linkd

commit ff736bc10baadea6c66cc4761ecb7de54571b4b8
parent dcecd3c0e451813776a15505a9d71ce99b449081
Author: finwo <finwo@pm.me>
Date:   Wed, 30 Sep 2026 15:16:08 +0200

Wake the poll loop on SIGCHLD so a dead dhcp client is reported at once

Diffstat:
M
src/dhcp/client.c
 | 
23
+++++++++++++++++++++++
M
src/netlink/netlink.c
 | 
3
++-
M
tests/unit/test_dhcp.sh
 | 
11
++++++++++-

3 files changed, 35 insertions(+), 2 deletions(-)
diff --git a/src/dhcp/client.c b/src/dhcp/client.c
@@ -73,6 +73,28 @@ static char *find_in_path(const char *bin) {
 }
 
 // ---------------------------------------------------------------------------
+
+// poll(2) is never restarted, so this drops nl_run out of its wait and the
+// reap at the top of that loop runs immediately. SA_RESTART keeps the
+// blocking waitpid in plugin.c from being interrupted.
+static void sigchld_noop(int sig) { (void)sig; }
+
+static void arm_sigchld(void) {
+  static int armed = 0;
+  if (armed) return;
+  armed = 1;
+
+  struct sigaction sa;
+  memset(&sa, 0, sizeof(sa));
+  sa.sa_handler = sigchld_noop;
+  sa.sa_flags   = SA_RESTART;
+  sigemptyset(&sa.sa_mask);
+  if (sigaction(SIGCHLD, &sa, NULL) != 0) {
+    log_warn("dhcp: sigaction SIGCHLD: %s", strerror(errno));
+  }
+}
+
+// ---------------------------------------------------------------------------
 // process table
 
 static struct dhcp_proc *proc_find(const char *ifname) {
@@ -107,6 +129,7 @@ int dhcp_start(const char *ifname) {
     return -1;
   }
   const char *const *tpl = (const char *const *)cfg->dhcp_client;
+  arm_sigchld();
 
   char **argv = calloc(argc + 1, sizeof(char *));
   if (!argv) {
diff --git a/src/netlink/netlink.c b/src/netlink/netlink.c
@@ -587,6 +587,8 @@ int nl_run(int fd, int resync_interval_s) {
   clock_gettime(CLOCK_MONOTONIC, &last_resync);
 
   while (!nl_stop) {
+    dhcp_reap();
+
     // Netlink occupies slot 0; the IPC server contributes its listeners and
     // every live connection after it.
     struct pollfd pfds[1 + IPC_MAX_POLLFDS];
@@ -622,7 +624,6 @@ int nl_run(int fd, int resync_interval_s) {
       log_error("netlink: poll failed: %s", strerror(errno));
       return -1;
     }
-    dhcp_reap();
     if (!pr) continue;
 
     if (nfds > 1) ipc_dispatch(&pfds[1], nfds - 1);
diff --git a/tests/unit/test_dhcp.sh b/tests/unit/test_dhcp.sh
@@ -30,8 +30,17 @@ iface swp-static inet static
     address 10.9.9.9/24
 EOF
 
+# An empty netns, because reaping is driven by SIGCHLD waking poll(2): on a
+# host with ambient netlink traffic the loop spins anyway and a broken reap
+# still passes. 3s for the same reason -- the wake is causal, not timed.
+if unshare --user --map-root-user --net true 2>/dev/null; then
+  in_ns() { unshare --user --map-root-user --net "$@"; }
+else
+  in_ns() { "$@"; }
+fi
+
 run_linkd() {
-  timeout 15 env PATH="${T}/emptybin" DHCP_LOG="${T}/dhcp.log" \
+  in_ns timeout 3 env PATH="${T}/emptybin" DHCP_LOG="${T}/dhcp.log" \
     "${BIN}" --config "$1" --resync-interval 0 \
     >"${T}/linkd.log" 2>&1 || true
   cat "${T}/linkd.log"