Inline vfs.VirtualFilesystem in Kernel struct

This saves one pointer dereference per VFS access.

Updates #1623

PiperOrigin-RevId: 295216176

20 files changed, 94 insertions, 44 deletions

diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go
index 8973754c8..5457ba5e7 100644
--- a/pkg/sentry/control/proc.go
+++ b/pkg/sentry/control/proc.go
@@ -199,7 +199,7 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI
 			}
 
 			paths := fs.GetPath(initArgs.Envv)
-			vfsObj := proc.Kernel.VFS
+			vfsObj := proc.Kernel.VFS()
 			file, err := ResolveExecutablePath(ctx, vfsObj, initArgs.WorkingDirectory, initArgs.Argv[0], paths)
 			if err != nil {
 				return nil, 0, nil, fmt.Errorf("error finding executable %q in PATH %v: %v", initArgs.Argv[0], paths, err)
diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go
index 73308a2b5..b6d52c015 100644
--- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go
+++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go
@@ -29,7 +29,10 @@ func TestDevtmpfs(t *testing.T) {
 	ctx := contexttest.Context(t)
 	creds := auth.CredentialsFromContext(ctx)
 
-	vfsObj := vfs.New()
+	vfsObj := &vfs.VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
 	// Register tmpfs just so that we can have a root filesystem that isn't
 	// devtmpfs.
 	vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
diff --git a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go
index 2015a8871..89caee3df 100644
--- a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go
+++ b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go
@@ -52,7 +52,10 @@ func setUp(b *testing.B, imagePath string) (context.Context, *vfs.VirtualFilesys
 	creds := auth.CredentialsFromContext(ctx)
 
 	// Create VFS.
-	vfsObj := vfs.New()
+	vfsObj := &vfs.VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		return nil, nil, nil, nil, err
+	}
 	vfsObj.MustRegisterFilesystemType("extfs", ext.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
diff --git a/pkg/sentry/fsimpl/ext/ext_test.go b/pkg/sentry/fsimpl/ext/ext_test.go
index 05f992826..ef6127f3c 100644
--- a/pkg/sentry/fsimpl/ext/ext_test.go
+++ b/pkg/sentry/fsimpl/ext/ext_test.go
@@ -65,7 +65,10 @@ func setUp(t *testing.T, imagePath string) (context.Context, *vfs.VirtualFilesys
 	creds := auth.CredentialsFromContext(ctx)
 
 	// Create VFS.
-	vfsObj := vfs.New()
+	vfsObj := &vfs.VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
 	vfsObj.MustRegisterFilesystemType("extfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go
index 96a16e654..0459fb305 100644
--- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go
+++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go
@@ -45,7 +45,10 @@ type RootDentryFn func(*auth.Credentials, *filesystem) *kernfs.Dentry
 func newTestSystem(t *testing.T, rootFn RootDentryFn) *testutil.System {
 	ctx := contexttest.Context(t)
 	creds := auth.CredentialsFromContext(ctx)
-	v := vfs.New()
+	v := &vfs.VirtualFilesystem{}
+	if err := v.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
 	v.MustRegisterFilesystemType("testfs", &fsType{rootFn: rootFn}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go
index 96c72cbc9..c5d531fe0 100644
--- a/pkg/sentry/fsimpl/proc/tasks_test.go
+++ b/pkg/sentry/fsimpl/proc/tasks_test.go
@@ -90,7 +90,7 @@ func setup(t *testing.T) *testutil.System {
 	ctx := k.SupervisorContext()
 	creds := auth.CredentialsFromContext(ctx)
 
-	k.VFS.MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+	k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
 	fsOpts := vfs.GetFilesystemOptions{
@@ -101,11 +101,11 @@ func setup(t *testing.T) *testutil.System {
 			},
 		},
 	}
-	mntns, err := k.VFS.NewMountNamespace(ctx, creds, "", Name, &fsOpts)
+	mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", Name, &fsOpts)
 	if err != nil {
 		t.Fatalf("NewMountNamespace(): %v", err)
 	}
-	return testutil.NewSystem(ctx, t, k.VFS, mntns)
+	return testutil.NewSystem(ctx, t, k.VFS(), mntns)
 }
 
 func TestTasksEmpty(t *testing.T) {
diff --git a/pkg/sentry/fsimpl/sys/sys_test.go b/pkg/sentry/fsimpl/sys/sys_test.go
index 5d1ba5867..4b3602d47 100644
--- a/pkg/sentry/fsimpl/sys/sys_test.go
+++ b/pkg/sentry/fsimpl/sys/sys_test.go
@@ -34,15 +34,15 @@ func newTestSystem(t *testing.T) *testutil.System {
 	}
 	ctx := k.SupervisorContext()
 	creds := auth.CredentialsFromContext(ctx)
-	k.VFS.MustRegisterFilesystemType(sys.Name, sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+	k.VFS().MustRegisterFilesystemType(sys.Name, sys.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
 
