small tooltip windows have transparency issues

[J. Max Mena]

Some websites have a small pop-up that will display information, such as time since a post, information about a link, etc. etc.. In Xpra certain browsers will cause them to display a cross-hatching of colored lines before moving the mouse causes the proper pop-up to appear, with the lines just at the very bottom of the pop-up.

Interestingly, the pop-ups are only a problem in Chrome, not Firefox. The Firefox pop-ups are a solid white color, rather than the black (possibly partially transparent?) color in Chrome.

Currently running Xpra 0.11.x r5209 in Windows 7 64-bit. I'll attach a few screenshots to explain further.

[J. Max Mena]

the bugged pop-up in Chrome.

[J. Max Mena]

The pop-up working properly in Firefox

[J. Max Mena]

The pop-up working on interfacelift in Chrome

[J. Max Mena]

The pop-up working on twitch in chrome

[J. Max Mena]

Pop up working semi-properly on bmwusa after moving the mouse

[J. Max Mena]

pop-up showing colors after moving mouse on twitch

[J. Max Mena]

I forgot to mention in the original post, but in some websites the pop-ups behave as expected in Chrome, and in others it shows up bugged. I'm not sure yet why.

[Antoine Martin]

I don't think the colour (black vs white) is relevant here: browsers are free to style alt tags as they wish, it may just be how those two browsers do it. You can verify this by running the browser natively (without involving xpra).

• does this also happen on OSX?
• does this also happen with opengl turned off?
• does the bug occur with the default transparency test applications:
  • tests.xpra.test_apps.transparent_colors
  • tests.xpra.test_apps.transparent_window
• does this happen with transparency turned off on the server:

  XPRA_ALPHA=0 xpra start ...
  

• xpra info|grep window captured when the tooltip window is showing could be useful too
• does this happen with all encodings that support transparency? (all the png encodings, webp and rgb) Does it happen with encodings that do not support transparency? (jpeg)

[Antoine Martin]

Raising the priority: I'm waiting to hear on this bug to release 0.11.1

[J. Max Mena]

To answer your questions:

• Yes, the same thing occurs in OSX.

Interestingly instead of turning normal when you move the mouse like it does in the screenshot, the text just shifts over a couple of pixels to the right.

• Yes, it behaves the exact same, both Windows and OSX with and without opengl

• The default transparency tests look as they should, both with and without opengl

• With XPRA_ALPHA=0, it behaves the exact same, both Windows and OSX

• I retested with all the available encodings, and all the png encodings work fine. raw RGB + zlib/lz4 has the bug. jpeg works fine.

How do I use xpra info|grep window?

[J. Max Mena]

Okay, Alex showed me how to use it. I'll attach the relevant .txt.

[J. Max Mena]

the file generated after running "xpra info | grep window > xprainfo.txt"

[Antoine Martin]

Thanks for the details, so this is not transparency or GL related, only raw RGB has this problem.
From the xpra info log, the tooltip window is (I'm only showing the most relevant bits):

window[34].window-type=('TOOLTIP',)
window[34].total_frames[rgb24]=1
window[34].total_pixels[rgb24]=4002
window[34].size=(174, 23)
window[34].override-redirect=True
window[34].has-alpha=True
window[34].encoding=h264
window[34].encoding.last_used=rgb24

Human readable summary: the window is a TOOLTIP override-redirect window and we only send a single pixel update (174x23=4002) which is using rgb24 because it is small enough, and not rgb32 because the client probably does not support it (win32 I assume). The global encoding seems to be set to h264, but we never create a video context for small transient windows like these.

And judging by the screenshot, it looks like the rgb24 data is being processed as if it was rgb32. This is strange but at least now I know where to look.

