Re: [PATCH v5 20/23] drm/xe: Reset VMA attributes to default in SVM garbage collector

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

On Tue, Jul 22, 2025 at 07:05:23PM +0530, Himal Prasad Ghimiray wrote:
> Restore default memory attributes for VMAs during garbage collection
> if they were modified by madvise. Reuse existing VMA if fully overlapping;
> otherwise, allocate a new mirror VMA.
> 
> Suggested-by: Matthew Brost <matthew.brost@intel.com>
> Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_svm.c |  34 +++++++++
>  drivers/gpu/drm/xe/xe_vm.c  | 140 +++++++++++++++++++++++++-----------
>  drivers/gpu/drm/xe/xe_vm.h  |   2 +
>  3 files changed, 135 insertions(+), 41 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index ba1233d0d5a2..79709dc066b9 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -255,7 +255,18 @@ static int __xe_svm_garbage_collector(struct xe_vm *vm,
>  static int xe_svm_garbage_collector(struct xe_vm *vm)
>  {
>  	struct xe_svm_range *range;
> +	struct xe_vma *vma;
> +	u64 range_start;
> +	u64 range_size;
> +	u64 range_end;
>  	int err;
> +	struct xe_vma_mem_attr default_attr = {
> +		.preferred_loc = {
> +			.devmem_fd = DRM_XE_PREFERRED_LOC_DEFAULT_DEVICE,
> +			.migration_policy = DRM_XE_MIGRATE_ALL_PAGES,
> +		},
> +		.atomic_access = DRM_XE_ATOMIC_UNDEFINED,
> +	};
>  
>  	lockdep_assert_held_write(&vm->lock);
>  
> @@ -270,6 +281,12 @@ static int xe_svm_garbage_collector(struct xe_vm *vm)
>  		if (!range)
>  			break;
>  
> +		range_start = xe_svm_range_start(range);
> +		range_size = xe_svm_range_size(range);
> +		range_end = xe_svm_range_end(range);
> +
> +		vma = xe_vm_find_vma_by_addr(vm, xe_svm_range_start(range));
> +

I'd find the VMA outside of the svm.garbage_collector.lock.

>  		list_del(&range->garbage_collector_link);
>  		spin_unlock(&vm->svm.garbage_collector.lock);
>  
> @@ -282,7 +299,24 @@ static int xe_svm_garbage_collector(struct xe_vm *vm)
>  			return err;
>  		}
>  
> +		if (!xe_vma_has_default_mem_attrs(vma)) {

It seems possible the VMA could be NULL in error cases. I'd check for
NULL and error out.

Also could this code be moved to a helper? Internal SVM seems ok, in
that case xe_vm_find_vma_by_addr could also be in the helper.

> +			vm_dbg(&vm->xe->drm, "Existing VMA start=0x%016llx, vma_end=0x%016llx",
> +			       xe_vma_start(vma), xe_vma_end(vma));
> +
> +			if (xe_vma_start(vma) == range_start && xe_vma_end(vma) == range_end) {
> +				default_attr.pat_index = vma->attr.default_pat_index;
> +				default_attr.default_pat_index  = vma->attr.default_pat_index;
> +				vma->attr = default_attr;
> +			} else {
> +				vm_dbg(&vm->xe->drm, "Split VMA start=0x%016llx, vma_end=0x%016llx",
> +				       range_start, range_end);
> +				err = xe_vm_alloc_cpu_addr_mirror_vma(vm, range_start, range_size);
> +				if (err)

On error, I'd print a message and kill the VM as it shouldn't be
possible to fail aside from a memory allocation failure and we can't
code with errors given this can be inside a worker.

I'll circle back to the rest of the patch a bit later.

