workloads/cln: patch coverage build to break deadlock

workloads/cln/Dockerfile.coverage

@@ -65,14 +65,13 @@ RUN git clone --depth 1 --branch ${CLN_VERSION} --recurse-submodules \
     https://github.com/ElementsProject/lightning.git /cln
 WORKDIR /cln
 RUN pip3 install --break-system-packages mako
-# Currently CLN does an unnecessary SIGKILL of per-peer subdaemons (openingd,
-# channeld, etc.) during node shutdown. If the SIGKILL arrives before the
-# subdaemon shuts down gracefully on its own, we lose coverage profile data
-# (profraw files) for that subdaemon.
-#
-# Remove the SIGKILL to ensure we get full coverage data.
-RUN grep -q 'kill(sd->pid, SIGKILL)' lightningd/subd.c && \
-    sed -i '/kill(sd->pid, SIGKILL)/d' lightningd/subd.c
+# When a per-peer subdaemon (openingd, channeld, etc.) dies mid-run (e.g. after
+# a channel error), lightningd SIGKILLs it. SIGKILL skips the atexit handler
+# that writes the profraw file, so we lose that subdaemon's entire coverage
+# profile. Close its socket instead, which makes it exit by itself, only killing
+# as a fallback.
+COPY workloads/cln/patches/subd-coverage-shutdown.patch .
+RUN git apply subd-coverage-shutdown.patch
 # Reduce the block-polling interval from the 30s default to 2s. CLN only
 # notices newly confirmed transactions when this poll runs, so shorten the
 # interval to make block-related events available to scenarios promptly.

workloads/cln/patches/subd-coverage-shutdown.patch

@@ -0,0 +1,33 @@
+Let a dying subdaemon exit by itself before SIGKILLing it.
+
+SIGKILL skips the atexit handler that writes the subdaemon's LLVM profile data,
+so a coverage build loses every line it executed.
+
+Closing our end of its socket gets it to exit on its own, flushing the profraw
+file.  This is what subd_shutdown() already does for the global daemons, so we
+apply the same approach here for destroy_subd().
+
+Closing is not always enough -- subdaemons can still block on the HSM fd, which
+causes lightningd to wait several minutes for it.  So keep the kill as a
+fallback after 5s.
+
+--- a/lightningd/subd.c
++++ b/lightningd/subd.c
+@@ -611,9 +611,17 @@
+ 			if (!sd->must_not_exit) {
+ 				log_debug(sd->log,
+ 					  "Status closed, but not exited. Killing");
++				sd->conn = tal_free(sd->conn);
++				for (size_t i = 0; i < 500; i++) {
++					if (waitpid(sd->pid, &status,
++						    WNOHANG) > 0)
++						goto reaped;
++					usleep(10000);
++				}
+ 				kill(sd->pid, SIGKILL);
+ 			}
+ 			waitpid(sd->pid, &status, 0);
++		reaped:
+ 			fail_if_subd_fails = false;
+ 			break;
+ 		case -1: