Client-side health checking support.

4 files changed, 915 insertions, 0 deletions

diff --git a/src/core/ext/filters/client_channel/health/health.pb.c b/src/core/ext/filters/client_channel/health/health.pb.c
new file mode 100644
index 0000000000..5499c549cc
--- /dev/null
+++ b/src/core/ext/filters/client_channel/health/health.pb.c
@@ -0,0 +1,23 @@
+/* Automatically generated nanopb constant definitions */
+/* Generated by nanopb-0.3.7-dev */
+
+#include "src/core/ext/filters/client_channel/health/health.pb.h"
+/* @@protoc_insertion_point(includes) */
+#if PB_PROTO_HEADER_VERSION != 30
+#error Regenerate this file with the current version of nanopb generator.
+#endif
+
+
+
+const pb_field_t grpc_health_v1_HealthCheckRequest_fields[2] = {
+    PB_FIELD(  1, STRING  , OPTIONAL, STATIC  , FIRST, grpc_health_v1_HealthCheckRequest, service, service, 0),
+    PB_LAST_FIELD
+};
+
+const pb_field_t grpc_health_v1_HealthCheckResponse_fields[2] = {
+    PB_FIELD(  1, UENUM   , OPTIONAL, STATIC  , FIRST, grpc_health_v1_HealthCheckResponse, status, status, 0),
+    PB_LAST_FIELD
+};
+
+
+/* @@protoc_insertion_point(eof) */
diff --git a/src/core/ext/filters/client_channel/health/health.pb.h b/src/core/ext/filters/client_channel/health/health.pb.h
new file mode 100644
index 0000000000..9d54ccd618
--- /dev/null
+++ b/src/core/ext/filters/client_channel/health/health.pb.h
@@ -0,0 +1,73 @@
+/* Automatically generated nanopb header */
+/* Generated by nanopb-0.3.7-dev */
+
+#ifndef PB_GRPC_HEALTH_V1_HEALTH_PB_H_INCLUDED
+#define PB_GRPC_HEALTH_V1_HEALTH_PB_H_INCLUDED
+#include "pb.h"
+/* @@protoc_insertion_point(includes) */
+#if PB_PROTO_HEADER_VERSION != 30
+#error Regenerate this file with the current version of nanopb generator.
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Enum definitions */
+typedef enum _grpc_health_v1_HealthCheckResponse_ServingStatus {
+    grpc_health_v1_HealthCheckResponse_ServingStatus_UNKNOWN = 0,
+    grpc_health_v1_HealthCheckResponse_ServingStatus_SERVING = 1,
+    grpc_health_v1_HealthCheckResponse_ServingStatus_NOT_SERVING = 2,
+    grpc_health_v1_HealthCheckResponse_ServingStatus_SERVICE_UNKNOWN = 3
+} grpc_health_v1_HealthCheckResponse_ServingStatus;
+#define _grpc_health_v1_HealthCheckResponse_ServingStatus_MIN grpc_health_v1_HealthCheckResponse_ServingStatus_UNKNOWN
+#define _grpc_health_v1_HealthCheckResponse_ServingStatus_MAX grpc_health_v1_HealthCheckResponse_ServingStatus_SERVICE_UNKNOWN
+#define _grpc_health_v1_HealthCheckResponse_ServingStatus_ARRAYSIZE ((grpc_health_v1_HealthCheckResponse_ServingStatus)(grpc_health_v1_HealthCheckResponse_ServingStatus_SERVICE_UNKNOWN+1))
+
+/* Struct definitions */
+typedef struct _grpc_health_v1_HealthCheckRequest {
+    bool has_service;
+    char service[200];
+/* @@protoc_insertion_point(struct:grpc_health_v1_HealthCheckRequest) */
+} grpc_health_v1_HealthCheckRequest;
+
+typedef struct _grpc_health_v1_HealthCheckResponse {
+    bool has_status;
+    grpc_health_v1_HealthCheckResponse_ServingStatus status;
+/* @@protoc_insertion_point(struct:grpc_health_v1_HealthCheckResponse) */
+} grpc_health_v1_HealthCheckResponse;
+
+/* Default values for struct fields */
+
+/* Initializer values for message structs */
+#define grpc_health_v1_HealthCheckRequest_init_default {false, ""}
+#define grpc_health_v1_HealthCheckResponse_init_default {false, (grpc_health_v1_HealthCheckResponse_ServingStatus)0}
+#define grpc_health_v1_HealthCheckRequest_init_zero {false, ""}
+#define grpc_health_v1_HealthCheckResponse_init_zero {false, (grpc_health_v1_HealthCheckResponse_ServingStatus)0}
+
+/* Field tags (for use in manual encoding/decoding) */
+#define grpc_health_v1_HealthCheckRequest_service_tag 1
+#define grpc_health_v1_HealthCheckResponse_status_tag 1
+
+/* Struct field encoding specification for nanopb */
+extern const pb_field_t grpc_health_v1_HealthCheckRequest_fields[2];
+extern const pb_field_t grpc_health_v1_HealthCheckResponse_fields[2];
+
+/* Maximum encoded size of messages (where known) */
+#define grpc_health_v1_HealthCheckRequest_size   203
+#define grpc_health_v1_HealthCheckResponse_size  2
+
+/* Message IDs (where set with "msgid" option) */
+#ifdef PB_MSGID
+
+#define HEALTH_MESSAGES \
+
+
+#endif
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+/* @@protoc_insertion_point(eof) */
+
+#endif
