11 files changed, 464 insertions, 38 deletions

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/anonfs.go b/pkg/sentry/vfs/anonfs.go
index a64d86122..981bd8caa 100644
--- a/pkg/sentry/vfs/anonfs.go
+++ b/pkg/sentry/vfs/anonfs.go
@@ -237,10 +237,13 @@ func (fs *anonFilesystem) UnlinkAt(ctx context.Context, rp *ResolvingPath) error
 }
 
 // BoundEndpointAt implements FilesystemImpl.BoundEndpointAt.
-func (fs *anonFilesystem) BoundEndpointAt(ctx context.Context, rp *ResolvingPath) (transport.BoundEndpoint, error) {
+func (fs *anonFilesystem) BoundEndpointAt(ctx context.Context, rp *ResolvingPath, opts BoundEndpointOptions) (transport.BoundEndpoint, error) {
 	if !rp.Final() {
 		return nil, syserror.ENOTDIR
 	}
+	if err := GenericCheckPermissions(rp.Credentials(), MayWrite, anonFileMode, anonFileUID, anonFileGID); err != nil {
+		return nil, err
+	}
 	return nil, syserror.ECONNREFUSED
 }
 
diff --git a/pkg/sentry/vfs/context.go b/pkg/sentry/vfs/context.go
index 82781e6d3..c9e724fef 100644
--- a/pkg/sentry/vfs/context.go
+++ b/pkg/sentry/vfs/context.go
@@ -49,3 +49,27 @@ func RootFromContext(ctx context.Context) VirtualDentry {
 	}
 	return VirtualDentry{}
 }
+
+type rootContext struct {
+	context.Context
+	root VirtualDentry
+}
+
+// WithRoot returns a copy of ctx with the given root.
+func WithRoot(ctx context.Context, root VirtualDentry) context.Context {
+	return &rootContext{
+		Context: ctx,
+		root:    root,
+	}
+}
+
+// Value implements Context.Value.
+func (rc rootContext) Value(key interface{}) interface{} {
+	switch key {
+	case CtxRoot:
+		rc.root.IncRef()
+		return rc.root
+	default:
+		return rc.Context.Value(key)
+	}
+}
diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go
index 8e0b40841..8297f964b 100644
--- a/pkg/sentry/vfs/epoll.go
+++ b/pkg/sentry/vfs/epoll.go
@@ -192,6 +192,7 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, event lin
 		mask:     mask,
 		userData: event.Data,
 	}
+	epi.waiter.Callback = epi
 	ep.interest[key] = epi
 	wmask := waiter.EventMaskFromLinux(mask)
 	file.EventRegister(&epi.waiter, wmask)
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")
+	}
+}
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go
index 418d69b96..cfabd936c 100644
--- a/pkg/sentry/vfs/file_description.go
+++ b/pkg/sentry/vfs/file_description.go
@@ -91,10 +91,6 @@ type FileDescriptionOptions struct {
 	// ESPIPE.
 	DenyPWrite bool
 
-	// if InvalidWrite is true, calls to FileDescription.Write() return
-	// EINVAL.
-	InvalidWrite bool
-
 	// If UseDentryMetadata is true, calls to FileDescription methods that
 	// interact with file and filesystem metadata (Stat, SetStat, StatFS,
 	// Listxattr, Getxattr, Setxattr, Removexattr) are implemented by calling
@@ -570,9 +566,6 @@ func (fd *FileDescription) PWrite(ctx context.Context, src usermem.IOSequence, o
 
 // Write is similar to PWrite, but does not specify an offset.
 func (fd *FileDescription) Write(ctx context.Context, src usermem.IOSequence, opts WriteOptions) (int64, error) {
-	if fd.opts.InvalidWrite {
-		return 0, syserror.EINVAL
-	}
 	if !fd.writable {
 		return 0, syserror.EBADF
 	}
diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go
index 20e5bb072..1edd584c9 100644
--- a/pkg/sentry/vfs/filesystem.go
+++ b/pkg/sentry/vfs/filesystem.go
@@ -494,8 +494,14 @@ type FilesystemImpl interface {
 
 	// BoundEndpointAt returns the Unix socket endpoint bound at the path rp.
 	//
-	// - If a non-socket file exists at rp, then BoundEndpointAt returns ECONNREFUSED.
-	BoundEndpointAt(ctx context.Context, rp *ResolvingPath) (transport.BoundEndpoint, error)
+	// Errors:
+	//
+	// - If the file does not have write permissions, then BoundEndpointAt
+	// returns EACCES.
+	//
+	// - If a non-socket file exists at rp, then BoundEndpointAt returns
+	// ECONNREFUSED.
+	BoundEndpointAt(ctx context.Context, rp *ResolvingPath, opts BoundEndpointOptions) (transport.BoundEndpoint, error)
 
