On 2026/1/15 下午9:47, Christian König wrote:
On 1/15/26 02:28, Tiezhu Yang wrote:
Currently, there are many Call Traces when booting kernel on LoongArch,
here are the trimmed kernel log messages:
amdgpu 0000:07:00.0: amdgpu: hw_init of IP block <gfx_v6_0> failed -110
amdgpu 0000:07:00.0: amdgpu: amdgpu_device_ip_init failed
amdgpu 0000:07:00.0: amdgpu: Fatal error during GPU init
amdgpu 0000:07:00.0: amdgpu: amdgpu: finishing device.
------------[ cut here ]------------
WARNING: drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c:639 at
amdgpu_irq_put+0xb0/0x140 [amdgpu], CPU#0: kworker/0:0/9
...
Call Trace:
[<90000000047a8524>] show_stack+0x64/0x190
[<90000000047a0614>] dump_stack_lvl+0x6c/0x9c
[<90000000047cef34>] __warn+0xa4/0x1b0
[<90000000060a4884>] __report_bug+0xa4/0x1d0
[<90000000060a4a88>] report_bug+0x38/0xd0
[<90000000060df330>] do_bp+0x260/0x410
[<90000000047a6bc0>] handle_bp+0x120/0x1c0
[<ffff8000028bff40>] amdgpu_irq_put+0xb0/0x140 [amdgpu]
[<ffff8000027b1a8c>] amdgpu_fence_driver_hw_fini+0x12c/0x180 [amdgpu]
[<ffff800002f2c04c>] amdgpu_device_fini_hw+0xf0/0x3fc [amdgpu]
[<ffff80000279e2ac>] amdgpu_driver_load_kms+0x7c/0xa0 [amdgpu]
[<ffff800002791128>] amdgpu_pci_probe+0x298/0x810 [amdgpu]
[<90000000054d04a4>] local_pci_probe+0x44/0xc0
[<90000000047f4ab0>] work_for_cpu_fn+0x20/0x40
[<90000000047f93e0>] process_one_work+0x170/0x4e0
[<90000000047fa14c>] worker_thread+0x3ac/0x4e0
[<9000000004806824>] kthread+0x154/0x170
[<90000000060df5b4>] ret_from_kernel_thread+0x24/0xd0
[<90000000047a62a4>] ret_from_kernel_thread_asm+0xc/0x88
---[ end trace 0000000000000000 ]---
amdgpu 0000:07:00.0: probe with driver amdgpu failed with error -110
amdgpu 0000:07:00.0: amdgpu: amdgpu: ttm finalized
This is because amdgpu_irq_enabled() is false in amdgpu_irq_put(), then
the condition of WARN_ON() is true.
In order to avoid the unnecessary Call Traces, it can remove the check of
amdgpu_irq_enabled() and only check atomic_read(&src->enabled_types[type]
for three reasons:
(1) The aim is to prevent refcount from being less than 0, it was added in
commit 1fa8d710573f ("drm/amdgpu: Fix desktop freezed after gpu-reset").
(2) There are already many useful failed log before the Call Trace, there
is no need to WARN_ON().
Well completely disagree. The call trace here is absolutely intentional.
If so, since the call trace is same, is it enough to use WARN_ON_ONCE()
here?
That you get a lot of other backtraces because the driver doesn't initialize at
all isn't a good rational to remove this one here.
Regards,
Christian.
(3) The following checks in amdgpu_irq_put() are same with the checks in
amdgpu_irq_enabled(), there is no need to do the redundant operations.
if (!adev->irq.installed)
return -ENOENT;
if (type >= src->num_types)
return -EINVAL;
if (!src->enabled_types || !src->funcs->set)
return -EINVAL;
Is this reasonable? Only check atomic_read(&src->enabled_types[type]?
That is to say, does it make sense to do the following change?
----->8-----
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c
index 8112ffc85995..d10d6fcc525e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_irq.c
@@ -636,7 +636,7 @@ int amdgpu_irq_put(struct amdgpu_device *adev,
struct amdgpu_irq_src *src,
if (!src->enabled_types || !src->funcs->set)
return -EINVAL;
- if (WARN_ON(!amdgpu_irq_enabled(adev, src, type)))
+ if (WARN_ON_ONCE(!atomic_read(&src->enabled_types[type])))
return -EINVAL;
if (atomic_dec_and_test(&src->enabled_types[type]))
Thanks,
Tiezhu