Per RFC 9110, [section 8.8.1][1], different representations of the same
resource should have different Etags:
> A strong validator is unique across all versions of all
> representations associated with a particular resource over time.
> However, there is no implication of uniqueness across representations
> of different resources (i.e., the same strong validator might be in
> use for representations of multiple resources at the same time and
> does not imply that those representations are equivalent)
When serving statically compressed files (ie, when there is an existing
corresponding .gz/.br/etc. file on disk), Nginx sends the Etag marked
as strong. These tags should be different for each compressed format
(as shown in an explicit example in section [8.8.3.3][2] of the RFC).
Upstream Etags are composed of the file modification timestamp and
content length, and the latter generally changes between these
representations.
Previous implementation of Nix-specific Etags for things served from
store used the store hash. This is fine to share between different
files, but it becomes a problem for statically compressed versions of
the same file, as it means Nginx was serving different representations
of the same resource with the same Etag, marked as strong.
This patch addresses this by imitating the upstream Nginx behavior, and
appending the value of content length to the store hash.
[1]: https://www.rfc-editor.org/rfc/rfc9110.html#name-validator-fields
[2]:
https://www.rfc-editor.org/rfc/rfc9110.html#name-example-entity-tags-varying
The old limit was only 128 and this breaks some applications like e.g.
Frigate where playlists become bigger than that. According to upstream
you should just change the variable yourself if needed.
See this issue: https://github.com/kaltura/nginx-vod-module/issues/238
nginx lua needs resty
the enableSandbox option of nginx was removed in 535896671b
the test fails with
```
vm-test-run-nginx-sandbox> machine # [ 47.753580] nginx[1142]: nginx: [alert] detected a LuaJIT version which is not OpenResty's; many optimizations will be disabled and performance will be compromised (see https://github.com/openresty/luajit2 for OpenResty's LuaJIT or, even better, consider using the OpenResty releases from https://openresty.org/en/download.html)
vm-test-run-nginx-sandbox> machine # [ 47.756064] nginx[1142]: nginx: [alert] failed to load the 'resty.core' module (https://github.com/openresty/lua-resty-core); ensure you are using an OpenResty release from https://openresty.org/en/download.html (reason: module 'resty.core' not found:
vm-test-run-nginx-sandbox> machine # [ 57.911766] systemd[1]: Failed to start Nginx Web Server.
```
When listening on unix sockets, it doesn't make sense to specify a port
for nginx's listen directive.
Since nginx defaults to port 80 when the port isn't specified (but the
address is), we can change the default for the option to null as well
without changing any behaviour.
PROXY protocol is a convenient way to carry information about the
originating address/port of a TCP connection across multiple layers of
proxies/NAT, etc.
Currently, it is possible to make use of it in NGINX's NixOS module, but
is painful when we want to enable it "globally".
Technically, this is achieved by reworking the defaultListen options and
the objective is to have a coherent way to specify default listeners in
the current API design.
See `mkDefaultListenVhost` and `defaultListen` for the details.
It adds a safeguard against running a NGINX with no HTTP listeners (e.g.
only PROXY listeners) while asking for ACME certificates over HTTP-01.
An interesting usecase of PROXY protocol is to enable seamless IPv4 to
IPv6 proxy with origin IPv4 address for IPv6-only NGINX servers, it is
demonstrated how to achieve this in the tests, using sniproxy.
Finally, the tests covers:
- NGINX `defaultListen` mechanisms are not broken by these changes;
- NGINX PROXY protocol listeners are working in a final usecase
(sniproxy);
- uses snakeoil TLS certs from ACME setup with wildcard certificates;
In the future, it is desirable to spoof-attack NGINX in this scenario to
ascertain that `set_real_ip_from` and all the layers are working as
intended and preventing any user from setting their origin IP address to
any arbitrary, opening up the NixOS module to bad™ vulnerabilities.
For now, it is quite hard to achieve while being minimalistic about the
tests dependencies.
