Local control channel: unprivileged users can trigger --force-update via the running service

cmd/gpu-turnstile/main.go

@@ -17,11 +17,13 @@ import (
 	"path/filepath"
 	"runtime"
 	"strings"
+	"sync"
 	"syscall"
 	"time"
 
 	"gpu-turnstile/internal/comfy"
 	"gpu-turnstile/internal/config"
+	"gpu-turnstile/internal/control"
 	"gpu-turnstile/internal/game"
 	"gpu-turnstile/internal/lock"
 	"gpu-turnstile/internal/metrics"
@@ -98,6 +100,7 @@ Usage:
   gpu-turnstile -v | --version                                print just the version
   gpu-turnstile --force-update                                check for a signed update now,
                                                               apply it and restart the service
+                                                              (no admin needed when the service runs)
   gpu-turnstile -h | --help                                   this help
 
 Options:
@@ -347,6 +350,13 @@ func forceUpdateCommand(configPath string, elevatedChild bool) int {
 		fmt.Printf("%s: APP_VER=dev, updates disabled\n", versionLine())
 		return 0
 	}
+	// A running service can do the privileged work itself (its account owns
+	// the install dir and it knows when the GPU is idle): ask it over the
+	// local control channel first, no admin rights needed. Fails fast when
+	// no service is listening, in which case we do the direct check below.
+	if reply, err := control.Ask(control.CmdUpdateNow); err == nil {
+		return printControlReply(reply)
+	}
 	u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}
 
 	// Single-shot: one attempt, fail fast when the server is unreachable
@@ -397,6 +407,17 @@ func forceUpdateCommand(configPath string, elevatedChild bool) int {
 	return 0
 }
 
+// printControlReply prints the service's answer to a control-channel
+// request: "OK ..." on stdout (exit 0), "ERR ..." on stderr (exit 1).
+func printControlReply(reply string) int {
+	if msg, ok := strings.CutPrefix(reply, "ERR "); ok {
+		fmt.Fprintf(os.Stderr, "gpu-turnstile: %s\n", msg)
+		return 1
+	}
+	fmt.Println("gpu-turnstile: " + strings.TrimPrefix(reply, "OK "))
+	return 0
+}
+
 // reportElevatedUpdate prints the parent's summary of an elevated
 // --force-update child: exitCodeStaged means the child staged a new binary,
 // 0 means it found nothing to do. to is the tag the parent's own check
@@ -688,7 +709,35 @@ func run(ctx context.Context, cfg config.Config, log *slog.Logger, logOut io.Wri
 	service.StartWatchdog(ctx)
 
 	if cfg.AutoUpdate {
-		go updateLoop(ctx, cfg, log, lk, isService)
+		exePath, err := os.Executable()
+		if err != nil {
+			log.Warn("auto-update disabled: cannot locate executable", "err", err)
+		} else {
+			u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}
+			// applyStaged is shared by the hourly loop and the control
+			// channel; the once guard keeps a second trigger from
+			// double-waiting on the GPU lock.
+			var once sync.Once
+			applyStaged := func(to string) {
+				if !isService {
+					log.Warn("auto-update: new binary staged; restart gpu-turnstile to apply", "version", to)
+					return
+				}
+				log.Warn("auto-update: staged; restarting once the GPU is idle", "version", to)
+				once.Do(func() {
+					go func() {
+						if waitForIdle(ctx, lk, 24*time.Hour) {
+							log.Warn("auto-update: restarting to apply update")
+							os.Exit(exitCodeUpdate)
+						}
+					}()
+				})
+			}
+			go updateLoop(ctx, cfg.UpdateInterval, log, u, exePath, applyStaged)
+			if isService {
+				serveControl(ctx, log, u, exePath, applyStaged)
+			}
+		}
 	}
 
 	select {
@@ -830,42 +879,63 @@ func healthLoop(ctx context.Context, interval, probeTimeout time.Duration, log *
 	}
 }
 
-// updateLoop checks for signed updates on startup and every UPDATE_INTERVAL.
-// In service mode a staged update is applied by exiting with exitCodeUpdate
-// once the GPU lock is idle; the service recovery configuration restarts the
-// process with the new binary. Interactively it only logs.
-func updateLoop(ctx context.Context, cfg config.Config, log *slog.Logger, lk *lock.Lock, isService bool) {
-	exePath, err := os.Executable()
-	if err != nil {
-		log.Warn("auto-update disabled: cannot locate executable", "err", err)
-		return
-	}
-	u := &update.Updater{Repo: cfg.UpdateRepo, Asset: cfg.UpdateAsset, Version: version, Desired: cfg.AppVersion, Log: log}
+// updateLoop checks for signed updates on startup and every interval;
+// applyStaged decides what a staged update means (restart when idle as a
+// service, log only interactively).
+func updateLoop(ctx context.Context, interval time.Duration, log *slog.Logger, u *update.Updater, exePath string, applyStaged func(to string)) {
 	for {
 		staged, to, err := u.Check(ctx, exePath)
 		if err != nil && ctx.Err() == nil {
 			log.Warn("auto-update check failed", "err", err)
 		}
 		if staged {
-			if !isService {
-				log.Warn("auto-update: new binary staged; restart gpu-turnstile to apply", "version", to)
-				return
-			}
-			log.Warn("auto-update: staged; restarting once the GPU is idle", "version", to)
-			if waitForIdle(ctx, lk, 24*time.Hour) {
-				log.Warn("auto-update: restarting to apply update")
-				os.Exit(exitCodeUpdate)
-			}
+			applyStaged(to)
 			return
 		}
 		select {
 		case <-ctx.Done():
 			return
-		case <-time.After(cfg.UpdateInterval):
+		case <-time.After(interval):
 		}
 	}
 }
 
+// serveControl opens the local control channel (named pipe on Windows,
+// unix socket on Linux) so unprivileged local users can trigger an update
+// check via --force-update without admin rights. The check payload is
+// signature-verified regardless of who asks; triggers are rate-limited to
+// one per minute so the channel cannot be used to spam restarts.
+func serveControl(ctx context.Context, log *slog.Logger, u *update.Updater, exePath string, applyStaged func(to string)) {
+	var mu sync.Mutex
+	var lastTrigger time.Time
+	h := func(cmd string) string {
+		if cmd != control.CmdUpdateNow {
+			return "ERR unknown command: " + cmd
+		}
+		mu.Lock()
+		if wait := time.Minute - time.Since(lastTrigger); wait > 0 {
+			mu.Unlock()
+			return fmt.Sprintf("ERR rate limited: retry in %ds", int(wait.Seconds())+1)
+		}
+		lastTrigger = time.Now()
+		mu.Unlock()
+		cctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+		defer cancel()
+		staged, to, err := u.Check(cctx, exePath)
+		if err != nil {
+			return "ERR update check failed: " + err.Error()
+		}
+		if !staged {
+			return "OK " + version + " is up to date"
+		}
+		applyStaged(to)
+		return "OK updated from " + version + " to " + to + "; the service restarts once the GPU is idle"
+	}
+	if err := control.Serve(ctx, h, log); err != nil {
+		log.Warn("control channel disabled", "err", err)
+	}
+}
+
 // waitForIdle polls the lock until no LLM or image work is active or
 // pending, max at most. Returns false on timeout or cancellation.
 func waitForIdle(ctx context.Context, lk *lock.Lock, max time.Duration) bool {

internal/control/control.go

@@ -0,0 +1,61 @@
+// Package control exposes a local-only command channel into a running
+// gpu-turnstile service: a named pipe on Windows, a unix socket on Linux.
+// It lets unprivileged local users ask the service to do privileged work
+// that is safe to offer — currently triggering an update check, whose
+// payload is signature-verified regardless of who asks. The channel never
+// accepts data beyond a one-word command, and the server rate-limits
+// triggers, so the worst a local user can cause is a cheap, throttled
+// check and a GPU-idle-gated restart onto a signed binary.
+//
+// Protocol: the client writes one command line, the server answers with
+// one reply line ("OK ..." or "ERR ...") and hangs up.
+package control
+
+import (
+	"bufio"
+	"errors"
+	"fmt"
+	"io"
+	"strings"
+)
+
+// CmdUpdateNow asks the service to check for, stage and (once the GPU is
+// idle) restart onto a signed update immediately.
+const CmdUpdateNow = "update-now"
+
+// ErrUnavailable means no running service offers the control channel.
+var ErrUnavailable = errors.New("control channel unavailable")
+
+// Handler answers one command; the returned string is sent back as one
+// line. It must start with "OK " or "ERR ".
+type Handler func(cmd string) string
+
+// serveConn runs the line protocol on one accepted connection.
+func serveConn(c io.ReadWriteCloser, h Handler) {
+	defer c.Close()
+	line, err := bufio.NewReader(io.LimitReader(c, 4096)).ReadString('\n')
+	cmd := strings.TrimSpace(line)
+	if cmd == "" {
+		if err != nil {
+			return
+		}
+		fmt.Fprintln(c, "ERR empty command")
+		return
+	}
+	fmt.Fprintln(c, h(cmd))
+}
+
+// readReply writes cmd and reads the server's one-line reply.
+func readReply(c io.ReadWriteCloser, cmd string) (string, error) {
+	if _, err := fmt.Fprintln(c, cmd); err != nil {
+		return "", err
+	}
+	// The server hangs up after its reply; a broken-pipe error after the
+	// last byte still leaves the reply in the buffer.
+	data, _ := io.ReadAll(io.LimitReader(c, 4096))
+	line := strings.TrimSpace(string(data))
+	if line == "" {
+		return "", ErrUnavailable
+	}
+	return line, nil
+}

