grpc 1.28.0 → 1.42.0

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

@@ -0,0 +1,2573 @@
+//
+// Copyright 2015 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/retry_filter.h"
+
+#include "absl/container/inlined_vector.h"
+#include "absl/status/statusor.h"
+#include "absl/strings/strip.h"
+
+#include <grpc/support/log.h>
+
+#include "src/core/ext/filters/client_channel/client_channel.h"
+#include "src/core/ext/filters/client_channel/retry_service_config.h"
+#include "src/core/ext/filters/client_channel/retry_throttle.h"
+#include "src/core/ext/service_config/service_config.h"
+#include "src/core/ext/service_config/service_config_call_data.h"
+#include "src/core/lib/backoff/backoff.h"
+#include "src/core/lib/channel/channel_args.h"
+#include "src/core/lib/channel/channel_stack.h"
+#include "src/core/lib/channel/status_util.h"
+#include "src/core/lib/gprpp/manual_constructor.h"
+#include "src/core/lib/iomgr/polling_entity.h"
+#include "src/core/lib/slice/slice_internal.h"
+#include "src/core/lib/slice/slice_string_helpers.h"
+#include "src/core/lib/transport/error_utils.h"
+#include "src/core/lib/transport/metadata.h"
+#include "src/core/lib/transport/metadata_batch.h"
+#include "src/core/lib/transport/static_metadata.h"
+#include "src/core/lib/transport/status_metadata.h"
+#include "src/core/lib/uri/uri_parser.h"
+
+//
+// Retry filter
+//
+
+// This filter is intended to be used in the DynamicFilter stack in the
+// client channel, which is situated between the name resolver and the
+// LB policy.  Normally, the last filter in the DynamicFilter stack is
+// the DynamicTerminationFilter (see client_channel.cc), which creates a
+// LoadBalancedCall and delegates to it.  However, when retries are
+// enabled, this filter is used instead of the DynamicTerminationFilter.
+//
+// In order to support retries, we act as a proxy for stream op batches.
+// When we get a batch from the surface, we add it to our list of pending
+// batches, and we then use those batches to construct separate "child"
+// batches to be started on an LB call.  When the child batches return, we
+// then decide which pending batches have been completed and schedule their
+// callbacks accordingly.  If a call attempt fails and we want to retry it,
+// we create a new LB call and start again, constructing new "child" batches
+// for the new LB call.
+//
+// Note that retries are committed when receiving data from the server
+// (except for Trailers-Only responses).  However, there may be many
+// send ops started before receiving any data, so we may have already
+// completed some number of send ops (and returned the completions up to
+// the surface) by the time we realize that we need to retry.  To deal
+// with this, we cache data for send ops, so that we can replay them on a
+// different LB call even after we have completed the original batches.
+//
+// The code is structured as follows:
+// - In CallData (in the parent channel), we maintain a list of pending
+//   ops and cached data for send ops.
+// - There is a CallData::CallAttempt object for each retry attempt.
+//   This object contains the LB call for that attempt and state to indicate
+//   which ops from the CallData object have already been sent down to that
+//   LB call.
+// - There is a CallData::CallAttempt::BatchData object for each "child"
+//   batch sent on the LB call.
+//
+// When constructing the "child" batches, we compare the state in the
+// CallAttempt object against the state in the CallData object to see
+// which batches need to be sent on the LB call for a given attempt.
+
+// TODO(roth): In subsequent PRs:
+// - add support for transparent retries (including initial metadata)
+// - implement hedging
+
+// By default, we buffer 256 KiB per RPC for retries.
+// TODO(roth): Do we have any data to suggest a better value?
+#define DEFAULT_PER_RPC_RETRY_BUFFER_SIZE (256 << 10)
+
+// This value was picked arbitrarily.  It can be changed if there is
+// any even moderately compelling reason to do so.
+#define RETRY_BACKOFF_JITTER 0.2
+
+namespace grpc_core {
+
+namespace {
+
+using internal::RetryGlobalConfig;
+using internal::RetryMethodConfig;
+using internal::RetryServiceConfigParser;
+using internal::ServerRetryThrottleData;
+
+TraceFlag grpc_retry_trace(false, "retry");
+
+//
+// RetryFilter
+//
+
+class RetryFilter {
+ public:
+  class CallData;
+
+  static grpc_error_handle Init(grpc_channel_element* elem,
+                                grpc_channel_element_args* args) {
+    GPR_ASSERT(args->is_last);
+    GPR_ASSERT(elem->filter == &kRetryFilterVtable);
+    grpc_error_handle error = GRPC_ERROR_NONE;
+    new (elem->channel_data) RetryFilter(args->channel_args, &error);
+    return error;
+  }
+
+  static void Destroy(grpc_channel_element* elem) {
+    auto* chand = static_cast<RetryFilter*>(elem->channel_data);
+    chand->~RetryFilter();
+  }
+
+  // Will never be called.
