Port devpts to VFS2.

PiperOrigin-RevId: 308164359

pkg/sentry/fs/tty/line_discipline.go

@@ -27,6 +27,8 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
+// LINT.IfChange
+
 const (
 	// canonMaxBytes is the number of bytes that fit into a single line of
 	// terminal input in canonical mode. This corresponds to N_TTY_BUF_SIZE
@@ -443,3 +445,5 @@ func (l *lineDiscipline) peek(b []byte) int {
 	}
 	return size
 }
+
+// LINT.ThenChange(../../fsimpl/devpts/line_discipline.go)

pkg/sentry/fs/tty/master.go

@@ -26,6 +26,8 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
+// LINT.IfChange
+
 // masterInodeOperations are the fs.InodeOperations for the master end of the
 // Terminal (ptmx file).
 //
@@ -232,3 +234,5 @@ func maybeEmitUnimplementedEvent(ctx context.Context, cmd uint32) {
 		unimpl.EmitUnimplementedEvent(ctx)
 	}
 }
+
+// LINT.ThenChange(../../fsimpl/devpts/master.go)

pkg/sentry/fs/tty/queue.go

@@ -25,6 +25,8 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
+// LINT.IfChange
+
 // waitBufMaxBytes is the maximum size of a wait buffer. It is based on
 // TTYB_DEFAULT_MEM_LIMIT.
 const waitBufMaxBytes = 131072
@@ -234,3 +236,5 @@ func (q *queue) waitBufAppend(b []byte) {
 	q.waitBuf = append(q.waitBuf, b)
 	q.waitBufLen += uint64(len(b))
 }
+
+// LINT.ThenChange(../../fsimpl/devpts/queue.go)

pkg/sentry/fs/tty/slave.go

@@ -25,6 +25,8 @@ import (
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
+// LINT.IfChange
+
 // slaveInodeOperations are the fs.InodeOperations for the slave end of the
 // Terminal (pts file).
 //
@@ -172,3 +174,5 @@ func (sf *slaveFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem
 		return 0, syserror.ENOTTY
 	}
 }
+
+// LINT.ThenChange(../../fsimpl/devpts/slave.go)

pkg/sentry/fs/tty/terminal.go

@@ -23,6 +23,8 @@ import (
 	"gvisor.dev/gvisor/pkg/usermem"
 )
 
+// LINT.IfChange
+
 // Terminal is a pseudoterminal.
 //
 // +stateify savable
@@ -126,3 +128,5 @@ func (tm *Terminal) tty(isMaster bool) *kernel.TTY {
 	}
 	return tm.slaveKTTY
 }
+
+// LINT.ThenChange(../../fsimpl/devpts/terminal.go)

pkg/sentry/fsimpl/devpts/BUILD

@@ -0,0 +1,43 @@
+load("//tools:defs.bzl", "go_library", "go_test")
+
+licenses(["notice"])
+
+go_library(
+    name = "devpts",
+    srcs = [
+        "devpts.go",
+        "line_discipline.go",
+        "master.go",
+        "queue.go",
+        "slave.go",
+        "terminal.go",
+    ],
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/context",
+        "//pkg/safemem",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/fsimpl/kernfs",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/unimpl",
+        "//pkg/sentry/vfs",
+        "//pkg/sync",
+        "//pkg/syserror",
+        "//pkg/usermem",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "devpts_test",
+    size = "small",
+    srcs = ["devpts_test.go"],
+    library = ":devpts",
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/contexttest",
+        "//pkg/usermem",
+    ],
+)

pkg/sentry/fsimpl/devpts/devpts.go

@@ -0,0 +1,207 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package devpts provides a filesystem implementation that behaves like
+// devpts.
+package devpts
+
+import (
+	"fmt"
+	"math"
+	"sort"
+	"strconv"
+	"sync"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// Name is the filesystem name.
+const Name = "devpts"
+
+// FilesystemType implements vfs.FilesystemType.
+type FilesystemType struct{}
+
+// Name implements vfs.FilesystemType.Name.
+func (FilesystemType) Name() string {
+	return Name
+}
+
+var _ vfs.FilesystemType = (*FilesystemType)(nil)
+
+// GetFilesystem implements vfs.FilesystemType.GetFilesystem.
+func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) {
+	// No data allowed.
+	if opts.Data != "" {
+		return nil, nil, syserror.EINVAL
+	}
+
+	fs, root := fstype.newFilesystem(vfsObj, creds)
+	return fs.VFSFilesystem(), root.VFSDentry(), nil
+}
+
+// newFilesystem creates a new devpts filesystem with root directory and ptmx
+// master inode. It returns the filesystem and root Dentry.
+func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*kernfs.Filesystem, *kernfs.Dentry) {
+	fs := &kernfs.Filesystem{}
+	fs.Init(vfsObj, fstype)
+
+	// Construct the root directory. This is always inode id 1.
+	root := &rootInode{
+		slaves: make(map[uint32]*slaveInode),
+	}
+	root.InodeAttrs.Init(creds, 1, linux.ModeDirectory|0555)
+	root.OrderedChildren.Init(kernfs.OrderedChildrenOptions{})
+	root.dentry.Init(root)
+
+	// Construct the pts master inode and dentry. Linux always uses inode
+	// id 2 for ptmx. See fs/devpts/inode.c:mknod_ptmx.
+	master := &masterInode{
+		root: root,
+	}
+	master.InodeAttrs.Init(creds, 2, linux.ModeCharacterDevice|0666)
+	master.dentry.Init(master)
+
+	// Add the master as a child of the root.
+	links := root.OrderedChildren.Populate(&root.dentry, map[string]*kernfs.Dentry{
+		"ptmx": &master.dentry,
+	})
+	root.IncLinks(links)
+
+	return fs, &root.dentry
+}
+
+// rootInode is the root directory inode for the devpts mounts.
+type rootInode struct {
+	kernfs.AlwaysValid
+	kernfs.InodeAttrs
+	kernfs.InodeDirectoryNoNewChildren
+	kernfs.InodeNotSymlink
+	kernfs.OrderedChildren
+
+	// Keep a reference to this inode's dentry.
+	dentry kernfs.Dentry
+
+	// master is the master pty inode. Immutable.
+	master *masterInode
+
+	// root is the root directory inode for this filesystem. Immutable.
+	root *rootInode
+
+	// mu protects the fields below.
+	mu sync.Mutex
+
+	// slaves maps pty ids to slave inodes.
+	slaves map[uint32]*slaveInode
+
+	// nextIdx is the next pty index to use. Must be accessed atomically.
+	//
+	// TODO(b/29356795): reuse indices when ptys are closed.
+	nextIdx uint32
+}
+
+var _ kernfs.Inode = (*rootInode)(nil)
+
+// allocateTerminal creates a new Terminal and installs a pts node for it.
+func (i *rootInode) allocateTerminal(creds *auth.Credentials) (*Terminal, error) {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	if i.nextIdx == math.MaxUint32 {
+		return nil, syserror.ENOMEM
+	}
+	idx := i.nextIdx
+	i.nextIdx++
+
+	// Sanity check that slave with idx does not exist.
+	if _, ok := i.slaves[idx]; ok {
+		panic(fmt.Sprintf("pty index collision; index %d already exists", idx))
+	}
+
+	// Create the new terminal and slave.
+	t := newTerminal(idx)
+	slave := &slaveInode{
+		root: i,
+		t:    t,
+	}
+	// Linux always uses pty index + 3 as the inode id. See
+	// fs/devpts/inode.c:devpts_pty_new().
+	slave.InodeAttrs.Init(creds, uint64(idx+3), linux.ModeCharacterDevice|0600)
+	slave.dentry.Init(slave)
+	i.slaves[idx] = slave
+
+	return t, nil
+}
+
+// masterClose is called when the master end of t is closed.
+func (i *rootInode) masterClose(t *Terminal) {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
+	// Sanity check that slave with idx exists.
+	if _, ok := i.slaves[t.n]; !ok {
+		panic(fmt.Sprintf("pty with index %d does not exist", t.n))
+	}
+	delete(i.slaves, t.n)
+}
+
+// Open implements kernfs.Inode.Open.
+func (i *rootInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) {
+	fd := &kernfs.GenericDirectoryFD{}
+	fd.Init(rp.Mount(), vfsd, &i.OrderedChildren, &opts)
+	return fd.VFSFileDescription(), nil
+}
+
+// Lookup implements kernfs.Inode.Lookup.
+func (i *rootInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) {
+	idx, err := strconv.ParseUint(name, 10, 32)
+	if err != nil {
+		return nil, syserror.ENOENT
+	}
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	if si, ok := i.slaves[uint32(idx)]; ok {
+		si.dentry.IncRef()
+		return si.dentry.VFSDentry(), nil
+
+	}
+	return nil, syserror.ENOENT
+}
+
+// IterDirents implements kernfs.Inode.IterDirents.
+func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	ids := make([]int, 0, len(i.slaves))
+	for id := range i.slaves {
+		ids = append(ids, int(id))
+	}
+	sort.Ints(ids)
+	for _, id := range ids[relOffset:] {
+		dirent := vfs.Dirent{
+			Name:    strconv.FormatUint(uint64(id), 10),
+			Type:    linux.DT_CHR,
+			Ino:     i.slaves[uint32(id)].InodeAttrs.Ino(),
+			NextOff: offset + 1,
+		}
+		if err := cb.Handle(dirent); err != nil {
+			return offset, err
+		}
+		offset++
+	}
+	return offset, nil
+}

pkg/sentry/fsimpl/devpts/devpts_test.go

@@ -0,0 +1,56 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devpts
+
+import (
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/contexttest"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+func TestSimpleMasterToSlave(t *testing.T) {
+	ld := newLineDiscipline(linux.DefaultSlaveTermios)
+	ctx := contexttest.Context(t)
+	inBytes := []byte("hello, tty\n")
+	src := usermem.BytesIOSequence(inBytes)
+	outBytes := make([]byte, 32)
+	dst := usermem.BytesIOSequence(outBytes)
+
+	// Write to the input queue.
+	nw, err := ld.inputQueueWrite(ctx, src)
+	if err != nil {
+		t.Fatalf("error writing to input queue: %v", err)
+	}
+	if nw != int64(len(inBytes)) {
+		t.Fatalf("wrote wrong length: got %d, want %d", nw, len(inBytes))
+	}
+
+	// Read from the input queue.
+	nr, err := ld.inputQueueRead(ctx, dst)
+	if err != nil {
+		t.Fatalf("error reading from input queue: %v", err)
+	}
+	if nr != int64(len(inBytes)) {
+		t.Fatalf("read wrong length: got %d, want %d", nr, len(inBytes))
+	}
+
+	outStr := string(outBytes[:nr])
+	inStr := string(inBytes)
+	if outStr != inStr {
+		t.Fatalf("written and read strings do not match: got %q, want %q", outStr, inStr)
+	}
+}

pkg/sentry/fsimpl/devpts/line_discipline.go

@@ -0,0 +1,449 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devpts
+
+import (
+	"bytes"
+	"unicode/utf8"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// LINT.IfChange
+
+const (
+	// canonMaxBytes is the number of bytes that fit into a single line of
+	// terminal input in canonical mode. This corresponds to N_TTY_BUF_SIZE
+	// in include/linux/tty.h.
+	canonMaxBytes = 4096
+
+	// nonCanonMaxBytes is the maximum number of bytes that can be read at
+	// a time in noncanonical mode.
+	nonCanonMaxBytes = canonMaxBytes - 1
+
+	spacesPerTab = 8
+)
+
+// lineDiscipline dictates how input and output are handled between the
+// pseudoterminal (pty) master and slave. It can be configured to alter I/O,
+// modify control characters (e.g. Ctrl-C for SIGINT), etc. The following man
+// pages are good resources for how to affect the line discipline:
+//
+//   * termios(3)
+//   * tty_ioctl(4)
+//
+// This file corresponds most closely to drivers/tty/n_tty.c.
+//
+// lineDiscipline has a simple structure but supports a multitude of options
+// (see the above man pages). It consists of two queues of bytes: one from the
+// terminal master to slave (the input queue) and one from slave to master (the
+// output queue). When bytes are written to one end of the pty, the line
+// discipline reads the bytes, modifies them or takes special action if
+// required, and enqueues them to be read by the other end of the pty:
+//
+//       input from terminal    +-------------+   input to process (e.g. bash)
+//    +------------------------>| input queue |---------------------------+
+//    |   (inputQueueWrite)     +-------------+     (inputQueueRead)      |
+//    |                                                                   |
+//    |                                                                   v
+// masterFD                                                            slaveFD
+//    ^                                                                   |
+//    |                                                                   |
+//    |   output to terminal   +--------------+    output from process    |
+//    +------------------------| output queue |<--------------------------+
+//        (outputQueueRead)    +--------------+    (outputQueueWrite)
+//
+// Lock order:
+//  termiosMu
+//    inQueue.mu
+//      outQueue.mu
+//
+// +stateify savable
+type lineDiscipline struct {
+	// sizeMu protects size.
+	sizeMu sync.Mutex `state:"nosave"`
+
+	// size is the terminal size (width and height).
+	size linux.WindowSize
+
+	// inQueue is the input queue of the terminal.
+	inQueue queue
+
+	// outQueue is the output queue of the terminal.
+	outQueue queue
+
+	// termiosMu protects termios.
+	termiosMu sync.RWMutex `state:"nosave"`
+
+	// termios is the terminal configuration used by the lineDiscipline.
+	termios linux.KernelTermios
+
+	// column is the location in a row of the cursor. This is important for
+	// handling certain special characters like backspace.
+	column int
+
+	// masterWaiter is used to wait on the master end of the TTY.
+	masterWaiter waiter.Queue `state:"zerovalue"`
+
+	// slaveWaiter is used to wait on the slave end of the TTY.
+	slaveWaiter waiter.Queue `state:"zerovalue"`
+}
+
+func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline {
+	ld := lineDiscipline{termios: termios}
+	ld.inQueue.transformer = &inputQueueTransformer{}
+	ld.outQueue.transformer = &outputQueueTransformer{}
+	return &ld
+}
+
+// getTermios gets the linux.Termios for the tty.
+func (l *lineDiscipline) getTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	l.termiosMu.RLock()
+	defer l.termiosMu.RUnlock()
+	// We must copy a Termios struct, not KernelTermios.
+	t := l.termios.ToTermios()
+	_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), t, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	return 0, err
+}
+
+// setTermios sets a linux.Termios for the tty.
+func (l *lineDiscipline) setTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	l.termiosMu.Lock()
+	defer l.termiosMu.Unlock()
+	oldCanonEnabled := l.termios.LEnabled(linux.ICANON)
+	// We must copy a Termios struct, not KernelTermios.
+	var t linux.Termios
+	_, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &t, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	l.termios.FromTermios(t)
+
+	// If canonical mode is turned off, move bytes from inQueue's wait
+	// buffer to its read buffer. Anything already in the read buffer is
+	// now readable.
+	if oldCanonEnabled && !l.termios.LEnabled(linux.ICANON) {
+		l.inQueue.mu.Lock()
+		l.inQueue.pushWaitBufLocked(l)
+		l.inQueue.readable = true
+		l.inQueue.mu.Unlock()
+		l.slaveWaiter.Notify(waiter.EventIn)
+	}
+
+	return 0, err
+}
+
+func (l *lineDiscipline) windowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error {
+	l.sizeMu.Lock()
+	defer l.sizeMu.Unlock()
+	_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), l.size, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	return err
+}
+
+func (l *lineDiscipline) setWindowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error {
+	l.sizeMu.Lock()
+	defer l.sizeMu.Unlock()
+	_, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &l.size, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	return err
+}
+
+func (l *lineDiscipline) masterReadiness() waiter.EventMask {
+	// We don't have to lock a termios because the default master termios
+	// is immutable.
+	return l.inQueue.writeReadiness(&linux.MasterTermios) | l.outQueue.readReadiness(&linux.MasterTermios)
+}
+
+func (l *lineDiscipline) slaveReadiness() waiter.EventMask {
+	l.termiosMu.RLock()
+	defer l.termiosMu.RUnlock()
+	return l.outQueue.writeReadiness(&l.termios) | l.inQueue.readReadiness(&l.termios)
+}
+
+func (l *lineDiscipline) inputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error {
+	return l.inQueue.readableSize(ctx, io, args)
+}
+
+func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) {
+	l.termiosMu.RLock()
+	defer l.termiosMu.RUnlock()
+	n, pushed, err := l.inQueue.read(ctx, dst, l)
+	if err != nil {
+		return 0, err
+	}
+	if n > 0 {
+		l.masterWaiter.Notify(waiter.EventOut)
+		if pushed {
+			l.slaveWaiter.Notify(waiter.EventIn)
+		}
+		return n, nil
+	}
+	return 0, syserror.ErrWouldBlock
+}
+
+func (l *lineDiscipline) inputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) {
+	l.termiosMu.RLock()
+	defer l.termiosMu.RUnlock()
+	n, err := l.inQueue.write(ctx, src, l)
+	if err != nil {
+		return 0, err
+	}
+	if n > 0 {
+		l.slaveWaiter.Notify(waiter.EventIn)
+		return n, nil
+	}
+	return 0, syserror.ErrWouldBlock
+}
+
+func (l *lineDiscipline) outputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error {
+	return l.outQueue.readableSize(ctx, io, args)
+}
+
+func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) {
+	l.termiosMu.RLock()
+	defer l.termiosMu.RUnlock()
+	n, pushed, err := l.outQueue.read(ctx, dst, l)
+	if err != nil {
+		return 0, err
+	}
+	if n > 0 {
+		l.slaveWaiter.Notify(waiter.EventOut)
+		if pushed {
+			l.masterWaiter.Notify(waiter.EventIn)
+		}
+		return n, nil
+	}
+	return 0, syserror.ErrWouldBlock
+}
+
+func (l *lineDiscipline) outputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) {
+	l.termiosMu.RLock()
+	defer l.termiosMu.RUnlock()
+	n, err := l.outQueue.write(ctx, src, l)
+	if err != nil {
+		return 0, err
+	}
+	if n > 0 {
+		l.masterWaiter.Notify(waiter.EventIn)
+		return n, nil
+	}
+	return 0, syserror.ErrWouldBlock
+}
+
+// transformer is a helper interface to make it easier to stateify queue.
+type transformer interface {
+	// transform functions require queue's mutex to be held.
+	transform(*lineDiscipline, *queue, []byte) int
+}
+
+// outputQueueTransformer implements transformer. It performs line discipline
+// transformations on the output queue.
+//
+// +stateify savable
+type outputQueueTransformer struct{}
+
+// transform does output processing for one end of the pty. See
+// drivers/tty/n_tty.c:do_output_char for an analogous kernel function.
+//
+// Preconditions:
+// * l.termiosMu must be held for reading.
+// * q.mu must be held.
+func (*outputQueueTransformer) transform(l *lineDiscipline, q *queue, buf []byte) int {
+	// transformOutput is effectively always in noncanonical mode, as the
+	// master termios never has ICANON set.
+
+	if !l.termios.OEnabled(linux.OPOST) {
+		q.readBuf = append(q.readBuf, buf...)
+		if len(q.readBuf) > 0 {
+			q.readable = true
+		}
+		return len(buf)
+	}
+
+	var ret int
+	for len(buf) > 0 {
+		size := l.peek(buf)
+		cBytes := append([]byte{}, buf[:size]...)
+		ret += size
+		buf = buf[size:]
+		// We're guaranteed that cBytes has at least one element.
+		switch cBytes[0] {
+		case '\n':
+			if l.termios.OEnabled(linux.ONLRET) {
+				l.column = 0
+			}
+			if l.termios.OEnabled(linux.ONLCR) {
+				q.readBuf = append(q.readBuf, '\r', '\n')
+				continue
+			}
+		case '\r':
+			if l.termios.OEnabled(linux.ONOCR) && l.column == 0 {
+				continue
+			}
+			if l.termios.OEnabled(linux.OCRNL) {
+				cBytes[0] = '\n'
+				if l.termios.OEnabled(linux.ONLRET) {
+					l.column = 0
+				}
+				break
+			}
+			l.column = 0
+		case '\t':
+			spaces := spacesPerTab - l.column%spacesPerTab
+			if l.termios.OutputFlags&linux.TABDLY == linux.XTABS {
+				l.column += spaces
+				q.readBuf = append(q.readBuf, bytes.Repeat([]byte{' '}, spacesPerTab)...)
+				continue
+			}
+			l.column += spaces
+		case '\b':
+			if l.column > 0 {
+				l.column--
+			}
+		default:
+			l.column++
+		}
+		q.readBuf = append(q.readBuf, cBytes...)
+	}
+	if len(q.readBuf) > 0 {
+		q.readable = true
+	}
+	return ret
+}
+
+// inputQueueTransformer implements transformer. It performs line discipline
+// transformations on the input queue.
+//
+// +stateify savable
+type inputQueueTransformer struct{}
+
+// transform does input processing for one end of the pty. Characters read are
+// transformed according to flags set in the termios struct. See
+// drivers/tty/n_tty.c:n_tty_receive_char_special for an analogous kernel
+// function.
+//
+// Preconditions:
+// * l.termiosMu must be held for reading.
+// * q.mu must be held.
+func (*inputQueueTransformer) transform(l *lineDiscipline, q *queue, buf []byte) int {
+	// If there's a line waiting to be read in canonical mode, don't write
+	// anything else to the read buffer.
+	if l.termios.LEnabled(linux.ICANON) && q.readable {
+		return 0
+	}
+
+	maxBytes := nonCanonMaxBytes
+	if l.termios.LEnabled(linux.ICANON) {
+		maxBytes = canonMaxBytes
+	}
+
+	var ret int
+	for len(buf) > 0 && len(q.readBuf) < canonMaxBytes {
+		size := l.peek(buf)
+		cBytes := append([]byte{}, buf[:size]...)
+		// We're guaranteed that cBytes has at least one element.
+		switch cBytes[0] {
+		case '\r':
+			if l.termios.IEnabled(linux.IGNCR) {
+				buf = buf[size:]
+				ret += size
+				continue
+			}
+			if l.termios.IEnabled(linux.ICRNL) {
+				cBytes[0] = '\n'
+			}
+		case '\n':
+			if l.termios.IEnabled(linux.INLCR) {
+				cBytes[0] = '\r'
+			}
+		}
+
+		// In canonical mode, we discard non-terminating characters
+		// after the first 4095.
+		if l.shouldDiscard(q, cBytes) {
+			buf = buf[size:]
+			ret += size
+			continue
+		}
+
+		// Stop if the buffer would be overfilled.
+		if len(q.readBuf)+size > maxBytes {
+			break
+		}
+		buf = buf[size:]
+		ret += size
+
+		// If we get EOF, make the buffer available for reading.
+		if l.termios.LEnabled(linux.ICANON) && l.termios.IsEOF(cBytes[0]) {
+			q.readable = true
+			break
+		}
+
+		q.readBuf = append(q.readBuf, cBytes...)
+
+		// Anything written to the readBuf will have to be echoed.
+		if l.termios.LEnabled(linux.ECHO) {
+			l.outQueue.writeBytes(cBytes, l)
+			l.masterWaiter.Notify(waiter.EventIn)
+		}
+
+		// If we finish a line, make it available for reading.
+		if l.termios.LEnabled(linux.ICANON) && l.termios.IsTerminating(cBytes) {
+			q.readable = true
+			break
+		}
+	}
+
+	// In noncanonical mode, everything is readable.
+	if !l.termios.LEnabled(linux.ICANON) && len(q.readBuf) > 0 {
+		q.readable = true
+	}
+
+	return ret
+}
+
+// shouldDiscard returns whether c should be discarded. In canonical mode, if
+// too many bytes are enqueued, we keep reading input and discarding it until
+// we find a terminating character. Signal/echo processing still occurs.
+//
+// Precondition:
+// * l.termiosMu must be held for reading.
+// * q.mu must be held.
+func (l *lineDiscipline) shouldDiscard(q *queue, cBytes []byte) bool {
+	return l.termios.LEnabled(linux.ICANON) && len(q.readBuf)+len(cBytes) >= canonMaxBytes && !l.termios.IsTerminating(cBytes)
+}
+
+// peek returns the size in bytes of the next character to process. As long as
+// b isn't empty, peek returns a value of at least 1.
+func (l *lineDiscipline) peek(b []byte) int {
+	size := 1
+	// If UTF-8 support is enabled, runes might be multiple bytes.
+	if l.termios.IEnabled(linux.IUTF8) {
+		_, size = utf8.DecodeRune(b)
+	}
+	return size
+}
+
+// LINT.ThenChange(../../fs/tty/line_discipline.go)

pkg/sentry/fsimpl/devpts/master.go

@@ -0,0 +1,226 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devpts
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/unimpl"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// LINT.IfChange
+
+// masterInode is the inode for the master end of the Terminal.
+type masterInode struct {
+	kernfs.InodeAttrs
+	kernfs.InodeNoopRefCount
+	kernfs.InodeNotDirectory
+	kernfs.InodeNotSymlink
+
+	// Keep a reference to this inode's dentry.
+	dentry kernfs.Dentry
+
+	// root is the devpts root inode.
+	root *rootInode
+}
+
+var _ kernfs.Inode = (*masterInode)(nil)
+
+// Open implements kernfs.Inode.Open.
+func (mi *masterInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) {
+	t, err := mi.root.allocateTerminal(rp.Credentials())
+	if err != nil {
+		return nil, err
+	}
+
+	mi.IncRef()
+	fd := &masterFileDescription{
+		inode: mi,
+		t:     t,
+	}
+	if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+		mi.DecRef()
+		return nil, err
+	}
+	return &fd.vfsfd, nil
+}
+
+// Stat implements kernfs.Inode.Stat.
+func (mi *masterInode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) {
+	statx, err := mi.InodeAttrs.Stat(vfsfs, opts)
+	if err != nil {
+		return linux.Statx{}, err
+	}
+	statx.Blksize = 1024
+	statx.RdevMajor = linux.TTYAUX_MAJOR
+	statx.RdevMinor = linux.PTMX_MINOR
+	return statx, nil
+}
+
+// SetStat implements kernfs.Inode.SetStat
+func (mi *masterInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error {
+	if opts.Stat.Mask&linux.STATX_SIZE != 0 {
+		return syserror.EINVAL
+	}
+	return mi.InodeAttrs.SetStat(ctx, vfsfs, creds, opts)
+}
+
+type masterFileDescription struct {
+	vfsfd vfs.FileDescription
+	vfs.FileDescriptionDefaultImpl
+
+	inode *masterInode
+	t     *Terminal
+}
+
+var _ vfs.FileDescriptionImpl = (*masterFileDescription)(nil)
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (mfd *masterFileDescription) Release() {
+	mfd.inode.root.masterClose(mfd.t)
+	mfd.inode.DecRef()
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (mfd *masterFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	mfd.t.ld.masterWaiter.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (mfd *masterFileDescription) EventUnregister(e *waiter.Entry) {
+	mfd.t.ld.masterWaiter.EventUnregister(e)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (mfd *masterFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return mfd.t.ld.masterReadiness()
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (mfd *masterFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) {
+	return mfd.t.ld.outputQueueRead(ctx, dst)
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (mfd *masterFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) {
+	return mfd.t.ld.inputQueueWrite(ctx, src)
+}
+
+// Ioctl implements vfs.FileDescriptionImpl.Ioctl.
+func (mfd *masterFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch cmd := args[1].Uint(); cmd {
+	case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ
+		// Get the number of bytes in the output queue read buffer.
+		return 0, mfd.t.ld.outputQueueReadSize(ctx, io, args)
+	case linux.TCGETS:
+		// N.B. TCGETS on the master actually returns the configuration
+		// of the slave end.
+		return mfd.t.ld.getTermios(ctx, io, args)
+	case linux.TCSETS:
+		// N.B. TCSETS on the master actually affects the configuration
+		// of the slave end.
+		return mfd.t.ld.setTermios(ctx, io, args)
+	case linux.TCSETSW:
+		// TODO(b/29356795): This should drain the output queue first.
+		return mfd.t.ld.setTermios(ctx, io, args)
+	case linux.TIOCGPTN:
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(mfd.t.n), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	case linux.TIOCSPTLCK:
+		// TODO(b/29356795): Implement pty locking. For now just pretend we do.
+		return 0, nil
+	case linux.TIOCGWINSZ:
+		return 0, mfd.t.ld.windowSize(ctx, io, args)
+	case linux.TIOCSWINSZ:
+		return 0, mfd.t.ld.setWindowSize(ctx, io, args)
+	case linux.TIOCSCTTY:
+		// Make the given terminal the controlling terminal of the
+		// calling process.
+		return 0, mfd.t.setControllingTTY(ctx, io, args, true /* isMaster */)
+	case linux.TIOCNOTTY:
+		// Release this process's controlling terminal.
+		return 0, mfd.t.releaseControllingTTY(ctx, io, args, true /* isMaster */)
+	case linux.TIOCGPGRP:
+		// Get the foreground process group.
+		return mfd.t.foregroundProcessGroup(ctx, io, args, true /* isMaster */)
+	case linux.TIOCSPGRP:
+		// Set the foreground process group.
+		return mfd.t.setForegroundProcessGroup(ctx, io, args, true /* isMaster */)
+	default:
+		maybeEmitUnimplementedEvent(ctx, cmd)
+		return 0, syserror.ENOTTY
+	}
+}
+
+// SetStat implements vfs.FileDescriptionImpl.SetStat.
+func (mfd *masterFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
+	creds := auth.CredentialsFromContext(ctx)
+	fs := mfd.vfsfd.VirtualDentry().Mount().Filesystem()
+	return mfd.inode.SetStat(ctx, fs, creds, opts)
+}
+
+// Stat implements vfs.FileDescriptionImpl.Stat.
+func (mfd *masterFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
+	fs := mfd.vfsfd.VirtualDentry().Mount().Filesystem()
+	return mfd.inode.Stat(fs, opts)
+}
+
+// maybeEmitUnimplementedEvent emits unimplemented event if cmd is valid.
+func maybeEmitUnimplementedEvent(ctx context.Context, cmd uint32) {
+	switch cmd {
+	case linux.TCGETS,
+		linux.TCSETS,
+		linux.TCSETSW,
+		linux.TCSETSF,
+		linux.TIOCGWINSZ,
+		linux.TIOCSWINSZ,
+		linux.TIOCSETD,
+		linux.TIOCSBRK,
+		linux.TIOCCBRK,
+		linux.TCSBRK,
+		linux.TCSBRKP,
+		linux.TIOCSTI,
+		linux.TIOCCONS,
+		linux.FIONBIO,
+		linux.TIOCEXCL,
+		linux.TIOCNXCL,
+		linux.TIOCGEXCL,
+		linux.TIOCGSID,
+		linux.TIOCGETD,
+		linux.TIOCVHANGUP,
+		linux.TIOCGDEV,
+		linux.TIOCMGET,
+		linux.TIOCMSET,
+		linux.TIOCMBIC,
+		linux.TIOCMBIS,
+		linux.TIOCGICOUNT,
+		linux.TCFLSH,
+		linux.TIOCSSERIAL,
+		linux.TIOCGPTPEER:
+
+		unimpl.EmitUnimplementedEvent(ctx)
+	}
+}
+
+// LINT.ThenChange(../../fs/tty/master.go)

