feat!: serve the actions cache from the runner (#1229)

Make cache requests go through the runner and remove outbound traffic from the cache server. The runner forwards them to `cache.external_server` when one is set, and attaches itself to each job network so jobs can reach it.

Benefits:

1. Jobs need no http proxy rules for the cache server, which `@actions/http-client` mishandles for CIDR ranges in `no_proxy` (https://github.com/actions/toolkit/pull/2495).
2. Only runners reach a shared cache server, which needs no route to Gitea.
3. A runner container sharing the host Docker socket now caches, where jobs previously could not reach it.

Breaks:

1. Jobs that reach `external_server` but not their runner, such as a remote `DOCKER_HOST`. The built-in cache has always needed that reachability, so both modes now agree. Set `cache.host` and `cache.port` to an address those jobs reach.
2. Jobs on `container.network: bridge`, where Docker rejects the alias, so the runner cannot attach and their cache fails as it does today. Leave `container.network` empty, or name a user-defined network.

*Written by Claude.*

---------

Co-authored-by: bircni <[email protected]>
Reviewed-on: https://gitea.com/gitea/runner/pulls/1229
Reviewed-by: bircni <[email protected]>
Co-authored-by: silverwind <[email protected]>

act/artifactcache/handler.go

@@ -17,6 +17,7 @@ import (
 	"io"
 	"net"
 	"net/http"
+	"net/http/httputil"
 	"net/url"
 	"os"
 	"path/filepath"
@@ -57,10 +58,9 @@ type JobCredential struct {
 	Repo string `json:"repo"`
 
 	// Results is the instance whose artifact service this server forwards for the job, and
-	// InsecureTLS how the runner reaches it; see results.go. The tags are the wire format a
-	// remote runner registers with.
-	Results     string `json:"results"`
-	InsecureTLS bool   `json:"insecure_tls"`
+	// InsecureTLS how the runner reaches it; see results.go.
+	Results     string `json:"-"`
+	InsecureTLS bool   `json:"-"`
 
 	// PublicURL is this server as a reverse proxy makes the job reach it, not the listen address.
 	PublicURL string `json:"public_url"`
@@ -117,6 +117,7 @@ type Options struct {
 	Dir        string
 	OutboundIP string
 	Port       uint16
+	Upstream   string
 
 	// InternalSecret, when non-empty, enables a control-plane API at
 	// /_internal/{register,revoke} that lets a remote runner pre-register the
@@ -129,7 +130,7 @@ type Options struct {
 	Logger logrus.FieldLogger
 }
 
-// StartHandler opens the on-disk cache store and starts the HTTP server.
+// StartHandler starts an embedded cache or forwards cache requests to Upstream.
 func StartHandler(opts Options) (*Handler, error) {
 	dir, logger := opts.Dir, opts.Logger
 	h := &Handler{
@@ -147,6 +148,25 @@ func StartHandler(opts Options) (*Handler, error) {
 	logger = logger.WithField("module", "artifactcache")
 	h.logger = logger
 
+	if opts.OutboundIP != "" {
+		h.outboundIP = opts.OutboundIP
+	} else if ip := common.GetOutboundIP(); ip == nil {
+		return nil, errors.New("unable to determine outbound IP address")
+	} else {
+		h.outboundIP = ip.String()
+	}
+
+	if opts.Upstream != "" {
+		handler, err := h.cacheForwarder(opts.Upstream)
+		if err != nil {
+			return nil, err
+		}
+		if err := h.serve(opts.Port, handler); err != nil {
+			return nil, err
+		}
+		return h, nil
+	}
+
 	if dir == "" {
 		home, err := os.UserHomeDir()
 		if err != nil {
@@ -166,14 +186,6 @@ func StartHandler(opts Options) (*Handler, error) {
 	}
 	h.storage = storage
 
-	if opts.OutboundIP != "" {
-		h.outboundIP = opts.OutboundIP
-	} else if ip := common.GetOutboundIP(); ip == nil {
-		return nil, errors.New("unable to determine outbound IP address")
-	} else {
-		h.outboundIP = ip.String()
-	}
-
 	secret, err := loadOrCreateSecret(dir)
 	if err != nil {
 		return nil, err
@@ -201,34 +213,69 @@ func StartHandler(opts Options) (*Handler, error) {
 
 	h.gcCache()
 
+	if err := h.serve(opts.Port, router); err != nil {
+		return nil, err
+	}
+	return h, nil
+}
+
+func (h *Handler) serve(port uint16, handler http.Handler) error {
 	// Listen on all interfaces. Binding to outboundIP only would give no real
 	// security benefit (it is the LAN/internet-facing address either way) and
 	// can break Docker Desktop variants where the host's outbound IP is not
 	// routable from inside the container network. Authentication is enforced
 	// by the bearer middleware and per-repo scoping, not by reachability.
-	listener, err := net.Listen("tcp", fmt.Sprintf(":%d", opts.Port))
+	listener, err := net.Listen("tcp", fmt.Sprintf(":%d", port))
 	if err != nil {
-		return nil, err
+		return err
 	}
 	addr, ok := listener.Addr().(*net.TCPAddr)
 	if !ok {
 		listener.Close()
-		return nil, fmt.Errorf("cache server listens on %T, want a TCP address", listener.Addr())
+		return fmt.Errorf("cache server listens on %T, want a TCP address", listener.Addr())
 	}
 	h.port = addr.Port
 	server := &http.Server{
 		ReadHeaderTimeout: 2 * time.Second,
-		Handler:           router,
+		Handler:           handler,
 	}
 	go func() {
 		if err := server.Serve(listener); err != nil && errors.Is(err, net.ErrClosed) {
-			logger.Errorf("http serve: %v", err)
+			h.logger.Errorf("http serve: %v", err)
 		}
 	}()
 	h.listener = listener
 	h.server = server
 
-	return h, nil
+	return nil
+}
+
+func (h *Handler) cacheForwarder(upstream string) (http.Handler, error) {
+	target, err := url.Parse(strings.TrimRight(upstream, "/"))
+	if err != nil {
+		return nil, fmt.Errorf("parse cache upstream: %w", err)
+	}
+	transport, _ := http.DefaultTransport.(*http.Transport)
+	transport = transport.Clone()
+	transport.MaxIdleConnsPerHost = 64 // parallel block uploads would otherwise redial past the default 2
+	proxy := &httputil.ReverseProxy{
+		Transport: transport,
+		Rewrite: func(r *httputil.ProxyRequest) {
+			r.SetURL(target)
+			r.Out.URL.RawQuery = r.In.URL.RawQuery // Rewrite drops unparsable query params, and signed URLs must pass verbatim.
+		},
+		ErrorHandler: func(w http.ResponseWriter, r *http.Request, err error) {
+			h.logger.Warnf("cache forward %s %s: %v", r.Method, r.URL.Path, err)
+			h.twirpError(w, r, twirpUnavailable, errors.New("cache upstream is unavailable"))
+		},
+	}
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, apiPath+"/") || strings.HasPrefix(r.URL.Path, cacheServiceV2Path+"/") {
+			proxy.ServeHTTP(w, r)
+		} else {
+			h.forwardOrNotFound(w, r)
+		}
+	}), nil
 }
 
 func (h *Handler) ExternalURL() string {
@@ -708,8 +755,7 @@ func (h *Handler) internalRegister(w http.ResponseWriter, r *http.Request, _ htt
 		return
 	}
 	h.RegisterJob(body.Token, body.JobCredential)
-	// A server too old to forward answers without this, which is how the caller knows.
-	h.responseJSON(w, r, http.StatusOK, map[string]any{"results_url": h.ResultsURL(body.JobCredential)})
+	h.responseJSON(w, r, http.StatusOK)
 }
 
 // POST /_internal/revoke

act/artifactcache/handler_test.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"net/http/httptest"
 	"os"
 	"path/filepath"
 	"strings"
@@ -56,6 +57,55 @@ func TestHandler_ExternalURL(t *testing.T) {
 	assert.Equal(t, "http://[2001:db8::1]:8080", (&Handler{outboundIP: "2001:db8::1", port: 8080}).ExternalURL())
 }
 
