Signals end every command, not only ssh to and ssh install (closes #48)

SIGINT, SIGTERM and SIGHUP were caught for the whole run, but only the
ssh and sftp children acted on them: the mnemonic prompt waited for
Enter, and an interrupted `age encrypt -o` put the encryption of the
cut-off input in place. Now `ssh to` and `ssh install` catch them from
once the mnemonic is read until their cleanup has run, and everywhere
else they end the tool at once, except while `age encrypt -o` or
`age decrypt -o` writes. There the work runs in the background, and a
signal that comes before it ends, or within a tenth of a second after,
removes the unfinished file and ends the tool with status 1, since
Ctrl-C on a pipeline can end the input just before the signal arrives.

Model: opus-5-5

README.md

@@ -288,6 +288,18 @@ girl mad pet galaxy egg matter matrix prison refuse sense ordinary nose
 Errors go to standard error and the exit status is 1, except for `ssh to`, which
 passes through `ssh`'s own exit status.
 
+SIGINT, SIGTERM and SIGHUP end any command at once, at the mnemonic prompt too,
+with the status a shell gives a program killed by that signal (130 for SIGINT).
+An interrupted `age encrypt -o` or `age decrypt -o` leaves no file: it removes
+the unfinished file it was writing, leaves a file already at the named path as
+it was, and exits with status 1. It puts the file in place a tenth of a second
+after its input ends, and a signal in that time still counts: Ctrl-C on a
+pipeline also ends the program feeding it, so the input can end just before the
+signal arrives. While `ssh to` or `ssh install` has `ssh` or `sftp` running, the
+signal ends that program instead, the tool removes its agent socket or working
+files, and it exits with status 1, or for `ssh to` with `ssh`'s own status if
+`ssh` reported one.
+
 ## Entrypoints
 
 The repo adheres to the

internal/cli/age/age.go

@@ -3,10 +3,14 @@
 package age
 
 import (
+	"errors"
 	"fmt"
 	"io"
 	"os"
+	"os/signal"
 	"path/filepath"
+	"syscall"
+	"time"
 
 	"github.com/spf13/cobra"
 	"sneak.berlin/go/keyfunc/internal/agekey"
@@ -14,6 +18,18 @@ import (
 	"sneak.berlin/go/keyfunc/internal/derive"
 )
 
+// ErrInterrupted is returned when SIGINT, SIGTERM or SIGHUP came before
+// the file --output names was put in place.
+var ErrInterrupted = errors.New(
+	"interrupted by a signal; the output file was left as it was",
+)
+
+// signalWait is how long after the work has ended a signal still keeps
+// the new file from being put in place. Ctrl-C on "producer | keyfunc
+// age encrypt -o file" ends the producer as well, and the end of the
+// input can reach the work a moment before the signal reaches the tool.
+const signalWait = 100 * time.Millisecond
+
 // Command returns the age command and everything under it.
 func Command() *cobra.Command {
 	group := &cobra.Command{
@@ -128,8 +144,9 @@ func runDecrypt(cmd *cobra.Command, args []string) error {
 	return through(cmd, args, key.Decrypt)
 }
 
-// through opens the input and the output the arguments ask for, hands
-// them to the work, and finishes the output afterwards either way.
+// through opens the input the arguments ask for and hands it to the
+// work, with the file --output names to write to, or the command's own
+// output when it names none.
 func through(
 	cmd *cobra.Command, args []string,
 	work func(io.Writer, io.Reader) error,
@@ -141,14 +158,16 @@ func through(
 
 	defer closeSrc()
 
-	dst, done, err := output(cmd)
+	name, err := cmd.Flags().GetString("output")
 	if err != nil {
-		return err
+		return fmt.Errorf("reading the output file: %w", err)
 	}
 
-	err = work(dst, src)
+	if name == "" {
+		return work(cmd.OutOrStdout(), src)
+	}
 
-	return done(err)
+	return output(name, src, work)
 }
 
 // input returns what to read from: the named file, or the command's
@@ -167,35 +186,45 @@ func input(cmd *cobra.Command, args []string) (io.Reader, func(), error) {
 	return file, func() { _ = file.Close() }, nil
 }
 
