GNU libreadline, and the libedit-based library shipped on stock OS-X
python, require different key-binding syntaxes to enable tab completion.
The previous commit to fix this (0977ef0) added both binding commands
Unfortunately when GNU libreadline is given the libedit-style
command (i.e. "bind ^I rl_complete"), it binds the letter "b" to a
non-existent command "ind", or something, and as a result the letter "b"
doesn't work anymore.
This patch uses the readline docstring to sense which flavor is
installed, and only runs the one binding command that's appropriate.
refs #37
The appveyor tests were failing because their VMs only have 127.0.0.1,
and stripping it out resulted in an empty hint list, which meant Transit
couldn't work at all.
With increased usage, I'm seeing a buildup of stale channels. Since the
channels aren't properly ephemeral yet (where they get closed as soon as
the last subscriber disconnects), clients which terminate without
calling close() tend to leave the channel lying around. We don't have
"persistent wormholes" yet, so channels should be much more ephemeral
than they currently are.
Apple's stock python doesn't use GNU libreadline, instead it uses BSD
libedit with a readline compatibility interface. The syntax to enable
tab completion is different for libedit. By including both bindings,
autocomplete should work on both flavors.
Closes#37. Thanks to @wsanchez for the catch and the fix.
(one displayed message per received welcome["motd"])
There's not much value in prohibiting the server from sending multiple
MOTD messages, and it would prevent us from using it to display a "your
client is using an old API, please upgrade" message after having already
sent a regular "please donate" MOTD message. (We could send a second
welcome message with ["error"] to kill the client, but ["motd"] is the
most convenient way to deliver a non-fatal warning).
This is an alias for the same host, so it's not really an incompatible
change. The new hostname is my personal domain, and seems a bit more
suitable for this service.
The reasoning is that this string is only ever likely to refer to the
version of the primary/initial client (the CLI application, written in
Python, that you get with "pip install magic-wormhole"). When there are
other implementations, with unrelated versions, they should obviously
not pay attention to a warning about the other implementation being out
of date.
This gives us room in the future to put other keys there, like one which
says we want to use Noise for the phase-message encryption instead of
our current HKDF scheme.
This will be useful for the upcoming "persistent wormhole" mode. A
client might send an allocation request, crash/terminate before
receiving a response, then restart, then re-send the request. If the
server sees a request with the same request_id a previous request, it
can return the same nameplate.
We'll need code changes on both sides to support this (nothing sends or
checks request_id yet), but this lands the schema change early to reduce
future disruption.
This will allow a future peer to figure out what transit modes we can
and cannot do, and thus avoid spinning up expensive modes that we won't
be able to use (e.g. WebRTC).
This enhances the ACK that wormhole-receive returns when it finishes
receiving all the data to be a dictionary. The dict includes the SHA256
hash of everything it received, and the sender checks this for a match
before declaring the transfer to be a success. This guards against data
being shuffled somehow during transit.
This better reflects the purpose of the message. Key confirmation is a
side-effect.
This patch only changes the "phase:" name and the key-derivation string.
A subsequent patch will modify the function and variable names to match.
The file-send protocol now sends a "hints-v1" key in the "transit"
message, which contains a list of JSON data structures that describe the
connection hints (a mixture of direct, tor, and relay hints, for now).
Previously the direct/tor and relay hints were sent in different keys,
and all were sent as strings like "tcp:hostname:1234" which had to be
parsed by the recipient.
The new structures include a version string, to make it easier to add
new types in the future. Transit logs+ignores hints it cannot
understand.
In the future, both sides should expect to receive "transit" messages at
any time, and they will add to the list of hints that they should try.
For now, each side only sends a single transit message, before they send
the offer (sender) or answer (receiver).
This moves us slowly towards a file-transfer protocol that exchanges
multiple messages, with a single offer (sender->receiver) and
answer (receiver->sender), and one or more connection hint messages (in
either direction) that appear gradually over time as connection
providers come online.
At present the protocol still expects the whole hint list to be present
in the offer/answer message.
This should enable forwards-compatibility with clients which send extra
data, like a pre-PAKE "auxdata" message that hints we should spin up a
tor client (because they can connect to it) while we're waiting for the
user to type in the wormhole code.
Previously the encryption key used for "phase messages" (anything sent
from one side to the other, protected by the shared PAKE-generated
session key) was derived just from the session key and the phase name.
The two sides would use the same key for their first message (but with
random, thus different, nonces).
This uses the sending side's string (a random 5-byte/10-character hex
string) in the derivation process too, so the two sides use different
keys. This gives us an easy way to reject reflected messages. We already
ignore messages that claim to use a "side" which matches our own (to
ignore server echoes of our own outbound messages). With this change, an
attacker (or the server) can't swap in the payload of an outbound
message, change the "side" to make it look like a peer message, and then
let us decrypt it correctly.
