This ended up turning into a rewrite/refactor of the background page.
For best results, view this diff with `-w` to ignore whitespace. In
order to support retrying message decryption, possibly at a much later
time than the message is received, we now implement the following:
Each message is saved before it is decrypted. This generates a unique
message_id which is later used to update the database entry with the
message contents, or with any errors generated during processing.
When an IncomingIdentityKeyError occurs, we catch it and save it on the
model, then update the front end as usual. When the user clicks to
accept the new key, the error is replayed, which causes the message to
be decrypted and then passed to the background page for normal
processing.
ReplayableErrors make it easy for the frontend to handle identity key
errors by wrapping the necessary steps into one convenient little
replay() callback function.
The frontend remains agnostic to what those steps are. It just calls
replay() once the user has acknowledged the key change.
The protocol layer is responsible for registering the callbacks needed
by the IncomingIdentityKeyError and OutgoingIdentityKeyError.
This commit provides the javascript complement to
[WebSocket-Resources](https://github.com/WhisperSystems/WebSocket-Resources),
allowing us to use a bi-directional request-response framework over
websockets.
See websocket-resources.js and websocket-resources_test.js
for usage details.
Along the way I also factored the websocket keepalive and reconnect
logic into its own file/wrapper object.
Eliminates the global Whisper.Messages object and consolidates shared
send/receive logic in Whisper.Threads.
To the latter end, note that the decrypted array buffer on an attachment
pointer is now named data instead of decrypted, in order to match the
format of outgoing attachments presented by
FileReader.readAsArrayBuffers and let us use the same handler to base64
encode them.
We only depend on cryptojs for this webcrypto polyfill, so let Grunt
concatenate them into one file.
The reference in the getString helper isn't needed since we use the
built in string converters on CryptoJS's word arrays.
Firstly, don't initialize textsecure.nativclient unless the browser
supports it. The mimetype-check trick is hewn from nacl-common.js.
Secondly, nativeclient crypto functions will all automatically wait for
the module to load before sending messages, so we needn't register any
onload callbacks outside nativeclient.js. (Previously, if you wanted to
do crypto with native client, you would have to register a call back and
wait for the module to load.) Now that the native client crypto is
encapsulated behind a nice interface, it can handle all that
onload-callback jazz internally: if the module isn't loaded when you
call a nativeclient function, return a promise that waits for the load
callback, and eventually resolves with the result of the requested
command. This removes the need for textsecure.registerOnLoadCallback.
Finally, although native client has its quirks, it's significantly
faster than the alternative (emscripten compiled js), so this commit
also lets the crypto backend use native client opportunistically, if
it's available, falling back to js if not, which should make us
compatible with older versions of chrome and chromium.
The nativeclient.js module overrides
window.textsecure.registerOnLoadFunction with its own version. Otherwise
helpers will define a trivial placeholder for same.
The flag textsecure.NATIVE_CLIENT can be set anywhere ahead of
nativeclient.js, but is only acted on in nativeclient.js,
and crypto.js.
NB: this diff is best viewed with --ignore-whitespace
Distills crypto.js down to the hard cryptoey bones. It pulls from
webcrypto for aes and hmac, and from native client for curve25519 stuff
or potentially another object implementing the handful of needed
curve25519 functions.
Everything else formerly known as crypto, including session storage and
management, axolotl, etc.. is now protocol.js. The separation is not
quite perfect, but it's a big step.
nativeclient.js now enables talking to the native client module through
a high level interface as well as registering callbacks that will be
executed once the module is loaded. And it has tests!
Finally, this commit removes all references to the "testing_only"
object, preferring to run tests on textsecure.crypto instead.
TypedArray.prototype.set doesn't handle ArrayBuffers correctly (it
writes all zeros). Instead, wrap each ArrayBuffer in a typed array
for concatenation.
Parse attachment ids out of the attachment pointer url and return them
as strings because the copy parsed by JSON suffers a loss of precision.
Convert them to and from the format expected by the protobuf using
facilities from decodeIO.Long.
DRY up protobuf declarations and move to a slightly briefer naming
convention.
Also dropped some ArrayBuffer -> string conversions as
ProtoBuf.js handles ArrayBuffers just fine, and in fact, more
efficiently than strings.
Finally, dropped the btoa() wrappers, because that incurs an extra
string -> string conversion before the protobuf's internal string ->
array buffer conversion. In lieu of btoa, we can simply pass in the
optional string encoding argument to the protobuf's decode method,
which in these cases should be 'binary'.
Related: #17
An exception is thrown when protobuf tries to encode a number as the
group id, which is declared to have type 'bytes'.
Fix by make it an ArrayBuffer instead, and increase the length to 16,
which is what the Android client uses:
c632b32ff8/src/org/thoughtcrime/securesms/database/GroupDatabase.java (L222)
Previously, if calling createNewGroup with an undefined groupId,
no groupId was generated.
This occurred because no entry for "group" + undefined exists in
localStorage, which caused this code to think undefined was a
valid group id.
Fixed by adding `|| groupId == undefined` to the while clause.
Also decoupled the groupId collision check for clarity.
