[PATCH] drm/amdgpu: Add a new module param to disable d3cold

[Mario Limonciello]

On 11/29/2023 02:51, Ma Jun wrote:
> Some platforms can't resume from d3cold state, So add a
> new module parameter to disable d3cold state for debugging
> purpose or workaround.
> 
> Signed-off-by: Ma Jun <Jun.Ma2 at amd.com>
> ---

This patch is essentially an 'amdgpu knob' for d3cold on the root port. 
At least for debugging purposes we also have a sysfs file 
'd3cold_allowed' that will enact the same behavior.

I do have a patch that I proposed to PCI core that stops d3cold_allowed 
from working in favor of requesting pcie_port_pm=off to be used instead 
for debugging purposes.

However that's a 'relatively big' debugging knob however as it will 
apply to all PCIe root ports.

Considering above I'm in favor of this being available as a localized 
debugging path for just the root port the dGPU is connected to.

Some comments below though:

>   drivers/gpu/drm/amd/amdgpu/amdgpu.h        | 1 +
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 7 +++++++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    | 8 ++++++++
>   3 files changed, 16 insertions(+)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index a9f54df9d33e..db9f60790267 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -166,6 +166,7 @@ extern char amdgpu_lockup_timeout[AMDGPU_MAX_TIMEOUT_PARAM_LENGTH];
>   extern int amdgpu_dpm;
>   extern int amdgpu_fw_load_type;
>   extern int amdgpu_aspm;
> +extern int amdgpu_d3cold;
>   extern int amdgpu_runtime_pm;
>   extern uint amdgpu_ip_block_mask;
>   extern int amdgpu_bapm;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 22b6a910b7f2..90501c44e7d0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -264,6 +264,13 @@ bool amdgpu_device_supports_px(struct drm_device *dev)
>   bool amdgpu_device_supports_boco(struct drm_device *dev)
>   {
>   	struct amdgpu_device *adev = drm_to_adev(dev);
> +	struct pci_dev *parent;
> +
> +	if (!amdgpu_d3cold) {
> +		parent = pcie_find_root_port(adev->pdev);
> +		pci_d3cold_disable(parent);
> +		return false;
> +	}
>   
>   	if (adev->has_pr3 ||
>   	    ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index 5f14f04cb553..c9fbb8bd4169 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -145,6 +145,7 @@ char amdgpu_lockup_timeout[AMDGPU_MAX_TIMEOUT_PARAM_LENGTH];
>   int amdgpu_dpm = -1;
>   int amdgpu_fw_load_type = -1;
>   int amdgpu_aspm = -1;
> +int amdgpu_d3cold = -1;

If this was chained to a larger workaround (such as automatically 
applying to a DMI quirk) it would make sense as int and with using
-1 for auto.  However there is a pretty dramatic downside for using this 
knob that it can break s2idle.

In my testing I've found that the following happens on an A+A design 
after s2idle with this parameter in use.

[   70.572270] pcieport 0000:01:00.0: Unable to change power state from 
D3cold to D0, device inaccessible
[   70.572481] pcieport 0000:02:00.0: Unable to change power state from 
D3cold to D0, device inaccessible
[   72.855769] amdgpu 0000:03:00.0: not ready 1023ms after resume; waiting
[   73.943545] amdgpu 0000:03:00.0: not ready 2047ms after resume; waiting
[   76.055602] amdgpu 0000:03:00.0: not ready 4095ms after resume; waiting
[   80.279550] amdgpu 0000:03:00.0: not ready 8191ms after resume; waiting
[   88.983562] amdgpu 0000:03:00.0: not ready 16383ms after resume; waiting
[  105.879581] amdgpu 0000:03:00.0: not ready 32767ms after resume; waiting
[  142.743646] amdgpu 0000:03:00.0: not ready 65535ms after resume; 
giving up
[  142.743793] amdgpu 0000:03:00.0: Unable to change power state from 
D3cold to D0, device inaccessible
[  142.804011] snd_hda_intel 0000:03:00.1: Unable to change power state 
from D3cold to D0, device inaccessible

So I don't see us ever automatically using this and it should be 
debugging only.  IOW this doesn't need to be integer; it can be boolean.

>   int amdgpu_runtime_pm = -1;
>   uint amdgpu_ip_block_mask = 0xffffffff;
>   int amdgpu_bapm = -1;
> @@ -359,6 +360,13 @@ module_param_named(fw_load_type, amdgpu_fw_load_type, int, 0444);
>   MODULE_PARM_DESC(aspm, "ASPM support (1 = enable, 0 = disable, -1 = auto)");
>   module_param_named(aspm, amdgpu_aspm, int, 0444);
>   
> +/**
> + * DOC: d3cold (int)

If you flip it to boolean as I suggested this should probably either 
rename to disable_d3cold or you should default to TRUE.

> + * To disable d3cold (1 = enable, 0 = disable). The default is -1 (auto, enabled).
> + */
> +MODULE_PARM_DESC(d3cold, "d3cold support (1 = enable, 0 = disable, -1 = auto)");
> +module_param_named(d3cold, amdgpu_d3cold, int, 0444);
> +
>   /**
>    * DOC: runpm (int)
>    * Override for runtime power management control for dGPUs. The amdgpu driver can dynamically power down