Fix race: Make informed watcher start wait for Add event 🏎️ (#2036)

* Fix race: Make informed watcher start wait for Add event 🏎️

When using the informed watcher to watch a config map, previously add
events were being processed in a goroutine with no syncrhonization
making it so that code may try to access the values backed by the
configmaps before they are initialized.

This commit makes it so that the Start method of the informer will wait
for the add event to occur at least once for all config maps it is
watching.

This commit also undoes the workaround added in #1929 which was working
around the race condition identified in #1907 (and in
https://github.com/tektoncd/pipeline/issues/3720). This means that if
the synchronization was removed, the impacted test would start flaking
again. If we wanted it to reliably fail in that case, we could introduce
a sleep in the callback but that doesn't seem worth it.

I also tested this change by manually patching the changes into my
clone of tektoncd/pipeline and following the repro steps at
https://github.com/tektoncd/pipeline/issues/2815#issuecomment-733207368
Before the change I can reproduce the issue, and after the change, I
can't! :D

Fixes #1960

* Make synced callback and named wait group private 🕵️

These utility objects don't really make sense to expose as part of the
informed watcher package and are only used by the informed watcher.
Writing tests for unexported code makes me a bit :( but hopefully these
will get moved to some other package one day. And there's really no
reason to expose these to users of knative/pkg at the moment.
This commit is contained in:
Christie Wilson 2021-03-02 14:58:15 -05:00 committed by GitHub
parent 9039dc189c
commit 2bc944bb68
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 311 additions and 37 deletions

View File

@ -107,7 +107,8 @@ var _ configmap.Watcher = (*InformedWatcher)(nil)
// Asserts that InformedWatcher implements DefaultingWatcher.
var _ configmap.DefaultingWatcher = (*InformedWatcher)(nil)
// WatchWithDefault implements DefaultingWatcher.
// WatchWithDefault implements DefaultingWatcher. Adding a default for the configMap being watched means that when
// Start is called, Start will not wait for the add event from the API server.
func (i *InformedWatcher) WatchWithDefault(cm corev1.ConfigMap, o ...configmap.Observer) {
i.defaults[cm.Name] = &cm
@ -126,31 +127,59 @@ func (i *InformedWatcher) WatchWithDefault(cm corev1.ConfigMap, o ...configmap.O
i.Watch(cm.Name, o...)
}
// Start implements Watcher.
func (i *InformedWatcher) Start(stopCh <-chan struct{}) error {
// Pretend that all the defaulted ConfigMaps were just created. This is done before we start
// the informer to ensure that if a defaulted ConfigMap does exist, then the real value is
// processed after the default one.
func (i *InformedWatcher) triggerAddEventForDefaultedConfigMaps(addConfigMapEvent func(obj interface{})) {
i.ForEach(func(k string, _ []configmap.Observer) error {
if def, ok := i.defaults[k]; ok {
i.addConfigMapEvent(def)
addConfigMapEvent(def)
}
return nil
})
}
if err := i.registerCallbackAndStartInformer(stopCh); err != nil {
func (i *InformedWatcher) getConfigMapNames() []string {
var configMaps []string
i.ForEach(func(k string, _ []configmap.Observer) error {
configMaps = append(configMaps, k)
return nil
})
return configMaps
}
// Start implements Watcher. Start will wait for all watched resources to exist and for the add event handler to be
// invoked at least once for each before continuing or for the stopCh to be signalled, whichever happens first. If
// the watched resource is defaulted, Start will invoke the add event handler directly and will not wait for a further
// add event from the API server.
func (i *InformedWatcher) Start(stopCh <-chan struct{}) error {
// using the synced callback wrapper around the add event handler will allow the caller
// to wait for the add event to be processed for all configmaps
s := newSyncedCallback(i.getConfigMapNames(), i.addConfigMapEvent)
addConfigMapEvent := func(obj interface{}) {
configMap := obj.(*corev1.ConfigMap)
s.Call(obj, configMap.Name)
}
// Pretend that all the defaulted ConfigMaps were just created. This is done before we start
// the informer to ensure that if a defaulted ConfigMap does exist, then the real value is
// processed after the default one.
i.triggerAddEventForDefaultedConfigMaps(addConfigMapEvent)
if err := i.registerCallbackAndStartInformer(addConfigMapEvent, stopCh); err != nil {
return err
}
// Wait until it has been synced (WITHOUT holing the mutex, so callbacks happen)
// Wait until the shared informer has been synced (WITHOUT holing the mutex, so callbacks happen)
if ok := cache.WaitForCacheSync(stopCh, i.informer.Informer().HasSynced); !ok {
return errors.New("error waiting for ConfigMap informer to sync")
}
return i.checkObservedResourcesExist()
if err := i.checkObservedResourcesExist(); err != nil {
return err
}
// Wait until all config maps have been at least initially processed
return s.WaitForAllKeys(stopCh)
}
func (i *InformedWatcher) registerCallbackAndStartInformer(stopCh <-chan struct{}) error {
func (i *InformedWatcher) registerCallbackAndStartInformer(addConfigMapEvent func(obj interface{}), stopCh <-chan struct{}) error {
i.Lock()
defer i.Unlock()
if i.started {
@ -159,13 +188,14 @@ func (i *InformedWatcher) registerCallbackAndStartInformer(stopCh <-chan struct{
i.started = true
i.informer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{
AddFunc: i.addConfigMapEvent,
AddFunc: addConfigMapEvent,
UpdateFunc: i.updateConfigMapEvent,
DeleteFunc: i.deleteConfigMapEvent,
})
// Start the shared informer factory (non-blocking).
i.sif.Start(stopCh)
return nil
}

