This commit ensures that input arrays are copied to new storage when
they are pushed onto the render queue. This ensures that they are
not overwritten before they are eventually used.
Fixes#571
If a token is already present in the path, the new variable
is ignored. In order to properly manipulate the path,
a new method, `WebUtil.injectParamIfMissing` was introduced.
Fixes#536
[@directxman12: fix up path manipulation logic]
This commit removes some unused variables in the TIGHT server init
section of the server init handler, replacing them with documenting
comments and calls to rQskipBytes.
Fixes#440.
the `copyWithin` function of typed arrays has performance issues
in some versions of Chromium, and doesn't doesn't have enough
of a performance impact to justify leaving it enabled.
Any config variable like host, port, password, token may be
specified either in the query string (like now), or in the URL hash
fragment. In case a given variable is present in both, the value in the
fragment takes precedence. Supplying variables in the fragment avoids
leaking them to the web server hosting the noVNC viewer HTML.
Previously, we used a fixed chunkSize of 100KiB for Pako's output
buffer. Using a hardcoded size caused issues, since Pako would assume
we wanted to use multiple chunks, and we didn't deal with this. Now,
`Inflator#inflate()` takes a new `expected` argument, which indicates
the expected output size. If this is bigger than the current chunkSize,
Inflator allocates a new output buffer that's big enough to hold the
output.
Fixes#531
For performance reasons, the `handle_tight` function skips the
use of the receive queue API and uses the raw receive queue directly.
Because of the way that typed array receive queue gets reused, this
introduced the potential for buffer over-reads.
To address this, a new function, `rQwhole`, was introduced.
`rQwhole` simply returns a new view into the receive queue that
starts at 0 and ends at the current recorded end of the queue.
`handle_tight` now makes use of this function.
Fixes#522
There was a bug caused by 38781d931e
which prevented color map look-ups sent by rfb.js from working properly,
since display.js expected a single-item array, and rfb.js sent just them
item value itself (a number) instead. This fixes that, and tweaks the
corresponding test to match that behavior.
This commit skips object creation for the render queue when not
needed. Instead of pushing an object onto the queue, and then
immediately running the result, you call the function directly.
Then, if the render queue is not empty, an object is created and
pushed onto the queue. Otherwise, the functionality is just run
directly.
This commit converts the send queue to use typed arrays, and converts
message creation functions in 'rfb.js' to create messages directly into
the socket's send queue. This commit also removes the separate mouse array,
which is no longer needed.
Creating lots of small objects frequently can drastically decrease
performance. This commit introduces three fixes which avoid this:
- Use a preallocated palette and indexed-to-rgb destination Typed Array
(the destination typed array is currently allocated at `4 * width *
height`).
- Inline `getTightCLength`, which returned a two-item array.
- Pass RGBX data directly in a Typed Array to the Display, which
avoids an extra loop, and only creates a new Typed Array View,
instead of a whole new ArrayBuffer.
**This commit removes Base64 (and Flash) support**
This commit converts websock.js to used Typed Arrays for the
receive queue (and tweaks rfb.js to ensure that it continues
to function, since only Firefox implements
`%TypedArray%.prototype.slice`). Base64 support was removed
to simplify code paths, and pave the way for using Typed Arrays
for the send queue as well.
This provides two advantages: first, we allocate a buffer ahead
of time, meaning the browser doesn't have to do any work dynamically
increasing the receive queue size. Secondly, we are now able to pass
around Typed Array Views (e.g. `Uint8Array`), which are lightweight, and
don't involve copying.
The downside is that we initially allocate more memory -- we currently
start out with 4 MiB, and then automatically double when it looks like
the amount unused is getting to small.
The commit also explicitly adds a check to the compacting logic that
avoids calling the copy functions if `_rQlen === _rQi`.
Previously, if an error was thrown from the Display constructor
in the RFB constructor, we would attempt to use `RFB#updateState`
to handle this. However, `RFB#updateState` attempts to close
the WebSocket connection, which doesn't exist at this point.
In the constructor, it's probably just better to raise an exception
instead (making sure to clean up anything relevant).
Fixes#460
In e543525faa, we switched to creating
a new RFB object on disconnect. This caused issues, however, since
any errors were only displayed briefly before the new "loaded" text
was displayed instead.
Now, we create the RFB object on connect. This essentially removes
the usefulness of the "loaded" state, but prevents the aforementioned
problem.
To facilitate this, the code which does detection of cursor URI support
was moved from this Display constructor (which now calls the new
function) into its own function, `Util.browserSupportsCursorURIs()`.
Fixes#467
* Renamed and reworked fbuClip to clippingDisplay
* Added tests for clippingDisplay
* Use the a noVNC_container which covers the entire page to get the full size
(Fixes#463)
* Added maxWidth and maxHeight to the canvas which can limit the viewport size
* Only show either the canvas or the logo, hide one when the other is shown
* Always center the canvas (previously it was only centered when not clipping)
* Removed iOS specific "position-fixed" fixes and start calling setBarPosition
on every resize
* Removed the noVNC_screen_pad
* Added better error handling in ExtendedDesktopSize
* Added helper function to share code with DesktopSize
* Update test.rfb.js to only check for error handling if we were the
ones requesting the resize
This commit adds two new addition scaling options. Both options do
local scaling. The first "Local Scaling", does both upscaling and
downscaling. The second option, "Local Downscaling", only downscales.
This is based on work by @mightypenguin (with an additional bug
reported by @glazik12).
samhed
Jonathan Bennett
Solly Ross
Solly Ross
Miguel Xavier Penha Neto
Ganesh Varadarajan
Daniël van de Giessen
Fabian Zaremba