 	// PrependPath prepends a path from vd to vd.Mount().Root() to b.
 	//
diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go
index 022bac127..53d364c5c 100644
--- a/pkg/sentry/vfs/options.go
+++ b/pkg/sentry/vfs/options.go
@@ -151,6 +151,24 @@ type SetStatOptions struct {
 	Stat linux.Statx
 }
 
+// BoundEndpointOptions contains options to VirtualFilesystem.BoundEndpointAt()
+// and FilesystemImpl.BoundEndpointAt().
+type BoundEndpointOptions struct {
+	// Addr is the path of the file whose socket endpoint is being retrieved.
+	// It is generally irrelevant: most endpoints are stored at a dentry that
+	// was created through a bind syscall, so the path can be stored on creation.
+	// However, if the endpoint was created in FilesystemImpl.BoundEndpointAt(),
+	// then we may not know what the original bind address was.
+	//
+	// For example, if connect(2) is called with address "foo" which corresponds
+	// a remote named socket in goferfs, we need to generate an endpoint wrapping
+	// that file. In this case, we can use Addr to set the endpoint address to
+	// "foo". Note that Addr is only a best-effort attempt--we still do not know
+	// the exact address that was used on the remote fs to bind the socket (it
+	// may have been "foo", "./foo", etc.).
+	Addr string
+}
+
 // GetxattrOptions contains options to VirtualFilesystem.GetxattrAt(),
 // FilesystemImpl.GetxattrAt(), FileDescription.Getxattr(), and
 // FileDescriptionImpl.Getxattr().
diff --git a/pkg/sentry/vfs/timerfd.go b/pkg/sentry/vfs/timerfd.go
index 42b880656..cc536ceaf 100644
--- a/pkg/sentry/vfs/timerfd.go
+++ b/pkg/sentry/vfs/timerfd.go
@@ -53,7 +53,6 @@ func (vfs *VirtualFilesystem) NewTimerFD(clock ktime.Clock, flags uint32) (*File
 		UseDentryMetadata: true,
 		DenyPRead:         true,
 		DenyPWrite:        true,
-		InvalidWrite:      true,
 	}); err != nil {
 		return nil, err
 	}
diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go
index 9015f2cc1..8d7f8f8af 100644
--- a/pkg/sentry/vfs/vfs.go
+++ b/pkg/sentry/vfs/vfs.go
@@ -351,33 +351,6 @@ func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentia
 	}
 }
 
-// BoundEndpointAt gets the bound endpoint at the given path, if one exists.
-func (vfs *VirtualFilesystem) BoundEndpointAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) (transport.BoundEndpoint, error) {
-	if !pop.Path.Begin.Ok() {
-		if pop.Path.Absolute {
-			return nil, syserror.ECONNREFUSED
-		}
-		return nil, syserror.ENOENT
-	}
-	rp := vfs.getResolvingPath(creds, pop)
-	for {
-		bep, err := rp.mount.fs.impl.BoundEndpointAt(ctx, rp)
-		if err == nil {
-			vfs.putResolvingPath(rp)
-			return bep, nil
-		}
-		if checkInvariants {
-			if rp.canHandleError(err) && rp.Done() {
-				panic(fmt.Sprintf("%T.BoundEndpointAt() consumed all path components and returned %v", rp.mount.fs.impl, err))
-			}
-		}
-		if !rp.handleError(err) {
-			vfs.putResolvingPath(rp)
-			return nil, err
-		}
-	}
-}
-
 // OpenAt returns a FileDescription providing access to the file at the given
 // path. A reference is taken on the returned FileDescription.
 func (vfs *VirtualFilesystem) OpenAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *OpenOptions) (*FileDescription, error) {
@@ -675,6 +648,33 @@ func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credenti
 	}
 }
 
+// BoundEndpointAt gets the bound endpoint at the given path, if one exists.
+func (vfs *VirtualFilesystem) BoundEndpointAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *BoundEndpointOptions) (transport.BoundEndpoint, error) {
+	if !pop.Path.Begin.Ok() {
+		if pop.Path.Absolute {
+			return nil, syserror.ECONNREFUSED
+		}
+		return nil, syserror.ENOENT
+	}
+	rp := vfs.getResolvingPath(creds, pop)
+	for {
+		bep, err := rp.mount.fs.impl.BoundEndpointAt(ctx, rp, *opts)
+		if err == nil {
+			vfs.putResolvingPath(rp)
+			return bep, nil
+		}
+		if checkInvariants {
+			if rp.canHandleError(err) && rp.Done() {
+				panic(fmt.Sprintf("%T.BoundEndpointAt() consumed all path components and returned %v", rp.mount.fs.impl, err))
+			}
+		}
+		if !rp.handleError(err) {
+			vfs.putResolvingPath(rp)
+			return nil, err
+		}
+	}
+}
+
 // ListxattrAt returns all extended attribute names for the file at the given
 // path.
 func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, size uint64) ([]string, error) {