on macOS — allow writes into $cur__install and $cur__target_dir only. That would be a good start as it will make build caching safe.

[yunxing]

@andreypopp namespaces on Linux uses mount. I can experiment with that a bit.