fix(app): detach daemon with double fork

commit: 4c5416faa2e59ba9b8492e47d7e4e292946e3f9a
author: brice · 3mo ago · 2026-07-24 16:09 UTC
parent: 253fc6a
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

3 files · +111 -6

M internal/app/run.go +5 -0


@@ -61,6 +61,7 @@ const (
61	61		kindKill
62	62		kindCmd
63	63		kindDaemon
	64	+	kindDaemonLauncher
64	65		kindStdio
65	66		kindUDPBootstrap
66	67		kindUDPProxy

@@ -141,6 +142,8 @@ func parseArgs(args []string) (command, error) {
141	142		switch args[0] {
142	143		case "--daemon":
143	144		return command{kind: kindDaemon}, nil
	145	+	case "--daemon-launcher":
	146	+	return command{kind: kindDaemonLauncher}, nil
144	147		case "_stdio":
145	148		if len(args) > 2 {
146	149		return command{}, usagef("`_stdio` accepts at most one session name")

@@ -244,6 +247,8 @@ func dispatch(ctx context.Context, cmd command) error {
244	247		return nil
245	248		case kindDaemon:
246	249		return runDaemon()
	250	+	case kindDaemonLauncher:
	251	+	return runDaemonLauncher()
247	252		case kindStdio:
248	253		return runStdio(ctx)
249	254		case kindUDPBootstrap:


@@ -61,6 +61,7 @@ const (
61	kindKill	61		kindKill
62	kindCmd	62		kindCmd
63	kindDaemon	63		kindDaemon
		64	+	kindDaemonLauncher
64	kindStdio	65		kindStdio
65	kindUDPBootstrap	66		kindUDPBootstrap
66	kindUDPProxy	67		kindUDPProxy

@@ -141,6 +142,8 @@ func parseArgs(args []string) (command, error) {
141	switch args[0] {	142		switch args[0] {
142	case "--daemon":	143		case "--daemon":
143	return command{kind: kindDaemon}, nil	144		return command{kind: kindDaemon}, nil
		145	+	case "--daemon-launcher":
		146	+	return command{kind: kindDaemonLauncher}, nil
144	case "_stdio":	147		case "_stdio":
145	if len(args) > 2 {	148		if len(args) > 2 {
146	return command{}, usagef("`_stdio` accepts at most one session name")	149		return command{}, usagef("`_stdio` accepts at most one session name")

@@ -244,6 +247,8 @@ func dispatch(ctx context.Context, cmd command) error {
244	return nil	247		return nil
245	case kindDaemon:	248		case kindDaemon:
246	return runDaemon()	249		return runDaemon()
		250	+	case kindDaemonLauncher:
		251	+	return runDaemonLauncher()
247	case kindStdio:	252		case kindStdio:
248	return runStdio(ctx)	253		return runStdio(ctx)
249	case kindUDPBootstrap:	254		case kindUDPBootstrap:

M internal/app/spawn.go +35 -6

@@ -1,6 +1,7 @@
1	1		package app
2	2
3	3		import (
	4	+	"bytes"
4	5		"context"
5	6		"errors"
6	7		"fmt"

@@ -152,10 +153,9 @@ func realDial(ctx context.Context, dir string) (ports.Transport, error) {
152	153		return ipc.DialContext(ctx, dir)
153	154		}
154	155
155		-	// realSpawn re-execs this binary as a detached daemon: a new session
156		-	// (Setsid) with stdio bound to /dev/null so it survives the client's exit
157		-	// and never writes to the client's terminal. It logs via slog to the shared
158		-	// log file instead.
	156	+	// realSpawn re-execs this binary through a short-lived launcher. Waiting for
	157	+	// that launcher to exit ensures the daemon has been re-parented before the
	158	+	// client continues, keeping process-tree termination from reaching it.
159	159		func realSpawn() error {
160	160		exePath, err := selfExePath()
161	161		if err != nil {

@@ -168,6 +168,37 @@ func realSpawn() error {
168	168		}
169	169		defer func() { _ = devNull.Close() }()
170	170
	171	+	var stderr bytes.Buffer
	172	+	cmd := exec.Command(exePath, "--daemon-launcher")
	173	+	cmd.Dir = platform.DirOrHome("")
	174	+	cmd.Stdin = devNull
	175	+	cmd.Stdout = devNull
	176	+	cmd.Stderr = &stderr
	177	+	if err := cmd.Run(); err != nil {
	178	+	if detail := bytes.TrimSpace(stderr.Bytes()); len(detail) > 0 {
	179	+	return fmt.Errorf("%w: %s", err, detail)
	180	+	}
	181	+	return err
	182	+	}
	183	+	slog.Info("daemon process detached")
	184	+	return nil
	185	+	}
	186	+
	187	+	// runDaemonLauncher is the intermediate half of the double-fork. It starts
	188	+	// the long-lived daemon in a new session and exits immediately; realSpawn
	189	+	// waits for this exit so the daemon is no longer descended from the client.
	190	+	func runDaemonLauncher() error {
	191	+	exePath, err := selfExePath()
	192	+	if err != nil {
	193	+	return fmt.Errorf("resolving executable path: %w", err)
	194	+	}
	195	+
	196	+	devNull, err := os.OpenFile(os.DevNull, os.O_RDWR, 0)
	197	+	if err != nil {
	198	+	return fmt.Errorf("opening %s: %w", os.DevNull, err)
	199	+	}
	200	+	defer func() { _ = devNull.Close() }()
	201	+
171	202		cmd := exec.Command(exePath, "--daemon")
172	203		cmd.Dir = platform.DirOrHome("")
173	204		cmd.Stdin = devNull

@@ -177,7 +208,5 @@ func realSpawn() error {
177	208		if err := cmd.Start(); err != nil {
178	209		return err
179	210		}
180		-	slog.Info("daemon process started", "pid", cmd.Process.Pid)
181		-	// Detach: we neither wait for nor signal the daemon after launch.
182	211		return cmd.Process.Release()
183	212		}

@@ -1,6 +1,7 @@
1		package app	1		package app
2			2
3		import (	3		import (
			4	+	"bytes"
4		"context"	5		"context"
5		"errors"	6		"errors"
6		"fmt"	7		"fmt"

@@ -152,10 +153,9 @@ func realDial(ctx context.Context, dir string) (ports.Transport, error) {
152		return ipc.DialContext(ctx, dir)	153		return ipc.DialContext(ctx, dir)
153		}	154		}
154			155
155	-	// realSpawn re-execs this binary as a detached daemon: a new session	156	+	// realSpawn re-execs this binary through a short-lived launcher. Waiting for
156	-	// (Setsid) with stdio bound to /dev/null so it survives the client's exit	157	+	// that launcher to exit ensures the daemon has been re-parented before the
157	-	// and never writes to the client's terminal. It logs via slog to the shared	158	+	// client continues, keeping process-tree termination from reaching it.
158	-	// log file instead.
159		func realSpawn() error {	159		func realSpawn() error {
160		exePath, err := selfExePath()	160		exePath, err := selfExePath()
161		if err != nil {	161		if err != nil {

@@ -168,6 +168,37 @@ func realSpawn() error {
168		}	168		}
169		defer func() { _ = devNull.Close() }()	169		defer func() { _ = devNull.Close() }()
170			170
			171	+	var stderr bytes.Buffer
			172	+	cmd := exec.Command(exePath, "--daemon-launcher")
			173	+	cmd.Dir = platform.DirOrHome("")
			174	+	cmd.Stdin = devNull
			175	+	cmd.Stdout = devNull
			176	+	cmd.Stderr = &stderr
			177	+	if err := cmd.Run(); err != nil {
			178	+	if detail := bytes.TrimSpace(stderr.Bytes()); len(detail) > 0 {
			179	+	return fmt.Errorf("%w: %s", err, detail)
			180	+	}
			181	+	return err
			182	+	}
			183	+	slog.Info("daemon process detached")
			184	+	return nil
			185	+	}
			186	+
			187	+	// runDaemonLauncher is the intermediate half of the double-fork. It starts
			188	+	// the long-lived daemon in a new session and exits immediately; realSpawn
			189	+	// waits for this exit so the daemon is no longer descended from the client.
			190	+	func runDaemonLauncher() error {
			191	+	exePath, err := selfExePath()
			192	+	if err != nil {
			193	+	return fmt.Errorf("resolving executable path: %w", err)
			194	+	}
			195	+
			196	+	devNull, err := os.OpenFile(os.DevNull, os.O_RDWR, 0)
			197	+	if err != nil {
			198	+	return fmt.Errorf("opening %s: %w", os.DevNull, err)
			199	+	}
			200	+	defer func() { _ = devNull.Close() }()
			201	+
171		cmd := exec.Command(exePath, "--daemon")	202		cmd := exec.Command(exePath, "--daemon")
172		cmd.Dir = platform.DirOrHome("")	203		cmd.Dir = platform.DirOrHome("")
173		cmd.Stdin = devNull	204		cmd.Stdin = devNull

