Re: [RFC v2 3/7] drm/xe/svm: Clear CPU_AUTORESET_ACTIVE on first GPU fault

[Matthew Brost <matthew.brost@intel.com>]

On Mon, Apr 06, 2026 at 02:28:26PM +0530, Arvind Yadav wrote:
> CPU address mirror VMAs start with cpu_autoreset_active set, indicating
> they are still CPU-only.
> 
> Clear cpu_autoreset_active after the first successful GPU fault so
> subsequent munmap follows the SVM path instead of the autoreset path.
> 
> Do this in xe_svm_handle_pagefault() on the success path only.
> Prefetch faults that install no PTEs must not transition this state.
> 
> v2:
>   - Move xe_vma_gpu_touch() to the success path in
>     xe_svm_handle_pagefault() so prefetch faults that find no range do
>     not transition the state. (Matt)
>   - Add xe_vma_gpu_touch() helper in xe_vm.h and use
>     vma->cpu_autoreset_active instead of vma->gpuva.flags. (Matt)
> 
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> Signed-off-by: Arvind Yadav <arvind.yadav@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_svm.c | 11 +++++++++++
>  drivers/gpu/drm/xe/xe_vm.h  | 13 +++++++++++++
>  2 files changed, 24 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index 5933b2b6392b..fd57c9d41db8 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -1388,6 +1388,9 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>  			    bool atomic)
>  {
>  	int need_vram, ret;
> +
> +	lockdep_assert_held_write(&vm->lock);
> +

This looks unrelated... and will change here [1] to read mode.

[1] https://patchwork.freedesktop.org/patch/707294/?series=162167&rev=4

So I'd say drop this part.

>  retry:
>  	need_vram = xe_vma_need_vram_for_atomic(vm->xe, vma, atomic);
>  	if (need_vram < 0)
> @@ -1406,6 +1409,14 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>  
>  		goto retry;
>  	}
> +
> +	/*
> +	 * Mark the VMA as GPU-touched only after a successful fault-in.
> +	 * Prefetch faults that find no range must not transition this state.
> +	 */
> +	if (!ret && xe_vma_has_cpu_autoreset_active(vma))
> +		xe_vma_gpu_touch(vma);
> +
>  	return ret;
>  }
>  
> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
> index bdf42083da86..8d45f896f90b 100644
> --- a/drivers/gpu/drm/xe/xe_vm.h
> +++ b/drivers/gpu/drm/xe/xe_vm.h
> @@ -441,4 +441,17 @@ static inline struct drm_exec *xe_vm_validation_exec(struct xe_vm *vm)
>  	((READ_ONCE(tile_present) & ~READ_ONCE(tile_invalidated)) & BIT((tile)->id))
>  
>  void xe_vma_mem_attr_copy(struct xe_vma_mem_attr *to, struct xe_vma_mem_attr *from);
> +
> +/**
> + * xe_vma_gpu_touch() - Mark a VMA as no longer CPU-only
> + * @vma: VMA to update
> + *
> + * Clear cpu_autoreset_active after the first successful GPU fault-in.
> + * Caller must hold vm->lock in write mode.
> + */
> +static inline void xe_vma_gpu_touch(struct xe_vma *vma)
> +{
> +	lockdep_assert_held_write(&xe_vma_vm(vma)->lock);
> +	vma->cpu_autoreset_active = false;

Write mode VM lock seems fine for now, but will likely have to change
this to vma->fault_lock with [1]...

Also IMO the caller adjusting this is too late. Adjusting auto-reset
likely should be basically be the first thing handler does, moved to the
xe_pagefault layer... I noticed in [1] the VMA more than likely needs to
be refcounted but that's a different issue.

Lastly - when we call xe_vma_gpu_touch and go from cpu_autoreset_active
1 -> 0 shouldn't we remove the auto-reset notifier [2] [3]? Keeping
notifier around is a non-zero cost as these are entiries in interval
tree the core MM walks for every notifier trigger, so I think deleting
the notifier part makes sense if this is possible - still wrapping my
head around [2] [3], so take this as a suggestion.

Matt

[2] https://patchwork.freedesktop.org/patch/716552/?series=161815&rev=2
[3] https://patchwork.freedesktop.org/patch/716553/?series=161815&rev=2

> +}
>  #endif
> -- 
> 2.43.0
>