[PATCH 5/5] drm/amdkfd: classify and map mixed svm range pages in GPU

[Felix Kuehling]

Am 2021-05-06 um 11:36 p.m. schrieb Alex Sierra:
> [Why]
> svm ranges can have mixed pages from device or system memory.
> A good example is, after a prange has been allocated in VRAM and a
> copy-on-write is triggered by a fork. This invalidates some pages
> inside the prange. Endding up in mixed pages.
>
> [How]
> By classifying each page inside a prange, based on its type. Device or
> system memory, during dma mapping call. If page corresponds
> to VRAM domain, a flag is set to its dma_addr entry for each GPU.
> Then, at the GPU page table mapping. All group of contiguous pages within
> the same type are mapped with their proper pte flags.
>
> Signed-off-by: Alex Sierra <alex.sierra at amd.com>
> ---
>  drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 51 ++++++++++++++++++----------
>  drivers/gpu/drm/amd/amdkfd/kfd_svm.h |  1 +
>  2 files changed, 35 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 940165f4437d..1440e4e555f1 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -141,6 +141,12 @@ svm_range_dma_map_dev(struct device *dev, dma_addr_t **dma_addr,
>  			dma_unmap_page(dev, addr[i], PAGE_SIZE, dir);
>  
>  		page = hmm_pfn_to_page(hmm_pfns[i]);
> +		if (is_zone_device_page(page)) {
> +			addr[i] = hmm_pfns[i] << PAGE_SHIFT;
> +			addr[i] |= SVM_RANGE_VRAM_DOMAIN;
> +			pr_debug("vram address detected: 0x%llx\n", addr[i] >> PAGE_SHIFT);
> +			continue;
> +		}
>  		addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
>  		r = dma_mapping_error(dev, addr[i]);
>  		if (r) {
> @@ -1131,32 +1137,43 @@ svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>  		     struct amdgpu_device *bo_adev, struct dma_fence **fence)
>  {
>  	struct amdgpu_bo_va bo_va;
> +	struct ttm_resource *ttm_res;
>  	uint64_t pte_flags;
> +	unsigned long last_start;
> +	int last_domain;
>  	int r = 0;
> +	int64_t i;
>  
>  	pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms, prange->start,
>  		 prange->last);
>  
> -	if (prange->svm_bo && prange->ttm_res) {
> +	if (prange->svm_bo && prange->ttm_res)
>  		bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
> -		prange->mapping.bo_va = &bo_va;
> -	}
>  
> -	prange->mapping.start = prange->start;
> -	prange->mapping.last = prange->last;
> -	prange->mapping.offset = prange->offset;
> -	pte_flags = svm_range_get_pte_flags(adev, prange);
> +	last_start = prange->start;
> +	for (i = 0; i < prange->npages; i++) {
> +		last_domain = dma_addr[i] & SVM_RANGE_VRAM_DOMAIN;
> +		if ((prange->start + i) < prange->last &&
> +		    last_domain == (dma_addr[i + 1] & SVM_RANGE_VRAM_DOMAIN))
> +			continue;
>  
> -	r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false, NULL,
> -					prange->mapping.start,
> -					prange->mapping.last, pte_flags,
> -					prange->mapping.offset,
> -					prange->ttm_res ?
> -						prange->ttm_res->mm_node : NULL,
> -					dma_addr, &vm->last_update);
> -	if (r) {
> -		pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
> -		goto out;
> +		pr_debug("Mapping range [0x%lx 0x%llx] on domain: %s\n",
> +			 last_start, prange->start + i, last_domain ? "GPU" : "CPU");
> +		ttm_res = last_domain ? prange->ttm_res : NULL;
> +		pte_flags = svm_range_get_pte_flags(adev, prange);
> +		r = amdgpu_vm_bo_update_mapping(adev, bo_adev, vm, false, false, NULL,
> +						last_start,
> +						prange->start + i, pte_flags,
> +						prange->offset +
> +						((last_start - prange->start) << PAGE_SHIFT),
> +						ttm_res ? ttm_res->mm_node : NULL,
> +						ttm_res ? NULL : dma_addr,
> +						&vm->last_update);
> +		if (r) {
> +			pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
> +			goto out;
> +		}
> +		last_start += prange->start + i + 1;

This looks wrong. It's either

    last_start += i + 1;

or

    last_start = prange->start + i + 1;

With that fixed, the series is
Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>


>  	}
>  
>  	r = amdgpu_vm_update_pdes(adev, vm, false);
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> index 08542fe39303..e68aa51322df 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> @@ -35,6 +35,7 @@
>  #include "amdgpu.h"
>  #include "kfd_priv.h"
>  
> +#define SVM_RANGE_VRAM_DOMAIN (1UL << 0)
>  #define SVM_ADEV_PGMAP_OWNER(adev)\
>  			((adev)->hive ? (void *)(adev)->hive : (void *)(adev))
>