Merge pull request #34736 from yuwata/network-mtu

network: wait for IPv6 MTU being synced to link MTU

6 files changed, 112 insertions, 24 deletions

diff --git a/src/network/networkd-link.c b/src/network/networkd-link.c
index 234e08680b..59240bbc36 100644
--- a/src/network/networkd-link.c
+++ b/src/network/networkd-link.c
@@ -288,6 +288,7 @@ static Link *link_free(Link *link) {
         network_unref(link->network);
 
         sd_event_source_disable_unref(link->carrier_lost_timer);
+        sd_event_source_disable_unref(link->ipv6_mtu_wait_synced_event_source);
 
         return mfree(link);
 }
@@ -1869,8 +1870,6 @@ static int link_carrier_lost(Link *link) {
 }
 
 static int link_admin_state_up(Link *link) {
-        int r;
-
         assert(link);
 
         /* This is called every time an interface admin state changes to up;
@@ -1886,9 +1885,7 @@ static int link_admin_state_up(Link *link) {
 
         /* We set the ipv6 mtu after the device mtu, but the kernel resets
          * ipv6 mtu on NETDEV_UP, so we need to reset it. */
-        r = link_set_ipv6_mtu(link, LOG_INFO);
-        if (r < 0)
-                log_link_warning_errno(link, r, "Cannot set IPv6 MTU, ignoring: %m");
+        (void) link_set_ipv6_mtu(link, LOG_INFO);
 
         return 0;
 }
@@ -2439,12 +2436,9 @@ static int link_update_mtu(Link *link, sd_netlink_message *message) {
 
         link->mtu = mtu;
 
-        if (IN_SET(link->state, LINK_STATE_CONFIGURING, LINK_STATE_CONFIGURED)) {
+        if (IN_SET(link->state, LINK_STATE_CONFIGURING, LINK_STATE_CONFIGURED))
                 /* The kernel resets IPv6 MTU after changing device MTU. So, we need to re-set IPv6 MTU again. */
-                r = link_set_ipv6_mtu(link, LOG_INFO);
-                if (r < 0)
-                        log_link_warning_errno(link, r, "Failed to set IPv6 MTU, ignoring: %m");
-        }
+                (void) link_set_ipv6_mtu_async(link);
 
         if (link->dhcp_client) {
                 r = sd_dhcp_client_set_mtu(link->dhcp_client, link->mtu);
diff --git a/src/network/networkd-link.h b/src/network/networkd-link.h
index 27bc299bc5..e86839af8e 100644
--- a/src/network/networkd-link.h
+++ b/src/network/networkd-link.h
@@ -74,6 +74,9 @@ typedef struct Link {
         sd_device *dev;
         char *driver;
 
+        sd_event_source *ipv6_mtu_wait_synced_event_source;
+        unsigned ipv6_mtu_wait_trial_count;
+
         /* bridge vlan */
         uint16_t bridge_vlan_pvid;
         bool bridge_vlan_pvid_is_untagged;
diff --git a/src/network/networkd-ndisc.c b/src/network/networkd-ndisc.c
index 529311f4a4..fc06e5c38b 100644
--- a/src/network/networkd-ndisc.c
+++ b/src/network/networkd-ndisc.c
@@ -1083,9 +1083,7 @@ static int ndisc_router_process_mtu(Link *link, sd_ndisc_router *rt) {
 
         link->ndisc_mtu = mtu;
 
-        r = link_set_ipv6_mtu(link, LOG_DEBUG);
-        if (r < 0)
-                log_link_warning_errno(link, r, "Failed to apply IPv6 MTU (%"PRIu32"), ignoring: %m", mtu);
+        (void) link_set_ipv6_mtu(link, LOG_DEBUG);
 
         return 0;
 }
diff --git a/src/network/networkd-route-util.c b/src/network/networkd-route-util.c
index 9bb1e0f707..dc9663e24b 100644
--- a/src/network/networkd-route-util.c
+++ b/src/network/networkd-route-util.c
@@ -16,23 +16,26 @@
 #include "strv.h"
 #include "sysctl-util.h"
 