pkg/sentry/fsimpl/devpts/queue.go

@@ -0,0 +1,240 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devpts
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/safemem"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// LINT.IfChange
+
+// waitBufMaxBytes is the maximum size of a wait buffer. It is based on
+// TTYB_DEFAULT_MEM_LIMIT.
+const waitBufMaxBytes = 131072
+
+// queue represents one of the input or output queues between a pty master and
+// slave. Bytes written to a queue are added to the read buffer until it is
+// full, at which point they are written to the wait buffer. Bytes are
+// processed (i.e. undergo termios transformations) as they are added to the
+// read buffer. The read buffer is readable when its length is nonzero and
+// readable is true.
+//
+// +stateify savable
+type queue struct {
+	// mu protects everything in queue.
+	mu sync.Mutex `state:"nosave"`
+
+	// readBuf is buffer of data ready to be read when readable is true.
+	// This data has been processed.
+	readBuf []byte
+
+	// waitBuf contains data that can't fit into readBuf. It is put here
+	// until it can be loaded into the read buffer. waitBuf contains data
+	// that hasn't been processed.
+	waitBuf    [][]byte
+	waitBufLen uint64
+
+	// readable indicates whether the read buffer can be read from.  In
+	// canonical mode, there can be an unterminated line in the read buffer,
+	// so readable must be checked.
+	readable bool
+
+	// transform is the the queue's function for transforming bytes
+	// entering the queue. For example, transform might convert all '\r's
+	// entering the queue to '\n's.
+	transformer
+}
+
+// readReadiness returns whether q is ready to be read from.
+func (q *queue) readReadiness(t *linux.KernelTermios) waiter.EventMask {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	if len(q.readBuf) > 0 && q.readable {
+		return waiter.EventIn
+	}
+	return waiter.EventMask(0)
+}
+
+// writeReadiness returns whether q is ready to be written to.
+func (q *queue) writeReadiness(t *linux.KernelTermios) waiter.EventMask {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	if q.waitBufLen < waitBufMaxBytes {
+		return waiter.EventOut
+	}
+	return waiter.EventMask(0)
+}
+
+// readableSize writes the number of readable bytes to userspace.
+func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	var size int32
+	if q.readable {
+		size = int32(len(q.readBuf))
+	}
+
+	_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), size, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	return err
+
+}
+
+// read reads from q to userspace. It returns the number of bytes read as well
+// as whether the read caused more readable data to become available (whether
+// data was pushed from the wait buffer to the read buffer).
+//
+// Preconditions:
+// * l.termiosMu must be held for reading.
+func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipline) (int64, bool, error) {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+
+	if !q.readable {
+		return 0, false, syserror.ErrWouldBlock
+	}
+
+	if dst.NumBytes() > canonMaxBytes {
+		dst = dst.TakeFirst(canonMaxBytes)
+	}
+
+	n, err := dst.CopyOutFrom(ctx, safemem.ReaderFunc(func(dst safemem.BlockSeq) (uint64, error) {
+		src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(q.readBuf))
+		n, err := safemem.CopySeq(dst, src)
+		if err != nil {
+			return 0, err
+		}
+		q.readBuf = q.readBuf[n:]
+
+		// If we read everything, this queue is no longer readable.
+		if len(q.readBuf) == 0 {
+			q.readable = false
+		}
+
+		return n, nil
+	}))
+	if err != nil {
+		return 0, false, err
+	}
+
+	// Move data from the queue's wait buffer to its read buffer.
+	nPushed := q.pushWaitBufLocked(l)
+
+	return int64(n), nPushed > 0, nil
+}
+
+// write writes to q from userspace.
+//
+// Preconditions:
+// * l.termiosMu must be held for reading.
+func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscipline) (int64, error) {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+
+	// Copy data into the wait buffer.
+	n, err := src.CopyInTo(ctx, safemem.WriterFunc(func(src safemem.BlockSeq) (uint64, error) {
+		copyLen := src.NumBytes()
+		room := waitBufMaxBytes - q.waitBufLen
+		// If out of room, return EAGAIN.
+		if room == 0 && copyLen > 0 {
+			return 0, syserror.ErrWouldBlock
+		}
+		// Cap the size of the wait buffer.
+		if copyLen > room {
+			copyLen = room
+			src = src.TakeFirst64(room)
+		}
+		buf := make([]byte, copyLen)
+
+		// Copy the data into the wait buffer.
+		dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf))
+		n, err := safemem.CopySeq(dst, src)
+		if err != nil {
+			return 0, err
+		}
+		q.waitBufAppend(buf)
+
+		return n, nil
+	}))
+	if err != nil {
+		return 0, err
+	}
+
+	// Push data from the wait to the read buffer.
+	q.pushWaitBufLocked(l)
+
+	return n, nil
+}
+
+// writeBytes writes to q from b.
+//
+// Preconditions:
+// * l.termiosMu must be held for reading.
+func (q *queue) writeBytes(b []byte, l *lineDiscipline) {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+
+	// Write to the wait buffer.
+	q.waitBufAppend(b)
+	q.pushWaitBufLocked(l)
+}
+
+// pushWaitBufLocked fills the queue's read buffer with data from the wait
+// buffer.
+//
+// Preconditions:
+// * l.termiosMu must be held for reading.
+// * q.mu must be locked.
+func (q *queue) pushWaitBufLocked(l *lineDiscipline) int {
+	if q.waitBufLen == 0 {
+		return 0
+	}
+
+	// Move data from the wait to the read buffer.
+	var total int
+	var i int
+	for i = 0; i < len(q.waitBuf); i++ {
+		n := q.transform(l, q, q.waitBuf[i])
+		total += n
+		if n != len(q.waitBuf[i]) {
+			// The read buffer filled up without consuming the
+			// entire buffer.
+			q.waitBuf[i] = q.waitBuf[i][n:]
+			break
+		}
+	}
+
+	// Update wait buffer based on consumed data.
+	q.waitBuf = q.waitBuf[i:]
+	q.waitBufLen -= uint64(total)
+
+	return total
+}
+
+// Precondition: q.mu must be locked.
+func (q *queue) waitBufAppend(b []byte) {
+	q.waitBuf = append(q.waitBuf, b)
+	q.waitBufLen += uint64(len(b))
+}
+
+// LINT.ThenChange(../../fs/tty/queue.go)