Matt

> +					return err;
> +			}
> +		}
>  		spin_lock(&vm->svm.garbage_collector.lock);
> +
>  	}
>  	spin_unlock(&vm->svm.garbage_collector.lock);
>  
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index d3f08bf9a3ee..003c8209f8bd 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -4254,34 +4254,24 @@ bool xe_vma_need_vram_for_atomic(struct xe_device *xe, struct xe_vma *vma, bool
>  	}
>  }
>  
> -/**
> - * xe_vm_alloc_madvise_vma - Allocate VMA's with madvise ops
> - * @vm: Pointer to the xe_vm structure
> - * @start: Starting input address
> - * @range: Size of the input range
> - *
> - * This function splits existing vma to create new vma for user provided input range
> - *
> - *  Return: 0 if success
> - */
> -int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
> +static int xe_vm_alloc_vma(struct xe_vm *vm,
> +			   u64 start, u64 range,
> +			   enum drm_gpuvm_sm_map_ops_flags flags)
>  {
>  	struct xe_vma_ops vops;
>  	struct drm_gpuva_ops *ops = NULL;
>  	struct drm_gpuva_op *__op;
>  	bool is_cpu_addr_mirror = false;
>  	bool remap_op = false;
> +	bool is_madvise = flags == DRM_GPUVM_SKIP_GEM_OBJ_VA_SPLIT_MADVISE;
>  	struct xe_vma_mem_attr tmp_attr;
> +	u16 default_pat;
>  	int err;
>  
> -	vm_dbg(&vm->xe->drm, "MADVISE IN: addr=0x%016llx, size=0x%016llx", start, range);
> -
>  	lockdep_assert_held_write(&vm->lock);
>  
> -	vm_dbg(&vm->xe->drm, "MADVISE_OPS_CREATE: addr=0x%016llx, size=0x%016llx", start, range);
>  	ops = drm_gpuvm_sm_map_ops_create(&vm->gpuvm, start, range,
> -					  DRM_GPUVM_SKIP_GEM_OBJ_VA_SPLIT_MADVISE,
> -					  NULL, start);
> +					  flags, NULL, start);
>  	if (IS_ERR(ops))
>  		return PTR_ERR(ops);
>  
> @@ -4292,33 +4282,56 @@ int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
>  
>  	drm_gpuva_for_each_op(__op, ops) {
>  		struct xe_vma_op *op = gpuva_op_to_vma_op(__op);
> +		struct xe_vma *vma = NULL;
>  
> -		if (__op->op == DRM_GPUVA_OP_REMAP) {
> -			xe_assert(vm->xe, !remap_op);
> -			remap_op = true;
> +		if (!is_madvise) {
> +			if (__op->op == DRM_GPUVA_OP_UNMAP) {
> +				vma = gpuva_to_vma(op->base.unmap.va);
> +				XE_WARN_ON(!xe_vma_has_default_mem_attrs(vma));
> +				default_pat = vma->attr.default_pat_index;
> +			}
>  
> -			if (xe_vma_is_cpu_addr_mirror(gpuva_to_vma(op->base.remap.unmap->va)))
> -				is_cpu_addr_mirror = true;
> -			else
> -				is_cpu_addr_mirror = false;
> -		}
> +			if (__op->op == DRM_GPUVA_OP_REMAP) {
> +				vma = gpuva_to_vma(op->base.remap.unmap->va);
> +				default_pat = vma->attr.default_pat_index;
> +			}
>  
> -		if (__op->op == DRM_GPUVA_OP_MAP) {
> -			xe_assert(vm->xe, remap_op);
> -			remap_op = false;
> +			if (__op->op == DRM_GPUVA_OP_MAP) {
> +				op->map.is_cpu_addr_mirror = true;
> +				op->map.pat_index = default_pat;
> +			}
> +		} else {
> +			if (__op->op == DRM_GPUVA_OP_REMAP) {
> +				vma = gpuva_to_vma(op->base.remap.unmap->va);
> +				xe_assert(vm->xe, !remap_op);
> +				remap_op = true;
>  
> -			/* In case of madvise ops DRM_GPUVA_OP_MAP is always after
> -			 * DRM_GPUVA_OP_REMAP, so ensure we assign op->map.is_cpu_addr_mirror true
> -			 * if REMAP is for xe_vma_is_cpu_addr_mirror vma
> -			 */
> -			op->map.is_cpu_addr_mirror = is_cpu_addr_mirror;
> -		}
> +				if (xe_vma_is_cpu_addr_mirror(vma))
> +					is_cpu_addr_mirror = true;
> +				else
> +					is_cpu_addr_mirror = false;
> +			}
>  
> +			if (__op->op == DRM_GPUVA_OP_MAP) {
> +				xe_assert(vm->xe, remap_op);
> +				remap_op = false;
> +				/*
> +				 * In case of madvise ops DRM_GPUVA_OP_MAP is
> +				 * always after DRM_GPUVA_OP_REMAP, so ensure
> +				 * we assign op->map.is_cpu_addr_mirror true
> +				 * if REMAP is for xe_vma_is_cpu_addr_mirror vma
> +				 */
