Port eventfd to VFS2.

And move sys_timerfd.go to just timerfd.go for consistency.

Updates #1475.

PiperOrigin-RevId: 309835029

8 files changed, 450 insertions, 15 deletions

diff --git a/pkg/sentry/syscalls/linux/sys_eventfd.go b/pkg/sentry/syscalls/linux/sys_eventfd.go
index 8a34c4e99..ed3413ca6 100644
--- a/pkg/sentry/syscalls/linux/sys_eventfd.go
+++ b/pkg/sentry/syscalls/linux/sys_eventfd.go
@@ -15,6 +15,7 @@
 package linux
 
 import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
@@ -22,32 +23,24 @@ import (
 	"gvisor.dev/gvisor/pkg/syserror"
 )
 
-const (
-	// EFD_SEMAPHORE is a flag used in syscall eventfd(2) and eventfd2(2). Please
-	// see its man page for more information.
-	EFD_SEMAPHORE = 1
-	EFD_NONBLOCK  = 0x800
-	EFD_CLOEXEC   = 0x80000
-)
-
 // Eventfd2 implements linux syscall eventfd2(2).
 func Eventfd2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
 	initVal := args[0].Int()
 	flags := uint(args[1].Uint())
-	allOps := uint(EFD_SEMAPHORE | EFD_NONBLOCK | EFD_CLOEXEC)
+	allOps := uint(linux.EFD_SEMAPHORE | linux.EFD_NONBLOCK | linux.EFD_CLOEXEC)
 
 	if flags & ^allOps != 0 {
 		return 0, nil, syserror.EINVAL
 	}
 
-	event := eventfd.New(t, uint64(initVal), flags&EFD_SEMAPHORE != 0)
+	event := eventfd.New(t, uint64(initVal), flags&linux.EFD_SEMAPHORE != 0)
 	event.SetFlags(fs.SettableFileFlags{
-		NonBlocking: flags&EFD_NONBLOCK != 0,
+		NonBlocking: flags&linux.EFD_NONBLOCK != 0,
 	})
 	defer event.DecRef()
 
