add new resolutions to Xdummy on the fly - dummy driver needs randr support added?

[Antoine Martin]

Instead of finding the nearest match, we should just add the modeline, assign it to the output then select it.
This would allow us to match the client's screen size exactly.

(somewhat blocked by #10)

[Antoine Martin]

this patch would add a new mode to Xdummy so we can use it... but does not work

[Antoine Martin]

After ​asking xorg-devel, it would seem that the dummy driver does not support randr - or at least not fully.

The patch above attempts to replicate xrandr's command line functionality and even includes its own modeline calculator, so from xpra's side it is pretty much ready.

[Antoine Martin]

Not heard back, and less critical now that the default config contains many of the typical resolutions found on desktops and even android devices.

[onlyjob]

refreshed patch to apply over 0.6.1

[Antoine Martin]

Closed #186 (duplicate - some info there)

[Antoine Martin]

Asked xorg dev again: ​RFC: dummy driver RandR improvements?

Some other links:

• ​Fake Xinerama for X devs
• ​Fake xinerama on github
• ​PATCH: make Xephyr resizeable

[Antoine Martin]

Forgot to update this ticket:

• we now have wiki/FakeXinerama support
• better DPI support (see #163)

[Antoine Martin]

Nicolas Boichat has just posted X11 patches to do exactly this!: ​PATCH dummy: Add support for custom resolutions (RandR 1.2)

Maybe this will fix #728
We'll just have to ensure this works with DPI (#163) - that we deal with #732 if possible, see also #349

[Antoine Martin]

Looks like the discussion of what to do for randr 1.2 support is ongoing: ​Cleanups and RandR 1.2 support, with different implementations suggested.

And so this will have to wait until this gets merged in one form or another, and maybe we only have to package it then which would be great (assuming that the new code can be built against older servers that is...)

[Antoine Martin]

Maybe this can help us: ​RandR 1.5 Brings Monitor Objects & Tile Support For X.Org

[Antoine Martin]

Re-post: ​xvfb: add randr support (v2)

I think we should update the dummy driver with the same changes.

[Antoine Martin]

In one of the responses to the patch above: ​http://lists.x.org/archives/xorg-devel/2015-June/046666.html links to: ​PATCH xf86-video-dummy 0/6: Cleanups and RandR 1.2 support - which is a better approach IMO.

[Antoine Martin]

See also #969, in particular the ramdac fix in ticket:969#comment:3

[Antoine Martin]

dummy randr 1.2: patch 1 Remove pointless empty functions

[Antoine Martin]

dummy randr 1.2: patch 2 Delete XV stuff

[Antoine Martin]

dummy randr 1.2: patch 3 Delete dPtr->screenSaver

[Antoine Martin]

dummy randr 1.2: patch 4 Remove DGA support

[Antoine Martin]

dummy randr 1.2: patch 5 Get rid of dPtr->FBBase

[Antoine Martin]

dummy randr 1.2: patch 6 Support RandR 1.2

[Antoine Martin]

just tried the patch series above and hit a snag: ​my xorg-devel reply

[Antoine Martin]

See also: ​MST monitors and the proposal for RandR 1.5 which would have the concept of "monitors" (without requiring us to have virtual CRTCs attached to them apparently)

[Antoine Martin]

No movement on this upstream, re-scheduling.

[Antoine Martin]

The randr v2 patch has been merged: ​http://lists.x.org/archives/xorg-devel/2015-December/048228.html.
Commit: ​3d68d1f26709ecb5ce22a9baa0d3d8162574ed6a.
So either we figure out how to use the same code in Xdummy, or we switch back to Xvfb with newer versions?
Will have to test this new stuff: re-scheduling.

[Antoine Martin]

re-scheduling

[Antoine Martin]

We're not alone in needing this: ​xf86-video-dummy: resize to exact resolution from Chrome Remote Desktop developer Erik Jensen.

[Antoine Martin]

Milestone renamed

[Antoine Martin]

I don't have time for this, new links:

• ​re: xf86-video-dummy: resize to exact resolution: Xvfb is as maintained as any other server we build
• ​re: xf86-video-dummy: resize to exact resolution: Which approach seems more likely to gain traction? Without a champion for either one, neither. :)

[Antoine Martin]

for testing the latest upstream patches for dummy

[Antoine Martin]

Recent discussion: ​remove dead code in dummy driver: In 2014 I had also modified the dummy driver while working with Teradici Corporation to support not only arbitrary pixel dimensions, but also multiple monitors. The latter feature might not make sense to some folks, but if you have a server-side Xserver mapped to a hardware thin client sitting across a network, which has multiple physical monitors attached, you want the Xserver to be configured in the exact same manner as the tin client. Even though there is just a virtual framebuffer in main memory, X applications need to know the number/size/location of monitors so that toolbars are placed properly, windows are fullscreen'ed properly, etc. And when the user moves from one hardware thin client to another, which may have a different monitor configuration, the Xserver session needs to change to that configuration.

