Forgot to mention:
The gpu's AMD Radeon R7 260X (Bonaire XTX) PCI ID: 1002:6658

Please attach the /etc/X11/xorg.conf file.

Created attachment 130327 [details]
xorg.conf that resulted in the crash

Huh, it also seems to work after deleting the: Option "DRI" "3"

Created attachment 130373 [details] [review]
Bail from drmmode_set_mode_major if it's called too early

(In reply to timon37 from comment #3)
> Huh, it also seems to work after deleting the: Option "DRI" "3"

That must be luck, DRI3 should be enabled by default anyway.

I can't seem to reproduce the problem. It looks like maybe crtc->enabled isn't always initialized to FALSE in drmmode_xf86crtc_resize at this point (valgrind doesn't complain about any access to uninitialized memory though).

Does this patch avoid the problem for you?

Patch didn't help. Actually it initially didn't compile for me (AMDGPUWindowExposures_oneshot was undefined) but I added a prototype for it.

Indeed DRI3 didn't seem to affect it on further tries (not sure what happened originally).

But when I comment out the "Virtual 4096 4096" line it consistently works, and if then uncomment it it crashes.

To clarify the crash is because pScreen->root is 0. It doesn't crash without rotation because that's not queried in the else case (I verified that it was also 0 without screen rotation).

And I'm guessing it's not 0 when the Virtual setting isn't used.
Alternatively the code never reaches that point, I can try to verify that or anything else you want. Though hopefully you'll be able to repro it now?

Comment on attachment 130373 [details] [review]
Bail from drmmode_set_mode_major if it's called too early

(In reply to timon37 from comment #5)
> Patch didn't help. Actually it initially didn't compile for me
> (AMDGPUWindowExposures_oneshot was undefined)

Oops, sorry, I thought I'd at least compile-tested the patch, but I obviously hadn't.

> but I added a prototype for it.

Did you also remove the static keyword from the function declaration src/amdgpu_kms.c? If not, that might explain why the patch didn't work.


> But when I comment out the "Virtual 4096 4096" line it consistently works,
> and if then uncomment it it crashes.

Thanks, this allowed me to reproduce the problem. https://patchwork.freedesktop.org/patch/145922/ fixes it for me.

BTW, does the Virtual line serve any purpose? Xorg seems to resize the screen to fit the rotated output anyway for me.

(In reply to Michel Dänzer from comment #6)
> Did you also remove the static keyword from the function declaration
> src/amdgpu_kms.c? If not, that might explain why the patch didn't work.

Hmm, I thought I did. But I guess it probably would only fail on load saying it can't find the symbol, and I probably stupidly didn't properly verify what the output was...

Really sorry about the sloppiness on my side.

> BTW, does the Virtual line serve any purpose? Xorg seems to resize the
> screen to fit the rotated output anyway for me.

Probably was need at some point a long time ago, and was simply leftover.
Right now doesn't seem to be necessary.

Thanks for the report, fixed in Git master:

https://cgit.freedesktop.org/xorg/driver/xf86-video-amdgpu/commit/?id=981bac185cfd74ae50dffc28f57cf34623a9595f