2026-10-19 01:11:42 +02:00
8 changed files with 215 additions and 46 deletions
@@ -21,6 +21,21 @@ Run `make frontend-viewport-test` in CI as its own step; it is not part of

 # Completed Steps

+- 2026-10-18: `netwatch-server prepare-data-dir` no longer changes the owner of
+  a file outside `/data` through a hard link in the volume
+  ([#143](https://git.eeqj.de/sneak/netwatch/issues/143)). A file with a second
+  name (hard link) keeps its owner, and the start logs its path. Owners are now
+  set on the opened directory, file or link, never by name, so a hard link
+  swapped in after the check is caught too; a symbolic link is opened without
+  following it and given to the netwatch user itself, so the server can still
+  follow one in a sticky directory that anyone can write to. Devices, named
+  pipes and sockets keep their owner. Setting the owner on an open file that is
+  not a directory is built only for Linux: `prepare-data-dir` runs only in the
+  Linux image, and macOS, where the backend is also built for development, has
+  neither `O_PATH` nor `AT_EMPTY_PATH`. On any other system the package still
+  builds, and `prepare-data-dir` fails at the first file that is not a
+  directory. The test that runs as root hard-links a file, a symbolic link and a
+  named pipe from outside the volume into it and finds their owners unchanged
 - 2026-10-18: from a trusted proxy, the backend takes the client address from
  the right-most `X-Forwarded-For` entry that is not in `TRUSTED_PROXIES`
  ([#144](https://git.eeqj.de/sneak/netwatch/issues/144)). It took the left-most
@@ -66,10 +66,13 @@ done > /etc/nginx/trusted-proxies.conf
 # netwatch-server keeps its report files in DATA_DIR, on the /data
 # volume, which may be a host directory owned by root or by another
 # uid. Here, as root, netwatch-server prepare-data-dir creates DATA_DIR
-# and gives /data and everything in it to the netwatch user, so the
-# host directory needs no preparing. It stops the start, naming
-# DATA_DIR, unless DATA_DIR is /data or a path below it, and it acts on
-# nothing outside /data, whatever symbolic links it meets there.
+# and gives /data and the directories, regular files and symbolic links
+# in it to the netwatch user, so the host directory needs no preparing.
+# A file with a second name (hard link) keeps its owner, as that name
+# could be outside /data, and its path is logged. It stops the start,
+# naming DATA_DIR, unless DATA_DIR is /data or a path below it, and it
+# acts on nothing outside /data, whatever symbolic or hard links it
+# meets there.
 export DATA_DIR="${DATA_DIR:-/data/reports}"
 netwatch-server prepare-data-dir "$DATA_DIR" || exit 1

@@ -17,6 +17,7 @@ require (
 	github.com/slok/go-http-metrics v0.13.0
 	github.com/spf13/viper v1.21.0
 	go.uber.org/fx v1.24.0
+	golang.org/x/sys v0.47.0
 )

