http: adding response flood protection (#113)

docs/root/configuration/http/http_conn_man/stats.rst

@@ -111,6 +111,7 @@ All http1 statistics are rooted at *http1.*
    :widths: 1, 1, 2
 
    metadata_not_supported_error, Counter, Total number of metadata dropped during HTTP/1 encoding
+   response_flood, Counter, Total number of connections closed due to response flooding
 
 Http2 codec statistics
 ~~~~~~~~~~~~~~~~~~~~~~

docs/root/intro/version_history.rst

@@ -8,6 +8,7 @@ Version history
 * http: added HTTP/1.1 flood protection. Can be temporarily disabled using the runtime feature `envoy.reloadable_features.http1_flood_protection`.
 * listeners: fixed issue where :ref:`TLS inspector listener filter <config_listener_filters_tls_inspector>` could have been bypassed by a client using only TLS 1.3.
 * sds: fix the SDS vulnerability that TLS validation context (e.g., subject alt name or hash) cannot be effectively validated in some cases.
+* http: added HTTP/1.1 flood protection. Can be temporarily disabled using the runtime feature `envoy.reloadable_features.http1_flood_protection`.
 
 1.13.0 (January 20, 2020)
 =========================

source/common/http/http1/codec_impl.cc

@@ -243,17 +243,49 @@ void StreamEncoderImpl::endEncode() {
     connection_.buffer().add(CRLF);
   }
 
-  connection_.flushOutput();
+  connection_.flushOutput(true);
   connection_.onEncodeComplete();
 }
 
-void ConnectionImpl::flushOutput() {
+void ServerConnectionImpl::maybeAddSentinelBufferFragment(Buffer::WatermarkBuffer& output_buffer) {
+  if (!flood_protection_) {
+    return;
+  }
+  // It's messy and complicated to try to tag the final write of an HTTP response for response
+  // tracking for flood protection. Instead, write an empty buffer fragment after the response,
+  // to allow for tracking.
+  // When the response is written out, the fragment will be deleted and the counter will be updated
+  // by ServerConnectionImpl::releaseOutboundResponse()
+  auto fragment =
+      Buffer::OwnedBufferFragmentImpl::create(absl::string_view("", 0), response_buffer_releasor_);
+  output_buffer.addBufferFragment(*fragment.release());
+  ASSERT(outbound_responses_ < max_outbound_responses_);
+  outbound_responses_++;
+}
+
+void ServerConnectionImpl::doFloodProtectionChecks() const {
+  if (!flood_protection_) {
+    return;
+  }
+  // Before sending another response, make sure it won't exceed flood protection thresholds.
+  if (outbound_responses_ >= max_outbound_responses_) {
+    ENVOY_CONN_LOG(trace, "error sending response: Too many pending responses queued", connection_);
+    stats_.response_flood_.inc();
+    throw FrameFloodException("Too many responses queued.");
+  }
+}
+
+void ConnectionImpl::flushOutput(bool end_encode) {
   if (reserved_current_) {
     reserved_iovec_.len_ = reserved_current_ - static_cast<char*>(reserved_iovec_.mem_);
     output_buffer_.commit(&reserved_iovec_, 1);
     reserved_current_ = nullptr;
   }
-
+  if (end_encode) {
+    // If this is an HTTP response in ServerConnectionImpl, track outbound responses for flood
+    // protection
+    maybeAddSentinelBufferFragment(output_buffer_);
+  }
   connection().write(output_buffer_, false);
   ASSERT(0UL == output_buffer_.length());
 }
@@ -309,6 +341,9 @@ static const char RESPONSE_PREFIX[] = "HTTP/1.1 ";
 static const char HTTP_10_RESPONSE_PREFIX[] = "HTTP/1.0 ";
 
 void ResponseStreamEncoderImpl::encodeHeaders(const HeaderMap& headers, bool end_stream) {
+  // Do flood checks before attempting to write any responses.
