Abstract mount tree cloning into its own mount method.

When copying a mount tree, new child mounts are added to a mount's
pending list. Each of these mounts has its pending mountpoint already set
with setKey. Eventually this method will be used to enable recursive binds.

PiperOrigin-RevId: 563236298

pkg/sentry/vfs/mount.go

@@ -108,6 +108,10 @@ type Mount struct {
 	// Mount.EndWrite(). The MSB of writers is set if MS_RDONLY is in effect.
 	// writers is accessed using atomic memory operations.
 	writers atomicbitops.Int64
+
+	// pendingChildren is a list of new child mounts that have not yet been
+	// connected to this mount as the parent.
+	pendingChildren []*Mount
 }
 
 func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *MountNamespace, opts *MountOptions) *Mount {
@@ -304,6 +308,39 @@ func (vfs *VirtualFilesystem) cloneMount(mnt *Mount, root *Dentry, mopts *MountO
 	return clone
 }
 
+type cloneTreeNode struct {
+	prevMount   *Mount
+	parentMount *Mount
+}
+
+// cloneMountTree creates a copy of mnt's tree with the specified root
+// dentry at root. The new descendents are added to mnt's pending mount list.
+//
+// +checklocks:vfs.mountMu
+func (vfs *VirtualFilesystem) cloneMountTree(ctx context.Context, mnt *Mount, root *Dentry) (*Mount, error) {
+	clone := vfs.cloneMount(mnt, root, nil)
+	queue := []cloneTreeNode{{mnt, clone}}
+	for len(queue) != 0 {
+		p := queue[len(queue)-1]
+		queue = queue[:len(queue)-1]
+		for c := range p.prevMount.children {
+			m := vfs.cloneMount(c, c.root, nil)
+			vfs.delayDecRef(m)
+			mp := VirtualDentry{
+				mount:  p.parentMount,
+				dentry: c.point(),
+			}
+			mp.IncRef()
+			m.setKey(mp)
+			p.parentMount.pendingChildren = append(p.parentMount.pendingChildren, m)
+			if len(c.children) != 0 {
+				queue = append(queue, cloneTreeNode{c, m})
+			}
+		}
+	}
+	return clone, nil
+}
+
 // BindAt creates a clone of the source path's parent mount and mounts it at
 // the target path. The new mount's root dentry is one pointed to by the source
 // path.

pkg/sentry/vfs/namespace.go

@@ -15,6 +15,8 @@
 package vfs
 
 import (
+	"fmt"
+
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/errors/linuxerr"
 	"gvisor.dev/gvisor/pkg/refs"
@@ -125,22 +127,28 @@ func (vfs *VirtualFilesystem) NewMountNamespaceFrom(
 	return mntns
 }
 
-type cloneEntry struct {
-	prevMount   *Mount
-	parentMount *Mount
-}
-
 // +checklocks:vfs.mountMu
-func (vfs *VirtualFilesystem) updateRootAndCWD(ctx context.Context, root *VirtualDentry, cwd *VirtualDentry, src *Mount, dst *Mount) {
+func (vfs *VirtualFilesystem) updateRootAndCWD(ctx context.Context, root *VirtualDentry, cwd *VirtualDentry, srcRoot *Mount, dstRoot *Mount) {
-	if root.mount == src {
+	// The mount trees are exact copies of each other so submountsLocked will
-		vfs.delayDecRef(root.mount)
+	// return corresponding mounts in the same order.
-		root.mount = dst
+	srcMounts := srcRoot.submountsLocked()
-		root.mount.IncRef()
+	dstMounts := dstRoot.submountsLocked()
+	if len(srcMounts) != len(dstMounts) {
+		panic(fmt.Sprintf("mount trees are not the same size: len(srcTree) = %d, len(dstTree) = %d", len(srcMounts), len(dstMounts)))
 	}
-	if cwd.mount == src {
+	for i := 0; i < len(srcMounts); i++ {
-		vfs.delayDecRef(cwd.mount)
+		old := srcMounts[i]
-		cwd.mount = dst
+		new := dstMounts[i]
-		cwd.mount.IncRef()
+		if root.mount == old {
+			vfs.delayDecRef(root.mount)
+			root.mount = new
+			root.mount.IncRef()
+		}
+		if cwd.mount == old {
+			vfs.delayDecRef(cwd.mount)
+			cwd.mount = new
+			cwd.mount.IncRef()
+		}
 	}
 }
 