@@ -177,7 +208,5 @@ func realSpawn() error {
177		if err := cmd.Start(); err != nil {	208		if err := cmd.Start(); err != nil {
178		return err	209		return err
179		}	210		}
180	-	slog.Info("daemon process started", "pid", cmd.Process.Pid)
181	-	// Detach: we neither wait for nor signal the daemon after launch.
182		return cmd.Process.Release()	211		return cmd.Process.Release()
183		}	212		}

M internal/app/spawn_test.go +71 -0


@@ -3,10 +3,14 @@ package app
3	3		import (
4	4		"context"
5	5		"errors"
	6	+	"fmt"
6	7		"os"
7	8		"path/filepath"
	9	+	"strconv"
	10	+	"strings"
8	11		"sync"
9	12		"sync/atomic"
	13	+	"syscall"
10	14		"testing"
11	15		"time"
12	16

@@ -16,6 +20,73 @@ import (
16	20
17	21		var errDialFailed = errors.New("dial failed")
18	22
	23	+	const spawnTestChildFileEnv = "VEV_SPAWN_TEST_CHILD_FILE"
	24	+
	25	+	func TestMain(m *testing.M) {
	26	+	if len(os.Args) == 2 {
	27	+	switch os.Args[1] {
	28	+	case "--daemon-launcher":
	29	+	if err := Run(os.Args[1:]); err != nil {
	30	+	os.Exit(2)
	31	+	}
	32	+	os.Exit(0)
	33	+	case "--daemon":
	34	+	path := os.Getenv(spawnTestChildFileEnv)
	35	+	if path == "" {
	36	+	os.Exit(2)
	37	+	}
	38	+	if err := os.WriteFile(path, []byte(strconv.Itoa(os.Getpid())), 0o600); err != nil {
	39	+	os.Exit(2)
	40	+	}
	41	+	for {
	42	+	time.Sleep(time.Hour)
	43	+	}
	44	+	}
	45	+	}
	46	+	os.Exit(m.Run())
	47	+	}
	48	+
	49	+	func TestRealSpawnReparentsDaemonBeforeReturning(t *testing.T) {
	50	+	childFile := filepath.Join(t.TempDir(), "child")
	51	+	t.Setenv(spawnTestChildFileEnv, childFile)
	52	+
	53	+	if err := realSpawn(); err != nil {
	54	+	t.Fatalf("realSpawn: %v", err)
	55	+	}
	56	+
	57	+	var pid int
	58	+	deadline := time.Now().Add(2 * time.Second)
	59	+	for time.Now().Before(deadline) {
	60	+	data, err := os.ReadFile(childFile)
	61	+	if err == nil {
	62	+	pid, err = strconv.Atoi(string(data))
	63	+	if err != nil {
	64	+	t.Fatalf("parse child PID: %v", err)
	65	+	}
	66	+	break
	67	+	}
	68	+	if !errors.Is(err, os.ErrNotExist) {
	69	+	t.Fatalf("read child process record: %v", err)
	70	+	}
	71	+	time.Sleep(10 * time.Millisecond)
	72	+	}
	73	+	if pid == 0 {
	74	+	t.Fatal("daemon child did not report its ancestry")
	75	+	}
	76	+	t.Cleanup(func() { _ = syscall.Kill(pid, syscall.SIGKILL) })
	77	+
	78	+	status, err := os.ReadFile(filepath.Join("/proc", strconv.Itoa(pid), "status"))
	79	+	if err != nil {
	80	+	t.Fatalf("read daemon status: %v", err)
	81	+	}
	82	+	if strings.Contains(string(status), fmt.Sprintf("PPid:\t%d\n", os.Getpid())) {
	83	+	t.Fatalf("daemon PPID = client PID %d; daemon remains in the client process tree", os.Getpid())
	84	+	}
	85	+	if err := syscall.Kill(pid, 0); err != nil {
	86	+	t.Fatalf("daemon exited with its intermediate parent: %v", err)
	87	+	}
	88	+	}
	89	+
19	90		// fastBackoff keeps the retry loop snappy for tests.
20	91		var fastBackoff = backoffConfig{initial: time.Millisecond, max: 5 * time.Millisecond, total: 100 * time.Millisecond}
21	92