diff --git a/src/core/ext/filters/client_channel/health/health_check_client.cc b/src/core/ext/filters/client_channel/health/health_check_client.cc
new file mode 100644
index 0000000000..400d99b07c
--- /dev/null
+++ b/src/core/ext/filters/client_channel/health/health_check_client.cc
@@ -0,0 +1,646 @@
+/*
+ *
+ * Copyright 2018 gRPC authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ */
+
+#include <grpc/support/port_platform.h>
+
+#include <stdint.h>
+
+#include "src/core/ext/filters/client_channel/health/health_check_client.h"
+
+#include "pb_decode.h"
+#include "pb_encode.h"
+#include "src/core/ext/filters/client_channel/health/health.pb.h"
+#include "src/core/lib/debug/trace.h"
+#include "src/core/lib/gprpp/mutex_lock.h"
+#include "src/core/lib/slice/slice_internal.h"
+#include "src/core/lib/transport/error_utils.h"
+#include "src/core/lib/transport/status_metadata.h"
+
+#define HEALTH_CHECK_INITIAL_CONNECT_BACKOFF_SECONDS 1
+#define HEALTH_CHECK_RECONNECT_BACKOFF_MULTIPLIER 1.6
+#define HEALTH_CHECK_RECONNECT_MAX_BACKOFF_SECONDS 120
+#define HEALTH_CHECK_RECONNECT_JITTER 0.2
+
+grpc_core::TraceFlag grpc_health_check_client_trace(false,
+                                                    "health_check_client");
+
+namespace grpc_core {
+
+//
+// HealthCheckClient
+//
+
+HealthCheckClient::HealthCheckClient(
+    const char* service_name,
+    RefCountedPtr<ConnectedSubchannel> connected_subchannel,
+    grpc_pollset_set* interested_parties,
+    grpc_core::RefCountedPtr<grpc_core::channelz::SubchannelNode> channelz_node)
+    : InternallyRefCountedWithTracing<HealthCheckClient>(
+          &grpc_health_check_client_trace),
+      service_name_(service_name),
+      connected_subchannel_(std::move(connected_subchannel)),
+      interested_parties_(interested_parties),
+      channelz_node_(std::move(channelz_node)),
+      retry_backoff_(
+          BackOff::Options()
+              .set_initial_backoff(
+                  HEALTH_CHECK_INITIAL_CONNECT_BACKOFF_SECONDS * 1000)
+              .set_multiplier(HEALTH_CHECK_RECONNECT_BACKOFF_MULTIPLIER)
+              .set_jitter(HEALTH_CHECK_RECONNECT_JITTER)
+              .set_max_backoff(HEALTH_CHECK_RECONNECT_MAX_BACKOFF_SECONDS *
+                               1000)) {
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "created HealthCheckClient %p", this);
+  }
+  GRPC_CLOSURE_INIT(&retry_timer_callback_, OnRetryTimer, this,
+                    grpc_schedule_on_exec_ctx);
+  gpr_mu_init(&mu_);
+  StartCall();
+}
+
+HealthCheckClient::~HealthCheckClient() {
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "destroying HealthCheckClient %p", this);
+  }
+  GRPC_ERROR_UNREF(error_);
+  gpr_mu_destroy(&mu_);
+}
+
+void HealthCheckClient::NotifyOnHealthChange(grpc_connectivity_state* state,
+                                             grpc_closure* closure) {
+  MutexLock lock(&mu_);
+  GPR_ASSERT(notify_state_ == nullptr);
+  if (*state != state_) {
+    *state = state_;
+    GRPC_CLOSURE_SCHED(closure, GRPC_ERROR_REF(error_));
+    return;
+  }
+  notify_state_ = state;
+  on_health_changed_ = closure;
+}
+
+void HealthCheckClient::SetHealthStatus(grpc_connectivity_state state,
+                                        grpc_error* error) {
+  MutexLock lock(&mu_);
+  SetHealthStatusLocked(state, error);
+}
+
+void HealthCheckClient::SetHealthStatusLocked(grpc_connectivity_state state,
+                                              grpc_error* error) {
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: setting state=%d error=%s", this,
+            state, grpc_error_string(error));
+  }
+  if (notify_state_ != nullptr && *notify_state_ != state) {
+    *notify_state_ = state;
+    notify_state_ = nullptr;
+    GRPC_CLOSURE_SCHED(on_health_changed_, GRPC_ERROR_REF(error));
+    on_health_changed_ = nullptr;
+  }
+  state_ = state;
+  GRPC_ERROR_UNREF(error_);
+  error_ = error;
+}
+
+void HealthCheckClient::Orphan() {
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: shutting down", this);
+  }
+  {
+    MutexLock lock(&mu_);
+    if (on_health_changed_ != nullptr) {
+      *notify_state_ = GRPC_CHANNEL_SHUTDOWN;
+      notify_state_ = nullptr;
+      GRPC_CLOSURE_SCHED(on_health_changed_, GRPC_ERROR_NONE);
+      on_health_changed_ = nullptr;
+    }
+    shutting_down_ = true;
+    call_state_.reset();
+    if (retry_timer_callback_pending_) {