+func TestHandler_Upstream(t *testing.T) {
+	var upstream *httptest.Server
+	upstream = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		assert.Equal(t, http.MethodPut, r.Method)
+		assert.Equal(t, "/cache%20server"+blobPath+"/%34%32", r.URL.EscapedPath())
+		assert.Equal(t, "sig=opaque%2Fvalue&raw=a;b", r.URL.RawQuery)
+		assert.Equal(t, strings.TrimPrefix(upstream.URL, "http://"), r.Host)
+		assert.Equal(t, "Bearer unregistered", r.Header.Get("Authorization"))
+		_, err := io.Copy(w, r.Body)
+		assert.NoError(t, err)
+	}))
+	t.Cleanup(upstream.Close)
+	dir := filepath.Join(t.TempDir(), "artifactcache")
+	handler, err := StartHandler(Options{
+		Dir:        dir,
+		OutboundIP: "127.0.0.1",
+		Upstream:   upstream.URL + "/cache%20server//",
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { require.NoError(t, handler.Close()) })
+
+	req, err := http.NewRequestWithContext(t.Context(), http.MethodPut,
+		handler.ExternalURL()+blobPath+"/%34%32?sig=opaque%2Fvalue&raw=a;b", strings.NewReader("cache content"))
+	require.NoError(t, err)
+	req.Header.Set("Authorization", "Bearer unregistered")
+	resp, err := http.DefaultClient.Do(req)
+	require.NoError(t, err)
+	forwarded, err := io.ReadAll(resp.Body)
+	resp.Body.Close()
+	require.NoError(t, err)
+	assert.Equal(t, "cache content", string(forwarded))
+
+	resp, err = http.Post(handler.ExternalURL()+internalPath+"/register", "application/json", nil)
+	require.NoError(t, err)
+	resp.Body.Close()
+	assert.Equal(t, http.StatusNotFound, resp.StatusCode)
+
+	upstream.Close()
+	resp, err = http.Post(handler.ExternalURL()+cacheServiceV2Path+"/CreateCacheEntry", "application/json", strings.NewReader(`{}`))
+	require.NoError(t, err)
+	var failure map[string]string
+	err = json.UnmarshalRead(resp.Body, &failure)
+	resp.Body.Close()
+	require.NoError(t, err)
+	assert.Equal(t, http.StatusServiceUnavailable, resp.StatusCode)
+	assert.Equal(t, map[string]string{"code": "unavailable", "msg": "cache upstream is unavailable"}, failure)
+	assert.NoDirExists(t, dir)
+}
+
 func TestHandler(t *testing.T) {
 	dir := filepath.Join(t.TempDir(), "artifactcache")
 	handler, err := StartHandler(Options{Dir: dir})
@@ -1522,7 +1572,10 @@ func TestHandler_InternalAPI_AuthAndUsage(t *testing.T) {
 		}
 
 		assert.Equal(t, http.StatusUnauthorized, probe("via-internal-api"))
-		assert.Equal(t, http.StatusOK, post("/_internal/register", secret, `{"token":"via-internal-api","repo":"owner/repo"}`))
+		assert.Equal(t, http.StatusOK, post("/_internal/register", secret, `{"token":"via-internal-api","repo":"owner/repo","public_url":"http://runner.example","results":"https://gitea.example","insecure_tls":true}`))
+		cred, ok := handler.lookupCredential("via-internal-api")
+		require.True(t, ok)
+		assert.Equal(t, JobCredential{Repo: testRepo, PublicURL: "http://runner.example"}, cred)
 		assert.NotEqual(t, http.StatusUnauthorized, probe("via-internal-api"))
 		assert.Equal(t, http.StatusOK, post("/_internal/revoke", secret, `{"token":"via-internal-api"}`))
 		assert.Equal(t, http.StatusUnauthorized, probe("via-internal-api"))

act/artifactcache/handler_v2.go

@@ -40,6 +40,7 @@ const (
 	blobUploadURLTTL = time.Hour
 
 	twirpInternal        = "internal"
+	twirpUnavailable     = "unavailable"
 	twirpUnauthenticated = "unauthenticated"
 )
 
