[Closed] Mesa/LLVM/Qutebrowser issue after profile switch

[Weezer]

Greetings!

New to forum here! I will say thanks in advance for your patience in the likely case I mess this up...

I have been fighting with an issue after switching my profile to 23.0. After updating the profile on three laptops, Qutebrowser began consuming just about all resources possible grinding them to a halt. I have narrowed down the culprit to Mesa and LLVM. Prior to the profile switch, I had successfully "de-LLVM'd" all three machines about a year and a half ago. Everything was working swimmingly and I no longer had to have LLVM installed on my machines.
After the profile switch, I had to re-enable LLVM for Mesa which restored performance to where it was prior.

Now, I am not necessarily against keeping LLVM but I am stuck figuring out as to why I now need it?

The reason I disabled LLVM was simply because all my laptops are "potatoes", to put it lightly, and I try to avoid anything with long emerge times when possible. I know, ironic, because my favorite web browser uses qtwebengine which I'm embarassed to even mention merge times for that.

Profile switch was: amd64/17.1 (stable) -> amd64/23.0/split-usr (stable)

Two laptops are are unstable (~amd64) and one is stable (amd64) and I run Xmonad as my WM. All three laptops differ in hardware variations so not polluting this post with all the specifications right now but they all use integrated Intel graphics of the i9*5 variety.

I have researched every angle including Mesa, LLVM, Qutebrowser, Qtwebengine, Qt6 and am now "lost".

It boils down to this: when I use Mesa without LLVM (bad); Mesa with LLVM (good). I don't even know what else to put in here and am just hoping for a "nudge" at this point...

TIA for any suggestions/input you may have!

Closing - Don't want to leave this hanging. Thanks.

[Hu]

Welcome to the forum. Could you elaborate on what resources are consumed? Do you observe that all CPU cores are running full out? If so, what process(es) are consuming the CPU time? Do you observe that the system exhausts available RAM? Can you reproduce the failure with a simpler program? In particular, does glxgears exhibit bad behavior under the same circumstances that Qutebrowser behaves badly? Aside from Qutebrowser, are there any other programs which exhibit problems?

[Weezer]

Hi Hu,

Thanks for jumping in here.

glxgears definitely performs much better with LLVM:

1110 frames in 5.0 seconds = 221.990 FPS
1119 frames in 5.0 seconds = 223.757 FPS
1141 frames in 5.0 seconds = 228.173 FPS

As opposed to without:

208 frames in 5.0 seconds = 41.429 FPS
208 frames in 5.0 seconds = 41.494 FPS
213 frames in 5.0 seconds = 42.449 FPS

When I launch qutebrowser (no LLVM), my CPU cores (all two of them) alternate between 0 and around 98%. When one is at 98%, the other is at 0 and vise versa. Fans immediately go full tilt.

As far as RAM goes, qutebrowser process itself uses around 998MB and then there are multiple QtWebEngineProc processes that launch with one at >150MB and then a couple more at 44MB. It does not exhaust the RAM however. CPU seems to be the major issue IMHO.

Basically I have to quit qutebrowser at that point as to not "melt" my laptop.

When I have LLVM in use, qutebrowser process uses around 337MB of RAM and the QtWebEngineProc process are basically the same as above. However, my CPU does not max out and fans don't even run.

I've no other issues to speak of and being fairly old hardware, I can't really do much that requires too much "power" as one might say.

I just don't quite know why the profile switch would have been the difference? It was the first thing I noticed after the re-emerge of all packages and I made zero use flag changes etc. This was the same on all my laptops. I thought maybe I had somehow not removed LLVM but one of my builds was more recent and never had LLVM installed/enabled... Thinking out loud here.

I'm happy to try or provide anything further. Just to reiterate, this is more of a why? type of issue as everything works fine as is (though I would prefer to not have to carry LLVM).

[Hu]

Those results suggest to me that your bad case is trying to do all the OpenGL processing on the CPU, and the good case delegates it to your GPU. (I saw similar behavior with a game when I accidentally denied it use of the GPU; it was so slow that it was unplayable and it kept one CPU core maxed out just trying to render itself. Restoring GPU access made it very playable, and CPU usage went down considerably.) We had at least one other thread recently where someone made a profile switch, and also updated many installed packages to newer versions. One of the updates (in that case, Docker) introduced a behavior change that prompted the user to post. Perhaps you also experienced an upgrade, rather than an in-place rebuild, when you did the profile switch, and your upgraded package now requires GPU assistance to work well.

[Weezer]

I agree with everything you mentioned.

I'm religious about updating my systems every day and I'm sure I did prior to the emptytree rebuild, but it certainly is possible nonetheless.

I will keep this open for a bit just to see if anyone might have other ideas/angles to look at.

Thanks Hu.