(I found "user request" in one place, and fixed that. I am not
currently going to try to unify "control message" and "meta message"
since both terms are misleading and we already have TODOs to try to
merge them into a third better term.)
This new function combines "sending a message" and "accepting
replies in a stream" into a single call, so that there is no gap
between when the message is sent and the replies are available.
There are a number of compromises here, in order to avoid API
proliferation. I've tried to contain them as best I can.
See comments for additional design discussion.
Now, the MetaCellHandler is responsible for consuming the messages
it gets, and reporting status to whatever task is waiting for a
status message.
Additionally, the MetaCellHandler can decide to remain installed or
shut down the circuit after a successful message. (Previously, it
could only uninstall itself on success and kill the circuit on
failure.)
These changes will enable MetaCellHandlers to be used as the basis
for handling more kinds of message.
(There is some moved and reformatted code here; you may want to
review it with `git {diff or show} --color-moved -b`.)
When I was trying to add HS support to these layers, I found I could
add a new variant to the `Host` enum but everything would still
compile even though I hadn't written the necessary implementation!
This method is a liability: when using it, one inevitably writes such
latent bugs.
Doing so doesn't seem like a good idea. It might even be some kind of
leak?
Found because I added a variant to `address::Host` for hidden
services, and noticed that the resolve code still compiled.
Doing this means that any attempt to use a read-only store would
crash as soon as it found that the consensus was usable.
It seems that this bug was introduced at some point doing all the
dirmgr refactors we did over the past year. Perhaps there should be
a test for running with a read-only store.
Fixes#779
Previously, if somebody wrote this code, an attacker could easily
use it to cause an OOM panic:
```
let n = r.take_u64();
let items: Vec<Foo> = r.extract_n(n as usize)?;
```
The first line of defense here is not to write protocols like that:
we don't actually _have_ any 32-bit counters in our protocol
AFAICT.
The second line of defense is to pre-check `n` for reasonableness
before calling `extract_n`.
Here we add a third line of defense: whereas previously we would do
`Vec::with_capacity(n)` in `extract_n`, we now allocate an initial
capacity of `min(n, r.remaining())`. This ensures that the size of
the allocation can't exceed the remaining length of the message,
which (for our cell types at least) should prevent it from
overflowing or running OOM.
These Arcs are all "downward", referencing items from layers lower in
the stack. So they don't cause cycles.
There was going to be a cycle involving the `OnionConnector` upcall
trait, but we have just abolished that.
Abolish CircMgr::get_or_launch_onion_client and everything to support
it. We have decided that `.onion` diversion ccan't/shouldn't occur in
tor-circmgr. Probably, it should occur much higher up - arti-client
maybe - since it will sometimes need ambient authority (KS_hsc_*).
Now all knowledge of HS connections is in tor-hsclient. This
gets rid of a layering inversion and the trait needed for tor-circmgr
to do the upcall to tor-hsclient.
Ian Jackson
Nick Mathewson
Ian Jackson
Dimitris Apostolou