Re: [PATCH v1] drm/amdgpu: Fix ISP regression issue in kernel v7.0

["Du, Bin" <bin.du@amd.com>]

Hi Pratap,

FYI, the following patch may resolve the AMD ISP driver not auto-loading 
issue (without using MODULE_SOFTDEP). Please double check.

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c 
b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index 920595f0d22c..95d26f086d54 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -3212,4 +3212,3 @@ module_exit(amdgpu_exit);
  MODULE_AUTHOR(DRIVER_AUTHOR);
  MODULE_DESCRIPTION(DRIVER_DESC);
  MODULE_LICENSE("GPL and additional rights");
-MODULE_SOFTDEP("post: amd_isp4_capture i2c-designware-amdisp 
pinctrl-amdisp");
\ No newline at end of file
diff --git a/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c 
b/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c
index 485ecdec9618..409c1ca5a5e1 100644
--- a/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c
+++ b/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c
@@ -118,6 +118,29 @@ static int isp_set_performance_state(struct 
generic_pm_domain *genpd,
         return 0;
  }

+/*
+ * MFD core sets the parent GPU's ACPI companion (LNXVIDEO) on each
+ * child, giving them modalias "acpi:LNXVIDEO:" instead of
+ * "platform:<name>".  Clear it and retrigger the uevent so udev
+ * sees the correct platform modalias for module autoloading.
+ */
+static int isp_clear_acpi_fwnode(struct device *dev, void *data)
+{
+       struct platform_device *pdev;
+
+       if (!dev->type || !dev->type->name ||
+           strcmp(dev->type->name, "mfd_device"))
+               return 0;
+
+       pdev = to_platform_device(dev);
+       if (pdev->mfd_cell) {
+               device_set_node(dev, NULL);
+               kobject_uevent(&dev->kobj, KOBJ_ADD);
+       }
+
+       return 0;
+}
+
  static int isp_genpd_add_device(struct device *dev, void *data)
  {
         struct generic_pm_domain *gpd = data;
@@ -375,6 +398,13 @@ static int isp_v4_1_1_hw_init(struct amdgpu_isp *isp)
                 goto failure;
         }

+       /*
+        * MFD core sets the parent GPU's ACPI companion (LNXVIDEO) on
+        * each child, giving them modalias "acpi:LNXVIDEO:" instead of
+        * "platform:<name>". Clear it so udev matches by platform alias.
+        */
+       device_for_each_child(isp->parent, NULL, isp_clear_acpi_fwnode);
+
         return 0;

  failure:

