[PATCH 1/3] drm/amdkfd: classify and map mixed svm range pages in GPU
Felix Kuehling
felix.kuehling at amd.com
Mon May 17 13:44:57 UTC 2021
A few more comments inline. Sorry I took so long.
Am 2021-05-12 um 1:34 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.
>
> v2:
> Instead of using ttm_res to calculate vram pfns in the svm_range. It is now
> done by setting the vram real physical address into drm_addr array.
> This makes more flexible VRAM management, plus removes the need to have
> a BO reference in the svm_range.
>
> Signed-off-by: Alex Sierra <alex.sierra at amd.com>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 71 ++++++++++++++++++----------
> drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 1 +
> 2 files changed, 46 insertions(+), 26 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 2b4318646a75..0ab10cb24205 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -119,11 +119,12 @@ static void svm_range_remove_notifier(struct svm_range *prange)
> }
>
> static int
> -svm_range_dma_map_dev(struct device *dev, dma_addr_t **dma_addr,
> +svm_range_dma_map_dev(struct amdgpu_device *adev, dma_addr_t **dma_addr,
> unsigned long *hmm_pfns, uint64_t npages)
> {
> enum dma_data_direction dir = DMA_BIDIRECTIONAL;
> dma_addr_t *addr = *dma_addr;
> + struct device *dev = adev->dev;
> struct page *page;
> int i, r;
>
> @@ -141,6 +142,14 @@ 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) +
> + adev->vm_manager.vram_base_offset -
> + adev->kfd.dev->pgmap.range.start;
> + addr[i] |= SVM_RANGE_VRAM_DOMAIN;
> + pr_debug("vram address detected: 0x%llx\n", addr[i]);
> + continue;
> + }
> addr[i] = dma_map_page(dev, page, 0, PAGE_SIZE, dir);
> r = dma_mapping_error(dev, addr[i]);
> if (r) {
> @@ -175,7 +184,7 @@ svm_range_dma_map(struct svm_range *prange, unsigned long *bitmap,
> }
> adev = (struct amdgpu_device *)pdd->dev->kgd;
>
> - r = svm_range_dma_map_dev(adev->dev, &prange->dma_addr[gpuidx],
> + r = svm_range_dma_map_dev(adev, &prange->dma_addr[gpuidx],
> hmm_pfns, prange->npages);
> if (r)
> break;
> @@ -1003,21 +1012,22 @@ svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
> }
>
> static uint64_t
> -svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange)
> +svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange,
> + int domain)
> {
> struct amdgpu_device *bo_adev;
> uint32_t flags = prange->flags;
> uint32_t mapping_flags = 0;
> uint64_t pte_flags;
> - bool snoop = !prange->ttm_res;
> + bool snoop = !!domain;
This is a bit obscure and it assumes that SVM_RANGE_VRAM_DOMAIN is 0.
And it's not actually correct. Make this
bool snoop = (domain != SVM_RANGE_VRAM_DOMAIN);
> bool coherent = flags & KFD_IOCTL_SVM_FLAG_COHERENT;
>
> - if (prange->svm_bo && prange->ttm_res)
> + if (domain == SVM_RANGE_VRAM_DOMAIN)
> bo_adev = amdgpu_ttm_adev(prange->svm_bo->bo->tbo.bdev);
>
> switch (adev->asic_type) {
> case CHIP_ARCTURUS:
> - if (prange->svm_bo && prange->ttm_res) {
> + if (domain == SVM_RANGE_VRAM_DOMAIN) {
> if (bo_adev == adev) {
> mapping_flags |= coherent ?
> AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
> @@ -1032,7 +1042,7 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange)
> }
> break;
> case CHIP_ALDEBARAN:
> - if (prange->svm_bo && prange->ttm_res) {
> + if (domain == SVM_RANGE_VRAM_DOMAIN) {
> if (bo_adev == adev) {
> mapping_flags |= coherent ?
> AMDGPU_VM_MTYPE_CC : AMDGPU_VM_MTYPE_RW;
> @@ -1061,14 +1071,14 @@ svm_range_get_pte_flags(struct amdgpu_device *adev, struct svm_range *prange)
> mapping_flags |= AMDGPU_VM_PAGE_EXECUTABLE;
>
> pte_flags = AMDGPU_PTE_VALID;
> - pte_flags |= prange->ttm_res ? 0 : AMDGPU_PTE_SYSTEM;
> + pte_flags |= (domain == SVM_RANGE_VRAM_DOMAIN) ? 0 : AMDGPU_PTE_SYSTEM;
> pte_flags |= snoop ? AMDGPU_PTE_SNOOPED : 0;
>
> pte_flags |= amdgpu_gem_va_map_flags(adev, mapping_flags);
>
> pr_debug("svms 0x%p [0x%lx 0x%lx] vram %d PTE 0x%llx mapping 0x%x\n",
> prange->svms, prange->start, prange->last,
> - prange->ttm_res ? 1:0, pte_flags, mapping_flags);
> + (domain == SVM_RANGE_VRAM_DOMAIN) ? 1:0, pte_flags, mapping_flags);
>
> return pte_flags;
> }
> @@ -1138,31 +1148,40 @@ svm_range_map_to_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,
> {
> struct amdgpu_bo_va bo_va;
> 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;
It looks like you could remove mapping from the svm_range structure now.
> - 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");
> + pte_flags = svm_range_get_pte_flags(adev, prange, last_domain);
> + 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),
> + NULL,
I think this still needs to be ttm_res->mm_node if last_domain == VRAM.
> + dma_addr,
And this needs to be NULL if last_domain == VRAM.
Regards,
Felix
> + &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;
> }
>
> 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))
>
More information about the amd-gfx
mailing list