For the purposes of debugging running nodes, it's useful to know which
threads are doing what. This allows tools like `htop` to reveal which
threads are running.
It turns out it's possible to create a deadlock when using
`child.wait()` with `Stdio::piped()`, if we don't read the output while
the child process is running. This is because the pipe has an internal
kernel buffer that will eventually fill up, causing the child process to
block on writing to stdout and never exit.
To remedy this situation, we read the stdout in a thread while the child
process is running.
This change updates to the latest radicle-git-ext, which required a
few changes within the radicle family of crates.
One set of changes is that the radicle-git-ext crate subsumes
git-commit, git-trailers, and git-ref-format -- so all those imports
go through radicle-git-ext.
The commit code requires that parent `Oid`s point to commits, which
made some tests fail. One of the reasons for these failures is that
the patch tests used fabricated `Oid`s and so the verification would
fail. This is fixed by allowing the test context code to create
`PatchRequest`s that create valid commits.
The other reason for failure was that the identity `Action` passed an
`Oid` that pointed to a blob. This is fixed by removing that code and
making a note of it in documentation.
The final major change was moving the Refspec type into git-ref-format
and removing the AsRefspecs trait. The trait was not required and
could its usage could be replaced by simpler code.
Signed-off-by: Fintan Halpenny <fintan.halpenny@gmail.com>
X-Clacks-Overhead: GNU Terry Pratchett
On macOS, shutting down a socket returns an error if the socket is
already closed. We don't consider that a problem, as it just returns
`Ok(())` on Linux.
We fix the failing fetch tests by:
1. Not verifying our own refs, unless we're cloning, since we're
otherwise not fetching our own refs.
2. Always force-fetching `sigrefs` from remotes into the staging copy.
3. Making sure that sigref updates are fast-forward before transfering
the remote into the production copy.
Given the scenario of 3 peers, A, B, and C each having a project R. If
peer A has B's fork, but C does not, then when peer A attempts to
fetch from C the fetch will be rejected since Git will fail when a
refspec that is asked for does not exist on the remote side.
To avoid this, the fetch logic is changed so that the client first calls
ls-remote to the remote side, which results in the references that the
remote has which the client is interested in. The follow-up fetch then
uses these refs as the refspecs -- which should succeed unless there
was a race for a deletion on the remote side.
The rest of the verification logic stays the same, so the storage
should still be in a working state before transferring from the staged
repository to the production repository.
Signed-off-by: Fintan Halpenny <fintan.halpenny@gmail.com>
X-Clacks-Overhead: GNU Terry Pratchett
Instead of returning the namespaces that we fetched with,
we return the remotes that were actually fetched from.
This can differ if some trusted peers were not available
on the remote node.
Previously, if that was the case, it would cause errors
since the remotes were looked up and that lookup failed.
If `git-fetch` returns a non-zero code on exit, we keep going in the
case of an existing repository.
The reason is that `git-fetch` will return an error code if any ref
is rejected during the fetch. But this can happen if someone
pushes conflicting `rad/sigrefs` branches for example, and the node
already has a copy of that branch.
Instead of parsing the worker requests and responses, which was causing issues
around packfile transmission, we use threads and pass the bytes
transparently.
To improve the reliability and flexibility of the protocol, we introduce
multiplexing over peer connections. This involves a new `Frame` type
that carries a stream-id and payload.
Three stream types are made available:
1. Control
2. Gossip
3. Git
This change brings the following improvements:
* Removed need to queue fetch requests
* Removed need to queue gossip messages
* Removed need for `Fetch` and `FetchOk` messages
* Service doesn't need to know about inbound fetches
* Removed one round-trip for fetch negotiation
* Removed special `done` git packet
* Removed session logic and state around Git/Fetch protocols
* Removed code around upgrading/downgrading transport
* Worker in responder mode is able to process any number of fetches
* Connections support any number of concurrent fetches
We had to introduce a few extra things however to make it all work:
* A `VarInt` type for variable-length integers, since we want the frames
to be lightweight.
* A custom "tunnel" implementation, since we couldn't use the existing
one anymore.
Overall the change removes more complexity than it adds, while improving
the protocol along the way.
The previous fetching logic had some pitfalls regarding its logic:
* Fetched 'rad' references could end up not being verifiable and leaving
the storage in a corrupt state
* Forced pushes would be prohibited, since they would be rejected by
the client side due to the refspecs being used.
The new approach is to stage a fetch in 2 steps. The first step is to
fetch the relevant `rad/id` and `rad/sigrefs` for verification. For
all remotes that are verified, fetch all references listed in their
respective `rad/sigrefs`, using the `+` marker in the refspec,
allowing to update for any force pushes -- note that this becomes a
safe operation since the `sigrefs` are signed by the remote that
created them.
To prevent unverifiable 'rad' references from polluting storage, all
fetch negotiations are done in a staging repository set up in a
temporary directory. All successful fetches are then transferred,
using the 'file://' protocol, from the temporary directory into the
radicle storage.
Signed-off-by: Fintan Halpenny <fintan.halpenny@gmail.com>
X-Clacks-Overhead: GNU Terry Pratchett
Better handle case where the connection to the daemon fails, in the
responder scenario.
Instead of trying to read the `done` packet once, we loop until we've
read it. This handles a scenario where the uploader fails before the
git protocol is initiated, and the fetcher is still sending git packets.
If there is an unexpected error, eg. `BrokenPipe` in the upload process
process of the worker, we may have more than four bytes to read from the
remote socket. This change ensures that we don't attempt to read more
into a buffer of size `4`.
It's possible that the daemon or stream closes the git connection
without reading the final `done` packet. If this is the case, we
try to read it after either of the connections is closed.
This may potentially mitigate the issue where the node then receives
the `done` packet while in gossip mode, and disconnects the remote
for misbehavior.
The fetch implementation will potentially fetch the local operator's
remote, which is unnecessary and could result in deleted data.
Fix this by threading the local operator's public key and add an
ignore refspec to the fetch arguments, i.e.
^refs/namespaces/<local>/*
The implementation also strictly fetches all or one namespace. This is
improved upon by introducing a new variant for specifying a set of
namespaces.
Signed-off-by: Fintan Halpenny <fintan.halpenny@gmail.com>
X-Clacks-Overhead: GNU Terry Pratchett
A fetch is a client-server interaction and so there are two sides to
this interaction. These sides are generally referred to as receive --
the client side -- and upload -- the server side.
The `Fetch` type indicated whether it was doing a receive or upload by
using a field `initiated: bool`. However, this can be confusing when
you pair it with the `namespaces: Namespaces` field, since only the
receive side should be sending what refspecs it wants and the upload
side responds with those matching refspecs.
A better way to represent this is using an enum for which direction
the fetch is being considered, where the receive side contains the
Namespaces. To borrow from Noise, the variant names are 'Initiator'
for the client and 'Responder' for the server.
The resulting code then makes decisions based on what variant of the
enum was passed.
Signed-off-by: Fintan Halpenny <fintan.halpenny@gmail.com>
X-Clacks-Overhead: GNU Terry Pratchett
Before this change, the identity document would be loaded for the
COB signer. This meant that a fork would be needed for that signer.
After this change, it's no longer necessary, since the identity doc
head is computed from available remotes.
While making those changes, it was also apparent that some of the
identity-related functions had bad names and error types, this
was fixed as well.
* Read git request line before starting daemon, so that we don't leave
git data in the socket read buffer in case of an error connecting to
the daemon.
* Send a response on the fetch channel even if we disconnect the remote.
* Don't send `done` packet if there's an error; it won't be expected.
Alexis Sellier
Fintan Halpenny
Alexis Sellier