smallwebwaf/docs/gitea.md

# smallwebwaf in front of gitea

`deploy/gitea/compose.yaml` runs gitea behind traefik, with `rproxy` and
`servewaf` in front of it. This document says what that setup sets and why, and
what the defaults would do without it. [`SPEC.md`](../SPEC.md) says how
`smallwebwaf` behaves.

## The setup

Four containers share one Docker network, the shared network:

- traefik, which routes the site's HTTPS requests to `rproxy`, and redirects
  plain HTTP to HTTPS;
- `rproxy`, in front of gitea;
- gitea 1.27.3, as its image comes;
- `servewaf`, with a volume for its state files.

To start it:

1. In `deploy/gitea/compose.yaml`, replace `git.example.org` with the site's
   host name.
2. Replace each `10.99.0` with the first three numbers of a `/24` subnet that no
   other network on the host uses. The `/24` must lie inside `10.0.0.0/8`,
   `172.16.0.0/12` or `192.168.0.0/16`, because `rproxy` believes
   `X-Forwarded-For` only from those by default. Replace each `fd10:99` with the
   first two hextets of an unused `/64` subnet whose third and fourth hextets
   are 0, as in `fd10:99::/64`.
3. Set `SMALLWEBWAF_IMAGE` to the `smallwebwaf` image, named by digest.
4. Set `GITEA_SECRET_KEY` to the output of
   `docker run --rm <gitea's image> gitea generate secret SECRET_KEY`, with
   gitea's image as `deploy/gitea/compose.yaml` names it. Do this once per
   install, and keep the key.
5. Run `docker compose up --detach` in `deploy/gitea`.
6. Make gitea's first user with `gitea admin user create`, in gitea's container,
   as the user `git`.

Sign-up is off: accounts are made with `gitea admin user create`. gitea starts
at its defaults, with SQLite. traefik answers with a certificate it made itself
until you add a certificate resolver, as for your other sites. SSH to gitea does
not pass through `smallwebwaf`; publish gitea's SSH port yourself if you use it.

`make gitea-setup` checks the setup. It builds the image, brings the setup up
under a name of its own, and checks:

- that a page load, a push and a clone sent through traefik pass through
  `rproxy`, and that `servewaf` checks them;
- that gitea logs the visitor's address, not `rproxy`'s;
- that a request from a container on the shared network skips the checks;
- that with `servewaf` stopped, gitea still answers through `rproxy`.

## Uploads and downloads

An upload is a request:

- a git push;
- an LFS object;
- a container image layer;
- a package file;
- a release attachment.

The setup lets an upload of up to 2 GiB, sent in up to 30 minutes, through:

- `SWRP_MAX_REQUEST_SIZE` is `2G`. At the default, an upload over 100 MiB is cut
  off.
- `SWRP_MAX_REQUEST_TIME_MS` is `1800000`, 30 minutes. At the default, an upload
  that takes over 60 seconds is cut off.
- traefik's `readTimeout` is 30 minutes. traefik's own default of 60 seconds for
  a whole request would cut off an upload first.

A clone or a download is a response:

- `SWRP_MAX_RESPONSE_TIME_MS` is `900000`, 15 minutes, half the default. A clone
  or download that takes longer is cut off.
- `SWRP_MAX_RESPONSE_SIZE` stays at its default, 1 GiB. A repository whose clone
  is larger needs it raised, and the byte limits with it ("Byte limits, per
  client" in `SPEC.md`).

The byte limits stay at their defaults. Two uploads of 2 GiB from one address in
one minute pass the minute's limit of 4 GiB, and ban the address.

## The checks

`SWWAF_WAF_MODE` is `detect`. The Core Rule Set then logs a match, with its rule
ids, and refuses nothing. In `block` mode, the default, it refuses everyday
gitea searches and file names. The section "What the Core Rule Set refuses in
`block` mode" below lists them. Each such refusal counts toward the error burst,
so 31 in one minute ban the address.