@@ -3,10 +3,14 @@ package app
3	import (	3		import (
4	"context"	4		"context"
5	"errors"	5		"errors"
		6	+	"fmt"
6	"os"	7		"os"
7	"path/filepath"	8		"path/filepath"
		9	+	"strconv"
		10	+	"strings"
8	"sync"	11		"sync"
9	"sync/atomic"	12		"sync/atomic"
		13	+	"syscall"
10	"testing"	14		"testing"
11	"time"	15		"time"
12		16

@@ -16,6 +20,73 @@ import (
16		20
17	var errDialFailed = errors.New("dial failed")	21		var errDialFailed = errors.New("dial failed")
18		22
		23	+	const spawnTestChildFileEnv = "VEV_SPAWN_TEST_CHILD_FILE"
		24	+
		25	+	func TestMain(m *testing.M) {
		26	+	if len(os.Args) == 2 {
		27	+	switch os.Args[1] {
		28	+	case "--daemon-launcher":
		29	+	if err := Run(os.Args[1:]); err != nil {
		30	+	os.Exit(2)
		31	+	}
		32	+	os.Exit(0)
		33	+	case "--daemon":
		34	+	path := os.Getenv(spawnTestChildFileEnv)
		35	+	if path == "" {
		36	+	os.Exit(2)
		37	+	}
		38	+	if err := os.WriteFile(path, []byte(strconv.Itoa(os.Getpid())), 0o600); err != nil {
		39	+	os.Exit(2)
		40	+	}
		41	+	for {
		42	+	time.Sleep(time.Hour)
		43	+	}
		44	+	}
		45	+	}
		46	+	os.Exit(m.Run())
		47	+	}
		48	+
		49	+	func TestRealSpawnReparentsDaemonBeforeReturning(t *testing.T) {
		50	+	childFile := filepath.Join(t.TempDir(), "child")
		51	+	t.Setenv(spawnTestChildFileEnv, childFile)
		52	+
		53	+	if err := realSpawn(); err != nil {
		54	+	t.Fatalf("realSpawn: %v", err)
		55	+	}
		56	+
		57	+	var pid int
		58	+	deadline := time.Now().Add(2 * time.Second)
		59	+	for time.Now().Before(deadline) {
		60	+	data, err := os.ReadFile(childFile)
		61	+	if err == nil {
		62	+	pid, err = strconv.Atoi(string(data))
		63	+	if err != nil {
		64	+	t.Fatalf("parse child PID: %v", err)
		65	+	}
		66	+	break
		67	+	}
		68	+	if !errors.Is(err, os.ErrNotExist) {
		69	+	t.Fatalf("read child process record: %v", err)
		70	+	}
		71	+	time.Sleep(10 * time.Millisecond)
		72	+	}
		73	+	if pid == 0 {
		74	+	t.Fatal("daemon child did not report its ancestry")
		75	+	}
		76	+	t.Cleanup(func() { _ = syscall.Kill(pid, syscall.SIGKILL) })
		77	+
		78	+	status, err := os.ReadFile(filepath.Join("/proc", strconv.Itoa(pid), "status"))
		79	+	if err != nil {
		80	+	t.Fatalf("read daemon status: %v", err)
		81	+	}
		82	+	if strings.Contains(string(status), fmt.Sprintf("PPid:\t%d\n", os.Getpid())) {
		83	+	t.Fatalf("daemon PPID = client PID %d; daemon remains in the client process tree", os.Getpid())
		84	+	}
		85	+	if err := syscall.Kill(pid, 0); err != nil {
		86	+	t.Fatalf("daemon exited with its intermediate parent: %v", err)
		87	+	}
		88	+	}
		89	+
19	// fastBackoff keeps the retry loop snappy for tests.	90		// fastBackoff keeps the retry loop snappy for tests.
20	var fastBackoff = backoffConfig{initial: time.Millisecond, max: 5 * time.Millisecond, total: 100 * time.Millisecond}	91		var fastBackoff = backoffConfig{initial: time.Millisecond, max: 5 * time.Millisecond, total: 100 * time.Millisecond}
21		92