]> git.ipfire.org Git - thirdparty/systemd.git/blobdiff - src/core/socket.c
Merge pull request #3202 from poettering/socket-fixes
[thirdparty/systemd.git] / src / core / socket.c
index 42260d872921a9d333dca31d335a7d7ff9572268..f6204d04bfd80f52935c2a9fa333ffae617e9ac5 100644 (file)
@@ -28,7 +28,6 @@
 #include <unistd.h>
 #include <linux/sctp.h>
 
-#include "sd-event.h"
 #include "alloc-util.h"
 #include "bus-error.h"
 #include "bus-util.h"
@@ -38,6 +37,7 @@
 #include "exit-status.h"
 #include "fd-util.h"
 #include "formats-util.h"
+#include "io-util.h"
 #include "label.h"
 #include "log.h"
 #include "missing.h"
@@ -100,7 +100,8 @@ static void socket_init(Unit *u) {
 
         s->control_command_id = _SOCKET_EXEC_COMMAND_INVALID;
 
-        RATELIMIT_INIT(s->trigger_limit, 5*USEC_PER_SEC, 2500);
+        s->trigger_limit.interval = USEC_INFINITY;
+        s->trigger_limit.burst = (unsigned) -1;
 }
 
 static void socket_unwatch_control_pid(Socket *s) {
@@ -229,7 +230,6 @@ int socket_instantiate_service(Socket *s) {
         if (r < 0)
                 return r;
 
-        u->no_gc = true;
         unit_ref_set(&s->service, u);
 
         return unit_add_two_dependencies(UNIT(s), UNIT_BEFORE, UNIT_TRIGGERS, u, false);
@@ -329,6 +329,25 @@ static int socket_add_extras(Socket *s) {
 
         assert(s);
 
+        /* Pick defaults for the trigger limit, if nothing was explicitly configured. We pick a relatively high limit
+         * in Accept=yes mode, and a lower limit for Accept=no. Reason: in Accept=yes mode we are invoking accept()
+         * ourselves before the trigger limit can hit, thus incoming connections are taken off the socket queue quickly
+         * and reliably. This is different for Accept=no, where the spawned service has to take the incoming traffic
+         * off the queues, which it might not necessarily do. Moreover, while Accept=no services are supposed to
+         * process whatever is queued in one go, and thus should normally never have to be started frequently. This is
+         * different for Accept=yes where each connection is processed by a new service instance, and thus frequent
+         * service starts are typical. */
+
+        if (s->trigger_limit.interval == USEC_INFINITY)
+                s->trigger_limit.interval = 2 * USEC_PER_SEC;
+
+        if (s->trigger_limit.burst == (unsigned) -1) {
+                if (s->accept)
+                        s->trigger_limit.burst = 200;
+                else
+                        s->trigger_limit.burst = 20;
+        }
+
         if (have_non_accept_socket(s)) {
 
                 if (!UNIT_DEREF(s->service)) {
@@ -621,8 +640,8 @@ static void socket_dump(Unit *u, FILE *f, const char *prefix) {
 
         if (!isempty(s->user) || !isempty(s->group))
                 fprintf(f,
-                        "%sOwnerUser: %s\n"
-                        "%sOwnerGroup: %s\n",
+                        "%sSocketUser: %s\n"
+                        "%sSocketGroup: %s\n",
                         prefix, strna(s->user),
                         prefix, strna(s->group));
 
@@ -671,6 +690,12 @@ static void socket_dump(Unit *u, FILE *f, const char *prefix) {
                         fprintf(f, "%sListenFIFO: %s\n", prefix, p->path);
         }
 
+        fprintf(f,
+                "%sTriggerLimitIntervalSec: %s\n"
+                "%sTriggerLimitBurst: %u\n",
+                prefix, format_timespan(time_string, FORMAT_TIMESPAN_MAX, s->trigger_limit.interval, USEC_PER_SEC),
+                prefix, s->trigger_limit.burst);
+
         exec_context_dump(&s->exec_context, f, prefix);
         kill_context_dump(&s->kill_context, f, prefix);
 
@@ -794,47 +819,45 @@ static void socket_close_fds(Socket *s) {
         assert(s);
 
         LIST_FOREACH(port, p, s->ports) {
+                bool was_open;
 
-                p->event_source = sd_event_source_unref(p->event_source);
-
-                if (p->fd < 0)
-                        continue;
+                was_open = p->fd >= 0;
 
+                p->event_source = sd_event_source_unref(p->event_source);
                 p->fd = safe_close(p->fd);
                 socket_cleanup_fd_list(p);
 
