On 10/22/2025 12:41 AM, Michał Winiarski wrote:
> In an upcoming change, the VF GGTT migration data will be handled as
> part of VF control state machine. Add the necessary helpers to allow the
> migration data transfer to/from the HW GGTT resource.
>
> Signed-off-by: Michał Winiarski <[email protected]>
> ---
> drivers/gpu/drm/xe/xe_ggtt.c | 100 +++++++++++++++++++++
> drivers/gpu/drm/xe/xe_ggtt.h | 3 +
> drivers/gpu/drm/xe/xe_ggtt_types.h | 2 +
> drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c | 44 +++++++++
> drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h | 5 ++
> 5 files changed, 154 insertions(+)
>
> diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
> index 40680f0c49a17..99fe891c7939e 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt.c
> +++ b/drivers/gpu/drm/xe/xe_ggtt.c
> @@ -151,6 +151,14 @@ static void xe_ggtt_set_pte_and_flush(struct xe_ggtt
> *ggtt, u64 addr, u64 pte)
> ggtt_update_access_counter(ggtt);
> }
>
> +static u64 xe_ggtt_get_pte(struct xe_ggtt *ggtt, u64 addr)
> +{
> + xe_tile_assert(ggtt->tile, !(addr & XE_PTE_MASK));
> + xe_tile_assert(ggtt->tile, addr < ggtt->size);
> +
> + return readq(&ggtt->gsm[addr >> XE_PTE_SHIFT]);
> +}
> +
> static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size)
> {
> u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB];
> @@ -233,16 +241,19 @@ void xe_ggtt_might_lock(struct xe_ggtt *ggtt)
> static const struct xe_ggtt_pt_ops xelp_pt_ops = {
> .pte_encode_flags = xelp_ggtt_pte_flags,
> .ggtt_set_pte = xe_ggtt_set_pte,
> + .ggtt_get_pte = xe_ggtt_get_pte,
> };
>
> static const struct xe_ggtt_pt_ops xelpg_pt_ops = {
> .pte_encode_flags = xelpg_ggtt_pte_flags,
> .ggtt_set_pte = xe_ggtt_set_pte,
> + .ggtt_get_pte = xe_ggtt_get_pte,
> };
>
> static const struct xe_ggtt_pt_ops xelpg_pt_wa_ops = {
> .pte_encode_flags = xelpg_ggtt_pte_flags,
> .ggtt_set_pte = xe_ggtt_set_pte_and_flush,
> + .ggtt_get_pte = xe_ggtt_get_pte,
> };
>
> static void __xe_ggtt_init_early(struct xe_ggtt *ggtt, u32 reserved)
> @@ -912,6 +923,22 @@ static void xe_ggtt_assign_locked(struct xe_ggtt *ggtt,
> const struct drm_mm_node
> xe_ggtt_invalidate(ggtt);
> }
>
> +/**
> + * xe_ggtt_pte_size() - Convert GGTT VMA size to page table entries size.
> + * @ggtt: the &xe_ggtt
> + * @size: GGTT VMA size in bytes
> + *
> + * Return: GGTT page table entries size in bytes.
> + */
> +size_t xe_ggtt_pte_size(struct xe_ggtt *ggtt, size_t size)
passing ggtt just for assert seems overkill
> +{
> + struct xe_device __maybe_unused *xe = tile_to_xe(ggtt->tile);
we try to avoid __maybe_unused
if you need xe/tile/gt just in the assert, then put to_xe/tile/gt inside it
> +
> + xe_assert(xe, size % XE_PAGE_SIZE == 0);
> +
> + return size / XE_PAGE_SIZE * sizeof(u64);
> +}
> +
> /**
> * xe_ggtt_assign - assign a GGTT region to the VF
> * @node: the &xe_ggtt_node to update
> @@ -927,6 +954,79 @@ void xe_ggtt_assign(const struct xe_ggtt_node *node, u16
> vfid)
> xe_ggtt_assign_locked(node->ggtt, &node->base, vfid);
> mutex_unlock(&node->ggtt->lock);
> }
> +
> +/**
> + * xe_ggtt_node_save() - Save a &xe_ggtt_node to a buffer.