internal/control/control_linux.go

@@ -0,0 +1,54 @@
+//go:build linux
+
+package control
+
+import (
+	"context"
+	"log/slog"
+	"net"
+	"os"
+	"time"
+)
+
+// sockPath lives in the unit's RuntimeDirectory; mode 0666 lets every
+// local user ask, nothing can reach it from off the machine.
+const sockPath = "/run/gpu-turnstile/control.sock"
+
+// Serve starts the socket listener in the background and returns; only a
+// setup failure is reported. Each client connection is answered in its own
+// goroutine.
+func Serve(ctx context.Context, h Handler, log *slog.Logger) error {
+	os.Remove(sockPath) // stale socket from a previous run
+	ln, err := net.Listen("unix", sockPath)
+	if err != nil {
+		return err
+	}
+	if err := os.Chmod(sockPath, 0o666); err != nil {
+		ln.Close()
+		return err
+	}
+	go func() {
+		<-ctx.Done()
+		ln.Close()
+	}()
+	go func() {
+		for {
+			c, err := ln.Accept()
+			if err != nil {
+				return // shutting down
+			}
+			go serveConn(c, h)
+		}
+	}()
+	return nil
+}
+
+// Ask sends one command to the running service and returns its reply.
+func Ask(cmd string) (string, error) {
+	c, err := net.DialTimeout("unix", sockPath, 2*time.Second)
+	if err != nil {
+		return "", ErrUnavailable
+	}
+	defer c.Close()
+	return readReply(c, cmd)
+}

internal/control/control_other.go

@@ -0,0 +1,18 @@
+//go:build !windows && !linux
+
+package control
+
+import (
+	"context"
+	"log/slog"
+)
+
+// Serve is a no-op on platforms without a control channel.
+func Serve(ctx context.Context, h Handler, log *slog.Logger) error {
+	return ErrUnavailable
+}
+
+// Ask always reports the channel as unavailable.
+func Ask(cmd string) (string, error) {
+	return "", ErrUnavailable
+}

internal/control/control_test.go

@@ -0,0 +1,46 @@
+package control
+
+import (
+	"errors"
+	"net"
+	"strings"
+	"testing"
+)
+
+func TestRoundTrip(t *testing.T) {
+	server, client := net.Pipe()
+	go serveConn(server, func(cmd string) string {
+		if cmd != CmdUpdateNow {
+			return "ERR unknown command: " + cmd
+		}
+		return "OK v0.2.2 is up to date"
+	})
+	reply, err := readReply(client, CmdUpdateNow)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if reply != "OK v0.2.2 is up to date" {
+		t.Fatalf("reply = %q", reply)
+	}
+
+	server2, client2 := net.Pipe()
+	go serveConn(server2, func(cmd string) string { return "ERR unknown command: " + cmd })
+	reply, err = readReply(client2, "bogus")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.HasPrefix(reply, "ERR ") {
+		t.Fatalf("reply = %q, want ERR prefix", reply)
+	}
+}
+
+func TestEmptyReplyIsUnavailable(t *testing.T) {
+	server, client := net.Pipe()
+	go serveConn(server, func(cmd string) string {
+		server.Close() // hang up without answering
+		return ""
+	})
+	if _, err := readReply(client, CmdUpdateNow); !errors.Is(err, ErrUnavailable) {
+		t.Fatalf("err = %v, want ErrUnavailable", err)
+	}
+}