[Antoine Martin]

Updated the bug title: it isn't with ALT tags, at least not in chrome, but with title tags (on both images and links).
A good page to test this is: ​links-testing.
I get the tooltip to show with chrome on tests: 13, 11, 10 (image only), 9, 8 (image only), 6, 5, 4, 3, 2

[Antoine Martin]

simplest html test case possible

[Antoine Martin]

Fixed in r5262 (r5263 for v0.11.x branch) by not sending alpha: we tell the server we don't handle RGBA on win32 or OSX, this means we strip the alpha server side.
The reason why we didn't do this normally is that:

• the bandwidth savings are tiny (since this is used for very small windows only)
• all pixels are stored as 32bpp anyway, so the empty alpha gets re-added client side anyway

In the process of debugging, I cleaned up the GL codepath: r5261

---

I am keeping this ticket open because we should be able to handle RGBA pixels client side, even when we don't actually paint the alpha channel.
May be related to #521

[alas]

I am still seeing this issue on chrome (and lazarus) with osx client 0.12.0 r5531 and fedora 19 server 0.12.0 r 5531. Interestingly though, with firefox the tooltips are behaving as expected.

With chrome:

window[56].window-type=('TOOLTIP',)
window[56].total_frames[png]=1
window[56].total_pixels[png]=3772
window[56].size=(164, 23)
window[56].override-redirect=True
window[56].has-alpha=True
window[56].encoding=png
window[56].encoding.last_used=png

With firefox:

window[57].window-type=('TOOLTIP',)
window[57].total_frames[png]=1
window[57].total_pixels[png]=3744
window[57].size=(156, 24)
window[57].override-redirect=True
window[57].has-alpha=False
window[57].encoding=png
window[57].encoding.last_used=png

[Antoine Martin]

As expected this is because of the alpha channel: firefox doesn't use alpha but chrome does (look for has-alpha).

[Antoine Martin]

Lots of fixes in this area: r5624, r5625, r5626, r5627 and some more in ticket:521#comment:4

So, the core issue was that we were converting BGRX to RGBX but still sending it as rgb24, and it needs to be rgb32 instead.

Most of the important fixes have been backported.

afarr: Feel free to close this ticket if everything works as expected, and if you have time please also check #521

[alas]

Bad news... with 0.12.0 r5692 osx client and 0.12.0 r5711 fedora 19 server I'm still seeing the buggy Alpha dsiplay on tooltips with chrome and lazarus. (I'll attach a screenshot.)

Are the fixes meant to fix that? Is there some way to override the alpha for tooltips?

The interesting (but unsurprising now) bits of xpra info | grep window:

window[26].has-alpha=True
window[26].window-type=('TOOLTIP',)
window[26].total_frames[png]=2
window[26].total_pixels[png]=12282
window[26].override-redirect=True
window[26].size=(267, 23)
window[26].override-redirect=True
window[26].encoding=png

[alas]

r5692 osx client chrome/lazarus tooltips still misbehaving

[Antoine Martin]

AFAICT, this is a different bug with the exact same symptom, as it only occurs on OSX now. win32 is now fine and it wasn't before.

Which means that I will probably have to blacklist OSX for 32-bit RGB support. But I am now blocked by #529. Oh joy.

[Antoine Martin]

Help... with r5760 on server and client, I just cannot reproduce this at all now. Which is odd because I haven't made any changes that should have fixed this (and I've just skimmed the list of changes since r5700), and I was seeing the problem 6 days ago!?

I also hacked my server to remove other encodings completely (to make sure only rgb would get used), tried other encodings, etc. No luck!?

Can you still reproduce?

[Antoine Martin]

Got it at last! (by trying various archived beta builds to narrow things down), counterintuitively this is a real bug that has been there for much longer but only came to the surface because packaging problems on OSX prevented the inclusion of PIL (Python Imaging Library) in the disk image (and recent build improvements fixed that - which made the bug disappear from view).
Without PIL, the code reverts to the slower GDK image loading code... which did not deal with rgb32 and assumed the data could only be rgb24... r5771 fixes that.

In order to test this alternative codepath, you need to run a client without PIL installed (for example by removing PIL from the installation image) or using the environment variable: XPRA_USE_PIL=0 xpra attach ...

Feel free to just test that things look good instead, and close this ticket.

[alas]

Ok, testing with XPRA_USE_PIL=0 the tooltips display as expected with all browsers (including with opengl=on).

Closing.

[migration script]

this ticket has been moved to: ​https://github.com/Xpra-org/xpra/issues/500