Re: [PATCH 2/3] drm/xe/tlb_inval: add FULL_INTRA_VF GGTT invalidation helper

[Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>]

On Fri, Aug 21, 2026 at 01:12:37PM +0530, tilak.tirumalesh.tangudu@intel.com wrote:
>From: Tangudu Tilak Tirumalesh <tilak.tirumalesh.tangudu@intel.com>
>
>Add xe_tlb_inval_ggtt_full() and its GuC backend
>send_tlb_inval_ggtt_full() for full engine TLB invalidation
>within the requesting VF partition across all engines.
>
>Assisted-by: Claude:claude-opus-4-8
>Signed-off-by: Tangudu Tilak Tirumalesh <tilak.tirumalesh.tangudu@intel.com>
>---
> drivers/gpu/drm/xe/xe_guc_tlb_inval.c   | 21 +++++++++++++++++++++
> drivers/gpu/drm/xe/xe_tlb_inval.c       | 25 +++++++++++++++++++++++++
> drivers/gpu/drm/xe/xe_tlb_inval.h       |  1 +
> drivers/gpu/drm/xe/xe_tlb_inval_types.h | 10 ++++++++++
> 4 files changed, 57 insertions(+)
>
>diff --git a/drivers/gpu/drm/xe/xe_guc_tlb_inval.c b/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
>index 046d0655122f..9c7a02827902 100644
>--- a/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
>+++ b/drivers/gpu/drm/xe/xe_guc_tlb_inval.c
>@@ -98,6 +98,25 @@ static int send_tlb_inval_ggtt(struct xe_tlb_inval *tlb_inval, u32 seqno)
> 	return -ECANCELED;
> }
>
>+/*
>+ * Emit FULL_INTRA_VF to invalidate engine TLBs across all engines
>+ * in the VF partition.
>+ */
>+static int send_tlb_inval_ggtt_full(struct xe_tlb_inval *tlb_inval, u32 seqno)
>+{
>+	struct xe_guc *guc = tlb_inval->private;
>+	u32 action[] = {
>+		XE_GUC_ACTION_TLB_INVALIDATION,
>+		seqno,
>+		MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL),
>+	};
>+
>+	if (!xe_guc_ct_enabled(&guc->ct) || !guc->submission_state.enabled)
>+		return -ECANCELED;
>+
>+	return send_tlb_inval(guc, action, ARRAY_SIZE(action));
>+}
>+
> static int send_page_reclaim(struct xe_guc *guc, u32 seqno,
> 			     u64 gpu_addr)
> {
>@@ -346,6 +365,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval)
> static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {
> 	.all = send_tlb_inval_all,
> 	.ggtt = send_tlb_inval_ggtt,
>+	.ggtt_full = send_tlb_inval_ggtt_full,
> 	.ppgtt = send_tlb_inval_asid_ppgtt,
> 	.initialized = tlb_inval_initialized,
> 	.flush = tlb_inval_flush,
>@@ -354,6 +374,7 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {
>
> static const struct xe_tlb_inval_ops guc_tlb_inval_ctx_ops = {
> 	.ggtt = send_tlb_inval_ggtt,
>+	.ggtt_full = send_tlb_inval_ggtt_full,
> 	.all = send_tlb_inval_all,
> 	.ppgtt = send_tlb_inval_ctx_ppgtt,
> 	.initialized = tlb_inval_initialized,
>diff --git a/drivers/gpu/drm/xe/xe_tlb_inval.c b/drivers/gpu/drm/xe/xe_tlb_inval.c
>index bbd21d393062..aa12c7d95448 100644
>--- a/drivers/gpu/drm/xe/xe_tlb_inval.c
>+++ b/drivers/gpu/drm/xe/xe_tlb_inval.c
>@@ -321,6 +321,31 @@ int xe_tlb_inval_ggtt(struct xe_tlb_inval *tlb_inval)
> 	return ret;
> }
>
>+/**
>+ * xe_tlb_inval_ggtt_full() - Full engine TLB invalidation within VF partition
>+ * @tlb_inval: TLB invalidation client
>+ *
>+ * Issue FULL_INTRA_VF to flush engine TLBs of all engines in the
>+ * requesting VF partition.
>+ *
>+ * Return: 0 on success, negative error code on error
>+ */
>+int xe_tlb_inval_ggtt_full(struct xe_tlb_inval *tlb_inval)
>+{
>+	struct xe_tlb_inval_fence fence, *fence_ptr = &fence;
>+	int ret;
>+
>+	if (!tlb_inval->ops->ggtt_full)
>+		return -EOPNOTSUPP;
>+
>+	xe_tlb_inval_fence_init(tlb_inval, fence_ptr, true);
>+	ret = xe_tlb_inval_issue(tlb_inval, fence_ptr,
>+				 tlb_inval->ops->ggtt_full);
>+	xe_tlb_inval_fence_wait(fence_ptr);
>+
>+	return ret;
>+}

NIT...it is lot similar to xe_tlb_inval_ggtt(), may be abstract out
common part?

Niranjana

>+
> /**
>  * xe_tlb_inval_range() - Issue a TLB invalidation for an address range
>  * @tlb_inval: TLB invalidation client
>diff --git a/drivers/gpu/drm/xe/xe_tlb_inval.h b/drivers/gpu/drm/xe/xe_tlb_inval.h
>index a76b7823a5f2..665072b4b81d 100644
>--- a/drivers/gpu/drm/xe/xe_tlb_inval.h
>+++ b/drivers/gpu/drm/xe/xe_tlb_inval.h
>@@ -20,6 +20,7 @@ void xe_tlb_inval_reset(struct xe_tlb_inval *tlb_inval);
> int xe_tlb_inval_all(struct xe_tlb_inval *tlb_inval,
> 		     struct xe_tlb_inval_fence *fence);
> int xe_tlb_inval_ggtt(struct xe_tlb_inval *tlb_inval);
>+int xe_tlb_inval_ggtt_full(struct xe_tlb_inval *tlb_inval);
> void xe_tlb_inval_vm(struct xe_tlb_inval *tlb_inval, struct xe_vm *vm);
> int xe_tlb_inval_range(struct xe_tlb_inval *tlb_inval,
> 		       struct xe_tlb_inval_fence *fence,
>diff --git a/drivers/gpu/drm/xe/xe_tlb_inval_types.h b/drivers/gpu/drm/xe/xe_tlb_inval_types.h
>index 3d1797d186fd..02a356bfef47 100644
>--- a/drivers/gpu/drm/xe/xe_tlb_inval_types.h
>+++ b/drivers/gpu/drm/xe/xe_tlb_inval_types.h
>@@ -36,6 +36,16 @@ struct xe_tlb_inval_ops {
> 	 */
> 	int (*ggtt)(struct xe_tlb_inval *tlb_inval, u32 seqno);
>
>+	/**
>+	 * @ggtt_full: Full engine TLB invalidation within the VF partition
>+	 * @tlb_inval: TLB invalidation client
>+	 * @seqno: Seqno of TLB invalidation
>+	 *
>+	 * Return 0 on success, -ECANCELED if backend is mid-reset, error on
>+	 * failure
>+	 */
>+	int (*ggtt_full)(struct xe_tlb_inval *tlb_inval, u32 seqno);
>+
> 	/**
> 	 * @ppgtt: Invalidate per-process translation TLBs
> 	 * @tlb_inval: TLB invalidation client
>-- 
>2.46.0
>