Skip to contentSkip to conversation

drm/amdgpu: Fix CSA buffer alloc failed on Vega

PATCH In review Rex Zhu 2 participants 3 messages Nov 12, 2018 2 messages not in the archive

Subsystems (2)
DRM drivers Radeon and amdgpu DRM drivers
lore
Share & tools
lore

Rex Zhu

author
#0 lore raw

PATCH 1/1 drm/amdgpu: Fix CSA buffer alloc failed on Vega

Alloc_pte failed when the VA address located in
the higher arrange of 256T.

so reserve the csa buffer under 128T as a work around.

[  122.979425] amdgpu 0000:03:00.0: va above limit (0xFFFFFFFFFFF1F >= 0x1000000000)
[  122.987080] BUG: unable to handle kernel paging request at ffff880e1a79fff8
Signed-off-by Rex Zhu <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c +2 -2
index fea4555..e2f325b 100644
@@ -36,9 +36,9 @@ uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev, uint32_t id)
 {
 	uint64_t addr = adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT;
 
-	addr -= AMDGPU_VA_RESERVED_SIZE * id;
+	addr = min(addr, AMDGPU_GMC_HOLE_START);
 
-	addr = amdgpu_gmc_sign_extend(addr);
+	addr -= (uint64_t)AMDGPU_VA_RESERVED_SIZE * id;
 
 	return addr;
 }
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -1
index 338a091..ea6a12a 100644
@@ -711,7 +711,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 
 		dev_info.virtual_address_offset = AMDGPU_VA_RESERVED_SIZE;
 		dev_info.virtual_address_max =
-			min(vm_size, AMDGPU_GMC_HOLE_START);
+			min(vm_size, AMDGPU_GMC_HOLE_START - adev->vm_manager.reserved_vm_size);
 
 		if (vm_size > AMDGPU_GMC_HOLE_START) {
 			dev_info.high_va_offset = AMDGPU_GMC_HOLE_END;
Signature · 6 lines
1.9.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1542018800-998-1-git-send-email-Rex.Zhu-5C7GfCeVMHo@public.gmane.org

#2 lore raw

Am 12.11.18 um 11:33 schrieb Rex Zhu:

Quoted 9 lines

Alloc_pte failed when the VA address located in
the higher arrange of 256T.

so reserve the csa buffer under 128T as a work around.

[  122.979425] amdgpu 0000:03:00.0: va above limit (0xFFFFFFFFFFF1F >= 0x1000000000)
[  122.987080] BUG: unable to handle kernel paging request at ffff880e1a79fff8
Signed-off-by Rex Zhu <[email protected]>

Well NAK, userspace wants to use the full address space below 128T for SVM.

The problem is rather that you incorrectly use amdgpu_csa_vaddr(). See the code in amdgpu_driver_open_kms() how to do it correctly:

Quoted 6 lines

                uint64_t csa_addr = amdgpu_csa_vaddr(adev) & AMDGPU_GMC_HOLE_MASK;

                r = amdgpu_map_static_csa(adev, &fpriv->vm, adev->virt.csa_obj, &fpriv->csa_va, csa_addr, AMDGPU_CSA_SIZE);
                if (r)
                        goto error_vm;

Christian.

Quoted 26 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 2 +-
  2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c +2 -2
index fea4555..e2f325b 100644
@@ -36,9 +36,9 @@ uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev, uint32_t id)
  {
  	uint64_t addr = adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT;
 
-	addr -= AMDGPU_VA_RESERVED_SIZE * id;
+	addr = min(addr, AMDGPU_GMC_HOLE_START);
 
-	addr = amdgpu_gmc_sign_extend(addr);
+	addr -= (uint64_t)AMDGPU_VA_RESERVED_SIZE * id;
 
  	return addr;
  }
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
index 338a091..ea6a12a 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file · line 717

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -1
@@ -711,7 +711,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 
  		dev_info.virtual_address_offset = AMDGPU_VA_RESERVED_SIZE;
  		dev_info.virtual_address_max =
-			min(vm_size, AMDGPU_GMC_HOLE_START);
+			min(vm_size, AMDGPU_GMC_HOLE_START - adev->vm_manager.reserved_vm_size);
 
  		if (vm_size > AMDGPU_GMC_HOLE_START) {
  			dev_info.high_va_offset = AMDGPU_GMC_HOLE_END;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): e7909a29-fddd-9c65-d748-49f51e72afa9-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

author
#4 lore raw

Got it .

Thanks.

Best Regards

Rex

________________________________
From: Christian König <[email protected]>
Sent: Monday, November 12, 2018 6:42 PM
To: Zhu, Rex; [email protected]
Subject: Re: [PATCH] drm/amdgpu: Fix CSA buffer alloc failed on Vega

Am 12.11.18 um 11:33 schrieb Rex Zhu:

Quoted 9 lines

Alloc_pte failed when the VA address located in
the higher arrange of 256T.

so reserve the csa buffer under 128T as a work around.

[  122.979425] amdgpu 0000:03:00.0: va above limit (0xFFFFFFFFFFF1F >= 0x1000000000)
[  122.987080] BUG: unable to handle kernel paging request at ffff880e1a79fff8
Signed-off-by Rex Zhu <[email protected]>

Well NAK, userspace wants to use the full address space below 128T for SVM.

The problem is rather that you incorrectly use amdgpu_csa_vaddr(). See the code in amdgpu_driver_open_kms() how to do it correctly:

Quoted 8 lines

uint64_t csa_addr = amdgpu_csa_vaddr(adev) &
AMDGPU_GMC_HOLE_MASK;

                r = amdgpu_map_static_csa(adev, &fpriv->vm,
adev->virt.csa_obj,
&fpriv->csa_va, csa_addr, AMDGPU_CSA_SIZE);
                if (r)
                        goto error_vm;

Christian.

Quoted 26 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 2 +-
  2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_csa.c +2 -2
index fea4555..e2f325b 100644
@@ -36,9 +36,9 @@ uint64_t amdgpu_csa_vaddr(struct amdgpu_device *adev, uint32_t id)
  {
       uint64_t addr = adev->vm_manager.max_pfn << AMDGPU_GPU_PAGE_SHIFT;
 
-     addr -= AMDGPU_VA_RESERVED_SIZE * id;
+     addr = min(addr, AMDGPU_GMC_HOLE_START);
 
-     addr = amdgpu_gmc_sign_extend(addr);
+     addr -= (uint64_t)AMDGPU_VA_RESERVED_SIZE * id;
 
       return addr;
  }
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
index 338a091..ea6a12a 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file · line 717

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -1
@@ -711,7 +711,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 
               dev_info.virtual_address_offset = AMDGPU_VA_RESERVED_SIZE;
               dev_info.virtual_address_max =
-                     min(vm_size, AMDGPU_GMC_HOLE_START);
+                     min(vm_size, AMDGPU_GMC_HOLE_START - adev->vm_manager.reserved_vm_size);
 
               if (vm_size > AMDGPU_GMC_HOLE_START) {
                       dev_info.high_va_offset = AMDGPU_GMC_HOLE_END;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx