[PATCH v2 5/5] drm/xe/uapi, drm/xe: Make the PT code handle placement per PTE rather than per vma / range

Matthew Brost matthew.brost at intel.com
Fri Mar 21 17:16:38 UTC 2025


On Fri, Mar 21, 2025 at 05:34:16PM +0100, Thomas Hellström wrote:
> With SVM, ranges forwarded to the PT code for binding can, mostly
> due to races when migrating, point to both VRAM and system / foreign
> device memory. Make the PT code able to handle that by checking,
> for each PTE set up, whether it points to local VRAM or to system
> memory.
> 
> The UAPI is changed implicitly in that before this patch,
> global atomics required a bo with VRAM/System placements. With
> this patch that is changed to requiring LR mode, and
> if the required placement is not available upon GPU atomic access
> pagefault, an error will be generated and the VM banned.

I think the queue will get banned if I'm reading the code correctly.

> 
> v2:
> - Fix system memory GPU atomic access.
> 
> Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> ---
>  drivers/gpu/drm/xe/xe_bo.c |  12 +++--
>  drivers/gpu/drm/xe/xe_pt.c | 106 ++++++++++++++++---------------------
>  2 files changed, 56 insertions(+), 62 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> index 9d043d2c30fd..3c7c2353d3c8 100644
> --- a/drivers/gpu/drm/xe/xe_bo.c
> +++ b/drivers/gpu/drm/xe/xe_bo.c
> @@ -2116,10 +2116,16 @@ uint64_t vram_region_gpu_offset(struct ttm_resource *res)
>  {
>  	struct xe_device *xe = ttm_to_xe_device(res->bo->bdev);
>  
> -	if (res->mem_type == XE_PL_STOLEN)
> +	switch (res->mem_type) {
> +	case XE_PL_STOLEN:
>  		return xe_ttm_stolen_gpu_offset(xe);
> -
> -	return res_to_mem_region(res)->dpa_base;
> +	case XE_PL_TT:
> +	case XE_PL_SYSTEM:
> +		return 0;
> +	default:
> +		return res_to_mem_region(res)->dpa_base;
> +	}
> +	return 0;
>  }
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
> index 9e719535a3bb..3ffe135c27f1 100644
> --- a/drivers/gpu/drm/xe/xe_pt.c
> +++ b/drivers/gpu/drm/xe/xe_pt.c
> @@ -278,13 +278,15 @@ struct xe_pt_stage_bind_walk {
>  	struct xe_vm *vm;
>  	/** @tile: The tile we're building for. */
>  	struct xe_tile *tile;
> -	/** @default_pte: PTE flag only template. No address is associated */
> -	u64 default_pte;
> +	/** @default_pte: PTE flag only template for VRAM. No address is associated */
> +	u64 default_vram_pte;
> +	/** @default_pte: PTE flag only template for VRAM. No address is associated */
> +	u64 default_system_pte;
>  	/** @dma_offset: DMA offset to add to the PTE. */
>  	u64 dma_offset;
>  	/**
>  	 * @needs_64k: This address range enforces 64K alignment and
> -	 * granularity.
> +	 * granularity on VRAM.
>  	 */
>  	bool needs_64K;
>  	/**
> @@ -515,13 +517,16 @@ xe_pt_stage_bind_entry(struct xe_ptw *parent, pgoff_t offset,
>  	if (level == 0 || xe_pt_hugepte_possible(addr, next, level, xe_walk)) {
>  		struct xe_res_cursor *curs = xe_walk->curs;
>  		bool is_null = xe_vma_is_null(xe_walk->vma);
> +		bool is_vram = is_null ? false : xe_res_is_vram(curs);

So xe_res_is_vram is still per range / VMA. I assume a follow on will
change this to be per page? 

