grpc 1.15.0 → 1.30.0

data/src/core/ext/filters/client_channel/connector.h

@@ -22,52 +22,58 @@
 #include <grpc/support/port_platform.h>
 
 #include "src/core/lib/channel/channel_stack.h"
+#include "src/core/lib/channel/channelz.h"
+#include "src/core/lib/gprpp/orphanable.h"
 #include "src/core/lib/iomgr/resolve_address.h"
 #include "src/core/lib/transport/transport.h"
 
-typedef struct grpc_connector grpc_connector;
-typedef struct grpc_connector_vtable grpc_connector_vtable;
+namespace grpc_core {
 
-struct grpc_connector {
-  const grpc_connector_vtable* vtable;
-};
+// Interface for connection-establishment functionality.
+// Each transport that supports client channels (e.g., not inproc) must
+// supply an implementation of this.
+class SubchannelConnector : public InternallyRefCounted<SubchannelConnector> {
+ public:
+  struct Args {
+    // Set of pollsets interested in this connection.
+    grpc_pollset_set* interested_parties;
+    // Deadline for connection.
+    grpc_millis deadline;
+    // Channel args to be passed to handshakers and transport.
+    const grpc_channel_args* channel_args;
+  };
+
+  struct Result {
+    // The connected transport.
+    grpc_transport* transport = nullptr;
+    // Channel args to be passed to filters.
+    const grpc_channel_args* channel_args = nullptr;
+    // Channelz socket node of the connected transport, if any.
+    RefCountedPtr<channelz::SocketNode> socket_node;
 
-typedef struct {
-  /** set of pollsets interested in this connection */
-  grpc_pollset_set* interested_parties;
-  /** deadline for connection */
-  grpc_millis deadline;
-  /** channel arguments (to be passed to transport) */
-  const grpc_channel_args* channel_args;
-} grpc_connect_in_args;
+    void Reset() {
+      transport = nullptr;
+      channel_args = nullptr;
+      socket_node.reset();
+    }
+  };
 
-typedef struct {
-  /** the connected transport */
-  grpc_transport* transport;
+  // Attempts to connect.
+  // When complete, populates *result and invokes notify.
+  // Only one connection attempt may be in progress at any one time.
+  virtual void Connect(const Args& args, Result* result,
+                       grpc_closure* notify) = 0;
 
-  /** channel arguments (to be passed to the filters) */
-  grpc_channel_args* channel_args;
-} grpc_connect_out_args;
+  // Cancels any in-flight connection attempt and shuts down the
+  // connector.
+  virtual void Shutdown(grpc_error* error) = 0;
 
-struct grpc_connector_vtable {
-  void (*ref)(grpc_connector* connector);
-  void (*unref)(grpc_connector* connector);
-  /** Implementation of grpc_connector_shutdown */
-  void (*shutdown)(grpc_connector* connector, grpc_error* why);
-  /** Implementation of grpc_connector_connect */
-  void (*connect)(grpc_connector* connector,
-                  const grpc_connect_in_args* in_args,
-                  grpc_connect_out_args* out_args, grpc_closure* notify);
+  void Orphan() override {
+    Shutdown(GRPC_ERROR_CREATE_FROM_STATIC_STRING("Subchannel disconnected"));
+    Unref();
+  }
 };
 
-grpc_connector* grpc_connector_ref(grpc_connector* connector);
-void grpc_connector_unref(grpc_connector* connector);
-/** Connect using the connector: max one outstanding call at a time */
-void grpc_connector_connect(grpc_connector* connector,
-                            const grpc_connect_in_args* in_args,
-                            grpc_connect_out_args* out_args,
-                            grpc_closure* notify);
-/** Cancel any pending connection */
-void grpc_connector_shutdown(grpc_connector* connector, grpc_error* why);
+}  // namespace grpc_core
 
 #endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_CONNECTOR_H */

data/src/core/ext/filters/client_channel/global_subchannel_pool.cc

@@ -0,0 +1,179 @@
+//
+//
+// 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 "src/core/ext/filters/client_channel/global_subchannel_pool.h"
+
+#include "src/core/ext/filters/client_channel/subchannel.h"
+
+namespace grpc_core {
+
+GlobalSubchannelPool::GlobalSubchannelPool() {
+  subchannel_map_ = grpc_avl_create(&subchannel_avl_vtable_);
+  gpr_mu_init(&mu_);
+}
+
+GlobalSubchannelPool::~GlobalSubchannelPool() {
+  gpr_mu_destroy(&mu_);
+  grpc_avl_unref(subchannel_map_, nullptr);
+}
+
+void GlobalSubchannelPool::Init() {
+  instance_ = new RefCountedPtr<GlobalSubchannelPool>(
+      MakeRefCounted<GlobalSubchannelPool>());
+}
+
+void GlobalSubchannelPool::Shutdown() {
+  // To ensure Init() was called before.
+  GPR_ASSERT(instance_ != nullptr);
+  // To ensure Shutdown() was not called before.
+  GPR_ASSERT(*instance_ != nullptr);
+  instance_->reset();
+  delete instance_;
+}
+
+RefCountedPtr<GlobalSubchannelPool> GlobalSubchannelPool::instance() {
+  GPR_ASSERT(instance_ != nullptr);
+  GPR_ASSERT(*instance_ != nullptr);
+  return *instance_;
+}
+
+Subchannel* GlobalSubchannelPool::RegisterSubchannel(SubchannelKey* key,
+                                                     Subchannel* constructed) {
+  Subchannel* c = nullptr;
+  // Compare and swap (CAS) loop:
+  while (c == nullptr) {
+    // Ref the shared map to have a local copy.
+    gpr_mu_lock(&mu_);
+    grpc_avl old_map = grpc_avl_ref(subchannel_map_, nullptr);
+    gpr_mu_unlock(&mu_);
+    // Check to see if a subchannel already exists.
+    c = static_cast<Subchannel*>(grpc_avl_get(old_map, key, nullptr));
+    if (c != nullptr) {
+      // The subchannel already exists. Try to reuse it.
+      c = GRPC_SUBCHANNEL_REF_FROM_WEAK_REF(c, "subchannel_register+reuse");
+      if (c != nullptr) {
+        GRPC_SUBCHANNEL_UNREF(constructed,
+                              "subchannel_register+found_existing");
+        // Exit the CAS loop without modifying the shared map.
+      }  // Else, reuse failed, so retry CAS loop.
+    } else {
+      // There hasn't been such subchannel. Add one.
+      // Note that we should ref the old map first because grpc_avl_add() will
+      // unref it while we still need to access it later.
+      grpc_avl new_map = grpc_avl_add(
+          grpc_avl_ref(old_map, nullptr), new SubchannelKey(*key),
+          GRPC_SUBCHANNEL_WEAK_REF(constructed, "subchannel_register+new"),
+          nullptr);
+      // Try to publish the change to the shared map. It may happen (but
+      // unlikely) that some other thread has changed the shared map, so compare
+      // to make sure it's unchanged before swapping. Retry if it's changed.
+      gpr_mu_lock(&mu_);
+      if (old_map.root == subchannel_map_.root) {
+        GPR_SWAP(grpc_avl, new_map, subchannel_map_);
+        c = constructed;
+      }
+      gpr_mu_unlock(&mu_);
+      grpc_avl_unref(new_map, nullptr);
+    }
+    grpc_avl_unref(old_map, nullptr);
+  }
+  return c;
+}
+
+void GlobalSubchannelPool::UnregisterSubchannel(SubchannelKey* key) {
+  bool done = false;
+  // Compare and swap (CAS) loop:
+  while (!done) {
+    // Ref the shared map to have a local copy.
+    gpr_mu_lock(&mu_);
+    grpc_avl old_map = grpc_avl_ref(subchannel_map_, nullptr);
+    gpr_mu_unlock(&mu_);
+    // Remove the subchannel.
+    // Note that we should ref the old map first because grpc_avl_remove() will
+    // unref it while we still need to access it later.
+    grpc_avl new_map =
+        grpc_avl_remove(grpc_avl_ref(old_map, nullptr), key, nullptr);
+    // Try to publish the change to the shared map. It may happen (but
+    // unlikely) that some other thread has changed the shared map, so compare
+    // to make sure it's unchanged before swapping. Retry if it's changed.
+    gpr_mu_lock(&mu_);
+    if (old_map.root == subchannel_map_.root) {
+      GPR_SWAP(grpc_avl, new_map, subchannel_map_);
+      done = true;
+    }
+    gpr_mu_unlock(&mu_);
+    grpc_avl_unref(new_map, nullptr);
+    grpc_avl_unref(old_map, nullptr);
+  }
+}
+
+Subchannel* GlobalSubchannelPool::FindSubchannel(SubchannelKey* key) {
+  // Lock, and take a reference to the subchannel map.
+  // We don't need to do the search under a lock as AVL's are immutable.
+  gpr_mu_lock(&mu_);
+  grpc_avl index = grpc_avl_ref(subchannel_map_, nullptr);
+  gpr_mu_unlock(&mu_);
+  Subchannel* c = static_cast<Subchannel*>(grpc_avl_get(index, key, nullptr));
+  if (c != nullptr) c = GRPC_SUBCHANNEL_REF_FROM_WEAK_REF(c, "found_from_pool");
+  grpc_avl_unref(index, nullptr);
+  return c;
+}
+
+RefCountedPtr<GlobalSubchannelPool>* GlobalSubchannelPool::instance_ = nullptr;
+
+namespace {
+
+void sck_avl_destroy(void* p, void* /*user_data*/) {
+  SubchannelKey* key = static_cast<SubchannelKey*>(p);
+  delete key;
+}
+
+void* sck_avl_copy(void* p, void* /*unused*/) {
+  const SubchannelKey* key = static_cast<const SubchannelKey*>(p);
+  auto* new_key = new SubchannelKey(*key);
+  return static_cast<void*>(new_key);
+}
+
+long sck_avl_compare(void* a, void* b, void* /*unused*/) {
+  const SubchannelKey* key_a = static_cast<const SubchannelKey*>(a);
+  const SubchannelKey* key_b = static_cast<const SubchannelKey*>(b);
+  return key_a->Cmp(*key_b);
+}
+
+void scv_avl_destroy(void* p, void* /*user_data*/) {
+  GRPC_SUBCHANNEL_WEAK_UNREF((Subchannel*)p, "global_subchannel_pool");
+}
+
+void* scv_avl_copy(void* p, void* /*unused*/) {
+  GRPC_SUBCHANNEL_WEAK_REF((Subchannel*)p, "global_subchannel_pool");
+  return p;
+}
+
+}  // namespace
+
+const grpc_avl_vtable GlobalSubchannelPool::subchannel_avl_vtable_ = {
+    sck_avl_destroy,  // destroy_key
+    sck_avl_copy,     // copy_key
+    sck_avl_compare,  // compare_keys
+    scv_avl_destroy,  // destroy_value
+    scv_avl_copy      // copy_value
+};
+
+}  // namespace grpc_core

