Browse Source

More fine grained locking in discovery cache (fixes #2667)

We only need to protect the integrity of the "finders" and "caches"
slices, and for that we only need an RLock except while actually
appending to them. The actual finders and caches are concurrency safe on
their own.
Jakob Borg 9 years ago
parent
commit
370b0fc5da
2 changed files with 60 additions and 8 deletions
  1. 8 8
      lib/discover/cache.go
  2. 52 0
      lib/discover/cache_test.go

+ 8 - 8
lib/discover/cache.go

@@ -26,7 +26,7 @@ type CachingMux struct {
 	*suture.Supervisor
 	finders []cachedFinder
 	caches  []*cache
-	mut     sync.Mutex
+	mut     sync.RWMutex
 }
 
 // A cachedFinder is a Finder with associated cache timeouts.
@@ -54,7 +54,7 @@ type cachedError interface {
 func NewCachingMux() *CachingMux {
 	return &CachingMux{
 		Supervisor: suture.NewSimple("discover.cachingMux"),
-		mut:        sync.NewMutex(),
+		mut:        sync.NewRWMutex(),
 	}
 }
 
@@ -75,7 +75,7 @@ func (m *CachingMux) Add(finder Finder, cacheTime, negCacheTime time.Duration, p
 func (m *CachingMux) Lookup(deviceID protocol.DeviceID) (direct []string, relays []Relay, err error) {
 	var pdirect []prioritizedAddress
 
-	m.mut.Lock()
+	m.mut.RLock()
 	for i, finder := range m.finders {
 		if cacheEntry, ok := m.caches[i].Get(deviceID); ok {
 			// We have a cache entry. Lets see what it says.
@@ -129,7 +129,7 @@ func (m *CachingMux) Lookup(deviceID protocol.DeviceID) (direct []string, relays
 			m.caches[i].Set(deviceID, entry)
 		}
 	}
-	m.mut.Unlock()
+	m.mut.RUnlock()
 
 	direct = uniqueSortedAddrs(pdirect)
 	relays = uniqueSortedRelays(relays)
@@ -149,12 +149,12 @@ func (m *CachingMux) Error() error {
 }
 
 func (m *CachingMux) ChildErrors() map[string]error {
-	m.mut.Lock()
 	children := make(map[string]error, len(m.finders))
+	m.mut.RLock()
 	for _, f := range m.finders {
 		children[f.String()] = f.Error()
 	}
-	m.mut.Unlock()
+	m.mut.RUnlock()
 	return children
 }
 
@@ -163,7 +163,7 @@ func (m *CachingMux) Cache() map[protocol.DeviceID]CacheEntry {
 	// children's caches.
 	res := make(map[protocol.DeviceID]CacheEntry)
 
-	m.mut.Lock()
+	m.mut.RLock()
 	for i := range m.finders {
 		// Each finder[i] has a corresponding cache at cache[i]. Go through it
 		// and populate the total, if it's newer than what's already in there.
@@ -183,7 +183,7 @@ func (m *CachingMux) Cache() map[protocol.DeviceID]CacheEntry {
 			}
 		}
 	}
-	m.mut.Unlock()
+	m.mut.RUnlock()
 
 	return res
 }

+ 52 - 0
lib/discover/cache_test.go

@@ -91,3 +91,55 @@ func (f *fakeDiscovery) String() string {
 func (f *fakeDiscovery) Cache() map[protocol.DeviceID]CacheEntry {
 	return nil
 }
+
+func TestCacheSlowLookup(t *testing.T) {
+	c := NewCachingMux()
+	c.ServeBackground()
+	defer c.Stop()
+
+	// Add a slow discovery service.
+
+	started := make(chan struct{})
+	f1 := &slowDiscovery{time.Second, started}
+	c.Add(f1, time.Minute, 0, 0)
+
+	// Start a lookup, which will take at least a second
+
+	t0 := time.Now()
+	go c.Lookup(protocol.LocalDeviceID)
+	<-started // The slow lookup method has been called so we're inside the lock
+
+	// It should be possible to get ChildErrors while it's running
+
+	c.ChildErrors()
+
+	// Only a small amount of time should have passed, not the full second
+
+	diff := time.Since(t0)
+	if diff > 500*time.Millisecond {
+		t.Error("ChildErrors was blocked for", diff)
+	}
+}
+
+type slowDiscovery struct {
+	delay   time.Duration
+	started chan struct{}
+}
+
+func (f *slowDiscovery) Lookup(deviceID protocol.DeviceID) (direct []string, relays []Relay, err error) {
+	close(f.started)
+	time.Sleep(f.delay)
+	return nil, nil, nil
+}
+
+func (f *slowDiscovery) Error() error {
+	return nil
+}
+
+func (f *slowDiscovery) String() string {
+	return "fake"
+}
+
+func (f *slowDiscovery) Cache() map[protocol.DeviceID]CacheEntry {
+	return nil
+}