Most (all?) new APIs will require a "secure context", which generally
means served over TLS. We can expect crashes because of missing
functions if this requirement isn't fulfilled, so try to warn the user.
Fixed eslint warnings
Fixing tests that failed
Added unit tests for ARD authentication
Fixed an issue with the ARD rfb version number in the unit tests
Fixed issue with username/password lengths
Username and password lengths are now capped at 63 characters each. Improved code for sign bit on public key bytes.
UTF Encoder username and password before packing it
Change UTF encoding to encode the username and password before packing it to prevent it from being expanded beyond the allowed size. Public key is truncated to proper key length.
Replaced forge with web crypto for ARD authentication
Changed the way in which the async methods are handled, added unit tests to verify ARD encryption output.
Update .eslintignore
If we increase the remote screen size from the server in such a way that
it no longer fits the browser window, the browser will probably want to
show scrollbars. The same happens if you enable 'clipping' while the
remote is larger than the browser window. These scrollbars do, in turn,
decrease the available space in the browser window. This causes our
ResizeObserver to trigger.
If the resize observation triggers a requestRemoteResize() we will
overwrite the size and request a new one just because scrollbars have
appeared. We don't want that.
We can save the expected client size after resizing, and then compare
the current client size with the expected one. If there is no change
compared to the expected size, we shouldn't send the request.
Fixes issue #1616.
This is no longer an issue on Google Chrome, tested on Chrome 96 on
Fedora 34, Windows 10, macOS 12 and Android 12. It is however an issue
on Safari on macOS 12.
Without this workaround we get scrollbars when making the browser
window smaller, despite remote resize being enabled.
Fixed eslint warnings
Improved memory usage of zrle decoding. Added unit tests for zrle decoding.
Added support for ZRLE encoding
Fixed eslint warnings
Reverted allowIncomplete changes to Inflator
Fixed failing tests for zrle decoder.
Chrome scrolls the view to show as much as possible of the canvas when
we call focus(), which is likely not the desired behaviour.
This also exposes the ability to pass on future options when focusing
the RFB object manually.
We need to do this in order to safely attach to existing WebSocket
objects. There may be message events already pending so we must set up
our event handlers before returning.
This means we will now throw errors instead of generating "disconnect"
events on problems as the caller no longer has the opportunity to set up
event handlers.
This might have been the correct approach from the start as it mimics
how e.g. the WebSocket constructor works.
We don't know if the caller is prepared to receive those events right
now as normally they would get them on a fresh new stack later. We also
can't delay delivery since then we might deliver the event after any
pending "message" events.
Better to push the problem one layer up to the caller, which now needs
to be more aware of the state of the WebSocket object it is trying to
use.
This is an error that presents itself with RTCDataChannel's, I suspect this could not
happen with a pre-existing WebSocket.
If the remote connection creates a data channel then the local (VNC) side gets a channel
created callback. It may also be the case that in that very same tick the socket is also
opened and buffered data received. This meant that (in my tests) about 1/3 of the time
noVNC would fail to respond to the initial message from the server because it was received
and subsequently not handled during that initial tick.
Also made the documentation reflect this new behaviour and document the existing behaviour.
Most places that check for the presence / absence of credentials compare
them against `undefined`, except the one for Plain authentication.
This change makes the very last place to use the same pattern (instead
of checking for falsiness) for consistency. Additionally, there are ways
to configure PAM to accept empty passwords, so it's possible for a user
to legitimately send an empty string as password.
This work is originally by Ryan Castner <castner.rr@gmail.com> and
submitted as a PR here https://github.com/novnc/noVNC/pull/1362
Architecturally it is much the same except it doesn't rename a lot
of variables to make this more reviewable. It also avoids unrelated
changes such as replacing .onclose with an event listener, which
caused numerous test failures.
It also adds in ppoffice's fix to initialise the buffers.
