With high DPI systems we can end up with a container with a size that
is not an integer number of CSS pixels. Make sure we can handle those
cases by allowing a fractional size for the output canvas. Framebuffer
size and viewport coordinates are still restricted to integer dimensions
though.
Based on initial patch by Alexander E. Patrakov.
Always use the shorthand notation if the function is a method of an object or class `{ foo() { ... } }` or `class bar { foo() { ... } }`
unless it's a callback in which case you a fat arrow function should be used `{ cb: () => { ... } }`
Some browsers don't support custom cursors, and there are cases
where the browsers refuse to show the cursor. Handle both of these
cases by letting the browser render the cursor via a floating
canvas.
This allows us to support a local cursor at all times.
Siemens' touch panels support Tight authentication as well as NOTUNNEL,
but they fail to advertise the latter. Work around this issue by detecting
a Siemens device (through their custom tunnel types) and assume NOTUNNEL
support even if not advertised.
Pasting clipboard texts that were larger than 10240 bytes didnt work and
caused a crash in noVNC. This commit fixes the crash and adds handling
for sending large clipboard texts. Fixes issue #1065.
Firefox no longer sends keyup events properly for the Alt keys. Try
to sniff out the state of the Alt key by monitoring other events that
include its state.
Try to properly detect the fake CtrlL+AltR sequence Windows sends
when pressing AltGr. This allows us to send more accurate key
events over to the server.
Tight PNG rects cannot use the basic compression variants, and PNG
cannot be used in a standard Tight rect.
This is a partial revert of 3e8b26a based on better understanding
of the encoding.
The function removeChild throws an exception when the element specified
doesn't exist. This can happen when opening the websocket throws an
exception during connect. Fixes issue #979
The API allowed strings to be passed from the RFB module to the
application using the disconnect reason. This caused problems since
the application didn't have control over translations for these
strings.
Most of the information being passed using this string was very
technical and not helpful to the end user. One exception to this was
the security result information regarding for example authentication
failures. The protocol allows the VNC server to pass a string
directly to the user in the security result.
So the disconnect reason is replaced by a boolean saying if the
disconnection was clean or not. And for the security result information
from the server, a new event has been added.
Instead of exposing all the internal connection states, the RFB module
will now only send events on connect and on disconnect. This makes it
simpler for the application and gets rid of the double events that were
being sent on disconnect (previously updatestate and disconnect).
An RFB object represents a single connection so it doesn't make
sense to have one without it trying to connect right away. Matches
the behaviour of other APIs, e.g. WebSocket.
Use normal properties with JavaScript setters and getters instead of
our homegrown stuff.
This also changes the properties to follow normal naming conventions.
Sort things by category, and organise everything in the same place.
We don't support reuse of RFB objects so we can safely init everything
in the constructor.
The previous method of retaining focus didn't work reliably when
the RFB object tried to move the focus to the canvas. Add a setting
to control "focus on click" behaviour instead of letting them try
to fight it out.
Give the canvas proper focus handling. This avoids messy logic that
needs to disable and enable event handling when we want to interact
with other UI elements.
It also makes sure we can properly inhibit the browser from triggering
local actions on key presses.
This restores basic support for Intel AMT servers. They refuse clients
that request more than 16 bits per pixels, so implement a fallback
in just the "Raw" encoding.
The VNC protocol can't handle different deltas or speeds for a mouse
wheel event. When using a device that sends a lot of small mouse wheel
events, instead of fewer larger steps, the effect was that mouse wheel
scrolling was way to sensitive.
This patch looks at the delta of wheel events and doesn't send events
until the combined delta has passed a threshold. Single events that
doesn't pass the threshold get sent after a timeout in order to not
loose any events.
Fixes#577.
Servers will assume that a scan code is present if this message type
is used, so fall back to the standard key event message if we don't
know the scan code.
IE and Edge have some corner cases (e.g. Ctrl+key) where we get
insufficient information in the keydown event, and we never get
a keypress event. Try to make a guess of the key in those cases.
iOS sends decent key down events, but junk key up events when a
hardware keyboard is used. This confuses the key tracking as a
corresponding release is then never detected. To work around this
we'll treat the hardware keyboard like the virtual ones and send
the key release right away.
It doesn't need to be this general as the issue is mostly about
Windows. Also use the same modifier shuffle that RealVNC and
TigerVNC uses to get macOS working well.
The fields provided cannot tell us if it is the left or right
version of the key that's pressed, so they are inherently unreliable.
It is also not a huge problem in practice as we'll get in sync on
the next press or release of the modifier.
Look up keys that are independent of layout and state first,
followed by keys that are only mild variations in layouts.
This is more robust as there might be multiple physical keys
generating the same symbols, and Keysyms don't map directly to
Unicode in all cases.
At the same time switch over to using the modern, standardised
'code' field for lookup.
Use the more modern 'key' field, and remove some legacy fallbacks
that are no longer required. This also removes the "stall" mechanism
as it is not needed with current browsers.
Setting a style to null does restore it in FF, Chrome, Safari and Edge.
