added TCP KeepAlive options to configuration

config/agent/agent.conf

@@ -42,6 +42,20 @@
 #LogIsQuiet=false
 
 # Enables or disables extended, reliable error message passing for the peer connection with the
-# controller. For example, if set to true, the peer connection will be dropped instantly on
-# Host unreachable errors.
+# controller by setting the IP_RECVERR socket option. For example, if set to true, the peer connection
+# will be dropped instantly on Host Unreachable errors.
 #IPReceiveErrors=true
+
+#
+# Number of seconds the TCP connection with the controller needs to be idle before keepalive packets are sent.
+# Value is set to socket option TCP_KEEPIDLE.
+#TCPKeepAliveTime=1
+
+#
+# Number of seconds between each keepalive packet. Value is set to socket option TCP_KEEPINTVL.
+#TCPKeepAliveInterval=1
+
+#
+# Number of keepalive packets without ACK from the controller till the connection is dropped. Value is
+# set to socket option TCP_KEEPCNT.
+#TCPKeepAliveCount=6

config/controller/controller.conf

@@ -28,7 +28,21 @@
 # If this flag is set to true, no logs are written by bluechi.
 #LogIsQuiet=false
 
-# Enables or disables extended, reliable error message passing for the peer connection with
-# the agent. For example, if set to true, the peer connection will be dropped instantly on
-# Host unreachable errors.
+# Enables or disables extended, reliable error message passing for the peer connection with the
+# agent by setting the IP_RECVERR socket option. For example, if set to true, the peer connection
+# will be dropped instantly on Host Unreachable errors.
 #IPReceiveErrors=true
+
+#
+# Number of seconds the TCP connection with the agent needs to be idle before keepalive packets are sent.
+# Value is set to socket option TCP_KEEPIDLE.
+#TCPKeepAliveTime=1
+
+#
+# Number of seconds between each keepalive packet. Value is set to socket option TCP_KEEPINTVL.
+#TCPKeepAliveInterval=1
+
+#
+# Number of keepalive packets without ACK from the agent till the peer connection is dropped. Value is
+# set to socket option TCP_KEEPCNT.
+#TCPKeepAliveCount=6

doc/man/bluechi-agent.conf.5.md

@@ -72,7 +72,27 @@ If this flag is set to `true`, it enables extended, reliable error message passi
 the peer connection with the controller. This results in BlueChi receiving errors such as
 host unreachable ICMP packets instantly and possibly dropping the connection. This is
 useful to detect disconnects faster, but should be used with care as this might cause
-unnecessary disconnects in less robut networks. Default: true.
+unnecessary disconnects in less robut networks.
+Default: true.
+
+#### **TCPKeepAliveTime** (long)
+
+The number of seconds the TCP connection of the agent with the controller needs to be idle before
+keepalive packets are sent. When `TCPKeepAliveTime` is set to 0, the system default will be used.
+Default: 1s.
+
+#### **TCPKeepAliveInterval** (long)
+
+The number of seconds between each keepalive packet. When `TCPKeepAliveInterval` is set to 0,
+the system default will be used.
+Default: 1s.
+
+#### **TCPKeepAliveCount** (long)
+
+The number of keepalive packets without ACK from the controller till the connection is
+dropped. When `TCPKeepAliveCount` is set to 0, the system default will be used.
+Default: 6.
+
 
 ## Example
 

doc/man/bluechi-controller.conf.5.md

@@ -58,8 +58,27 @@ If this flag is set to `true`, it enables extended, reliable error message passi
 the peer connection with all agents. This results in BlueChi receiving errors such as
 host unreachable ICMP packets instantly and possibly dropping the connection. This is
 useful to detect disconnects faster, but should be used with care as this might cause
-unnecessary disconnects in less robut networks. Default: true.
+unnecessary disconnects in less robut networks.
+Default: true.
 
+#### **TCPKeepAliveTime** (long)
+
+The number of seconds the individual TCP connection with an agent needs to be idle
+before keepalive packets are sent. When `TCPKeepAliveTime` is set to 0, the system
+default will be used.
+Default: 1s.
+
+#### **TCPKeepAliveInterval** (long)
+
+The number of seconds between each keepalive packet. When `TCPKeepAliveInterval` is set to 0,
+the system default will be used.
+Default: 1s.
+
+#### **TCPKeepAliveCount** (long)
+
+The number of keepalive packets without ACK from an agent till the individual connection is dropped.
+When `TCPKeepAliveCount` is set to 0, the system default will be used.
+Default: 6.
 
 ## Example
 

src/agent/agent.c

@@ -390,10 +390,24 @@ Agent *agent_new(void) {
                 return NULL;
         }
 
+        _cleanup_free_ SocketOptions *socket_opts = socket_options_new();
+        if (socket_opts == NULL) {
+                bc_log_error("Out of memory");
+                return NULL;
+        }
+
+        struct hashmap *unit_infos = hashmap_new(
+                        sizeof(AgentUnitInfo), 0, 0, 0, unit_info_hash, unit_info_compare, unit_info_clear, NULL);
+        if (unit_infos == NULL) {
+                return NULL;
+        }
+
         _cleanup_agent_ Agent *agent = malloc0(sizeof(Agent));
         agent->ref_count = 1;
         agent->event = steal_pointer(&event);
         agent->api_bus_service_name = steal_pointer(&service_name);
+        agent->peer_socket_options = steal_pointer(&socket_opts);
+        agent->unit_infos = unit_infos;
         LIST_HEAD_INIT(agent->outstanding_requests);
         LIST_HEAD_INIT(agent->tracked_jobs);
         LIST_HEAD_INIT(agent->proxy_services);
@@ -408,7 +422,6 @@ Agent *agent_new(void) {
         agent->connection_retry_count = 0;
         agent->wildcard_subscription_active = false;
         agent->metrics_enabled = false;
-        agent->ip_receive_errors = false;
         agent->disconnect_timestamp = 0;
 
         return steal_pointer(&agent);
@@ -624,7 +637,35 @@ bool agent_parse_config(Agent *agent, const char *configfile) {
                 }
         }
 
-        agent->ip_receive_errors = cfg_get_bool_value(agent->config, CFG_IP_RECEIVE_ERRORS);
+        /* Set socket options used for peer connections with the agents */
+        const char *keepidle = cfg_get_value(agent->config, CFG_TCP_KEEPALIVE_TIME);
+        if (keepidle) {
+                if (socket_options_set_tcp_keepidle(agent->peer_socket_options, keepidle) < 0) {
+                        bc_log_error("Failed to set TCP KEEPIDLE");
+                        return false;
+                }
+        }
+        const char *keepintvl = cfg_get_value(agent->config, CFG_TCP_KEEPALIVE_INTERVAL);
+        if (keepintvl) {
+                if (socket_options_set_tcp_keepintvl(agent->peer_socket_options, keepintvl) < 0) {
+                        bc_log_error("Failed to set TCP KEEPINTVL");
+                        return false;
+                }
+        }
+        const char *keepcnt = cfg_get_value(agent->config, CFG_TCP_KEEPALIVE_COUNT);
+        if (keepcnt) {
+                if (socket_options_set_tcp_keepcnt(agent->peer_socket_options, keepcnt) < 0) {
+                        bc_log_error("Failed to set TCP KEEPCNT");
+                        return false;
+                }
+        }
+        if (socket_options_set_ip_recverr(
+                            agent->peer_socket_options,
+                            cfg_get_bool_value(agent->config, CFG_IP_RECEIVE_ERRORS)) < 0) {
+                bc_log_error("Failed to set IP RECVERR");
+                return false;
+        }
+
 
         _cleanup_free_ const char *dumped_cfg = cfg_dump(agent->config);
         bc_log_debug_with_data("Final configuration used", "\n%s", dumped_cfg);
@@ -2394,23 +2435,9 @@ static bool agent_connect(Agent *agent) {
                 return false;
         }
 
