perf: SSH uploads fetch pieces ahead and downloads post them side by side

internal/agent/ordered_test.go

@@ -0,0 +1,45 @@
+package agent
+
+import (
+	"context"
+	"math/rand"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// Pieces fetched side by side are used in their order, never more than
+// depth at once in flight.
+func TestOrderedKeepsOrder(t *testing.T) {
+	var inFlight, most int32
+	var got []int
+	err := ordered(context.Background(), 20, 3, func(ctx context.Context, i int) ([]byte, error) {
+		n := atomic.AddInt32(&inFlight, 1)
+		for {
+			m := atomic.LoadInt32(&most)
+			if n <= m || atomic.CompareAndSwapInt32(&most, m, n) {
+				break
+			}
+		}
+		time.Sleep(time.Duration(rand.Intn(5)) * time.Millisecond)
+		atomic.AddInt32(&inFlight, -1)
+		return []byte{byte(i)}, nil
+	}, func(i int, b []byte) error {
+		if int(b[0]) != i {
+			t.Errorf("piece %d carries %d", i, b[0])
+		}
+		got = append(got, i)
+		return nil
+	})
+	if err != nil || len(got) != 20 {
+		t.Fatalf("got %v, %v", got, err)
+	}
+	for i, v := range got {
+		if v != i {
+			t.Fatalf("order: %v", got)
+		}
+	}
+	if most > 3 || most < 2 {
+		t.Fatalf("at most %d in flight", most)
+	}
+}