cuda-ab/soak: accept a terminal STATUS line as completion - the guest wrapper's EXIT= marker is lost when the guest's process reaping stalls (observed: victory_mail reached STATUS OK but the zombie was never reaped; soak reported a false still-running failure)

1 files changed, 13 insertions, 0 deletions

diff --git a/tools/cuda-ab/soak.sh b/tools/cuda-ab/soak.sh
index f870fda..c150692 100755
--- a/tools/cuda-ab/soak.sh
+++ b/tools/cuda-ab/soak.sh
@@ -135,6 +135,19 @@ run_one() {
 		r="$(gssh "cat $out 2>/dev/null")" || true
 		[ -n "$r" ] && text="$r"
 		printf '%s' "$text" | grep -q 'EXIT=' && break
+		# The guest wrapper's EXIT= marker can be lost when the guest's
+		# process reaping stalls (observed: the client reached STATUS OK
+		# but its zombie was never reaped and `timeout` never wrote the
+		# marker). A terminal STATUS line is proof of completion by
+		# itself - accept it after a short grace so the EXIT= marker
+		# still wins when it does arrive.
+		if printf '%s' "$text" | grep -qE 'STATUS (OK|FAIL)'; then
+			sleep 10
+			r="$(gssh "cat $out 2>/dev/null")" || true
+			[ -n "$r" ] && text="$r"
+			printf '%s' "$text" | grep -q 'EXIT=' && break
+			break
+		fi
 		tnow=$(date +%s)
 		if [ $((tnow - t0)) -gt $((TMO + 120)) ]; then
 			if client_alive "$bin"; then