Patch LGTM though. With that:
Reviewed-by: Matthew Brost <matthew.brost at intel.com>

>  
>  		XE_WARN_ON(xe_walk->va_curs_start != addr);
>  
>  		pte = vm->pt_ops->pte_encode_vma(is_null ? 0 :
>  						 xe_res_dma(curs) + xe_walk->dma_offset,
>  						 xe_walk->vma, pat_index, level);
> -		pte |= xe_walk->default_pte;
> +		if (!is_null)
> +			pte |= is_vram ? xe_walk->default_vram_pte :
> +				xe_walk->default_system_pte;
>  
>  		/*
>  		 * Set the XE_PTE_PS64 hint if possible, otherwise if
> @@ -531,7 +536,7 @@ xe_pt_stage_bind_entry(struct xe_ptw *parent, pgoff_t offset,
>  			if (xe_pt_is_pte_ps64K(addr, next, xe_walk)) {
>  				xe_walk->vma->gpuva.flags |= XE_VMA_PTE_64K;
>  				pte |= XE_PTE_PS64;
> -			} else if (XE_WARN_ON(xe_walk->needs_64K)) {
> +			} else if (XE_WARN_ON(xe_walk->needs_64K && is_vram)) {
>  				return -EINVAL;
>  			}
>  		}
> @@ -603,6 +608,31 @@ static const struct xe_pt_walk_ops xe_pt_stage_bind_ops = {
>  	.pt_entry = xe_pt_stage_bind_entry,
>  };
>  
> +/* The GPU can always to atomics in VRAM */
> +static bool xe_atomic_for_vram(struct xe_vm *vm)
> +{
> +	return true;
> +}
> +
> +/*
> + * iGFX always expect to be able to do atomics in system.
> + *
> + * For DGFX, 3D clients want to do atomics in system that is
> + * not coherent with CPU atomics. Compute clients want
> + * atomics that look coherent with CPU atomics. We
> + * distinguish the two by checking for lr mode. For
> + * compute we then disallow atomics in system.
> + * Compute attempts to perform atomics in system memory would
> + * then cause an unrecoverable page-fault in preempt-fence
> + * mode, but in fault mode the data would be migrated to VRAM
> + * for GPU atomics and to system for CPU atomics.
> + */
> +static bool xe_atomic_for_system(struct xe_vm *vm)
> +{
> +	return (!IS_DGFX(vm->xe) || !xe_vm_in_lr_mode(vm)) &&
> +		vm->xe->info.has_device_atomics_on_smem;
> +}
> +
>  /**
>   * xe_pt_stage_bind() - Build a disconnected page-table tree for a given address
>   * range.
> @@ -629,9 +659,8 @@ xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
>  {
>  	struct xe_device *xe = tile_to_xe(tile);
>  	struct xe_bo *bo = xe_vma_bo(vma);
> -	bool is_devmem = !xe_vma_is_userptr(vma) && bo &&
> -		(xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo));
>  	struct xe_res_cursor curs;
> +	struct xe_vm *vm = xe_vma_vm(vma);
>  	struct xe_pt_stage_bind_walk xe_walk = {
>  		.base = {
>  			.ops = &xe_pt_stage_bind_ops,
> @@ -639,7 +668,7 @@ xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
>  			.max_level = XE_PT_HIGHEST_LEVEL,
>  			.staging = true,
>  		},
> -		.vm = xe_vma_vm(vma),
> +		.vm = vm,
>  		.tile = tile,
>  		.curs = &curs,
>  		.va_curs_start = range ? range->base.itree.start :
> @@ -647,26 +676,22 @@ xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
>  		.vma = vma,
>  		.wupd.entries = entries,
>  	};
> -	struct xe_pt *pt = xe_vma_vm(vma)->pt_root[tile->id];
> +	struct xe_pt *pt = vm->pt_root[tile->id];
>  	int ret;
>  
>  	if (range) {
>  		/* Move this entire thing to xe_svm.c? */
> -		xe_svm_notifier_lock(xe_vma_vm(vma));
> +		xe_svm_notifier_lock(vm);
>  		if (!xe_svm_range_pages_valid(range)) {