+      grpc_timer_cancel(&retry_timer_);
+    }
+  }
+  Unref(DEBUG_LOCATION, "orphan");
+}
+
+void HealthCheckClient::StartCall() {
+  MutexLock lock(&mu_);
+  StartCallLocked();
+}
+
+void HealthCheckClient::StartCallLocked() {
+  if (shutting_down_) return;
+  GPR_ASSERT(call_state_ == nullptr);
+  SetHealthStatusLocked(GRPC_CHANNEL_CONNECTING, GRPC_ERROR_NONE);
+  call_state_ = MakeOrphanable<CallState>(Ref(), interested_parties_);
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: created CallState %p", this,
+            call_state_.get());
+  }
+  call_state_->StartCall();
+}
+
+void HealthCheckClient::StartRetryTimer() {
+  MutexLock lock(&mu_);
+  SetHealthStatusLocked(
+      GRPC_CHANNEL_TRANSIENT_FAILURE,
+      GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+          "health check call failed; will retry after backoff"));
+  grpc_millis next_try = retry_backoff_.NextAttemptTime();
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: health check call lost...", this);
+    grpc_millis timeout = next_try - ExecCtx::Get()->Now();
+    if (timeout > 0) {
+      gpr_log(GPR_INFO,
+              "HealthCheckClient %p: ... will retry in %" PRId64 "ms.", this,
+              timeout);
+    } else {
+      gpr_log(GPR_INFO, "HealthCheckClient %p: ... retrying immediately.",
+              this);
+    }
+  }
+  // Ref for callback, tracked manually.
+  Ref(DEBUG_LOCATION, "health_retry_timer").release();
+  retry_timer_callback_pending_ = true;
+  grpc_timer_init(&retry_timer_, next_try, &retry_timer_callback_);
+}
+
+void HealthCheckClient::OnRetryTimer(void* arg, grpc_error* error) {
+  HealthCheckClient* self = static_cast<HealthCheckClient*>(arg);
+  {
+    MutexLock lock(&self->mu_);
+    self->retry_timer_callback_pending_ = false;
+    if (!self->shutting_down_ && error == GRPC_ERROR_NONE &&
+        self->call_state_ == nullptr) {
+      if (grpc_health_check_client_trace.enabled()) {
+        gpr_log(GPR_INFO, "HealthCheckClient %p: restarting health check call",
+                self);
+      }
+      self->StartCallLocked();
+    }
+  }
+  self->Unref(DEBUG_LOCATION, "health_retry_timer");
+}
+
+//
+// protobuf helpers
+//
+
+namespace {
+
+void EncodeRequest(const char* service_name,
+                   ManualConstructor<SliceBufferByteStream>* send_message) {
+  grpc_health_v1_HealthCheckRequest request_struct;
+  request_struct.has_service = true;
+  snprintf(request_struct.service, sizeof(request_struct.service), "%s",
+           service_name);
+  pb_ostream_t ostream;
+  memset(&ostream, 0, sizeof(ostream));
+  pb_encode(&ostream, grpc_health_v1_HealthCheckRequest_fields,
+            &request_struct);
+  grpc_slice request_slice = GRPC_SLICE_MALLOC(ostream.bytes_written);
+  ostream = pb_ostream_from_buffer(GRPC_SLICE_START_PTR(request_slice),
+                                   GRPC_SLICE_LENGTH(request_slice));
+  GPR_ASSERT(pb_encode(&ostream, grpc_health_v1_HealthCheckRequest_fields,
+                       &request_struct) != 0);
+  grpc_slice_buffer slice_buffer;
+  grpc_slice_buffer_init(&slice_buffer);
+  grpc_slice_buffer_add(&slice_buffer, request_slice);
+  send_message->Init(&slice_buffer, 0);
+  grpc_slice_buffer_destroy_internal(&slice_buffer);
+}
+
+// Returns true if healthy.
+// If there was an error parsing the response, sets *error and returns false.
+bool DecodeResponse(grpc_slice_buffer* slice_buffer, grpc_error** error) {
+  // If message is empty, assume unhealthy.
+  if (slice_buffer->length == 0) {
+    *error =
+        GRPC_ERROR_CREATE_FROM_STATIC_STRING("health check response was empty");
+    return false;
+  }
+  // Concatenate the slices to form a single string.
+  UniquePtr<uint8_t> recv_message_deleter;
+  uint8_t* recv_message;
+  if (slice_buffer->count == 1) {
+    recv_message = GRPC_SLICE_START_PTR(slice_buffer->slices[0]);
+  } else {
+    recv_message = static_cast<uint8_t*>(gpr_malloc(slice_buffer->length));
+    recv_message_deleter.reset(recv_message);
+    size_t offset = 0;
+    for (size_t i = 0; i < slice_buffer->count; ++i) {
+      memcpy(recv_message + offset,
+             GRPC_SLICE_START_PTR(slice_buffer->slices[i]),
+             GRPC_SLICE_LENGTH(slice_buffer->slices[i]));
+      offset += GRPC_SLICE_LENGTH(slice_buffer->slices[i]);
+    }
+  }
+  // Deserialize message.