@@ -170,38 +178,16 @@ func (vfs *VirtualFilesystem) CloneMountNamespace(
 	vfs.lockMounts()
 	defer vfs.unlockMounts(ctx)
 
-	ns.root.root.IncRef()
+	newRoot, err := vfs.cloneMountTree(ctx, ns.root, ns.root.root)
-	ns.root.fs.IncRef()
+	if err != nil {
-	newns.root = newMount(vfs, ns.root.fs, ns.root.root, newns, &MountOptions{Flags: ns.root.Flags, ReadOnly: ns.root.ReadOnly()})
+		newns.DecRef(ctx)
-	if ns.root.isShared {
+		vfs.abortTree(ctx, newRoot)
-		vfs.addPeer(ns.root, newns.root)
+		return nil, err
 	}
+	newns.root = newRoot
+	newns.root.ns = newns
+	vfs.commitTree(ctx, newRoot)
 	vfs.updateRootAndCWD(ctx, root, cwd, ns.root, newns.root)
-
-	queue := []cloneEntry{cloneEntry{ns.root, newns.root}}
-	for len(queue) != 0 {
-		p := queue[0]
-		queue = queue[1:]
-		for c := range p.prevMount.children {
-			m := vfs.cloneMount(c, c.root, nil)
-			vd := VirtualDentry{
-				mount:  p.parentMount,
-				dentry: c.point(),
-			}
-			vd.IncRef()
-
-			err := vfs.connectMountAtLocked(ctx, m, vd)
-			vfs.delayDecRef(m)
-			if err != nil {
-				newns.DecRef(ctx)
-				return nil, err
-			}
-			vfs.updateRootAndCWD(ctx, root, cwd, c, m)
-			if len(c.children) != 0 {
-				queue = append(queue, cloneEntry{c, m})
-			}
-		}
-	}
 	return newns, nil
 }
 

pkg/sentry/vfs/propagation.go

@@ -153,6 +153,43 @@ func (vfs *VirtualFilesystem) abortPropagationTree(ctx context.Context, tree map
 	}
 }
 
+// +checklocks:vfs.mountMu
+func (vfs *VirtualFilesystem) commitTree(ctx context.Context, mnt *Mount) {
+	vfs.mounts.seq.BeginWrite()
+	for _, c := range mnt.pendingChildren {
+		vfs.commitTreeSeqed(ctx, c)
+	}
+	mnt.pendingChildren = nil
+	vfs.mounts.seq.EndWrite()
+}
+
+// +checklocks:vfs.mountMu
+func (vfs *VirtualFilesystem) commitTreeSeqed(ctx context.Context, mnt *Mount) {
+	mp := mnt.getKey()
+	mp.dentry.mu.Lock()
+	vfs.connectLocked(mnt, mp, mp.mount.ns)
+	mp.dentry.mu.Unlock()
+	for _, c := range mnt.pendingChildren {
+		vfs.commitTreeSeqed(ctx, c)
+	}
+	mnt.pendingChildren = nil
+}
+
+// abortTree releases references on a pending mount and all its pending
+// descendants.
+//
+// +checklocks:vfs.mountMu
+func (vfs *VirtualFilesystem) abortTree(ctx context.Context, mnt *Mount) {
+	mp := mnt.getKey()
+	vfs.delayDecRef(mnt)
+	vfs.delayDecRef(mp.dentry)
+	vfs.setPropagation(mnt, linux.MS_PRIVATE)
+	for _, c := range mnt.pendingChildren {
+		vfs.abortTree(ctx, c)
+	}
+	mnt.pendingChildren = nil
+}
+
 // SetMountPropagationAt changes the propagation type of the mount pointed to by
 // pop.
 func (vfs *VirtualFilesystem) SetMountPropagationAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, propFlags uint32) error {