+  static void StartTransportOp(grpc_channel_element* /*elem*/,
+                               grpc_transport_op* /*op*/) {}
+  static void GetChannelInfo(grpc_channel_element* /*elem*/,
+                             const grpc_channel_info* /*info*/) {}
+
+ private:
+  static size_t GetMaxPerRpcRetryBufferSize(const grpc_channel_args* args) {
+    return static_cast<size_t>(grpc_channel_args_find_integer(
+        args, GRPC_ARG_PER_RPC_RETRY_BUFFER_SIZE,
+        {DEFAULT_PER_RPC_RETRY_BUFFER_SIZE, 0, INT_MAX}));
+  }
+
+  RetryFilter(const grpc_channel_args* args, grpc_error_handle* error)
+      : client_channel_(grpc_channel_args_find_pointer<ClientChannel>(
+            args, GRPC_ARG_CLIENT_CHANNEL)),
+        per_rpc_retry_buffer_size_(GetMaxPerRpcRetryBufferSize(args)) {
+    // Get retry throttling parameters from service config.
+    auto* service_config = grpc_channel_args_find_pointer<ServiceConfig>(
+        args, GRPC_ARG_SERVICE_CONFIG_OBJ);
+    if (service_config == nullptr) return;
+    const auto* config = static_cast<const RetryGlobalConfig*>(
+        service_config->GetGlobalParsedConfig(
+            RetryServiceConfigParser::ParserIndex()));
+    if (config == nullptr) return;
+    // Get server name from target URI.
+    const char* server_uri =
+        grpc_channel_args_find_string(args, GRPC_ARG_SERVER_URI);
+    if (server_uri == nullptr) {
+      *error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+          "server URI channel arg missing or wrong type in client channel "
+          "filter");
+      return;
+    }
+    absl::StatusOr<URI> uri = URI::Parse(server_uri);
+    if (!uri.ok() || uri->path().empty()) {
+      *error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+          "could not extract server name from target URI");
+      return;
+    }
+    std::string server_name(absl::StripPrefix(uri->path(), "/"));
+    // Get throttling config for server_name.
+    retry_throttle_data_ = internal::ServerRetryThrottleMap::GetDataForServer(
+        server_name, config->max_milli_tokens(), config->milli_token_ratio());
+  }
+
+  ClientChannel* client_channel_;
+  size_t per_rpc_retry_buffer_size_;
+  RefCountedPtr<ServerRetryThrottleData> retry_throttle_data_;
+};
+
+//
+// RetryFilter::CallData
+//
+
+class RetryFilter::CallData {
+ public:
+  static grpc_error_handle Init(grpc_call_element* elem,
+                                const grpc_call_element_args* args);
+  static void Destroy(grpc_call_element* elem,
+                      const grpc_call_final_info* /*final_info*/,
+                      grpc_closure* then_schedule_closure);
+  static void StartTransportStreamOpBatch(
+      grpc_call_element* elem, grpc_transport_stream_op_batch* batch);
+  static void SetPollent(grpc_call_element* elem, grpc_polling_entity* pollent);
+
+ private:
+  class CallStackDestructionBarrier;
+
+  // Pending batches stored in call data.
+  struct PendingBatch {
+    // The pending batch.  If nullptr, this slot is empty.
+    grpc_transport_stream_op_batch* batch = nullptr;
+    // Indicates whether payload for send ops has been cached in CallData.
+    bool send_ops_cached = false;
+  };
+
+  // State associated with each call attempt.
+  class CallAttempt : public RefCounted<CallAttempt> {
+   public:
+    explicit CallAttempt(CallData* calld);
+    ~CallAttempt() override;
+
+    bool lb_call_committed() const { return lb_call_committed_; }
+
+    // Constructs and starts whatever batches are needed on this call
+    // attempt.
+    void StartRetriableBatches();
+
+    // Frees cached send ops that have already been completed after
+    // committing the call.
+    void FreeCachedSendOpDataAfterCommit();
+
+    // Cancels the call attempt.
+    void CancelFromSurface(grpc_transport_stream_op_batch* cancel_batch);
+
+   private:
+    // State used for starting a retryable batch on the call attempt's LB call.
+    // This provides its own grpc_transport_stream_op_batch and other data
+    // structures needed to populate the ops in the batch.
+    // We allocate one struct on the arena for each attempt at starting a
+    // batch on a given LB call.
+    class BatchData
+        : public RefCounted<BatchData, PolymorphicRefCount, kUnrefCallDtor> {
+     public:
+      BatchData(RefCountedPtr<CallAttempt> call_attempt, int refcount,
+                bool set_on_complete);
+      ~BatchData() override;
+
+      grpc_transport_stream_op_batch* batch() { return &batch_; }
+
+      // Adds retriable send_initial_metadata op.
+      void AddRetriableSendInitialMetadataOp();
+      // Adds retriable send_message op.
+      void AddRetriableSendMessageOp();
+      // Adds retriable send_trailing_metadata op.
+      void AddRetriableSendTrailingMetadataOp();
+      // Adds retriable recv_initial_metadata op.
+      void AddRetriableRecvInitialMetadataOp();
+      // Adds retriable recv_message op.
+      void AddRetriableRecvMessageOp();
+      // Adds retriable recv_trailing_metadata op.
+      void AddRetriableRecvTrailingMetadataOp();
+      // Adds cancel_stream op.
+      void AddCancelStreamOp(grpc_error_handle error);
+
+     private:
+      // Frees cached send ops that were completed by the completed batch in
+      // batch_data.  Used when batches are completed after the call is
+      // committed.
+      void FreeCachedSendOpDataForCompletedBatch();
+
+      // If there is a pending recv_initial_metadata op, adds a closure
+      // to closures for recv_initial_metadata_ready.
+      void MaybeAddClosureForRecvInitialMetadataCallback(
+          grpc_error_handle error, CallCombinerClosureList* closures);
+      // Intercepts recv_initial_metadata_ready callback for retries.
+      // Commits the call and returns the initial metadata up the stack.
+      static void RecvInitialMetadataReady(void* arg, grpc_error_handle error);
+
+      // If there is a pending recv_message op, adds a closure to closures
+      // for recv_message_ready.
+      void MaybeAddClosureForRecvMessageCallback(
+          grpc_error_handle error, CallCombinerClosureList* closures);
+      // Intercepts recv_message_ready callback for retries.
+      // Commits the call and returns the message up the stack.
+      static void RecvMessageReady(void* arg, grpc_error_handle error);
+
+      // If there is a pending recv_trailing_metadata op, adds a closure to
+      // closures for recv_trailing_metadata_ready.
+      void MaybeAddClosureForRecvTrailingMetadataReady(
+          grpc_error_handle error, CallCombinerClosureList* closures);
+      // Adds any necessary closures for deferred batch completion
+      // callbacks to closures.
+      void AddClosuresForDeferredCompletionCallbacks(
+          CallCombinerClosureList* closures);
+      // For any pending batch containing an op that has not yet been started,
+      // adds the pending batch's completion closures to closures.
+      void AddClosuresToFailUnstartedPendingBatches(
+          grpc_error_handle error, CallCombinerClosureList* closures);
+      // Runs necessary closures upon completion of a call attempt.
+      void RunClosuresForCompletedCall(grpc_error_handle error);
+      // Intercepts recv_trailing_metadata_ready callback for retries.
+      // Commits the call and returns the trailing metadata up the stack.
+      static void RecvTrailingMetadataReady(void* arg, grpc_error_handle error);
+
+      // Adds the on_complete closure for the pending batch completed in
+      // batch_data to closures.
+      void AddClosuresForCompletedPendingBatch(
+          grpc_error_handle error, CallCombinerClosureList* closures);
+
+      // If there are any cached ops to replay or pending ops to start on the
+      // LB call, adds them to closures.
+      void AddClosuresForReplayOrPendingSendOps(
+          CallCombinerClosureList* closures);
+
+      // Callback used to intercept on_complete from LB calls.
+      static void OnComplete(void* arg, grpc_error_handle error);
+
+      // Callback used to handle on_complete for internally generated
+      // cancel_stream op.
+      static void OnCompleteForCancelOp(void* arg, grpc_error_handle error);
+
+      RefCountedPtr<CallAttempt> call_attempt_;
+      // The batch to use in the LB call.
+      // Its payload field points to CallAttempt::batch_payload_.
+      grpc_transport_stream_op_batch batch_;
+      // For intercepting on_complete.
+      grpc_closure on_complete_;
+    };
+
+    class AttemptDispatchController
+        : public ConfigSelector::CallDispatchController {
+     public:
+      explicit AttemptDispatchController(CallAttempt* call_attempt)
+          : call_attempt_(call_attempt) {}
+
+      // Will never be called.
