sema: r420d bounded 2D-copy geometry witness (sema-op2d)

Every copy op logs dst/src/bytes/rows/dpitch/spitch beside the decoded
line_length/line_count/pitches, so the memset2d readback forensics can tell
'one row per push' from 'N rows in one op' and whether the pitch methods
arrived. Bounded by the existing opadd budget.

1 files changed, 21 insertions, 0 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 3b7c5e8..931669c 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -10424,6 +10424,27 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf,
 						lm->elem, lm->count,
 						lm->aperture, lm->src_aperture,
 						lm->phys, lm->src_phys);
+					/*
+					 * r420d: 2D geometry of every copy op
+					 * (the memset2d readback forensics -
+					 * the driver decomposes cudaMemcpy2D
+					 * into per-push ops whose LINE_COUNT/
+					 * PITCH decode must be visible to
+					 * tell "one row per op" from "N rows
+					 * in one op"). Bounded.
+					 */
+					if (lm->type == 1)
+						pr_info("sema-op2d: h=0x%x dst=0x%llx src=0x%llx bytes=%u rows=%u dpitch=%u spitch=%u line=%u lc=%u have_po=%u po=%u have_pi=%u pi=%u\n",
+							ch->h_object, lm->dst,
+							lm->src, lm->bytes,
+							lm->rows, lm->dpitch,
+							lm->spitch,
+							st.line_length,
+							st.line_count,
+							st.have_pitch_out,
+							st.pitch_out,
+							st.have_pitch_in,
+							st.pitch_in);
 				}
 			}