Files
kube-vip/pkg/services/processor.go
Dan Finneran 332a23e543 Bump k8s api and endpointslices as default
Signed-off-by: Dan Finneran <dan@thebsdbox.co.uk>
2025-07-29 13:36:32 +00:00

345 lines
10 KiB
Go

package services
import (
"context"
"fmt"
log "log/slog"
"reflect"
"sync"
"github.com/kube-vip/kube-vip/pkg/arp"
"github.com/kube-vip/kube-vip/pkg/bgp"
"github.com/kube-vip/kube-vip/pkg/endpoints"
"github.com/kube-vip/kube-vip/pkg/endpoints/providers"
"github.com/kube-vip/kube-vip/pkg/instance"
"github.com/kube-vip/kube-vip/pkg/kubevip"
"github.com/kube-vip/kube-vip/pkg/networkinterface"
"github.com/kube-vip/kube-vip/pkg/servicecontext"
"github.com/kube-vip/kube-vip/pkg/vip"
"github.com/prometheus/client_golang/prometheus"
"github.com/vishvananda/netlink"
v1 "k8s.io/api/core/v1"
"k8s.io/apimachinery/pkg/types"
"k8s.io/apimachinery/pkg/watch"
"k8s.io/client-go/kubernetes"
)
type Processor struct {
config *kubevip.Config
lbClassFilter func(svc *v1.Service, config *kubevip.Config) bool
svcMap sync.Map
// Keeps track of all running instances
ServiceInstances []*instance.Instance
mutex sync.Mutex
bgpServer *bgp.Server
clientSet *kubernetes.Clientset
rwClientSet *kubernetes.Clientset
shutdownChan chan struct{}
// This is a prometheus counter used to count the number of events received
// from the service watcher
CountServiceWatchEvent *prometheus.CounterVec
intfMgr *networkinterface.Manager
arpMgr *arp.Manager
}
func NewServicesProcessor(config *kubevip.Config, bgpServer *bgp.Server,
clientSet *kubernetes.Clientset, rwClientSet *kubernetes.Clientset, shutdownChan chan struct{},
intfMgr *networkinterface.Manager, arpMgr *arp.Manager) *Processor {
lbClassFilterFunc := lbClassFilter
if config.LoadBalancerClassLegacyHandling {
lbClassFilterFunc = lbClassFilterLegacy
}
return &Processor{
config: config,
lbClassFilter: lbClassFilterFunc,
ServiceInstances: []*instance.Instance{},
bgpServer: bgpServer,
clientSet: clientSet,
rwClientSet: rwClientSet,
shutdownChan: shutdownChan,
CountServiceWatchEvent: prometheus.NewCounterVec(prometheus.CounterOpts{
Namespace: "kube_vip",
Subsystem: "manager",
Name: "all_services_events",
Help: "Count all events fired by the service watcher categorised by event type",
}, []string{"type"}),
intfMgr: intfMgr,
arpMgr: arpMgr,
}
}
func (p *Processor) AddOrModify(ctx context.Context, event watch.Event, serviceFunc func(context.Context, *v1.Service) error) (bool, error) {
// log.Debugf("Endpoints for service [%s] have been Created or modified", s.service.ServiceName)
svc, ok := event.Object.(*v1.Service)
if !ok {
return false, fmt.Errorf("unable to parse Kubernetes services from API watcher")
}
// We only care about LoadBalancer services
if svc.Spec.Type != v1.ServiceTypeLoadBalancer {
return true, nil
}
// Check if we ignore this service
if svc.Annotations["kube-vip.io/ignore"] == "true" {
log.Info("ignore annotation for kube-vip", "service name", svc.Name)
return true, nil
}
// Check the loadBalancer class
if p.lbClassFilter(svc, p.config) {
return true, nil
}
svcAddresses := instance.FetchServiceAddresses(svc)
// We only care about LoadBalancer services that have been allocated an address
if len(svcAddresses) <= 0 {
return true, nil
}
svcCtx, err := p.getServiceContext(svc.UID)
if err != nil {
return false, fmt.Errorf("failed to get service context: %w", err)
}
// The modified event should only be triggered if the service has been modified (i.e. moved somewhere else)
if event.Type == watch.Modified {
i := instance.FindServiceInstance(svc, p.ServiceInstances)
var originalService []string
shouldGarbageCollect := true
if i != nil {
originalService = instance.FetchServiceAddresses(i.ServiceSnapshot)
shouldGarbageCollect = !reflect.DeepEqual(originalService, svcAddresses)
}
if shouldGarbageCollect {
for _, addr := range svcAddresses {
// log.Debugf("(svcs) Retreiving local addresses, to ensure that this modified address doesn't exist: %s", addr)
f, err := vip.GarbageCollect(p.config.Interface, addr, p.intfMgr)
if err != nil {
log.Error("(svcs) cleaning existing address error", "err", err)
}
if f {
log.Warn("(svcs) already found existing config", "address", addr, "adapter", p.config.Interface)
}
}
}
// This service has been modified, but it was also active.