+      bool ShouldRetry() override { return false; }
+
+      void Commit() override {
+        call_attempt_->lb_call_committed_ = true;
+        auto* calld = call_attempt_->calld_;
+        if (calld->retry_committed_) {
+          auto* service_config_call_data =
+              static_cast<ClientChannelServiceConfigCallData*>(
+                  calld->call_context_[GRPC_CONTEXT_SERVICE_CONFIG_CALL_DATA]
+                      .value);
+          service_config_call_data->call_dispatch_controller()->Commit();
+        }
+      }
+
+     private:
+      CallAttempt* call_attempt_;
+    };
+
+    // Creates a BatchData object on the call's arena with the
+    // specified refcount.  If set_on_complete is true, the batch's
+    // on_complete callback will be set to point to on_complete();
+    // otherwise, the batch's on_complete callback will be null.
+    BatchData* CreateBatch(int refcount, bool set_on_complete) {
+      return calld_->arena_->New<BatchData>(Ref(DEBUG_LOCATION, "CreateBatch"),
+                                            refcount, set_on_complete);
+    }
+
+    // If there are any cached send ops that need to be replayed on this
+    // call attempt, creates and returns a new batch to replay those ops.
+    // Otherwise, returns nullptr.
+    BatchData* MaybeCreateBatchForReplay();
+
+    // Adds a closure to closures that will execute batch in the call combiner.
+    void AddClosureForBatch(grpc_transport_stream_op_batch* batch,
+                            const char* reason,
+                            CallCombinerClosureList* closures);
+
+    // Helper function used to start a recv_trailing_metadata batch.  This
+    // is used in the case where a recv_initial_metadata or recv_message
+    // op fails in a way that we know the call is over but when the application
+    // has not yet started its own recv_trailing_metadata op.
+    void AddBatchForInternalRecvTrailingMetadata(
+        CallCombinerClosureList* closures);
+
+    // Adds a batch to closures to cancel this call attempt.
+    void AddBatchForCancelOp(grpc_error_handle error,
+                             CallCombinerClosureList* closures);
+
+    // Adds batches for pending batches to closures.
+    void AddBatchesForPendingBatches(CallCombinerClosureList* closures);
+
+    // Adds whatever batches are needed on this attempt to closures.
+    void AddRetriableBatches(CallCombinerClosureList* closures);
+
+    // Returns true if any send op in the batch was not yet started on this
+    // attempt.
+    bool PendingBatchContainsUnstartedSendOps(PendingBatch* pending);
+
+    // Returns true if there are cached send ops to replay.
+    bool HaveSendOpsToReplay();
+
+    // If our retry state is no longer needed, switch to fast path by moving
+    // our LB call into calld_->committed_call_ and having calld_ drop
+    // its ref to us.
+    void MaybeSwitchToFastPath();
+
+    // Returns true if the call should be retried.
+    // If server_pushback_md is non-null, sets *server_pushback_ms.
+    bool ShouldRetry(absl::optional<grpc_status_code> status, bool is_lb_drop,
+                     grpc_mdelem* server_pushback_md,
+                     grpc_millis* server_pushback_ms);
+
+    // Abandons the call attempt.  Unrefs any deferred batches.
+    void Abandon();
+
+    static void OnPerAttemptRecvTimer(void* arg, grpc_error_handle error);
+    static void OnPerAttemptRecvTimerLocked(void* arg, grpc_error_handle error);
+    void MaybeCancelPerAttemptRecvTimer();
+
+    CallData* calld_;
+    AttemptDispatchController attempt_dispatch_controller_;
+    OrphanablePtr<ClientChannel::LoadBalancedCall> lb_call_;
+    bool lb_call_committed_ = false;
+
+    grpc_timer per_attempt_recv_timer_;
+    grpc_closure on_per_attempt_recv_timer_;
+    bool per_attempt_recv_timer_pending_ = false;
+
+    // BatchData.batch.payload points to this.
+    grpc_transport_stream_op_batch_payload batch_payload_;
+    // For send_initial_metadata.
+    grpc_linked_mdelem retry_attempts_metadata_;
+    grpc_metadata_batch send_initial_metadata_{calld_->arena_};
+    // For send_message.
+    // TODO(roth): Restructure this to eliminate use of ManualConstructor.
+    ManualConstructor<ByteStreamCache::CachingByteStream> send_message_;
+    // For send_trailing_metadata.
+    grpc_metadata_batch send_trailing_metadata_{calld_->arena_};
+    // For intercepting recv_initial_metadata.
+    grpc_metadata_batch recv_initial_metadata_{calld_->arena_};
+    grpc_closure recv_initial_metadata_ready_;
+    bool trailing_metadata_available_ = false;
+    // For intercepting recv_message.
+    grpc_closure recv_message_ready_;
+    OrphanablePtr<ByteStream> recv_message_;
+    // For intercepting recv_trailing_metadata.
+    grpc_metadata_batch recv_trailing_metadata_{calld_->arena_};
+    grpc_transport_stream_stats collect_stats_;
+    grpc_closure recv_trailing_metadata_ready_;
+    // These fields indicate which ops have been started and completed on
+    // this call attempt.
+    size_t started_send_message_count_ = 0;
+    size_t completed_send_message_count_ = 0;
+    size_t started_recv_message_count_ = 0;
+    size_t completed_recv_message_count_ = 0;
+    bool started_send_initial_metadata_ : 1;
+    bool completed_send_initial_metadata_ : 1;
+    bool started_send_trailing_metadata_ : 1;
+    bool completed_send_trailing_metadata_ : 1;
+    bool started_recv_initial_metadata_ : 1;
+    bool completed_recv_initial_metadata_ : 1;
+    bool started_recv_trailing_metadata_ : 1;
+    bool completed_recv_trailing_metadata_ : 1;
+    // State for callback processing.
+    RefCountedPtr<BatchData> recv_initial_metadata_ready_deferred_batch_;
+    grpc_error_handle recv_initial_metadata_error_ = GRPC_ERROR_NONE;
+    RefCountedPtr<BatchData> recv_message_ready_deferred_batch_;
+    grpc_error_handle recv_message_error_ = GRPC_ERROR_NONE;
+    struct OnCompleteDeferredBatch {
+      OnCompleteDeferredBatch(RefCountedPtr<BatchData> batch,
+                              grpc_error_handle error)
+          : batch(std::move(batch)), error(error) {}
+      RefCountedPtr<BatchData> batch;
+      grpc_error_handle error;
+    };
+    // There cannot be more than 3 pending send op batches at a time.
+    absl::InlinedVector<OnCompleteDeferredBatch, 3>
+        on_complete_deferred_batches_;
+    RefCountedPtr<BatchData> recv_trailing_metadata_internal_batch_;
+    grpc_error_handle recv_trailing_metadata_error_ = GRPC_ERROR_NONE;
+    bool seen_recv_trailing_metadata_from_surface_ : 1;
+    // NOTE: Do not move this next to the metadata bitfields above. That would
+    //       save space but will also result in a data race because compiler
+    //       will generate a 2 byte store which overwrites the meta-data
+    //       fields upon setting this field.
+    bool abandoned_ : 1;
+  };
+
+  CallData(RetryFilter* chand, const grpc_call_element_args& args);
+  ~CallData();
+
+  void StartTransportStreamOpBatch(grpc_transport_stream_op_batch* batch);
+
+  // Returns the index into pending_batches_ to be used for batch.
+  static size_t GetBatchIndex(grpc_transport_stream_op_batch* batch);
+  PendingBatch* PendingBatchesAdd(grpc_transport_stream_op_batch* batch);
+  void PendingBatchClear(PendingBatch* pending);
+  void MaybeClearPendingBatch(PendingBatch* pending);
+  static void FailPendingBatchInCallCombiner(void* arg,
+                                             grpc_error_handle error);
+  // Fails all pending batches.  Does NOT yield call combiner.
+  void PendingBatchesFail(grpc_error_handle error);
+  // Returns a pointer to the first pending batch for which predicate(batch)
+  // returns true, or null if not found.
+  template <typename Predicate>
+  PendingBatch* PendingBatchFind(const char* log_message, Predicate predicate);
+
+  // Caches data for send ops so that it can be retried later, if not
+  // already cached.
+  void MaybeCacheSendOpsForBatch(PendingBatch* pending);
+  void FreeCachedSendInitialMetadata();
+  // Frees cached send_message at index idx.
+  void FreeCachedSendMessage(size_t idx);
+  void FreeCachedSendTrailingMetadata();
+  void FreeAllCachedSendOpData();
+
+  // Commits the call so that no further retry attempts will be performed.
+  void RetryCommit(CallAttempt* call_attempt);
+
+  // Starts a timer to retry after appropriate back-off.
+  // If server_pushback_ms is -1, retry_backoff_ is used.
+  void StartRetryTimer(grpc_millis server_pushback_ms);
+
+  static void OnRetryTimer(void* arg, grpc_error_handle error);
+  static void OnRetryTimerLocked(void* arg, grpc_error_handle error);
+
+  OrphanablePtr<ClientChannel::LoadBalancedCall> CreateLoadBalancedCall(
+      ConfigSelector::CallDispatchController* call_dispatch_controller);
+
+  void CreateCallAttempt();
+
+  RetryFilter* chand_;
+  grpc_polling_entity* pollent_;
+  RefCountedPtr<ServerRetryThrottleData> retry_throttle_data_;
+  const RetryMethodConfig* retry_policy_ = nullptr;
+  BackOff retry_backoff_;
+
+  grpc_slice path_;  // Request path.
