The previous commit laid out the foundation of the quota engine, this
one builds on top of it, and implements the actual enforcement.
Enforcement happens at the route decoration level, whenever possible. In
case of the API, when over quota, a 413 error is returned, with an
appropriate JSON payload. In case of web routes, a 413 HTML page is
rendered with similar information.
This implementation is for a **soft quota**: quota usage is checked
before an operation is to be performed, and the operation is *only*
denied if the user is already over quota. This makes it possible to go
over quota, but has the significant advantage of being practically
implementable within the current Forgejo architecture.
The goal of enforcement is to deny actions that can make the user go
over quota, and allow the rest. As such, deleting things should - in
almost all cases - be possible. A prime exemption is deleting files via
the web ui: that creates a new commit, which in turn increases repo
size, thus, is denied if the user is over quota.
Limitations
-----------
Because we generally work at a route decorator level, and rarely
look *into* the operation itself, `size:repos:public` and
`size:repos:private` are not enforced at this level, the engine enforces
against `size:repos:all`. This will be improved in the future.
AGit does not play very well with this system, because AGit PRs count
toward the repo they're opened against, while in the GitHub-style fork +
pull model, it counts against the fork. This too, can be improved in the
future.
There's very little done on the UI side to guard against going over
quota. What this patch implements, is enforcement, not prevention. The
UI will still let you *try* operations that *will* result in a denial.
Signed-off-by: Gergely Nagy <forgejo@gergo.csillger.hu>
This is an implementation of a quota engine, and the API routes to
manage its settings. This does *not* contain any enforcement code: this
is just the bedrock, the engine itself.
The goal of the engine is to be flexible and future proof: to be nimble
enough to build on it further, without having to rewrite large parts of
it.
It might feel a little more complicated than necessary, because the goal
was to be able to support scenarios only very few Forgejo instances
need, scenarios the vast majority of mostly smaller instances simply do
not care about. The goal is to support both big and small, and for that,
we need a solid, flexible foundation.
There are thee big parts to the engine: counting quota use, setting
limits, and evaluating whether the usage is within the limits. Sounds
simple on paper, less so in practice!
Quota counting
==============
Quota is counted based on repo ownership, whenever possible, because
repo owners are in ultimate control over the resources they use: they
can delete repos, attachments, everything, even if they don't *own*
those themselves. They can clean up, and will always have the permission
and access required to do so. Would we count quota based on the owning
user, that could lead to situations where a user is unable to free up
space, because they uploaded a big attachment to a repo that has been
taken private since. It's both more fair, and much safer to count quota
against repo owners.
This means that if user A uploads an attachment to an issue opened
against organization O, that will count towards the quota of
organization O, rather than user A.
One's quota usage stats can be queried using the `/user/quota` API
endpoint. To figure out what's eating into it, the
`/user/repos?order_by=size`, `/user/quota/attachments`,
`/user/quota/artifacts`, and `/user/quota/packages` endpoints should be
consulted. There's also `/user/quota/check?subject=<...>` to check
whether the signed-in user is within a particular quota limit.
Quotas are counted based on sizes stored in the database.
Setting quota limits
====================
There are different "subjects" one can limit usage for. At this time,
only size-based limits are implemented, which are:
- `size:all`: As the name would imply, the total size of everything
Forgejo tracks.
- `size:repos:all`: The total size of all repositories (not including
LFS).
- `size:repos:public`: The total size of all public repositories (not
including LFS).
- `size:repos:private`: The total size of all private repositories (not
including LFS).
- `sizeall`: The total size of all git data (including all
repositories, and LFS).
- `sizelfs`: The size of all git LFS data (either in private or
public repos).
- `size:assets:all`: The size of all assets tracked by Forgejo.
- `size:assets:attachments:all`: The size of all kinds of attachments
tracked by Forgejo.
- `size:assets:attachments:issues`: Size of all attachments attached to
issues, including issue comments.
- `size:assets:attachments:releases`: Size of all attachments attached
to releases. This does *not* include automatically generated archives.
- `size:assets:artifacts`: Size of all Action artifacts.
- `size:assets:packages:all`: Size of all Packages.
- `size:wiki`: Wiki size
Wiki size is currently not tracked, and the engine will always deem it
within quota.
These subjects are built into Rules, which set a limit on *all* subjects
within a rule. Thus, we can create a rule that says: "1Gb limit on all
release assets, all packages, and git LFS, combined". For a rule to
stand, the total sum of all subjects must be below the rule's limit.
Rules are in turn collected into groups. A group is just a name, and a
list of rules. For a group to stand, all of its rules must stand. Thus,
if we have a group with two rules, one that sets a combined 1Gb limit on
release assets, all packages, and git LFS, and another rule that sets a
256Mb limit on packages, if the user has 512Mb of packages, the group
will not stand, because the second rule deems it over quota. Similarly,
if the user has only 128Mb of packages, but 900Mb of release assets, the
group will not stand, because the combined size of packages and release
assets is over the 1Gb limit of the first rule.
Groups themselves are collected into Group Lists. A group list stands
when *any* of the groups within stand. This allows an administrator to
set conservative defaults, but then place select users into additional
groups that increase some aspect of their limits.
To top it off, it is possible to set the default quota groups a user
belongs to in `app.ini`. If there's no explicit assignment, the engine
will use the default groups. This makes it possible to avoid having to
assign each and every user a list of quota groups, and only those need
to be explicitly assigned who need a different set of groups than the
defaults.
If a user has any quota groups assigned to them, the default list will
not be considered for them.
The management APIs
===================
This commit contains the engine itself, its unit tests, and the quota
management APIs. It does not contain any enforcement.
The APIs are documented in-code, and in the swagger docs, and the
integration tests can serve as an example on how to use them.
Signed-off-by: Gergely Nagy <forgejo@gergo.csillger.hu>