 require (
@@ -41,7 +42,6 @@ require (
 	go.uber.org/multierr v1.10.0 // indirect
 	go.uber.org/zap v1.26.0 // indirect
 	go.yaml.in/yaml/v3 v3.0.4 // indirect
-	golang.org/x/sys v0.47.0 // indirect
 	golang.org/x/text v0.40.0 // indirect
 	google.golang.org/protobuf v1.36.11 // indirect
 )
@@ -18,9 +18,15 @@ var ErrDataDirOutsideVolume = errors.New(
 // PrepareDataDir gets dir, the server's DATA_DIR, ready for owner, the
 // user the server runs as, so that a host directory mounted at volume,
 // /data in the image, needs no preparing: it creates dir, gives volume
-// and everything in it to owner, and gives volume and dir the mode the
-// server gives a directory it creates. dir must be volume or a path
-// below it, with no '.', '..', empty part or '/' at the end.
+// and the directories, regular files and symbolic links in it to
+// owner, and gives volume and dir the mode the server gives a directory
+// it creates. It leaves a file with a second name (hard link) as it
+// is, as that name could be outside volume, and logs its path; it
+// leaves a device, named pipe or socket as it is too (see chownFile).
+// dir must be volume or a path below it, with no '.', '..', empty part
+// or '/' at the end. On any system but Linux, the system of the image, it
+// fails at the first file in volume that is not a directory (see
+// chownFile).
 //
 // bin/entrypoint.sh runs this as root, which would follow a symbolic
 // link anywhere, so every step goes through an os.Root opened on
@@ -32,10 +38,9 @@ var ErrDataDirOutsideVolume = errors.New(
 // open, never following a link at the name it creates, and follows a
 // link on the path only as the os.Root allows, so a relative link
 // inside volume can lead it to create directories elsewhere inside
-// volume. Lchown never changes what a link points to, and the modes
-// are set on directories already opened (see chmodDir), so the most
-// that process can do is make a step fail or wait, or act on
-// something else inside volume.
+// volume. The owners and modes are set on files already opened (see
+// chownAll and chmodDir), so the most that process can do is make a
+// step fail or wait, or act on something else inside volume.
 func PrepareDataDir(volume, dir string, owner *user.User) error {
 	// rel is dir as a path from volume; IsLocal is false for one that
 	// leads out of it.
@@ -66,7 +71,7 @@ func PrepareDataDir(volume, dir string, owner *user.User) error {
 		return err
 	}

-	err = lchownAll(root, ".", uid, gid)
+	err = chownAll(root, ".", uid, gid)
 	if err != nil {
 		return err
 	}
@@ -79,28 +84,23 @@ func PrepareDataDir(volume, dir string, owner *user.User) error {
 	return chmodDir(root, rel)
 }

-// lchownAll gives name, a directory inside root, and everything in it
-// to uid and gid. It reads each directory opened through root, not
-// through root.FS(), which refuses a name that is not valid UTF-8, and
-// calls Lchown on every entry, which gives a symbolic link itself to
-// them, not what it points to. It goes into an entry only when the
-// read found a directory there, so it follows no link it finds; one
-// swapped in for that directory afterwards is followed only as the
-// os.Root allows.
-func lchownAll(root *os.Root, name string, uid, gid int) error {
-	err := root.Lchown(name, uid, gid)
-	if err != nil {
-		return err
-	}
-
-	dir, err := root.Open(name)
-	if err != nil {
-		return err
-	}
-
-	entries, err := dir.ReadDir(-1)
-	_ = dir.Close()
-
+// chownAll gives name, a directory inside root, and the directories,
+// regular files and symbolic links in it to uid and gid, but not the
+// files chownFile leaves. It opens each through root and gives the
+// open file to them, as chmodDir sets a mode on the open directory.
+// Root.Lchown would not hold: it acts on whatever is at the name by
+// then, and a hard link swapped in there shares its owner with a file
+// that could be outside root. A symbolic link is given to them itself,
+// never what it points to, because in a directory that anyone can
+// write to and that has the sticky bit set, as /tmp has, the kernel
+// follows a link only for the link's owner or when the link and the
+// directory have the same owner. chownAll goes into an entry only when
+// the read found a directory there, so it follows no link it finds;
+// one swapped in afterwards for a directory is followed only as the
+// os.Root allows, and one swapped in for any other file is not
+// followed at all (see chownFile).
+func chownAll(root *os.Root, name string, uid, gid int) error {
+	entries, err := chownDir(root, name, uid, gid)
 	if err != nil {
 		return err
 	}
@@ -108,9 +108,9 @@ func lchownAll(root *os.Root, name string, uid, gid int) error {
 	for _, entry := range entries {
 		entryName := filepath.Join(name, entry.Name())
 		if entry.IsDir() {
-			err = lchownAll(root, entryName, uid, gid)
+			err = chownAll(root, entryName, uid, gid)
 		} else {
-			err = root.Lchown(entryName, uid, gid)
+			err = chownFile(root, entryName, uid, gid)
 		}

 		if err != nil {
@@ -121,6 +121,36 @@ func lchownAll(root *os.Root, name string, uid, gid int) error {
 	return nil
 }

+// chownDir gives name, a directory inside root, to uid and gid, and
+// returns what it holds. Like chmodDir, it opens name through root and
+// refuses anything but a directory, which has no second name (hard
+// link). It reads the open directory, not root.FS(), which refuses a
+// name that is not valid UTF-8.
+func chownDir(root *os.Root, name string, uid, gid int) ([]fs.DirEntry, error) {
+	dir, err := root.Open(name)
+	if err != nil {
+		return nil, err
+	}
+
+	defer func() { _ = dir.Close() }()
+
+	info, err := dir.Stat()
+	if err != nil {
+		return nil, err
+	}
+
+	if !info.IsDir() {
+		return nil, &fs.PathError{Op: "chown", Path: name, Err: syscall.ENOTDIR}
+	}
+
+	err = dir.Chown(uid, gid)
+	if err != nil {
+		return nil, err
+	}
+
+	return dir.ReadDir(-1)
+}
+
 // chmodDir gives name, a directory inside root, the mode the server
 // gives a directory it creates. Root.Chmod would not hold: on Linux it
 // checks that name is not a symbolic link, then sets the mode by name,
@@ -0,0 +1,65 @@
+//go:build linux
+
+package reportbuf
+
+import (
+	"io/fs"
+	"log/slog"
+	"os"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+)
+
+// chownFile gives name, a file inside root that is not a directory, to
+// uid and gid if it is a regular file or symbolic link with one name.
+// It leaves, and logs, a file of any kind with a second name (hard
+// link): all the names of a file share its owner, and the other could
+// be outside root. It leaves any other file with one name, such as a
+// device, named pipe or socket: the server needs to own none of them.
+// It opens name through root with O_PATH, which reads nothing from the
+// file, so opening a device or named pipe neither waits nor acts on
+// it; os.Root adds O_NOFOLLOW, and the two together open a symbolic
+// link itself, never what it points to. It reads the kind of file and
+// its number of names from the open file and gives that file to them,
+// so a hard link swapped in at name is caught too.
+//
+// This file is built only on Linux: prepare-data-dir runs only in the
+// Linux image, and macOS, where the backend is also built for
+// development, has neither O_PATH nor AT_EMPTY_PATH. On any other
+// system chownFile returns an error (see datadir_other.go).
+func chownFile(root *os.Root, name string, uid, gid int) error {
+	file, err := root.OpenFile(name, unix.O_PATH, 0)
+	if err != nil {
+		return err
+	}
+
+	defer func() { _ = file.Close() }()
+
+	info, err := file.Stat()
+	if err != nil {
+		return err
+	}
+
+	stat, ok := info.Sys().(*syscall.Stat_t)
+	if !ok || stat.Nlink > 1 {
+		slog.Warn("owner not changed: file has a second name (hard link)",
+			"path", file.Name())
+
+		return nil
+	}
+
+	if !info.Mode().IsRegular() && info.Mode().Type() != fs.ModeSymlink {
+		return nil
+	}
+
+	// File.Chown would fail: fchown refuses a file opened with O_PATH.
+	// Fchownat with an empty name and AT_EMPTY_PATH acts on the open
+	// file instead.
+	err = unix.Fchownat(int(file.Fd()), "", uid, gid, unix.AT_EMPTY_PATH)
+	if err != nil {
+		return &fs.PathError{Op: "chown", Path: file.Name(), Err: err}
+	}
+
+	return nil
+}
@@ -0,0 +1,21 @@
+//go:build !linux
+
+package reportbuf
+
+import (
+	"errors"
+	"io/fs"
+	"os"
+)
+
+// chownFile returns an error on every system but Linux. Opening a file
+// that is not a directory without reading it or following it, and
+// giving it to a user on the open file, never by name, is built only
+// for Linux (see datadir_linux.go): prepare-data-dir runs only in the
+// Linux image, and macOS, where the backend is also built for
+// development, has neither O_PATH nor AT_EMPTY_PATH. This file lets
+// the package build elsewhere, such as on macOS, where a developer
+// runs make build and make run.
+func chownFile(_ *os.Root, name string, _, _ int) error {
+	return &fs.PathError{Op: "chown", Path: name, Err: errors.ErrUnsupported}
+}
@@ -254,10 +254,13 @@ func TestPrepareDataDirTakesDirectoryNamedInLatin1(t *testing.T) {
 	requireOwner(t, filepath.Join(latin1, "f"), uint32(uid), uint32(gid))
 }