-#define ROUTES_DEFAULT_MAX_PER_FAMILY 4096U
+#define ROUTES_DEFAULT_MAX_PER_FAMILY 4096
 
 unsigned routes_max(void) {
         static thread_local unsigned cached = 0;
-        _cleanup_free_ char *s4 = NULL, *s6 = NULL;
-        unsigned val4 = ROUTES_DEFAULT_MAX_PER_FAMILY, val6 = ROUTES_DEFAULT_MAX_PER_FAMILY;
+        int val4 = ROUTES_DEFAULT_MAX_PER_FAMILY, val6 = ROUTES_DEFAULT_MAX_PER_FAMILY;
 
         if (cached > 0)
                 return cached;
 
-        if (sysctl_read_ip_property(AF_INET, NULL, "route/max_size", &s4) >= 0)
-                if (safe_atou(s4, &val4) >= 0 && val4 == 2147483647U)
+        /* The kernel internally stores these maximum size in int. */
+
+        if (sysctl_read_ip_property_int(AF_INET, /* ifname = */ NULL, "route/max_size", &val4) >= 0)
+                if (val4 == INT_MAX)
                         /* This is the default "no limit" value in the kernel */
                         val4 = ROUTES_DEFAULT_MAX_PER_FAMILY;
 
-        if (sysctl_read_ip_property(AF_INET6, NULL, "route/max_size", &s6) >= 0)
-                (void) safe_atou(s6, &val6);
+        if (sysctl_read_ip_property_int(AF_INET6, /* ifname = */ NULL, "route/max_size", &val6) >= 0)
+                if (val6 == INT_MAX)
+                        /* This is the default "no limit" value in the kernel */
+                        val6 = ROUTES_DEFAULT_MAX_PER_FAMILY;
 
         cached = MAX(ROUTES_DEFAULT_MAX_PER_FAMILY, val4) +
                  MAX(ROUTES_DEFAULT_MAX_PER_FAMILY, val6);
diff --git a/src/network/networkd-sysctl.c b/src/network/networkd-sysctl.c
index 7330c2cc76..c0762b9b70 100644
--- a/src/network/networkd-sysctl.c
+++ b/src/network/networkd-sysctl.c
@@ -8,6 +8,7 @@
 
 #include "af-list.h"
 #include "cgroup-util.h"
+#include "event-util.h"
 #include "fd-util.h"
 #include "format-util.h"
 #include "missing_network.h"
@@ -503,6 +504,7 @@ static int link_set_ipv6_proxy_ndp(Link *link) {
 
 int link_set_ipv6_mtu(Link *link, int log_level) {
         uint32_t mtu = 0;
+        int r;
 
         assert(link);
         assert(link->manager);
@@ -510,6 +512,14 @@ int link_set_ipv6_mtu(Link *link, int log_level) {
         if (!link_is_configured_for_family(link, AF_INET6))
                 return 0;
 
+        if (!IN_SET(link->state, LINK_STATE_CONFIGURING, LINK_STATE_CONFIGURED))
+                return 0;
+
+        if (sd_event_source_get_enabled(link->ipv6_mtu_wait_synced_event_source, /* ret = */ NULL) > 0) {
+                log_link_debug(link, "Waiting for IPv6 MTU is synced to link MTU, delaying to set IPv6 MTU.");
+                return 0;
+        }
+
         assert(link->network);
 
         if (link->network->ndisc_use_mtu)
@@ -526,7 +536,88 @@ int link_set_ipv6_mtu(Link *link, int log_level) {
                 mtu = link->mtu;
         }
 
