I've identified a critical issue with the existing GPU reset mechanism that
causes a BSOD on the Windows Hyper-V platform. The current function:
static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
incorrectly sets the AMDGPU_HOST_FLR flag if any engine is hanging. This
approach wrongly assumes that the Host PF is always responsible for triggering
FLR. However, a VF (VM-guest) can also cause a GPU hang, which results in an
unsuccessful VM reset. This ultimately causes a FULL_ACCESS_TIMEOUT on the host
side, leading to six attempts to retrigger a Whole Guest Reset (WGR), which
results in a BSOD after five to six failed restarts.
Additionally, the current sequence sends a READY_TO_RESTART event and then
requests FULL_ACCESS, which seems incorrect to me.
My fix addresses this problem by using REQ_GPU_RESET to initiate the necessary
restart while appropriately handling the FULL ACCESS request. My implementation
has successfully passed 100 loop tests, confirming its effectiveness.
Signed-off-by: Nikola Petrovic <nipet...@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 7f354cd532dc..a2a436707200 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -5314,11 +5314,12 @@ static int amdgpu_device_reset_sriov(struct
amdgpu_device *adev,
struct amdgpu_hive_info *hive = NULL;
if (test_bit(AMDGPU_HOST_FLR, &reset_context->flags)) {
+ r = amdgpu_virt_wait_reset(adev);
+ if (r)
+ return r;
if (!amdgpu_ras_get_fed_status(adev))
- amdgpu_virt_ready_to_reset(adev);
- amdgpu_virt_wait_reset(adev);
+ amdgpu_virt_reset_gpu(adev);
clear_bit(AMDGPU_HOST_FLR, &reset_context->flags);
- r = amdgpu_virt_request_full_gpu(adev, true);
} else {
r = amdgpu_virt_reset_gpu(adev);
}
--
2.43.0