It also changes the derivation function to combine the phase and side
values safely. This didn't matter much when we only had one
externally-provided string, but with two, there's an opportunity for
format confusion if they were combined with a simple delimiter. Now we
hash both values before concatenating them.
This breaks interoperability with clients from before this change. They
will always get WrongPasswordErrors.
* add "released" ack-response for "release" command, to sync w.close()
* move websocket URL to root
* relayurl= should now be a "ws://" URL
* many tests pass (except for test_twisted, which will be removed, and
test_scripts)
* still moving integration tests from test_twisted to
test_wormhole.Wormholes
This made sense for ServerSentEvent channels (which has no purpose once
the channel was gone), but not so much for websockets. And it prevented
testing duplicate-close.
Pass in a handle and a pair of functions, rather than an object with two
well-known methods. This should make it easier to subscribe to multiple
channels in the future.
but only if the client is modern enough to include "id" in the message,
which lets us avoid sending acks to an 0.7.5 client (which would cause
them to abort, they don't like unrecognized server messages).
The acks let the client learn the server_rx time of messages that
terminate on the server, like "allocate" and "claim".
This improves the error behavior when --verify is used but there's a
WrongPasswordError: the mismatch is detected before the verifiers are
displayed or confirmation is requested.
It requires that the far end sends a "_confirm" message, which was
introduced in release 0.6.0. Use with older versions (if it doesn't
break for other reasons) will cause a hang.
This patch also deletes test_twisted.Basic.test_verifier_mismatch, since
both sides now detect this on their own. It changes
test_wrong_password() too, since we might now notice the error during
send_data (previously we'd only see it in get_data).
One downside is that we keep the wormhole channel allocated longer (we
have to finish the file transfer before we can deallocate it, which
could take a while for large files). Maybe we can fix this in the
future.
Also clean up test_scripts.PregeneratedCode:
* fetch results from both sides at the same time
* only check rc when using a subprocess, since the direct call doesn't
use rc=0 anymore
* no need to cancel the other side's Deferred when one errors
* provide more information if stderr was non-empty
And provide a close() that can live at the end of a Deferred chain, so
callers can do d.addBoth(w.close).
I like auto-close-on-error in general, but I'm removing it so I can
clean up the error-handling pathways. It will probably come back later.
The constraint is that it must be possible to wait on the return
Deferred that close() gives you (to synchronize tests, or keep the CLI
program running long enough to deallocate the channel) even if something
else (and error handler) called close() earlier. This will require
either a OneShotObserverList, or keeping a "deallocated" Deferred around
in case more callers want to wait on it later.
If we're closing because of an error, we need to sleep through the old
error, to be able to wait for the "deallocated" message. This might want
to be different: maybe clear the error first, or store the errors in a
list and sleep until a second error happens.
These were split out to make the blocking- and twisted- based
implementations share some code, but now that we're down to just
Twisted, it's clearer to merge them back in.
Hitting Control-C (which sends SIGINT) while we're waiting in the
readline-based input_code() function didn't shut down the process
properly: the reactor would wait for the readline thread to exit, which
wouldn't happen until it finished getting a code, which requires the
user to hit Return. I haven't found a good way to force the thread to
exit, or to synthetically inject a newline into stdin. So my compromise
is to tell the user that they need to hit Return to finish interrupting
the command.
See the _warn_readline() function for a list of other potential
approaches.
This uses a single TCP connection to the relay server for all
requests (although it probably uses a second one for the downstream
EventSource feed). This should squeeze out some of the round-trip times.
This adds an expected= argument to Connection.connectConsumer(), which
then returns a Deferred that fires when enough bytes have been written
to the consumer. It also adds Connection.writeToFile(), a helper method
that writes bytes to a filehandle.
I made the classic dataReceived() mistake, and exited the function after
delivering the first record. Keep at it until there are no complete
records left.
The previous commits improve test failures by dropping relay connections
at shutdown, and flunking a test quickly when one client fails but the
other one hangs.
If that doesn't work (say, some client has a time.sleep(), or other
stall that isn't affected by the relay shutdown), we'll be left with an
active thread holding that hanging client.
This patch adds a check to wormhole.test.common.ServerBase.tearDown that
looks for active threads, waits a second (after stopService), then
checks the threadpool again. If the threadpool is empty, everything is
fine. If not, it prints a message (to stdout) to inform the impatient
user why the test is probably hanging.
When test_scripts ran two clients at the same time, an error in one
could leave the other hanging (in a thread). One Deferred would errback,
the other would hang. Tests wait on one Deferred at a time, so if we're
unlucky and were waiting on the hanging Deferred (instead of the
erroring one), we'll wait forever, or at least until the default test
timeout of 180 seconds.
This adds an errback to notice when either client has errored, and
cancels the other Deferred, so it doesn't matter which one we wait upon
first.
'readline' is part of the python stdlib, so declaring a dependency on it
doesn't help. It doesn't exist on windows, and the pypi 'readline'
module doesn't work on windows. So instead, just attempt to import
readline, and if that fails, fall back to a non-completion flavor.
This ensures that we'll be ready for them. Previously there was a race
between us revealing the direct hints to the peer, and us setting the
transit key (thus allowing us to check inbound handshake requests). The
Transit instance didn't handle the race, causing errors to be thrown
when the other side connected quickly.
This ensures that we'll be ready for them. Previously there was a race
between us revealing the direct hints to the peer, and us setting the
transit key (thus allowing us to check inbound handshake requests). The
Transit instance handles this race (with an interlock on the transit
key), but it's still nicer to do it cleanly.
This exposed a new race in Transit, where the inbound connection would
complete before transit.connect() had been called. The previous commit
adds an interlock to wait for that too. Until this change, the transit
key lock was covering that one up.
Some tests failed to override --transit-helper, which meant they
intermittently talked to the real transit server (briefly, before
deciding the local+direct connection was better).
This limits the buffering to about 10MB (per connection*direction).
Previously, if the sender had more bandwidth than the receiver, the
transit relay would buffer the entire file. With this change, the sender
will be throttled to match the receiver's downstream speed.
The latest Twisted fixes the web.Agent code we need for proper async
support. There's still a daemonization bug that prevents 'wormhole
server start' from succeeding (it hangs).
I'm planning to leave non-EventSource "/get" in until after 0.6.0, then
remove it. I think it's cleaner for the logs to have the two
forms (EventSource and immediate) use different URLs.
In the twisted-style code, the close_on_error() decorator forces the
return value to be a Deferred, which is all wrong for internal uses of
derive_key() (verification string and confirmation message). It might be
useful to have a synchronous form of close_on_error(), but since the
actual close() is async, that's not very straightforward.
So for now, tolerate unclosed Wormhole objects when someone calls
derive_key() too early, or with a non-unicode type string.
This removes the 'allocations' table entirely, and cleans up the way we
prune old messages. This should make it easier to summarize each
connection (for usage stats) when it gets deallocated, as well as making
pruning more reliable.
Without this, old senders will throw a messy 404 traceback when talking
to a modern server.
Unfortunately 0.4.0 receivers don't make API calls in the right order,
so they throw a 404 before seeing our "you need to upgrade" message.
I was really confused about the Server-Sent Events syntax. This new one
is compatible with actual web browsers and the spec:
http://www.w3.org/TR/eventsource/
This requires a DB delete/recreate when upgrading. It changes the server
protocol, and app IDs, so clients cannot interoperate with each other
across this change, nor with the server. Flag day for everyone!
Now apps do not share channel IDs, so a lot of usage of app1 will not
cause the wormhole codes for app2 to get longer.
There are now three ways to invoke send:
* "wormhole send": ask for a text message
* "wormhole send FILENAME": send a file
* "wormhole send --text TEXT": send text message
This removes "side" and "msgnum" from the URLs, and puts them in a JSON
request body instead. The server now maintains a simple set of messages
for each channel-id, and isn't responsible for removing duplicates.
The client now fetches all messages, and just ignores everything it sent
itself. This removes the "reflection attack".
Deallocate now returns JSON, for consistency. DB and API use "phase" and
"body" instead of msgnum/message.
This changes the DB schema, so delete the DB before upgrading the server.
The "bytes % bytes" syntax only appeared on py3.5, so don't use it.
Updated travis to expect py3.4 works.
The twisted side is probably even more broken for py3.4 than it is for
py3.5.
This roughly parallels the way that blocking/eventsource.py and the pypi
"requests" modules work: the server can set the encoding (with
"Content-Type: text/event-stream; charset=utf-8"), and the EventSource
parser will decode accordingly. However eventsource_twisted.py *always*
returns unicode (on both py2/py3), even when the server hasn't set an
encoding. blocking/eventsource.py returns bytes (on py3, and str on py2)
when the server doesn't set an encoding.
In the future, eventsource_twisted.py should return bytes when the
server doesn't set an encoding.
eventsource_twisted.py includes an alternate approach that might be
necessary (a to_unicode() function instead of always using .decode), but
I won't be sure until enough of Twisted has been ported to allow the
EventSourceParser to be tested.
Also fix demo.py for python3.