Like the original author I don't have enough time available to
refactor this project to the new style event listeners.
Review cleanup for RTCDataChannel support (see below)
* More descriptive error when url or channel not set.
* Moved websocket property check to WebSock.
This had unintended consequences in the tests that required some
fixup. Mostly due to some tests not always passing FakeWebsocket.
FakeWebsocket also needs to set the listeners to null to be compatible
with what is in thw browser and expected by the property check code.
The property check code now also takes into account class prototypes
for test compatibility.
* Removed unreachable code.
* Reverted comment.
* Cleanup raw channel reference in rfb on websock close.
* Use readyState to check whether a socket is open rather than assuming.
* Updated RFB constructor documentation
Removed an unused boolean passed to attach
Some people have longer passwords than 256 characters (hooray for
password managers!). Server implementations also allow longer passwords:
TigerVNC allows up to 1024 characters.
Windows behaves very oddly for some Japanese IM keys in that it won't
send a key release event when the key is released. In some keys it never
sends the event, and in some cases it sends the release as the key is
pressed the subsequent time.
Windows doesn't give us stable symbols for a bunch of Japanese IM keys,
instead alternating between two symbols. This state is not synchronised
with the IM running on the remote server so to have stable behaviour we
have to collapse these multiple symbols in to a single keysym.
These are very pointless for the server to send, but not a violation of
the protocol so we need to be able to handle them. We've seen this
happen in real world scenarios a few times.
This timer might fire after the Cursor object has detached from a DOM
element, causing crashes. This will likely not happen in real scenarios,
but the tests are quick enough to trigger this.
The new gesture detection code will always prevent the default behaviour
of touchstart, so this check no longer works properly. We might want to
add something similar to GestureHandler in the future, but let's wait
and see what use cases are requested.
With the new gestures we will simulate the cursor being in a different
location than any of the touch points. This is a bit too complex for the
Cursor class, so let's just explicitly tell it where we want the cursor
rendered.
The previous value made the detection too sensitive and it was very
difficult to scroll precisely. A value of 50 pixels should give similar
behaviour to systems that don't do fine grained scrolling.
This isn't really expected behaviour from a user, i.e. that an extremely
small wheel movement still gives a large scroll event in the remote application.
Add several single and multitouch gestures to simulate various mouse
actions that would otherwise be impossible to perform.
This replaces the old system where you could select which mouse button
a single touch would generate.
This is what the browser wants so it avoids having to spend time
converting everything. Unfortunately it usually means the server instead
needs to convert it for us, but we assume it has more power than we do.
The code that used these were removed in the following commits:
* 9ff86fb718 (RFB._mouse_arr)
* bb6965f2e6 (old_requestAnimationFrame)
* 490d471c53 (Display._c_forceCanvas)
This allows using TigerVNC server with PAM authentication (e.g. agains
LDAP or other extensible authentication mechanisms)
Tested with TigerVNC server (Xvnc -SecurityTypes Plain -PlainUsers '*')
Should not break anything else, this method is tried last when all
other fail.
Tested in Firefox 74 and Chromium 80
If too much text is copied in the session, String.fromCharCode.apply()
would crash in Safari on macOS and Chrome on Linux. This commit fixes
this issue by avoiding apply() altogether. Also added test to cover this
issue.
As a rule, instead of hard-coding a behavior on specific platforms we
should do dynamic detection.
This commit moves away from always hiding scrollbars on Android and iOS
and instead detects the rendered width of scrollbars in the browser.
Internet Explorer seems to flag images as loaded prematurely, which
can result in rendering bugs. We can detect this by looking at the
dimensions though.
Pierre Ossman
pdlan
Paul Dumais
Samuel Mannehed
Benjamin Leeds
Tim Stableford
lhchavez
Quentin Dreyer
Uwe Klatt
Filip Stedronsky
Andrey Trebler
Alex Tanskanen
Niko Lehto
Juanjo Diaz