Every other check stays at its default. A client may make 500 requests a minute,
20,000 an hour and 300,000 a day. Archive downloads and blame and history pages
are what scrapers hammer, and these limits do most of the work there.

## The shared network

Containers on the shared network skip `servewaf`'s checks:

- The shared network's subnets are `10.99.0.0/24` and `fd10:99::/64`. Docker
  gives its containers addresses in the upper half of each, `10.99.0.128/25` and
  `fd10:99:0:0:8000::/65`, which `SWWAF_ALLOW_NETS` names.
- traefik is known on the shared network by the site's host name. So a container
  there that asks for the site reaches traefik from its own address. An
  automation fleet in such a container skips the checks.
- Each subnet's gateway, `10.99.0.1` and `fd10:99::1`, is the host's own address
  on it. Requests docker passes on from a published port can come from the
  gateway, so the gateways are left out of `SWWAF_ALLOW_NETS`. The network has
  IPv6, so a visitor over IPv6 reaches traefik from its own address, and
  `servewaf` checks it as itself.
- A container on the shared network that asks for the site at the host's public
  address arrives from the gateway. It is checked like any visitor.
- A fleet on another host can have its address added to `SWWAF_ALLOW_NETS`.
  Without that, the default request limits are sized for such a fleet.
- traefik replaces the `X-Forwarded-For` a visitor sends, by default. Keep it
  so: otherwise a visitor could name an address on the shared network, and skip
  the checks.
- Anything on the shared network can reach gitea and `servewaf` around `rproxy`.
  Put only containers you trust on it.

## gitea's settings

- `REVERSE_PROXY_TRUSTED_PROXIES` names `rproxy`'s fixed addresses, `10.99.0.2`
  and `fd10:99::2`. gitea then takes each visitor's address from the `X-Real-Ip`
  header, which traefik sets and `rproxy` passes on. Without it, gitea logs
  every visitor as `rproxy`.
- `INSTALL_LOCK` is `true`, as "The setup" above says.

## Methods

`rproxy` passes on only these methods: `GET`, `HEAD`, `POST`, `OPTIONS`, `PUT`,
`DELETE`, `PATCH` and `QUERY`. gitea 1.27.3 needs no other. Its routes use:

- pages and the API: `GET`, `HEAD`, `POST`, `PUT`, `PATCH` and `DELETE`;
- git's push, clone and fetch over HTTP: `GET`, `POST` and `OPTIONS`;
- LFS: `GET`, `POST` and `PUT`;
- the container registry and the other package registries: `GET`, `HEAD`,
  `POST`, `PUT`, `PATCH` and `DELETE`;
- Actions runners: `POST`.

## The log

`rproxy` and `servewaf` write each query string to the log whole. A token sent
in the query string, as `?token=`, is then in the log. Send it in the
`Authorization` header instead.

## Actions runners

A runner on the shared network skips the checks. One on another host is counted
like any client. A runner older than version 0.4 (April 2026) asks for work
every 2 seconds, 43,200 requests a day. While it runs a job, it reports the
job's log and state every second instead, two requests a second. So one runner
busy all day sends 120 requests a minute and 172,800 a day, under the request
limits.

Several runners on one address can pass them:

- Two busy all day make 345,600 requests, and pass the day limit of 300,000.
- Three busy at once make 21,600 an hour, and pass the hour limit of 20,000.
- Five busy at once make 600 a minute, and pass the minute limit of 500.

Their address is then banned, their running jobs fail, and each repeat within a
day makes the ban longer. Newer runners ask for work less often. Put the
runners' addresses in `SWWAF_RATE_LIMIT_EXEMPT_NETS`.