-// TestPrepareDataDirGivesVolumeToOwner gives everything in the volume
-// to a uid and gid that own nothing, which only root can do. A
-// symbolic link in the volume to a directory outside it is given to
-// them itself; what it points to is left as it was.
+// TestPrepareDataDirGivesVolumeToOwner gives the directories, regular
+// files and symbolic links in the volume to a uid and gid that own
+// nothing, which only root can do. A symbolic link in the volume to a
+// directory outside it is given to them itself; what it points to is
+// left as it was. A hard link to a file outside the volume, one to a
+// symbolic link outside it and one to a named pipe outside it are left
+// as they were, and so is a named pipe in the volume.
 func TestPrepareDataDirGivesVolumeToOwner(t *testing.T) {
 	t.Parallel()

@@ -274,6 +277,16 @@ func TestPrepareDataDirGivesVolumeToOwner(t *testing.T) {
 		t.Fatal(err)
 	}

+	err = os.Symlink("f", filepath.Join(outside, "link"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	err = syscall.Mkfifo(filepath.Join(outside, "fifo"), 0o600)
+	if err != nil {
+		t.Fatal(err)
+	}
+
 	err = os.Mkdir(old, 0o700)
 	if err != nil {
 		t.Fatal(err)
@@ -289,6 +302,21 @@ func TestPrepareDataDirGivesVolumeToOwner(t *testing.T) {
 		t.Fatal(err)
 	}

+	err = syscall.Mkfifo(filepath.Join(old, "fifo"), 0o600)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// os.Link does not follow a symbolic link: it gives the link itself
+	// a second name.
+	for _, name := range []string{"f", "link", "fifo"} {
+		err = os.Link(filepath.Join(outside, name),
+			filepath.Join(old, "hardlink-to-"+name))
+		if err != nil {
+			t.Fatal(err)
+		}
+	}
+
 	err = reportbuf.PrepareDataDir(volume, filepath.Join(volume, reports),
 		&user.User{Uid: "4242", Gid: "4343"})
 	if err != nil {
@@ -296,12 +324,16 @@ func TestPrepareDataDirGivesVolumeToOwner(t *testing.T) {
 	}

 	for _, path := range []string{
-		volume, filepath.Join(volume, reports), old,
-		filepath.Join(old, "f"), filepath.Join(old, "link"),
+		volume, filepath.Join(volume, reports), old, filepath.Join(old, "f"),
+		filepath.Join(old, "link"),
 	} {
 		requireOwner(t, path, 4242, 4343)
 	}

-	requireOwner(t, outside, 0, 0)
-	requireOwner(t, filepath.Join(outside, "f"), 0, 0)
+	for _, path := range []string{
+		outside, filepath.Join(outside, "f"), filepath.Join(outside, "link"),
+		filepath.Join(outside, "fifo"), filepath.Join(old, "fifo"),
+	} {
+		requireOwner(t, path, 0, 0)
+	}
 }
@@ -1,6 +1,9 @@
 // Package reportbuf accumulates telemetry reports in memory
 // and periodically flushes them to zstd-compressed JSONL files,
-// deleting the oldest files to keep them under a size cap.
+// deleting the oldest files to keep them under a size cap. Before the
+// server starts, PrepareDataDir gets the directory those files are
+// kept in, DATA_DIR, ready for the user the server runs as: it creates
+// it and gives it and the volume it is on to that user.
 package reportbuf

 import (