-	mns, err := k.VFS.NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{})
+	mns, err := k.VFS().NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{})
 	if err != nil {
 		t.Fatalf("Failed to create new mount namespace: %v", err)
 	}
-	return testutil.NewSystem(ctx, t, k.VFS, mns)
+	return testutil.NewSystem(ctx, t, k.VFS(), mns)
 }
 
 func TestReadCPUFile(t *testing.T) {
diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go
index a91b3ec4d..d0be32e72 100644
--- a/pkg/sentry/fsimpl/testutil/kernel.go
+++ b/pkg/sentry/fsimpl/testutil/kernel.go
@@ -102,12 +102,13 @@ func Boot() (*kernel.Kernel, error) {
 
 	kernel.VFS2Enabled = true
 
-	vfsObj := vfs.New()
-	vfsObj.MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+	if err := k.VFS().Init(); err != nil {
+		return nil, fmt.Errorf("VFS init: %v", err)
+	}
+	k.VFS().MustRegisterFilesystemType(tmpfs.Name, &tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 		AllowUserList:  true,
 	})
-	k.VFS = vfsObj
 
 	ls, err := limits.NewLinuxLimitSet()
 	if err != nil {
diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
index 9fce5e4b4..383133e44 100644
--- a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
+++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go
@@ -175,7 +175,10 @@ func BenchmarkVFS2MemfsStat(b *testing.B) {
 			creds := auth.CredentialsFromContext(ctx)
 
 			// Create VFS.
-			vfsObj := vfs.New()
+			vfsObj := vfs.VirtualFilesystem{}
+			if err := vfsObj.Init(); err != nil {
+				b.Fatalf("VFS init: %v", err)
+			}
 			vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 				AllowUserMount: true,
 			})
@@ -366,7 +369,10 @@ func BenchmarkVFS2MemfsMountStat(b *testing.B) {
 			creds := auth.CredentialsFromContext(ctx)
 
 			// Create VFS.
-			vfsObj := vfs.New()
+			vfsObj := vfs.VirtualFilesystem{}
+			if err := vfsObj.Init(); err != nil {
+				b.Fatalf("VFS init: %v", err)
+			}
 			vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 				AllowUserMount: true,
 			})
diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go
index 5ee7f2a72..1614f2c39 100644
--- a/pkg/sentry/fsimpl/tmpfs/pipe_test.go
+++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go
@@ -151,7 +151,10 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy
 	creds := auth.CredentialsFromContext(ctx)
 
 	// Create VFS.
-	vfsObj := vfs.New()
+	vfsObj := &vfs.VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
 	vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go
index e9f71e334..0399725cf 100644
--- a/pkg/sentry/fsimpl/tmpfs/regular_file_test.go
+++ b/pkg/sentry/fsimpl/tmpfs/regular_file_test.go
@@ -40,7 +40,11 @@ var nextFileID int64
 func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentry, func(), error) {
 	creds := auth.CredentialsFromContext(ctx)
 
-	vfsObj := vfs.New()
+	vfsObj := &vfs.VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("VFS init: %v", err)
+	}
+
 	vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
 		AllowUserMount: true,
 	})
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index 2665f057c..ea21af33f 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -246,7 +246,7 @@ type Kernel struct {
 	SpecialOpts
 
 	// VFS keeps the filesystem state used across the kernel.
-	VFS *vfs.VirtualFilesystem
+	vfs vfs.VirtualFilesystem
 }
 
 // InitKernelArgs holds arguments to Init.
@@ -815,7 +815,7 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID,
 				FollowFinalSymlink: true,
 			}
 			var err error
-			wd, err = k.VFS.GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{
+			wd, err = k.VFS().GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{
 				CheckSearchable: true,
 			})
 			if err != nil {
@@ -1506,3 +1506,8 @@ func (k *Kernel) EmitUnimplementedEvent(ctx context.Context) {
 		Registers: t.Arch().StateData().Proto(),
 	})
 }
+
+// VFS returns the virtual filesystem for the kernel.
+func (k *Kernel) VFS() *vfs.VirtualFilesystem {
+	return &k.vfs
+}
diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go
index 486a76475..35b208721 100644
--- a/pkg/sentry/vfs/dentry.go
+++ b/pkg/sentry/vfs/dentry.go
@@ -71,6 +71,8 @@ import (
 // lifetime. Dentry reference counts only indicate the extent to which VFS
 // requires Dentries to exist; Filesystems may elect to cache or discard
 // Dentries with zero references.
+//
+// +stateify savable
 type Dentry struct {
 	// parent is this Dentry's parent in this Filesystem. If this Dentry is
 	// independent, parent is nil.
@@ -89,7 +91,7 @@ type Dentry struct {
 	children map[string]*Dentry
 