> + * @node: the &xe_ggtt_node to be saved
> + * @dst: destination buffer
> + * @size: destination buffer size in bytes
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size)
> +{
> + struct xe_ggtt *ggtt;
> + u64 start, end;
> + u64 *buf = dst;
> +
> + if (!node)
> + return -ENOENT;
> +
> + guard(mutex)(&node->ggtt->lock);
> +
> + ggtt = node->ggtt;
> + start = node->base.start;
> + end = start + node->base.size - 1;
> +
> + if (xe_ggtt_pte_size(ggtt, node->base.size) > size)
> + return -EINVAL;
> +
> + while (start < end) {
> + *buf++ = ggtt->pt_ops->ggtt_get_pte(ggtt, start) &
> ~GGTT_PTE_VFID;
up to this point function is generic, non-IOV, so maybe leave PTEs as-is and do
not sanitize VFID ?
or, similar to node_load(), also pass vfid to enforce additional checks ?
pte = ggtt->pt_ops->ggtt_get_pte(ggtt, start);
if (vfid != u64_get_bits(pte, GGTT_PTE_VFID))
return -EPERM;
then optionally sanitize using:
*buf++ = u64_replace_bits(pte, 0, GGTT_PTE_VFID);
> + start += XE_PAGE_SIZE;
> + }
> +
> + return 0;
> +}
> +
> +/**
> + * xe_ggtt_node_load() - Load a &xe_ggtt_node from a buffer.
> + * @node: the &xe_ggtt_node to be loaded
> + * @src: source buffer
> + * @size: source buffer size in bytes
> + * @vfid: VF identifier
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t
> size, u16 vfid)
> +{
> + u64 vfid_pte = xe_encode_vfid_pte(vfid);
> + const u64 *buf = src;
> + struct xe_ggtt *ggtt;
> + u64 start, end;
> +
> + if (!node)
> + return -ENOENT;
> +
> + guard(mutex)(&node->ggtt->lock);
> +
> + ggtt = node->ggtt;
> + start = node->base.start;
> + end = start + size - 1;
> +
> + if (xe_ggtt_pte_size(ggtt, node->base.size) != size)
> + return -EINVAL;
> +
> + while (start < end) {
> + ggtt->pt_ops->ggtt_set_pte(ggtt, start, (*buf & ~GGTT_PTE_VFID)
> | vfid_pte);
pte = u64_replace_bits(*buf++, vfid, GGTT_PTE_VFID));
ggtt_set_pte(ggtt, start, pte);
> + start += XE_PAGE_SIZE;
> + buf++;
> + }
> + xe_ggtt_invalidate(ggtt);
> +
> + return 0;
> +}
> +
> #endif
>
> /**
> diff --git a/drivers/gpu/drm/xe/xe_ggtt.h b/drivers/gpu/drm/xe/xe_ggtt.h
> index 75fc7a1efea76..5f55f80fe3adc 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt.h
> +++ b/drivers/gpu/drm/xe/xe_ggtt.h
> @@ -42,7 +42,10 @@ int xe_ggtt_dump(struct xe_ggtt *ggtt, struct drm_printer
> *p);
> u64 xe_ggtt_print_holes(struct xe_ggtt *ggtt, u64 alignment, struct
> drm_printer *p);
>
> #ifdef CONFIG_PCI_IOV
> +size_t xe_ggtt_pte_size(struct xe_ggtt *ggtt, size_t size);
this could be generic (non PCI-IOV only) inline helper or macro here or in .c
size_t to_xe_ggtt_pt_size(size_t size);
and then more elegant solution would be to expose
size_t xe_ggtt_node_pt_size(const struct xe_ggtt_node *node);
and yes, that would require to additionally expose something from
gt_sriov_pf_config
as migration code doesn't have access to this node,
but maybe xe_gt_sriov_pf_config_ggtt_save() can be updated to also support
'query' mode?
