regression bisected RX580 GPU crash chipset overheating (kernel 6.8 and later)

Brief summary of the problem:

It is impossible to use the discrete GPU as primary render with recent kernel in a decade old all AMD computer. Trying results in the system becoming unresponsive, then a black screen and after a few seconds an corrupted image in the screen. The problem occur with the Manjaro kernel as well as the upstream compiled kernels.

All 6.6+ kernels are affected (6.13.9, 6.14 e LTS kernel 6.12)

All old LTS kernels are not affected (6.6, 6.1, 5.15, 5.10 e 5.4), with a smooth function of the machine not showing any issue at all.

The tip of Linus's mainline tree is also affected by the issue.

Using only the integrated GPU (AMD RS880 [Radeon HD 4250], from Phenom II x6 1090T processor), result in no issue with any of the kernels compiled of the versions supported (last and stable ones) in kernel.org. Changing to the discrete GPU as primary render on bios (the system is a decade old and still uses the bios not being compatible the UEFI system) result in Black screen after the login screen on KDE Plasma. Keeping the integrated GPU in the bios and offloading the graphic applications to the discrete GPU (RX580) with DRI_PRIME=1 results in the application not appearing in the screen and after some minutes the screen start alternating between black screen and flickered screen (corrupted image), with unresponsive system. The moment you offload it to the discrete GPU you can’t access anymore the KDE launcher (menu). Both Wayland and X11 show the same behavior. Trying Steam games with Steam Proton results in the same behavior: game won't show in the screen and the system become unresponsive. When starting the game sometimes the Xwayland crash menu (when using wayland) or X crash menu (when using X11) appears. Starting the glxinfo or eglinfo with the command DRI_PRIME=1 glxinfo -B or DRI_PRIME=1 eglinfo -B result in the same crash behavior after some minutes. The crash occurs for a year but only after changing, this month, the power supply and the case for one with easy access I could verify that when using the discrete GPU (bios or DRI_PRIME don’t matter) the heatsink over the chipset (AMD 880G on M5A88-M Asus motherboard) starts to get hot. When it get too hot to withstand with the bare hand the system starts to crash showing the black screen and the corrupted image. The motherboard and the CPU sensors show always temperatures below 43ºC. The heatsink hotness do not occur with 6.6 LTS kernel and older ones.

Using the upstream git kernel I could bisect the problem to the commit “[466a7d115326ece682c2b60d1c77d1d0b9010b4f] drm/amd: Use the first non-dGPU PCI device for BW limits”

Possible related to #3761 issue: same behavior (crash not showing on 6.7 launch and start to appear with 6.8 launch. Same screen problems).

Searching for open bugs with “RX580” show many been reported in the last year (after 6.8 kernel launch) with the same screen problems. Since none noted chipset overheating (that I too did not before having easy access to motherboard this month) it was open another issue, for confirmation if it is really the same issue.

I can confirm it doesn’t affect modern, all AMD, computer (ryzen 7000 series and radeon RX6750 XT) nor Intel + Nvidia hybrid laptops.

Hardware description:

• CPU: AMD Phenom(tm) II X6 1090T Processor
• Integrated GPU: AMD RS880
• Discrete GPU: AMD Radeon RX 580 Series
• System Memory: total: 16 GiB available: 15.37 GiB used: 4.18 GiB (27.2%)
• Display(s): AOC D26W931 res: 1360x768@60hz, 100% DPI scaling
• Type of Display Connection: HDMI

System information:

• Distro name and Version: Manjaro
• Kernel version: 6.14.0-1
• Custom kernel: Kernel compiled from git+https://kernel.googlesource.com/pub/scm/linux/kernel/git/torvalds/linux for bisecting
• AMD official driver version: N/A, only Linux kernel drivers

How to reproduce the issue:

On a M5A88-M Asus motherboard try to use the the discrete GPU (RX 580) as primary render through bios set or try to start an application or Steam Proton game on the discrete GPU with DRI_PRIME=1once started the system with integrated GPU.

