Add <~cond> condition modifier for tightly coupled services

Conditions in Finit are dependencies: if A is asserted, service B is
allowed to run.  When A goes through FLUX (e.g., upstream reloads),
dependents are PAUSED and then simply resumed when the condition is
reasserted -- this is the correct behavior for barrier-style deps
like <pid/syslogd>.

However, some setups have tightly coupled services where dependents
must be reloaded/restarted when an upstream service reloads, not just
resumed.  E.g., the FRR routing stack on Infix OS:

    netd <pid/mgmtd> ← zebra <!pid/netd> ← {staticd,ripd} <!pid/zebra>

When netd reloads (SIGHUP), zebra and its dependents must be restarted
to pick up the new configuration.

The new '~' condition prefix marks a dependency as flux-sensitive:

    service <!~pid/netd> name:zebra ...

When the upstream condition goes FLUX and returns to ON, the dependent
is reloaded (SIGHUP) or restarted (noreload '!') instead of merely
resumed.  Transitivity follows naturally through the condition chain.

Closes #416
Closes #476

Signed-off-by: Joachim Wiberg <troglobit@gmail.com>

Author: troglobit

doc/conditions.md

@@ -14,11 +14,11 @@ specified separated by comma.  Multiple conditions are logically AND'ed
 during evaluation, i.e. all conditions must be satisfied in order for a
 service to run.
 
-A special syntax, using a leading `!` in run/task/service conditions,
-denote if a:
+Two special prefixes can be used inside the angle brackets:
 
- - service does not support `SIGHUP`
- - run/task should not block runlevel changes (i.e., bootstrap)
+ - `!` -- service does not support `SIGHUP` (noreload), or run/task
+   should not block runlevel changes (i.e., bootstrap)
+ - `~` -- propagate reload from this dependency, see below
 
 Finit guarantees by default that all run/tasks run (at least) once
 per runlevel.  For most tasks this is a good default, for example
@@ -53,6 +53,30 @@ moving to the next runlevel after bootstrap, so we set `<!>`:
     task [S0123456789] <!sys/pwr/fail> name:pwrfail initctl poweroff -- Power failure, shutting down
 
 
+Propagating Reload in Dependencies
+-----------------------------------
+
+By default, when a service reloads during `initctl reload`, dependent
+services are paused (`SIGSTOP`) and simply resumed (`SIGCONT`) when the
+condition is reasserted.  This is correct for barrier-style dependencies
+like `<pid/syslogd>`, where dependents just need syslogd running and do
+not care if it reloads its config.
+
+For services that need to react when their upstream reloads, the `~`
+prefix propagates the reload from the dependency:
+
+    service <pid/svc_a>   name:svc_b /sbin/svc_b -- Needs A (barrier)
+    service <!~pid/svc_b> name:svc_c /sbin/svc_c -- Propagate reload from B
+
+Here, `<~pid/svc_b>` means: propagate a reload of `svc_b` to `svc_c`.
+When `svc_b` reloads, `svc_c` will be restarted (because of `!`,
+noreload) instead of merely resumed.  If `svc_c` supported `SIGHUP`
+(no `!` prefix), it would be sent `SIGHUP` instead.
+
+This is similar to systemd's `PropagatesReloadTo=` directive, but
+declared on the consumer side rather than the provider side.
+
+
 Triggering
 ----------
 
@@ -281,6 +305,11 @@ This STOP/CONT handling minimizes the number of unnecessary service
 restarts that would otherwise occur because a depending service was sent
 `SIGHUP` for example.
 
+Services with the `~` prefix are an exception to this rule: when their
+conditions return to `on` after being in `flux`, the reload is propagated
+-- the service is reloaded (SIGHUP) or restarted (noreload `!`) instead
+of simply being resumed.
+
 Therefore, any plugin that supplies Finit with conditions must ensure
 that their state is updated after each reconfiguration.  This can be
 done by binding to the `HOOK_SVC_RECONF` hook.  For an example of how

doc/config/services.md

@@ -47,6 +47,14 @@ waits for another service, `zebra`, to have created its PID file in
 *all* files in `/var/run`, for each file named `*.pid`, or `*/pid`,
 Finit opens it and find the matching `NAME:ID` using the PID.
 
