On Mon, Sep 15, 2025 at 12:21:08AM -0700, Vivek Kasireddy wrote:
> For BOs of type ttm_bo_type_sg, that are backed by PCI BAR addresses
> associated with a VF, we need to adjust and translate these addresses
> to LMEM addresses to make the BOs usable by the PF. Otherwise, the
> BOs (i.e, PCI BAR addresses) are only accessible by the CPU and not
> by the GPU.
>
> In order to do the above, we first need to identify if the addresses
> associated with an imported BO (type ttm_bo_type_sg) belong to System
> RAM or a VF or other PCI devices. After we confirm that they belong to
> a VF, we convert the BAR addresses to DPAs and create a new dma_addr
> array (of type drm_pagemap_dma_addr) and populate it with the new
> addresses along with the segment sizes.
>
> v2:
> - Use dma_addr array instead of sg table to store translated addresses
> (Matt)
>
> v3:
> - Remove the usage of iommu_iova_to_phys() as the imported BO would no
> longer contain IOVAs and would instead have BAR addresses.
>
> v4:
> - Take a reference on the VF's backing object in LMEM (Michal)
> - Create a new type for storing dma data
>
> Cc: Matthew Brost <[email protected]>
> Cc: Thomas Hellström <[email protected]>
> Signed-off-by: Vivek Kasireddy <[email protected]>
> ---
> drivers/gpu/drm/xe/xe_bo.c | 101 ++++++++++++++++++++++++-
> drivers/gpu/drm/xe/xe_bo_types.h | 12 +++
> drivers/gpu/drm/xe/xe_sriov_pf_types.h | 19 +++++
> 3 files changed, 131 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> index e6d16b34e5b5..a4e09fa88274 100644
> --- a/drivers/gpu/drm/xe/xe_bo.c
> +++ b/drivers/gpu/drm/xe/xe_bo.c
> @@ -21,11 +21,13 @@
>
> #include <trace/events/gpu_mem.h>
>
> +#include "regs/xe_bars.h"
> #include "xe_device.h"
> #include "xe_dma_buf.h"
> #include "xe_drm_client.h"
> #include "xe_ggtt.h"
> #include "xe_gt.h"
> +#include "xe_gt_sriov_pf_config.h"
> #include "xe_map.h"
> #include "xe_migrate.h"
> #include "xe_pm.h"
> @@ -33,6 +35,7 @@
> #include "xe_pxp.h"
> #include "xe_res_cursor.h"
> #include "xe_shrinker.h"
> +#include "xe_sriov_pf_helpers.h"
> #include "xe_sriov_vf_ccs.h"
> #include "xe_trace_bo.h"
> #include "xe_ttm_stolen_mgr.h"
> @@ -679,6 +682,90 @@ static int xe_bo_trigger_rebind(struct xe_device *xe,
> struct xe_bo *bo,
> return ret;
> }
>
> +static struct pci_dev *xe_find_vf_dev(struct xe_device *xe,
> + phys_addr_t phys)
> +{
> + struct pci_dev *pdev, *pf_pdev = to_pci_dev(xe->drm.dev);
> + resource_size_t io_start, io_size;
> +
> + list_for_each_entry(pdev, &pf_pdev->bus->devices, bus_list) {
> + if (pdev->is_physfn)
> + continue;
> +
> + io_start = pci_resource_start(pdev, LMEM_BAR);
> + io_size = pci_resource_len(pdev, LMEM_BAR);
> +
> + if (phys >= io_start &&
> + phys < (io_start + io_size - PAGE_SIZE))
> + return pdev;
> + }
> +
> + return NULL;
> +}
> +
> +
> +static void xe_bo_translate_io_to_dpa(struct xe_bo *bo, struct sg_table *sg,
> + resource_size_t io_start, int vfid)
> +{
> + struct xe_device *xe = xe_bo_device(bo);
> + struct xe_gt *gt = xe_root_mmio_gt(xe);
> + struct scatterlist *sgl;
> + struct xe_bo *lmem_obj;
> + phys_addr_t phys;
> + dma_addr_t addr;
> + u64 offset, i;
> +
> + lmem_obj = xe_gt_sriov_pf_config_get_lmem_obj(gt, ++vfid);
> + bo->dma_data.lmem_obj = xe_bo_get(lmem_obj);
Since we're taking references here, and if those leak it could be quite
problematic, I'd assert that bo->dma_data.lmem_obj is NULL before
assigning it.
Additionally, I've updated the flow so that
xe_gt_sriov_pf_config_get_lmem_obj takes the BO reference under the
xe_gt_sriov_pf_master_mutex lock. This is typically the correct approach
when looking up an object under a lock to ensure it doesn't immediately
disappear after dropping the lock.
