Filing good bug reports for MESA / radeonsi, AMDGPU? (FS-UAE fences bug)

[lem79]

Hi all,

I think I've pretty much found a bug with mesa/amdgpu that has been introduced recently via mesa-git. I want to know how to file a good bug report so it has a chance of getting fixed.

Hardware: AMD FX-8350, Asus Strix Radeon 380X 4Gb, Asus MG279 27" 2560x1440 144Hz IPS monitor.
Software: Ubuntu 16.04 plus Padoka PPA, fully updated.

I'm using FS-UAE (an Amiga emulator by Frode Solheim), which uses OpenGL to draw its window contents. During this month, I noticed that the fs-uae window would lock up upon starting, then the fs-uae process would use 24% CPU (at whatever pstate the CPU was in - it didn't cause the CPU to clock up), and be unresponsive until a kill -9 was issued.

fs-uae allows for a couple of different rendering modes. I narrowed this problem down to any render mode using fences. Any mode that doesn't use fences works fine.

Here's my partial bug report (including apitrace details) on the EAB FS-UAE dev thread: http://eab.abime.net/showpost.php?p=...postcount=1146

I'm wondering if the recent changes to mesa-git (this month) to do with fences has introduced this bug? Either way, I'd like to be able to submit a good quality bug, so pointers in the right direction would be really helpful!

On the topic of reporting bugs .. how about the amdgpu kernel driver? On the same hardware above, 144Hz video modes do not work. I must use 120Hz. 144Hz renders about 1/5th of the screen, then it's like the signal gets lost and every second/third line is drawn, then the rest of the screen is black. It's like the monitor can't follow the sync of the card. I'd like to report that somewhere too.

Cheers

[droste]

Just copy what you've written here to the bug report on bugs.freedesktop.org (product "Mesa"), attach the apitrace and that's it :-)
Same goes for the other bug, but the product is "DRI" for kernel bugs.

[droste]

And btw. you could bisect mesa for the offending commit. That would be even better.

From a quick look there's some fence related commits from start of august:

https://cgit.freedesktop.org/mesa/mesa/commit/?id=54272e18a682c8b82d4a86b2c07b51c303d8cead

https://cgit.freedesktop.org/mesa/mesa/commit/?id=1cc95a1255d5ba6d44e48360c6ad7e5b97c6493c

https://cgit.freedesktop.org/mesa/mesa/commit/?id=3f100b77f9a80828fbb3651ea013f9dd3913d047

So maybe starting with c6043e7d54fc48771856a50235f9141b711151f3 if you don't have a known good version might be a good idea :-)

[dungeon]

Who knows, your first bug might be even related to the second... but since you said it is regression then bisection result would make good bug report even better.

[lem79]

Thanks guys. Yes, dungeon, indeed a bisection would be excellent .. and probably pretty easy to do since the problem was only introduced this month. I searched mesa git for the keyword "fence", and there's a fair bit of work just this month on it. I don't have the time to do it right now though (I'd have to learn how to do it first ).

For those who are interested (or who might land here via searching), the FS-UAE hang / fence sync bug is here: https://bugs.freedesktop.org/show_bug.cgi?id=97461

And the 144Hz display corruption bug had already been reported here: https://bugs.freedesktop.org/show_bug.cgi?id=96868

[dungeon]

Yeah as suggestion, if not checking latest kernel/mes/llvm code first as there thing might be already fixed i would probably try to run an app with LIBGL_DRI3_DISABLE=1 variable first and if it is not that, would checked out some code before and after Marek's unflushed fence optimization for Bioshock... that sounds like candidates on first tought to me.

But bisect will show it anyway as you have easy repro case if in mesa bisection is usually easy, but llvm and kernel might not be... at least because more time needed for those.