View File

@ -20,7 +20,6 @@ import (
"context"
"sync"
"testing"
"time"
corev1 "k8s.io/api/core/v1"
"k8s.io/apimachinery/pkg/api/equality"
@ -28,7 +27,6 @@ import (
"k8s.io/apimachinery/pkg/labels"
"k8s.io/apimachinery/pkg/runtime"
"k8s.io/apimachinery/pkg/selection"
"k8s.io/apimachinery/pkg/util/wait"
fakekubeclientset "k8s.io/client-go/kubernetes/fake"
)
@ -42,6 +40,7 @@ type counter struct {
func (c *counter) callback(cm *corev1.ConfigMap) {
c.mu.Lock()
defer c.mu.Unlock()
c.cfg = append(c.cfg, cm)
if c.wg != nil {
c.wg.Done()
@ -54,26 +53,6 @@ func (c *counter) count() int {
return len(c.cfg)
}
func (c *counter) eventuallyEquals(t *testing.T, want int) {
got := 0
err := wait.Poll(
// interval
100*time.Millisecond,
// timeout
5*time.Second,
func() (done bool, err error) {
got = c.count()
return got == want, nil
},
)
if err != nil {
t.Errorf("%v.count = %d, want %d", c.name, got, want)
}
}
func TestInformedWatcher(t *testing.T) {
fooCM := &corev1.ConfigMap{
ObjectMeta: metav1.ObjectMeta{
@ -105,10 +84,12 @@ func TestInformedWatcher(t *testing.T) {
t.Fatal("cm.Start() =", err)
}
// When Start returns the callbacks will eventually be called with the
// When Start returns the callbacks should have been called with the
// version of the objects that is available.
for _, count := range []*counter{foo1, foo2, bar} {
count.eventuallyEquals(t, 1)
for _, obj := range []*counter{foo1, foo2, bar} {
if got, want := obj.count(), 1; got != want {
t.Errorf("%v.count = %d, want %d", obj.name, got, want)
}
}
// After a "foo" event, the "foo" watchers should have 2,

View File

@ -0,0 +1,112 @@
/*
Copyright 2021 The Knative Authors
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package informer
import (
"sync"
"k8s.io/apimachinery/pkg/util/sets"
"k8s.io/apimachinery/pkg/util/wait"
)
// namedWaitGroup is used to increment and decrement a WaitGroup by name
type namedWaitGroup struct {
waitGroup sync.WaitGroup
keys sets.String
mu sync.Mutex
}
// newNamedWaitGroup returns an instantiated namedWaitGroup.
func newNamedWaitGroup() *namedWaitGroup {
return &namedWaitGroup{
keys: sets.NewString(),
}
}
// Add will add the key to the list of keys being tracked and increment the wait group.
// If the key has already been added, the wait group will not be incremented again.
func (n *namedWaitGroup) Add(key string) {
n.mu.Lock()
defer n.mu.Unlock()
if !n.keys.Has(key) {
n.keys.Insert(key)
n.waitGroup.Add(1)
}
}
// Done will decrement the counter if the key is present in the tracked keys. If it is not present
// it will be ignored.
func (n *namedWaitGroup) Done(key string) {
n.mu.Lock()
defer n.mu.Unlock()
if n.keys.Has(key) {
n.keys.Delete(key)
n.waitGroup.Done()
}
}
// Wait will wait for the underlying waitGroup to complete.
func (n *namedWaitGroup) Wait() {
n.waitGroup.Wait()
}
// syncedCallback can be used to wait for a callback to be called at least once for a list of keys.
type syncedCallback struct {
// namedWaitGroup will block until the callback has been called for all tracked entities
namedWaitGroup *namedWaitGroup
// callback is the callback that is intended to be called at least once for each key
// being tracked via WaitGroup
callback func(obj interface{})
}
// newSyncedCallback will return a syncedCallback that will track the provided keys.
func newSyncedCallback(keys []string, callback func(obj interface{})) *syncedCallback {
s := &syncedCallback{
callback: callback,
namedWaitGroup: newNamedWaitGroup(),
}
for _, key := range keys {
s.namedWaitGroup.Add(key)
}
return s
}
// Event is intended to be a wrapper for the actual event handler; this wrapper will signal via
// the wait group that the event handler has been called at least once for the key.
func (s *syncedCallback) Call(obj interface{}, key string) {
s.callback(obj)
s.namedWaitGroup.Done(key)
}
// WaitForAllKeys will block until s.Call has been called for all the keys we are tracking or the stop signal is
// received.
func (s *syncedCallback) WaitForAllKeys(stopCh <-chan struct{}) error {
c := make(chan struct{})
go func() {
defer close(c)
s.namedWaitGroup.Wait()
}()
select {
case <-c:
return nil
case <-stopCh:
return wait.ErrWaitTimeout
}
}

