* fix(kubevip): reject out-of-range routing protocol values
Netlink carries the address and route protocol in a single byte, so a
configured value above 255 was silently truncated on the wire and never
matched again on readback. Reject it during config validation instead.
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* feat(wireguard): qualify service tunnel IDs by protocol
Sanitisation maps '-' onto the '_' separator, so "a-b/c" and "a/b-c"
shared one nftables chain, and TCP and UDP on the same port collided.
ServicePortIDs appends the protocol and, when sanitisation changed the
name or the ID grew too long, a hash of the raw name. It also returns
the previous port-only ID so existing chains can be migrated.
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* fix(arp): guard manager state behind a single mutex
Instances were kept in a sync.Map with a per-instance mutex for the
refcount, so lookup and refcount update were not atomic: concurrent
Insert and Remove could resurrect a deleted instance or drop a live one.
Hold one manager mutex across both, and buffer link subscriptions so a
netlink sender is never parked on an unread channel during shutdown.
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* fix(instance): roll back partially created instances
Instance creation added addresses, VLAN or macvlan links and DHCP
clients incrementally, so a failure part way through left the node
holding state nobody owned. Initialization now unwinds what it created,
and link cleanup only deletes attachments this instance created that no
remaining instance still uses.
Namespace-dependent tests now skip unless KUBE_VIP_REQUIRE_NETNS is set,
which CI sets on the privileged job so lost capabilities turn it red
instead of silently skipping.
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* test(e2e): give docker kill more time under parallel load
The ARP suite runs four kind clusters against one Docker daemon, so
acknowledging a leader kill regularly exceeded the 5s budget and failed
the IPv6 failover specs before any assertion ran.
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* fix: regression on preserveOnLeadershipLoss
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* fix: use the introduced wireguard service_id
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
* fix(services): reuse link attachment ownership on service delete
deleteService removed VLAN and macvlan links unconditionally, which tore
down interfaces kube-vip had only adopted and interfaces another Service
still used. Route the delete path through CleanupLinkAttachments and pass
the remaining instances so ownership is handed over instead.
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
---------
Signed-off-by: Marcel Fest <marcel.fest@telekom.de>
shouldAddServiceIP() checks BGPAttachIPToInterface on a per-service
config struct. NewInstance() builds that struct fresh for each Service,
copying over most fields from the global config, missing this one. As
a result the flag has no effect regardless of its value, and BGP-mode
Service VIPs are never bound to the interface.
Copy the field at both construction sites where it's built, and add
a regression test covering the propagation.
Signed-off-by: Justin Cichra <jrcichra@yahoo.com>
This PR adds an explicit configuration option `vip_skipdad`. This allows
for a proper behaviour in IPv6 networks with RT mode and ECMP. Without
this option only one node can ever hold the same IPv6 VIP what defeats
the purpose of ECMP.
While the ARP mode currently has a recovery mode for failed DAD, RT mode
misses it. It is also more feasible to use config option, as recovery
mode in RT mode would always introduce a ~5s blackholing (we know there
will be a conflict, so first attempt to add a VIP will fail and only the
next one will force IFA_F_NODAD).
For simplicitly we ignore IPv4/IPv6 checks, as kernel transparently
strips IFA_F_NODAD from v4 addresses without throwing errors.
For systems with IPv4 duplicate detection (called ACD and used e.g. in
RHEL10's NetworkManager) this code has no effect as it's only userspace
action and addresses we add via netlink bypass any check.
First discovered on a dual-stack cluster with two routing-table-mode
advertisers: the ingress VIP answered or timed out depending on which
ECMP path the flow hashed to; with NODAD both advertisers accept
traffic.
Signed-off-by: Mat Kowalski <mko@redhat.com>
When in arp mode, when preserveVipOnLeadershipLoss is true, when a node loses leadership (or can't keep it),
this will stop performing ARPs but will wait until a new leader is elected before deleting the arp.
when the flag is set to false (default) it continue to function like normal.
Signed-off-by: Brian Davis <slimm609@gmail.com>
Signed-off-by: Brian Davis <dbrian@vmware.com>
* Moved functions from mod pkg vip to pkg utils
* remove the dependency of the kubevip config pkg on the bgp pkg
* introduce BGPConfig and BGPPeer to kubevip package and migrate label from node labeler
* Use the new BGPConfig and BGPPeer type of the kubevip pkg
* Removed utils functions from vip pkg and refactor to call utils instead
* migrate from pkg vip to utils for generic IP funcs
* migrate common annotations and labels into a central place + support for hostnames
* if any change happens we want to reconcile it here
* fix e2e tests which relied also on the vip package instead of utils
* Better debug logging
* Added info if it is egress
* added missing condition to dhcp specific actions
* Disable the service before cancelling the context
* lets try auto,auto instead of /32 and nothing
* Let's reset svcCtx to be nil to ensure a new context when the previous was garbage collected
* Added a space for project wide constants
---------
Signed-off-by: Cellebyte <marcel.fest@live.de>
