web: add feature flag PropagateCancels

This allow client-initiated cancels to propagate through gRPC.

features/features.go

@@ -103,6 +103,15 @@ type Config struct {
 	// This flag should only be used in conjunction with UseKvLimitsForNewOrder.
 	DisableLegacyLimitWrites bool
 
+	// PropagateCancels controls whether the WFE and ocsp-responder allows cancellation
+	// of an inbound request to cancel downstream gRPC and other queries. In practice,
+	// cancellation of an inbound request is achieved by Nginx closing the connection
+	// on which the request was happening. This may help shed load in overcapacity
+	// situations. However, not that in-progress database queries (for instance, in the
+	// SA) are not cancelled. Database queries waiting for an available connection may
+	// be cancelled.
+	PropagateCancels bool
+
 	// InsertAuthzsIndividually causes the SA's NewOrderAndAuthzs method to
 	// create each new authz one at a time, rather than using MultiInserter.
 	// Although this is expected to be a performance penalty, it is necessary to

test/config-next/wfe2.json

@@ -127,6 +127,7 @@
 			"Overrides": "test/config-next/wfe2-ratelimit-overrides.yml"
 		},
 		"features": {
+      "PropagateCancels": true,
 			"ServeRenewalInfo": true,
 			"CheckIdentifiersPaused": true,
 			"UseKvLimitsForNewOrder": true

web/context.go

@@ -12,6 +12,7 @@ import (
 	"strings"
 	"time"
 
+	"github.com/letsencrypt/boulder/features"
 	blog "github.com/letsencrypt/boulder/log"
 )
 
@@ -127,11 +128,13 @@ func (th *TopHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		Origin:    r.Header.Get("Origin"),
 		Extra:     make(map[string]interface{}),
 	}
-	// We specifically override the default r.Context() because we would prefer
-	// for clients to not be able to cancel our operations in arbitrary places.
-	// Instead we start a new context, and apply timeouts in our various RPCs.
-	ctx := context.WithoutCancel(r.Context())
-	r = r.WithContext(ctx)
+	if !features.Get().PropagateCancels {
+		// We specifically override the default r.Context() because we would prefer
+		// for clients to not be able to cancel our operations in arbitrary places.
+		// Instead we start a new context, and apply timeouts in our various RPCs.
+		ctx := context.WithoutCancel(r.Context())
+		r = r.WithContext(ctx)
+	}
 
 	// Some clients will send a HTTP Host header that includes the default port
 	// for the scheme that they are using. Previously when we were fronted by

web/context_test.go

@@ -2,13 +2,16 @@ package web
 
 import (
 	"bytes"
+	"context"
 	"crypto/tls"
 	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"strings"
 	"testing"
+	"time"
 
+	"github.com/letsencrypt/boulder/features"
 	blog "github.com/letsencrypt/boulder/log"
 	"github.com/letsencrypt/boulder/test"
 )
@@ -117,3 +120,36 @@ func TestHostHeaderRewrite(t *testing.T) {
 	req.Host = "localhost:123"
 	th.ServeHTTP(httptest.NewRecorder(), req)
 }
+
+type cancelHandler struct {
+	res chan string
+}
+
+func (ch cancelHandler) ServeHTTP(e *RequestEvent, w http.ResponseWriter, r *http.Request) {
+	select {
+	case <-r.Context().Done():
+		ch.res <- r.Context().Err().Error()
+	case <-time.After(300 * time.Millisecond):
+		ch.res <- "300 ms passed"
+	}
+}
+
+func TestPropagateCancel(t *testing.T) {
+	mockLog := blog.UseMock()
+	res := make(chan string)
+	features.Set(features.Config{PropagateCancels: true})
+	th := NewTopHandler(mockLog, cancelHandler{res})
+	ctx, cancel := context.WithCancel(context.Background())
+	go func() {
+		req, err := http.NewRequestWithContext(ctx, "GET", "/thisisignored", &bytes.Reader{})
+		if err != nil {
+			t.Error(err)
+		}
+		th.ServeHTTP(httptest.NewRecorder(), req)
+	}()
+	cancel()
+	result := <-res
+	if result != "context canceled" {
+		t.Errorf("expected 'context canceled', got %q", result)
+	}
+}