[Antoine Martin]

This will also make #1008 redundant.

[Antoine Martin]

Raising again.

This is the commit that added randr to Xvfb: ​vfb: add randr support (v2)

[Antoine Martin]

Done for Xvfb in r16994.
To use it, switch to Xvfb in browser/xpra/trunk/src/etc/xpra/conf.d/55_server_x11.conf.in.
The server should then create new resolutions for every client, the "-d screen,randr" debug output looks like this (shown here for 1080p, but this works for any reasonable resolution):

dpi=120, dpi.x=120, dpi.y=120, double_click_time=400, double_click_distance=(-1, -1), \
    antialias={'hinting': True, 'enabled': True, 'orientation': 'NONE', 'contrast': 1000, 'hintstyle': 'hintslight'}, \
    cursor_size=24
Python/Gtk2 Linux Fedora 26 TwentySix x11 client version 2.2-r16976 64-bit
 connected from 'desktop' as 'antoine' - 'Antoine Martin'
 mmap is enabled using 256MB area in /run/user/1000/xpra/xpra.7Ap0oy.mmap
 client root window size is 3840x2160 with 1 display:
  :1.0 (1016x572 mm - DPI: 96x95) workarea: 3840x2126 at 0x34
    monitor 1 (708x398 mm - DPI: 137x137)
maximum client resolution is 3840x2160 (current server resolution is 5760x2560)
set_screen_size(3840, 2160)
set_screen_size(3840, 2160) xdpi=120, ydpi=120
set_dpi(120, 120)
add_screen_size(3840, 2160)
Modeline 3840x2160 230 3840 4108 4492 5068 2160 2161 2164 2293
XRRCreateMode returned 0x275
screen_resources: crtcs=1, outputs=1, modes=2
adding mode 0x275 to output 0x3d
randr_added_sizes=[(3840, 2160)]
using XRRSetScreenConfigAndRate with 3840x2160
XRRSetScreenSize(0x555a8dbc8240, 0x25c, 3840, 2160, 813, 457)

Still TODO:

• ideally, replace the synchronous sleep with an asynchronous handler: we have to wait for the X11 server to reconfigure its outputs, this makes the connection setup slower than it used to be...
• maybe start removing old resolutions after we have added a few hundred, which can happen quite quickly
• update the HTML5 client to remove the black borders

[Antoine Martin]

r16996: improvements, client fixes (inc html5)

[Antoine Martin]

@maxmylyn: ready for some testing as it is.

[J. Max Mena]

With a Trunk r17074 Fedora 25 server and client, I played around with this for a few minutes:

I edited my conf file to uncomment the default Xvfb command for Ubuntu, so it read:

Xvfb +extension Composite -nolisten tcp -noreset \
        -auth $XAUTHORITY \
        -screen 0 5760x2560x24+32

And played around with it for a bit. I made sure to turn on -d screen,randr and saw I got the relevant prints. Nothing seems to have been broken, and resizing the HTML5 client seems to work fine. But, I didn't do much other than some xterms and glxgears. Is there anything in particular we should be looking at?

Of note:

I booted my local test bot up, but I need to reconfigure its display settings to work headless. Once I do that, I can do a run-through of the automated tests with Xvfb, maybe that'll catch something I wouldn't.

(Thanks for mentioning that, Smo)

[J. Max Mena]

NOTE:

test bot up and running. The tests are failing after the first run because the server does not get shut down properly. This will require editing the test_measure_perf.py file.

[Antoine Martin]

• r17075 adds GLX to Xvfb command line, not sure if that makes any difference. There are other switches that may impact performance: +iglx / -iglx, etc.
• r17080 removes old modes so we don't end up with thousands of modes (which could cause problems with some applications - known to cause crashes with some gnome applications for example), currently set to keep a maximum of 32 resolutions, for testing this can be changed using the env var XPRA_RANDR_MAX_NEW_MODES=VALUE
• when comparing the performance of Xvfb vs Xdummy, make sure that you don't use virtualgl unless you also intend to use virtualgl in production
• x11perf may not show much difference between the two (see attachment/wiki/Xdummy/Xvfb-vs-Xorg-x11perf.txt), but other tests might

Note: rather than editing the xvfb= option, you can just build with --without-Xdummy (easier)

[Antoine Martin]

Lowering priority: we have support for resizing via xvfb (see comment:25).
See also #1467