-        int r = bus_socket_set_no_delay(agent->peer_dbus);
-        if (r < 0) {
-                bc_log_warn("Failed to set NO_DELAY on socket");
-        }
-
-        r = bus_socket_set_keepalive(agent->peer_dbus);
-        if (r < 0) {
-                bc_log_warn("Failed to set KEEPALIVE on socket");
-        }
-        if (agent->ip_receive_errors) {
-                r = bus_socket_enable_recv_err(agent->peer_dbus);
-                if (r < 0) {
-                        bc_log_warnf("Failed to enable receiving errors on socket: %s", strerror(-r));
-                }
-        }
+        bus_socket_set_options(agent->peer_dbus, agent->peer_socket_options);
 
-        r = sd_bus_add_object_vtable(
+        int r = sd_bus_add_object_vtable(
                         agent->peer_dbus,
                         NULL,
                         INTERNAL_AGENT_OBJECT_PATH,

src/agent/agent.h

@@ -9,6 +9,7 @@
 
 #include "libbluechi/common/cfg.h"
 #include "libbluechi/common/common.h"
+#include "libbluechi/socket.h"
 
 #include "types.h"
 
@@ -65,7 +66,7 @@ struct Agent {
 
         bool metrics_enabled;
 
-        bool ip_receive_errors;
+        SocketOptions *peer_socket_options;
 
         sd_event *event;
 

src/libbluechi/bus/utils.c

@@ -1,18 +1,11 @@
 /* SPDX-License-Identifier: LGPL-2.1-or-later */
 #include <errno.h>
 #include <netinet/in.h>
-#include <netinet/tcp.h>
 #include <sys/socket.h>
 
-#include "utils.h"
-
 #include "libbluechi/common/string-util.h"
 
-/* Number of seconds idle before sending keepalive packets */
-#define AGENT_KEEPALIVE_SOCKET_KEEPIDLE_SECS 1
-
-/* Number of seconds idle between each keepalive packet */
-#define AGENT_KEEPALIVE_SOCKET_KEEPINTVL_SECS 1
+#include "utils.h"
 
 int bus_parse_properties_foreach(sd_bus_message *m, bus_property_cb cb, void *userdata) {
         bool stop = false;
@@ -280,82 +273,13 @@ char *bus_path_escape(const char *s) {
         return r;
 }
 
-static bool is_socket_tcp(int fd) {
-        int type = 0;
-        socklen_t length = sizeof(int);
-
-        getsockopt(fd, SOL_SOCKET, SO_DOMAIN, &type, &length);
-
-        return type == AF_INET || type == AF_INET6;
-}
-
-int bus_socket_set_no_delay(sd_bus *bus) {
-        int fd = sd_bus_get_fd(bus);
-        if (fd < 0) {
-                return fd;
-        }
-
-        if (!is_socket_tcp(fd)) {
-                return 0;
-        }
-
-        int flag = 1;
-        int r = setsockopt(fd, SOL_TCP, TCP_NODELAY, (char *) &flag, sizeof(int));
-        if (r < 0) {
-                return -errno;
-        }
-
-        return 0;
-}
-
-int bus_socket_set_keepalive(sd_bus *bus) {
+int bus_socket_set_options(sd_bus *bus, SocketOptions *opts) {
         int fd = sd_bus_get_fd(bus);
         if (fd < 0) {
                 return fd;
         }
 