 	fd, err := t.NewFDFrom(0, event, kernel.FDFlags{
-		CloseOnExec: flags&EFD_CLOEXEC != 0,
+		CloseOnExec: flags&linux.EFD_CLOEXEC != 0,
 	})
 	if err != nil {
 		return 0, nil, err
diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD
index f6fb0f219..ffca627d4 100644
--- a/pkg/sentry/syscalls/linux/vfs2/BUILD
+++ b/pkg/sentry/syscalls/linux/vfs2/BUILD
@@ -6,6 +6,7 @@ go_library(
     name = "vfs2",
     srcs = [
         "epoll.go",
+        "eventfd.go",
         "execve.go",
         "fd.go",
         "filesystem.go",
@@ -26,7 +27,7 @@ go_library(
         "stat_amd64.go",
         "stat_arm64.go",
         "sync.go",
-        "sys_timerfd.go",
+        "timerfd.go",
         "xattr.go",
     ],
     marshal = True,
diff --git a/pkg/sentry/syscalls/linux/vfs2/eventfd.go b/pkg/sentry/syscalls/linux/vfs2/eventfd.go
new file mode 100644
index 000000000..bd2194972
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/vfs2/eventfd.go
@@ -0,0 +1,59 @@
+// 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 vfs2
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// Eventfd2 implements linux syscall eventfd2(2).
+func Eventfd2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	initVal := uint64(args[0].Uint())
+	flags := uint(args[1].Uint())
+	allOps := uint(linux.EFD_SEMAPHORE | linux.EFD_NONBLOCK | linux.EFD_CLOEXEC)
+
+	if flags & ^allOps != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	fileFlags := uint32(linux.O_RDWR)
+	if flags&linux.EFD_NONBLOCK != 0 {
+		fileFlags |= linux.O_NONBLOCK
+	}
+	semMode := flags&linux.EFD_SEMAPHORE != 0
+	eventfd, err := t.Kernel().VFS().NewEventFD(initVal, semMode, fileFlags)
+	if err != nil {
+		return 0, nil, err
+	}
+	defer eventfd.DecRef()
+
+	fd, err := t.NewFDFromVFS2(0, eventfd, kernel.FDFlags{
+		CloseOnExec: flags&linux.EFD_CLOEXEC != 0,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// Eventfd implements linux syscall eventfd(2).
+func Eventfd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	args[1].Value = 0
+	return Eventfd2(t, args)
+}
diff --git a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go
index 74920f785..074f58e5d 100644
--- a/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go
+++ b/pkg/sentry/syscalls/linux/vfs2/linux64_override_amd64.go
@@ -141,14 +141,14 @@ func Override(table map[uintptr]kernel.Syscall) {
 	table[281] = syscalls.Supported("epoll_pwait", EpollPwait)
 	delete(table, 282) // signalfd
 	table[283] = syscalls.Supported("timerfd_create", TimerfdCreate)
-	delete(table, 284) // eventfd
+	table[284] = syscalls.Supported("eventfd", Eventfd)
 	delete(table, 285) // fallocate
 	table[286] = syscalls.Supported("timerfd_settime", TimerfdSettime)
 	table[287] = syscalls.Supported("timerfd_gettime", TimerfdGettime)
 	// TODO(gvisor.dev/issue/1485): Port all socket variants to VFS2.
 	table[288] = syscalls.PartiallySupported("accept4", Accept4, "In process of porting socket syscalls to VFS2.", nil)
 	delete(table, 289) // signalfd4
-	delete(table, 290) // eventfd2
+	table[290] = syscalls.Supported("eventfd2", Eventfd2)
 	table[291] = syscalls.Supported("epoll_create1", EpollCreate1)
 	table[292] = syscalls.Supported("dup3", Dup3)
 	table[293] = syscalls.Supported("pipe2", Pipe2)
diff --git a/pkg/sentry/syscalls/linux/vfs2/sys_timerfd.go b/pkg/sentry/syscalls/linux/vfs2/timerfd.go
index 7938a5249..7938a5249 100644
--- a/pkg/sentry/syscalls/linux/vfs2/sys_timerfd.go
+++ b/pkg/sentry/syscalls/linux/vfs2/timerfd.go
diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD
index 9aeb83fb0..c62505fe2 100644
--- a/pkg/sentry/vfs/BUILD
+++ b/pkg/sentry/vfs/BUILD
@@ -25,6 +25,7 @@ go_library(
         "device.go",
         "epoll.go",
         "epoll_interest_list.go",
+        "eventfd.go",
         "file_description.go",
         "file_description_impl_util.go",
         "filesystem.go",
@@ -44,6 +45,7 @@ go_library(
         "//pkg/abi/linux",
         "//pkg/context",
         "//pkg/fd",
+        "//pkg/fdnotifier",
         "//pkg/fspath",
         "//pkg/gohacks",
         "//pkg/log",
@@ -68,6 +70,7 @@ go_test(
     name = "vfs_test",
     size = "small",
     srcs = [
+        "eventfd_test.go",
         "file_description_impl_util_test.go",
         "mount_test.go",
     ],
@@ -79,5 +82,6 @@ go_test(
         "//pkg/sync",
         "//pkg/syserror",
         "//pkg/usermem",
+        "//pkg/waiter",
     ],
 )
diff --git a/pkg/sentry/vfs/eventfd.go b/pkg/sentry/vfs/eventfd.go
new file mode 100644
index 000000000..f39dacacf
--- /dev/null
