Merge pull request #6543 from garlick/no_child_watcher

libflux: drop child watchers and the FLUX_REACTOR_SIGCHLD flag

Author: mergify[bot]

doc/Makefile.am

@@ -119,7 +119,6 @@ MAN3_FILES_PRIMARY = \
 	man3/flux_msg_cmp.3 \
 	man3/flux_msg_create.3 \
 	man3/flux_msg_handler_addvec.3 \
-	man3/flux_child_watcher_create.3 \
 	man3/flux_signal_watcher_create.3 \
 	man3/flux_stat_watcher_create.3 \
 	man3/flux_respond.3 \
@@ -222,8 +221,6 @@ MAN3_FILES_SECONDARY = \
 	man3/flux_msg_handler_stop.3 \
 	man3/flux_get_handle_watcher.3 \
 	man3/flux_msg_handler_delvec.3 \
-	man3/flux_child_watcher_get_rpid.3 \
-	man3/flux_child_watcher_get_rstatus.3 \
 	man3/flux_signal_watcher_get_signum.3 \
 	man3/flux_stat_watcher_get_rstat.3 \
 	man3/flux_respond_raw.3 \

doc/man3/flux_child_watcher_create.rst

@@ -27,14 +27,8 @@ DESCRIPTION
 ===========
 
 :func:`flux_reactor_create` creates a :type:`flux_reactor_t` object which can
-be used to monitor for events on file descriptors, ZeroMQ sockets, timers, and
-:type:`flux_t` broker handles.
-
-There is currently only one possible flag for reactor creation:
-
-FLUX_REACTOR_SIGCHLD
-   The reactor will internally register a SIGCHLD handler and be capable
-   of handling flux child watchers (see :man3:`flux_child_watcher_create`).
+be used to monitor for events on file descriptors, timers, and
+:type:`flux_t` broker handles.  :var:`flags` should be set to zero.
 
 For each event source and type that is to be monitored, a :type:`flux_watcher_t`
 object is created using a type-specific create function, and started

doc/man3/flux_reactor_create.rst

@@ -37,6 +37,10 @@ When one :var:`callback` is shared by multiple watchers, the signal number that
 triggered the event can be obtained with
 :func:`flux_signal_watcher_get_signum`.
 
+Signal handling can be tricky in multi-threaded programs.  It is advisable
+to handle signals in the main thread only.  For example, block signals by
+calling :linux:man2:`sigprocmask` before spawning other threads, and register
+signal watchers only in the main thread.
 
 RETURN VALUE
 ============

doc/man3/flux_signal_watcher_create.rst

@@ -75,9 +75,6 @@
     ('man3/flux_attr_get', 'flux_attr_get', 'get/set Flux broker attributes', [author], 3),
     ('man3/flux_aux_set', 'flux_aux_get', 'get/set auxiliary handle data', [author], 3),
     ('man3/flux_aux_set', 'flux_aux_set', 'get/set auxiliary handle data', [author], 3),
-    ('man3/flux_child_watcher_create', 'flux_child_watcher_get_rpid', 'create child watcher', [author], 3),
-    ('man3/flux_child_watcher_create', 'flux_child_watcher_get_rstatus', 'create child watcher', [author], 3),
-    ('man3/flux_child_watcher_create', 'flux_child_watcher_create', 'create child watcher', [author], 3),
     ('man3/flux_core_version', 'flux_core_version_string', 'get flux-core version', [author], 3),
     ('man3/flux_core_version', 'flux_core_version', 'get flux-core version', [author], 3),
     ('man3/flux_event_decode', 'flux_event_decode_raw', 'encode/decode a Flux event message', [author], 3),

doc/manpages.py

@@ -940,3 +940,4 @@ cPATH
 SATTR
 myprogram
 unref
+sigprocmask

doc/test/spell.en.pws

@@ -248,16 +248,15 @@ int main (int argc, char *argv[])
      * ctx.h is used conventionally within the broker for RPCs, message
      * handlers, etc.  ctx.h_internal belongs to the broker's routing logic
      * and is accessed using flux_send() and flux_recv() only.  Both handles
-     * share a reactor, which is created with FLUX_REACTOR_SIGCHLD in order
-     * to support libsubprocess.
+     * share a reactor.
      *
      * N.B. since both handles are in the same thread, synchronous RPCs on
      * ctx.h will deadlock.  The main broker reactor must run in order
      * to move messages from the interthread queue to the routing logic.
      * Careful with flux_attr_get(), which hides a synchronous RPC if the
      * requested value is not cached.
      */
-    if (!(ctx.reactor = flux_reactor_create (FLUX_REACTOR_SIGCHLD))
+    if (!(ctx.reactor = flux_reactor_create (0))
         || !(ctx.h = flux_open ("interthread://broker", 0))
         || flux_set_reactor (ctx.h, ctx.reactor) < 0
         || !(ctx.h_internal = flux_open ("interthread://broker", 0))

src/broker/broker.c

@@ -350,7 +350,6 @@ void badinput (flux_t *h)
 int main (int argc, char *argv[])
 {
     flux_t *h;
-    flux_reactor_t *r;
 
     plan (NO_PLAN);
 
@@ -360,12 +359,8 @@ int main (int argc, char *argv[])
 
     if (!(logs = zlist_new ()))
         BAIL_OUT ("zlist_new failed");
-    if (!(r = flux_reactor_create (FLUX_REACTOR_SIGCHLD)))
-        BAIL_OUT ("flux_reactor_create failed");
     if (!(h = flux_open ("loop://", 0)))
         BAIL_OUT ("could not create loop handle");
-    if (flux_set_reactor (h, r) < 0)
-        BAIL_OUT ("flux_set_reactor failed");
     if (flux_attr_set_cacheonly (h, "rank", "0") < 0)
         BAIL_OUT ("flux_attr_set_cacheonly rank failed");
     flux_log_set_redirect (h, diag_logger, NULL);
@@ -374,7 +369,6 @@ int main (int argc, char *argv[])
     basic (h);
     badinput (h);
 
-    flux_reactor_destroy (r);
     flux_close (h);
 
     clear_list (logs);

src/broker/test/runat.c

@@ -385,10 +385,8 @@ static int cmd_proxy (optparse_t *p, int ac, char *av[])
     free (uri);
     flux_log_set_appname (ctx.h, "proxy");
     ctx.proxy_user = getuid ();
-    if (!(r = flux_reactor_create (FLUX_REACTOR_SIGCHLD)))
-        log_err_exit ("flux_reactor_create");
-    if (flux_set_reactor (ctx.h, r) < 0)
-        log_err_exit ("flux_set_reactor");
+    if (!(r = flux_get_reactor (ctx.h)))
+        log_err_exit ("flux_get_reactor");
 
     /* Register handler for loss of broker connection if --reconnect
      */

src/cmd/builtin/proxy.c

@@ -1074,7 +1074,7 @@ int start_session (const char *cmd_argz,
         if (signal (SIGTTOU, SIG_IGN) == SIG_ERR)
             log_err_exit ("signal");
     }
-    if (!(ctx.reactor = flux_reactor_create (FLUX_REACTOR_SIGCHLD)))
+    if (!(ctx.reactor = flux_reactor_create (0)))
         log_err_exit ("flux_reactor_create");
     if (!(ctx.timer = flux_timer_watcher_create (ctx.reactor,
                                                  ctx.exit_timeout,