pkg/sentry/fsimpl/devpts/slave.go

@@ -0,0 +1,186 @@
+// Copyright 2020 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devpts
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// LINT.IfChange
+
+// slaveInode is the inode for the slave end of the Terminal.
+type slaveInode struct {
+	kernfs.InodeAttrs
+	kernfs.InodeNoopRefCount
+	kernfs.InodeNotDirectory
+	kernfs.InodeNotSymlink
+
+	// Keep a reference to this inode's dentry.
+	dentry kernfs.Dentry
+
+	// root is the devpts root inode.
+	root *rootInode
+
+	// t is the connected Terminal.
+	t *Terminal
+}
+
+var _ kernfs.Inode = (*slaveInode)(nil)
+
+// Open implements kernfs.Inode.Open.
+func (si *slaveInode) Open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) {
+	si.IncRef()
+	fd := &slaveFileDescription{
+		inode: si,
+	}
+	if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil {
+		si.DecRef()
+		return nil, err
+	}
+	return &fd.vfsfd, nil
+
+}
+
+// Valid implements kernfs.Inode.Valid.
+func (si *slaveInode) Valid(context.Context) bool {
+	// Return valid if the slave still exists.
+	si.root.mu.Lock()
+	defer si.root.mu.Unlock()
+	_, ok := si.root.slaves[si.t.n]
+	return ok
+}
+
+// Stat implements kernfs.Inode.Stat.
+func (si *slaveInode) Stat(vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) {
+	statx, err := si.InodeAttrs.Stat(vfsfs, opts)
+	if err != nil {
+		return linux.Statx{}, err
+	}
+	statx.Blksize = 1024
+	statx.RdevMajor = linux.UNIX98_PTY_SLAVE_MAJOR
+	statx.RdevMinor = si.t.n
+	return statx, nil
+}
+
+// SetStat implements kernfs.Inode.SetStat
+func (si *slaveInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error {
+	if opts.Stat.Mask&linux.STATX_SIZE != 0 {
+		return syserror.EINVAL
+	}
+	return si.InodeAttrs.SetStat(ctx, vfsfs, creds, opts)
+}
+
+type slaveFileDescription struct {
+	vfsfd vfs.FileDescription
+	vfs.FileDescriptionDefaultImpl
+
+	inode *slaveInode
+}
+
+var _ vfs.FileDescriptionImpl = (*slaveFileDescription)(nil)
+
+// Release implements fs.FileOperations.Release.
+func (sfd *slaveFileDescription) Release() {
+	sfd.inode.DecRef()
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (sfd *slaveFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	sfd.inode.t.ld.slaveWaiter.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (sfd *slaveFileDescription) EventUnregister(e *waiter.Entry) {
+	sfd.inode.t.ld.slaveWaiter.EventUnregister(e)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (sfd *slaveFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return sfd.inode.t.ld.slaveReadiness()
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (sfd *slaveFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) {
+	return sfd.inode.t.ld.inputQueueRead(ctx, dst)
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (sfd *slaveFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) {
+	return sfd.inode.t.ld.outputQueueWrite(ctx, src)
+}
+
+// Ioctl implements vfs.FileDescripionImpl.Ioctl.
+func (sfd *slaveFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch cmd := args[1].Uint(); cmd {
+	case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ
+		// Get the number of bytes in the input queue read buffer.
+		return 0, sfd.inode.t.ld.inputQueueReadSize(ctx, io, args)
+	case linux.TCGETS:
+		return sfd.inode.t.ld.getTermios(ctx, io, args)
+	case linux.TCSETS:
+		return sfd.inode.t.ld.setTermios(ctx, io, args)
+	case linux.TCSETSW:
+		// TODO(b/29356795): This should drain the output queue first.
+		return sfd.inode.t.ld.setTermios(ctx, io, args)
+	case linux.TIOCGPTN:
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(sfd.inode.t.n), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	case linux.TIOCGWINSZ:
+		return 0, sfd.inode.t.ld.windowSize(ctx, io, args)
+	case linux.TIOCSWINSZ:
+		return 0, sfd.inode.t.ld.setWindowSize(ctx, io, args)
+	case linux.TIOCSCTTY:
+		// Make the given terminal the controlling terminal of the
+		// calling process.
+		return 0, sfd.inode.t.setControllingTTY(ctx, io, args, false /* isMaster */)
+	case linux.TIOCNOTTY:
+		// Release this process's controlling terminal.
+		return 0, sfd.inode.t.releaseControllingTTY(ctx, io, args, false /* isMaster */)
+	case linux.TIOCGPGRP:
+		// Get the foreground process group.
+		return sfd.inode.t.foregroundProcessGroup(ctx, io, args, false /* isMaster */)
+	case linux.TIOCSPGRP:
+		// Set the foreground process group.
+		return sfd.inode.t.setForegroundProcessGroup(ctx, io, args, false /* isMaster */)
+	default:
+		maybeEmitUnimplementedEvent(ctx, cmd)
+		return 0, syserror.ENOTTY
+	}
+}
+
+// SetStat implements vfs.FileDescriptionImpl.SetStat.
+func (sfd *slaveFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
+	creds := auth.CredentialsFromContext(ctx)
+	fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem()
+	return sfd.inode.SetStat(ctx, fs, creds, opts)
+}
+
+// Stat implements vfs.FileDescriptionImpl.Stat.
+func (sfd *slaveFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
+	fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem()
+	return sfd.inode.Stat(fs, opts)
+}
+
+// LINT.ThenChange(../../fs/tty/slave.go)