@@ -261,6 +262,8 @@ func (h *Handler) twirpError(w http.ResponseWriter, r *http.Request, code string
 	switch code {
 	case twirpInternal:
 		status = http.StatusInternalServerError
+	case twirpUnavailable:
+		status = http.StatusServiceUnavailable
 	case twirpUnauthenticated:
 		status = http.StatusUnauthorized
 	}

act/artifactcache/results.go

@@ -17,9 +17,6 @@ import (
 // Gitea implements only the artifact half of it. Forwarding that half from here makes this origin
 // the whole service, so ACTIONS_RESULTS_URL can point at it truthfully, which is what the clients
 // this runner cannot patch need, docker buildx among them.
-//
-// The instance to forward to travels with the job registration rather than with configuration, so
-// a cache server shared between runners serves each of their instances.
 const artifactServicePath = "/twirp/github.actions.results.api.v1.ArtifactService/"
 
 // forwardOrNotFound is the router's fallback: the artifact service of the instance the job
@@ -81,9 +78,6 @@ func (h *Handler) mustProxy(r *http.Request, cred JobCredential, target *url.URL
 		return true
 	}
 	scheme := r.Header.Get("X-Forwarded-Proto")
-	if base, err := url.Parse(cred.PublicURL); err == nil && base.Scheme != "" {
-		scheme = base.Scheme
-	}
 	if scheme == "" {
 		scheme = "http"
 	}
@@ -98,4 +92,4 @@ func mergeQuery(target, request string) string {
 }
 
 // insecureTransport is shared, because a transport per request would pool no connections.
-var insecureTransport = &http.Transport{TLSClientConfig: &tls.Config{InsecureSkipVerify: true}} //nolint:gosec // the runner reaches its instance on the operator's say-so
+var insecureTransport = &http.Transport{Proxy: http.ProxyFromEnvironment, TLSClientConfig: &tls.Config{InsecureSkipVerify: true}} //nolint:gosec // the runner reaches its instance on the operator's say-so

act/artifactcache/results_test.go

@@ -54,11 +54,11 @@ func TestFrontResultsService(t *testing.T) {
 	assert.Contains(t, body["msg"], "bearer")
 	assert.NotEmpty(t, body["msg"])
 
-	defer handler.RegisterJob(token, JobCredential{Repo: "owner/repo", Results: gitea.URL, PublicURL: "https://cache.example"})()
+	defer handler.RegisterJob(token, JobCredential{Repo: "owner/repo", Results: selfSigned.URL, InsecureTLS: true})()
 
 	status, _ = post(rpc)
-	require.Equal(t, http.StatusOK, status, "a registered https public URL onto an http instance must proxy, not redirect")
-	assert.Equal(t, strings.TrimPrefix(gitea.URL, "http://"), gotHost, "Gitea must see the host it mints its URLs from")
+	require.Equal(t, http.StatusOK, status, "an http job reaching an https instance must proxy")
+	assert.Equal(t, strings.TrimPrefix(selfSigned.URL, "https://"), gotHost, "Gitea must see the host it mints its URLs from")
 	assert.Empty(t, gotProto, "a forwarded scheme would make an https Gitea mint http URLs")
 	assert.Equal(t, rpc, gotPath)
 
@@ -87,15 +87,15 @@ func TestFrontResultsService(t *testing.T) {
 		},
 		{
 			name: "a trusted https instance is reached by the job itself",
-			cred: JobCredential{Results: selfSigned.URL, PublicURL: "https://cache.example"}, want: http.StatusTemporaryRedirect,
+			cred: JobCredential{Results: selfSigned.URL}, forwardedProto: "https", want: http.StatusTemporaryRedirect,
 		},
 		{
 			name: "the job does not share this server's disregard for the certificate",
-			cred: JobCredential{Results: selfSigned.URL, PublicURL: "https://cache.example", InsecureTLS: true}, want: http.StatusOK,
+			cred: JobCredential{Results: selfSigned.URL, InsecureTLS: true}, forwardedProto: "https", want: http.StatusOK,
 		},
 		{name: "an http job is not redirected to an https instance", cred: JobCredential{Results: selfSigned.URL}, want: http.StatusInternalServerError},
 		{
-			name: "a runner too old to send its public URL leaves the terminator to say so",
+			name: "a forwarded https scheme onto an http instance is proxied",
 			cred: JobCredential{Results: gitea.URL}, forwardedProto: "https", want: http.StatusOK,
 		},
 		{