if svcCtx != nil && svcCtx.IsActive {
if i != nil {
originalService := instance.FetchServiceAddresses(i.ServiceSnapshot)
newService := instance.FetchServiceAddresses(svc)
if !reflect.DeepEqual(originalService, newService) {
// Calls the cancel function of the context
if svcCtx != nil {
log.Warn("(svcs) The load balancer has changed, cancelling original load balancer")
svcCtx.Cancel()
log.Warn("(svcs) waiting for load balancer to finish")
<-svcCtx.Ctx.Done()
}
err = p.deleteService(svc.UID)
if err != nil {
log.Error("(svc) unable to remove", "service", svc.UID)
}
p.svcMap.Delete(svc.UID)
}
// in theory this should never fail
}
}
}
// Architecture walkthrough: (Had to do this as this code path is making my head hurt)
// Is the service active (bool), if not then process this new service
// Does this service use an election per service?
//
if svcCtx == nil || svcCtx != nil && !svcCtx.IsActive {
log.Debug("(svcs) has been added/modified with addresses", "service name", svc.Name, "ip", instance.FetchServiceAddresses(svc))
if svcCtx == nil {
svcCtx = servicecontext.New(ctx)
p.svcMap.Store(svc.UID, svcCtx)
}
if p.config.EnableServicesElection || // Service Election
((p.config.EnableRoutingTable || p.config.EnableBGP) && // Routing table mode or BGP
(!p.config.EnableLeaderElection && !p.config.EnableServicesElection)) { // No leaderelection or services election
// If this load balancer Traffic Policy is "local"
if svc.Spec.ExternalTrafficPolicy == v1.ServiceExternalTrafficPolicyTypeLocal {
// Start an endpoint watcher if we're not watching it already
if !svcCtx.IsWatched {
// background the endpoint watcher
if (p.config.EnableRoutingTable || p.config.EnableBGP) && (!p.config.EnableLeaderElection && !p.config.EnableServicesElection) {
err = serviceFunc(svcCtx.Ctx, svc)
if err != nil {
log.Error(err.Error())
}
}
go func() {
if svc.Spec.ExternalTrafficPolicy == v1.ServiceExternalTrafficPolicyTypeLocal {
// Add Endpoint or EndpointSlices watcher
var provider providers.Provider
if p.config.EnableEndpoints {
provider = providers.NewEndpoints()
} else {
provider = providers.NewEndpointslices()
}
if err = p.watchEndpoint(svcCtx, p.config.NodeName, svc, provider); err != nil {
log.Error(err.Error())
}
}
}()
// We're now watching this service
svcCtx.IsWatched = true
}
} else if (p.config.EnableBGP || p.config.EnableRoutingTable) && (!p.config.EnableLeaderElection && !p.config.EnableServicesElection) {
err = serviceFunc(svcCtx.Ctx, svc)
if err != nil {
log.Error(err.Error())
}
go func() {
if svc.Spec.ExternalTrafficPolicy == v1.ServiceExternalTrafficPolicyTypeCluster {
// Add Endpoint watcher
var provider providers.Provider
if p.config.EnableEndpoints {
provider = providers.NewEndpoints()
} else {
provider = providers.NewEndpointslices()
}
if err = p.watchEndpoint(svcCtx, p.config.NodeName, svc, provider); err != nil {
log.Error(err.Error())
}
}
}()
// We're now watching this service
svcCtx.IsWatched = true
} else {
go func() {
for {
select {
case <-svcCtx.Ctx.Done():
log.Warn("(svcs) restartable service watcher ending", "uid", svc.UID)
return
default:
log.Info("(svcs) restartable service watcher starting", "uid", svc.UID)
err = serviceFunc(svcCtx.Ctx, svc)
if err != nil {
log.Error(err.Error())
}
}
}
}()
}
} else {
// Increment the waitGroup before the service Func is called (Done is completed in there)
err = serviceFunc(svcCtx.Ctx, svc)
if err != nil {
log.Error(err.Error())
}
}
svcCtx.IsActive = true
}
return false, nil
}
func (p *Processor) Delete(event watch.Event) (bool, error) {
svc, ok := event.Object.(*v1.Service)
if !ok {
return false, fmt.Errorf("unable to parse Kubernetes services from API watcher")
}
svcCtx, err := p.getServiceContext(svc.UID)
if err != nil {
return false, fmt.Errorf("(svcs) unable to get context: %w", err)
}
if svcCtx != nil && svcCtx.IsActive {
// We only care about LoadBalancer services
if svc.Spec.Type != v1.ServiceTypeLoadBalancer {
return true, nil
}
// We can ignore this service
if svc.Annotations["kube-vip.io/ignore"] == "true" {
log.Info("(svcs)ignore annotation for kube-vip", "service name", svc.Name)
return true, nil
}
// If no leader election is enabled, delete routes here
if !p.config.EnableLeaderElection && !p.config.EnableServicesElection &&
p.config.EnableRoutingTable && svcCtx.HasConfiguredNetworks() {
if errs := endpoints.ClearRoutes(svc, &p.ServiceInstances); len(errs) == 0 {
svcCtx.ConfiguredNetworks.Clear()
}
}
// If this is an active service then and additional leaderElection will handle stopping
err = p.deleteService(svc.UID)
if err != nil {
log.Error(err.Error())
}
// Calls the cancel function of the context
log.Warn("(svcs) The load balancer was deleted, cancelling context")
svcCtx.Cancel()
log.Warn("(svcs) waiting for load balancer to finish")
<-svcCtx.Ctx.Done()
p.svcMap.Delete(svc.UID)
}
if p.config.EnableLeaderElection && !p.config.EnableServicesElection {
if p.config.EnableBGP {
endpoints.ClearBGPHosts(svc, &p.ServiceInstances, p.bgpServer)
} else if p.config.EnableRoutingTable {
endpoints.ClearRoutes(svc, &p.ServiceInstances)
}
}
log.Info("(svcs) deleted", "service name", svc.Name, "namespace", svc.Namespace)
return true, nil
}
func (p *Processor) Stop() {
for _, instance := range p.ServiceInstances {
for _, cluster := range instance.Clusters {
cluster.Stop()
}
}
}
func (p *Processor) getServiceContext(uid types.UID) (*servicecontext.Context, error) {
svcCtx, ok := p.svcMap.Load(uid)
if !ok {
return nil, nil
}
ctx, ok := svcCtx.(*servicecontext.Context)
if !ok {
return nil, fmt.Errorf("failed to cast service context pointer - UID: %s", uid)
}
return ctx, nil
}
func (p *Processor) CountRouteReferences(route *netlink.Route) int {
return endpoints.CountRouteReferences(route, &p.ServiceInstances)
}