rc6 related freeze/crash back in 5.5+

Hello,

Forgive me for dragging up a very old issue here, but the depreciation of the "enable_rc6" parameter back in 4.16 has reared its ugly head again. Here's a summary of the issue:

o Removing the option for i915.enable_rc6=0 caused many issues, some severe enough to render computers unusable (see, e.g., #100 (closed), #1344 (closed))

o In the most severe cases there appears to be no way to catch a log of the crash/freeze. The system acts as though the power was suddenly disconnected (see many attempts to capture logs in #100 (closed))

o Issue #100 (closed) was closed with "And so it came to pass it was to be disabled for all gen9." I'm not sure what that means, but in kernel 5.4 the i915.enable_rc6 parameter was still disabled but the crashes didn't happen. Yay!

o All kernels since 5.5, however, still crash in the same way. This is true for the current kernel (5.10 as of this writing).

I was able to get by with the linux-lts kernel, which was fortunately 5.4, until recently when it switched to 5.10 (and the associated crashes).

So clearly something magic happened in 5.4 that was undone or changed in 5.5 onward.

After reading lots of comments on similar issues I have tried the following kernel parameters, none of which work:

• i915.edp_vswing=0 (from here)
• intel_idle.max_cstate=0 (#100 (closed))
• intel_idle.max_cstate=1 (#100 (closed), #272 (closed))
• intel_idle.max_cstate=4 (#272 (closed))
• i915.disable_power_well=0 (with and without intel_idle.max_cstate=0 and processor.max_cstate=0) (#100 (closed))
• i915.enable_dc=0 (#100 (closed))
• i915.enable_psr=0 (#272 (closed))

Really I'm just flailing blindly because I don't really understand what any of these do. I am more that happy to help with any debugging, but at this point I'm way past my depth on what I can do on my own.

Thanks,

Arthur

#4858 just leaving this here.

@eigenbrot can you please try to reproduce this issue on latest drm-tip and post your observations on it described at https://gitlab.freedesktop.org/drm/intel/-/wikis/How-to-file-i915-bugs and https://01.org/linuxgraphics/documentation/how-get-gpu-error-state ?

added Community label

There is a new change to this matter, at least for the rc6p power saving thing. It is again disabled by default in Linux kernel 6.2, see "Linux Takes Another Shot At Fixing Visual Glitches & GPU Hangs For Intel Sandy Bridge".

However, it would be really helpful to know if there exist still a kernel parameter to adjust those rc6 related settings at a Sandy Bridge based GPU. The current state is unclear. I am at an iMac12,2 based system which includes an Intel HD 2000 iGPU and a TURKS Radeon HD 6770M. There is present some strange behavior with the Intel iGPU since years. Maybe this has something to do with the rc6 topic here.

Whatever, pinging @ickle for an explanation about his last comment in #100 (closed):

And so it came to pass it was to be disabled for all gen9.

Note, I use Kubuntu 22.04 LTS with Mesa 23.1.0-devel (git-7bed7d9 2023-01-20 jammy-oibaf-ppa) and kernel 5.15.0-58-generic.

@jewins @alan_previn_intel Please have a look.

added platform: SNB label

@eigenbrot enable_rc6 is just a modparam, removing this modparam should not have any impact on overall functionality. Could you try disabling rc6 with echo 0 > /sys/class/drm/card0/gt/gt0/rc6_enable and check if you still see the issues.

@anshuma1 Question, how can I search for the enable_rc6 option on my system?

The path seems to be different in my case. It is identical until /sys/class/drm/card0 and after that it starts to follow an other naming. Most likely also because I am at an dual-GPU system...

@lorn10 you may search in path with major/minor number of your intel gpu. Example /sys/dev/char/226:0/

Thanks, I have found it under /sys/dev/char/226:0/subsystem/card1/power/.

Will look into that somewhat deeper when I find the time. At least it is now clear that there exist still a way to disable the rc6 functionality.

there is a working solution for #4858, mentioned in the first post. could be a different situation, but worth a try.

By the way, - thanks for the hint.

I can confirm that intel_gpu_frequency -m sets the clock rate of the Intel iGPU to the max possible value while intel_gpu_frequency -d sets it back to the default one.

However, maybe I should add here that in my case I have no GPU HANG: errors in dmesg. That seems to be a different issue.

At the Apple iMac12,2 computer line there seems to be present a very rare multi-GPU config setting where the Radeon dGPU is the primary one and the Intel iGPU the secondary. (Usually the opposite is the case.) With older kernel and the classic Mesa i965 driver this resulted in conjunction with the r600 Gallium driver in some strange behavior like freezes and empty/black screen. But this all improved clearly with newer kernels and the introduction of the Gallium crocus driver.

The last remaining issue is some weird VRAM/GTT memory/performance mismatch between the Intel HD 2000 iGPU and the TURKS Radeon. I was hoping that this was somehow rc6 / rc6p related but it looks that changing those power saving settings doesn't have any impact in my case.