Re: [PATCH v4 28/33] drm/xe: Add SVM VRAM migration

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Wed, 2025-01-29 at 11:52 -0800, Matthew Brost wrote:
> Migration is implemented with range granularity, with VRAM backing
> being
> a VM private TTM BO (i.e., shares dma-resv with VM). The lifetime of
> the
> TTM BO is limited to when the SVM range is in VRAM (i.e., when a VRAM
> SVM range is migrated to SRAM, the TTM BO is destroyed).
> 
> The design choice for using TTM BO for VRAM backing store, as opposed
> to
> direct buddy allocation, is as follows:
> 
> - DRM buddy allocations are not at page granularity, offering no
>   advantage over a BO.
> - Unified eviction is required (SVM VRAM and TTM BOs need to be able
> to
>   evict each other).
> - For exhaustive eviction [1], SVM VRAM allocations will almost
> certainly
>   require a dma-resv.
> - Likely allocation size is 2M which makes of size of BO (872)
>   acceptable per allocation (872 / 2M == .0004158).
> 
> With this, using TTM BO for VRAM backing store seems to be an obvious
> choice as it allows leveraging of the TTM eviction code.
> 
> Current migration policy is migrate any SVM range greater than or
> equal
> to 64k once.
> 
> [1] https://patchwork.freedesktop.org/series/133643/
> 
> v2:
>  - Rebase on latest GPU SVM
>  - Retry page fault on get pages returning mixed allocation
>  - Use drm_gpusvm_devmem
> v3:
>  - Use new BO flags
>  - New range structure (Thomas)
>  - Hide migration behind Kconfig
>  - Kernel doc (Thomas)
>  - Use check_pages_threshold
> v4:
>  - Don't evict partial unmaps in garbage collector (Thomas)
>  - Use %pe to print errors (Thomas)
>  - Use %p to print pointers (Thomas)
> 
> Signed-off-by: Matthew Brost <matthew.brost@intel.com>

