vgsp-sema r614 diag: name the first shadow-lane device-dst copies and untruncate the arg-miss walk list - the r613b matmul wedge began with an H2D taking the shadow silently (no ledger/pin/walk) and the post-mortem could not see which registered pdb answered what (240B buffer, 3 of 8 entries)

1 files changed, 35 insertions, 6 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 27818b8..3e5b9dd 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -6803,6 +6803,7 @@ static int svc_copy_1d(struct nv_vgpu_device *vgpu, u64 dst, u64 src, size_t tot
 	bool audit = total >= SVC_COPY_AUDIT_MIN;
 	u8 *snap = NULL;
 	static u64 bigcopy_log;
+	static u32 shadow_dst_log;	/* r614 diag: shadow-lane dst witness */
 	bool trace64 = (total == 65536u);
 	const char *dir = "?";
 
@@ -6919,6 +6920,26 @@ static int svc_copy_1d(struct nv_vgpu_device *vgpu, u64 dst, u64 src, size_t tot
 			} else {
 				dst_shadow = true;
 				svc.vas_shadowed++;
+				/*
+				 * r614 diag: the r613b red boot's matmul
+				 * wedge began here - the H2D into the
+				 * client's cudaMalloc'd buffer found no
+				 * ledger/pin/walk answer, took the shadow
+				 * silently, and the later GR fill could not
+				 * consume the shadowed bytes (probe=none for
+				 * every arg base; the launch held 15 s, then
+				 * the degenerate refusal consumed it and the
+				 * client spun). Name the first shadow-lane
+				 * device-dst copies of a boot (bounded), so
+				 * the next red boot is greppable where it
+				 * starts instead of only at the launch.
+				 */
+				if (shadow_dst_log < 64 &&
+				    dst_orig >= SVC_VAS_FLOOR) {
+					shadow_dst_log++;
+					pr_info("sema-shadow-dst: h=0x%x dst=0x%llx total=%zu (device VA with no ledger/pin/walk answer; copy runs into the shadow, a GR fill cannot consume it)\n",
+						owner_ch, dst_orig, total);
+				}
 			}
 		} else {
 			dst = pa;
@@ -20818,7 +20839,14 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu,
 		if (is_arg[w] && svc_fwd_req.gr_n_data == before &&
 		    svc_gr_src_log < 64) {
 			int vi;
-			char vb[240];
+			/* r614 diag: the r613b red-boot post-mortem could not
+			 * tell which registered pdb would have answered an
+			 * unresolved arg base (the 240-byte buffer truncated
+			 * the walk list at three entries, and only the red
+			 * boot's misses were ever printed). A larger buffer
+			 * with a clamped append keeps every registered
+			 * VASpace's walk verdict visible. */
+			char vb[512];
 			int o = 0;
 
 			svc_gr_src_log++;
@@ -20833,11 +20861,12 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu,
 						    svc.vas[vi].entries, base,
 						    svc.vas[vi].ap, &wa, &wap,
 						    gsp_sema_walk_trace != 0);
-				o += snprintf(vb + o, sizeof(vb) - o,
-					      " vas[%d] pdb=0x%llx ent=%u ap=%d rc=%d pa=0x%llx",
-					      vi, svc.vas[vi].pdb,
-					      svc.vas[vi].entries,
-					      svc.vas[vi].ap, wrc, wa);
+				if (o < (int)sizeof(vb) - 1)
+					o += snprintf(vb + o, sizeof(vb) - o,
+						      " vas[%d] pdb=0x%llx ent=%u ap=%d rc=%d pa=0x%llx",
+						      vi, svc.vas[vi].pdb,
+						      svc.vas[vi].entries,
+						      svc.vas[vi].ap, wrc, wa);
 			}
 			pr_info("vgsp-sema: svc-gr arg-miss base=0x%llx probe=%s n_cand=%u pd_found=%u ipin_n=%u disc=0x%llx/%d%s\n",
 				base, pba_src, n_cand, svc_pd_found,