vgsp: revert m170 per-arg stall persistence to aggregate semantics (benign 0-page junk candidates in fresh clients' launches stalled mail forever, r176); the per-arg fail-closed line stays as diagnostics

1 files changed, 9 insertions, 8 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 79faf61..0217594 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -13405,15 +13405,16 @@ static void svc_gr_data_fill(struct nv_vgpu_device *vgpu)
 		}
 	}
 	/*
-	 * 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).
+	 * m170 revert: the per-argument fail-closed signal is kept as
+	 * DIAGNOSTIC only (the held fail-closed line above identifies the
+	 * exact unresolved operand), but the launch-level stall keeps the
+	 * historical aggregate semantics: a fresh client's launch stream
+	 * carries BENIGN 0-page candidates (junk pointer words that never
+	 * translate), and persisting the per-arg flag held mail's first
+	 * launch forever (r176). Only a launch with NO resolvable high
+	 * argument holds for the bounded m150 retry.
 	 */
-	if (n_arg_hi > 0 && n_arg_hi_res == 0)
-		svc_gr_data_stalled = true;
+	svc_gr_data_stalled = (n_arg_hi > 0 && n_arg_hi_res == 0);
 	if (svc_gr_data_stalled)
 		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);