Re: [PATCH 4/4] drm/amdgpu: enable huge page handling in the VM v3

[zhoucm1 <david1.zhou-5C7GfCeVMHo@public.gmane.org>]

On 2017年05月31日 20:38, Christian König wrote:
> From: Christian König <christian.koenig@amd.com>
>
> The hardware can use huge pages to map 2MB of address space with only one PDE.
>
> v2: few cleanups and rebased
> v3: skip PT updates if we are using the PDE
>
> Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 103 +++++++++++++++++++++++++++------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |   4 ++
>   2 files changed, 88 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index d07a28c..4cfef3c 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -323,6 +323,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
>   
>   			entry->bo = pt;
>   			entry->addr = 0;
> +			entry->huge_page = false;
>   		}
>   
>   		if (level < adev->vm_manager.num_level) {
> @@ -990,7 +991,8 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
>   
>   		pt = amdgpu_bo_gpu_offset(bo);
>   		pt = amdgpu_gart_get_vm_pde(adev, pt);
> -		if (parent->entries[pt_idx].addr == pt)
> +		if (parent->entries[pt_idx].addr == pt ||
> +		    parent->entries[pt_idx].huge_page)
>   			continue;
>   
>   		parent->entries[pt_idx].addr = pt;
> @@ -1122,29 +1124,83 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
>   }
>   
>   /**
> - * amdgpu_vm_find_pt - find the page table for an address
> + * amdgpu_vm_find_entry - find the entry for an address
>    *
>    * @p: see amdgpu_pte_update_params definition
>    * @addr: virtual address in question
> + * @entry: resulting entry or NULL
> + * @parent: parent entry
>    *
> - * Find the page table BO for a virtual address, return NULL when none found.
> + * Find the vm_pt entry and it's parent for the given address.
>    */
> -static struct amdgpu_bo *amdgpu_vm_get_pt(struct amdgpu_pte_update_params *p,
> -					  uint64_t addr)
> +void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr,
> +			 struct amdgpu_vm_pt **entry,
> +			 struct amdgpu_vm_pt **parent)
>   {
> -	struct amdgpu_vm_pt *entry = &p->vm->root;
>   	unsigned idx, level = p->adev->vm_manager.num_level;
>   
> -	while (entry->entries) {
> +	*parent = NULL;
> +	*entry = &p->vm->root;
> +	while ((*entry)->entries) {
>   		idx = addr >> (p->adev->vm_manager.block_size * level--);
> -		idx %= amdgpu_bo_size(entry->bo) / 8;
> -		entry = &entry->entries[idx];
> +		idx %= amdgpu_bo_size((*entry)->bo) / 8;
> +		*parent = *entry;
> +		*entry = &(*entry)->entries[idx];
>   	}
>   
>   	if (level)
> -		return NULL;
> +		*entry = NULL;
> +}
> +
> +/**
> + * amdgpu_vm_handle_huge_pages - handle updating the PD with huge pages
> + *
> + * @p: see amdgpu_pte_update_params definition
> + * @entry: vm_pt entry to check
> + * @parent: parent entry
> + * @nptes: number of PTEs updated with this operation
> + * @dst: destination address where the PTEs should point to
> + * @flags: access flags fro the PTEs
> + *
> + * Check if we can update the PD with a huge page.
> + */
> +static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
> +					struct amdgpu_vm_pt *entry,
> +					struct amdgpu_vm_pt *parent,
> +					unsigned nptes, uint64_t dst,
> +					uint64_t flags)
> +{
> +	uint64_t pd_addr, pde;
> +
> +	/* In the case of a mixed PT the PDE must point to it*/
> +	if (p->adev->asic_type < CHIP_VEGA10 ||
> +	    nptes != AMDGPU_VM_PTE_COUNT(p->adev) ||
> +	    p->func != amdgpu_vm_do_set_ptes ||
