When we verify an HTTP signature,
* If we know the key, check in the DB whether we know the actor lists it. If it
doesn't, and there's room left for keys, HTTP GET the actor and update the DB
accordingly.
* If we know the key but had to update it, do the same, check usage in DB and
update DB if needed
* If we don't know the key, record usage in DB
However,
* If we're GETing a key and discovering it's a shared key, we GET the actor to
verify it lists the key. When we don't know the key at all yet, that's fine
(can be further optimized but it's marginal), but if it's a key we do know,
it means we already know the actor and for now it's enough for us to rely
only on the DB to test usage.
Previously, when verifying an HTTP signature and we fetched the key and
discovered it's shared, we'd fetch the actor and make sure it lists the key URI
in the `publicKey` field. But if we already knew the key, had it cached in our
DB, we wouldn't check the actor at all, despite not knowing whether it lists
the key.
With this patch, we now always GET the actor when the key is shared,
determining the actor URI from the `ActivityPub-Actor` request header, and we
verify that the actor lists the key URI. We do that regardless of whether or
not we have the key in the DB, although these two cases and handled in
different parts of the code right now (for a new key, it's in Web.ActivityPub
fetchKey; for a known key, it's in Vervis.Foundation httpVerifySig).
Previously, when verifying an HTTP signature and we find out we have the
provided keyid in the DB, and this key is a personal key, we would just grab
the key owner from the DB and ignore the ActivityPub-Actor header.
This patch adds a check: If we find the key in the DB and it's a personal key,
do grab the owner from that DB row, but also check the actor header: If it's
provided, it has to be identical to the key owner ID URI.
If the key we fetched is a shared key, the only way to determine the actor to
which the signature applies is to read the HTTP header ActivityPub-Actor. But
if it's a personal key, we can detect the actor by checking the key's owner
field. Still, if that actor header is provided, we now compare it to the key
owner and make sure they're identical.
When fetching a key that is embedded in the actor document, we were already
comparing the actor ID with the actor header, so that part didn't require
changes.
When a local user wants to publish an activity, we were always GETing the
recipient actor, so that we could determine their inbox and POST the activity
to it. But now, instead, whenever we GET an actor (whether it's for the key sig
verification or for determining inbox URI), we keep their inbox URI in the
database, and we don't need to GET it again next time.
Using a dedicated type allows to record in the type the guarantees that we
provide, such as scheme being HTTPS and authority being present. Allows to
replace ugly `fromJust` and such with direct field access.
Before, there was a single key used as a personal key for all actors. Now,
things work like this:
- There are 2 keys, each time one is rotated, this way the old key remains
valid and we can freely rotate without a risk of race conditions on other
servers and end up with our posts being rejected
- The keys are explicitly instance-scope keys, all actors refer to them
- We add the ActivityPub-Actor header to all activity POSTs we send, to declare
for which specific actor our signature applies. Activities and otherwise
different payloads may have varying ways to specify attribution; using this
header will be a standard uniform way to specify the actor, regardless of
payload format. Of course, servers should make sure the actual activity is
attributed to the same actor we specified in the header. (This is important
with instance-scope keys; for personal keys it's not critical)
Allow keys to specify expiration time using w3c security vocabulary. If a key
has expired, we treat it like sig validation failure and re-fetch the key from
the other server. And we never accept a sig, even a valid sig, if the key has
expired.
Since servers keep actors and keys in the DB, expiration can be a nice way to
ask that keys aren't used more than we want them to. The security vocab spec
also recommends to set expiration time on keys, so it's nice to support this
feature.
It's now possible for activities we be attributed to actors that have more than
one key. We allow up to 2 keys. We also store in the DB. Scaling to support any
number of keys is trivial, but I'm limiting to 2 to avoid potential trouble and
because 2 is the actual number we need.