pkg/sentry/fsimpl/devpts/terminal.go

@@ -0,0 +1,124 @@
+// Copyright 2018 The gVisor Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package devpts
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// LINT.IfChanges
+
+// Terminal is a pseudoterminal.
+//
+// +stateify savable
+type Terminal struct {
+	// n is the terminal index. It is immutable.
+	n uint32
+
+	// ld is the line discipline of the terminal. It is immutable.
+	ld *lineDiscipline
+
+	// masterKTTY contains the controlling process of the master end of
+	// this terminal. This field is immutable.
+	masterKTTY *kernel.TTY
+
+	// slaveKTTY contains the controlling process of the slave end of this
+	// terminal. This field is immutable.
+	slaveKTTY *kernel.TTY
+}
+
+func newTerminal(n uint32) *Terminal {
+	termios := linux.DefaultSlaveTermios
+	t := Terminal{
+		n:          n,
+		ld:         newLineDiscipline(termios),
+		masterKTTY: &kernel.TTY{Index: n},
+		slaveKTTY:  &kernel.TTY{Index: n},
+	}
+	return &t
+}
+
+// setControllingTTY makes tm the controlling terminal of the calling thread
+// group.
+func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error {
+	task := kernel.TaskFromContext(ctx)
+	if task == nil {
+		panic("setControllingTTY must be called from a task context")
+	}
+
+	return task.ThreadGroup().SetControllingTTY(tm.tty(isMaster), args[2].Int())
+}
+
+// releaseControllingTTY removes tm as the controlling terminal of the calling
+// thread group.
+func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error {
+	task := kernel.TaskFromContext(ctx)
+	if task == nil {
+		panic("releaseControllingTTY must be called from a task context")
+	}
+
+	return task.ThreadGroup().ReleaseControllingTTY(tm.tty(isMaster))
+}
+
+// foregroundProcessGroup gets the process group ID of tm's foreground process.
+func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) {
+	task := kernel.TaskFromContext(ctx)
+	if task == nil {
+		panic("foregroundProcessGroup must be called from a task context")
+	}
+
+	ret, err := task.ThreadGroup().ForegroundProcessGroup(tm.tty(isMaster))
+	if err != nil {
+		return 0, err
+	}
+
+	// Write it out to *arg.
+	_, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(ret), usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	return 0, err
+}
+
+// foregroundProcessGroup sets tm's foreground process.
+func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) {
+	task := kernel.TaskFromContext(ctx)
+	if task == nil {
+		panic("setForegroundProcessGroup must be called from a task context")
+	}
+
+	// Read in the process group ID.
+	var pgid int32
+	if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgid, usermem.IOOpts{
+		AddressSpaceActive: true,
+	}); err != nil {
+		return 0, err
+	}
+
+	ret, err := task.ThreadGroup().SetForegroundProcessGroup(tm.tty(isMaster), kernel.ProcessGroupID(pgid))
+	return uintptr(ret), err
+}
+
+func (tm *Terminal) tty(isMaster bool) *kernel.TTY {
+	if isMaster {
+		return tm.masterKTTY
+	}
+	return tm.slaveKTTY
+}
+
+// LINT.ThenChange(../../fs/tty/terminal.go)

pkg/sentry/fsimpl/devtmpfs/devtmpfs.go

@@ -163,16 +163,25 @@ func (a *Accessor) CreateDeviceFile(ctx context.Context, pathname string, kind v
 func (a *Accessor) UserspaceInit(ctx context.Context) error {
 	actx := a.wrapContext(ctx)
 
-	// systemd: src/shared/dev-setup.c:dev_setup()
+	// Initialize symlinks.
 	for _, symlink := range []struct {
 		source string
 		target string
 	}{
-		// /proc/kcore is not implemented.
+		// systemd: src/shared/dev-setup.c:dev_setup()
 		{source: "fd", target: "/proc/self/fd"},
 		{source: "stdin", target: "/proc/self/fd/0"},
 		{source: "stdout", target: "/proc/self/fd/1"},
 		{source: "stderr", target: "/proc/self/fd/2"},
+		// /proc/kcore is not implemented.
+
+		// Linux implements /dev/ptmx as a device node, but advises
+		// container implementations to create /dev/ptmx as a symlink
+		// to pts/ptmx (Documentation/filesystems/devpts.txt). Systemd
+		// follows this advice (src/nspawn/nspawn.c:setup_pts()), while
+		// LXC tries to create a bind mount and falls back to a symlink
+		// (src/lxc/conf.c:lxc_setup_devpts()).
+		{source: "ptmx", target: "pts/ptmx"},
 	} {
 		if err := a.vfsObj.SymlinkAt(actx, a.creds, a.pathOperationAt(symlink.source), symlink.target); err != nil {
 			return fmt.Errorf("failed to create symlink %q => %q: %v", symlink.source, symlink.target, err)

pkg/sentry/fsimpl/kernfs/filesystem.go

@@ -391,7 +391,7 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf
 	// O_NOFOLLOW have no effect here (they're handled by VFS by setting
 	// appropriate bits in rp), but are returned by
 	// FileDescriptionImpl.StatusFlags().
-	opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW
+	opts.Flags &= linux.O_ACCMODE | linux.O_CREAT | linux.O_EXCL | linux.O_TRUNC | linux.O_DIRECTORY | linux.O_NOFOLLOW | linux.O_NONBLOCK
 	ats := vfs.AccessTypesForOpenFlags(&opts)
 
 	// Do not create new file.

pkg/sentry/fsimpl/kernfs/inode_impl_util.go

@@ -216,6 +216,11 @@ func (a *InodeAttrs) Init(creds *auth.Credentials, ino uint64, mode linux.FileMo
 	atomic.StoreUint32(&a.nlink, nlink)
 }
 
+// Ino returns the inode id.
+func (a *InodeAttrs) Ino() uint64 {
+	return atomic.LoadUint64(&a.ino)
+}
+
 // Mode implements Inode.Mode.
 func (a *InodeAttrs) Mode() linux.FileMode {
 	return linux.FileMode(atomic.LoadUint32(&a.mode))
@@ -359,8 +364,8 @@ func (o *OrderedChildren) Destroy() {
 // cache. Populate returns the number of directories inserted, which the caller
 // may use to update the link count for the parent directory.
 //
-// Precondition: d.Impl() must be a kernfs Dentry. d must represent a directory
-// inode. children must not contain any conflicting entries already in o.
+// Precondition: d must represent a directory inode. children must not contain
+// any conflicting entries already in o.
 func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint32 {
 	var links uint32
 	for name, child := range children {