We want each ID to have a unique form every time it is given out,
so that you can't use ID==ID to check whether Object==Object. (See
discussions leading to #848.)
We'd also like the form of object IDs to be a little annoying to
analyze, to discourage people from writing programs that depends on
their particular format. (We are reserving the right to change the
format whenever we want.)
We _don't_ want to use any cryptography here (yet), lest somebody
think that this is an actual security mechanism. (This isn't for
security; it's for encouraging developers to treat IDs as opaque.)
With that in mind, we now lightly obfuscate our generational indices
before returning them.
Per discussion referenced at #848, we want each operation that
returns a strong object ID to return a new, distinct strong ID.
Note that we no longer need to put strong and weak references in the
same arena; we can clean this code up a lot down the road.
Now we generate object IDs that we can parse. This is about to be
obsolete once we change how we generate objects and their IDs for #848,
but we may as well start from a working state.
This does the following:
- Gives every crate a `full`.
- Cause every `full` to depend on `full` from the lower-level
crates.
- Makes every feature listed _directly_ in `experimental` depend
on `__is_experimental`.
Now there can be one of each, since we want references and handles
to be conceptually separate.
(The alternative would be to say that an operation either "returns a
handle" or "returns a reference (which may become a handle) unless a
handle already exists.")
To me, "Session" suggests that we're authenticated, when we are
not necessarily authenticated. Also, we may eventually want to have
some kind of persistent session object; if we do, then we'll want
Connections to be separate.
(This is the correct capabilities-based behavior. For now it will only
work if the TorClient uses a PreferredRuntime, but with luck we will
find a solution for #837 soon.)
Now there's a module in `arti` that runs the loop for an RPC
listener. The part of the old `listener` module that made
the framed connections is now part of the `Session` object.
There is now yet another a temporary location for the pipe; we
should pick something better. At least now it's configurable.
In the future, this will probably hold more data as well, like a
TorClient and some configuration info.
The TorClient will present an issue; I've made comments about that.
Closes#820
Right now, this lets us say whether the method was unrecognized or
whether the parameter type was incorrect.
We'll use this to enforce correct method names later on.
(I have to add another `inventory` here, since the `typetag`
maintainer does not want to expose this functionality: see
https://github.com/dtolnay/typetag/issues/57#issuecomment-1506106290
)
Previously the main loop received updates via a `mpsc::channel`, and
final responses via a `futures::unordered`. This could lead to
final responses being transmitted to the user before the updates
were all flushed.
Now all of the responses are sent to the main loop via the same channel,
and they can't get out-of-sequence.
Closes#817 and (IMO) simplifies the code a bit.
Now instead of hoping that buggy clients will detect a magic `id`,
we can simply tell them that they will get no `id` at all. If they
can't handle that case, no major harm is done: the connection will
get closed anyway.
Since we're serializing everything in this format, let's enforce it.
With this change, we can no longer cram arbitrary junk into an
RPC error, so we have to clean up our handling of cancelled requests.
This is a bit big, but it's not that _complicated_.
The idea here is that we use serde's "untagged" enum facility
when parsing our `Request`s, such that if parsing as a `Request`
fails, we parse as an `InvalidRequest` and try to report
what the problem was exactly.
This lets us determine the ID of a request (if it had one),
so we can report that ID in our error message. We can also
recover from a much broader variety of errors.
We now also conform with the spec reporting errors about
completely wrong json, requests without IDs, and so on.
Even though json-rpc uses "result" to mean "a successful return value
from a method", we can't: Rust's `Result` type is so pervasive
that confusion would be inevitable.
Per our design, every connection starts out unauthenticated, and
needs one authenticate command to become authenticated.
Right now the only authentication type is "This is a unix named
socket where everybody who can connect has permission."
This code uses some kludges (discussed with Ian previously and
hopefully well documented here) to get a type-identifier for each
type in a const context. It then defines a macro to declare a
type-erased versions of a concrete implementation functions, and
register those implementations to be called later.
We will probably want to tweak a bunch of this code as we move ahead.
Ordinarily you can cancel a future just by dropping it, but we'll
want the ability to cancel futures that we no longer own (because we
gave them to a `FuturesUnordered`).