But it does not work in Internet Explorer. The proper way to restore to
default values is to set it to the empty string. This works in all
browsers. Fixes issue #808.
If down is false we can't just toggle the current internal state, even
though this is correct most of the time. There are cases where we don't
get the down event and thus won't have a correct internal state. For
example, when clicking in the session after using the clipboard
textarea.
If down is false, we always want to send a mouse event with the button
in 'up'-state.
This commit restructures many of the ES6 modules, splitting them
up to actual export multiple functions instead of a single object.
It also splits up Util into multiple sub-modules, to make it easier
to maintain.
Finally, localisation is renamed to localization.
The only remaining user of WebUtil.load_scripts was for loading
localisation. Instead, we now load the localization information
over XHR as a JSON blob.
The browsers' support for Microsoft's cursor format is a bit spotty,
so use the more common PNG format instead. This also allows us to
use a Canvas to generate the image, rather than coding it by hand.
This removes the special comment part of the ES6 module syntax,
opting to enable ES6 module syntax by default.
It also appends `.js` to all import paths to better support in-browser
loading.
`app/ui.js` had an incorrect import path which caused issues
when using the ES6 and/or CommonJS builds of noVNC.
`core/util.js` had a non-strict-compatible declaration of a variable
without a `let` or `var` (it now uses `let`).
This fixes both issues.
It mostly dealt with scrolling which we don't use. It also made mistakes
in some cases. Remove it and compute the coordinates directly in the
calling code.
The native WebSocket is in a much better position to do queue
management than us. We also failed to check the return value and
set up a timer, causing stalls.
This gets us in sync with websockify as of 40238b00.
We don't have to check for _display or context here since this is a
private function which is never called under such circumstances. This
solves problems caused by display.get_context() which was previously
removed in e549ae074f.
The old default was to ask for the maximum compression level. This
is against the recommendations in libvncserver/tight.c due to excessive
CPU load. It also causes Vino 3.8.1 (still shipped with Ubuntu 16.04
LTS) to prefer the blurry JPEG compression too much - e.g. red text on
the default background in MATE terminal becomes almost unreadable.
The new default is the recommended compression level for low-color
workloads, according to libvncserver source. Also, it is the maximum
compression level that doesn't trigger the Vino bug with red text in
most cases.
Fixes issue #737.
Do all rendering to a hidden canvas and then copy over the finished
frame to the visible canvas once everything is done. This simplifies
things and solves some bugs as we can retain a copy of the entire
frame buffer.
Supports server configurations that might prefer wss:// connections
on the default port, then proxies them through the web server to the
VNC server.
This proxy configuration is helpful on servers using self-signed
certificates. Accessing the https://host/vnc_auto.html page and
adding an exception for that host is sufficient to also satisfy the
wss://host/ request, unlike requests to wss://host:port/ which
may require an extra certificate exception.
RFB's _fail function logs the error, disconnects the session and sets
disconnect_reason. The disconnect_reason is upon disconnection sent to
the user interface. It is thus not suitable for including error details
that aren't user friendly. The idea is that you will look in the
browser console for a full log with details of the error.
RFB._fail() can be called at any time in any state, it is not certain
that we will get a close event on the socket since the socket might not
be open. This caused us to hit the disconnect timeout in such cases.
Fixes issue #678
For example, previously if the user typed in illegal characters in the
port field, no error would be displayed in the interface and the page
would stop at "connecting".
There is a specific event for when an image has finished loading,
so trigger on that rather than polling. The polling interval of
requestAnimationFrame() can also be very large.
The callers don't need to concern themselves with how images are
rendered, so hide the details behind the API. This also avoids
exposing the render queue.
Mouse wheel event handling has now been standardised and has broad
support. Use this event rather than the older, non-standard ones.
At the same time fix up support for horisontal mouse wheel events.
The states 'loaded', 'failed' and 'fatal' were all variations of the
'disconnected' state. Removing these states allows us to get rid of
many ugly workarounds and special cases. Error messages to the UI can
now instead be delivered via a new onDisconnected callback.
Names such as 'disconnect' and 'disconnected' are inconsistent in the
way that one describes an action and the other a state. The state that
was called 'normal' didn't fit in with the others because the other
names describe a connection state. The new names are: 'disconnecting',
'connecting' and 'connected'
We already have a callback mechanism for this, so let's use that.
Adds an optional parameter 'msg' to the callback.
Fixes vnc_auto.html (#646) which was broken after
4e0c36dda7
According to the RFB protocol, when in TIGHT auth negotation,
if the client receives a sub-auth count of zero, it should proceed
as if the `None` sub-auth type had already been selected (and not
send a message selecting that type).
Closes#564
Pierre Ossman
Alexander E. Patrakov
Samuel Mannehed
Juanjo Diaz
Vlastimil Sadilek
Leslie Qi Wang
Solly Ross
Juanjo Diaz
Josiah White
shaneharris
Solly Ross
Lee-Feng
Kyle Russell
Pierre Ossman
Giannis Kosmas
Loek Janssen
Asad Saeed
zarmhast
Dmitry Sankevich