prepare-data-dir leaves files with a second name to their owner (closes #143)

A hard link in /data shares its owner with a file that can be anywhere
on the same filesystem, so chowning it gave that file to the netwatch
user. A regular file or symbolic link with more than one name now
keeps its owner and its path is logged.

Owners are now set on the opened directory, file or link after
checking it, as chmodDir sets modes, so a hard link swapped in
afterwards is caught. A symbolic link is opened without following it
and given to the netwatch user itself, as the server must own a link
it follows in a sticky directory anyone can write to. Other files,
such as devices, keep their owner.

Model: opus-5-5

TODO.md

@@ -21,6 +21,17 @@ 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 regular file or
+  symbolic link 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. Other files that are neither directories, regular files nor symbolic links
+  keep their owner. The test that runs as root hard-links a file from outside
+  the volume into it and finds its owner unchanged
 - 2026-10-18: after an interval change, the summary's Reachable slot reads
   `--/--` in white, as at page load, until the second round since the change
   ends ([#140](https://git.eeqj.de/sneak/netwatch/issues/140)). It used to read

bin/entrypoint.sh

@@ -66,10 +66,12 @@ 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
+# and gives /data and the directories, regular files and symbolic links
+# in it to the netwatch user, so the host directory needs no preparing,
+# except a file with a second name (hard link), which keeps its owner,
+# as that name could be outside /data. 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.
+# 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
 

go.mod

@@ -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
 )

internal/reportbuf/datadir.go

@@ -3,11 +3,14 @@ package reportbuf
 import (
 	"errors"
 	"io/fs"
+	"log/slog"
 	"os"
 	"os/user"
 	"path/filepath"
 	"strconv"
 	"syscall"
+
+	"golang.org/x/sys/unix"
 )
 
 // ErrDataDirOutsideVolume is returned by PrepareDataDir for a DATA_DIR
@@ -18,9 +21,13 @@ 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 as they are a regular file or symbolic link
+// with a second name (hard link), as that name could be outside volume,
+// and any other file, such as a device (see chownAll). dir must be
+// volume or a path below it, with no '.', '..', empty part or '/' at
+// the end.
 //
 // 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 +39,9 @@ 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 +72,7 @@ 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
 	}
@@ -79,38 +85,36 @@ 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. 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 a
+// regular file or symbolic link is not followed at all (see
+// chownFile). It leaves any other file, such as a device: the server
+// needs to own none of them.
+func chownAll(root *os.Root, name string, uid, gid int) error {
+	entries, err := chownDir(root, name, uid, gid)
 	if err != nil {
 		return err
 	}
 
 	for _, entry := range entries {
 		entryName := filepath.Join(name, entry.Name())
-		if entry.IsDir() {
-			err = lchownAll(root, entryName, uid, gid)
-		} else {
-			err = root.Lchown(entryName, uid, gid)
+
+		switch {
+		case entry.IsDir():
+			err = chownAll(root, entryName, uid, gid)
+		case entry.Type().IsRegular(), entry.Type() == fs.ModeSymlink:
+			err = chownFile(root, entryName, uid, gid)
 		}
 
 		if err != nil {
@@ -121,6 +125,71 @@ 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)
+}
+
+// chownFile gives name, a regular file or symbolic link inside root, to
+// uid and gid, unless it has a second name (hard link): all the names
+// of a file share its owner, and the other could be outside root. It
+// opens name through root with O_PATH, which reads nothing from the
+// file; os.Root adds O_NOFOLLOW, and the two together open a symbolic
+// link itself, never what it points to. It reads the number of names
+// from the open file and gives that file to them, so a hard link
+// swapped in at name is caught too. It logs a file it leaves.
+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
+	}
+
+	// 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.
+	stat, ok := info.Sys().(*syscall.Stat_t)
+	if ok && stat.Nlink == 1 {
+		return unix.Fchownat(int(file.Fd()), "", uid, gid, unix.AT_EMPTY_PATH)
+	}
+
+	slog.Warn("owner not changed: file has a second name (hard link)",
+		"path", file.Name())
+
+	return nil
+}
+
 // 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,

internal/reportbuf/datadir_test.go

@@ -254,10 +254,12 @@ 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, and one to
+// a symbolic link outside it, are left as they were.
 func TestPrepareDataDirGivesVolumeToOwner(t *testing.T) {
 	t.Parallel()
 
@@ -289,6 +291,24 @@ func TestPrepareDataDirGivesVolumeToOwner(t *testing.T) {
 		t.Fatal(err)
 	}
 
+	err = os.Link(filepath.Join(outside, "f"), filepath.Join(old, "hardlink"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// os.Link does not follow a symbolic link: it gives the link itself
+	// a second name.
+	err = os.Symlink("f", filepath.Join(outside, "link"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	err = os.Link(filepath.Join(outside, "link"),
+		filepath.Join(old, "hardlink-to-link"))
+	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 +316,15 @@ 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"),
+	} {
+		requireOwner(t, path, 0, 0)
+	}
 }

internal/reportbuf/reportbuf.go

@@ -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 (