 	// mu synchronizes disowning and mounting over this Dentry.
-	mu sync.Mutex
+	mu sync.Mutex `state:"nosave"`
 
 	// impl is the DentryImpl associated with this Dentry. impl is immutable.
 	// This should be the last field in Dentry.
diff --git a/pkg/sentry/vfs/device.go b/pkg/sentry/vfs/device.go
index 3af2aa58d..bda5576fa 100644
--- a/pkg/sentry/vfs/device.go
+++ b/pkg/sentry/vfs/device.go
@@ -56,6 +56,7 @@ type Device interface {
 	Open(ctx context.Context, mnt *Mount, d *Dentry, opts OpenOptions) (*FileDescription, error)
 }
 
+// +stateify savable
 type registeredDevice struct {
 	dev  Device
 	opts RegisterDeviceOptions
@@ -63,6 +64,8 @@ type registeredDevice struct {
 
 // RegisterDeviceOptions contains options to
 // VirtualFilesystem.RegisterDevice().
+//
+// +stateify savable
 type RegisterDeviceOptions struct {
 	// GroupName is the name shown for this device registration in
 	// /proc/devices. If GroupName is empty, this registration will not be
diff --git a/pkg/sentry/vfs/file_description_impl_util_test.go b/pkg/sentry/vfs/file_description_impl_util_test.go
index 8fa26418e..3a75d4d62 100644
--- a/pkg/sentry/vfs/file_description_impl_util_test.go
+++ b/pkg/sentry/vfs/file_description_impl_util_test.go
@@ -107,7 +107,10 @@ func (fd *testFD) SetStat(ctx context.Context, opts SetStatOptions) error {
 func TestGenCountFD(t *testing.T) {
 	ctx := contexttest.Context(t)
 
-	vfsObj := New() // vfs.New()
+	vfsObj := &VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
 	fd := newTestFD(vfsObj, linux.O_RDWR, &genCount{})
 	defer fd.DecRef()
 
@@ -162,7 +165,10 @@ func TestGenCountFD(t *testing.T) {
 func TestWritable(t *testing.T) {
 	ctx := contexttest.Context(t)
 
-	vfsObj := New() // vfs.New()
+	vfsObj := &VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
 	fd := newTestFD(vfsObj, linux.O_RDWR, &storeData{data: "init"})
 	defer fd.DecRef()
 
diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go
index a06a6caf3..556976d0b 100644
--- a/pkg/sentry/vfs/filesystem.go
+++ b/pkg/sentry/vfs/filesystem.go
@@ -29,6 +29,8 @@ import (
 // Filesystem methods require that a reference is held.
 //
 // Filesystem is analogous to Linux's struct super_block.
+//
+// +stateify savable
 type Filesystem struct {
 	// refs is the reference count. refs is accessed using atomic memory
 	// operations.
diff --git a/pkg/sentry/vfs/filesystem_type.go b/pkg/sentry/vfs/filesystem_type.go
index c58b70728..bb9cada81 100644
--- a/pkg/sentry/vfs/filesystem_type.go
+++ b/pkg/sentry/vfs/filesystem_type.go
@@ -44,6 +44,7 @@ type GetFilesystemOptions struct {
 	InternalData interface{}
 }
 
+// +stateify savable
 type registeredFilesystemType struct {
 	fsType FilesystemType
 	opts   RegisterFilesystemTypeOptions
diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go
index ad2c9fcf4..9912df799 100644
--- a/pkg/sentry/vfs/mount.go
+++ b/pkg/sentry/vfs/mount.go
@@ -38,6 +38,8 @@ import (
 //
 // Mount is analogous to Linux's struct mount. (gVisor does not distinguish
 // between struct mount and struct vfsmount.)
+//
+// +stateify savable
 type Mount struct {
 	// vfs, fs, and root are immutable. References are held on fs and root.
 	//
@@ -85,6 +87,8 @@ type Mount struct {
 // MountNamespace methods require that a reference is held.
 //
 // MountNamespace is analogous to Linux's struct mnt_namespace.
+//
+// +stateify savable
 type MountNamespace struct {
 	// root is the MountNamespace's root mount. root is immutable.
 	root *Mount
diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go
index bd90d36c4..1fe766a44 100644
--- a/pkg/sentry/vfs/mount_unsafe.go
+++ b/pkg/sentry/vfs/mount_unsafe.go
@@ -64,6 +64,8 @@ func (mnt *Mount) storeKey(vd VirtualDentry) {
 // (provided mutation is sufficiently uncommon).
 //
 // mountTable.Init() must be called on new mountTables before use.
+//
+// +stateify savable
 type mountTable struct {
 	// mountTable is implemented as a seqcount-protected hash table that
 	// resolves collisions with linear probing, featuring Robin Hood insertion
@@ -75,8 +77,8 @@ type mountTable struct {
 	// intrinsics and inline assembly, limiting the performance of this
 	// approach.)
 