+  flood_checks_();
+
   started_response_ = true;
   uint64_t numeric_status = Utility::getResponseStatus(headers);
 
@@ -678,7 +713,18 @@ ServerConnectionImpl::ServerConnectionImpl(Network::Connection& connection, Stat
                                            const uint32_t max_request_headers_count)
     : ConnectionImpl(connection, stats, HTTP_REQUEST, max_request_headers_kb,
                      max_request_headers_count, formatter(settings), settings.enable_trailers_),
-      callbacks_(callbacks), codec_settings_(settings) {}
+      callbacks_(callbacks), codec_settings_(settings),
+      response_buffer_releasor_([this](const Buffer::OwnedBufferFragmentImpl* fragment) {
+        releaseOutboundResponse(fragment);
+      }),
+      // Pipelining is generally not well supported on the internet and has a series of dangerous
+      // overflow bugs. As such we are disabling it for now, and removing this temporary override if
+      // no one objects. If you use this integer to restore prior behavior, contact the
+      // maintainer team as it will otherwise be removed entirely soon.
+      max_outbound_responses_(
+          Runtime::getInteger("envoy.do_not_use_going_away_max_http2_outbound_responses", 2)),
+      flood_protection_(
+          Runtime::runtimeFeatureEnabled("envoy.reloadable_features.http1_flood_protection")) {}
 
 void ServerConnectionImpl::onEncodeComplete() {
   ASSERT(active_request_);
@@ -788,7 +834,8 @@ int ServerConnectionImpl::onHeadersComplete(HeaderMapImplPtr&& headers) {
 void ServerConnectionImpl::onMessageBegin() {
   if (!resetStreamCalled()) {
     ASSERT(!active_request_);
-    active_request_ = std::make_unique<ActiveRequest>(*this, header_key_formatter_.get());
+    active_request_ =
+        std::make_unique<ActiveRequest>(*this, header_key_formatter_.get(), flood_checks_);
     active_request_->request_decoder_ = &callbacks_.newStream(active_request_->response_encoder_);
   }
 }
@@ -863,6 +910,13 @@ void ServerConnectionImpl::onBelowLowWatermark() {
   }
 }
 
+void ServerConnectionImpl::releaseOutboundResponse(
+    const Buffer::OwnedBufferFragmentImpl* fragment) {
+  ASSERT(outbound_responses_ >= 1);
+  --outbound_responses_;
+  delete fragment;
+}
+
 ClientConnectionImpl::ClientConnectionImpl(Network::Connection& connection, Stats::Scope& stats,
                                            ConnectionCallbacks&, const Http1Settings& settings,
                                            const uint32_t max_response_headers_count)

source/common/http/http1/codec_impl.h

@@ -27,7 +27,9 @@ namespace Http1 {
 /**
  * All stats for the HTTP/1 codec. @see stats_macros.h
  */
-#define ALL_HTTP1_CODEC_STATS(COUNTER) COUNTER(metadata_not_supported_error)
+#define ALL_HTTP1_CODEC_STATS(COUNTER)                                                             \
+  COUNTER(metadata_not_supported_error)                                                            \
+  COUNTER(response_flood)
 
 /**
  * Wrapper struct for the HTTP/1 codec stats. @see stats_macros.h
@@ -112,15 +114,19 @@ class StreamEncoderImpl : public StreamEncoder,
  */
 class ResponseStreamEncoderImpl : public StreamEncoderImpl {
 public:
-  ResponseStreamEncoderImpl(ConnectionImpl& connection, HeaderKeyFormatter* header_key_formatter)
-      : StreamEncoderImpl(connection, header_key_formatter) {}
+  using FloodChecks = std::function<void()>;
+
+  ResponseStreamEncoderImpl(ConnectionImpl& connection, HeaderKeyFormatter* header_key_formatter,
+                            FloodChecks& flood_checks)
+      : StreamEncoderImpl(connection, header_key_formatter), flood_checks_(flood_checks) {}
 
   bool startedResponse() { return started_response_; }
 
   // Http::StreamEncoder
   void encodeHeaders(const HeaderMap& headers, bool end_stream) override;
 
 private:
+  FloodChecks& flood_checks_;
   bool started_response_{};
 };
 
@@ -172,7 +178,7 @@ class ConnectionImpl : public virtual Connection, protected Logger::Loggable<Log
   /**
    * Flush all pending output from encoding.