-        if (!is_socket_tcp(fd)) {
-                return 0;
-        }
-
-        int flag = 1;
-        int r = setsockopt(fd, SOL_SOCKET, SO_KEEPALIVE, (char *) &flag, sizeof(int));
-        if (r < 0) {
-                return -errno;
-        }
-
-        int keepidle = AGENT_KEEPALIVE_SOCKET_KEEPIDLE_SECS;
-        r = setsockopt(fd, IPPROTO_TCP, TCP_KEEPIDLE, &keepidle, sizeof(int));
-        if (r < 0) {
-                return -errno;
-        }
-
-        int keepintvl = AGENT_KEEPALIVE_SOCKET_KEEPINTVL_SECS;
-        r = setsockopt(fd, IPPROTO_TCP, TCP_KEEPINTVL, &keepintvl, sizeof(int));
-        if (r < 0) {
-                return -errno;
-        }
-
-        return 0;
-}
-
-int bus_socket_enable_recv_err(sd_bus *bus) {
-        int fd = sd_bus_get_fd(bus);
-        if (fd < 0) {
-                return fd;
-        }
-
-        if (!is_socket_tcp(fd)) {
-                return -EINVAL;
-        }
-
-        int flag = 1;
-        int r = setsockopt(fd, IPPROTO_IP, IP_RECVERR, &flag, sizeof(int));
-        if (r < 0) {
-                return -errno;
-        }
-
-        return 0;
+        return socket_set_options(fd, opts);
 }
 
 /*

src/libbluechi/bus/utils.h

@@ -5,6 +5,7 @@
 #include <systemd/sd-bus.h>
 
 #include "libbluechi/common/common.h"
+#include "libbluechi/socket.h"
 
 /* return < 0 for error, 0 to continue, 1 to stop, 2 to continue and skip (if value was not consumed) */
 typedef int (*bus_property_cb)(const char *key, const char *value_type, sd_bus_message *m, void *userdata);
@@ -40,9 +41,7 @@ void unit_unref(UnitInfo *unit);
 int bus_parse_unit_info(sd_bus_message *message, UnitInfo *u);
 int bus_parse_unit_on_node_info(sd_bus_message *message, UnitInfo *u);
 
-int bus_socket_set_no_delay(sd_bus *bus);
-int bus_socket_set_keepalive(sd_bus *bus);
-int bus_socket_enable_recv_err(sd_bus *bus);
+int bus_socket_set_options(sd_bus *bus, SocketOptions *opts);
 
 bool bus_id_is_valid(const char *name);
 

src/libbluechi/common/cfg.c

@@ -417,6 +417,19 @@ static int cfg_def_conf(struct config *config) {
                 return result;
         }
 
+        if ((result = cfg_set_value(config, CFG_TCP_KEEPALIVE_TIME, BC_DEFAULT_TCP_KEEPALIVE_TIME)) != 0) {
+                return result;
+        }
+
+        if ((result = cfg_set_value(config, CFG_TCP_KEEPALIVE_INTERVAL, BC_DEFAULT_TCP_KEEPALIVE_INTERVAL)) !=
+            0) {
+                return result;
+        }
+
+        if ((result = cfg_set_value(config, CFG_TCP_KEEPALIVE_COUNT, BC_DEFAULT_TCP_KEEPALIVE_COUNT)) != 0) {
+                return result;
+        }
+
         return 0;
 }
 
@@ -439,7 +452,8 @@ int cfg_agent_def_conf(struct config *config) {
                 return result;
         }
 
-        if ((result = cfg_set_value(config, CFG_HEARTBEAT_INTERVAL, AGENT_HEARTBEAT_INTERVAL_MSEC)) != 0) {
+        if ((result = cfg_set_value(config, CFG_HEARTBEAT_INTERVAL, AGENT_DEFAULT_HEARTBEAT_INTERVAL_MSEC)) !=
+            0) {
                 return result;
         }
 

src/libbluechi/common/cfg.h

@@ -18,6 +18,9 @@
 #define CFG_ALLOWED_NODE_NAMES "AllowedNodeNames"
 #define CFG_HEARTBEAT_INTERVAL "HeartbeatInterval"
 #define CFG_IP_RECEIVE_ERRORS "IPReceiveErrors"
+#define CFG_TCP_KEEPALIVE_TIME "TCPKeepAliveTime"
+#define CFG_TCP_KEEPALIVE_INTERVAL "TCPKeepAliveInterval"
+#define CFG_TCP_KEEPALIVE_COUNT "TCPKeepAliveCount"
 
 /*
  * Global section - this is used, when configuration options are specified in the configuration file