Remove O(n) lookup on unlink/rename

Currently, the path tracking in the gofer involves an O(n) lookup of child fidRefs. This causes a significant overhead on unlinks in directories with lots of child fidRefs (<4k). In this transition, pathNode moves from sync.Map to normal synchronized maps. There is a small chance of contention in walk, but the lock is held for a very short time (and sync.Map also had a chance of requiring locking). OTOH, sync.Map makes it very difficult to add a fidRef reverse map. PiperOrigin-RevId: 254489952
2026-05-22 17:12:49 -07:00 · 2019-06-21 16:27:26 -07:00
parent ae4ef32b8c
commit 6f933a934f
3 changed files with 193 additions and 83 deletions
@@ -224,7 +224,7 @@ func (t *Tattach) handle(cs *connState) message {
 		file:     sf,
 		refs:     1,
 		mode:     attr.Mode.FileType(),
-		pathNode: &cs.server.pathTree,
+		pathNode: cs.server.pathTree,
 	}
 	defer root.DecRef()

@@ -552,8 +552,8 @@ func (t *Tunlinkat) handle(cs *connState) message {
 		// since we always acquire deeper in the hierarchy, we know
 		// that we are free of lock cycles.
 		childPathNode := ref.pathNode.pathNodeFor(t.Name)
-		childPathNode.mu.Lock()
-		defer childPathNode.mu.Unlock()
+		childPathNode.opMu.Lock()
+		defer childPathNode.opMu.Unlock()

 		// Do the unlink.
 		err = ref.file.UnlinkAt(t.Name, t.Flags)
@@ -22,93 +22,200 @@ import (
 // pathNode is a single node in a path traversal.
 //
 // These are shared by all fidRefs that point to the same path.
+//
+// Lock ordering:
+//   opMu
+//     childMu
+//
+//   Two different pathNodes may only be locked if Server.renameMu is held for
+//   write, in which case they can be acquired in any order.
 type pathNode struct {
-	// fidRefName is a map[*fidRef]string mapping child fidRefs to their
-	// path component name.
-	fidRefNames sync.Map
-
-	// childNodes is a map[string]*pathNode mapping child path component
-	// names to their pathNode.
-	childNodes sync.Map
-
-	// mu does *not* protect the fields above.
-	//
-	// They are not synchronized because we allow certain operations (file
-	// walk) to proceed without having to acquire a write lock. mu exists
-	// to synchronize high-level, semantic operations, such as the
+	// opMu synchronizes high-level, sematic operations, such as the
 	// simultaneous creation and deletion of a file.
-	mu sync.RWMutex
+	//
+	// opMu does not directly protect any fields in pathNode.
+	opMu sync.RWMutex
+
+	// childMu protects the fields below.
+	childMu sync.RWMutex
+
+	// childNodes maps child path component names to their pathNode.
+	childNodes map[string]*pathNode
+
+	// childRefs maps child path component names to all of the their
+	// references.
+	childRefs map[string]map[*fidRef]struct{}
+
+	// childRefNames maps child references back to their path component
+	// name.
+	childRefNames map[*fidRef]string
+}
+
+func newPathNode() *pathNode {
+	return &pathNode{
+		childNodes:    make(map[string]*pathNode),
+		childRefs:     make(map[string]map[*fidRef]struct{}),
+		childRefNames: make(map[*fidRef]string),
+	}
+}
+
+// forEachChildRef calls fn for each child reference.
+func (p *pathNode) forEachChildRef(fn func(ref *fidRef, name string)) {
+	p.childMu.RLock()
+	defer p.childMu.RUnlock()
+
+	for name, m := range p.childRefs {
+		for ref := range m {
+			fn(ref, name)
+		}
+	}
+}
+
+// forEachChildNode calls fn for each child pathNode.
+func (p *pathNode) forEachChildNode(fn func(pn *pathNode)) {
+	p.childMu.RLock()
+	defer p.childMu.RUnlock()
+
+	for _, pn := range p.childNodes {
+		fn(pn)
+	}
 }

 // pathNodeFor returns the path node for the given name, or a new one.
-//
-// Precondition: This call is synchronized w.r.t. other adding or removing of
-// children.
 func (p *pathNode) pathNodeFor(name string) *pathNode {
-	// Load the existing path node.
-	if pn, ok := p.childNodes.Load(name); ok {
-		return pn.(*pathNode)
+	p.childMu.RLock()
+	// Fast path, node already exists.
+	if pn, ok := p.childNodes[name]; ok {
+		p.childMu.RUnlock()
+		return pn
+	}
+	p.childMu.RUnlock()
+
+	// Slow path, create a new pathNode for shared use.
+	p.childMu.Lock()
+
+	// Re-check after re-lock.
+	if pn, ok := p.childNodes[name]; ok {
+		p.childMu.Unlock()
+		return pn
 	}

-	// Create a new pathNode for shared use.
-	pn, _ := p.childNodes.LoadOrStore(name, new(pathNode))
-	return pn.(*pathNode)
+	pn := newPathNode()
+	p.childNodes[name] = pn
+	p.childMu.Unlock()
+	return pn
 }

 // nameFor returns the name for the given fidRef.
 //
 // Precondition: addChild is called for ref before nameFor.
 func (p *pathNode) nameFor(ref *fidRef) string {
-	if s, ok := p.fidRefNames.Load(ref); ok {
-		return s.(string)
+	p.childMu.RLock()
+	n, ok := p.childRefNames[ref]
+	p.childMu.RUnlock()
+
+	if !ok {
+		// This should not happen, don't proceed.
+		panic(fmt.Sprintf("expected name for %+v, none found", ref))
 	}

-	// This should not happen, don't proceed.
-	panic(fmt.Sprintf("expected name for %+v, none found", ref))
+	return n
 }

-// addChild adds a child to p.
+// addChildLocked adds a child reference to p.
 //
-// This applies only to an individual fidRef.
-//
-// Precondition: ref is added only once unless it is removed before adding with
-// a new name.
-func (p *pathNode) addChild(ref *fidRef, name string) {
-	if s, ok := p.fidRefNames.Load(ref); ok {
+// Precondition: As addChild, plus childMu is locked for write.
+func (p *pathNode) addChildLocked(ref *fidRef, name string) {
+	if n, ok := p.childRefNames[ref]; ok {
 		// This should not happen, don't proceed.
-		panic(fmt.Sprintf("unexpected fidRef %+v with path %q, wanted %q", ref, s, name))
+		panic(fmt.Sprintf("unexpected fidRef %+v with path %q, wanted %q", ref, n, name))
 	}

-	p.fidRefNames.Store(ref, name)
+	p.childRefNames[ref] = name
+
+	m, ok := p.childRefs[name]
+	if !ok {
+		m = make(map[*fidRef]struct{})
+		p.childRefs[name] = m
+	}
+
+	m[ref] = struct{}{}
+}
+
+// addChild adds a child reference to p.
+//
+// Precondition: ref may only be added once at a time.
+func (p *pathNode) addChild(ref *fidRef, name string) {
+	p.childMu.Lock()
+	p.addChildLocked(ref, name)
+	p.childMu.Unlock()
 }

 // removeChild removes the given child.
 //
-// This applies only to an individual fidRef.
+// This applies only to an individual fidRef, which is not required to exist.
 func (p *pathNode) removeChild(ref *fidRef) {
-	p.fidRefNames.Delete(ref)
+	p.childMu.Lock()
+
+	// This ref may not exist anymore. This can occur, e.g., in unlink,
+	// where a removeWithName removes the ref, and then a DecRef on the ref
+	// attempts to remove again.
+	if name, ok := p.childRefNames[ref]; ok {
+		m, ok := p.childRefs[name]
+		if !ok {
+			// This should not happen, don't proceed.
+			p.childMu.Unlock()
+			panic(fmt.Sprintf("name %s missing from childfidRefs", name))
+		}
+
+		delete(m, ref)
+		if len(m) == 0 {
+			delete(p.childRefs, name)
+		}
+	}
+
+	delete(p.childRefNames, ref)
+
+	p.childMu.Unlock()
+}
+
+// addPathNodeFor adds an existing pathNode as the node for name.
+//
+// Preconditions: newName does not exist.
+func (p *pathNode) addPathNodeFor(name string, pn *pathNode) {
+	p.childMu.Lock()
+
+	if opn, ok := p.childNodes[name]; ok {
+		p.childMu.Unlock()
+		panic(fmt.Sprintf("unexpected pathNode %+v with path %q", opn, name))
+	}
+
+	p.childNodes[name] = pn
+	p.childMu.Unlock()
 }

 // removeWithName removes all references with the given name.
 //
-// The original pathNode is returned by this function, and removed from this
-// pathNode. Any operations on the removed tree must use this value.
+// The provided function is executed after reference removal. The only method
+// it may (transitively) call on this pathNode is addChildLocked.
 //
-// The provided function is executed after removal.
-//
-// Precondition: This call is synchronized w.r.t. other adding or removing of
-// children.
+// If a child pathNode for name exists, it is removed from this pathNode and
+// returned by this function. Any operations on the removed tree must use this
+// value.
 func (p *pathNode) removeWithName(name string, fn func(ref *fidRef)) *pathNode {
-	p.fidRefNames.Range(func(key, value interface{}) bool {
-		if value.(string) == name {
-			p.fidRefNames.Delete(key)
-			fn(key.(*fidRef))
-		}
-		return true
-	})
+	p.childMu.Lock()
+	defer p.childMu.Unlock()

-	// Return the original path node.
-	origPathNode := p.pathNodeFor(name)
-	p.childNodes.Delete(name)
+	if m, ok := p.childRefs[name]; ok {
+		for ref := range m {
+			delete(m, ref)
+			delete(p.childRefNames, ref)
+			fn(ref)
+		}
+	}
+
+	// Return the original path node, if it exists.
+	origPathNode := p.childNodes[name]
+	delete(p.childNodes, name)
 	return origPathNode
 }
@@ -35,7 +35,7 @@ type Server struct {
 	// These may be across different connections, but rename operations
 	// must be serialized globally for safely. There is a single pathTree
 	// for the entire server, and not per connection.
-	pathTree pathNode
+	pathTree *pathNode

