libimage: pull: increase timeout running under systemd

Set the `EXTEND_TIMEOUT_USEC` over DBUS when pulling an image from a
registry and when running under systemd.  This will prevent a frequent
issue when running Quadlets and exceeding the default systemd start
timeout of 90 seconds when pulling the image takes too long.

Fixes: containers/podman/issues/18353
Signed-off-by: Valentin Rothberg <[email protected]>

libimage/pull.go

@@ -5,6 +5,8 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"net"
+	"os"
 	"runtime"
 	"strings"
 	"time"
@@ -379,6 +381,43 @@ func (r *Runtime) copyFromRegistry(ctx context.Context, ref types.ImageReference
 		return nil, err
 	}
 
+	// Avoid running out of time when running inside a systemd unit by
+	// regularly increasing the timeout.
+	if socketPath, ok := os.LookupEnv("NOTIFY_SOCKET"); ok {
+		socketAddr := &net.UnixAddr{
+			Name: socketPath,
+			Net:  "unixgram",
+		}
+		conn, err := net.DialUnix(socketAddr.Net, nil, socketAddr)
+		if err != nil {
+			return nil, err
+		}
+		defer conn.Close()
+
+		timeout := 20 * time.Second
+		increaseTimeout := func() {
+			// See upstream systemd change adding support for EXTEND_TIMEOUT_USEC:
+			// https://github.com/systemd/systemd/commit/a327431bd168b2f327f3cd422379e213c643f2a5
+			if _, err := conn.Write([]byte("EXTEND_TIMEOUT_USEC=" + timeout.Microseconds())); err != nil {
+				logrus.Errorf("Increasing EXTEND_TIMEOUT_USEC failed: %v", err)
+			}
+		}
+
+		increaseTimeout()
+		socketCtx, cancel := context.WithCancel(ctx)
+		defer cancel()
+		go func() {
+			for {
+				select {
+				case <-socketCtx.Done():
+					return
+				case <-time.After(timeout):
+					increaseTimeout()
+				}
+			}
+		}()
+	}
+
 	if !options.AllTags {
 		return r.copySingleImageFromRegistry(ctx, inputName, pullPolicy, options)
 	}