Attached files:

build_config: complete hardware and system information.

inxi.txt: .config file used to compile the git linux kernel.

BISECT_LOG: log of the bisect process starting with LTS kernels 6.6 and 6.12.

changed the description

added Polaris label

added bisected label

I'm a bit surprised that 6.12 is still affected and that https://git.kernel.org/torvalds/c/87dfeb47a5f48 didn't help.

Maybe the simplest solution is another condition in amdgpu_device_pcie_dynamic_switching_supported for this old system.

Could you please try this patch?

0001-drm-amd-Add-AMD-Phenom-2-to-pcie-dynamic-switching-e.patch

If this patch works, would you mind sharing your email address so I can CC you on the submission and code review?

Thank you so much Mario. I tested the patch with the upstream linux git (linux-git-6.14.r12245.91e5bfe317d8) and I can confirm it solve the issue. I will test it against the 6.12 stable branch tomorrow.

As about the email is it just post the email address in the comments? I am a new account user and because and may be on the restrictions (like due to an influx of spam message above says). I sorry for the silly questions, but it's my first interaction on gitlab (or any git service) at all, and I still don't know how it works. Anyway I set public email on the profile after searching the web. Inform if it is enough.

With your good test results and confusion to how to add an email, I just sent it out to mailing list last night.

[PATCH] drm/amd: Add AMD Phenom 2 to pcie dynamic switching exclusions

If you would like to add Tested-by tag you can reply to that post (there are reply instructions linked in it). Once it's reviewed we'll get it merged.

assigned to @superm1

Thanks. I could confirm that backporting the patch to the 6.12 git branch worked perfectly reverting the regression completely.

My only concern is if the test condition should be restricted to only my model of CPU or be applied to the AM3, AM3+ old processors in general.

I can’t test because I don’t have the same hardware and @Kaxydunyjousiz did not provided a bisect but it seems the issue affects other decade old AM3, AM3+ hardware too. By the comments of @Kaxydunyjousiz in issue #3864, the “journal 6.12.log” provided and the screen problems apparently equal I suspect it affects other AM3 processors paired with RX500 GPUs (Polaris).

hardware	mine	Kaxydunyjousiz
Motherboard	Asus M5A88-M	Asus M5A78L-M PLUS
Socket	AM3+	AM3+
Chipset	AMD 880G	AMD 760G
Processor	Phenom II x6 1090T	AMD FX(tm)-8300
Integrated GPU	AMD RS880 [Radeon HD 4250]	AMD RS780L [ Radeon HD 3000]
Discrete GPU	AMD Radeon RX 580	AMD Radeon RX 550

I don’t understand the internal of the AMD GPUs but looking the region of the code you patched (PCIe dynamic switching) to make it work again, and the configuration of system affected I wonder if it is not related with conflict with ATI Hybrid Graphics technology (Hybrid CrossFire X) (chipsets 790GX, 780G, 760G, 785G, 880G and 890GX) since both the function patched and the ATI technology take care of dynamic switching the GPUs (As I understood reading ATI Hybrid Graphics entry on Wikipedia).

Sorry for bothering you with that. My side is perfect, but I would like to avoid similar issues been reported by others in the future finding the extent of the problem.

Maybe here is not the place for it (let me know).

That's a good line of thought. Let's try this instead. It will cover a lot more hardware and should still be safe.

v2-0001-drm-amd-Add-pre-zen-AMD-hardware-to-PCIe-dynamic-.patch

mentioned in issue #3864

I tested the second version of the patch with the upstream linux git (linux-git-6.14.r12966.a2cc6ff5ec8f) and I can confirm it works perfectly solving the issue.

Great, thanks! If you would like to leave a Tested-by: tag, here is the upstream submission for it.

[PATCH v2] drm/amd: Add pre-zen AMD hardware to PCIe dynamic switching exclusions