[PATCH 1/2] drm/amdgpu: read FB through BAR0 when aperture is unavailable

[PATCH 1/2] drm/amdgpu: read FB through BAR0 when aperture is unavailable

[chong li]

Allow early VRAM reads to fall back to a temporary BAR0 mapping
when the normal aperture mapping has not been established yet.

Signed-off-by: chong li <chongli2@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 43 +++++++++++++++++++++-
 1 file changed, 42 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 5442a1fc1c37..610d82b79de3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -743,6 +743,44 @@ void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
 	drm_dev_exit(idx);
 }
 
+static int amdgpu_device_read_fb_via_bar0(struct amdgpu_device *adev,
+					  u64 offset, void *buf, size_t size)
+{
+	resource_size_t bar_start, bar_size, map_base;
+	void __iomem *vram;
+	size_t map_offset, map_size;
+	unsigned long flags;
+	u64 end;
+
+	if (!buf || !size)
+		return -EINVAL;
+
+	flags = pci_resource_flags(adev->pdev, 0);
+	if ((flags & IORESOURCE_UNSET) || !(flags & IORESOURCE_MEM))
+		return -EINVAL;
+
+	bar_size = pci_resource_len(adev->pdev, 0);
+	if (!bar_size)
+		return -ENODEV;
+
+	if (check_add_overflow(offset, size, &end) || end > bar_size)
+		return -EINVAL;
+
+	bar_start = pci_resource_start(adev->pdev, 0);
+	map_offset = offset_in_page(offset);
+	map_base = bar_start + (offset & PAGE_MASK);
+	map_size = PAGE_ALIGN(map_offset + size);
+
+	vram = ioremap_wc(map_base, map_size);
+	if (!vram)
+		return -ENOMEM;
+
+	memcpy_fromio(buf, (u8 __iomem *)vram + map_offset, size);
+	iounmap(vram);
+
+	return 0;
+}
+
 /**
  * amdgpu_device_aper_access - access vram by vram aperture
  *
@@ -762,8 +800,11 @@ size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
 	size_t count = 0;
 	uint64_t last;
 
-	if (!adev->mman.aper_base_kaddr)
+	if (!adev->mman.aper_base_kaddr) {
+		if (!write && !amdgpu_device_read_fb_via_bar0(adev, pos, buf, size))
+			return size;
 		return 0;
+	}
 
 	last = min(pos + size, adev->gmc.visible_vram_size);
 	if (last > pos) {
-- 
2.48.1

[PATCH 2/2] drm/amdgpu: improve the amdgpu device init progress in sriov mode

[chong li]

Move the initialization of non-GPU resources
out of the full GPU access region during AMDGPU device initialization

Signed-off-by: chong li <chongli2@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 610d82b79de3..c2ce4659ddc7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -1994,10 +1994,6 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
 	amdgpu_device_enable_virtual_display(adev);
 
 	if (amdgpu_sriov_vf(adev)) {
-		r = amdgpu_virt_request_full_gpu(adev, true);
-		if (r)
-			return r;
-
 		r = amdgpu_virt_init_critical_region(adev);
 		if (r)
 			return r;
@@ -2159,6 +2155,12 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
 	if (!total)
 		return -ENODEV;
 
+	if (amdgpu_sriov_vf(adev)) {
+		r = amdgpu_virt_request_full_gpu(adev, true);
+		if (r)
+			return r;
+	}
+
 	if (adev->gmc.xgmi.supported)
 		amdgpu_xgmi_early_init(adev);
 
-- 
2.48.1

Re: [PATCH 2/2] drm/amdgpu: improve the amdgpu device init progress in sriov mode

[Christian König]

On 7/1/26 07:53, chong li wrote:
> Move the initialization of non-GPU resources
> out of the full GPU access region during AMDGPU device initialization

As far as I can see that won't work like this.

There are a lot of steps which require full GPU access.

What exactly is the justification of the change?

Regards,
Christian.

> 
> Signed-off-by: chong li <chongli2@amd.com>
> Co-authored-by: Cursor <cursoragent@cursor.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 610d82b79de3..c2ce4659ddc7 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -1994,10 +1994,6 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
>  	amdgpu_device_enable_virtual_display(adev);
>  
>  	if (amdgpu_sriov_vf(adev)) {
> -		r = amdgpu_virt_request_full_gpu(adev, true);
> -		if (r)
> -			return r;
> -
>  		r = amdgpu_virt_init_critical_region(adev);
>  		if (r)
>  			return r;
> @@ -2159,6 +2155,12 @@ static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
>  	if (!total)
>  		return -ENODEV;
>  
> +	if (amdgpu_sriov_vf(adev)) {
> +		r = amdgpu_virt_request_full_gpu(adev, true);
> +		if (r)
> +			return r;
> +	}
> +
>  	if (adev->gmc.xgmi.supported)
>  		amdgpu_xgmi_early_init(adev);
>  

Re: [PATCH 1/2] drm/amdgpu: read FB through BAR0 when aperture is unavailable

[Christian König]

On 7/1/26 07:53, chong li wrote:
> Allow early VRAM reads to fall back to a temporary BAR0 mapping
> when the normal aperture mapping has not been established yet.

Clear NAK.

You are just trying to fix a symptom of incorrect initialization order here.

Regards,
Christian.

> 
> Signed-off-by: chong li <chongli2@amd.com>
> Co-authored-by: Cursor <cursoragent@cursor.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 43 +++++++++++++++++++++-
>  1 file changed, 42 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 5442a1fc1c37..610d82b79de3 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -743,6 +743,44 @@ void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
>  	drm_dev_exit(idx);
>  }
>  
> +static int amdgpu_device_read_fb_via_bar0(struct amdgpu_device *adev,
> +					  u64 offset, void *buf, size_t size)
> +{
> +	resource_size_t bar_start, bar_size, map_base;
> +	void __iomem *vram;
> +	size_t map_offset, map_size;
> +	unsigned long flags;
> +	u64 end;
> +
> +	if (!buf || !size)
> +		return -EINVAL;
> +
> +	flags = pci_resource_flags(adev->pdev, 0);
> +	if ((flags & IORESOURCE_UNSET) || !(flags & IORESOURCE_MEM))
> +		return -EINVAL;
> +
> +	bar_size = pci_resource_len(adev->pdev, 0);
> +	if (!bar_size)
> +		return -ENODEV;
> +
> +	if (check_add_overflow(offset, size, &end) || end > bar_size)
> +		return -EINVAL;
> +
> +	bar_start = pci_resource_start(adev->pdev, 0);
> +	map_offset = offset_in_page(offset);
> +	map_base = bar_start + (offset & PAGE_MASK);
> +	map_size = PAGE_ALIGN(map_offset + size);
> +
> +	vram = ioremap_wc(map_base, map_size);
> +	if (!vram)
> +		return -ENOMEM;
> +
> +	memcpy_fromio(buf, (u8 __iomem *)vram + map_offset, size);
> +	iounmap(vram);
> +
> +	return 0;
> +}
> +
>  /**
>   * amdgpu_device_aper_access - access vram by vram aperture
>   *
> @@ -762,8 +800,11 @@ size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
>  	size_t count = 0;
>  	uint64_t last;
>  
> -	if (!adev->mman.aper_base_kaddr)
> +	if (!adev->mman.aper_base_kaddr) {
> +		if (!write && !amdgpu_device_read_fb_via_bar0(adev, pos, buf, size))
> +			return size;
>  		return 0;
> +	}
>  
>  	last = min(pos + size, adev->gmc.visible_vram_size);
>  	if (last > pos) {