+  grpc_health_v1_HealthCheckResponse response_struct;
+  pb_istream_t istream =
+      pb_istream_from_buffer(recv_message, slice_buffer->length);
+  if (!pb_decode(&istream, grpc_health_v1_HealthCheckResponse_fields,
+                 &response_struct)) {
+    // Can't parse message; assume unhealthy.
+    *error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+        "cannot parse health check response");
+    return false;
+  }
+  if (!response_struct.has_status) {
+    // Field not present; assume unhealthy.
+    *error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+        "status field not present in health check response");
+    return false;
+  }
+  return response_struct.status ==
+         grpc_health_v1_HealthCheckResponse_ServingStatus_SERVING;
+}
+
+}  // namespace
+
+//
+// HealthCheckClient::CallState
+//
+
+HealthCheckClient::CallState::CallState(
+    RefCountedPtr<HealthCheckClient> health_check_client,
+    grpc_pollset_set* interested_parties)
+    : InternallyRefCountedWithTracing<CallState>(
+          &grpc_health_check_client_trace),
+      health_check_client_(std::move(health_check_client)),
+      pollent_(grpc_polling_entity_create_from_pollset_set(interested_parties)),
+      arena_(gpr_arena_create(health_check_client_->connected_subchannel_
+                                  ->GetInitialCallSizeEstimate(0))) {
+  memset(&call_combiner_, 0, sizeof(call_combiner_));
+  grpc_call_combiner_init(&call_combiner_);
+  memset(context_, 0, sizeof(context_));
+  gpr_atm_rel_store(&seen_response_, static_cast<gpr_atm>(0));
+}
+
+HealthCheckClient::CallState::~CallState() {
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: destroying CallState %p",
+            health_check_client_.get(), this);
+  }
+  if (call_ != nullptr) GRPC_SUBCHANNEL_CALL_UNREF(call_, "call_ended");
+  // Unset the call combiner cancellation closure.  This has the
+  // effect of scheduling the previously set cancellation closure, if
+  // any, so that it can release any internal references it may be
+  // holding to the call stack. Also flush the closures on exec_ctx so that
+  // filters that schedule cancel notification closures on exec_ctx do not
+  // need to take a ref of the call stack to guarantee closure liveness.
+  grpc_call_combiner_set_notify_on_cancel(&call_combiner_, nullptr);
+  grpc_core::ExecCtx::Get()->Flush();
+  grpc_call_combiner_destroy(&call_combiner_);
+  gpr_arena_destroy(arena_);
+}
+
+void HealthCheckClient::CallState::Orphan() {
+  grpc_call_combiner_cancel(&call_combiner_, GRPC_ERROR_CANCELLED);
+  Cancel();
+}
+
+void HealthCheckClient::CallState::StartCall() {
+  ConnectedSubchannel::CallArgs args = {
+      &pollent_,
+      GRPC_MDSTR_SLASH_GRPC_DOT_HEALTH_DOT_V1_DOT_HEALTH_SLASH_WATCH,
+      gpr_now(GPR_CLOCK_MONOTONIC),  // start_time
+      GRPC_MILLIS_INF_FUTURE,        // deadline
+      arena_,
+      context_,
+      &call_combiner_,
+      0,  // parent_data_size
+  };
+  grpc_error* error =
+      health_check_client_->connected_subchannel_->CreateCall(args, &call_);
+  if (error != GRPC_ERROR_NONE) {
+    gpr_log(GPR_ERROR,
+            "HealthCheckClient %p CallState %p: error creating health "
+            "checking call on subchannel (%s); will retry",
+            health_check_client_.get(), this, grpc_error_string(error));
+    GRPC_ERROR_UNREF(error);
+    // Schedule instead of running directly, since we must not be
+    // holding health_check_client_->mu_ when CallEnded() is called.
+    Ref(DEBUG_LOCATION, "call_end_closure").release();
+    GRPC_CLOSURE_SCHED(
+        GRPC_CLOSURE_INIT(&batch_.handler_private.closure, CallEndedRetry, this,
+                          grpc_schedule_on_exec_ctx),
+        GRPC_ERROR_NONE);
+    return;
+  }
+  // Initialize payload and batch.
+  memset(&batch_, 0, sizeof(batch_));
+  batch_.payload = &payload_;
+  // on_complete callback takes ref, handled manually.
+  Ref(DEBUG_LOCATION, "on_complete").release();
+  batch_.on_complete = GRPC_CLOSURE_INIT(&on_complete_, OnComplete, this,
+                                         grpc_schedule_on_exec_ctx);
+  // Add send_initial_metadata op.
