Re: [V2] drm/xe: Skip clearing purged page-table BOs

["Ghimiray, Himal Prasad" <himal.prasad.ghimiray@intel.com>]

On 08-09-2026 09:25, Matthew Brost wrote:
> On Tue, Sep 08, 2026 at 09:22:35AM +0530, Ghimiray, Himal Prasad wrote:
>>
>>
>> On 08-09-2026 00:58, Ghimiray, Himal Prasad wrote:
>>>
>>>
>>> On 07-09-2026 21:07, Tejas Upadhyay wrote:
>>>> VRAM page offlining can clear a page-table BO's vmap before VM teardown
>>>> calls xe_pt_clear(). Check for a purged BO under its dma-resv lock before
>>>> writing to the mapping.
>>>>
>>>> v2(Sashiko):
>>>> - Fix lock ordering
>>>>
>>>> Signed-off-by: Tejas Upadhyay <tejas.upadhyay@intel.com>
>>>> ---
>>>>    drivers/gpu/drm/xe/xe_pt.c | 6 ++++--
>>>>    drivers/gpu/drm/xe/xe_vm.c | 8 +++++---
>>>>    2 files changed, 9 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
>>>> index 5d990c1c3740..a9b9cd7ac2c8 100644
>>>> --- a/drivers/gpu/drm/xe/xe_pt.c
>>>> +++ b/drivers/gpu/drm/xe/xe_pt.c
>>>> @@ -236,9 +236,11 @@ void xe_pt_destroy(struct xe_pt *pt, u32 flags,
>>>> struct llist_head *deferred)
>>>>     */
>>>>    void xe_pt_clear(struct xe_device *xe, struct xe_pt *pt)
>>>>    {
>>>> -    struct iosys_map *map = &pt->bo->vmap;
>>>> +    struct xe_bo *bo = pt->bo;
>>>> -    xe_map_memset(xe, map, 0, 0, SZ_4K);
>>>> +    xe_bo_assert_held(bo);
>>>> +    if (!xe_bo_is_purged(bo))
>>>> +        xe_map_memset(xe, &bo->vmap, 0, 0, SZ_4K);
>>>>    }
>>>>    /**
>>>> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
>>>> index 9e71567b27cc..e77ff3cf3ca3 100644
>>>> --- a/drivers/gpu/drm/xe/xe_vm.c
>>>> +++ b/drivers/gpu/drm/xe/xe_vm.c
>>>> @@ -1876,10 +1876,13 @@ static void xe_vm_close(struct xe_vm *vm)
>>>>        bound = drm_dev_enter(&xe->drm, &idx);
>>>>        down_write(&vm->lock);
>>>> +    xe_vm_lock(vm, false);
>>>> +
>>>>        if (xe_vm_in_fault_mode(vm))
>>>>            xe_svm_notifier_lock(vm);
>>>> -
>>>>        vm->size = 0;
>>>> +    if (xe_vm_in_fault_mode(vm))
>>>> +        xe_svm_notifier_unlock(vm);
>>>>        if (!((vm->flags & XE_VM_FLAG_MIGRATION))) {
>>>>            struct xe_tile *tile;
>>>> @@ -1901,8 +1904,7 @@ static void xe_vm_close(struct xe_vm *vm)
>>>>            }
>>>>        }
>>>> -    if (xe_vm_in_fault_mode(vm))
>>>> -        xe_svm_notifier_unlock(vm);
>>>> +    xe_vm_unlock(vm);
>>>>        up_write(&vm->lock);
>>>
>>> LGTM
>>> Reviewed-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
>>
>> Sashiko pointed a valid scenario where the check breaks and could lead to
>> NPD.  Please move to iosys_map_is_null check as suggested.
>>
> 
> Yes, indeed. Also if it isn't clear we still need the VM dma-resv lock
> like in this series even though iosys_map_is_null won't have the needed
> lockdep assertions.

+1>
> Matt
> 
>>>
>>>>        if (bound)
>>>
>>