Re: [PATCH v5 23/25] drm/xe: Reset VMA attributes to default in SVM garbage collector

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

On Wed, Jul 30, 2025 at 06:30:48PM +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.
> 
> v2 (Matthew Brost)
> - Add helper for vma split
> - Add retry to get updated 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 | 114 +++++++++++++++++++++-----
>  drivers/gpu/drm/xe/xe_vm.c  | 155 ++++++++++++++++++++++++++----------
>  drivers/gpu/drm/xe/xe_vm.h  |   2 +
>  3 files changed, 206 insertions(+), 65 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index aef76e08b460..9b3a3f61758c 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -253,9 +253,55 @@ static int __xe_svm_garbage_collector(struct xe_vm *vm,
>  	return 0;
>  }
>  
> +static int xe_svm_range_set_default_attr(struct xe_vm *vm, u64 range_start, u64 range_end)
> +{
> +	struct xe_vma *vma;
> +	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,
> +	};
> +	int err = 0;
> +
> +	vma = xe_vm_find_vma_by_addr(vm, range_start);
> +	if (!vma)
> +		return -EINVAL;
> +
> +	if (xe_vma_has_default_mem_attrs(vma))
> +		return 0;
> +
> +	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_end - range_start);
> +		if (err) {
> +			drm_warn(&vm->xe->drm, "VMA SPLIT failed: %pe\n", ERR_PTR(err));
> +			xe_vm_kill(vm, true);
> +			return err;
> +		}
> +	}
> +
> +	/*
> +	 * On call from xe_svm_handle_pagefault original VMA might be changed
> +	 * signal this to lookup for VMA again.
> +	 */
> +	return -EAGAIN;
> +}
> +
>  static int xe_svm_garbage_collector(struct xe_vm *vm)
>  {
>  	struct xe_svm_range *range;
> +	u64 range_start;
> +	u64 range_end;
>  	int err;
>  
>  	lockdep_assert_held_write(&vm->lock);
> @@ -271,6 +317,9 @@ static int xe_svm_garbage_collector(struct xe_vm *vm)
>  		if (!range)
>  			break;
>  
> +		range_start = xe_svm_range_start(range);
> +		range_end = xe_svm_range_end(range);
> +
>  		list_del(&range->garbage_collector_link);
>  		spin_unlock(&vm->svm.garbage_collector.lock);
>  
> @@ -283,6 +332,10 @@ static int xe_svm_garbage_collector(struct xe_vm *vm)
>  			return err;
>  		}
>  
> +		err = xe_svm_range_set_default_attr(vm, range_start, range_end);
> +		if (err)
> +			return err;

You don't want to return on -EAGAIN here, rather collect it, continue
and return -EAGAIN once the garbage collector list is empty. No need to
contiously lookup the VMA in xe_svm_handle_pagefault (in next rev
__xe_svm_handle_pagefault), this only need be done once.

> +
>  		spin_lock(&vm->svm.garbage_collector.lock);
>  	}
>  	spin_unlock(&vm->svm.garbage_collector.lock);
> @@ -793,40 +846,59 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>  			    struct xe_gt *gt, u64 fault_addr,
>  			    bool atomic)
>  {
> -	int need_vram = xe_vma_need_vram_for_atomic(vm->xe, vma, atomic);
> -
> -	if (need_vram < 0)
> -		return need_vram;
> -
> -	struct drm_gpusvm_ctx ctx = {
> -		.read_only = xe_vma_read_only(vma),
> -		.devmem_possible = IS_DGFX(vm->xe) &&
> -			IS_ENABLED(CONFIG_DRM_XE_PAGEMAP),
> -		.check_pages_threshold = IS_DGFX(vm->xe) &&
> -			IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) ? SZ_64K : 0,
> -		.devmem_only = need_vram && IS_ENABLED(CONFIG_DRM_XE_PAGEMAP),
> -		.timeslice_ms = atomic && IS_DGFX(vm->xe) &&
> -			IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) ?
> -			vm->xe->atomic_svm_timeslice_ms : 0,
> -	};
> +	struct drm_gpusvm_ctx ctx = { };
> +	struct drm_pagemap *dpagemap;
>  	struct xe_svm_range *range;
>  	struct dma_fence *fence;
> -	struct drm_pagemap *dpagemap;
>  	struct xe_tile *tile = gt_to_tile(gt);
> -	int migrate_try_count = ctx.devmem_only ? 3 : 1;
> +	bool vma_updated = false;
> +	int need_vram;
> +	int migrate_try_count;
>  	ktime_t end = 0;
>  	int err;
>  
> -	lockdep_assert_held_write(&vm->lock);
> +find_vma:
> +	if (vma_updated) {
> +		vma = xe_vm_find_vma_by_addr(vm, fault_addr);
> +		if (!vma)
> +			return -EINVAL;
> +	}
> +
>  	xe_assert(vm->xe, xe_vma_is_cpu_addr_mirror(vma));
> +	vma_updated = false;
> +
> +	need_vram = xe_vma_need_vram_for_atomic(vm->xe, vma, atomic);
> +	if (need_vram < 0)
> +		return need_vram;

This is a bit ugly. I think if you have __xe_svm_handle_pagefault and
xe_svm_handle_pagefault as here [1] this can be handled cleaner (i.e.
still a static setup of drm_gpusvm_ctx).

