Re: [PATCH v4 3/4] drm/xe: Bump prefetch BO LRU via GPUSVM devmem callback

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

On Thu, Oct 01, 2026 at 09:06:06AM +0530, Himal Prasad Ghimiray wrote:
> Add xe_svm_devmem_lru_bump() and set it as drm_gpusvm_ctx.devmem_fn in
> the prefetch worker so each backing BO is moved to the LRU tail during
> get_pages.
> 
> v2:
> - Guard xe_svm_devmem_lru_bump() against foreign device-memory
>   allocations before casting to xe_bo: return early when
>   devmem_allocation->ops != &dpagemap_devmem_ops. A shared
>   device_private_page_owner can hand us another driver's allocation.
>   (Matthew Brost / Sashiko AI review)
> - Add stub for function (CI)
> 
> v3:
> - Guard xe_svm_devmem_lru_bump() body with CONFIG_DRM_XE_PAGEMAP; it
>   references dpagemap_devmem_ops and to_xe_bo(), which only exist when
>   PAGEMAP is enabled. Fixes the build with GPUSVM=y, PAGEMAP=n.
>   (Sashiko AI review)
> 
> v4:
> - Fix xe_svm_devmem_lru_bump() kernel-doc to explain the locking (Matt)
> 
> Suggested-by: Matthew Brost <matthew.brost@intel.com>

Reviewed-by: Matthew Brost <matthew.brost@intel.com>

Sashiko comment for series are theoretically possible with different
coherent devicesm one of which uses GPUSVM, another of which does not
but I don't think we need to solve theoretical problem today, nor is
this in scope.

Himal - go ahead and merge, ack on doing this through drm-xe-next.

Matt

> Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_svm.c | 34 ++++++++++++++++++++++++++++++++++
>  drivers/gpu/drm/xe/xe_svm.h |  7 +++++++
>  drivers/gpu/drm/xe/xe_vm.c  |  1 +
>  3 files changed, 42 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index f39e647512ad..d07dd95ab451 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -9,6 +9,7 @@
>  #include <drm/drm_managed.h>
>  #include <drm/drm_pagemap.h>
>  #include <drm/drm_pagemap_util.h>
> +#include <drm/ttm/ttm_bo.h>
>  
>  #include "xe_bo.h"
>  #include "xe_exec_queue_types.h"
> @@ -1625,6 +1626,39 @@ int xe_svm_range_get_pages(struct xe_vm *vm, struct xe_svm_range *range,
>  	return err;
>  }
>  
> +/**
> + * xe_svm_devmem_lru_bump() - Move a range's backing BO to the TTM LRU tail
> + * @devmem_allocation: The device-memory allocation backing the range's pages
> + *
> + * Intended as a &drm_gpusvm_ctx.devmem_fn, so drm_gpusvm_get_pages() calls it
> + * under the GPUSVM notifier lock. That lock sits in the reclaim path, so
> + * dma-resv cannot be acquired blocking beneath it; dma-resv is also the outer
> + * lock of the notifier when memory is allocated and migrated, so a blocking
> + * acquire here would invert that order. Hence the trylock. The BO is private
> + * (dedicated dma-resv), so the trylock succeeds unless the BO is being moved
> + * elsewhere (e.g. eviction); a contended BO is simply skipped.
> + *
> + * A shared device_private_page_owner can hand us allocations backed by a
> + * foreign driver; skip anything that is not an Xe BO before casting.
> + */
> +void xe_svm_devmem_lru_bump(struct drm_pagemap_devmem *devmem_allocation)
> +{
> +#if IS_ENABLED(CONFIG_DRM_XE_PAGEMAP)
> +	struct xe_bo *bo;
> +
> +	if (devmem_allocation->ops != &dpagemap_devmem_ops)
> +		return;
> +
> +	bo = to_xe_bo(devmem_allocation);
> +
> +	if (!dma_resv_trylock(bo->ttm.base.resv))
> +		return;
> +
> +	ttm_bo_move_to_lru_tail_unlocked(&bo->ttm);
> +	dma_resv_unlock(bo->ttm.base.resv);
> +#endif
> +}
> +
>  /**
>   * xe_svm_ranges_zap_ptes_in_range - clear ptes of svm ranges in input range
>   * @vm: Pointer to the xe_vm structure
> diff --git a/drivers/gpu/drm/xe/xe_svm.h b/drivers/gpu/drm/xe/xe_svm.h
> index 2ef4ef026ccd..585ad64b672b 100644
> --- a/drivers/gpu/drm/xe/xe_svm.h
> +++ b/drivers/gpu/drm/xe/xe_svm.h
> @@ -128,6 +128,8 @@ struct xe_svm_range *xe_svm_range_find_or_insert(struct xe_vm *vm, u64 addr,
>  int xe_svm_range_get_pages(struct xe_vm *vm, struct xe_svm_range *range,
>  			   struct drm_gpusvm_ctx *ctx);
>  
> +void xe_svm_devmem_lru_bump(struct drm_pagemap_devmem *devmem_allocation);
> +
>  bool xe_svm_range_needs_migrate_to_vram(struct xe_svm_range *range, struct xe_vma *vma,
>  					const struct drm_pagemap *dpagemap);
>  
> @@ -356,6 +358,11 @@ int xe_svm_range_get_pages(struct xe_vm *vm, struct xe_svm_range *range,
>  	return -EINVAL;
>  }
>  
> +static inline
> +void xe_svm_devmem_lru_bump(struct drm_pagemap_devmem *devmem_allocation)
> +{
> +}
> +
>  static inline struct xe_svm_range *to_xe_range(struct drm_gpusvm_range *r)
>  {
>  	return NULL;
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index ea5bfb307db1..986e0ebfdacd 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -3246,6 +3246,7 @@ static int prefetch_ranges(struct xe_vm *vm, struct xe_vma_ops *vops,
>  	ctx.devmem_possible = devmem_possible;
>  	ctx.check_pages_threshold = devmem_possible ? SZ_64K : 0;
>  	ctx.device_private_page_owner = xe_svm_private_page_owner(vm, !dpagemap);
> +	ctx.devmem_fn = xe_svm_devmem_lru_bump;
>  
>  	skip_threads =  op->prefetch_range.ranges_count == 1 ||
>  		(!dpagemap && !(vops->flags &
> -- 
> 2.43.0
>