> +	    !(flags & AMDGPU_PTE_VALID)) {
> +
> +		dst = amdgpu_bo_gpu_offset(entry->bo);
> +		dst = amdgpu_gart_get_vm_pde(p->adev, dst);
> +		flags = AMDGPU_PTE_VALID;
> +	} else {
> +		flags |= AMDGPU_PDE_PTE;
> +	}
>   
> -	return entry->bo;
> +	if (entry->addr == dst &&
> +	    entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
> +		return;
> +
> +	entry->addr = dst;
> +	entry->huge_page = !!(flags & AMDGPU_PDE_PTE);
> +
> +	if (parent->bo->shadow) {
> +		pd_addr = amdgpu_bo_gpu_offset(parent->bo->shadow);
> +		pde = pd_addr + (entry - parent->entries) * 8;
> +		amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
> +	}
> +
> +	pd_addr = amdgpu_bo_gpu_offset(parent->bo);
> +	pde = pd_addr + (entry - parent->entries) * 8;
> +	amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
>   }
>   
>   /**
> @@ -1172,18 +1228,25 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
>   	unsigned nptes;
>   
>   	/* walk over the address space and update the page tables */
> -	for (addr = start; addr < end; addr += nptes) {
> -		pt = amdgpu_vm_get_pt(params, addr);
> -		if (!pt) {
> -			pr_err("PT not found, aborting update_ptes\n");
> -			return -EINVAL;
> -		}
> +	for (addr = start; addr < end; addr += nptes,
> +	     dst += nptes * AMDGPU_GPU_PAGE_SIZE) {
> +		struct amdgpu_vm_pt *entry, *parent;
> +
> +		amdgpu_vm_get_entry(params, addr, &entry, &parent);
> +		if (!entry)
> +			return -ENOENT;
>   
>   		if ((addr & ~mask) == (end & ~mask))
>   			nptes = end - addr;
>   		else
>   			nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
>   
> +		amdgpu_vm_handle_huge_pages(params, entry, parent,
> +					    nptes, dst, flags);
> +		if (entry->huge_page)
> +			continue;
> +
> +		pt = entry->bo;
>   		if (pt->shadow) {
>   			pe_start = amdgpu_bo_gpu_offset(pt->shadow);
>   			pe_start += (addr & mask) * 8;
> @@ -1195,8 +1258,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
>   		pe_start += (addr & mask) * 8;
>   		params->func(params, pe_start, dst, nptes,
>   			     AMDGPU_GPU_PAGE_SIZE, flags);
> -
> -		dst += nptes * AMDGPU_GPU_PAGE_SIZE;
>   	}
>   
>   	return 0;
> @@ -1330,6 +1391,9 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
>   	/* padding, etc. */
>   	ndw = 64;
>   
> +	/* one PDE write for each huge page */
> +	ndw += ((nptes >> adev->vm_manager.block_size) + 1) * 7;
> +
>   	if (src) {
>   		/* only copy commands needed */
>   		ndw += ncmds * 7;
> @@ -1409,6 +1473,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
>   
>   error_free:
>   	amdgpu_job_free(job);
> +	amdgpu_vm_invalidate_level(&vm->root);
>   	return r;
>   }
>   
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index aabe815..79145e6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -70,6 +70,9 @@ struct amdgpu_bo_list_entry;
>   /* TILED for VEGA10, reserved for older ASICs  */
>   #define AMDGPU_PTE_PRT		(1ULL << 51)
>   
> +/* PDE is handled as PTE for VEGA10 */
> +#define AMDGPU_PDE_PTE		(1ULL << 54)
> +
>   /* VEGA10 only */
>   #define AMDGPU_PTE_MTYPE(a)    ((uint64_t)a << 57)
>   #define AMDGPU_PTE_MTYPE_MASK	AMDGPU_PTE_MTYPE(3ULL)
> @@ -92,6 +95,7 @@ struct amdgpu_bo_list_entry;
>   struct amdgpu_vm_pt {
>   	struct amdgpu_bo	*bo;
>   	uint64_t		addr;
> +	bool			huge_page;
>   
>   	/* array of page tables, one for each directory entry */
>   	struct amdgpu_vm_pt	*entries;

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx