resume fails after hibernate amdgpu (black screen - rx 5700)

s|mon

I'm facing issues using hibernate with my radeon 5700 rx.

It goes to hibernate (on 5.8 kernel screen goes black, then comes back short before real hibernation) but when resume is triggered from power button it ends with black screen and unresponsive system after image loading is completed (screen is visible till then). If my interpretation of dmesg output is correct it seems to be related to amdgpu (Radeon 5700 Navi 10 ).

SSH to machine is possible.
Suspend to memory worked so far. As well as testing levels: freezer, devices, platform, processors, core to /sys/power/pm_test with disk state according to:
https://www.kernel.org/doc/Documentation/power/basic-pm-debugging.txt

I did an update to a newer bios without change of the behavior.

Harware is a MSI X570 board, Ryzen 3700X, 32G and said Radeon RX 5700.
Further details in dmesg and lspci.

Juippisi · Developer Joined: 30 Sep 2005 Posts: 761 Location: /home

Maybe you're running buggy kernel or firmware? Try to upgrade/downgrade either/both.

I've heard especially with 5700 series amdgpu has been buggy on 5.8. I also heard that two latest firmware packages have been buggy for these cards.
Try 5.7 or 5.9 kernel first and if they fail, try older firmware versions? And if you incorporate firmware drivers to your kernel, remember to rebuild kernel after emerging other firmware versions.

halcon · l33t Joined: 15 Dec 2019 Posts: 649

s|mon · Posted: Wed Oct 14, 2020 2:19 pm Post subject:

Good hint on the firmware (right now it is 20200918) - i already tried 5.9-rc8 (vanilla) as well, but of course with the same firmware. Will try later today with 5.7 and older firmware.
And thanks for the link to the amd specific stuff, i was not aware that this was seperated from the kernel (will search their bugtracker as well)
Would that be covering issues within kernel code or only the firmware parts?

halcon · l33t Joined: 15 Dec 2019 Posts: 649

s|mon · Posted: Wed Oct 14, 2020 3:55 pm Post subject:

I now tried with
5.6 and linux-firmware-20200721
5.7.8 and linux-firmware-20200721
5.7.8 and linux-firmware-20200817
5.8.11 and linux-firmware-20200817
5.8.11 and linux-firmware-20200918
5.9-rc8 and linux-firmware-20200918
all fail with hibernation.

On you last link halcon, was this ment to be a specific related issue? As it is about a throttling problem which is don't see here?

halcon · l33t Joined: 15 Dec 2019 Posts: 649

Goverp · Advocate Joined: 07 Mar 2007 Posts: 2202

With different hardware (oldish HP laptop with HD3400 graphics), I've never had hibernate working in the 3 years I've had it.
_________________
Greybeard

s|mon · Posted: Thu Oct 15, 2020 11:41 am Post subject:

Never working would be not nice - especially since suspend seems to work quite fine i'd assume it should not be that complex, but who knows.
FYI: i created a entry there as well 1335

Goverp · Advocate Joined: 07 Mar 2007 Posts: 2202

A couple of years ago I tried to nail down my hibernate problem. Searching on Google doesn't hack it for my hardware - there are loads of reports, but in essence they come up to "the only way to hibernate is use one of the basic screen drivers instead of amdgpu". Diagnosing the problem requires building a kernel with some extra debugging, trying to hibernate, restarting after the failure, digging an address out of the extra debug info, and mapping that to a module - tedious. What it did was blame amdgpu, which wasn't really a surprise.

Unfortunately it doesn's tell you any more, so it would be a problem for upstream. Now upstream would probably want a kernel bisect (or equivalent for their driver) to find out which change broke hibernate. But that's only possible if it ever worked :-(

More to the point, AMD wrote somewhere that they don't do QA on hibernate 'cos (a) there's so many hardware configurations, and hibernate is very hardware dependent, and (b) nobody uses it anyway. The latter is, of course, putting the cart before the horse; nobody uses hibernate because it doesn't work.

I did note that the current AMD drivers complain that I should have checkpoint restart support built into my kernel; I tried adding that, but it didn't cure the problem, and I presume (as it triggered a complete kernel rebuild) it adds lots of code, so I've left it out, and stuck with suspend.
_________________
Greybeard

s|mon · Posted: Fri Oct 16, 2020 11:17 am Post subject:

Nice background information - doesn't raise to much hopes though as what you describe seems to go beyong my skill/time-to-spent ratio as well.
Let's hope they pick up that nevertheless. Suspend is nice but limiting. With hibernate i could e.g. keep running ffmpeg conversions, hibernate and boot another OS, come back and continue as i didn't leave. A very special scenario but would work otherwise.
Anyhow thanks for sharing that details.