View File

@ -0,0 +1,151 @@
/*
Copyright 2021 The Knative Authors
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package informer
import (
"testing"
"time"
)
func TestNamedWaitGroup(t *testing.T) {
nwg := newNamedWaitGroup()
// nothing has been added so wait returns immediately
initiallyDone := make(chan struct{})
go func() {
defer close(initiallyDone)
nwg.Wait()
}()
select {
case <-time.After(1 * time.Second):
t.Fatalf("Wait should have returned immediately but still hadn't after timeout elapsed")
case <-initiallyDone:
// the Wait returned as expected since nothing was tracked
}
// Add some keys to track
nwg.Add("foo")
nwg.Add("bar")
// Adding keys multiple times shouldn't increment the counter again
nwg.Add("bar")
// Now that we've added keys, when we Wait, it should block
done := make(chan struct{})
go func() {
defer close(done)
nwg.Wait()
}()
// Indicate that this key is done
nwg.Done("foo")
// Indicating done on a key that doesn't exist should do nothing
nwg.Done("doesnt exist")
// Only one of the tracked keys has completed, so the channel should not yet have closed
select {
case <-done:
t.Fatalf("Wait returned before all keys were done")
default:
// as expected, the channel is still open (waiting for the final key to be done)
}
// Indicate the final key is done
nwg.Done("bar")
// Now that all keys are done, the Wait should return
select {
case <-time.After(1 * time.Second):
t.Fatalf("Wait should have returned immediately but still hadn't after timeout elapsed")
case <-done:
// completed successfully
}
}
func TestSyncedCallback(t *testing.T) {
keys := []string{"foo", "bar"}
objs := []interface{}{"fooobj", "barobj"}
var seen []interface{}
callback := func(obj interface{}) {
seen = append(seen, obj)
}
sc := newSyncedCallback(keys, callback)
// Wait for the callback to be called for all of the keys
stopCh := make(chan struct{})
done := make(chan struct{})
go func() {
defer close(done)
sc.WaitForAllKeys(stopCh)
}()
// Call the callback for one of the keys
sc.Call(objs[0], "foo")
// Only one of the tracked keys has been synced so we should still be waiting
select {
case <-done:
t.Fatalf("Wait returned before all keys were done")
default:
// as expected, the channel is still open (waiting for the final key to be done)
}
// Call the callback for the other key
sc.Call(objs[1], "bar")
// Now that all keys are done, the Wait should return
select {
case <-time.After(1 * time.Second):
t.Fatalf("WaitForAllKeys should have returned but still hadn't after timeout elapsed")
case <-done:
// completed successfully
}
if len(seen) != 2 || seen[0] != objs[0] || seen[1] != objs[1] {
t.Errorf("callback wasn't called as expected, expected to see %v but saw %v", objs, seen)
}
}
func TestSyncedCallbackStops(t *testing.T) {
sc := newSyncedCallback([]string{"somekey"}, func(obj interface{}) {})
// Wait for the callback to be called - which it won't be!
stopCh := make(chan struct{})
done := make(chan struct{})
go func() {
defer close(done)
sc.WaitForAllKeys(stopCh)
}()
// Nothing has been synced so we should still be waiting
select {
case <-done:
t.Fatalf("Wait returned before all keys were done")
default:
// as expected, the channel is still open
}
// signal to stop via the stop channel
close(stopCh)
// Even though the callback wasn't called, the Wait should return b/c of the stop channel
select {
case <-time.After(1 * time.Second):
t.Fatalf("WaitForAllKeys should have returned because of the stop channel but still hadn't after timeout elapsed")
case <-done:
// stopped successfully
}
}