## What the Core Rule Set refuses in `block` mode

These notes record what the Core Rule Set does in front of gitea in `block`
mode, at the other defaults. They are observations, not requirements.
`script/gitea-requests` runs gitea behind `rproxy` and `servewaf` at the
defaults. It sends through them a sample of gitea's ordinary use and the
requests named below. It reports each request the Core Rule Set refuses, so that
these notes can be checked again when gitea or the Core Rule Set changes
version. In `detect` mode, as the setup runs it, the same requests are logged
with their rule ids, and pass.

Each refusal answers only that request ("Bans" in `SPEC.md` says what it counts
toward). The request log names the rule in `waf_rule_ids`, which
`SWWAF_WAF_DISABLED_RULES` can switch off.

### What it lets through

The Core Rule Set lets gitea's ordinary use through, apart from the refusals in
the sections after this one:

- browsing and views of files in a repository, with their history, blame and the
  file tree;
- diffs, including their hidden lines and large files, and pull request review;
- git's clone, fetch and push over HTTP;
- signing in, including the return to the page a visitor came from, and sign-in
  with Git Credential Manager, git-credential-oauth or tea;
- the API's calls for a file and its commits;
- pushing and pulling container images and packages;
- Actions runners, and artifacts uploaded with `actions/upload-artifact@v4`;
- posting issues, pull requests, comments, wiki pages and files saved in the web
  editor, code included, since no body is read;
- the page gitea shows after a change, with its message naming what was done.

### A query string that reads as an attack

Most often this is a search. A search term is a query parameter, which the Core
Rule Set checks against its lists of system files and commands (see the fourth
change under "Attack detection" in `SPEC.md`). A release attachment uploaded
through the API as `docker-compose.yml` is refused the same way. Examples:

- a name on those lists, such as `package.json`, `.gitignore` or
  `docker-compose.yml`;
- text that starts with a command name, such as `python3` or `ssh key`, or has
  one right after a `/` in its first word, such as `feat/docker-support`;
- a shell command with its options or a system path (`ls -la`, `sed -i`,
  `/bin/sh`), or a command in backticks;
- script code (`fetch(`, `${VAR}`, `process.env`), HTML (`<img src=`) or an SQL
  statement (`SELECT * FROM users WHERE`);
- a URL naming an IP address or `localhost`.

### A branch, tag or file name that starts with a command name

The Core Rule Set's list of commands (932260) holds names such as `docker-`,
`python3`, `ansible`, `base64` and `whoami`. A query parameter holding a branch,
tag or file name that starts with one is refused. So is one that has such a name
right after a `/`, such as `feat/docker-support`, `fix/docker-build` or
`renovate/docker-build-push-action-6.x`. gitea's own pages send the name in a
query parameter that 930120, 932160 and 932260 still check:

- `refSubUrl`, the branch or tag of a directory listing. A listing sends it when
  it asks for its entries' last commits in a second request. It does so whenever
  gitea takes more than a second to work them out;
- `name`, when a branch is deleted or restored on the branches page;
- `tag`, when a release is started from a tag;
- `template`, the file of an issue template, when an issue is opened from it;
- `rule_name`, when a branch protection rule is opened for editing in the
  repository's settings or has just been saved.

What a visitor sees, for a branch named `docker-build`:

- A directory listing that makes that second request leaves those last commits
  out and shows an error.
- The branches page shows an error instead of deleting or restoring the branch.
- A release started from such a tag gets the `403` answer of `smallwebwaf` in
  place of its form. So does an issue opened from such a template, and such a
  rule opened for editing.
- Saving such a rule stores it. gitea then returns to the branch settings page,
  which gets the `403` answer, so the save looks as if it failed.

### A file name in a query parameter that trips the code rules

The Core Rule Set's rules for Java, script, PowerShell and SQL find what they
look for in some file names. This holds too in the ten query parameters that
930120, 932160 and 932260 do not check (see the fourth change under "Attack
detection" in `SPEC.md`). Examples:

- `runtime` in OpenJDK's `src/java.base/share/classes/java/lang/Runtime.java`
  (944110);
- a Java class, as in `java.lang.String.md` (944130);
- `document.write.js` (941180);
- `process.env.d.ts` (934100);
- `__proto__` (934130);
- the PowerShell command `Install-Script`, as in `install-script.sh` (932120);
- a database's name, as in PostgreSQL's `information_schema.sql` (942140).