If xe_svm_garbage_collector returns an in __xe_svm_handle_pagefault kick
it up to xe_svm_handle_pagefault, you catch -EAGAIN there, relookup the
VMA and call __xe_svm_handle_pagefault again. I think that would look
quite a bit better.

Matt

[1] https://patchwork.freedesktop.org/patch/666222/?series=149550&rev=5#comment_1222471 

> +
> +	ctx.read_only = xe_vma_read_only(vma);
> +	ctx.devmem_possible = IS_DGFX(vm->xe) && IS_ENABLED(CONFIG_DRM_XE_PAGEMAP);
> +	ctx.check_pages_threshold = IS_DGFX(vm->xe) && IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) ?
> +				    SZ_64K : 0;
> +	ctx.devmem_only = need_vram && IS_ENABLED(CONFIG_DRM_XE_PAGEMAP);
> +	ctx.timeslice_ms = atomic && IS_DGFX(vm->xe) && IS_ENABLED(CONFIG_DRM_XE_PAGEMAP) ?
> +			   vm->xe->atomic_svm_timeslice_ms : 0;
>  
> +	migrate_try_count = ctx.devmem_only ? 3 : 1;
> +
> +	lockdep_assert_held_write(&vm->lock);
>  	xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_PAGEFAULT_COUNT, 1);
>  
>  retry:
>  	/* Always process UNMAPs first so view SVM ranges is current */
>  	err = xe_svm_garbage_collector(vm);
> -	if (err)
> -		return err;
> +	if (err) {
> +		if (err == -EAGAIN) {
> +			/*
> +			 * VMA might have changed due to garbage
> +			 * collection; retry lookup
> +			 */
> +			vma_updated = true;
> +			goto find_vma;
> +		} else {
> +			return err;
> +		}
> +	}
>  
>  	range = xe_svm_range_find_or_insert(vm, fault_addr, vma, &ctx);
>  
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index 5ee38e9cf6c6..e77c04f92d0b 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -4263,36 +4263,24 @@ int xe_vma_need_vram_for_atomic(struct xe_device *xe, struct xe_vma *vma, bool i
>  	}
>  }
>  
> -/**
> - * 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, struct drm_gpuva_op_map *map_req)
>  {
> -	struct drm_gpuva_op_map map_req = {
> -		.va.addr = start,
> -		.va.range = range,
> -		.flags = DRM_GPUVM_SM_MAP_OPS_FLAG_SPLIT_MADVISE,
> -	};
> -
>  	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 = (map_req->flags & DRM_GPUVM_SM_MAP_OPS_FLAG_SPLIT_MADVISE);
>  	struct xe_vma_mem_attr tmp_attr;
> +	u16 default_pat;
>  	int err;
>  
>  	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, &map_req);
> +	vm_dbg(&vm->xe->drm, "MADVISE_OPS_CREATE: addr=0x%016llx, size=0x%016llx",
> +	       map_req->va.addr, map_req->va.range);
> +
> +	ops = drm_gpuvm_sm_map_ops_create(&vm->gpuvm, map_req);
>  	if (IS_ERR(ops))
>  		return PTR_ERR(ops);
>  
> @@ -4303,33 +4291,56 @@ 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;
> @@ -4341,15 +4352,20 @@ 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;
> @@ -4357,7 +4373,8 @@ 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;
>  		}
>  	}
>  
> @@ -4371,3 +4388,53 @@ 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)
> +{
> +	struct drm_gpuva_op_map map_req = {
> +		.va.addr = start,
> +		.va.range = range,
> +		.flags = DRM_GPUVM_SM_MAP_OPS_FLAG_SPLIT_MADVISE,
> +	};
> +
> +	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, &map_req);
> +}
> +
> +/**
> + * 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)
> +{
> +	struct drm_gpuva_op_map map_req = {
> +		.va.addr = start,
> +		.va.range = 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, &map_req);
> +}
> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
> index f735d994806d..6538cddf158b 100644
> --- a/drivers/gpu/drm/xe/xe_vm.h
> +++ b/drivers/gpu/drm/xe/xe_vm.h
> @@ -177,6 +177,8 @@ int xe_vma_need_vram_for_atomic(struct xe_device *xe, struct xe_vma *vma, bool i
>  
>  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
>