-// output returns what to write to: a new file beside the one --output
-// names, or the command's own output when it names none. The second
-// result finishes the write, and is given whatever the work returned:
-// the new file takes the named file's place only when the work
-// succeeded, so a file that is already there survives a run that
-// failed.
-func output(cmd *cobra.Command) (io.Writer, func(error) error, error) {
-	name, err := cmd.Flags().GetString("output")
-	if err != nil {
-		return nil, nil, fmt.Errorf("reading the output file: %w", err)
-	}
+// output has the work write a new file beside the named one, and puts
+// the new file in the named file's place only when the work succeeded,
+// so a file that is already there survives a run that failed.
+//
+// Meanwhile SIGINT, SIGTERM and SIGHUP are caught. One that comes while
+// the work runs, or within signalWait after it has ended, wins: the new
+// file is removed and ErrInterrupted returned at once, without waiting
+// for the work, which may be blocked reading its input.
+func output(
+	name string, src io.Reader, work func(io.Writer, io.Reader) error,
+) error {
+	signals := make(chan os.Signal, 1)
+	signal.Notify(signals, syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP)
 
-	if name == "" {
-		return cmd.OutOrStdout(), func(failed error) error {
-			return failed
-		}, nil
-	}
+	defer signal.Stop(signals)
 
 	// The file is made in the same directory so that putting it in
 	// place is a rename and never a copy, and it is readable only by
 	// its owner, which is the mode it keeps once renamed.
 	file, err := os.CreateTemp(filepath.Dir(name), filepath.Base(name)+".")
 	if err != nil {
-		return nil, nil, fmt.Errorf("creating a file beside %s: %w", name, err)
+		return fmt.Errorf("creating a file beside %s: %w", name, err)
 	}
 
-	return file, func(failed error) error {
-		return finish(file, name, failed)
-	}, nil
+	worked := make(chan error, 1)
+
+	go func() { worked <- work(file, src) }()
+
+	select {
+	case failed := <-worked:
+		select {
+		case <-signals:
+		case <-time.After(signalWait):
+			return finish(file, name, failed)
+		}
+	case <-signals:
+	}
+
+	return finish(file, name, ErrInterrupted)
 }
 
 // finish closes the new file and puts it in the named file's place, or

internal/cli/age_test.go

@@ -1,10 +1,14 @@
 package cli_test
 
 import (
+	"io"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"strings"
+	"syscall"
 	"testing"
+	"time"
 
 	"github.com/stretchr/testify/require"
 	"sneak.berlin/go/keyfunc/internal/agekey"
@@ -90,6 +94,102 @@ func TestARefusedDecryptionLeavesTheOutputFileAlone(t *testing.T) {
 	require.Equal(t, "what was already there\n", string(kept))
 }
 
+func TestASignalStopsAnEncryptionAndLeavesNoFile(t *testing.T) {
+	t.Setenv(mnemonic.Variable, example())
+
+	for _, ending := range []os.Signal{
+		syscall.SIGTERM, syscall.SIGINT, syscall.SIGHUP,
+	} {
+		interrupted(t, ending, "encrypt", "the start of the secret\n", false)
+	}
+}
+
+func TestASignalStopsADecryptionAndLeavesNoFile(t *testing.T) {
+	t.Setenv(mnemonic.Variable, example())
+
+	// All of an encryption but its last byte, so the tool reads the
+	// header and then waits for the rest.
+	sealed := run(t, "age", "encrypt", written(t, "notes.txt", "the secret\n"))
+	cut := sealed[:len(sealed)-1]
+
+	for _, ending := range []os.Signal{
+		syscall.SIGTERM, syscall.SIGINT, syscall.SIGHUP,
+	} {
+		interrupted(t, ending, "decrypt", cut, false)
+	}
+}
+
+func TestASignalAsTheInputEndsLeavesNoFile(t *testing.T) {
+	t.Setenv(mnemonic.Variable, example())
+
+	sealed := run(t, "age", "encrypt", written(t, "notes.txt", "the secret\n"))
+
+	// Ctrl-C on "producer | keyfunc age encrypt -o file" ends the
+	// producer too, so the input ends just as the signal comes, with
+	// enough of it in hand for a whole encryption or decryption. Which
+	// of the two reaches the tool first varies, so it is tried often.
+	for range 25 {
+		interrupted(t, syscall.SIGINT, "encrypt", "the start of the secret\n", true)
+		interrupted(t, syscall.SIGINT, "decrypt", sealed, true)
+	}
+}
+
+// interrupted runs "age encrypt -o" or "age decrypt -o", as the
+// operation says, as a subprocess writing into a directory of its own
+// and reading the input from a pipe. It waits until the tool has begun
+// writing the file beside the one it was named, and sends it the
+// signal, then ends the input if endInput says so and otherwise leaves
+// it open. The tool has to end with status 1 and leave the directory
+// empty. A tool that went on reading would not end until the input
+// did; one that did not remove the file it was writing would leave it
+// there, with what it had written so far; one that put that file in
+// place because the input ended would leave the named file.
+func interrupted(
+	t *testing.T, signal os.Signal, operation, input string, endInput bool,
+) {
+	t.Helper()
+
+	name := operation + " " + signal.String()
+	directory := t.TempDir()
+
+	//nolint:gosec // the binary is this test's own, re-run as the tool
+	command := exec.CommandContext(
+		t.Context(), os.Args[0], "age", operation,
+		"-o", filepath.Join(directory, "notes"),
+	)
+
+	command.Env = append(os.Environ(), runAsTool+"=1")
+
+	producer, err := command.StdinPipe()
+	require.NoError(t, err)
+	require.NoError(t, command.Start())
+
+	_, err = io.WriteString(producer, input)
+	require.NoError(t, err)
+
+	// The file beside the named one is made once the mnemonic has been
+	// read, before any input is.
+	require.Eventually(t, func() bool {
+		entries, err := os.ReadDir(directory)
+
+		return err == nil && len(entries) > 0
+	}, 5*time.Second, 5*time.Millisecond)
+
+	require.NoError(t, command.Process.Signal(signal))
+
+	if endInput {
+		require.NoError(t, producer.Close())
+	}
+
+	waitForTool(t, name, command)
+
+	require.Equal(t, 1, command.ProcessState.ExitCode(), name)
+
+	left, err := os.ReadDir(directory)
+	require.NoError(t, err)
+	require.Empty(t, left, name)
+}
+
 // written puts the contents in a file of that name in a directory of
 // this test's own and returns the path to it.
 func written(t *testing.T, name, contents string) string {

internal/cli/cli.go

@@ -2,13 +2,10 @@
 package cli
 
 import (
-	"context"
 	"errors"
 	"fmt"
 	"os"
-	"os/signal"
 	"runtime/debug"
-	"syscall"
 
 	"github.com/spf13/cobra"
 	"sneak.berlin/go/keyfunc/internal/cli/age"
@@ -70,18 +67,15 @@ func Root() *cobra.Command {
 // ended with. ssh has already said whatever it had to say in that
 // case, so nothing more is printed.
 //
-// SIGINT, SIGTERM and SIGHUP cancel the command's context instead of
-// killing the process outright, so the child ssh or sftp ends and the
-// deferred cleanup that removes the agent socket and the install
-// working directory still runs.
+// SIGINT, SIGTERM and SIGHUP end the tool at once, as they end any Go
+// program, so a command waiting at the mnemonic prompt or reading what
+// it encrypts or decrypts goes no further. The exceptions catch the
+// signals to clean up first: "ssh to" and "ssh install" while they
+// have ssh or sftp running, so the child ends and their own cleanup
+// still runs, and "age encrypt -o" and "age decrypt -o" while they
+// write, so the unfinished file is removed.
 func Main() int {
-	ctx, stop := signal.NotifyContext(
-		context.Background(),
-		syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP,
-	)
-	defer stop()
-
-	err := Root().ExecuteContext(ctx)
+	err := Root().Execute()
 	if err == nil {
 		return 0
 	}

internal/cli/ssh/install.go

@@ -8,9 +8,11 @@ import (
 	"fmt"
 	"os"
 	"os/exec"
+	"os/signal"
 	"path/filepath"
 	"slices"
 	"strings"
+	"syscall"
 
 	"github.com/spf13/cobra"
 )
@@ -62,6 +64,17 @@ func install() *cobra.Command {
 				return err
 			}
 
+			// From here on a signal cancels the context, which
+			// sftp runs under, instead of ending the tool, so sftp
+			// ends and the working directory is still removed.
+			ctx, stop := signal.NotifyContext(
+				cmd.Context(),
+				syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP,
+			)
+			defer stop()
+
+			cmd.SetContext(ctx)
+
 			return add(cmd, args[0], args[1:], line)
 		},
 	}

internal/cli/ssh/to.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"os"
 	"os/exec"
+	"os/signal"
 	"slices"
 	"syscall"
 
@@ -42,7 +43,17 @@ func to() *cobra.Command {
 				return err
 			}
 
-			served, err := key.Serve(cmd.Context(), comment)
+			// From here until the agent is taken down, a signal
+			// cancels the context instead of ending the tool, so
+			// ssh ends and the socket and its directory are still
+			// removed.
+			ctx, stop := signal.NotifyContext(
+				cmd.Context(),
+				syscall.SIGINT, syscall.SIGTERM, syscall.SIGHUP,
+			)
+			defer stop()
+
+			served, err := key.Serve(ctx, comment)
 			if err != nil {
 				return err
 			}
@@ -53,7 +64,7 @@ func to() *cobra.Command {
 				"-o", "IdentityAgent=" + served.Socket(),
 			}, args)
 
-			return connect(cmd.Context(), argv)
+			return connect(ctx, argv)
 		},
 	}
 
@@ -76,7 +87,7 @@ func connect(ctx context.Context, argv []string) error {
 	command.Stdout = os.Stdout
 	command.Stderr = os.Stderr
 
-	// A cancelled context means a signal ended the tool. Send ssh a
+	// A cancelled context means a signal arrived. Send ssh a
 	// SIGTERM rather than the default kill, so it puts the terminal
 	// back the way it found it before it goes.
 	command.Cancel = func() error {

internal/cli/ssh_test.go

@@ -20,13 +20,13 @@ import (
 
 // runAsTool, set in the environment of a re-executed test binary, tells
 // TestMain to run the tool through Main rather than the suite, so the
-// signal test can drive the real signal path in a process it can send a
-// signal to.
+// signal tests can drive the real signal path in a process they can
+// send a signal to.
 const runAsTool = "KEYFUNC_TEST_RUN_AS_TOOL"
 
 // TestMain re-executes the test binary as the tool when runAsTool is
-// set, and otherwise runs the suite. The signal test starts the tool
-// this way, as a subprocess it can signal and watch clean up.
+// set, and otherwise runs the suite. The signal tests start the tool
+// this way, as a subprocess they can signal and watch end.
 func TestMain(m *testing.M) {
 	if os.Getenv(runAsTool) == "1" {
 		os.Exit(cli.Main())
@@ -209,6 +209,16 @@ fi
 sleep 5
 `
 
+// stalled is a stand-in for the system sftp that notes it has started
+// and then blocks, so a test can signal the tool while sftp is running
+// and watch it remove its working directory. The exec keeps the shell
+// from leaving a sleep behind that holds the output the tool reads sftp
+// through.
+const stalled = `
+touch "$KEYFUNC_TEST_STARTED"
+exec sleep 5
+`
+
 // pretended is where a stand-in writes down what it was asked to do.
 type pretended struct {
 	// home stands in for the home directory on the host.
@@ -544,7 +554,7 @@ func TestASignalTakesTheAgentDirectoryDown(t *testing.T) {
 // ssh that blocks, waits until the agent is up and ssh is running
 // against it, sends the tool the signal, and requires the agent socket
 // and its directory to be gone once the tool has ended. The subprocess
-// goes through Main and its signal handling, so with that handling
+// goes through Main and the command's signal handling, so with that handling
 // removed the signal kills the tool outright, no deferred cleanup runs,
 // the directory is left behind, and the check fails.
 func signalEndsTheTool(t *testing.T, name string, signal os.Signal) {
@@ -611,6 +621,56 @@ func waitForSocket(t *testing.T, noted string) string {
 	return socket
 }
 
+func TestASignalTakesTheInstallWorkingDirectoryDown(t *testing.T) {
+	t.Setenv(mnemonic.Variable, example())
+
+	for _, ending := range []os.Signal{
+		syscall.SIGTERM, syscall.SIGINT, syscall.SIGHUP,
+	} {
+		signalEndsTheInstall(t, ending.String(), ending)
+	}
+}
+
+// signalEndsTheInstall runs "ssh install" as a subprocess against a
+// stand-in sftp that blocks, with a temporary directory of the test's
+// own, waits until sftp is running, sends the tool the signal, and
+// requires the working directory the tool made there to be gone once
+// the tool has ended.
+func signalEndsTheInstall(t *testing.T, name string, signal os.Signal) {
+	t.Helper()
+
+	temporary := t.TempDir()
+	started := filepath.Join(t.TempDir(), "started")
+	t.Setenv("KEYFUNC_TEST_STARTED", started)
+	standIn(t, "sftp", stalled)
+
+	//nolint:gosec // the binary is this test's own, re-run as the tool
+	command := exec.CommandContext(
+		t.Context(), os.Args[0], subcommand, installing, host,
+	)
+
+	command.Env = append(os.Environ(), runAsTool+"=1", "TMPDIR="+temporary)
+	require.NoError(t, command.Start())
+
+	// sftp is started only once the working directory has been made.
+	require.Eventually(t, func() bool {
+		_, err := os.Stat(started)
+
+		return err == nil
+	}, 5*time.Second, 5*time.Millisecond)
+
+	working, err := os.ReadDir(temporary)
+	require.NoError(t, err)
+	require.Len(t, working, 1, name)
+
+	require.NoError(t, command.Process.Signal(signal))
+	waitForTool(t, name, command)
+
+	left, err := os.ReadDir(temporary)
+	require.NoError(t, err)
+	require.Empty(t, left, name)
+}
+
 func TestTheMnemonicIsNotHandedToSFTP(t *testing.T) {
 	t.Setenv(mnemonic.CommandVariable, "echo "+example())
 	t.Setenv(mnemonic.Variable, example())