For such a file, these requests are refused:

- the API's list of its commits (`path`);
- the API's call for the contents of files that names it in its query (`body`);
- the commit graph of that file alone (`file`);
- the hidden lines of its diff (`path`), which stay hidden;
- the sign-in page, for a visitor who opens a page of the file that needs
  signing in, such as its edit page (`redirect_to`). The visitor can still sign
  in by opening the sign-in page itself.

### A path

- A file name ending in `~`, or an `.xhtml` file whose path holds a space.
- A path that holds what the Core Rule Set takes for a Java class (944130), as
  `java.lang.String.md` and
  `META-INF/services/org.apache.commons.logging.LogFactory` do.
- A path that holds what it takes for Node.js code (934100), as
  `process.env.d.ts` does.
- A path with a directory named `...` (930100, 930110), such as `.../notes.txt`.

For such a file, these are refused:

- the page of the file;
- its raw file;
- its history and its blame;
- its edit page.

The directory that holds it lists it, and a clone carries it.

### A container image push that reuses a layer

A push can reuse a layer of an image docker has pushed to that gitea before.
docker then asks gitea to take the layer from that image, naming it in the query
parameter `from`, which 932260 checks. The request is refused when the image's
owner starts with an entry on the list of commands. It is refused too when the
image's name after the owner starts with one. Examples are `owner/python3-app`,
`owner/docker-base`, and any image of an owner named `ansible` or `docker-team`.

What happens next depends on docker's image store. A docker that keeps its
images in containerd's image store fails the push, and fails each new try the
same way. A new install of Docker Engine 29 does this. A docker that keeps them
in its own older image store uploads the layer instead. The push then succeeds
after one refusal for each layer it reuses, and each of those refusals counts
toward the error burst ("Attack detection" in `SPEC.md`).

### Artifact uploads from `actions/upload-artifact@v3`

This action sends the header `Content-Range`, which the Core Rule Set refuses
(920450). The action sends two files at a time, does not retry a `403`, and
sends no further file once one is refused. So an upload is refused once or
twice, however many files it holds, and the step fails. One upload does not ban
the runner. The error burst ("Attack detection" in `SPEC.md`) bans it after more
than 30 such refusals from its address in a minute, as when many jobs upload at
once. `actions/upload-artifact@v4` does not send the header.

### What body inspection refuses

With `SWWAF_WAF_BODY_LIMIT` set, the Core Rule Set reads request bodies too. In
front of gitea it then refuses these, until their rule ids, named in the request
log, are added to `SWWAF_WAF_DISABLED_RULES`:

- issue and comment text, wiki pages and files saved in the web editor that hold
  shell commands or code (932125, 932235, 932250 and others);
- package descriptions that show code;
- PyPI uploads (922130);
- attachments named like `debug.log` or `config.yml` (932180).

### What `script/gitea-requests` sends

Of the requests above, the script sends:

- for each file named under "A file name in a query parameter that trips the
  code rules", the API's list of its commits;
- for `Runtime.java` alone, the commit graph of that file, the hidden lines of
  its diff, and the API's call for the contents of files that names it;
- the sign-in page for a visitor who opens the edit page of `document.write.js`;
- the pages of `java.lang.String.md`,
  `META-INF/services/org.apache.commons.logging.LogFactory`, `process.env.d.ts`
  and `.../notes.txt`;
- one container image push that reuses a layer, which it does not try again.

It does not send the raw file, history, blame or edit page of those four files,
or the pages of the directories `META-INF/services` and `...`. It also sends a
short sample of gitea's ordinary use:

- signing in;
- pages of files and commits, and a diff;
- the API;
- a push, and clone and fetch over HTTP;
- two container image pushes.