+  grpc_metadata_batch_init(&send_initial_metadata_);
+  error = grpc_metadata_batch_add_head(
+      &send_initial_metadata_, &path_metadata_storage_,
+      grpc_mdelem_from_slices(
+          GRPC_MDSTR_PATH,
+          GRPC_MDSTR_SLASH_GRPC_DOT_HEALTH_DOT_V1_DOT_HEALTH_SLASH_WATCH));
+  GPR_ASSERT(error == GRPC_ERROR_NONE);
+  payload_.send_initial_metadata.send_initial_metadata =
+      &send_initial_metadata_;
+  payload_.send_initial_metadata.send_initial_metadata_flags = 0;
+  payload_.send_initial_metadata.peer_string = nullptr;
+  batch_.send_initial_metadata = true;
+  // Add send_message op.
+  EncodeRequest(health_check_client_->service_name_, &send_message_);
+  payload_.send_message.send_message.reset(send_message_.get());
+  batch_.send_message = true;
+  // Add send_trailing_metadata op.
+  grpc_metadata_batch_init(&send_trailing_metadata_);
+  payload_.send_trailing_metadata.send_trailing_metadata =
+      &send_trailing_metadata_;
+  batch_.send_trailing_metadata = true;
+  // Add recv_initial_metadata op.
+  grpc_metadata_batch_init(&recv_initial_metadata_);
+  payload_.recv_initial_metadata.recv_initial_metadata =
+      &recv_initial_metadata_;
+  payload_.recv_initial_metadata.recv_flags = nullptr;
+  payload_.recv_initial_metadata.trailing_metadata_available = nullptr;
+  payload_.recv_initial_metadata.peer_string = nullptr;
+  // recv_initial_metadata_ready callback takes ref, handled manually.
+  Ref(DEBUG_LOCATION, "recv_initial_metadata_ready").release();
+  payload_.recv_initial_metadata.recv_initial_metadata_ready =
+      GRPC_CLOSURE_INIT(&recv_initial_metadata_ready_, RecvInitialMetadataReady,
+                        this, grpc_schedule_on_exec_ctx);
+  batch_.recv_initial_metadata = true;
+  // Add recv_message op.
+  payload_.recv_message.recv_message = &recv_message_;
+  // recv_message callback takes ref, handled manually.
+  Ref(DEBUG_LOCATION, "recv_message_ready").release();
+  payload_.recv_message.recv_message_ready = GRPC_CLOSURE_INIT(
+      &recv_message_ready_, RecvMessageReady, this, grpc_schedule_on_exec_ctx);
+  batch_.recv_message = true;
+  // Start batch.
+  StartBatch(&batch_);
+  // Initialize recv_trailing_metadata batch.
+  memset(&recv_trailing_metadata_batch_, 0,
+         sizeof(recv_trailing_metadata_batch_));
+  recv_trailing_metadata_batch_.payload = &payload_;
+  // Add recv_trailing_metadata op.
+  grpc_metadata_batch_init(&recv_trailing_metadata_);
+  payload_.recv_trailing_metadata.recv_trailing_metadata =
+      &recv_trailing_metadata_;
+  payload_.recv_trailing_metadata.collect_stats = &collect_stats_;
+  // This callback signals the end of the call, so it relies on the
+  // initial ref instead of taking a new ref.  When it's invoked, the
+  // initial ref is released.
+  payload_.recv_trailing_metadata.recv_trailing_metadata_ready =
+      GRPC_CLOSURE_INIT(&recv_trailing_metadata_ready_,
+                        RecvTrailingMetadataReady, this,
+                        grpc_schedule_on_exec_ctx);
+  recv_trailing_metadata_batch_.recv_trailing_metadata = true;
+  // Start recv_trailing_metadata batch.
+  StartBatch(&recv_trailing_metadata_batch_);
+}
+
+void HealthCheckClient::CallState::StartBatchInCallCombiner(void* arg,
+                                                            grpc_error* error) {
+  grpc_transport_stream_op_batch* batch =
+      static_cast<grpc_transport_stream_op_batch*>(arg);
+  grpc_subchannel_call* call =
+      static_cast<grpc_subchannel_call*>(batch->handler_private.extra_arg);
+  grpc_subchannel_call_process_op(call, batch);
+}
+
+void HealthCheckClient::CallState::StartBatch(
+    grpc_transport_stream_op_batch* batch) {
+  batch->handler_private.extra_arg = call_;
+  GRPC_CLOSURE_INIT(&batch->handler_private.closure, StartBatchInCallCombiner,
+                    batch, grpc_schedule_on_exec_ctx);
+  GRPC_CALL_COMBINER_START(&call_combiner_, &batch->handler_private.closure,
+                           GRPC_ERROR_NONE, "start_subchannel_batch");
+}
+
+void HealthCheckClient::CallState::OnCancelComplete(void* arg,
+                                                    grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  GRPC_CALL_COMBINER_STOP(&self->call_combiner_, "health_cancel");
+  self->Unref(DEBUG_LOCATION, "cancel");
+}
+
+void HealthCheckClient::CallState::StartCancel(void* arg, grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  auto* batch = grpc_make_transport_stream_op(
+      GRPC_CLOSURE_CREATE(OnCancelComplete, self, grpc_schedule_on_exec_ctx));
+  batch->cancel_stream = true;
