The previous fix had only been tested locally through a runtime edit of
the service, and the order in which @chown had been re-added was
different so commit cf498c1a61 ("nixos/cryptpad: fix service with
nodejs 22.11") did not actually fix the issue.
This properly orders @chown after @privileged so the rule is respected,
and also properly denies with EPERM instead of allowing the chown family
of syscalls: this will properly prevent seccomp from killing nodejs
while still disallowing fchown()
Fixes https://github.com/NixOS/nixpkgs/issues/370717
Previously some modules used `config.environment.etc."ssl/certs/ca-certificates.crt".source`, some used `"/etc/ssl/certs/ca-certificates.crt"`, and some used `"${pkgs.cacert}/etc/ssl/certs/ca-bundle.crt"`. These were all bad in one way or another:
- `config.environment.etc."ssl/certs/ca-certificates.crt".source` relies on `source` being set; if `text` is set instead this breaks, introducing a weird undocumented requirement
- `"/etc/ssl/certs/ca-certificates.crt"` is probably okay but very un-nix. It's a magic string, and the path doesn't change when the file changes (and so you can't trigger service reloads, for example, when the contents change in a new system activation)
- `"${pkgs.cacert}/etc/ssl/certs/ca-bundle.crt"` silently doesn't include the options from `security.pki`
Co-authored-by: Shelvacu <git@shelvacu.com>
nixosTests.cryptpad started failing recently.
Investigating the issue shows that seccomp has become problematic during
the init phase, (e.g. this can be reproduced by removing the customize
directory in /var/lib/cryptpad):
machine # [ 10.774365] systemd-coredump[864]: Process 756 (node) of user 65513 dumped core.
machine #
machine # Module libgcc_s.so.1 without build-id.
machine # Module libstdc++.so.6 without build-id.
machine # Module libicudata.so.74 without build-id.
machine # Module libicuuc.so.74 without build-id.
machine # Module libicui18n.so.74 without build-id.
machine # Module libz.so.1 without build-id.
machine # Module node without build-id.
machine # Stack trace of thread 756:
machine # #0 0x00007ff951974dcb fchown (libc.so.6 + 0x107dcb)
machine # #1 0x00007ff95490d0c0 uv__fs_copyfile (libuv.so.1 + 0x150c0)
machine # #2 0x00007ff95490d89a uv__fs_work (libuv.so.1 + 0x1589a)
machine # #3 0x00007ff954910c76 uv_fs_copyfile (libuv.so.1 + 0x18c76)
machine # #4 0x0000000000eb8a39 _ZN4node2fsL8CopyFileERKN2v820FunctionCallbackInfoINS1_5ValueEEE (node + 0xab8a39)
machine # #5 0x0000000001cda5e2 Builtins_CallApiCallbackGeneric (node + 0x18da5e2)
[...]
machine # [ 10.877468] cryptpad[685]: /nix/store/h4yhhxpfm03c5rgz91q7jrvknh596ly2-cryptpad-2024.12.0/bin/cryptpad: line 3: 756 Bad system call (core dumped) "/nix/store/fkyp1bm5gll9adnfcj92snyym524mdrj-nodejs-22.11.0/bin/node" "/nix/store/h4yhhxpfm03c5rgz91q7jrvknh596ly2-cryptpad-2024.12.0/lib/node_modules/cryptpad/scripts/build.js"
nodejs 20.18 rightly did not require chown when the source and
destination are the same owner (heck, the script does not run as
root so even if it is not blocked there is no way it'd work with a
different owner...)
For now just allow chown calls again, this is not worth wasting more
time.
Fixes https://github.com/NixOS/nixpkgs/issues/370717
This is a full rewrite independent of the previously removed cryptpad
module, managing cryptpad's config in RFC0042 along with a shiny test.
Upstream cryptpad provides two nginx configs, with many optimizations
and complex settings; this uses the easier variant for now but
improvements (e.g. serving blocks and js files directly through nginx)
should be possible with a bit of work and care about http headers.
the /checkup page of cryptpad passes all tests except HSTS, we don't
seem to have any nginx config with HSTS enabled in nixpkgs so leave this
as is for now.
Co-authored-by: Pol Dellaiera <pol.dellaiera@protonmail.com>
Co-authored-by: Michael Smith <shmitty@protonmail.com>