    */
-  void flushOutput();
+  void flushOutput(bool end_encode = false);
 
   void addCharToBuffer(char c);
   void addIntToBuffer(uint64_t i);
@@ -195,6 +201,7 @@ class ConnectionImpl : public virtual Connection, protected Logger::Loggable<Log
   virtual bool supports_http_10() { return false; }
 
   bool maybeDirectDispatch(Buffer::Instance& data);
+  virtual void maybeAddSentinelBufferFragment(Buffer::WatermarkBuffer&) {}
 
   CodecStats& stats() { return stats_; }
 
@@ -327,6 +334,7 @@ class ConnectionImpl : public virtual Connection, protected Logger::Loggable<Log
  */
 class ServerConnectionImpl : public ServerConnection, public ConnectionImpl {
 public:
+  using FloodChecks = std::function<void()>;
   ServerConnectionImpl(Network::Connection& connection, Stats::Scope& stats,
                        ServerConnectionCallbacks& callbacks, const Http1Settings& settings,
                        uint32_t max_request_headers_kb, const uint32_t max_request_headers_count);
@@ -338,8 +346,9 @@ class ServerConnectionImpl : public ServerConnection, public ConnectionImpl {
    * An active HTTP/1.1 request.
    */
   struct ActiveRequest {
-    ActiveRequest(ConnectionImpl& connection, HeaderKeyFormatter* header_key_formatter)
-        : response_encoder_(connection, header_key_formatter) {}
+    ActiveRequest(ConnectionImpl& connection, HeaderKeyFormatter* header_key_formatter,
+                  FloodChecks& flood_checks)
+        : response_encoder_(connection, header_key_formatter, flood_checks) {}
 
     HeaderString request_url_;
     StreamDecoder* request_decoder_{};
@@ -370,9 +379,21 @@ class ServerConnectionImpl : public ServerConnection, public ConnectionImpl {
   void onAboveHighWatermark() override;
   void onBelowLowWatermark() override;
 
+  void releaseOutboundResponse(const Buffer::OwnedBufferFragmentImpl* fragment);
+  void maybeAddSentinelBufferFragment(Buffer::WatermarkBuffer& output_buffer) override;
+  void doFloodProtectionChecks() const;
+
   ServerConnectionCallbacks& callbacks_;
+  std::function<void()> flood_checks_{[&]() { this->doFloodProtectionChecks(); }};
   std::unique_ptr<ActiveRequest> active_request_;
   Http1Settings codec_settings_;
+  const Buffer::OwnedBufferFragmentImpl::Releasor response_buffer_releasor_;
+  uint32_t outbound_responses_{};
+  // This defaults to 2, which functionally disables pipelining. If any users
+  // of Envoy wish to enable pipelining (which is dangerous and ill supported)
+  // we could make this configurable.
+  uint32_t max_outbound_responses_{};
+  bool flood_protection_{};
 };
 
 /**

source/common/runtime/runtime_features.cc

@@ -24,6 +24,7 @@ namespace Runtime {
 // problem of the bugs being found after the old code path has been removed.