data/src/core/ext/filters/client_channel/global_subchannel_pool.h

@@ -0,0 +1,68 @@
+/*
+ *
+ * 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_GLOBAL_SUBCHANNEL_POOL_H
+#define GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_GLOBAL_SUBCHANNEL_POOL_H
+
+#include <grpc/support/port_platform.h>
+
+#include "src/core/ext/filters/client_channel/subchannel_pool_interface.h"
+
+namespace grpc_core {
+
+// The global subchannel pool. It shares subchannels among channels. There
+// should be only one instance of this class. Init() should be called once at
+// the filter initialization time; Shutdown() should be called once at the
+// filter shutdown time.
+// TODO(juanlishen): Enable subchannel retention.
+class GlobalSubchannelPool final : public SubchannelPoolInterface {
+ public:
+  // The ctor and dtor are not intended to use directly.
+  GlobalSubchannelPool();
+  ~GlobalSubchannelPool() override;
+
+  // Should be called exactly once at filter initialization time.
+  static void Init();
+  // Should be called exactly once at filter shutdown time.
+  static void Shutdown();
+
+  // Gets the singleton instance.
+  static RefCountedPtr<GlobalSubchannelPool> instance();
+
+  // Implements interface methods.
+  Subchannel* RegisterSubchannel(SubchannelKey* key,
+                                 Subchannel* constructed) override;
+  void UnregisterSubchannel(SubchannelKey* key) override;
+  Subchannel* FindSubchannel(SubchannelKey* key) override;
+
+ private:
+  // The singleton instance. (It's a pointer to RefCountedPtr so that this
+  // non-local static object can be trivially destructible.)
+  static RefCountedPtr<GlobalSubchannelPool>* instance_;
+
+  // The vtable for subchannel operations in an AVL tree.
+  static const grpc_avl_vtable subchannel_avl_vtable_;
+  // A map from subchannel key to subchannel.
+  grpc_avl subchannel_map_;
+  // To protect subchannel_map_.
+  gpr_mu mu_;
+};
+
+}  // namespace grpc_core
+
+#endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_GLOBAL_SUBCHANNEL_POOL_H */