+  batch->payload->cancel_stream.cancel_error = GRPC_ERROR_CANCELLED;
+  grpc_subchannel_call_process_op(self->call_, batch);
+}
+
+void HealthCheckClient::CallState::Cancel() {
+  if (call_ != nullptr) {
+    Ref(DEBUG_LOCATION, "cancel").release();
+    GRPC_CALL_COMBINER_START(
+        &call_combiner_,
+        GRPC_CLOSURE_CREATE(StartCancel, this, grpc_schedule_on_exec_ctx),
+        GRPC_ERROR_NONE, "health_cancel");
+  }
+}
+
+void HealthCheckClient::CallState::OnComplete(void* arg, grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  GRPC_CALL_COMBINER_STOP(&self->call_combiner_, "on_complete");
+  grpc_metadata_batch_destroy(&self->send_initial_metadata_);
+  grpc_metadata_batch_destroy(&self->send_trailing_metadata_);
+  self->Unref(DEBUG_LOCATION, "on_complete");
+}
+
+void HealthCheckClient::CallState::RecvInitialMetadataReady(void* arg,
+                                                            grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  GRPC_CALL_COMBINER_STOP(&self->call_combiner_, "recv_initial_metadata_ready");
+  grpc_metadata_batch_destroy(&self->recv_initial_metadata_);
+  self->Unref(DEBUG_LOCATION, "recv_initial_metadata_ready");
+}
+
+void HealthCheckClient::CallState::DoneReadingRecvMessage(grpc_error* error) {
+  recv_message_.reset();
+  if (error != GRPC_ERROR_NONE) {
+    GRPC_ERROR_UNREF(error);
+    Cancel();
+    grpc_slice_buffer_destroy_internal(&recv_message_buffer_);
+    Unref(DEBUG_LOCATION, "recv_message_ready");
+    return;
+  }
+  const bool healthy = DecodeResponse(&recv_message_buffer_, &error);
+  const grpc_connectivity_state state =
+      healthy ? GRPC_CHANNEL_READY : GRPC_CHANNEL_TRANSIENT_FAILURE;
+  if (error == GRPC_ERROR_NONE && !healthy) {
+    error = GRPC_ERROR_CREATE_FROM_STATIC_STRING("backend unhealthy");
+  }
+  health_check_client_->SetHealthStatus(state, error);
+  gpr_atm_rel_store(&seen_response_, static_cast<gpr_atm>(1));
+  grpc_slice_buffer_destroy_internal(&recv_message_buffer_);
+  // Start another recv_message batch.
+  // This re-uses the ref we're holding.
+  // Note: Can't just reuse batch_ here, since we don't know that all
+  // callbacks from the original batch have completed yet.
+  memset(&recv_message_batch_, 0, sizeof(recv_message_batch_));
+  recv_message_batch_.payload = &payload_;
+  payload_.recv_message.recv_message = &recv_message_;
+  payload_.recv_message.recv_message_ready = GRPC_CLOSURE_INIT(
+      &recv_message_ready_, RecvMessageReady, this, grpc_schedule_on_exec_ctx);
+  recv_message_batch_.recv_message = true;
+  StartBatch(&recv_message_batch_);
+}
+
+grpc_error* HealthCheckClient::CallState::PullSliceFromRecvMessage() {
+  grpc_slice slice;
+  grpc_error* error = recv_message_->Pull(&slice);
+  if (error == GRPC_ERROR_NONE) {
+    grpc_slice_buffer_add(&recv_message_buffer_, slice);
+  }
+  return error;
+}
+
+void HealthCheckClient::CallState::ContinueReadingRecvMessage() {
+  while (recv_message_->Next(SIZE_MAX, &recv_message_ready_)) {
+    grpc_error* error = PullSliceFromRecvMessage();
+    if (error != GRPC_ERROR_NONE) {
+      DoneReadingRecvMessage(error);
+      return;
+    }
+    if (recv_message_buffer_.length == recv_message_->length()) {
+      DoneReadingRecvMessage(GRPC_ERROR_NONE);
+      break;
+    }
+  }
+}
+
+void HealthCheckClient::CallState::OnByteStreamNext(void* arg,
+                                                    grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  if (error != GRPC_ERROR_NONE) {
+    self->DoneReadingRecvMessage(GRPC_ERROR_REF(error));
+    return;
+  }
+  error = self->PullSliceFromRecvMessage();
+  if (error != GRPC_ERROR_NONE) {
+    self->DoneReadingRecvMessage(error);
+    return;
+  }
+  if (self->recv_message_buffer_.length == self->recv_message_->length()) {
+    self->DoneReadingRecvMessage(GRPC_ERROR_NONE);
+  } else {
+    self->ContinueReadingRecvMessage();
+  }
+}
+
+void HealthCheckClient::CallState::RecvMessageReady(void* arg,
+                                                    grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  GRPC_CALL_COMBINER_STOP(&self->call_combiner_, "recv_message_ready");
+  if (self->recv_message_ == nullptr) {
+    self->Unref(DEBUG_LOCATION, "recv_message_ready");
+    return;
+  }
+  grpc_slice_buffer_init(&self->recv_message_buffer_);
+  GRPC_CLOSURE_INIT(&self->recv_message_ready_, OnByteStreamNext, self,
+                    grpc_schedule_on_exec_ctx);
+  self->ContinueReadingRecvMessage();
+  // Ref will continue to be held until we finish draining the byte stream.