 constexpr const char* runtime_features[] = {
     // Enabled
+    "envoy.reloadable_features.http1_flood_protection",
     "envoy.reloadable_features.test_feature_true",
     "envoy.reloadable_features.strict_header_validation",
     "envoy.reloadable_features.buffer_filter_populate_content_length",

source/common/runtime/runtime_impl.cc

@@ -65,7 +65,7 @@ bool runtimeFeatureEnabled(absl::string_view feature) {
 }
 
 uint64_t getInteger(absl::string_view feature, uint64_t default_value) {
-  ASSERT(absl::StartsWith(feature, "envoy.reloadable_features"));
+  ASSERT(absl::StartsWith(feature, "envoy."));
   if (Runtime::LoaderSingleton::getExisting()) {
     return Runtime::LoaderSingleton::getExisting()->threadsafeSnapshot()->getInteger(
         std::string(feature), default_value);

test/common/http/http1/codec_impl_test.cc

@@ -607,6 +607,92 @@ TEST_F(Http1ServerConnectionImplTest, BadRequestStartedStream) {
   EXPECT_EQ("HTTP/1.1 400 Bad Request\r\ncontent-length: 0\r\nconnection: close\r\n\r\n", output);
 }
 
+TEST_F(Http1ServerConnectionImplTest, FloodProtection) {
+  initialize();
+
+  NiceMock<Http::MockStreamDecoder> decoder;
+  Buffer::OwnedImpl local_buffer;
+  // Read a request and send a response, without draining the response from the
+  // connection buffer. The first two should not cause problems.
+  for (int i = 0; i < 2; ++i) {
+    Http::StreamEncoder* response_encoder = nullptr;
+    EXPECT_CALL(callbacks_, newStream(_, _))
+        .WillOnce(Invoke([&](Http::StreamEncoder& encoder, bool) -> Http::StreamDecoder& {
+          response_encoder = &encoder;
+          return decoder;
+        }));
+
+    Buffer::OwnedImpl buffer("GET / HTTP/1.1\r\n\r\n");
+    codec_->dispatch(buffer);
+    EXPECT_EQ(0U, buffer.length());
+
+    // In most tests the write output is serialized to a buffer here it is
+    // ignored to build up queued "end connection" sentinels.
+    EXPECT_CALL(connection_, write(_, _))
+        .Times(1)
+        .WillOnce(Invoke([&](Buffer::Instance& data, bool) -> void {
+          // Move the response out of data while preserving the buffer fragment sentinels.
+          local_buffer.move(data);
+        }));
+
+    TestHeaderMapImpl headers{{":status", "200"}};
+    response_encoder->encodeHeaders(headers, true);
+  }
+
+  // Trying to shove a third response in the queue should trigger flood protection.
+  {
+    Http::StreamEncoder* response_encoder = nullptr;
+    EXPECT_CALL(callbacks_, newStream(_, _))
+        .WillOnce(Invoke([&](Http::StreamEncoder& encoder, bool) -> Http::StreamDecoder& {
+          response_encoder = &encoder;
+          return decoder;
+        }));
+
+    Buffer::OwnedImpl buffer("GET / HTTP/1.1\r\n\r\n");
+    codec_->dispatch(buffer);
+
+    TestHeaderMapImpl headers{{":status", "200"}};
+    EXPECT_THROW_WITH_MESSAGE(response_encoder->encodeHeaders(headers, true), FrameFloodException,
+                              "Too many responses queued.");
+    EXPECT_EQ(1, store_.counter("http1.response_flood").value());
+  }
+}
+
+TEST_F(Http1ServerConnectionImplTest, FloodProtectionOff) {
+  TestScopedRuntime scoped_runtime;
+  Runtime::LoaderSingleton::getExisting()->mergeValues(
+      {{"envoy.reloadable_features.http1_flood_protection", "false"}});
+  initialize();
+
+  NiceMock<Http::MockStreamDecoder> decoder;
+  Buffer::OwnedImpl local_buffer;
+  // With flood protection off, many responses can be queued up.
+  for (int i = 0; i < 4; ++i) {
+    Http::StreamEncoder* response_encoder = nullptr;
+    EXPECT_CALL(callbacks_, newStream(_, _))
+        .WillOnce(Invoke([&](Http::StreamEncoder& encoder, bool) -> Http::StreamDecoder& {
+          response_encoder = &encoder;
+          return decoder;
+        }));
+
+    Buffer::OwnedImpl buffer("GET / HTTP/1.1\r\n\r\n");
+    codec_->dispatch(buffer);
+    EXPECT_EQ(0U, buffer.length());
+
+    // In most tests the write output is serialized to a buffer here it is
+    // ignored to build up queued "end connection" sentinels.