data/src/core/ext/filters/client_channel/health/health_check_client.cc

@@ -0,0 +1,606 @@
+/*
+ *
+ * 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 <stdio.h>
+
+#include "src/core/ext/filters/client_channel/health/health_check_client.h"
+
+#include "src/core/lib/debug/trace.h"
+#include "src/core/lib/gprpp/sync.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"
+#include "src/proto/grpc/health/v1/health.upb.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
+
+namespace grpc_core {
+
+TraceFlag grpc_health_check_client_trace(false, "health_check_client");
+
+//
+// HealthCheckClient
+//
+
+HealthCheckClient::HealthCheckClient(
+    const char* service_name,
+    RefCountedPtr<ConnectedSubchannel> connected_subchannel,
+    grpc_pollset_set* interested_parties,
+    RefCountedPtr<channelz::SubchannelNode> channelz_node,
+    RefCountedPtr<ConnectivityStateWatcherInterface> watcher)
+    : InternallyRefCounted<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)),
+      watcher_(std::move(watcher)),
+      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_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    gpr_log(GPR_INFO, "created HealthCheckClient %p", this);
+  }
+  GRPC_CLOSURE_INIT(&retry_timer_callback_, OnRetryTimer, this,
+                    grpc_schedule_on_exec_ctx);
+  StartCall();
+}
+
+HealthCheckClient::~HealthCheckClient() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    gpr_log(GPR_INFO, "destroying HealthCheckClient %p", this);
+  }
+}
+
+void HealthCheckClient::SetHealthStatus(grpc_connectivity_state state,
+                                        const char* reason) {
+  MutexLock lock(&mu_);
+  SetHealthStatusLocked(state, reason);
+}
+
+void HealthCheckClient::SetHealthStatusLocked(grpc_connectivity_state state,
+                                              const char* reason) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: setting state=%s reason=%s", this,
+            ConnectivityStateName(state), reason);
+  }
+  if (watcher_ != nullptr) watcher_->Notify(state);
+}
+
+void HealthCheckClient::Orphan() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: shutting down", this);
+  }
+  {
+    MutexLock lock(&mu_);
+    shutting_down_ = true;
+    watcher_.reset();
+    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, "starting health watch");
+  call_state_ = MakeOrphanable<CallState>(Ref(), interested_parties_);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: created CallState %p", this,
+            call_state_.get());
+  }
+  call_state_->StartCall();
+}
+
+void HealthCheckClient::StartRetryTimerLocked() {
+  SetHealthStatusLocked(GRPC_CHANNEL_TRANSIENT_FAILURE,
+                        "health check call failed; will retry after backoff");
+  grpc_millis next_try = retry_backoff_.NextAttemptTime();
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    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_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+        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) {
+  upb::Arena arena;
+  grpc_health_v1_HealthCheckRequest* request_struct =
+      grpc_health_v1_HealthCheckRequest_new(arena.ptr());
+  grpc_health_v1_HealthCheckRequest_set_service(
+      request_struct, upb_strview_makez(service_name));
+  size_t buf_length;
+  char* buf = grpc_health_v1_HealthCheckRequest_serialize(
+      request_struct, arena.ptr(), &buf_length);
+  grpc_slice request_slice = GRPC_SLICE_MALLOC(buf_length);
+  memcpy(GRPC_SLICE_START_PTR(request_slice), buf, buf_length);
+  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.
+  std::unique_ptr<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.
+  upb::Arena arena;
+  grpc_health_v1_HealthCheckResponse* response_struct =
+      grpc_health_v1_HealthCheckResponse_parse(
+          reinterpret_cast<char*>(recv_message), slice_buffer->length,
+          arena.ptr());
+  if (response_struct == nullptr) {
+    // Can't parse message; assume unhealthy.
+    *error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+        "cannot parse health check response");
+    return false;
+  }
+  int32_t status = grpc_health_v1_HealthCheckResponse_status(response_struct);
+  return status == grpc_health_v1_HealthCheckResponse_SERVING;
+}
+
+}  // namespace
+
+//
+// HealthCheckClient::CallState
+//
+
+HealthCheckClient::CallState::CallState(
+    RefCountedPtr<HealthCheckClient> health_check_client,
+    grpc_pollset_set* interested_parties)
+    : health_check_client_(std::move(health_check_client)),
+      pollent_(grpc_polling_entity_create_from_pollset_set(interested_parties)),
