The amdgpu_vm_set_pasid function now requires the VM root PD's dma_resv lock
to be held, as enforced by the amdgpu_vm_assert_locked check added in the
lock refactoring.
This patch adds the necessary dma_resv_lock/dma_resv_unlock calls around
both the initial amdgpu_vm_set_pasid call and the cleanup path in the error
handler to properly satisfy the lock assertion.
Without these locks, we hit lockdep warnings or potential race conditions
when modifying the VM's PASID mapping, as the underlying VM state changes
must be protected by the root PD's reservation lock.
Fixes: ebe038089be("drm/amdgpu: revert to old status lock handling v3")
Signed-off-by: Jesse Zhang <[email protected]>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 10 +++++++++-
1 file changed, 9 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
index 8676400834fc..e411dccbe612 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
@@ -1386,6 +1386,7 @@ int amdgpu_driver_open_kms(struct drm_device *dev, struct
drm_file *file_priv)
{
struct amdgpu_device *adev = drm_to_adev(dev);
struct amdgpu_fpriv *fpriv;
+ struct dma_resv *resv;
int r, pasid;
/* Ensure IB tests are run on ring */
@@ -1425,9 +1426,14 @@ int amdgpu_driver_open_kms(struct drm_device *dev,
struct drm_file *file_priv)
if (r)
goto error_pasid;
+ resv = fpriv->vm.root.bo->tbo.base.resv;
+ dma_resv_lock(resv, NULL);
r = amdgpu_vm_set_pasid(adev, &fpriv->vm, pasid);
- if (r)
+ if (r) {
+ dma_resv_unlock(resv);
goto error_vm;
+ }
+ dma_resv_unlock(resv);
fpriv->prt_va = amdgpu_vm_bo_add(adev, &fpriv->vm, NULL);
if (!fpriv->prt_va) {
@@ -1470,7 +1476,9 @@ int amdgpu_driver_open_kms(struct drm_device *dev, struct
drm_file *file_priv)
error_pasid:
if (pasid) {
amdgpu_pasid_free(pasid);
+ dma_resv_lock(resv, NULL);
amdgpu_vm_set_pasid(adev, &fpriv->vm, 0);
+ dma_resv_unlock(resv);
}
kfree(fpriv);
--
2.49.0
