6 files changed, 38 insertions, 23 deletions
diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD
index 0b4f18ab5..07c8383e6 100644
--- a/pkg/sentry/vfs/BUILD
+++ b/pkg/sentry/vfs/BUILD
@@ -43,6 +43,7 @@ go_library(
         "//pkg/abi/linux",
         "//pkg/context",
         "//pkg/fspath",
+        "//pkg/gohacks",
         "//pkg/log",
         "//pkg/sentry/arch",
         "//pkg/sentry/fs/lock",
diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go
index eed41139b..3da45d744 100644
--- a/pkg/sentry/vfs/epoll.go
+++ b/pkg/sentry/vfs/epoll.go
@@ -202,6 +202,9 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, event lin
 	// Add epi to file.epolls so that it is removed when the last
 	// FileDescription reference is dropped.
 	file.epollMu.Lock()
+	if file.epolls == nil {
+		file.epolls = make(map[*epollInterest]struct{})
+	}
 	file.epolls[epi] = struct{}{}
 	file.epollMu.Unlock()
 
diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go
index 9912df799..31a4e5480 100644
--- a/pkg/sentry/vfs/mount.go
+++ b/pkg/sentry/vfs/mount.go
@@ -139,6 +139,23 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth
 	return mntns, nil
 }
 
+// NewDisconnectedMount returns a Mount representing fs with the given root
+// (which may be nil). The new Mount is not associated with any MountNamespace
+// and is not connected to any other Mounts. References are taken on fs and
+// root.
+func (vfs *VirtualFilesystem) NewDisconnectedMount(fs *Filesystem, root *Dentry, opts *MountOptions) (*Mount, error) {
+	fs.IncRef()
+	if root != nil {
+		root.IncRef()
+	}
+	return &Mount{
+		vfs:  vfs,
+		fs:   fs,
+		root: root,
+		refs: 1,
+	}, nil
+}
+
 // MountAt creates and mounts a Filesystem configured by the given arguments.
 func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error {
 	rft := vfs.getFilesystemType(fsTypeName)
diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go
index 1fe766a44..bc7581698 100644
--- a/pkg/sentry/vfs/mount_unsafe.go
+++ b/pkg/sentry/vfs/mount_unsafe.go
@@ -26,6 +26,7 @@ import (
 	"sync/atomic"
 	"unsafe"
 
+	"gvisor.dev/gvisor/pkg/gohacks"
 	"gvisor.dev/gvisor/pkg/sync"
 )
 
@@ -160,7 +161,7 @@ func newMountTableSlots(cap uintptr) unsafe.Pointer {
 // Lookup may be called even if there are concurrent mutators of mt.
 func (mt *mountTable) Lookup(parent *Mount, point *Dentry) *Mount {
 	key := mountKey{parent: unsafe.Pointer(parent), point: unsafe.Pointer(point)}
-	hash := memhash(noescape(unsafe.Pointer(&key)), uintptr(mt.seed), mountKeyBytes)
+	hash := memhash(gohacks.Noescape(unsafe.Pointer(&key)), uintptr(mt.seed), mountKeyBytes)
 
 loop:
 	for {
@@ -361,12 +362,3 @@ func memhash(p unsafe.Pointer, seed, s uintptr) uintptr
 
 //go:linkname rand32 runtime.fastrand
 func rand32() uint32
-
-// This is copy/pasted from runtime.noescape(), and is needed because arguments
-// apparently escape from all functions defined by linkname.
-//
-//go:nosplit
-func noescape(p unsafe.Pointer) unsafe.Pointer {
-	x := uintptr(p)
-	return unsafe.Pointer(x ^ 0)
-}
diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go
index 8a0b382f6..eb4ebb511 100644
--- a/pkg/sentry/vfs/resolving_path.go
+++ b/pkg/sentry/vfs/resolving_path.go
@@ -228,7 +228,7 @@ func (rp *ResolvingPath) Advance() {
 		rp.pit = next
 	} else { // at end of path segment, continue with next one
 		rp.curPart--
-		rp.pit = rp.parts[rp.curPart-1]
+		rp.pit = rp.parts[rp.curPart]
 	}
 }
 
diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go
index 8f29031b2..bde81e1ef 100644
--- a/pkg/sentry/vfs/vfs.go
+++ b/pkg/sentry/vfs/vfs.go
@@ -126,17 +126,23 @@ func (vfs *VirtualFilesystem) Init() error {
 	// Construct vfs.anonMount.
 	anonfsDevMinor, err := vfs.GetAnonBlockDevMinor()
 	if err != nil {
-		return err
+		// This shouldn't be possible since anonBlockDevMinorNext was
+		// initialized to 1 above (no device numbers have been allocated yet).
+		panic(fmt.Sprintf("VirtualFilesystem.Init: device number allocation for anonfs failed: %v", err))
 	}
 	anonfs := anonFilesystem{
 		devMinor: anonfsDevMinor,
 	}
 	anonfs.vfsfs.Init(vfs, &anonfs)
-	vfs.anonMount = &Mount{
-		vfs:  vfs,
-		fs:   &anonfs.vfsfs,
-		refs: 1,
+	defer anonfs.vfsfs.DecRef()
+	anonMount, err := vfs.NewDisconnectedMount(&anonfs.vfsfs, nil, &MountOptions{})
+	if err != nil {
+		// We should not be passing any MountOptions that would cause
+		// construction of this mount to fail.
+		panic(fmt.Sprintf("VirtualFilesystem.Init: anonfs mount failed: %v", err))
 	}
+	vfs.anonMount = anonMount
+
 	return nil
 }
 
@@ -385,15 +391,11 @@ func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credential
 				// Only a regular file can be executed.
 				stat, err := fd.Stat(ctx, StatOptions{Mask: linux.STATX_TYPE})
 				if err != nil {
+					fd.DecRef()
 					return nil, err
 				}
-				if stat.Mask&linux.STATX_TYPE != 0 {
-					// This shouldn't happen, but if type can't be retrieved, file can't
-					// be executed.
-					return nil, syserror.EACCES
-				}
-				if t := linux.FileMode(stat.Mode).FileType(); t != linux.ModeRegular {
-					ctx.Infof("%q is not a regular file: %v", pop.Path, t)
+				if stat.Mask&linux.STATX_TYPE == 0 || stat.Mode&linux.S_IFMT != linux.S_IFREG {
+					fd.DecRef()
 					return nil, syserror.EACCES
 				}
 			}