+      arena_(Arena::Create(health_check_client_->connected_subchannel_
+                               ->GetInitialCallSizeEstimate(0))),
+      payload_(context_) {}
+
+HealthCheckClient::CallState::~CallState() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    gpr_log(GPR_INFO, "HealthCheckClient %p: destroying CallState %p",
+            health_check_client_.get(), this);
+  }
+  for (size_t i = 0; i < GRPC_CONTEXT_COUNT; i++) {
+    if (context_[i].destroy != nullptr) {
+      context_[i].destroy(context_[i].value);
+    }
+  }
+  // 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.
+  call_combiner_.SetNotifyOnCancel(nullptr);
+  ExecCtx::Get()->Flush();
+  arena_->Destroy();
+}
+
+void HealthCheckClient::CallState::Orphan() {
+  call_combiner_.Cancel(GRPC_ERROR_CANCELLED);
+  Cancel();
+}
+
+void HealthCheckClient::CallState::StartCall() {
+  SubchannelCall::Args args = {
+      health_check_client_->connected_subchannel_,
+      &pollent_,
+      GRPC_MDSTR_SLASH_GRPC_DOT_HEALTH_DOT_V1_DOT_HEALTH_SLASH_WATCH,
+      gpr_get_cycle_counter(),  // start_time
+      GRPC_MILLIS_INF_FUTURE,   // deadline
+      arena_,
+      context_,
+      &call_combiner_,
+      0,  // parent_data_size
+  };
+  grpc_error* error = GRPC_ERROR_NONE;
+  call_ = SubchannelCall::Create(std::move(args), &error).release();
+  // Register after-destruction callback.
+  GRPC_CLOSURE_INIT(&after_call_stack_destruction_, AfterCallStackDestruction,
+                    this, grpc_schedule_on_exec_ctx);
+  call_->SetAfterCallStackDestroy(&after_call_stack_destruction_);
+  // Check if creation failed.
+  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);
+    CallEndedLocked(/*retry=*/true);
+    return;
+  }
+  // Initialize payload and batch.
+  payload_.context = context_;
+  batch_.payload = &payload_;
+  // on_complete callback takes ref, handled manually.
+  call_->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),
+      GRPC_BATCH_PATH);
+  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.
+  call_->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.
+  call_->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.
+  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);
+  SubchannelCall* call =
+      static_cast<SubchannelCall*>(batch->handler_private.extra_arg);
+  call->StartTransportStreamOpBatch(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::AfterCallStackDestruction(
+    void* arg, grpc_error* /*error*/) {
+  HealthCheckClient::CallState* self =
+      static_cast<HealthCheckClient::CallState*>(arg);
+  delete self;
+}
+
+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->call_->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;
+  self->call_->StartTransportStreamOpBatch(batch);
+}
+
+void HealthCheckClient::CallState::Cancel() {
+  bool expected = false;
+  if (cancelled_.CompareExchangeStrong(&expected, true, MemoryOrder::ACQ_REL,
+                                       MemoryOrder::ACQUIRE)) {
+    call_->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->call_->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->call_->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_);
+    call_->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;
+  const char* reason = error == GRPC_ERROR_NONE && !healthy
+                           ? "backend unhealthy"
+                           : grpc_error_string(error);
+  health_check_client_->SetHealthStatus(state, reason);
+  seen_response_.Store(true, MemoryOrder::RELEASE);
+  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.
+  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->call_->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_TRACE_FLAG_ENABLED(grpc_health_check_client_trace)) {
+    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,
+                                                kErrorMessage);
+    retry = false;
+  }
+  MutexLock lock(&self->health_check_client_->mu_);
+  self->CallEndedLocked(retry);
+}
+
+void HealthCheckClient::CallState::CallEndedLocked(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 (seen_response_.Load(MemoryOrder::ACQUIRE)) {
+        // 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_->StartCallLocked();
+      } else {
+        // If the call failed without receiving any messages, retry later.
+        health_check_client_->StartRetryTimerLocked();
+      }
+    }
+  }
+  // When the last ref to the call stack goes away, the CallState object
+  // will be automatically destroyed.
+  call_->Unref(DEBUG_LOCATION, "call_ended");
+}
+
+}  // namespace grpc_core