 	// renameMu is a global lock protecting rename operations. With this
 	// lock, we can be certain that any given rename operation can safely
@@ -49,6 +49,7 @@ type Server struct {
 func NewServer(attacher Attacher) *Server {
 	return &Server{
 		attacher: attacher,
+		pathTree: newPathNode(),
 	}
 }

@@ -204,16 +205,13 @@ func (f *fidRef) maybeParent() *fidRef {
 // Precondition: this must be called via safelyWrite or safelyGlobal.
 func notifyDelete(pn *pathNode) {
 	// Call on all local references.
-	pn.fidRefNames.Range(func(key, _ interface{}) bool {
-		ref := key.(*fidRef)
+	pn.forEachChildRef(func(ref *fidRef, _ string) {
 		atomic.StoreUint32(&ref.deleted, 1)
-		return true
 	})

 	// Call on all subtrees.
-	pn.childNodes.Range(func(_, value interface{}) bool {
-		notifyDelete(value.(*pathNode))
-		return true
+	pn.forEachChildNode(func(pn *pathNode) {
+		notifyDelete(pn)
 	})
 }

@@ -225,8 +223,10 @@ func (f *fidRef) markChildDeleted(name string) {
 		atomic.StoreUint32(&ref.deleted, 1)
 	})

-	// Mark everything below as deleted.
-	notifyDelete(origPathNode)
+	if origPathNode != nil {
+		// Mark all children as deleted.
+		notifyDelete(origPathNode)
+	}
 }

 // notifyNameChange calls the relevant Renamed method on all nodes in the path,
@@ -236,17 +236,13 @@ func (f *fidRef) markChildDeleted(name string) {
 // Precondition: this must be called via safelyGlobal.
 func notifyNameChange(pn *pathNode) {
 	// Call on all local references.
-	pn.fidRefNames.Range(func(key, value interface{}) bool {
-		ref := key.(*fidRef)
-		name := value.(string)
+	pn.forEachChildRef(func(ref *fidRef, name string) {
 		ref.file.Renamed(ref.parent.file, name)
-		return true
 	})

 	// Call on all subtrees.
-	pn.childNodes.Range(func(name, value interface{}) bool {
-		notifyNameChange(value.(*pathNode))
-		return true
+	pn.forEachChildNode(func(pn *pathNode) {
+		notifyNameChange(pn)
 	})
 }

@@ -256,18 +252,25 @@ func notifyNameChange(pn *pathNode) {
 func (f *fidRef) renameChildTo(oldName string, target *fidRef, newName string) {
 	target.markChildDeleted(newName)
 	origPathNode := f.pathNode.removeWithName(oldName, func(ref *fidRef) {
+		// N.B. DecRef can take f.pathNode's parent's childMu. This is
+		// allowed because renameMu is held for write via safelyGlobal.
 		ref.parent.DecRef() // Drop original reference.
 		ref.parent = target // Change parent.
 		ref.parent.IncRef() // Acquire new one.
-		target.pathNode.addChild(ref, newName)
+		if f.pathNode == target.pathNode {
+			target.pathNode.addChildLocked(ref, newName)
+		} else {
+			target.pathNode.addChild(ref, newName)
+		}
 		ref.file.Renamed(target.file, newName)
 	})

-	// Replace the previous (now deleted) path node.
-	target.pathNode.childNodes.Store(newName, origPathNode)
-
-	// Call Renamed on everything above.
-	notifyNameChange(origPathNode)
+	if origPathNode != nil {
+		// Replace the previous (now deleted) path node.
+		target.pathNode.addPathNodeFor(newName, origPathNode)
+		// Call Renamed on all children.
+		notifyNameChange(origPathNode)
+	}
 }

 // safelyRead executes the given operation with the local path node locked.
@@ -275,8 +278,8 @@ func (f *fidRef) renameChildTo(oldName string, target *fidRef, newName string) {
 func (f *fidRef) safelyRead(fn func() error) (err error) {
 	f.server.renameMu.RLock()
 	defer f.server.renameMu.RUnlock()
-	f.pathNode.mu.RLock()
-	defer f.pathNode.mu.RUnlock()
+	f.pathNode.opMu.RLock()
+	defer f.pathNode.opMu.RUnlock()
 	return fn()
 }

@@ -285,8 +288,8 @@ func (f *fidRef) safelyRead(fn func() error) (err error) {
 func (f *fidRef) safelyWrite(fn func() error) (err error) {
 	f.server.renameMu.RLock()
 	defer f.server.renameMu.RUnlock()
-	f.pathNode.mu.Lock()
-	defer f.pathNode.mu.Unlock()
+	f.pathNode.opMu.Lock()
+	defer f.pathNode.opMu.Unlock()
 	return fn()
 }