+  grpc_millis deadline_;
+  Arena* arena_;
+  grpc_call_stack* owning_call_;
+  CallCombiner* call_combiner_;
+  grpc_call_context_element* call_context_;
+
+  grpc_error_handle cancelled_from_surface_ = GRPC_ERROR_NONE;
+
+  RefCountedPtr<CallStackDestructionBarrier> call_stack_destruction_barrier_;
+
+  // TODO(roth): As part of implementing hedging, we will need to maintain a
+  // list of all pending attempts, so that we can cancel them all if the call
+  // gets cancelled.
+  RefCountedPtr<CallAttempt> call_attempt_;
+
+  // LB call used when we've committed to a call attempt and the retry
+  // state for that attempt is no longer needed.  This provides a fast
+  // path for long-running streaming calls that minimizes overhead.
+  OrphanablePtr<ClientChannel::LoadBalancedCall> committed_call_;
+
+  // When are are not yet fully committed to a particular call (i.e.,
+  // either we might still retry or we have committed to the call but
+  // there are still some cached ops to be replayed on the call),
+  // batches received from above will be added to this list, and they
+  // will not be removed until we have invoked their completion callbacks.
+  size_t bytes_buffered_for_retry_ = 0;
+  PendingBatch pending_batches_[MAX_PENDING_BATCHES];
+  bool pending_send_initial_metadata_ : 1;
+  bool pending_send_message_ : 1;
+  bool pending_send_trailing_metadata_ : 1;
+
+  // Retry state.
+  bool retry_committed_ : 1;
+  bool retry_timer_pending_ : 1;
+  int num_attempts_completed_ = 0;
+  grpc_timer retry_timer_;
+  grpc_closure retry_closure_;
+
+  // Cached data for retrying send ops.
+  // send_initial_metadata
+  bool seen_send_initial_metadata_ = false;
+  grpc_metadata_batch send_initial_metadata_{arena_};
+  uint32_t send_initial_metadata_flags_;
+  // TODO(roth): As part of implementing hedging, we'll probably need to
+  // have the LB call set a value in CallAttempt and then propagate it
+  // from CallAttempt to the parent call when we commit.  Otherwise, we
+  // may leave this with a value for a peer other than the one we
+  // actually commit to.  Alternatively, maybe see if there's a way to
+  // change the surface API such that the peer isn't available until
+  // after initial metadata is received?  (Could even change the
+  // transport API to return this with the recv_initial_metadata op.)
+  gpr_atm* peer_string_;
+  // send_message
+  // When we get a send_message op, we replace the original byte stream
+  // with a CachingByteStream that caches the slices to a local buffer for
+  // use in retries.
+  // Note: We inline the cache for the first 3 send_message ops and use
+  // dynamic allocation after that.  This number was essentially picked
+  // at random; it could be changed in the future to tune performance.
+  // TODO(roth): As part of implementing hedging, we may need some
+  // synchronization here, since ByteStreamCache does not provide any
+  // synchronization, so it's not safe to have multiple
+  // CachingByteStreams read from the same ByteStreamCache concurrently.
+  absl::InlinedVector<ByteStreamCache*, 3> send_messages_;
+  // send_trailing_metadata
+  bool seen_send_trailing_metadata_ = false;
+  grpc_metadata_batch send_trailing_metadata_{arena_};
+};
+
+//
+// RetryFilter::CallData::CallStackDestructionBarrier
+//
+
+// A class to track the existence of LoadBalancedCall call stacks that
+// we've created.  We wait until all such call stacks have been
+// destroyed before we return the on_call_stack_destruction closure up
+// to the surface.
+//
+// The parent RetryFilter::CallData object holds a ref to this object.
+// When it is destroyed, it will store the on_call_stack_destruction
+// closure from the surface in this object and then release its ref.
+// We also take a ref to this object for each LB call we create, and
+// those refs are not released until the LB call stack is destroyed.
+// When this object is destroyed, it will invoke the
+// on_call_stack_destruction closure from the surface.
+class RetryFilter::CallData::CallStackDestructionBarrier
+    : public RefCounted<CallStackDestructionBarrier, PolymorphicRefCount,
+                        kUnrefCallDtor> {
+ public:
+  CallStackDestructionBarrier() {}
+
+  ~CallStackDestructionBarrier() override {
+    // TODO(yashkt) : This can potentially be a Closure::Run
+    ExecCtx::Run(DEBUG_LOCATION, on_call_stack_destruction_, GRPC_ERROR_NONE);
+  }
+
+  // Set the closure from the surface.  This closure will be invoked
+  // when this object is destroyed.
