fix: drop stalled HTTP/2 connections in action clones (#1266)

go-git sends all concurrent action clones from a host over one HTTP/2 connection, and Go sends no health-check pings by default. When that connection stalls, every clone on it hangs until the job timeout and cancellation can't free them, as in https://gitea.com/gitea/runner/actions/runs/1003885/jobs/1361593.

Ping idle connections so a stalled one is dropped after about 45s.

---------

Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1266
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>

act/common/git/git.go

@@ -9,11 +9,13 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	nethttp "net/http"
 	"os"
 	"path"
 	"regexp"
 	"strings"
 	"sync"
+	"time"
 
 	"gitea.com/gitea/runner/act/common"
 	"gitea.com/gitea/runner/internal/pkg/lock"
@@ -22,6 +24,8 @@ import (
 	"github.com/go-git/go-git/v5/config"
 	"github.com/go-git/go-git/v5/plumbing"
 	"github.com/go-git/go-git/v5/plumbing/storer"
+	"github.com/go-git/go-git/v5/plumbing/transport"
+	"github.com/go-git/go-git/v5/plumbing/transport/client"
 	"github.com/go-git/go-git/v5/plumbing/transport/http"
 	"github.com/mattn/go-isatty"
 	log "github.com/sirupsen/logrus"
@@ -38,6 +42,20 @@ var (
 	ErrShortRef = errors.New("short SHA references are not supported")
 )
 
+func init() {
+	httpClient := newHTTPClient(&nethttp.HTTP2Config{SendPingTimeout: 30 * time.Second})
+	client.InstallProtocol("http", httpClient)
+	client.InstallProtocol("https", httpClient)
+}
+
+// newHTTPClient pings idle HTTP/2 connections to drop stalled ones.
+func newHTTPClient(h2 *nethttp.HTTP2Config) transport.Transport {
+	defaultTransport, _ := nethttp.DefaultTransport.(*nethttp.Transport)
+	httpTransport := defaultTransport.Clone()
+	httpTransport.HTTP2 = h2
+	return http.NewClient(&nethttp.Client{Transport: httpTransport})
+}
+
 // AcquireCloneLock returns an unlock function after locking the per-directory mutex for dir.
 // Only concurrent operations targeting the same directory are serialized; clones into different directories run in parallel.
 // Callers reading files inside dir (e.g. tarring a checked-out action into a job container) must hold this lock too,

act/common/git/git_test.go

@@ -6,7 +6,9 @@ package git
 
 import (
 	"context"
+	"crypto/tls"
 	"fmt"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -756,3 +758,25 @@ func TestNewGitCloneExecutorFetchHonoursContext(t *testing.T) {
 		t.Fatal("fetch ignored context cancellation")
 	}
 }
+
+func TestNewGitCloneExecutorFailsOnSilentHTTP2Connection(t *testing.T) {
+	server := httptest.NewUnstartedServer(nil)
+	server.EnableHTTP2 = true
+	server.Config.TLSNextProto = map[string]func(*http.Server, *tls.Conn, http.Handler){
+		"h2": func(_ *http.Server, conn *tls.Conn, _ http.Handler) { _, _ = io.Copy(io.Discard, conn) },
+	}
+	server.StartTLS()
+	t.Cleanup(server.Close)
+
+	previous := gogitclient.Protocols["https"]
+	gogitclient.InstallProtocol("https", newHTTPClient(&http.HTTP2Config{SendPingTimeout: 50 * time.Millisecond, PingTimeout: 50 * time.Millisecond}))
+	t.Cleanup(func() { gogitclient.InstallProtocol("https", previous) })
+
+	ctx, cancel := context.WithTimeout(t.Context(), 10*time.Second)
+	defer cancel()
+	err := NewGitCloneExecutor(NewGitCloneExecutorInput{
+		URL: server.URL + "/action", Ref: "main", Dir: filepath.Join(t.TempDir(), "action"), InsecureSkipTLS: true,
+	})(ctx)
+	require.Error(t, err)
+	require.NoError(t, ctx.Err())
+}