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

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

On Thu, Feb 19, 2026 at 02:43:08PM +0530, Arvind Yadav wrote:
> Clear XE_VMA_CPU_AUTORESET_ACTIVE before installing GPU PTEs for CPU
> address mirror VMAs.
> 
> This marks the one-way transition from CPU-only to GPU-touched so munmap
> handling can switch from the MADVISE autoreset notifier to the existing
> SVM notifier.
> 
> 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 | 10 ++++++++++
>  drivers/gpu/drm/xe/xe_vm.h  | 11 +++++++++++
>  2 files changed, 21 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index cda3bf7e2418..b9dbbb245779 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -1209,6 +1209,9 @@ static int __xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>  	lockdep_assert_held_write(&vm->lock);
>  	xe_assert(vm->xe, xe_vma_is_cpu_addr_mirror(vma));
>  
> +	/* Invariant: CPU_AUTORESET_ACTIVE cleared before reaching here. */
> +	WARN_ON_ONCE(xe_vma_has_cpu_autoreset_active(vma));
> +
>  	xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_PAGEFAULT_COUNT, 1);
>  
>  retry:
> @@ -1360,6 +1363,13 @@ 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);
> +
> +	/* Transition CPU-only -> GPU-touched before installing PTEs. */
> +	if (xe_vma_has_cpu_autoreset_active(vma))
> +		xe_vma_gpu_touch(vma);

I don’t think this will work going forward. I plan on making the fault
handler run under vm->lock in read mode [1], and VMA state will only be
allowed to be modified under the lockdep constraints in [2], which are
vm->lock in write mode or vm->lock in read mode plus the garbage
collector lock. Maybe this is fine for now, and we can rework it once
[1] and [2] land—most likely by taking the garbage collector mutex
introduced in [1] before touching this VMA’s flags.

Another issue is what happens if we don’t want to taint the VMA unless
we actually fault in a range. It is valid to not find a range to fault
in if this is a prefetch, as that fault just gets suppressed on the
device. So at a minimum, this needs to be moved to where the function
returns zero (i.e., by the out label).

[1] https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/08fa2b95800583e804a91caf477f9c30b3440a33#33e7d2d9323cd529c8d587d9d3801e353439d783_181_179
[2] https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/08fa2b95800583e804a91caf477f9c30b3440a33#71bf077daec46f3ebd785235e8ecc786681aff99_1112_1128


> +
>  retry:
>  	need_vram = xe_vma_need_vram_for_atomic(vm->xe, vma, atomic);
>  	if (need_vram < 0)
> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
> index 7bf400f068ce..3dc549550c91 100644
> --- a/drivers/gpu/drm/xe/xe_vm.h
> +++ b/drivers/gpu/drm/xe/xe_vm.h
> @@ -423,4 +423,15 @@ 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 VMA as GPU-touched
> + * @vma: VMA to mark
> + *
> + * Clear XE_VMA_CPU_AUTORESET_ACTIVE. Must be done before first GPU PTE install.
> + */
> +static inline void xe_vma_gpu_touch(struct xe_vma *vma)
> +{
> +	vma->gpuva.flags &= ~XE_VMA_CPU_AUTORESET_ACTIVE;

Thinking out loud — not strictly related to your series, but I think we
should route all accesses to vma->gpuva.flags through helpers with
lockdep annotations to prove we aren’t violating the rules I mentioned
above (right now this would just require vm->lock in write mode).

Perhaps when I post [1] and [2], I can clean all of that up, or if you
want to transition all access to vma->gpuva.flags to helpers (e.g.,
xe_vma_write_flags(vma, mask), xe_vma_read_flags(vma, mask),
xe_vma_clear_flags(vma, mask)), I wouldn’t complain.

Matt

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