Skip to contentSkip to conversation

drm/amd/ras: Avoid ECC status update in hw_fini for VF unload

PATCH In review Ce Sun 2 participants 2 messages Apr 13

Subsystems (2)
lore
Share & tools
lore

Ce Sun

author
#0 lore raw

PATCH 1/1 drm/amd/ras: Avoid ECC status update in hw_fini for VF unload

VF sends IDH_REQ_GPU_FINI_ACCESS before hw_fini during unload. PF no longer accepts requests, so skip ECC status update to prevent mailbox timeout.

---
 drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_virt_ras_cmd.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_virt_ras_cmd.c +2 -7
index 28f94b92716a..838eb91aef39 100644
@@ -517,14 +517,9 @@ int amdgpu_virt_ras_hw_fini(struct amdgpu_device *adev)
 			(struct amdgpu_virt_ras_cmd *)ras_mgr->virt_ras_cmd;
 	struct vram_blocks_ecc *blks_ecc = &virt_ras->blocks_ecc;
 
-	if (blks_ecc->shared_mem.cpu_addr) {
-		__set_cmd_auto_update(adev,
-			RAS_CMD__GET_ALL_BLOCK_ECC_STATUS,
-			blks_ecc->shared_mem.gpa,
-			blks_ecc->shared_mem.size, false);
-
+	if (blks_ecc->shared_mem.cpu_addr)
 		memset(blks_ecc->shared_mem.cpu_addr, 0, blks_ecc->shared_mem.size);
-	}
+
 	memset(blks_ecc, 0, sizeof(*blks_ecc));
 
 	return 0;
Signature · 1 line
2.34.1
Reviewed-by on v1 1/1
#1 lore raw

[AMD Official Use Only - AMD Internal Distribution Only]

Fine to merge this as an initial step toward parity with the legacy implementation. Let's follow up on the design: we may be able to fix this by adjusting the current sequence rather than relying on this approach long term.

The patch is

Regards,
Hawking

-----Original Message-----
From: Sun, Ce(Overlord) <[email protected]>
Sent: Monday, April 13, 2026 8:12 PM
To: [email protected]
Cc: Zhang, Hawking <[email protected]>; Chai, Thomas <[email protected]>; Zhou1, Tao <[email protected]>; Sun, Ce(Overlord) <[email protected]>
Subject: [PATCH] drm/amd/ras: Avoid ECC status update in hw_fini for VF unload

VF sends IDH_REQ_GPU_FINI_ACCESS before hw_fini during unload.
PF no longer accepts requests, so skip ECC status update to prevent mailbox timeout.

---
 drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_virt_ras_cmd.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/ras/ras_mgr/amdgpu_virt_ras_cmd.c +2 -7
index 28f94b92716a..838eb91aef39 100644
@@ -517,14 +517,9 @@ int amdgpu_virt_ras_hw_fini(struct amdgpu_device *adev)
                        (struct amdgpu_virt_ras_cmd *)ras_mgr->virt_ras_cmd;
        struct vram_blocks_ecc *blks_ecc = &virt_ras->blocks_ecc;
 
-       if (blks_ecc->shared_mem.cpu_addr) {
-               __set_cmd_auto_update(adev,
-                       RAS_CMD__GET_ALL_BLOCK_ECC_STATUS,
-                       blks_ecc->shared_mem.gpa,
-                       blks_ecc->shared_mem.size, false);
-
+       if (blks_ecc->shared_mem.cpu_addr)
                memset(blks_ecc->shared_mem.cpu_addr, 0, blks_ecc->shared_mem.size);
-       }
+
        memset(blks_ecc, 0, sizeof(*blks_ecc));
 
        return 0;

--
2.34.1