amdgpu (RAVEN) X server crash: "Failed to make 1x1x32bpp GBM bo", bo conflict (probably a kernel or Mesa bug?)
This happened with an Athlon 3000G APU (RAVEN), with kernel 5.8.9, X.org X server 1.20.9, xf86-video-amdgpu commit 442efe73. Dual monitors, one DVI, one HDMI, identical resolutions (1080x1050), effective screen 0 res 3360x1050. After a few days of normal usage (almost no 3D other than the usual glamor usage), I popped up calibre's ebook-viewer, and got an instant black screen, though the machine was still responsive over the network so I could shut down cleanly. There was a spray of tens of thousands of lines of this in the X server log:
[419221.702] (EE) AMDGPU(0): Failed to make 1x1x32bpp GBM bo
At the same time, the kernel log filled up with fifty thousand lines or so of this:
[418985.683944] amdgpu 0000:0a:00.0: amdgpu: bo 00000000a83d33d2 va 0x080010176e-0x080010176f conflict with 0x080010176e-0x0800101770
[418985.684129] amdgpu 0000:0a:00.0: amdgpu: bo 00000000a83d33d2 va 0x080010176e-0x080010176f conflict with 0x080010176e-0x0800101770
[418985.690683] amdgpu 0000:0a:00.0: amdgpu: bo 00000000c4a6faf4 va 0x080010176e-0x080010176f conflict with 0x080010176e-0x0800101770
(different bo address every time, the other two addresses were unchanged.)
This is probably a Mesa or amdgpu kernel bug, but I have no idea which it might be, nor really any idea how to replicate this. (I mean, possibly just rerunning the ebook-viewer would do it, but if I try now and it works I lose this bug report :) )