>-----Original Message-----
>From: Nirujogi, Pratap <Pratap.Nirujogi@amd.com>
>Sent: Wednesday, March 11, 2026 11:29 PM
>To: Limonciello, Mario <Mario.Limonciello@amd.com>; Nirujogi, Pratap
><Pratap.Nirujogi@amd.com>; amd-gfx@lists.freedesktop.org; Deucher,
>Alexander <Alexander.Deucher@amd.com>; Limonciello, Mario
><Mario.Limonciello@amd.com>; Koenig, Christian
><Christian.Koenig@amd.com>
>Cc: rafael.j.wysocki@intel.com; Chan, Benjamin (Koon Pan)
><Benjamin.Chan@amd.com>; Du, Bin <Bin.Du@amd.com>; Li, King
><King.Li@amd.com>
>Subject: Re: [PATCH v1] drm/amdgpu: Fix ISP regression issue in kernel v7.0
>
>
>
>On 3/11/2026 1:26 AM, Mario Limonciello wrote:
>>
>>
>> On 3/10/2026 5:52 PM, Nirujogi, Pratap wrote:
>>>
>>>
>>> On 3/9/2026 5:58 PM, Mario Limonciello wrote:
>>>>
>>>>
>>>> On 3/9/2026 4:50 PM, Pratap Nirujogi wrote:
>>>>> Add NULL pointer checks for dev->type before accessing
>>>>> dev->type->name in ISP genpd add/remove functions to
>>>>> prevent kernel crashes. Also add MODULE_SOFTDEP to ensure ISP
>>>>> driver dependencies are loaded in correct order.
>>>>>
>>>>> The regression was introduced in kernel v7.0 where MFD ISP device
>>>>> enumeration doesn't complete by the time it is added to gendp. The
>>>>> timing of ISP device enumeration has changed because of the changes
>>>>> in registering the device sources in the device hierarchy.
>>>>
>>>> It's a little bit pedantic; but I /think/ there are two different
>>>> problems here with two different root causes that both happened in
>>>> 7.0-rc.
>>>>
>>>> As a consequence I think you should have this split out as two
>>>> separate patches in a series linked to a Fixes tag with the reason
>>>> for each of them.
>>>>
>>> Hi Mario,
>>>
>>> Few things to clarify before I split into 2 patches and submit v2.
>>>
>>> - Yes, you are correct, the changes in this patch are not directly
>>> related to the Fixes tag 02c057ddefef mentioned. But since it has the
>>> dependency on the below patch for automatic modprobe of isp to work,
>>> I have used the same tag to cover the dependency. My apologies if
>>> this approach is incorrect and misleading.
>>>
>>> https://lore.kernel.org/all/5986516.DvuYhMxLoT@rafael.j.wysocki/
>>>
>>> - NULL dereferencing issue with dev->type observed on v7.0 is
>>> specific to this commit 057edc58aa5926d63840c7f30afe0953d3994fa3. As
>>> the wakeup sources are registered using physical device instead of
>>> ACPI device, wakeup source device (wakeup14) is added as the first
>>> child of AMDGPU device; and since its dev->type is not initialized
>>> properly it has resulted in segfault.
>>
>> Sure this makes sense and is a no brainer to get in.  I would just
>> send this one right now and we can keep noodling on MODULE_SOFTDEP
>> (more comments below).
>>
>thanks, I’ll go ahead and send this one out now and will submit the
>MODULE_SOFTDEP change separately after the root cause is identified ( more
>details below ).
>
>>>
>>> In 6.19-rc4 or earlier versions, this issue was not observed as the
>>> wakeup source device was never part of AMDGPU children list.
>>>
>>> For the changes in isp_v4_1_1.c, I will use Fixes tag 057edc58aa59 in v2.
>>>
>>> - MODULE_SOFTDEP change in amdgpu_drv.c is needed for automatic
>>> modprobe of isp (and other amdgpu mfd child devices) to work in v7.0.
>>> But couldn't identify the specific commit in v7.0 that is causing the
>>> issue. I can confirm it is not because of commit 057edc58aa59 as the
>>> automatic modprobe doesn't work even on reverting this commit. Can I
>>> submit this as the fix needed for isp probe to work in v7.0 without
>>> the fixes tag?
>>>
>>
>> MODULE_SOFTDEP is generally for ordering, but I don't think you have
>> an issue with those modules loading before amdgpu do you?  I'm not
>> really following why the modaliases stopped working and I'm a bit
>> worried that it's papering over a more nuanced issue still.
>>
>yes, this explicit load‑order dependency was not required earlier. I agree that
>the regression point is not clearly identified. I will identify the bisected commit
>and submit the patch later.
>
>> To identify the root cause, it might be helpful to do a bisect,
>> although it's a bit complicated.
>> 1) At any point that has that platform driver conversion you need to
>> either revert or add the change to auxillary
>> 2) At any point that has 057edc58aa5926d63840c7f30afe0953d3994fa3 you
>> need to apply the NULL pointer derf fix
>> 3) At any point that has the changes in linux-media for API
>> adjustments in 7.0 you'll need to pick what version of ISP series to apply.
>>
>I tried #1, #2 and few other combinations, but none of them helped. I will
>investigate further to identify the commit causing the issue.
>>
>>> Thanks,
>>> Pratap
>>>
>>>>>
>>>>> Co-developed-by: Bin Du <Bin.Du@amd.com>
>>>>> Fixes: 02c057ddefef ("ACPI: video: Convert the driver to a platform
>>>>> one")
>>>>> Signed-off-by: Pratap Nirujogi <pratap.nirujogi@amd.com>
>>>>> ---
>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 1 +
>>>>>   drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c | 4 ++--
>>>>>   2 files changed, 3 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/
>>>>> drm/amd/amdgpu/amdgpu_drv.c index 95d26f086d545..920595f0d22ca
>>>>> 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>> @@ -3212,3 +3212,4 @@ module_exit(amdgpu_exit);
>>>>>   MODULE_AUTHOR(DRIVER_AUTHOR);
>>>>>   MODULE_DESCRIPTION(DRIVER_DESC);
>>>>>   MODULE_LICENSE("GPL and additional rights");
>>>>> +MODULE_SOFTDEP("post: amd_isp4_capture i2c-designware-amdisp
>>>>> pinctrl- amdisp");
>>>>> \ No newline at end of file
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c b/drivers/gpu/
>>>>> drm/amd/amdgpu/isp_v4_1_1.c index b3590b33cab9e..485ecdec96184
>>>>> 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/isp_v4_1_1.c
>>>>> @@ -129,7 +129,7 @@ static int isp_genpd_add_device(struct device
>>>>> *dev, void *data)
>>>>>       if (!pdev)
>>>>>           return -EINVAL;
>>>>> -    if (!dev->type->name) {
>>>>> +    if (!dev->type || !dev->type->name) {
>>>>>           drm_dbg(&adev->ddev, "Invalid device type to add\n");
>>>>>           goto exit;
>>>>>       }
>>>>> @@ -165,7 +165,7 @@ static int isp_genpd_remove_device(struct
>>>>> device *dev, void *data)
>>>>>       if (!pdev)
>>>>>           return -EINVAL;
>>>>> -    if (!dev->type->name) {
>>>>> +    if (!dev->type || !dev->type->name) {
>>>>>           drm_dbg(&adev->ddev, "Invalid device type to remove\n");
>>>>>           goto exit;
>>>>>       }
>>>>
>>>>
>>>
>>

Regards,
Bin