> +				op->map.is_cpu_addr_mirror = is_cpu_addr_mirror;
> +			}
> +		}
>  		print_op(vm->xe, __op);
>  	}
>  
>  	xe_vma_ops_init(&vops, vm, NULL, NULL, 0);
> -	vops.flags |= XE_VMA_OPS_FLAG_MADVISE;
> +
> +	if (is_madvise)
> +		vops.flags |= XE_VMA_OPS_FLAG_MADVISE;
> +
>  	err = vm_bind_ioctl_ops_parse(vm, ops, &vops);
>  	if (err)
>  		goto unwind_ops;
> @@ -4330,15 +4343,20 @@ int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
>  		struct xe_vma *vma;
>  
>  		if (__op->op == DRM_GPUVA_OP_UNMAP) {
> -			/* There should be no unmap */
> -			XE_WARN_ON("UNEXPECTED UNMAP");
> -			xe_vma_destroy(gpuva_to_vma(op->base.unmap.va), NULL);
> +			vma = gpuva_to_vma(op->base.unmap.va);
> +			/* There should be no unmap for madvise */
> +			if (is_madvise)
> +				XE_WARN_ON("UNEXPECTED UNMAP");
> +
> +			xe_vma_destroy(vma, NULL);
>  		} else if (__op->op == DRM_GPUVA_OP_REMAP) {
>  			vma = gpuva_to_vma(op->base.remap.unmap->va);
> -			/* Store attributes for REMAP UNMAPPED VMA, so they can be assigned
> -			 * to newly MAP created vma.
> +			/* In case of madvise ops Store attributes for REMAP UNMAPPED
> +			 * VMA, so they can be assigned to newly MAP created vma.
>  			 */
> -			tmp_attr = vma->attr;
> +			if (is_madvise)
> +				tmp_attr = vma->attr;
> +
>  			xe_vma_destroy(gpuva_to_vma(op->base.remap.unmap->va), NULL);
>  		} else if (__op->op == DRM_GPUVA_OP_MAP) {
>  			vma = op->map.vma;
> @@ -4346,7 +4364,8 @@ int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
>  			 * Therefore temp_attr will always have sane values, making it safe to
>  			 * copy them to new vma.
>  			 */
> -			vma->attr = tmp_attr;
> +			if (is_madvise)
> +				vma->attr = tmp_attr;
>  		}
>  	}
>  
> @@ -4360,3 +4379,42 @@ int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
>  	drm_gpuva_ops_free(&vm->gpuvm, ops);
>  	return err;
>  }
> +
> +/**
> + * xe_vm_alloc_madvise_vma - Allocate VMA's with madvise ops
> + * @vm: Pointer to the xe_vm structure
> + * @start: Starting input address
> + * @range: Size of the input range
> + *
> + * This function splits existing vma to create new vma for user provided input range
> + *
> + * Return: 0 if success
> + */
> +int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
> +{
> +	lockdep_assert_held_write(&vm->lock);
> +
> +	vm_dbg(&vm->xe->drm, "MADVISE_OPS_CREATE: addr=0x%016llx, size=0x%016llx", start, range);
> +
> +	return xe_vm_alloc_vma(vm, start, range, DRM_GPUVM_SKIP_GEM_OBJ_VA_SPLIT_MADVISE);
> +}
> +
> +/**
> + * xe_vm_alloc_cpu_addr_mirror_vma - Allocate CPU addr mirror vma
> + * @vm: Pointer to the xe_vm structure
> + * @start: Starting input address
> + * @range: Size of the input range
> + *
> + * This function splits/merges existing vma to create new vma for user provided input range
> + *
> + * Return: 0 if success
> + */
> +int xe_vm_alloc_cpu_addr_mirror_vma(struct xe_vm *vm, uint64_t start, uint64_t range)
> +{
> +	lockdep_assert_held_write(&vm->lock);
> +
> +	vm_dbg(&vm->xe->drm, "CPU_ADDR_MIRROR_VMA_OPS_CREATE: addr=0x%016llx, size=0x%016llx",
> +	       start, range);
> +
> +	return xe_vm_alloc_vma(vm, start, range, DRM_GPUVM_SM_MAP_NOT_MADVISE);
> +}
> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
> index a4db843de540..f7b9ad83685a 100644
> --- a/drivers/gpu/drm/xe/xe_vm.h
> +++ b/drivers/gpu/drm/xe/xe_vm.h
> @@ -177,6 +177,8 @@ bool xe_vma_need_vram_for_atomic(struct xe_device *xe, struct xe_vma *vma, bool
>  
>  int xe_vm_alloc_madvise_vma(struct xe_vm *vm, uint64_t addr, uint64_t size);
>  
> +int xe_vm_alloc_cpu_addr_mirror_vma(struct xe_vm *vm, uint64_t addr, uint64_t size);
> +
>  /**
>   * to_userptr_vma() - Return a pointer to an embedding userptr vma
>   * @vma: Pointer to the embedded struct xe_vma
> -- 
> 2.34.1
>