-                /* One little note: we should normally not delete any
-                 * sockets in the file system here! After all some
-                 * other process we spawned might still have a
-                 * reference of this fd and wants to continue to use
-                 * it. Therefore we delete sockets in the file system
-                 * before we create a new one, not after we stopped
-                 * using one! */
+                /* One little note: we should normally not delete any sockets in the file system here! After all some
+                 * other process we spawned might still have a reference of this fd and wants to continue to use
+                 * it. Therefore we normally delete sockets in the file system before we create a new one, not after we
+                 * stopped using one! That all said, if the user explicitly requested this, we'll delete them here
+                 * anyway, but only then. */
 
-                if (s->remove_on_stop) {
-                        switch (p->type) {
+                if (!was_open || !s->remove_on_stop)
+                        continue;
 
-                        case SOCKET_FIFO:
-                                unlink(p->path);
-                                break;
+                switch (p->type) {
 
-                        case SOCKET_MQUEUE:
-                                mq_unlink(p->path);
-                                break;
+                case SOCKET_FIFO:
+                        (void) unlink(p->path);
+                        break;
 
-                        case SOCKET_SOCKET:
-                                socket_address_unlink(&p->address);
-                                break;
+                case SOCKET_MQUEUE:
+                        (void) mq_unlink(p->path);
+                        break;
 
-                        default:
-                                break;
-                        }
+                case SOCKET_SOCKET:
+                        (void) socket_address_unlink(&p->address);
+                        break;
+
+                default:
+                        break;
                 }
         }
 
         if (s->remove_on_stop)
                 STRV_FOREACH(i, s->symlinks)
-                        unlink(*i);
+                        (void) unlink(*i);
 }
 
 static void socket_apply_socket_options(Socket *s, int fd) {
@@ -1224,6 +1247,45 @@ fail:
         return r;
 }
 
+static int socket_determine_selinux_label(Socket *s, char **ret) {
+        ExecCommand *c;
+        int r;
+
+        assert(s);
+        assert(ret);
+
+        if (s->selinux_context_from_net) {
+                /* If this is requested, get label from the network label */
+
+                r = mac_selinux_get_our_label(ret);
+                if (r == -EOPNOTSUPP)
+                        goto no_label;
+
+        } else {
+                /* Otherwise, get it from the executable we are about to start */
+                r = socket_instantiate_service(s);
+                if (r < 0)
+                        return r;
+
+                if (!UNIT_ISSET(s->service))
+                        goto no_label;
+
+                c = SERVICE(UNIT_DEREF(s->service))->exec_command[SERVICE_EXEC_START];
+                if (!c)
+                        goto no_label;
+
+                r = mac_selinux_get_create_label_from_exe(c->path, ret);
+                if (r == -EPERM || r == -EOPNOTSUPP)
+                        goto no_label;
+        }
+
+        return r;
+
+no_label:
+        *ret = NULL;
+        return 0;
+}
+
 static int socket_open_fds(Socket *s) {
         _cleanup_(mac_selinux_freep) char *label = NULL;
         bool know_label = false;
@@ -1242,46 +1304,28 @@ static int socket_open_fds(Socket *s) {
                 case SOCKET_SOCKET:
 
                         if (!know_label) {
-                                /* Figure out label, if we don't it know
-                                 * yet. We do it once, for the first
-                                 * socket where we need this and
-                                 * remember it for the rest. */
-
-                                if (s->selinux_context_from_net) {
-                                        /* Get it from the network label */
-
-                                        r = mac_selinux_get_our_label(&label);
-                                        if (r < 0 && r != -EOPNOTSUPP)
-                                                goto rollback;
+                                /* Figure out label, if we don't it know yet. We do it once, for the first socket where
+                                 * we need this and remember it for the rest. */
 
-                                } else {
-                                        /* Get it from the executable we are about to start */
-
-                                        r = socket_instantiate_service(s);
-                                        if (r < 0)
-                                                goto rollback;
-
-                                        if (UNIT_ISSET(s->service) &&
-                                            SERVICE(UNIT_DEREF(s->service))->exec_command[SERVICE_EXEC_START]) {
-                                                r = mac_selinux_get_create_label_from_exe(SERVICE(UNIT_DEREF(s->service))->exec_command[SERVICE_EXEC_START]->path, &label);
-                                                if (r < 0 && r != -EPERM && r != -EOPNOTSUPP)
-                                                        goto rollback;
-                                        }
-                                }
+                                r = socket_determine_selinux_label(s, &label);
+                                if (r < 0)
+                                        goto rollback;
 