+The condition can be prefixed with `!` and/or `~`:
+
+ - `<!pid/zebra>` -- `ospfd` does not support `SIGHUP` (noreload)
+ - `<~pid/zebra>` -- propagate reload from `zebra` to `ospfd`
+ - `<!~pid/zebra>` -- both: noreload and propagate reload
+
+For details, see the [Finit Conditions](../conditions.md) document.
+
 Some services do not maintain a PID file and rather than patching each
 application Finit provides a workaround.  A `pid` keyword can be set
 to have Finit automatically create (when starting) and later remove

src/conf.c

@@ -745,8 +745,21 @@ void conf_parse_cond(svc_t *svc, char *cond)
 		return;
 	}
 
+	/*
+	 * The '~' prefix means a reload of the upstream service is
+	 * propagated to this service -- it will be reloaded (SIGHUP)
+	 * or restarted (noreload), not just paused and resumed.
+	 * Syntax: <!~pid/foo,~pid/bar> or <~pid/foo>
+	 *
+	 * Strip '~' from each condition and set the service-level
+	 * flag.  This allows '~' on any condition in the list.
+	 */
 	svc->cond[0] = 0;
 	for (i = 0, c = strtok(ptr, ","); c; c = strtok(NULL, ","), i++) {
+		if (c[0] == '~') {
+			c++;
+			svc->flux_reload = 1;
+		}
 		devmon_add_cond(c);
 		if (i)
 			strlcat(svc->cond, ",", sizeof(svc->cond));

src/service.c

@@ -3146,6 +3146,20 @@ int service_step(svc_t *svc)
 		switch (cond) {
 		case COND_ON:
 			kill(svc->pid, SIGCONT);
+
+			/*
+			 * Propagate reload (~): upstream reloaded, so
+			 * we must also reload/restart, not just resume.
+			 */
+			if (svc->flux_reload && !svc_is_changed(svc)) {
+				svc_set_state(svc, SVC_RUNNING_STATE);
+				if (svc_is_noreload(svc))
+					service_stop(svc);
+				else
+					service_reload(svc);
+				break;
+			}
+
 			svc_set_state(svc, SVC_RUNNING_STATE);
 			/* Reassert condition if we go from waiting and no change */
 			if (!svc_is_changed(svc)) {

src/svc.h

@@ -146,6 +146,7 @@ typedef struct svc {
 	int            starting;       /* ... waiting for pidfile to be re-asserted */
 	int	       runlevels;
 	int            sighup;	       /* This service supports SIGHUP :) */
+	int            flux_reload;    /* Propagate reload from dependency, '~' prefix */
 	int	       forking;	       /* This is a service/sysv daemon that forks, wait for it ... */
 	svc_block_t    block;	       /* Reason that this service is currently stopped */
 	char           cond[MAX_COND_LEN];

test/dep-chain-reload.sh

@@ -4,17 +4,18 @@
 # Four services in a chain: A → B → C → D, all using notify:pid.
 # B and C are placed in separate sub-config files so we can use
 # 'initctl touch' on them independently.  B supports SIGHUP (reload),
-# while C and D use the '!' condition prefix (noreload), matching
-# a common pattern in real-world setups (e.g., FRR daemons on Infix).
+# while C and D use '!' (noreload) and '~' (propagate reload),
+# matching a common pattern in real-world setups (e.g., FRR routing
+# daemons on Infix OS).
 #
-# Chain: A ← B <pid/A> ← C <!pid/B> ← D <!pid/C>
+# Chain: A ← B <pid/A> ← C <!~pid/B> ← D <!~pid/C>
 #
 # Test 1 - touch C + reload (the "Infix" scenario):
 #   C is in the middle of the chain.  After 'initctl touch svc_c.conf'
 #   + 'initctl reload':
 #     - A and B should be unaffected (same PID)
 #     - C is stopped/started (config was touched, noreload)
-#     - D must be restarted (reverse dep of C)
+#     - D must be restarted (reload propagated from C)
 #
 # Test 2 - crash (kill -9):
 #   When B is killed with SIGKILL the crash path (RUNNING → HALTED)
@@ -26,7 +27,7 @@
 #   svc_b.conf' + 'initctl reload':
 #     - A should be unaffected (same PID)
 #     - B is SIGHUP'd (same PID, config was touched)
-#     - C and D must be restarted (transitive reverse deps)
+#     - C and D must be restarted (reload propagated, '~' prefix)
 
 set -eu
 
@@ -47,13 +48,13 @@ test_setup()
 {
     run "cat >> $FINIT_CONF" <<EOF
 service log:stdout notify:pid               name:svc_a serv -np -i svc_a -- Chain root
-service log:stdout notify:pid <!pid/svc_c>  name:svc_d serv -np -i svc_d -- Needs C
+service log:stdout notify:pid <!~pid/svc_c>  name:svc_d serv -np -i svc_d -- Needs C
 EOF
     run "cat >> $FINIT_RCSD/svc_b.conf" <<EOF
-service log:stdout notify:pid <pid/svc_a>   name:svc_b serv -np -i svc_b -- Needs A
+service log:stdout notify:pid <pid/svc_a>    name:svc_b serv -np -i svc_b -- Needs A
 EOF
     run "cat >> $FINIT_RCSD/svc_c.conf" <<EOF
-service log:stdout notify:pid <!pid/svc_b>  name:svc_c serv -np -i svc_c -- Needs B
+service log:stdout notify:pid <!~pid/svc_b>  name:svc_c serv -np -i svc_c -- Needs B
 EOF
 }