+++ b/pkg/sentry/vfs/eventfd.go
@@ -0,0 +1,282 @@
+// 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 vfs
+
+import (
+	"math"
+	"sync"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/fdnotifier"
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+// EventFileDescription implements FileDescriptionImpl for file-based event
+// notification (eventfd). Eventfds are usually internal to the Sentry but in
+// certain situations they may be converted into a host-backed eventfd.
+type EventFileDescription struct {
+	vfsfd FileDescription
+	FileDescriptionDefaultImpl
+	DentryMetadataFileDescriptionImpl
+
+	// queue is used to notify interested parties when the event object
+	// becomes readable or writable.
+	queue waiter.Queue `state:"zerovalue"`
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// val is the current value of the event counter.
+	val uint64
+
+	// semMode specifies whether the event is in "semaphore" mode.
+	semMode bool
+
+	// hostfd indicates whether this eventfd is passed through to the host.
+	hostfd int
+}
+
+var _ FileDescriptionImpl = (*EventFileDescription)(nil)
+
+// NewEventFD creates a new event fd.
+func (vfs *VirtualFilesystem) NewEventFD(initVal uint64, semMode bool, flags uint32) (*FileDescription, error) {
+	vd := vfs.NewAnonVirtualDentry("[eventfd]")
+	defer vd.DecRef()
+	efd := &EventFileDescription{
+		val:     initVal,
+		semMode: semMode,
+		hostfd:  -1,
+	}
+	if err := efd.vfsfd.Init(efd, flags, vd.Mount(), vd.Dentry(), &FileDescriptionOptions{
+		UseDentryMetadata: true,
+		DenyPRead:         true,
+		DenyPWrite:        true,
+	}); err != nil {
+		return nil, err
+	}
+	return &efd.vfsfd, nil
+}
+
+// HostFD returns the host eventfd associated with this event.
+func (efd *EventFileDescription) HostFD() (int, error) {
+	efd.mu.Lock()
+	defer efd.mu.Unlock()
+	if efd.hostfd >= 0 {
+		return efd.hostfd, nil
+	}
+
+	flags := linux.EFD_NONBLOCK
+	if efd.semMode {
+		flags |= linux.EFD_SEMAPHORE
+	}
+
+	fd, _, errno := syscall.Syscall(syscall.SYS_EVENTFD2, uintptr(efd.val), uintptr(flags), 0)
+	if errno != 0 {
+		return -1, errno
+	}
+
+	if err := fdnotifier.AddFD(int32(fd), &efd.queue); err != nil {
+		if closeErr := syscall.Close(int(fd)); closeErr != nil {
+			log.Warningf("close(%d) eventfd failed: %v", fd, closeErr)
+		}
+		return -1, err
+	}
+
+	efd.hostfd = int(fd)
+	return efd.hostfd, nil
+}
+
+// Release implements FileDescriptionImpl.Release()
+func (efd *EventFileDescription) Release() {
+	efd.mu.Lock()
+	defer efd.mu.Unlock()
+	if efd.hostfd >= 0 {
+		fdnotifier.RemoveFD(int32(efd.hostfd))
+		if closeErr := syscall.Close(int(efd.hostfd)); closeErr != nil {
+			log.Warningf("close(%d) eventfd failed: %v", efd.hostfd, closeErr)
+		}
+		efd.hostfd = -1
+	}
+}
+
+// Read implements FileDescriptionImpl.Read.
+func (efd *EventFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ ReadOptions) (int64, error) {
+	if dst.NumBytes() < 8 {
+		return 0, syscall.EINVAL
+	}
+	if err := efd.read(ctx, dst); err != nil {
+		return 0, err
+	}
+	return 8, nil
+}
+
+// Write implements FileDescriptionImpl.Write.
+func (efd *EventFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ WriteOptions) (int64, error) {
+	if src.NumBytes() < 8 {
+		return 0, syscall.EINVAL
+	}
+	if err := efd.write(ctx, src); err != nil {
+		return 0, err
+	}
+	return 8, nil
+}
+
+// Preconditions: Must be called with efd.mu locked.
+func (efd *EventFileDescription) hostReadLocked(ctx context.Context, dst usermem.IOSequence) error {
+	var buf [8]byte
+	if _, err := syscall.Read(efd.hostfd, buf[:]); err != nil {
+		if err == syscall.EWOULDBLOCK {
+			return syserror.ErrWouldBlock
+		}
+		return err
+	}
+	_, err := dst.CopyOut(ctx, buf[:])
+	return err
+}
+
+func (efd *EventFileDescription) read(ctx context.Context, dst usermem.IOSequence) error {
+	efd.mu.Lock()
+	if efd.hostfd >= 0 {
+		defer efd.mu.Unlock()
+		return efd.hostReadLocked(ctx, dst)
+	}
+
+	// We can't complete the read if the value is currently zero.
+	if efd.val == 0 {
+		efd.mu.Unlock()
+		return syserror.ErrWouldBlock
+	}
+
+	// Update the value based on the mode the event is operating in.
+	var val uint64
+	if efd.semMode {
+		val = 1
+		// Consistent with Linux, this is done even if writing to memory fails.
+		efd.val--
+	} else {
+		val = efd.val
+		efd.val = 0
+	}
+
+	efd.mu.Unlock()
+
+	// Notify writers. We do this even if we were already writable because
+	// it is possible that a writer is waiting to write the maximum value
+	// to the event.
+	efd.queue.Notify(waiter.EventOut)
