Merge pull request 'Keep gitea's forge-account provider running when a watch is refused, so a later grant is picked up (issue 955)' (#342) from fix/955-provider-keeps-running into main

modules/gitea/cmd/forge-account-provider/main.go

@@ -115,9 +115,7 @@ func main() {
 		pass()
 	}()
 	go func() {
-		if err := p.Run(ctx); err != nil {
+		svc.KeepRunning(ctx, p.Run)
-			say("the provider stopped: %v", err)
-		}
 	}()
 	if err := stdio.Serve("", nil); err != nil {
 		say("%v", err)

modules/gitea/cmd/forge-account-provider/service.go

@@ -407,3 +407,41 @@ func (b *valuesBus) Delete(ctx context.Context, bucket, key string) error {
 	b.mu.Unlock()
 	return b.Bus.Delete(ctx, bucket, key)
 }
+
+// Backoff of KeepRunning: the first pause and the longest.
+const (
+	restartFirst = 5 * time.Second
+	restartLast  = 2 * time.Minute
+)
+
+// KeepRunning runs the SDK provider and starts it again whenever it ends, however it ended (a watch or a read the
+// runtime refuses, because a grant has not reached it yet, ends Run): a provider stopped for good answers no consumer
+// until something restarts the process, and nothing does when its bundle is unchanged (issue 955). It says so once
+// per spell of the same error, pauses 5 s doubling to 2 min between starts, and returns only when ctx ends.
+func (s *Service) KeepRunning(ctx context.Context, run func(context.Context) error) {
+	pause := restartFirst
+	lastSaid := ""
+	for ctx.Err() == nil {
+		started := time.Now()
+		err := run(ctx)
+		if ctx.Err() != nil {
+			return
+		}
+		if time.Since(started) >= restartLast {
+			// It served a while: a new spell.
+			pause, lastSaid = restartFirst, ""
+		}
+		why := "it returned without an error"
+		if err != nil {
+			why = err.Error()
+		}
+		if why != lastSaid {
+			s.cfg.Log("the provider stopped: %s; starting it again every %s up to %s until it stays up", why, pause, restartLast)
+			lastSaid = why
+		}
+		s.Pause(pause)
+		if pause *= 2; pause > restartLast {
+			pause = restartLast
+		}
+	}
+}

modules/gitea/cmd/forge-account-provider/service_test.go

@@ -471,3 +471,72 @@ func TestAPassBeforeTheKeysAreLoadedRemovesAndDeletesNothing(t *testing.T) {
 		t.Fatalf("keys = %s", got)
 	}
 }
+
+// Issue 955: a watch the runtime refuses (a grant that comes later) must not end the provider for good.
+func TestKeepRunningStartsTheProviderAgainAfterARefusedWatch(t *testing.T) {
+	r := newRig(t)
+	var pauses []time.Duration
+	var said []string
+	r.svc.Pause = func(d time.Duration) { pauses = append(pauses, d) }
+	r.svc.cfg.Log = func(f string, a ...any) { said = append(said, fmt.Sprintf(f, a...)) }
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	calls := 0
+	serving := make(chan struct{})
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		r.svc.KeepRunning(ctx, func(ctx context.Context) error {
+			calls++
+			if calls <= 2 {
+				return errors.New(`watching mesh-controller_assignments: gitea reads no state called "mesh-controller.assignments"`)
+			}
+			close(serving)
+			<-ctx.Done()
+			return ctx.Err()
+		})
+	}()
+	select {
+	case <-serving:
+	case <-time.After(5 * time.Second):
+		t.Fatal("the provider was not started again after its watch was refused")
+	}
+	stop()
+	<-done
+	if calls != 3 {
+		t.Fatalf("run %d times", calls)
+	}
+	if len(pauses) != 2 || pauses[0] != 5*time.Second || pauses[1] != 10*time.Second {
+		t.Fatalf("the backoff is %v", pauses)
+	}
+	if len(said) == 0 || !strings.Contains(said[0], "mesh-controller.assignments") {
+		t.Fatalf("a refusal is said loudly: %v", said)
+	}
+	loud := 0
+	for _, s := range said {
+		if strings.Contains(s, "stopped") {
+			loud++
+		}
+	}
+	if loud != 1 {
+		t.Fatalf("the same refusal is said %d times in one spell: %v", loud, said)
+	}
+}
+
+func TestKeepRunningBacksOffToTwoMinutes(t *testing.T) {
+	r := newRig(t)
+	var pauses []time.Duration
+	r.svc.Pause = func(d time.Duration) { pauses = append(pauses, d) }
+	ctx, stop := context.WithCancel(context.Background())
+	calls := 0
+	r.svc.KeepRunning(ctx, func(context.Context) error {
+		if calls++; calls == 9 {
+			stop()
+		}
+		return errors.New("refused")
+	})
+	last := pauses[len(pauses)-1]
+	if pauses[0] != 5*time.Second || last != 2*time.Minute {
+		t.Fatalf("the backoff is %v", pauses)
+	}
+}