+  void set_on_call_stack_destruction(grpc_closure* on_call_stack_destruction) {
+    on_call_stack_destruction_ = on_call_stack_destruction;
+  }
+
+  // Invoked to get an on_call_stack_destruction closure for a new LB call.
+  grpc_closure* MakeLbCallDestructionClosure(CallData* calld) {
+    Ref().release();  // Ref held by callback.
+    grpc_closure* on_lb_call_destruction_complete =
+        calld->arena_->New<grpc_closure>();
+    GRPC_CLOSURE_INIT(on_lb_call_destruction_complete,
+                      OnLbCallDestructionComplete, this, nullptr);
+    return on_lb_call_destruction_complete;
+  }
+
+ private:
+  static void OnLbCallDestructionComplete(void* arg,
+                                          grpc_error_handle /*error*/) {
+    auto* self = static_cast<CallStackDestructionBarrier*>(arg);
+    self->Unref();
+  }
+
+  grpc_closure* on_call_stack_destruction_ = nullptr;
+};
+
+//
+// RetryFilter::CallData::CallAttempt
+//
+
+RetryFilter::CallData::CallAttempt::CallAttempt(CallData* calld)
+    : RefCounted(GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace) ? "CallAttempt"
+                                                           : nullptr),
+      calld_(calld),
+      attempt_dispatch_controller_(this),
+      batch_payload_(calld->call_context_),
+      started_send_initial_metadata_(false),
+      completed_send_initial_metadata_(false),
+      started_send_trailing_metadata_(false),
+      completed_send_trailing_metadata_(false),
+      started_recv_initial_metadata_(false),
+      completed_recv_initial_metadata_(false),
+      started_recv_trailing_metadata_(false),
+      completed_recv_trailing_metadata_(false),
+      seen_recv_trailing_metadata_from_surface_(false),
+      abandoned_(false) {
+  lb_call_ = calld->CreateLoadBalancedCall(&attempt_dispatch_controller_);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: create lb_call=%p",
+            calld->chand_, calld, this, lb_call_.get());
+  }
+  // If per_attempt_recv_timeout is set, start a timer.
+  if (calld->retry_policy_ != nullptr &&
+      calld->retry_policy_->per_attempt_recv_timeout().has_value()) {
+    grpc_millis per_attempt_recv_deadline =
+        ExecCtx::Get()->Now() +
+        *calld->retry_policy_->per_attempt_recv_timeout();
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: per-attempt timeout in %" PRId64
+              " ms",
+              calld->chand_, calld, this,
+              *calld->retry_policy_->per_attempt_recv_timeout());
+    }
+    // Schedule retry after computed delay.
+    GRPC_CLOSURE_INIT(&on_per_attempt_recv_timer_, OnPerAttemptRecvTimer, this,
+                      nullptr);
+    GRPC_CALL_STACK_REF(calld->owning_call_, "OnPerAttemptRecvTimer");
+    Ref(DEBUG_LOCATION, "OnPerAttemptRecvTimer").release();
+    per_attempt_recv_timer_pending_ = true;
+    grpc_timer_init(&per_attempt_recv_timer_, per_attempt_recv_deadline,
+                    &on_per_attempt_recv_timer_);
+  }
+}
+
+RetryFilter::CallData::CallAttempt::~CallAttempt() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: destroying call attempt",
+            calld_->chand_, calld_, this);
+  }
+}
+
+void RetryFilter::CallData::CallAttempt::FreeCachedSendOpDataAfterCommit() {
+  // TODO(roth): When we implement hedging, this logic will need to get
+  // a bit more complex, because there may be other (now abandoned) call
+  // attempts still using this data.  We may need to do some sort of
+  // ref-counting instead.
+  if (completed_send_initial_metadata_) {
+    calld_->FreeCachedSendInitialMetadata();
+  }
+  for (size_t i = 0; i < completed_send_message_count_; ++i) {
+    calld_->FreeCachedSendMessage(i);
+  }
+  if (completed_send_trailing_metadata_) {
+    calld_->FreeCachedSendTrailingMetadata();
+  }
+}
+
+bool RetryFilter::CallData::CallAttempt::PendingBatchContainsUnstartedSendOps(
+    PendingBatch* pending) {
+  if (pending->batch->on_complete == nullptr) return false;
+  if (pending->batch->send_initial_metadata &&
+      !started_send_initial_metadata_) {
+    return true;
+  }
+  if (pending->batch->send_message &&
+      started_send_message_count_ < calld_->send_messages_.size()) {
+    return true;
+  }
+  if (pending->batch->send_trailing_metadata &&
+      !started_send_trailing_metadata_) {
+    return true;
+  }
+  return false;
+}
+
+bool RetryFilter::CallData::CallAttempt::HaveSendOpsToReplay() {
+  // We don't check send_initial_metadata here, because that op will always
+  // be started as soon as it is received from the surface, so it will
+  // never need to be started at this point.