-        return sysctl_write_ip_property_uint32(AF_INET6, link->ifname, "mtu", mtu, manager_get_sysctl_shadow(link->manager));
+        r = sysctl_write_ip_property_uint32(AF_INET6, link->ifname, "mtu", mtu, manager_get_sysctl_shadow(link->manager));
+        if (r < 0)
+                return log_link_warning_errno(link, r, "Failed to set IPv6 MTU to %"PRIu32": %m", mtu);
+
+        return 0;
+}
+
+static int ipv6_mtu_wait_synced_handler(sd_event_source *s, uint64_t usec, void *userdata);
+
+static int link_set_ipv6_mtu_async_impl(Link *link) {
+        uint32_t current_mtu;
+        int r;
+
+        assert(link);
+
+        /* When the link MTU is updated, it seems that the kernel IPv6 MTU of the interface is asynchronously
+         * reset to the link MTU. Hence, we need to check if it is already reset, and wait for a while if not. */
+
+        if (++link->ipv6_mtu_wait_trial_count >= 10) {
+                log_link_debug(link, "Timed out waiting for IPv6 MTU being synced to link MTU, proceeding anyway.");
+                r = link_set_ipv6_mtu(link, LOG_INFO);
+                if (r < 0)
+                        return r;
+
+                return 1; /* done */
+        }
+
+        /* Check if IPv6 MTU is synced. */
+        r = sysctl_read_ip_property_uint32(AF_INET6, link->ifname, "mtu", &current_mtu);
+        if (r < 0)
+                return log_link_warning_errno(link, r, "Failed to read IPv6 MTU: %m");
+
+        if (current_mtu == link->mtu) {
+                /* Already synced. Update IPv6 MTU now. */
+                r = link_set_ipv6_mtu(link, LOG_INFO);
+                if (r < 0)
+                        return r;
+
+                return 1; /* done */
+        }
+
+        /* If not, set up a timer event source. */
+        r = event_reset_time_relative(
+                        link->manager->event, &link->ipv6_mtu_wait_synced_event_source,
+                        CLOCK_BOOTTIME, 100 * USEC_PER_MSEC, 0,
+                        ipv6_mtu_wait_synced_handler, link,
+                        /* priority = */ 0, "ipv6-mtu-wait-synced", /* force = */ true);
+        if (r < 0)
+                return log_link_warning_errno(link, r, "Failed to configure timer event source for waiting for IPv6 MTU being synced: %m");
+
+        /* Check again. */
+        r = sysctl_read_ip_property_uint32(AF_INET6, link->ifname, "mtu", &current_mtu);
+        if (r < 0)
+                return log_link_warning_errno(link, r, "Failed to read IPv6 MTU: %m");
+
+        if (current_mtu == link->mtu) {
+                /* Synced while setting up the timer event source. Disable it and update IPv6 MTU now. */
+                r = sd_event_source_set_enabled(link->ipv6_mtu_wait_synced_event_source, SD_EVENT_OFF);
+                if (r < 0)
+                        log_link_debug_errno(link, r, "Failed to disable timer event source for IPv6 MTU, ignoring: %m");
+
+                r = link_set_ipv6_mtu(link, LOG_INFO);
+                if (r < 0)
+                        return r;
+
+                return 1; /* done */
+        }
+
+        log_link_debug(link, "IPv6 MTU is not synced to the link MTU after it is changed. Waiting for a while.");
+        return 0; /* waiting */
+}
+
+static int ipv6_mtu_wait_synced_handler(sd_event_source *s, uint64_t usec, void *userdata) {
+        (void) link_set_ipv6_mtu_async_impl(ASSERT_PTR(userdata));
+        return 0;
+}
+
+int link_set_ipv6_mtu_async(Link *link) {
+        assert(link);
+
+        link->ipv6_mtu_wait_trial_count = 0;
+        return link_set_ipv6_mtu_async_impl(link);
 }
 
 static int link_set_ipv4_accept_local(Link *link) {
@@ -616,9 +707,7 @@ int link_set_sysctl(Link *link) {
         if (r < 0)
                 log_link_warning_errno(link, r, "Cannot set IPv6 proxy NDP, ignoring: %m");
 
-        r = link_set_ipv6_mtu(link, LOG_INFO);
-        if (r < 0)
-                log_link_warning_errno(link, r, "Cannot set IPv6 MTU, ignoring: %m");
+        (void) link_set_ipv6_mtu(link, LOG_INFO);
 
         r = link_set_ipv6ll_stable_secret(link);
         if (r < 0)
diff --git a/src/network/networkd-sysctl.h b/src/network/networkd-sysctl.h
index 446b835555..1c19fbd2d6 100644
--- a/src/network/networkd-sysctl.h
+++ b/src/network/networkd-sysctl.h
@@ -42,6 +42,7 @@ void manager_set_sysctl(Manager *manager);
 int link_get_ip_forwarding(Link *link, int family);
 int link_set_sysctl(Link *link);
 int link_set_ipv6_mtu(Link *link, int log_level);
+int link_set_ipv6_mtu_async(Link *link);
 
 const char* ipv6_privacy_extensions_to_string(IPv6PrivacyExtensions i) _const_;
 IPv6PrivacyExtensions ipv6_privacy_extensions_from_string(const char *s) _pure_;