>  			xe_svm_range_debug(range, "BIND PREPARE - RETRY");
> -			xe_svm_notifier_unlock(xe_vma_vm(vma));
> +			xe_svm_notifier_unlock(vm);
>  			return -EAGAIN;
>  		}
>  		if (xe_svm_range_has_dma_mapping(range)) {
>  			xe_res_first_dma(range->base.dma_addr, 0,
>  					 range->base.itree.last + 1 - range->base.itree.start,
>  					 &curs);
> -			is_devmem = xe_res_is_vram(&curs);
> -			if (is_devmem)
> -				xe_svm_range_debug(range, "BIND PREPARE - DMA VRAM");
> -			else
> -				xe_svm_range_debug(range, "BIND PREPARE - DMA");
> +			xe_svm_range_debug(range, "BIND PREPARE - MIXED");
>  		} else {
>  			xe_assert(xe, false);
>  		}
> @@ -674,54 +699,17 @@ xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
>  		 * Note, when unlocking the resource cursor dma addresses may become
>  		 * stale, but the bind will be aborted anyway at commit time.
>  		 */
> -		xe_svm_notifier_unlock(xe_vma_vm(vma));
> +		xe_svm_notifier_unlock(vm);
>  	}
>  
> -	xe_walk.needs_64K = (xe_vma_vm(vma)->flags & XE_VM_FLAG_64K) && is_devmem;
> -
> -	/**
> -	 * Default atomic expectations for different allocation scenarios are as follows:
> -	 *
> -	 * 1. Traditional API: When the VM is not in LR mode:
> -	 *    - Device atomics are expected to function with all allocations.
> -	 *
> -	 * 2. Compute/SVM API: When the VM is in LR mode:
> -	 *    - Device atomics are the default behavior when the bo is placed in a single region.
> -	 *    - In all other cases device atomics will be disabled with AE=0 until an application
> -	 *      request differently using a ioctl like madvise.
> -	 */
> +	xe_walk.needs_64K = (vm->flags & XE_VM_FLAG_64K);
>  	if (vma->gpuva.flags & XE_VMA_ATOMIC_PTE_BIT) {
> -		if (xe_vm_in_lr_mode(xe_vma_vm(vma))) {
> -			if (bo && xe_bo_has_single_placement(bo))
> -				xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
> -			/**
> -			 * If a SMEM+LMEM allocation is backed by SMEM, a device
> -			 * atomics will cause a gpu page fault and which then
> -			 * gets migrated to LMEM, bind such allocations with
> -			 * device atomics enabled.
> -			 */
> -			else if (is_devmem)
> -				xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
> -		} else {
> -			xe_walk.default_pte |= XE_USM_PPGTT_PTE_AE;
> -		}
> -
> -		/**
> -		 * Unset AE if the platform(PVC) doesn't support it on an
> -		 * allocation
> -		 */
> -		if (!xe->info.has_device_atomics_on_smem && !is_devmem)
> -			xe_walk.default_pte &= ~XE_USM_PPGTT_PTE_AE;
> +		xe_walk.default_vram_pte = xe_atomic_for_vram(vm) ? XE_USM_PPGTT_PTE_AE : 0;
> +		xe_walk.default_system_pte = xe_atomic_for_system(vm) ? XE_USM_PPGTT_PTE_AE : 0;
>  	}
>  
> -	if (is_devmem) {
> -		xe_walk.default_pte |= XE_PPGTT_PTE_DM;
> -		xe_walk.dma_offset = bo ? vram_region_gpu_offset(bo->ttm.resource) : 0;
> -	}
> -
> -	if (!xe_vma_has_no_bo(vma) && xe_bo_is_stolen(bo))
> -		xe_walk.dma_offset = xe_ttm_stolen_gpu_offset(xe_bo_device(bo));
> -
> +	xe_walk.default_vram_pte |= XE_PPGTT_PTE_DM;
> +	xe_walk.dma_offset = bo ? vram_region_gpu_offset(bo->ttm.resource) : 0;
>  	if (!range)
>  		xe_bo_assert_held(bo);
>  
> -- 
> 2.48.1
> 


More information about the Intel-xe mailing list