Kernel hang when amdgpu driver is loaded on old radeon card

[Sitsofe Wheeler]

Hello,

(Apologies if I have CC'd the wrong people/places - I just went by
what get_maintainer.pl -f drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
said)

I recently upgraded from Ubuntu 20.04 (5.15.0-119.129~20.04.1-generic
kernel) to Ubuntu 24.04 (6.8.0-44-generic kernel) and found that while
booting the kernel hangs for around 15 seconds just before the amdgpu
driver is loaded:

[    4.459519] radeon 0000:01:05.0: [drm] Cannot find any crtc or sizes
[    4.460118] probe of 0000:01:05.0 returned 0 after 902266 usecs
[    4.460184] initcall radeon_module_init+0x0/0xff0 [radeon] returned
0 after 902473 usecs
[    4.465797] calling  drm_buddy_module_init+0x0/0xff0 [drm_buddy] @ 122
[    4.465853] initcall drm_buddy_module_init+0x0/0xff0 [drm_buddy]
returned 0 after 29 usecs
[    4.469419] radeon 0000:01:05.0: [drm] Cannot find any crtc or sizes
[    4.473831] calling  drm_sched_fence_slab_init+0x0/0xff0 [gpu_sched] @ 122
[    4.473892] initcall drm_sched_fence_slab_init+0x0/0xff0
[gpu_sched] returned 0 after 31 usecs
[   18.724442] calling  amdgpu_init+0x0/0xff0 [amdgpu] @ 122
[   18.726303] [drm] amdgpu kernel modesetting enabled.
[   18.726576] amdgpu: Virtual CRAT table created for CPU
[   18.726609] amdgpu: Topology: Add CPU node
[   18.726787] initcall amdgpu_init+0x0/0xff0 [amdgpu] returned 0
after 528 usecs

I've checked and the problem still exists in 6.11.0-061100rc7-generic
(which is close to vanilla upstream).

The graphics card I have is:
01:05.0 VGA compatible controller: Advanced Micro Devices, Inc.
[AMD/ATI] RS880M [Mobility Radeon HD 4225/4250] (prog-if 00 [VGA
controller])
01:05.0 0300: 1002:9712 (prog-if 00 [VGA controller])
Subsystem: 103c:1609

At first I thought the problem was related to the change
https://github.com/torvalds/linux/commit/eb4fd29afd4aa1c98d882800ceeee7d1f5262803
("drm/amdgpu: bind to any 0x1002 PCI diplay [sic] class device") which
now means my card is claimed by two drivers (radeon and amdgpu). That
change complicated things because:
- The amdgpu module and its dependencies remain permanently present (which
  never used to happen)
- It took some time for me to realise that the amdgpu driver hadn't suddenly
  grown the ability to support this old card :-) There is a nice table on
  https://www.x.org/wiki/RadeonFeature/#decoderringforengineeringvsmarketingnames
  that shows it is part of the R600 family and
  https://www.x.org/wiki/RadeonFeature/#featurematrixforfreeradeondrivers shows
  that R600 is only supported by the radeon driver.

However, testing a 5.16.20-051620-generic kernel showed that while the
amdgpu module is loaded, there is no 15 second hang... So far my
testing has the following results:
- 5.16.20-051620-generic - amdgpu loaded, no hang
- 5.18.19-051819-generic - amdgpu loaded, no hang
- 6.0.0-060000-generic - amdgpu loaded, hang
- 6.2.0-060200-generic - amdgpu loaded, hang
- 6.8.0-44-generic - amdgpu loaded, hang
- 6.11.0-061100rc7-generic - amdgpu loaded, hang

To work around the problem I've taken to blacklisting amdgpu in
/etc/modprobe.d/ which makes the hang disappear.

Does anyone else see this issue? Is there something better than my
current workaround? What do other drivers that want to bind to such a
large set of devices do? Further, while I'm already using
initcall_debug, is there any other kernel boot parameter to make
what's happening more visible?

--
Sitsofe