                                 know_label = true;
                         }
 
                         /* Apply the socket protocol */
-                        switch(p->address.type) {
+                        switch (p->address.type) {
+
                         case SOCK_STREAM:
                         case SOCK_SEQPACKET:
-                                if (p->socket->socket_protocol == IPPROTO_SCTP)
-                                        p->address.protocol = p->socket->socket_protocol;
+                                if (s->socket_protocol == IPPROTO_SCTP)
+                                        p->address.protocol = s->socket_protocol;
                                 break;
+
                         case SOCK_DGRAM:
-                                if (p->socket->socket_protocol == IPPROTO_UDPLITE)
-                                        p->address.protocol = p->socket->socket_protocol;
+                                if (s->socket_protocol == IPPROTO_UDPLITE)
+                                        p->address.protocol = s->socket_protocol;
                                 break;
                         }
 
@@ -1342,8 +1386,7 @@ static int socket_open_fds(Socket *s) {
                         }
                         break;
 
-                case SOCKET_USB_FUNCTION:
-                {
+                case SOCKET_USB_FUNCTION: {
                         _cleanup_free_ char *ep = NULL;
 
                         ep = path_make_absolute("ep0", p->path);
@@ -1426,6 +1469,34 @@ fail:
         return r;
 }
 
+enum {
+        SOCKET_OPEN_NONE,
+        SOCKET_OPEN_SOME,
+        SOCKET_OPEN_ALL,
+};
+
+static int socket_check_open(Socket *s) {
+        bool have_open = false, have_closed = false;
+        SocketPort *p;
+
+        assert(s);
+
+        LIST_FOREACH(port, p, s->ports) {
+                if (p->fd < 0)
+                        have_closed = true;
+                else
+                        have_open = true;
+
+                if (have_open && have_closed)
+                        return SOCKET_OPEN_SOME;
+        }
+
+        if (have_open)
+                return SOCKET_OPEN_ALL;
+
+        return SOCKET_OPEN_NONE;
+}
+
 static void socket_set_state(Socket *s, SocketState state) {
         SocketState old_state;
         assert(s);
@@ -1505,14 +1576,24 @@ static int socket_coldplug(Unit *u) {
                    SOCKET_START_CHOWN,
                    SOCKET_START_POST,
                    SOCKET_LISTENING,
-                   SOCKET_RUNNING,
-                   SOCKET_STOP_PRE,
-                   SOCKET_STOP_PRE_SIGTERM,
-                   SOCKET_STOP_PRE_SIGKILL)) {
-
-                r = socket_open_fds(s);
-                if (r < 0)
-                        return r;
+                   SOCKET_RUNNING)) {
+
+                /* Originally, we used to simply reopen all sockets here that we didn't have file descriptors
+                 * for. However, this is problematic, as we won't traverse throught the SOCKET_START_CHOWN state for
+                 * them, and thus the UID/GID wouldn't be right. Hence, instead simply check if we have all fds open,
+                 * and if there's a mismatch, warn loudly. */
+
+                r = socket_check_open(s);
+                if (r == SOCKET_OPEN_NONE)
+                        log_unit_warning(UNIT(s),
+                                         "Socket unit configuration has changed while unit has been running, "
+                                         "no open socket file descriptor left. "
+                                         "The socket unit is not functional until restarted.");
+                else if (r == SOCKET_OPEN_SOME)
+                        log_unit_warning(UNIT(s),
+                                         "Socket unit configuration has changed while unit has been running, "
+                                         "and some socket file descriptors have not been opened yet. "
+                                         "The socket unit is not fully functional until restarted.");
         }
 
         if (s->deserialized_state == SOCKET_LISTENING) {
@@ -1885,6 +1966,21 @@ fail:
         socket_enter_dead(s, SOCKET_FAILURE_RESOURCES);
 }
 
+static void flush_ports(Socket *s) {
+        SocketPort *p;
+
+        /* Flush all incoming traffic, regardless if actual bytes or new connections, so that this socket isn't busy
+         * anymore */
+
+        LIST_FOREACH(port, p, s->ports) {
+                if (p->fd < 0)
+                        continue;
+
+                (void) flush_accept(p->fd);
+                (void) flush_fd(p->fd);
+        }
+}
+
 static void socket_enter_running(Socket *s, int cfd) {
         _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
         int r;
@@ -1894,31 +1990,15 @@ static void socket_enter_running(Socket *s, int cfd) {
 
         assert(s);
 
