Am 17.05.2017 um 10:01 schrieb Michel Dänzer:

On 17/05/17 04:13 PM, zhoucm1 wrote:

On 2017年05月17日 14:57, Michel Dänzer wrote:

On 17/05/17 01:28 PM, zhoucm1 wrote:

On 2017年05月17日 11:15, Michel Dänzer wrote:

On 17/05/17 12:04 PM, zhoucm1 wrote:

On 2017年05月17日 09:18, Michel Dänzer wrote:

On 16/05/17 06:25 PM, Chunming Zhou wrote:

Change-Id: I8eb6d7f558da05510e429d3bf1d48c8cec6c1977
Signed-off-by: Chunming Zhou <David1.Zhou-5C7GfCeVMHo@public.gmane.org>

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index bca1fb5..f3e7525 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -2547,6 +2547,9 @@ int amdgpu_vm_ioctl(struct drm_device *dev,
void *data, struct drm_file *filp)
         case AMDGPU_VM_OP_UNRESERVE_VMID:
             amdgpu_vm_free_reserved_vmid(adev, &fpriv->vm,
AMDGPU_GFXHUB);
             break;
+    case AMDGPU_VM_OP_RESET:
+        fpriv->vram_lost_counter =
atomic_read(&adev->vram_lost_counter);
+        break;

How do you envision the UMDs using this? I can mostly think of them
calling this ioctl when a context is created or destroyed. But that
would also allow any other remaining contexts using the same DRM file
descriptor to use all ioctls again. So, I think there needs to be a
vram_lost_counter in struct amdgpu_ctx instead of in struct
amdgpu_fpriv.

struct amdgpu_fpriv for vram_lost_counter is proper place, especially
for ioctl return value.
if you need to reset ctx one by one, we can mark all contexts of that
vm, and then reset by userspace.

I'm not following. With vram_lost_counter in amdgpu_fpriv, if any
context calls this ioctl, all other contexts using the same file
descriptor will also be considered safe again, right?

Yes, but it really depends on userspace requirement, if you need to
reset ctx one by one, we can mark all contexts of that vm to guilty, and
then reset one context by userspace.

Still not sure what you mean by that.

E.g. what do you mean by "guilty"? I thought that refers to the context
which caused a hang. But it seems like you're using it to refer to any
context which hasn't reacted yet to VRAM contents being lost.

When vram is lost, we treat all contexts need to reset.

Essentially, your patches only track VRAM contents being lost per file
descriptor, not per context. I'm not sure (rather skeptical) that this
is suitable for OpenGL UMDs, since state is usually tracked per context.
Marek / Nicolai?

Oh, yeah that's a good point.

The problem with tracking it per context is that Vulkan also wants the ENODEV on the amdgpu_gem_va_ioct() and amdgpu_info_ioctl() which are context less.

But thinking more about this blocking those two doesn't make much sense. The VM content can be restored and why should be disallow reading GPU info?

The Vulkan APIs listed below could return VK_ERROR_DEVICE_LOST according to the spec.

I tries to provide a list of u/k interfaces that could be called for each vk API.

vkCreateDevice

-          amdgpu_device_initialize.

-          amdgpu_query_gpu_info

vkQueueSubmit

-          amdgpu_cs_submit

vkWaitForFences

                amdgpu_cs_wait_fences

vkGetEventStatus

vkQueueWaitIdle

vkDeviceWaitIdle

vkGetQueryPoolResults

                amdgpu_cs_query_Fence_status

vkQueueBindSparse

                amdgpu_bo_va_op

                amdgpu_bo_va_op_raw

vkCreateSwapchainKHR

vkAcquireNextImageKHR

vkQueuePresentKHR

                Not related with u/k interface.

Christian.