vgsp: m170 fix - the aggregate stall check overwrote the per-argument fail-closed signal (one resolved operand masked every unresolved one; r175 matmul held per-arg then queued anyway, bad=65536)

1 files changed, 12 insertions, 3 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 8303c87..79faf61 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -13404,10 +13404,19 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu)
 				svc.ipin_n, svc_pd_base, svc_pd_ap, vb);
 		}
 	}
-	svc_gr_data_stalled = (n_arg_hi > 0 && n_arg_hi_res == 0);
+	/*
+	 * m170 fix: never CLEAR the per-argument stall signal set above. The
+	 * original aggregate condition only held a launch when NO high
+	 * argument resolved at all; one resolved kernel operand masked every
+	 * unresolved one (r175: matmul cand[5]/[6] probe=none held fail-closed
+	 * per-arg, then this line reset the flag because cand[13]/[14] had
+	 * pins - the launch queued with two stale operands, bad=65536).
+	 */
+	if (n_arg_hi > 0 && n_arg_hi_res == 0)
+		svc_gr_data_stalled = true;
 	if (svc_gr_data_stalled)
-		pr_info("vgsp-sema: svc-gr data fill: %u candidate ptrs, %u pages resolved; %u high argument ptr(s), 0 resolved -> launch held for a bounded retry\n",
-			n_cand, svc_fwd_req.gr_n_data, n_arg_hi);
+		pr_info("vgsp-sema: svc-gr data fill: %u candidate ptrs, %u pages resolved; %u high argument ptr(s), %u resolved -> launch held for a bounded retry\n",
+			n_cand, svc_fwd_req.gr_n_data, n_arg_hi, n_arg_hi_res);
 	else
 		pr_info("vgsp-sema: svc-gr data fill: %u candidate ptrs, %u pages resolved (reloc 0x%llx+)\n",
 			n_cand, svc_fwd_req.gr_n_data, reloc_base);