| 30 Nov 2025 |
 hexa | possibly https://github.com/DavHau/nix-portable | 21:38:38 |
 Jules Lamur | yeah sorry forgot about the important part: with the sandbox :) | 21:38:49 |
| Jules Lamur | (ie. sandbox-fallback = false) | 21:39:03 |
 raitobezarius | AFAIK the podman rootless thing has a seccomp policy that prevents all unshare calls with any relevant arg | 21:39:13 |
| raitobezarius | If you get rid of that and you have subuid delegation, you can run with sandbox | 21:39:30 |
| raitobezarius | Otherwise hexa gave you the 50% performance penalty solution by using syscall interception | 21:39:46 |
| Jules Lamur | yep you're right, running with eg podman run --cap-add=SYS_ADMIN --security-opt unmask=/proc/* --rm -it works | 21:39:53 |
| raitobezarius | In reply to @jlamur:matrix.org
yep you're right, running with eg podman run --cap-add=SYS_ADMIN --security-opt unmask=/proc/* --rm -it works I think you can get remove just the unshare blacklist | 21:40:13 |
| raitobezarius | Don't need to unmask proc and cap add | 21:40:20 |
| raitobezarius | For proc, you only need a partial view, not a full view anyway, but you need a proper procfs I suppose | 21:41:03 |
| Jules Lamur | AFAIU, podman sets a policy that prevents any mount related syscall in /proc, so remounting a procfs should not work at all without the --security-opt flag above (again, if I understand correctly -- I only have a high level understanding of all that) | 21:43:18 |
| raitobezarius | Argh, yeah, we need *a* procfs | 21:44:11 |
| raitobezarius | Not necessarily *the* procfs | 21:44:16 |
| Jules Lamur | I'm not sure that I understand what you're suggesting here | 21:44:51 |
| raitobezarius | In reply to @jlamur:matrix.org
I'm not sure that I understand what you're suggesting here --security-opt seccomp reduced-seccomp.json | 21:45:29 |
| raitobezarius | The JSON file is a fine grained control of which syscall is allowed | 21:46:23 |
| raitobezarius | Allowing unshare makes sense if you want the sandbox to work at all | 21:46:45 |
| raitobezarius | Perhaps, allowing the mount procfs is sufficiently risk free too | 21:46:56 |
| Jules Lamur | what do you mean by allowing unshare? user namespaces are allowed by default in podman rootless containers so unshare should work out of the box | 21:48:43 |
| raitobezarius | I was pretty sure the default seccomp policy forbids the unshare syscall | 21:49:08 |
| Jules Lamur | (that does not solve my problem though because of the procfs mask) | 21:49:13 |
| raitobezarius | So maybe the question is: if you try to allow the mount syscall for procfs, does it work? | 21:49:35 |
| raitobezarius | Without any cap or proc unmasking | 21:49:41 |
| Jules Lamur | If I try to run with only the unmask, then I get that:
$ podman run --security-opt unmask=/proc/* --rm -it foo unshare -r nix-build
error: changing ownership of path '/nix/store': Invalid argument
| 21:52:54 |
| Jules Lamur | mhh I think that's because /nix/store is root:nogroup in the new user ns | 21:55:10 |
| raitobezarius | That's very weird to have EINVAL on chown? Do you have a clean mount namespace? | 21:55:12 |
| raitobezarius | In reply to @jlamur:matrix.org
mhh I think that's because /nix/store is root:nogroup in the new user ns So subuid delegation | 21:55:24 |
 WeetHet | WYM? | 21:56:57 |
| raitobezarius | In reply to @weethet:catgirl.cloud
WYM? The surface of API calls used for nix-doc is extremely low and trivial | 21:57:48 |
| raitobezarius | It's absolutely not representative of general purpose FFI for CLI alas | 21:58:05 |