+}
+
+void HealthCheckClient::CallState::RecvTrailingMetadataReady(
+    void* arg, grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  GRPC_CALL_COMBINER_STOP(&self->call_combiner_,
+                          "recv_trailing_metadata_ready");
+  // Get call status.
+  grpc_status_code status = GRPC_STATUS_UNKNOWN;
+  if (error != GRPC_ERROR_NONE) {
+    grpc_error_get_status(error, GRPC_MILLIS_INF_FUTURE, &status,
+                          nullptr /* slice */, nullptr /* http_error */,
+                          nullptr /* error_string */);
+  } else if (self->recv_trailing_metadata_.idx.named.grpc_status != nullptr) {
+    status = grpc_get_status_code_from_metadata(
+        self->recv_trailing_metadata_.idx.named.grpc_status->md);
+  }
+  if (grpc_health_check_client_trace.enabled()) {
+    gpr_log(GPR_INFO,
+            "HealthCheckClient %p CallState %p: health watch failed with "
+            "status %d",
+            self->health_check_client_.get(), self, status);
+  }
+  // Clean up.
+  grpc_metadata_batch_destroy(&self->recv_trailing_metadata_);
+  // For status UNIMPLEMENTED, give up and assume always healthy.
+  bool retry = true;
+  if (status == GRPC_STATUS_UNIMPLEMENTED) {
+    static const char kErrorMessage[] =
+        "health checking Watch method returned UNIMPLEMENTED; "
+        "disabling health checks but assuming server is healthy";
+    gpr_log(GPR_ERROR, kErrorMessage);
+    if (self->health_check_client_->channelz_node_ != nullptr) {
+      self->health_check_client_->channelz_node_->AddTraceEvent(
+          channelz::ChannelTrace::Error,
+          grpc_slice_from_static_string(kErrorMessage));
+    }
+    self->health_check_client_->SetHealthStatus(GRPC_CHANNEL_READY,
+                                                GRPC_ERROR_NONE);
+    retry = false;
+  }
+  self->CallEnded(retry);
+}
+
+void HealthCheckClient::CallState::CallEndedRetry(void* arg,
+                                                  grpc_error* error) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  self->CallEnded(true /* retry */);
+  self->Unref(DEBUG_LOCATION, "call_end_closure");
+}
+
+void HealthCheckClient::CallState::CallEnded(bool retry) {
+  // If this CallState is still in use, this call ended because of a failure,
+  // so we need to stop using it and optionally create a new one.
+  // Otherwise, we have deliberately ended this call, and no further action
+  // is required.
+  if (this == health_check_client_->call_state_.get()) {
+    health_check_client_->call_state_.reset();
+    if (retry) {
+      GPR_ASSERT(!health_check_client_->shutting_down_);
+      if (static_cast<bool>(gpr_atm_acq_load(&seen_response_))) {
+        // If the call fails after we've gotten a successful response, reset
+        // the backoff and restart the call immediately.
+        health_check_client_->retry_backoff_.Reset();
+        health_check_client_->StartCall();
+      } else {
+        // If the call failed without receiving any messages, retry later.
+        health_check_client_->StartRetryTimer();
+      }
+    }
+  }
+  Unref(DEBUG_LOCATION, "call_ended");
+}
+
+}  // namespace grpc_core
diff --git a/src/core/ext/filters/client_channel/health/health_check_client.h b/src/core/ext/filters/client_channel/health/health_check_client.h
new file mode 100644
index 0000000000..7f77348f18
--- /dev/null
+++ b/src/core/ext/filters/client_channel/health/health_check_client.h
@@ -0,0 +1,173 @@
+/*
+ *
+ * Copyright 2018 gRPC authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ */
+
+#ifndef GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_HEALTH_HEALTH_CHECK_CLIENT_H
+#define GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_HEALTH_HEALTH_CHECK_CLIENT_H
+
+#include <grpc/support/port_platform.h>
+
+#include <grpc/grpc.h>
+#include <grpc/support/atm.h>
+#include <grpc/support/sync.h>
+
+#include "src/core/ext/filters/client_channel/client_channel_channelz.h"
+#include "src/core/ext/filters/client_channel/subchannel.h"
+#include "src/core/lib/backoff/backoff.h"
+#include "src/core/lib/gpr/arena.h"
+#include "src/core/lib/gprpp/orphanable.h"
+#include "src/core/lib/gprpp/ref_counted_ptr.h"
+#include "src/core/lib/iomgr/call_combiner.h"
+#include "src/core/lib/iomgr/closure.h"
+#include "src/core/lib/iomgr/polling_entity.h"
+#include "src/core/lib/iomgr/timer.h"
+#include "src/core/lib/transport/byte_stream.h"
+#include "src/core/lib/transport/metadata_batch.h"
+#include "src/core/lib/transport/transport.h"
+
+namespace grpc_core {
+
+class HealthCheckClient
+    : public InternallyRefCountedWithTracing<HealthCheckClient> {
+ public:
+  HealthCheckClient(const char* service_name,
+                    RefCountedPtr<ConnectedSubchannel> connected_subchannel,
+                    grpc_pollset_set* interested_parties,
+                    RefCountedPtr<channelz::SubchannelNode> channelz_node);
+
+  ~HealthCheckClient();
+
+  // When the health state changes from *state, sets *state to the new
+  // value and schedules closure.