By having 2 keys, and replacing only one of them in each rotation, we avoid
race conditions. With 1 key, the following can happen:
1. We send an activity to another server
2. We rotate our key
3. The server reaches the activity in its processing queue, tries to verify our
request signature, but fails because it can't fetch the key. It's the old
key and we discarded it already, replaced it with the new one
When we use 2 keys, the previous key remains available and other servers have
time to finish processing our requests signed with that key. We can safely
rotate, without worrying about whether the user sent anything right before the
rotation time.
Caveat: With this feature, we allow OTHER servers to rotate freely. It's safe
because it's optional, but it's just Vervis right now. Once Vervis itself
starts using 2 keys, it will be able to rotate freely without race condition
risk, but probably Mastodon etc. won't accept its signatures because of the use
of 2 keys and because they're server-scope keys.
Maybe I can get these features adopted by the fediverse?
Shared key means the key is used for multiple actors. I'm not sure explicitly
specifying this will be necessary, but I prefer to have it in place to help
with debugging in case something unexpected comes from other servers, or my
format overlaps with stuff used in other software and encodes a different
meaning.
Each public key can specify whether it's shared or personal, and this patch
checks for that when verifying a request signature. It rejects shared keys,
accepting valid sigs only from personal keys.
Very soon I'll add shared key support.
* Repo collab now supports basic default roles developer/user/guest like
project collab does
* User/Anon collab for repos and projects are now stored as fields instead of
in dedicated tables, there was never a need for dedicated tables but I didn't
see that before
* Repo push op is now part of `ProjectOperation`
* `RepoRole` and related code has been entirely removed, only project roles
remain and they're used for both repos and projects
* This is the first not-totally-trivial DB migration in Vervis, it's automatic
but please be careful and report errors
* When adding collaborators, you don't need a custom role. If you don't choose
one, a basic default "developer" role will be used
* If you don't assign a `ProjectCollabUser` role, a default "user" role is
assumed for logged in users, otherwise a "guest" role
* The "guest" role currently has no access at all
* Theoretically there may also be a "maintainer" role allowing project
sharers/maintainers to give maintainer-level access to more people, but right
now maintainer role would be the same as developer so I haven't added it yet
It already had one, but it didn't have a public key and it was using the old
mess of the Vervis.ActivityStreams module, which I'll possibly remove soon.
It's hopefully more elegant now.
This patch includes some ugliness and commented out code. Sorry for that. I'll
clean it up soon.
Basically there's a TVar holding a Vector of at most 10 AP activities. You can
freely POST stuff to /inbox, and then GET /inbox and see what you posted, or an
error description saying why your activity was rejected.
The actor key will be used for all actors on the server. It's held in a `TVar`
so that it can always be safely updated and safely retrieved (technically there
is a single writer so IORef and MVar could work, but they require extra care
while TVar is by design suited for this sort of thing).
In Haskell by default if a thread has an exception, the main thread isn't
notified at all. This patch changes service thread launching to re-throw their
exceptions in the main thread, so that their failure is noticed.
I suppose there's no performance difference in using one, but it requires
`http-conduit` as a build dependency, so potentially we may be reducing build
time by removing unnecessary deps.
Git pull uses a POST request, which is treated as a write request and the CSRF
token is checked. However, no modification to the server is made by git pulls,
as far as I know (actually I'm not sure why it uses a POST). The entire
response is handled by the git command, and the client side is usually the git
command running in the terminal, there's no session and no cookies (as far as I
know). So I'm just disabling CSRF token checking for this route.
The sharer and repo were being taken and used as is to check push permissions,
which is how it's supposed to be, *but* they were also being used as is to
build the repo path! So sharer and repo names that aren't all lowercase were
getting "No such repository" errors when trying to push.
I changed `RepoSpec` to hold `ShrIdent` and `RpIdent` instead of plain `Text`,
to avoid confusions like that and be clear and explicit about the
representation, and failures to find a repo after verifying it against the DB
are now logged as errors to help with debugging.
I hope this fixes the problem.
