[PATCH 4/6] drm/amdgpu: enable pdb0 for hibernation on SRIOV
Christian König
ckoenig.leichtzumerken at gmail.com
Wed Apr 16 13:52:23 UTC 2025
Am 14.04.25 um 12:46 schrieb Samuel Zhang:
> When switching to new GPU index after hibernation and then resume,
> VRAM offset of each VRAM BO will be changed, and the cached gpu
> addresses needed to updated.
>
> This is to enable pdb0 and switch to use pdb0-based virtual gpu
> address by default in amdgpu_bo_create_reserved(). since the virtual
> addresses do not change, this can avoid the need to update all
> cached gpu addresses all over the codebase.
>
> Signed-off-by: Emily Deng <Emily.Deng at amd.com>
> Signed-off-by: Samuel Zhang <guoqing.zhang at amd.com>
> Change-Id: I2b20b9b94f1e41820a013ce5d05bb3fa96859b21
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c | 43 +++++++++++++++-------
> drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h | 1 +
> drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 8 +++-
> drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c | 2 +-
> drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 29 +++++++++------
> drivers/gpu/drm/amd/amdgpu/mmhub_v9_4.c | 30 ++++++++++++---
> 6 files changed, 82 insertions(+), 31 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c
> index 5b60d714e089..e706afcb7e95 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c
> @@ -248,18 +248,25 @@ void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
> void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
> {
> u64 hive_vram_start = 0;
> - u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
> - mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
> - mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
> + u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes;
> +
> + hive_vram_end = ALIGN(hive_vram_end, (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21) - 1;
> +
> + if (!mc->vram_start) {
> + mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
> + mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
> + dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
> + mc->mc_vram_size >> 20, mc->vram_start,
> + mc->vram_end, mc->real_vram_size >> 20);
> + }
> +
> mc->gart_start = hive_vram_end + 1;
> mc->gart_end = mc->gart_start + mc->gart_size - 1;
> mc->fb_start = hive_vram_start;
> mc->fb_end = hive_vram_end;
> - dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
> - mc->mc_vram_size >> 20, mc->vram_start,
> - mc->vram_end, mc->real_vram_size >> 20);
> - dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
> - mc->gart_size >> 20, mc->gart_start, mc->gart_end);
> +
> + dev_info(adev->dev, "FB 0x%016llX - 0x%016llX, GART: %lluM 0x%016llX - 0x%016llX\n",
> + mc->fb_start, mc->fb_end, mc->gart_size >> 20, mc->gart_start, mc->gart_end);
> }
>
> /**
> @@ -677,8 +684,9 @@ void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
> &job);
> if (r)
> goto error_alloc;
> -
> - job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
> + job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
> + adev->gmc.pdb0_bo :
> + adev->gart.bo);
> job->vm_needs_flush = true;
> job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
> amdgpu_ring_pad_ib(ring, &job->ibs[0]);
> @@ -1041,8 +1049,9 @@ void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
> */
> u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
> u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
> - u64 vram_addr = adev->vm_manager.vram_base_offset -
> + u64 vram_addr_first = adev->vm_manager.vram_base_offset -
> adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
> + u64 vram_addr = adev->vm_manager.vram_base_offset;
> u64 vram_end = vram_addr + vram_size;
> u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
> int idx;
> @@ -1056,11 +1065,19 @@ void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
> flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
> flags |= AMDGPU_PDE_PTE_FLAG(adev);
>
> + if (adev->gmc.xgmi.connected_to_cpu) {
> + vram_addr = vram_addr_first;
> + vram_end = vram_addr + vram_size;
> + }
> +
> /* The first n PDE0 entries are used as PTE,
> * pointing to vram
> */
> - for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
> - amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
> + for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size) {
> + amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i,
> + (vram_addr >= vram_addr_first + vram_size) ? (vram_addr - vram_size) : vram_addr,
> + flags);
> + }
>
> /* The n+1'th PDE0 entry points to a huge
> * PTB who has more than 512 entries each
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
> index bd7fc123b8f9..758b47240c6f 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
> @@ -307,6 +307,7 @@ struct amdgpu_gmc {
> struct amdgpu_bo *pdb0_bo;
> /* CPU kmapped address of pdb0*/
> void *ptr_pdb0;
> + bool enable_pdb0;
>
> /* MALL size */
> u64 mall_size;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index d90e9daf5a50..83a3444c69d9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -287,8 +287,14 @@ int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
> goto error_unpin;
> }
>
> - if (gpu_addr)
> + if (gpu_addr) {
> *gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
> + if (!adev->gmc.xgmi.connected_to_cpu && adev->gmc.enable_pdb0) {
> + if ((*bo_ptr)->tbo.resource->mem_type == TTM_PL_VRAM) {
> + *gpu_addr -= amdgpu_ttm_domain_start(adev, TTM_PL_VRAM);
> + }
> + }
> + }
Please NAK to that approach here. The GPU offset should still point into the mapped VRAM.
>
> if (cpu_addr) {
> r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c b/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c
> index cb25f7f0dfc1..5ebb92ac9fd7 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfxhub_v1_2.c
> @@ -180,7 +180,7 @@ gfxhub_v1_2_xcc_init_system_aperture_regs(struct amdgpu_device *adev,
> /* In the case squeezing vram into GART aperture, we don't use
> * FB aperture and AGP aperture. Disable them.