-	seq  sync.SeqCount
-	seed uint32 // for hashing keys
+	seq  sync.SeqCount `state:"nosave"`
+	seed uint32        // for hashing keys
 
 	// size holds both length (number of elements) and capacity (number of
 	// slots): capacity is stored as its base-2 log (referred to as order) in
@@ -89,7 +91,7 @@ type mountTable struct {
 	// length and cap in separate uint32s) for ~free.
 	size uint64
 
-	slots unsafe.Pointer // []mountSlot; never nil after Init
+	slots unsafe.Pointer `state:"nosave"` // []mountSlot; never nil after Init
 }
 
 type mountSlot struct {
diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go
index 51deae313..8f29031b2 100644
--- a/pkg/sentry/vfs/vfs.go
+++ b/pkg/sentry/vfs/vfs.go
@@ -46,11 +46,13 @@ import (
 //
 // There is no analogue to the VirtualFilesystem type in Linux, as the
 // equivalent state in Linux is global.
+//
+// +stateify savable
 type VirtualFilesystem struct {
 	// mountMu serializes mount mutations.
 	//
 	// mountMu is analogous to Linux's namespace_sem.
-	mountMu sync.Mutex
+	mountMu sync.Mutex `state:"nosave"`
 
 	// mounts maps (mount parent, mount point) pairs to mounts. (Since mounts
 	// are uniquely namespaced, including mount parent in the key correctly
@@ -89,44 +91,42 @@ type VirtualFilesystem struct {
 
 	// devices contains all registered Devices. devices is protected by
 	// devicesMu.
-	devicesMu sync.RWMutex
+	devicesMu sync.RWMutex `state:"nosave"`
 	devices   map[devTuple]*registeredDevice
 
 	// anonBlockDevMinor contains all allocated anonymous block device minor
 	// numbers. anonBlockDevMinorNext is a lower bound for the smallest
 	// unallocated anonymous block device number. anonBlockDevMinorNext and
 	// anonBlockDevMinor are protected by anonBlockDevMinorMu.
-	anonBlockDevMinorMu   sync.Mutex
+	anonBlockDevMinorMu   sync.Mutex `state:"nosave"`
 	anonBlockDevMinorNext uint32
 	anonBlockDevMinor     map[uint32]struct{}
 
 	// fsTypes contains all registered FilesystemTypes. fsTypes is protected by
 	// fsTypesMu.
-	fsTypesMu sync.RWMutex
+	fsTypesMu sync.RWMutex `state:"nosave"`
 	fsTypes   map[string]*registeredFilesystemType
 
 	// filesystems contains all Filesystems. filesystems is protected by
 	// filesystemsMu.
-	filesystemsMu sync.Mutex
+	filesystemsMu sync.Mutex `state:"nosave"`
 	filesystems   map[*Filesystem]struct{}
 }
 
-// New returns a new VirtualFilesystem with no mounts or FilesystemTypes.
-func New() *VirtualFilesystem {
-	vfs := &VirtualFilesystem{
-		mountpoints:           make(map[*Dentry]map[*Mount]struct{}),
-		devices:               make(map[devTuple]*registeredDevice),
-		anonBlockDevMinorNext: 1,
-		anonBlockDevMinor:     make(map[uint32]struct{}),
-		fsTypes:               make(map[string]*registeredFilesystemType),
-		filesystems:           make(map[*Filesystem]struct{}),
-	}
+// Init initializes a new VirtualFilesystem with no mounts or FilesystemTypes.
+func (vfs *VirtualFilesystem) Init() error {
+	vfs.mountpoints = make(map[*Dentry]map[*Mount]struct{})
+	vfs.devices = make(map[devTuple]*registeredDevice)
+	vfs.anonBlockDevMinorNext = 1
+	vfs.anonBlockDevMinor = make(map[uint32]struct{})
+	vfs.fsTypes = make(map[string]*registeredFilesystemType)
+	vfs.filesystems = make(map[*Filesystem]struct{})
 	vfs.mounts.Init()
 
 	// Construct vfs.anonMount.
 	anonfsDevMinor, err := vfs.GetAnonBlockDevMinor()
 	if err != nil {
-		panic(fmt.Sprintf("VirtualFilesystem.GetAnonBlockDevMinor() failed during VirtualFilesystem construction: %v", err))
+		return err
 	}
 	anonfs := anonFilesystem{
 		devMinor: anonfsDevMinor,
@@ -137,8 +137,7 @@ func New() *VirtualFilesystem {
 		fs:   &anonfs.vfsfs,
 		refs: 1,
 	}
-
-	return vfs
+	return nil
 }
 
 // PathOperation specifies the path operated on by a VFS method.