SR-IOV VFs use amdgpu_virt_ras_hw_init() which skips ras_core_hw_init(),
leaving fw_gpu_mem unallocated. ras_psp_reload_firmwares() then hits the
NULL buffer and returns -ENOMEM (-12) for every VF on resume, aborting
ip_resume_phase2 before amdgpu_fence_driver_hw_init() runs. The next
hibernate double-puts IRQ refcounts causing a WARN storm.
Skip the reload for VFs — there is nothing local to reload. The
post-reset path (amdgpu_ras_mgr_resume_after_reset) is unaffected.
Fixes: 5e5be56a5b76 ("drm/amdgpu/ras: reload RAS TA from ras mgr resume")
Signed-off-by: Chi Hong <[email protected]>
---
drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_ras_mgr.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_ras_mgr.c
b/drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_ras_mgr.c
index a02167397ce2..5bac792554f5 100644
--- a/drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_ras_mgr.c
+++ b/drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_ras_mgr.c
@@ -529,6 +529,9 @@ static int amdgpu_ras_mgr_resume(struct amdgpu_ip_block
*ip_block)
if (!ras_mgr->ras_is_ready)
return 0;
+ if (amdgpu_sriov_vf(adev))
+ return 0;
+
ret = ras_psp_reload_firmwares(ras_mgr->ras_core, 0);
if (ret)
RAS_DEV_ERR(adev,
--
2.43.0