-        /* We don't take connections anymore if we are supposed to
-         * shut down anyway */
+        /* We don't take connections anymore if we are supposed to shut down anyway */
         if (unit_stop_pending(UNIT(s))) {
 
                 log_unit_debug(UNIT(s), "Suppressing connection request since unit stop is scheduled.");
 
                 if (cfd >= 0)
                         cfd = safe_close(cfd);
-                else  {
-                        /* Flush all sockets by closing and reopening them */
-                        socket_close_fds(s);
-
-                        r = socket_open_fds(s);
-                        if (r < 0) {
-                                log_unit_warning_errno(UNIT(s), r, "Failed to listen on sockets: %m");
-                                socket_enter_stop_pre(s, SOCKET_FAILURE_RESOURCES);
-                                return;
-                        }
-
-                        r = socket_watch_fds(s);
-                        if (r < 0) {
-                                log_unit_warning_errno(UNIT(s), r, "Failed to watch sockets: %m");
-                                socket_enter_stop_pre(s, SOCKET_FAILURE_RESOURCES);
-                        }
-                }
+                else
+                        flush_ports(s);
 
                 return;
         }
@@ -1996,10 +2076,8 @@ static void socket_enter_running(Socket *s, int cfd) {
 
                 service = SERVICE(UNIT_DEREF(s->service));
                 unit_ref_unset(&s->service);
-                s->n_accepted++;
-
-                UNIT(service)->no_gc = false;
 
+                s->n_accepted++;
                 unit_choose_id(UNIT(service), name);
 
                 r = service_set_socket_fd(service, cfd, s, s->selinux_context_from_net);
@@ -2010,8 +2088,12 @@ static void socket_enter_running(Socket *s, int cfd) {
                 s->n_connections++;
 
                 r = manager_add_job(UNIT(s)->manager, JOB_START, UNIT(service), JOB_REPLACE, &error, NULL);
-                if (r < 0)
+                if (r < 0) {
+                        /* We failed to activate the new service, but it still exists. Let's make sure the service
+                         * closes and forgets the connection fd again, immediately. */
+                        service_close_socket_fd(service);
                         goto fail;
+                }
 
                 /* Notify clients about changed counters */
                 unit_add_to_dbus_queue(UNIT(s));
@@ -2058,6 +2140,7 @@ fail:
 
 static int socket_start(Unit *u) {
         Socket *s = SOCKET(u);
+        int r;
 
         assert(s);
 
@@ -2102,6 +2185,12 @@ static int socket_start(Unit *u) {
 
         assert(s->state == SOCKET_DEAD || s->state == SOCKET_FAILED);
 
+        r = unit_start_limit_test(u);
+        if (r < 0) {
+                socket_enter_dead(s, SOCKET_FAILURE_START_LIMIT_HIT);
+                return r;
+        }
+
         s->result = SOCKET_SUCCESS;
         s->reset_cpu_usage = true;
 
@@ -2736,17 +2825,26 @@ static void socket_trigger_notify(Unit *u, Unit *other) {
         assert(u);
         assert(other);
 
-        /* Don't propagate state changes from the service if we are
-           already down or accepting connections */
-        if (!IN_SET(s->state, SOCKET_RUNNING, SOCKET_LISTENING) || s->accept)
+        /* Filter out invocations with bogus state */
+        if (other->load_state != UNIT_LOADED || other->type != UNIT_SERVICE)
                 return;
 
+        /* Don't propagate state changes from the service if we are already down */
+        if (!IN_SET(s->state, SOCKET_RUNNING, SOCKET_LISTENING))
+                return;
+
+        /* We don't care for the service state if we are in Accept=yes mode */
+        if (s->accept)
+                return;
+
+        /* Propagate start limit hit state */
         if (other->start_limit_hit) {
                 socket_enter_stop_pre(s, SOCKET_FAILURE_SERVICE_START_LIMIT_HIT);
                 return;
         }
 
-        if (other->load_state != UNIT_LOADED || other->type != UNIT_SERVICE)
+        /* Don't propagate anything if there's still a job queued */
+        if (other->job)
                 return;
 
         if (IN_SET(SERVICE(other)->state,
@@ -2819,6 +2917,7 @@ static const char* const socket_result_table[_SOCKET_RESULT_MAX] = {
         [SOCKET_FAILURE_EXIT_CODE] = "exit-code",
         [SOCKET_FAILURE_SIGNAL] = "signal",
         [SOCKET_FAILURE_CORE_DUMP] = "core-dump",
+        [SOCKET_FAILURE_START_LIMIT_HIT] = "start-limit-hit",
         [SOCKET_FAILURE_TRIGGER_LIMIT_HIT] = "trigger-limit-hit",
         [SOCKET_FAILURE_SERVICE_START_LIMIT_HIT] = "service-start-limit-hit"
 };