Add support for spawning panes in separate cgroups with systemd and a configure

flag to disable. From Eric T Johnson yut23 AT gvljohnsons DOT com in GitHub
issue 3514.

4 files changed, 178 insertions, 1 deletions

diff --git a/compat.h b/compat.h
index aefc0d38..cabdf3ad 100644
--- a/compat.h
+++ b/compat.h
@@ -425,6 +425,7 @@ void		*recallocarray(void *, size_t, size_t, size_t);
 /* systemd.c */
 int		 systemd_activated(void);
 int		 systemd_create_socket(int, char **);
+int		 systemd_move_pid_to_new_cgroup(pid_t, char **);
 #endif
 
 #ifdef HAVE_UTF8PROC
diff --git a/compat/systemd.c b/compat/systemd.c
index cce42ed4..6f790a32 100644
--- a/compat/systemd.c
+++ b/compat/systemd.c
@@ -19,7 +19,10 @@
 #include <sys/types.h>
 #include <sys/un.h>
 
+#include <systemd/sd-bus.h>
 #include <systemd/sd-daemon.h>
+#include <systemd/sd-login.h>
+#include <systemd/sd-id128.h>
 
 #include <string.h>
 
@@ -64,3 +67,149 @@ fail:
 		xasprintf(cause, "systemd socket error (%s)", strerror(errno));
 	return (-1);
 }
+
+int
+systemd_move_pid_to_new_cgroup(pid_t pid, char **cause)
+{
+	sd_bus_error	 error = SD_BUS_ERROR_NULL;
+	sd_bus_message	*m = NULL, *reply = NULL;
+	sd_bus 		*bus = NULL;
+	char		*name, *desc, *slice;
+	sd_id128_t	 uuid;
+	int		 r;
+	pid_t		 parent_pid;
+
+	/* Connect to the session bus. */
+	r = sd_bus_default_user(&bus);
+	if (r < 0) {
+		xasprintf(cause, "failed to connect to session bus: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* Start building the method call. */
+	r = sd_bus_message_new_method_call(bus, &m,
+	    "org.freedesktop.systemd1",
+	    "/org/freedesktop/systemd1",
+	    "org.freedesktop.systemd1.Manager",
+	    "StartTransientUnit");
+	if (r < 0) {
+		xasprintf(cause, "failed to create bus message: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* Generate a unique name for the new scope, to avoid collisions. */
+	r = sd_id128_randomize(&uuid);
+	if (r < 0) {
+		xasprintf(cause, "failed to generate uuid: %s", strerror(-r));
+		goto finish;
+	}
+	xasprintf(&name, "tmux-spawn-" SD_ID128_UUID_FORMAT_STR ".scope",
+	    SD_ID128_FORMAT_VAL(uuid));
+	r = sd_bus_message_append(m, "s", name);
+	free(name);
+	if (r < 0) {
+		xasprintf(cause, "failed to append to bus message: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* Mode: fail if there's a queued unit with the same name. */
+	r = sd_bus_message_append(m, "s", "fail");
+	if (r < 0) {
+		xasprintf(cause, "failed to append to bus message: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* Start properties array. */
+	r = sd_bus_message_open_container(m, 'a', "(sv)");
+	if (r < 0) {
+		xasprintf(cause, "failed to start properties array: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	parent_pid = getpid();
+	xasprintf(&desc, "tmux child pane %ld launched by process %ld",
+	    (long)pid, (long)parent_pid);
+	r = sd_bus_message_append(m, "(sv)", "Description", "s", desc);
+	free(desc);
+	if (r < 0) {
+		xasprintf(cause, "failed to append to properties: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/*
+	 * Inherit the slice from the parent process, or default to
+	 * "app-tmux.slice" if that fails.
+	 */
+	r = sd_pid_get_user_slice(parent_pid, &slice);
+	if (r < 0) {
+		slice = xstrdup("app-tmux.slice");
+	}
+	r = sd_bus_message_append(m, "(sv)", "Slice", "s", slice);
+	free(slice);
+	if (r < 0) {
+		xasprintf(cause, "failed to append to properties: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* PIDs to add to the scope: length - 1 array of uint32_t. */
+	r = sd_bus_message_append(m, "(sv)", "PIDs", "au", 1, pid);
+	if (r < 0) {
+		xasprintf(cause, "failed to append to properties: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* Clean up the scope even if it fails. */
+	r = sd_bus_message_append(m, "(sv)", "CollectMode", "s",
+	    "inactive-or-failed");
+	if (r < 0) {
+		xasprintf(cause, "failed to append to properties: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* End properties array. */
+	r = sd_bus_message_close_container(m);
+	if (r < 0) {
+		xasprintf(cause, "failed to end properties array: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* aux is currently unused and should be passed an empty array. */
+	r = sd_bus_message_append(m, "a(sa(sv))", 0);
+	if (r < 0) {
+		xasprintf(cause, "failed to append to bus message: %s",
+		    strerror(-r));
+		goto finish;
+	}
+
+	/* Call the method with a timeout of 1 second = 1e6 us. */
+	r = sd_bus_call(bus, m, 1000000, &error, &reply);
+	if (r < 0) {
+		if (error.message != NULL) {
+			/* We have a specific error message from sd-bus. */
+			xasprintf(cause, "StartTransientUnit call failed: %s",
+			    error.message);
+		} else {
+			xasprintf(cause, "StartTransientUnit call failed: %s",
+			    strerror(-r));
+		}
+		goto finish;
+	}
+
+finish:
+	sd_bus_error_free(&error);
+	sd_bus_message_unref(m);
+	sd_bus_message_unref(reply);
+	sd_bus_unref(bus);
+
+	return (r);
+}
diff --git a/configure.ac b/configure.ac
index 8e846042..4b9d75b3 100644
--- a/configure.ac
+++ b/configure.ac
@@ -420,6 +420,21 @@ if test x"$enable_systemd" = xyes; then
 	fi
 fi
 AM_CONDITIONAL(HAVE_SYSTEMD, [test "x$found_systemd" = xyes])
+AC_ARG_ENABLE(
+	cgroups,
+	AS_HELP_STRING(--disable-cgroups, disable adding panes to new cgroups with systemd)
+)
+if test "x$enable_cgroups" = x; then
+	# Default to the same as $enable_systemd.
+	enable_cgroups=$enable_systemd
+fi
+if test "x$enable_cgroups" = xyes; then
+	if test "x$found_systemd" = xyes; then
+		AC_DEFINE(ENABLE_CGROUPS)
+	else
+		AC_MSG_ERROR("cgroups requires systemd to be enabled")
+	fi
+fi
 
 # Check for b64_ntop. If we have b64_ntop, we assume b64_pton as well.
 AC_MSG_CHECKING(for b64_ntop)
diff --git a/spawn.c b/spawn.c
index e63133fe..10604028 100644
--- a/spawn.c
+++ b/spawn.c
@@ -380,8 +380,20 @@ spawn_pane(struct spawn_context *sc, char **cause)
 	}
 
 	/* In the parent process, everything is done now. */
-	if (new_wp->pid != 0)
+	if (new_wp->pid != 0) {
+#if defined(HAVE_SYSTEMD) && defined(ENABLE_CGROUPS)
+		/*
+		 * Move the child process into a new cgroup for systemd-oomd
+		 * isolation.
+		 */
+		if (systemd_move_pid_to_new_cgroup(new_wp->pid, cause) < 0) {
+			log_debug("%s: moving pane to new cgroup failed: %s",
+			    __func__, *cause);
+			free (*cause);
+		}
+#endif
 		goto complete;
+	}
 
 	/*
 	 * Child process. Change to the working directory or home if that