> */
> - if (adev->gmc.pdb0_bo) {
> + if (adev->gmc.pdb0_bo && !amdgpu_sriov_vf(adev)) {
> WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_FB_LOCATION_TOP, 0);
> WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_FB_LOCATION_BASE, 0x00FFFFFF);
> WREG32_SOC15(GC, GET_INST(GC, i), regMC_VM_AGP_TOP, 0);
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> index 7c7a9fe6be6d..73ac05b9a1bf 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
> @@ -1677,6 +1677,10 @@ static int gmc_v9_0_early_init(struct amdgpu_ip_block *ip_block)
> adev->gmc.private_aperture_start + (4ULL << 30) - 1;
> adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF;
>
> + if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) ||
> + amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 4) ||
> + amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 5, 0))
> + adev->gmc.enable_pdb0 = amdgpu_sriov_vf(adev);
> return 0;
> }
>
> @@ -1719,6 +1723,14 @@ static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
> {
> u64 base = adev->mmhub.funcs->get_fb_location(adev);
>
> + if (adev->gmc.xgmi.connected_to_cpu || adev->gmc.enable_pdb0) {
> + adev->gmc.vmid0_page_table_depth = 1;
> + adev->gmc.vmid0_page_table_block_size = 12;
> + } else {
> + adev->gmc.vmid0_page_table_depth = 0;
> + adev->gmc.vmid0_page_table_block_size = 0;
> + }
> +
What is the justification to moving that stuff around?
> amdgpu_gmc_set_agp_default(adev, mc);
>
> /* add the xgmi offset of the physical node */
> @@ -1727,7 +1739,10 @@ static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
> amdgpu_gmc_sysvm_location(adev, mc);
> } else {
> amdgpu_gmc_vram_location(adev, mc, base);
> - amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT);
> + if (!adev->gmc.enable_pdb0)
> + amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT);
> + else
> + amdgpu_gmc_sysvm_location(adev, mc);
> if (!amdgpu_sriov_vf(adev) && (amdgpu_agp == 1))
> amdgpu_gmc_agp_location(adev, mc);
> }
> @@ -1838,14 +1853,6 @@ static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
> return 0;
> }
>
> - if (adev->gmc.xgmi.connected_to_cpu) {
> - adev->gmc.vmid0_page_table_depth = 1;
> - adev->gmc.vmid0_page_table_block_size = 12;
> - } else {
> - adev->gmc.vmid0_page_table_depth = 0;
> - adev->gmc.vmid0_page_table_block_size = 0;
> - }
> -
> /* Initialize common gart structure */
> r = amdgpu_gart_init(adev);
> if (r)
> @@ -1864,7 +1871,7 @@ static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
> if (r)
> return r;
>
> - if (adev->gmc.xgmi.connected_to_cpu)
> + if (adev->gmc.xgmi.connected_to_cpu || adev->gmc.enable_pdb0)
Drop the connected_to_cpu check here.
> r = amdgpu_gmc_pdb0_alloc(adev);
> }
>
> @@ -2361,7 +2368,7 @@ static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
> {
> int r;
>
> - if (adev->gmc.xgmi.connected_to_cpu)
> + if (adev->gmc.xgmi.connected_to_cpu || adev->gmc.enable_pdb0)
And here.
> amdgpu_gmc_init_pdb0(adev);
>
> if (adev->gart.bo == NULL) {
> diff --git a/drivers/gpu/drm/amd/amdgpu/mmhub_v9_4.c b/drivers/gpu/drm/amd/amdgpu/mmhub_v9_4.c
> index fe0710b55c3a..13b229d07ac4 100644
> --- a/drivers/gpu/drm/amd/amdgpu/mmhub_v9_4.c
> +++ b/drivers/gpu/drm/amd/amdgpu/mmhub_v9_4.c
> @@ -74,27 +74,47 @@ static void mmhub_v9_4_setup_hubid_vm_pt_regs(struct amdgpu_device *adev, int hu
> static void mmhub_v9_4_init_gart_aperture_regs(struct amdgpu_device *adev,
> int hubid)
> {
> - uint64_t pt_base = amdgpu_gmc_pd_addr(adev->gart.bo);
> + uint64_t pt_base = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ? adev->gmc.pdb0_bo : adev->gart.bo);
That can be written as adev->gmc.pdb0_bo ?: adev->gart.bo
>
> mmhub_v9_4_setup_hubid_vm_pt_regs(adev, hubid, 0, pt_base);
>
> - WREG32_SOC15_OFFSET(MMHUB, 0,
> + if (adev->gmc.pdb0_bo) {
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> + mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32,
> + hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> + (u32)(adev->gmc.fb_start >> 12));
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> + mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32,
> + hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> + (u32)(adev->gmc.fb_start >> 44));
> +
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> + mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32,
> + hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> + (u32)(adev->gmc.gart_end >> 12));
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> + mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32,
> + hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> + (u32)(adev->gmc.gart_end >> 44));
> + } else {
> ++ WREG32_SOC15_OFFSET(MMHUB, 0,
> mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32,
> hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> (u32)(adev->gmc.gart_start >> 12));
> - WREG32_SOC15_OFFSET(MMHUB, 0,
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32,
> hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> (u32)(adev->gmc.gart_start >> 44));
When you indent the WREG32_SOC15_OFFSET() you need to indent the following lines as well.
>
> - WREG32_SOC15_OFFSET(MMHUB, 0,
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32,
> hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> (u32)(adev->gmc.gart_end >> 12));
> - WREG32_SOC15_OFFSET(MMHUB, 0,
> + WREG32_SOC15_OFFSET(MMHUB, 0,
> mmVML2VC0_VM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32,
> hubid * MMHUB_INSTANCE_REGISTER_OFFSET,
> (u32)(adev->gmc.gart_end >> 44));
> + }
The programming of the end addr is still the same, you don't need to change anything here.
Regards,
Christian.
> }
>
> static void mmhub_v9_4_setup_vm_pt_regs(struct amdgpu_device *adev, uint32_t vmid,
More information about the amd-gfx
mailing list