+
+	var buf [8]byte
+	usermem.ByteOrder.PutUint64(buf[:], val)
+	_, err := dst.CopyOut(ctx, buf[:])
+	return err
+}
+
+// Preconditions: Must be called with efd.mu locked.
+func (efd *EventFileDescription) hostWriteLocked(val uint64) error {
+	var buf [8]byte
+	usermem.ByteOrder.PutUint64(buf[:], val)
+	_, err := syscall.Write(efd.hostfd, buf[:])
+	if err == syscall.EWOULDBLOCK {
+		return syserror.ErrWouldBlock
+	}
+	return err
+}
+
+func (efd *EventFileDescription) write(ctx context.Context, src usermem.IOSequence) error {
+	var buf [8]byte
+	if _, err := src.CopyIn(ctx, buf[:]); err != nil {
+		return err
+	}
+	val := usermem.ByteOrder.Uint64(buf[:])
+
+	return efd.Signal(val)
+}
+
+// Signal is an internal function to signal the event fd.
+func (efd *EventFileDescription) Signal(val uint64) error {
+	if val == math.MaxUint64 {
+		return syscall.EINVAL
+	}
+
+	efd.mu.Lock()
+
+	if efd.hostfd >= 0 {
+		defer efd.mu.Unlock()
+		return efd.hostWriteLocked(val)
+	}
+
+	// We only allow writes that won't cause the value to go over the max
+	// uint64 minus 1.
+	if val > math.MaxUint64-1-efd.val {
+		efd.mu.Unlock()
+		return syserror.ErrWouldBlock
+	}
+
+	efd.val += val
+	efd.mu.Unlock()
+
+	// Always trigger a notification.
+	efd.queue.Notify(waiter.EventIn)
+
+	return nil
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (efd *EventFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask {
+	efd.mu.Lock()
+	defer efd.mu.Unlock()
+
+	if efd.hostfd >= 0 {
+		return fdnotifier.NonBlockingPoll(int32(efd.hostfd), mask)
+	}
+
+	ready := waiter.EventMask(0)
+	if efd.val > 0 {
+		ready |= waiter.EventIn
+	}
+
+	if efd.val < math.MaxUint64-1 {
+		ready |= waiter.EventOut
+	}
+
+	return mask & ready
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (efd *EventFileDescription) EventRegister(entry *waiter.Entry, mask waiter.EventMask) {
+	efd.queue.EventRegister(entry, mask)
+
+	efd.mu.Lock()
+	defer efd.mu.Unlock()
+	if efd.hostfd >= 0 {
+		fdnotifier.UpdateFD(int32(efd.hostfd))
+	}
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (efd *EventFileDescription) EventUnregister(entry *waiter.Entry) {
+	efd.queue.EventUnregister(entry)
+
+	efd.mu.Lock()
+	defer efd.mu.Unlock()
+	if efd.hostfd >= 0 {
+		fdnotifier.UpdateFD(int32(efd.hostfd))
+	}
+}
diff --git a/pkg/sentry/vfs/eventfd_test.go b/pkg/sentry/vfs/eventfd_test.go
new file mode 100644
index 000000000..2dff2d10b
--- /dev/null
+++ b/pkg/sentry/vfs/eventfd_test.go
@@ -0,0 +1,96 @@
+// 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 vfs
+
+import (
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/contexttest"
+	"gvisor.dev/gvisor/pkg/usermem"
+	"gvisor.dev/gvisor/pkg/waiter"
+)
+
+func TestEventFD(t *testing.T) {
+	initVals := []uint64{
+		0,
+		// Using a non-zero initial value verifies that writing to an
+		// eventfd signals when the eventfd's counter was already
+		// non-zero.
+		343,
+	}
+
+	for _, initVal := range initVals {
+		ctx := contexttest.Context(t)
+		vfsObj := &VirtualFilesystem{}
+		if err := vfsObj.Init(); err != nil {
+			t.Fatalf("VFS init: %v", err)
+		}
+
+		// Make a new eventfd that is writable.
+		eventfd, err := vfsObj.NewEventFD(initVal, false, linux.O_RDWR)
+		if err != nil {
+			t.Fatalf("NewEventFD failed: %v", err)
+		}
+		defer eventfd.DecRef()
+
+		// Register a callback for a write event.
+		w, ch := waiter.NewChannelEntry(nil)
+		eventfd.EventRegister(&w, waiter.EventIn)
+		defer eventfd.EventUnregister(&w)
+
+		data := []byte("00000124")
+		// Create and submit a write request.
+		n, err := eventfd.Write(ctx, usermem.BytesIOSequence(data), WriteOptions{})
+		if err != nil {
+			t.Fatal(err)
+		}
+		if n != 8 {
+			t.Errorf("eventfd.write wrote %d bytes, not full int64", n)
+		}
+
+		// Check if the callback fired due to the write event.
+		select {
+		case <-ch:
+		default:
+			t.Errorf("Didn't get notified of EventIn after write")
+		}
+	}
+}
+
+func TestEventFDStat(t *testing.T) {
+	ctx := contexttest.Context(t)
+	vfsObj := &VirtualFilesystem{}
+	if err := vfsObj.Init(); err != nil {
+		t.Fatalf("VFS init: %v", err)
+	}
+
+	// Make a new eventfd that is writable.
+	eventfd, err := vfsObj.NewEventFD(0, false, linux.O_RDWR)
+	if err != nil {
+		t.Fatalf("NewEventFD failed: %v", err)
+	}
+	defer eventfd.DecRef()
+
+	statx, err := eventfd.Stat(ctx, StatOptions{
+		Mask: linux.STATX_BASIC_STATS,
+	})
+	if err != nil {
+		t.Fatalf("eventfd.Stat failed: %v", err)
+	}
+	if statx.Size != 0 {
+		t.Errorf("eventfd size should be 0")
+	}
+}