internal/control/control_windows.go

@@ -0,0 +1,112 @@
+//go:build windows
+
+package control
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"os"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+// pipePath is a kernel-local named pipe: no TCP, no firewall prompt.
+const pipePath = `\\.\pipe\gpu-turnstile`
+
+// sddlPipe grants full access to Administrators, SYSTEM and the pipe owner,
+// and read+write to authenticated users — except network logons, so the
+// pipe cannot be reached from another machine over SMB.
+const sddlPipe = "D:(D;;GRGW;;;NU)(A;;GA;;;BA)(A;;GA;;;SY)(A;;GA;;;OW)(A;;GRGW;;;AU)"
+
+var (
+	procConvertSDDL = windows.NewLazySystemDLL("advapi32.dll").
+			NewProc("ConvertStringSecurityDescriptorToSecurityDescriptorW")
+	procWaitNamedPipe = windows.NewLazySystemDLL("kernel32.dll").
+				NewProc("WaitNamedPipeW")
+)
+
+func waitNamedPipe(name *uint16, timeout uint32) error {
+	r, _, err := procWaitNamedPipe.Call(uintptr(unsafe.Pointer(name)), uintptr(timeout))
+	if r == 0 {
+		return err
+	}
+	return nil
+}
+
+func securityAttributesFromSDDL(sddl string) (*windows.SecurityAttributes, error) {
+	s, err := windows.UTF16PtrFromString(sddl)
+	if err != nil {
+		return nil, err
+	}
+	var sd *uint16 // SECURITY_DESCRIPTOR*, kept for the process lifetime
+	r, _, callErr := procConvertSDDL.Call(
+		uintptr(unsafe.Pointer(s)), 1, /* SDDL_REVISION_1 */
+		uintptr(unsafe.Pointer(&sd)), 0)
+	if r == 0 {
+		return nil, fmt.Errorf("invalid SDDL: %w", callErr)
+	}
+	sa := &windows.SecurityAttributes{
+		Length:             uint32(unsafe.Sizeof(windows.SecurityAttributes{})),
+		SecurityDescriptor: (*windows.SECURITY_DESCRIPTOR)(unsafe.Pointer(sd)),
+	}
+	return sa, nil
+}
+
+// Serve starts the pipe listener in the background and returns; only a
+// setup failure is reported. Each client connection is answered in its own
+// goroutine. On shutdown the process exit reaps everything.
+func Serve(ctx context.Context, h Handler, log *slog.Logger) error {
+	sa, err := securityAttributesFromSDDL(sddlPipe)
+	if err != nil {
+		return err
+	}
+	name, err := windows.UTF16PtrFromString(pipePath)
+	if err != nil {
+		return err
+	}
+	go func() {
+		for ctx.Err() == nil {
+			pipe, err := windows.CreateNamedPipe(name,
+				windows.PIPE_ACCESS_DUPLEX,
+				windows.PIPE_TYPE_BYTE|windows.PIPE_READMODE_BYTE|windows.PIPE_WAIT,
+				16, 4096, 4096, 0, sa)
+			if err != nil {
+				log.Warn("control channel stopped", "err", err)
+				return
+			}
+			go func() {
+				// Blocks until a client connects; on process exit the
+				// handle goes away with everything else.
+				if err := windows.ConnectNamedPipe(pipe, nil); err != nil {
+					windows.CloseHandle(pipe)
+					return
+				}
+				f := os.NewFile(uintptr(pipe), pipePath)
+				serveConn(f, h) // closes f, and with it the pipe handle
+			}()
+		}
+	}()
+	return nil
+}
+
+// Ask sends one command to the running service and returns its reply.
+func Ask(cmd string) (string, error) {
+	name, err := windows.UTF16PtrFromString(pipePath)
+	if err != nil {
+		return "", err
+	}
+	if err := waitNamedPipe(name, 2000); err != nil {
+		return "", ErrUnavailable
+	}
+	handle, err := windows.CreateFile(name,
+		windows.GENERIC_READ|windows.GENERIC_WRITE, 0, nil,
+		windows.OPEN_EXISTING, 0, 0)
+	if err != nil {
+		return "", ErrUnavailable
+	}
+	f := os.NewFile(uintptr(handle), pipePath)
+	defer f.Close()
+	return readReply(f, cmd)
+}

internal/service/service_linux.go

@@ -95,6 +95,8 @@ RestartSec=5s
 
 DynamicUser=yes
 StateDirectory=%s
+RuntimeDirectory=%s
+RuntimeDirectoryMode=0755
 %sProtectSystem=strict
 ProtectHome=yes
 PrivateTmp=yes
@@ -117,7 +119,7 @@ SystemCallErrorNumber=EPERM
 
 [Install]
 WantedBy=multi-user.target
-`, exePath, configPath, Name, bind)
+`, exePath, configPath, Name, Name, bind)
 }
 
 // copyFile copies src to dst, creating dst with the given mode.

internal/service/service_linux_test.go

@@ -17,6 +17,7 @@ func TestRenderUnit(t *testing.T) {
 		"WantedBy=multi-user.target",
 		"DynamicUser=yes",
 		"StateDirectory=gpu-turnstile",
+		"RuntimeDirectory=gpu-turnstile",
 		"ProtectSystem=strict",
 		"NoNewPrivileges=yes",
 		"RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6",