+    EXPECT_CALL(connection_, write(_, _))
+        .Times(1)
+        .WillOnce(Invoke([&](Buffer::Instance& data, bool) -> void {
+          // Move the response out of data while preserving the buffer fragment sentinels.
+          local_buffer.move(data);
+        }));
+
+    TestHeaderMapImpl headers{{":status", "200"}};
+    response_encoder->encodeHeaders(headers, true);
+  }
+}
+
 TEST_F(Http1ServerConnectionImplTest, HostHeaderTranslation) {
   initialize();
 

test/integration/integration_admin_test.cc

@@ -267,8 +267,9 @@ TEST_P(IntegrationAdminTest, Admin) {
   case Http::CodecClient::Type::HTTP1:
     EXPECT_EQ("   Count Lookup\n"
               "       1 http1.metadata_not_supported_error\n"
+              "       1 http1.response_flood\n"
               "\n"
-              "total: 1\n",
+              "total: 2\n",
               response->body());
     break;
   case Http::CodecClient::Type::HTTP2:

test/integration/integration_test.cc

@@ -1152,4 +1152,53 @@ TEST_P(UpstreamEndpointIntegrationTest, TestUpstreamEndpointAddress) {
                Network::Test::getLoopbackAddressString(GetParam()).c_str());
 }
 
+// Send continuous pipelined requests while not reading responses, to check
+// HTTP/1.1 response flood protection.
+TEST_P(IntegrationTest, TestFlood) {
+  initialize();
+
+  // Set up a raw connection to easily send requests without reading responses.
+  Network::ClientConnectionPtr raw_connection = makeClientConnection(lookupPort("http"));
+  raw_connection->connect();
+
+  // Read disable so responses will queue up.
+  uint32_t bytes_to_send = 0;
+  raw_connection->readDisable(true);
+  // Track locally queued bytes, to make sure the outbound client queue doesn't back up.
+  raw_connection->addBytesSentCallback([&](uint64_t bytes) { bytes_to_send -= bytes; });
+
+  // Keep sending requests until flood protection kicks in and kills the connection.
+  while (raw_connection->state() == Network::Connection::State::Open) {
+    // These requests are missing the host header, so will provoke an internally generated error
+    // response from Envoy.
+    Buffer::OwnedImpl buffer("GET / HTTP/1.1\r\n\r\nGET / HTTP/1.1\r\n\r\nGET / HTTP/1.1\r\n\r\n");
+    bytes_to_send += buffer.length();
+    raw_connection->write(buffer, false);
+    // Loop until all bytes are sent.
+    while (bytes_to_send > 0 && raw_connection->state() == Network::Connection::State::Open) {
+      raw_connection->dispatcher().run(Event::Dispatcher::RunType::NonBlock);
+    }
+  }
+
+  // Verify the connection was closed due to flood protection.
+  EXPECT_EQ(1, test_server_->counter("http1.response_flood")->value());
+}
+
+// Make sure flood protection doesn't kick in with many requests sent serially.
+TEST_P(IntegrationTest, TestManyBadRequests) {
+  initialize();
+
+  codec_client_ = makeHttpConnection(lookupPort("http"));
+  Http::TestHeaderMapImpl bad_request{
+      {":method", "GET"}, {":path", "/test/long/url"}, {":scheme", "http"}};
+
+  for (int i = 0; i < 1; ++i) {
+    IntegrationStreamDecoderPtr response = codec_client_->makeHeaderOnlyRequest(bad_request);
+    response->waitForEndStream();
+    ASSERT_TRUE(response->complete());
+    EXPECT_THAT(response->headers(), HttpStatusIs("400"));
+  }
+  EXPECT_EQ(0, test_server_->counter("http1.response_flood")->value());
+}
+
 } // namespace Envoy