Reviewed-by: Thomas Hellström <thomas.hellstrom@lists.freedesktop.org>

> ---
>  drivers/gpu/drm/xe/xe_svm.c | 99
> +++++++++++++++++++++++++++++++++++--
>  drivers/gpu/drm/xe/xe_svm.h |  5 ++
>  2 files changed, 100 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_svm.c
> b/drivers/gpu/drm/xe/xe_svm.c
> index ba1db030bf33..fc030855d078 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -502,7 +502,6 @@ static int xe_svm_populate_devmem_pfn(struct
> drm_gpusvm_devmem *devmem_allocatio
>  	return 0;
>  }
>  
> -__maybe_unused
>  static const struct drm_gpusvm_devmem_ops gpusvm_devmem_ops = {
>  	.devmem_release = xe_svm_devmem_release,
>  	.populate_devmem_pfn = xe_svm_populate_devmem_pfn,
> @@ -582,6 +581,64 @@ static bool xe_svm_range_is_valid(struct
> xe_svm_range *range,
>  	return (range->tile_present & ~range->tile_invalidated) &
> BIT(tile->id);
>  }
>  
> +static struct xe_mem_region *tile_to_mr(struct xe_tile *tile)
> +{
> +	return &tile->mem.vram;
> +}
> +
> +static struct xe_bo *xe_svm_alloc_vram(struct xe_vm *vm, struct
> xe_tile *tile,
> +				       struct xe_svm_range *range,
> +				       const struct drm_gpusvm_ctx
> *ctx)
> +{
> +	struct xe_mem_region *mr = tile_to_mr(tile);
> +	struct drm_buddy_block *block;
> +	struct list_head *blocks;
> +	struct xe_bo *bo;
> +	ktime_t end = 0;
> +	int err;
> +
> +retry:
> +	xe_vm_lock(vm, false);
> +	bo = xe_bo_create(tile_to_xe(tile), tile, vm, range-
> >base.itree.last + 1 -
> +			  range->base.itree.start,
> ttm_bo_type_device,
> +			  XE_BO_FLAG_VRAM_IF_DGFX(tile) |
> +			  XE_BO_FLAG_CPU_ADDR_MIRROR);
> +	xe_vm_unlock(vm);
> +	if (IS_ERR(bo)) {
> +		err = PTR_ERR(bo);
> +		if (xe_vm_validate_should_retry(NULL, err, &end))
> +			goto retry;
> +		return bo;
> +	}
> +
> +	drm_gpusvm_devmem_init(&bo->devmem_allocation,
> +			       vm->xe->drm.dev, vm->svm.gpusvm.mm,
> +			       &gpusvm_devmem_ops,
> +			       &tile->mem.vram.dpagemap,
> +			       range->base.itree.last + 1 -
> +			       range->base.itree.start);
> +
> +	blocks = &to_xe_ttm_vram_mgr_resource(bo->ttm.resource)-
> >blocks;
> +	list_for_each_entry(block, blocks, link)
> +		block->private = mr;
> +
> +	/*
> +	 * Take ref because as soon as drm_gpusvm_migrate_to_devmem
> succeeds the
> +	 * creation ref can be dropped upon CPU fault or unmap.
> +	 */
> +	xe_bo_get(bo);
> +
> +	err = drm_gpusvm_migrate_to_devmem(&vm->svm.gpusvm, &range-
> >base,
> +					   &bo->devmem_allocation,
> ctx);
> +	if (err) {
> +		xe_bo_put(bo);	/* Local ref */
> +		xe_bo_put(bo);	/* Creation ref */
> +		return ERR_PTR(err);
> +	}
> +
> +	return bo;
> +}
> +
>  /**
>   * xe_svm_handle_pagefault() - SVM handle page fault
>   * @vm: The VM.
> @@ -590,7 +647,8 @@ static bool xe_svm_range_is_valid(struct
> xe_svm_range *range,
>   * @fault_addr: The GPU fault address.
>   * @atomic: The fault atomic access bit.
>   *
> - * Create GPU bindings for a SVM page fault.
> + * Create GPU bindings for a SVM page fault. Optionally migrate to
> device
> + * memory.
>   *
>   * Return: 0 on success, negative error code on error.
>   */
> @@ -598,11 +656,18 @@ int xe_svm_handle_pagefault(struct xe_vm *vm,
> struct xe_vma *vma,
>  			    struct xe_tile *tile, u64 fault_addr,
>  			    bool atomic)
>  {
> -	struct drm_gpusvm_ctx ctx = { .read_only =
> xe_vma_read_only(vma), };
> +	struct drm_gpusvm_ctx ctx = {
> +		.read_only = xe_vma_read_only(vma),
> +		.devmem_possible = IS_DGFX(vm->xe) &&
> +			IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR),
> +		.check_pages_threshold = IS_DGFX(vm->xe) &&
> +			IS_ENABLED(CONFIG_DRM_XE_DEVMEM_MIRROR) ?
> SZ_64K : 0,
> +	};
>  	struct xe_svm_range *range;
>  	struct drm_gpusvm_range *r;
>  	struct drm_exec exec;
>  	struct dma_fence *fence;
> +	struct xe_bo *bo = NULL;
>  	ktime_t end = 0;
>  	int err;
>  
> @@ -610,6 +675,9 @@ int xe_svm_handle_pagefault(struct xe_vm *vm,
> struct xe_vma *vma,
>  	xe_assert(vm->xe, xe_vma_is_cpu_addr_mirror(vma));
>  
>  retry:
> +	xe_bo_put(bo);
> +	bo = NULL;
> +
>  	/* Always process UNMAPs first so view SVM ranges is current
> */
>  	err = xe_svm_garbage_collector(vm);
>  	if (err)
> @@ -625,9 +693,31 @@ int xe_svm_handle_pagefault(struct xe_vm *vm,
> struct xe_vma *vma,
>  	if (xe_svm_range_is_valid(range, tile))
>  		return 0;
>  
> +	/* XXX: Add migration policy, for now migrate range once */
> +	if (!range->migrated && range->base.flags.migrate_devmem &&
> +	    (range->base.itree.last + 1 - range->base.itree.start)
> >= SZ_64K) {
> +		range->migrated = true;
> +
> +		bo = xe_svm_alloc_vram(vm, tile, range, &ctx);
> +		if (IS_ERR(bo)) {
> +			drm_info(&vm->xe->drm,
> +				 "VRAM allocation failed, falling
> back to retrying, asid=%u, errno %pe\n",
> +				 vm->usm.asid, bo);
> +			bo = NULL;
> +			goto retry;
> +		}
> +	}
> +
>  	err = drm_gpusvm_range_get_pages(&vm->svm.gpusvm, r, &ctx);
> -	if (err == -EFAULT || err == -EPERM)	/* Corner where CPU
> mappings have changed */
> +	/* Corner where CPU mappings have changed */
> +	if (err == -EOPNOTSUPP || err == -EFAULT || err == -EPERM) {
> +		if (err == -EOPNOTSUPP)
> +			drm_gpusvm_range_evict(&vm->svm.gpusvm,
> &range->base);
> +		drm_info(&vm->xe->drm,
> +			 "Get pages failed, falling back to
> retrying, asid=%u, gpusvm=%p, errno %pe\n",
> +			 vm->usm.asid, &vm->svm.gpusvm,
> ERR_PTR(err));
>  		goto retry;
> +	}
>  	if (err)
>  		goto err_out;
>  
> @@ -658,6 +748,7 @@ int xe_svm_handle_pagefault(struct xe_vm *vm,
> struct xe_vma *vma,
>  	dma_fence_put(fence);
>  
>  err_out:
> +	xe_bo_put(bo);
>  
>  	return err;
>  }
> diff --git a/drivers/gpu/drm/xe/xe_svm.h
> b/drivers/gpu/drm/xe/xe_svm.h
> index 63daffdfdbf6..4c2576162c39 100644
> --- a/drivers/gpu/drm/xe/xe_svm.h
> +++ b/drivers/gpu/drm/xe/xe_svm.h
> @@ -35,6 +35,11 @@ struct xe_svm_range {
>  	 * range. Protected by GPU SVM notifier lock.
>  	 */
>  	u8 tile_invalidated;
> +	/**
> +	 * @migrated: Range has been migrated to device memory,
> protected by
> +	 * GPU fault handler locking.
> +	 */
> +	u8 migrated	:1;
>  };
>  
>  int xe_devm_add(struct xe_tile *tile, struct xe_mem_region *mr);