Matt
> +
> + for_each_sgtable_dma_sg(sg, sgl, i) {
> + phys = sg_dma_address(sgl);
> + offset = phys - io_start;
> + addr = xe_bo_addr(lmem_obj, offset, sg_dma_len(sgl));
> +
> + bo->dma_data.dma_addr[i] = drm_pagemap_addr_encode(addr,
> + DRM_INTERCONNECT_DRIVER,
> + get_order(sg_dma_len(sgl)),
> + DMA_BIDIRECTIONAL);
> + }
> +}
> +
> +static int xe_bo_sg_to_dma_addr_array(struct sg_table *sg, struct xe_bo *bo)
> +{
> + struct xe_device *xe = xe_bo_device(bo);
> + struct drm_pagemap_addr *dma_addr;
> + resource_size_t io_start;
> + struct pci_dev *pdev;
> + phys_addr_t phys;
> + int vfid;
> +
> + if (!IS_SRIOV_PF(xe))
> + return 0;
> +
> + phys = sg_dma_address(sg->sgl);
> + if (page_is_ram(PFN_DOWN(phys)))
> + return 0;
> +
> + pdev = xe_find_vf_dev(xe, phys);
> + if (!pdev)
> + return 0;
> +
> + vfid = pci_iov_vf_id(pdev);
> + if (vfid < 0)
> + return 0;
> +
> + dma_addr = kmalloc_array(sg->nents, sizeof(*dma_addr), GFP_KERNEL);
> + if (!dma_addr)
> + return -ENOMEM;
> +
> + bo->is_devmem_external = true;
> + bo->dma_data.dma_addr = dma_addr;
> + io_start = pci_resource_start(pdev, LMEM_BAR);
> + xe_bo_translate_io_to_dpa(bo, sg, io_start, vfid);
> +
> + return 0;
> +}
> +
> /*
> * The dma-buf map_attachment() / unmap_attachment() is hooked up here.
> * Note that unmapping the attachment is deferred to the next
> @@ -697,6 +784,7 @@ static int xe_bo_move_dmabuf(struct ttm_buffer_object
> *ttm_bo,
> struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
> bool device_unplugged = drm_dev_is_unplugged(&xe->drm);
> struct sg_table *sg;
> + int ret = 0;
>
> xe_assert(xe, attach);
> xe_assert(xe, ttm_bo->ttm);
> @@ -721,13 +809,19 @@ static int xe_bo_move_dmabuf(struct ttm_buffer_object
> *ttm_bo,
> if (IS_ERR(sg))
> return PTR_ERR(sg);
>
> + ret = xe_bo_sg_to_dma_addr_array(sg, ttm_to_xe_bo(ttm_bo));
> + if (ret < 0) {
> + dma_buf_unmap_attachment(attach, sg, DMA_BIDIRECTIONAL);
> + return ret;
> + }
> +
> ttm_bo->sg = sg;
> xe_tt->sg = sg;
>
> out:
> ttm_bo_move_null(ttm_bo, new_res);
>
> - return 0;
> + return ret;
> }
>
> /**
> @@ -1554,6 +1648,11 @@ static void xe_ttm_bo_delete_mem_notify(struct
> ttm_buffer_object *ttm_bo)
>
> dma_buf_unmap_attachment(ttm_bo->base.import_attach, ttm_bo->sg,
> DMA_BIDIRECTIONAL);
> +
> + if (bo->is_devmem_external) {
> + xe_bo_put(bo->dma_data.lmem_obj);
> + kfree(bo->dma_data.dma_addr);
> + }
> ttm_bo->sg = NULL;
> xe_tt->sg = NULL;
> }
> diff --git a/drivers/gpu/drm/xe/xe_bo_types.h
> b/drivers/gpu/drm/xe/xe_bo_types.h
> index d4fe3c8dca5b..8e416c4ffbca 100644
> --- a/drivers/gpu/drm/xe/xe_bo_types.h
> +++ b/drivers/gpu/drm/xe/xe_bo_types.h
> @@ -108,6 +108,18 @@ struct xe_bo {
> * from default
> */
> u64 min_align;
> +
> + /**
> + * @is_devmem_external: Whether this BO is an imported dma-buf that
> + * is LMEM based.
> + */
> + bool is_devmem_external;
> +
> + /**
> + * @dma_data: DMA related data for an imported dmabuf BO that is LMEM
> + * based.
> + */
> + struct xe_sriov_dma_data dma_data;
> };
>
> #endif
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> index 956a88f9f213..6d5f923f7fc4 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> @@ -11,6 +11,8 @@
>
> #include "xe_sriov_pf_service_types.h"
>
> +struct xe_bo;
> +
> /**
> * struct xe_sriov_metadata - per-VF device level metadata
> */
> @@ -42,4 +44,21 @@ struct xe_device_pf {
> struct xe_sriov_metadata *vfs;
> };
>
> +/**
> + * struct xe_sriov_dma_data - DMA related data for LMEM based imported dmabuf
> + * BOs that are associated with a sriov VF.
> + *
> + * The data in this structure is valid only if driver is running in the
> + * @XE_SRIOV_MODE_PF mode.
> + */
> +struct xe_sriov_dma_data {
> + /**
> + * @dma_addr: An array to store DMA addresses (DPAs) for imported
> + * dmabuf BOs that are LMEM based.
> + */
> + struct drm_pagemap_addr *dma_addr;
> +
> + /** @lmem_obj: Ref taken on the LMEM obj that backs a VF's quota */
> + struct xe_bo *lmem_obj;
> +};
> #endif
> --
> 2.50.1
>