size_t xe_gt_sriov_pf_config_ggtt_save(gt, vfid, buf, size) -> bytes
saved
size_t xe_gt_sriov_pf_config_ggtt_save(gt, vfid, NULL, 0) -> size to be
saved
> void xe_ggtt_assign(const struct xe_ggtt_node *node, u16 vfid);
> +int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size);
> +int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t
> size, u16 vfid);
> #endif
>
> #ifndef CONFIG_LOCKDEP
> diff --git a/drivers/gpu/drm/xe/xe_ggtt_types.h
> b/drivers/gpu/drm/xe/xe_ggtt_types.h
> index c5e999d58ff2a..dacd796f81844 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt_types.h
> +++ b/drivers/gpu/drm/xe/xe_ggtt_types.h
> @@ -78,6 +78,8 @@ struct xe_ggtt_pt_ops {
> u64 (*pte_encode_flags)(struct xe_bo *bo, u16 pat_index);
> /** @ggtt_set_pte: Directly write into GGTT's PTE */
> void (*ggtt_set_pte)(struct xe_ggtt *ggtt, u64 addr, u64 pte);
> + /** @ggtt_get_pte: Directly read from GGTT's PTE */
> + u64 (*ggtt_get_pte)(struct xe_ggtt *ggtt, u64 addr);
> };
>
> #endif
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> index c0c0215c07036..c857879e28fe5 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> @@ -726,6 +726,50 @@ int xe_gt_sriov_pf_config_set_fair_ggtt(struct xe_gt
> *gt, unsigned int vfid,
> return xe_gt_sriov_pf_config_bulk_set_ggtt(gt, vfid, num_vfs, fair);
> }
>
> +/**
> + * xe_gt_sriov_pf_config_ggtt_save() - Save a VF provisioned GGTT data into
> a buffer.
> + * @gt: the &xe_gt
> + * @vfid: VF identifier (can't be 0)
> + * @buf: the GGTT data destination buffer
> + * @size: the size of the buffer
> + *
> + * This function can only be called on PF.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_config_ggtt_save(struct xe_gt *gt, unsigned int vfid,
> + void *buf, size_t size)
> +{
> + xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> + xe_gt_assert(gt, vfid);
> +
> + guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
> +
> + return xe_ggtt_node_save(pf_pick_vf_config(gt, vfid)->ggtt_region, buf,
> size);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_config_ggtt_restore() - Restore a VF provisioned GGTT data
> from a buffer.
> + * @gt: the &xe_gt
> + * @vfid: VF identifier (can't be 0)
> + * @buf: the GGTT data source buffer
> + * @size: the size of the buffer
> + *
> + * This function can only be called on PF.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_config_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> + const void *buf, size_t size)
> +{
> + xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> + xe_gt_assert(gt, vfid);
> +
> + guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
> +
> + return xe_ggtt_node_load(pf_pick_vf_config(gt, vfid)->ggtt_region, buf,
> size, vfid);
> +}
> +
> static u32 pf_get_min_spare_ctxs(struct xe_gt *gt)
> {
> /* XXX: preliminary */
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> index 513e6512a575b..6916b8f58ebf2 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> @@ -61,6 +61,11 @@ ssize_t xe_gt_sriov_pf_config_save(struct xe_gt *gt,
> unsigned int vfid, void *bu
> int xe_gt_sriov_pf_config_restore(struct xe_gt *gt, unsigned int vfid,
> const void *buf, size_t size);
>
> +int xe_gt_sriov_pf_config_ggtt_save(struct xe_gt *gt, unsigned int vfid,
> + void *buf, size_t size);
> +int xe_gt_sriov_pf_config_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> + const void *buf, size_t size);
> +
> bool xe_gt_sriov_pf_config_is_empty(struct xe_gt *gt, unsigned int vfid);
>
> int xe_gt_sriov_pf_config_init(struct xe_gt *gt);