+  // Only one closure can be outstanding at a time.
+  void NotifyOnHealthChange(grpc_connectivity_state* state,
+                            grpc_closure* closure);
+
+  void Orphan() override;
+
+ private:
+  // Contains a call to the backend and all the data related to the call.
+  class CallState : public InternallyRefCountedWithTracing<CallState> {
+   public:
+    CallState(RefCountedPtr<HealthCheckClient> health_check_client,
+              grpc_pollset_set* interested_parties_);
+    ~CallState();
+
+    void Orphan() override;
+
+    void StartCall();
+
+   private:
+    void Cancel();
+
+    void StartBatch(grpc_transport_stream_op_batch* batch);
+    static void StartBatchInCallCombiner(void* arg, grpc_error* error);
+
+    static void CallEndedRetry(void* arg, grpc_error* error);
+    void CallEnded(bool retry);
+
+    static void OnComplete(void* arg, grpc_error* error);
+    static void RecvInitialMetadataReady(void* arg, grpc_error* error);
+    static void RecvMessageReady(void* arg, grpc_error* error);
+    static void RecvTrailingMetadataReady(void* arg, grpc_error* error);
+    static void StartCancel(void* arg, grpc_error* error);
+    static void OnCancelComplete(void* arg, grpc_error* error);
+
+    static void OnByteStreamNext(void* arg, grpc_error* error);
+    void ContinueReadingRecvMessage();
+    grpc_error* PullSliceFromRecvMessage();
+    void DoneReadingRecvMessage(grpc_error* error);
+
+    RefCountedPtr<HealthCheckClient> health_check_client_;
+    grpc_polling_entity pollent_;
+
+    gpr_arena* arena_;
+    grpc_call_combiner call_combiner_;
+    grpc_call_context_element context_[GRPC_CONTEXT_COUNT];
+
+    // The streaming call to the backend. Always non-NULL.
+    grpc_subchannel_call* call_;
+
+    grpc_transport_stream_op_batch_payload payload_;
+    grpc_transport_stream_op_batch batch_;
+    grpc_transport_stream_op_batch recv_message_batch_;
+    grpc_transport_stream_op_batch recv_trailing_metadata_batch_;
+
+    grpc_closure on_complete_;
+
+    // send_initial_metadata
+    grpc_metadata_batch send_initial_metadata_;
+    grpc_linked_mdelem path_metadata_storage_;
+
+    // send_message
+    ManualConstructor<SliceBufferByteStream> send_message_;
+
+    // send_trailing_metadata
+    grpc_metadata_batch send_trailing_metadata_;
+
+    // recv_initial_metadata
+    grpc_metadata_batch recv_initial_metadata_;
+    grpc_closure recv_initial_metadata_ready_;
+
+    // recv_message
+    OrphanablePtr<ByteStream> recv_message_;
+    grpc_closure recv_message_ready_;
+    grpc_slice_buffer recv_message_buffer_;
+    gpr_atm seen_response_;
+
+    // recv_trailing_metadata
+    grpc_metadata_batch recv_trailing_metadata_;
+    grpc_transport_stream_stats collect_stats_;
+    grpc_closure recv_trailing_metadata_ready_;
+  };
+
+  void StartCall();
+  void StartCallLocked();  // Requires holding mu_.
+
+  void StartRetryTimer();
+  static void OnRetryTimer(void* arg, grpc_error* error);
+
+  void SetHealthStatus(grpc_connectivity_state state, grpc_error* error);
+  void SetHealthStatusLocked(grpc_connectivity_state state,
+                             grpc_error* error);  // Requires holding mu_.
+
+  const char* service_name_;  // Do not own.
+  RefCountedPtr<ConnectedSubchannel> connected_subchannel_;
+  grpc_pollset_set* interested_parties_;  // Do not own.
+  RefCountedPtr<channelz::SubchannelNode> channelz_node_;
+
+  gpr_mu mu_;
+  grpc_connectivity_state state_ = GRPC_CHANNEL_CONNECTING;
+  grpc_error* error_ = GRPC_ERROR_NONE;
+  grpc_connectivity_state* notify_state_ = nullptr;
+  grpc_closure* on_health_changed_ = nullptr;
+  bool shutting_down_ = false;
+
+  // The data associated with the current health check call.  It holds a ref
+  // to this HealthCheckClient object.
+  OrphanablePtr<CallState> call_state_;
+
+  // Call retry state.
+  BackOff retry_backoff_;
+  grpc_timer retry_timer_;
+  grpc_closure retry_timer_callback_;
+  bool retry_timer_callback_pending_ = false;
+};
+
+}  // namespace grpc_core
+
+#endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_HEALTH_HEALTH_CHECK_CLIENT_H */