+  return started_send_message_count_ < calld_->send_messages_.size() ||
+         (calld_->seen_send_trailing_metadata_ &&
+          !started_send_trailing_metadata_);
+}
+
+void RetryFilter::CallData::CallAttempt::MaybeSwitchToFastPath() {
+  // If we're not yet committed, we can't switch yet.
+  // TODO(roth): As part of implementing hedging, this logic needs to
+  // check that *this* call attempt is the one that we've committed to.
+  // Might need to replace abandoned_ with an enum indicating whether we're
+  // in flight, abandoned, or the winning call attempt.
+  if (!calld_->retry_committed_) return;
+  // If we've already switched to fast path, there's nothing to do here.
+  if (calld_->committed_call_ != nullptr) return;
+  // If the perAttemptRecvTimeout timer is pending, we can't switch yet.
+  if (per_attempt_recv_timer_pending_) return;
+  // If there are still send ops to replay, we can't switch yet.
+  if (HaveSendOpsToReplay()) return;
+  // If we started an internal batch for recv_trailing_metadata but have not
+  // yet seen that op from the surface, we can't switch yet.
+  if (recv_trailing_metadata_internal_batch_ != nullptr) return;
+  // Switch to fast path.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p: retry state no longer needed; "
+            "moving LB call to parent and unreffing the call attempt",
+            calld_->chand_, calld_, this);
+  }
+  calld_->committed_call_ = std::move(lb_call_);
+  calld_->call_attempt_.reset(DEBUG_LOCATION, "MaybeSwitchToFastPath");
+}
+
+// If there are any cached send ops that need to be replayed on the
+// current call attempt, creates and returns a new batch to replay those ops.
+// Otherwise, returns nullptr.
+RetryFilter::CallData::CallAttempt::BatchData*
+RetryFilter::CallData::CallAttempt::MaybeCreateBatchForReplay() {
+  BatchData* replay_batch_data = nullptr;
+  // send_initial_metadata.
+  if (calld_->seen_send_initial_metadata_ && !started_send_initial_metadata_ &&
+      !calld_->pending_send_initial_metadata_) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: replaying previously completed "
+              "send_initial_metadata op",
+              calld_->chand_, calld_, this);
+    }
+    replay_batch_data = CreateBatch(1, true /* set_on_complete */);
+    replay_batch_data->AddRetriableSendInitialMetadataOp();
+  }
+  // send_message.
+  // Note that we can only have one send_message op in flight at a time.
+  if (started_send_message_count_ < calld_->send_messages_.size() &&
+      started_send_message_count_ == completed_send_message_count_ &&
+      !calld_->pending_send_message_) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: replaying previously completed "
+              "send_message op",
+              calld_->chand_, calld_, this);
+    }
+    if (replay_batch_data == nullptr) {
+      replay_batch_data = CreateBatch(1, true /* set_on_complete */);
+    }
+    replay_batch_data->AddRetriableSendMessageOp();
+  }
+  // send_trailing_metadata.
+  // Note that we only add this op if we have no more send_message ops
+  // to start, since we can't send down any more send_message ops after
+  // send_trailing_metadata.
+  if (calld_->seen_send_trailing_metadata_ &&
+      started_send_message_count_ == calld_->send_messages_.size() &&
+      !started_send_trailing_metadata_ &&
+      !calld_->pending_send_trailing_metadata_) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: replaying previously completed "
+              "send_trailing_metadata op",
+              calld_->chand_, calld_, this);
+    }
+    if (replay_batch_data == nullptr) {
+      replay_batch_data = CreateBatch(1, true /* set_on_complete */);
+    }
+    replay_batch_data->AddRetriableSendTrailingMetadataOp();
+  }
+  return replay_batch_data;
+}
+
+namespace {
+
+void StartBatchInCallCombiner(void* arg, grpc_error_handle /*ignored*/) {
+  grpc_transport_stream_op_batch* batch =
+      static_cast<grpc_transport_stream_op_batch*>(arg);
+  auto* lb_call = static_cast<ClientChannel::LoadBalancedCall*>(
+      batch->handler_private.extra_arg);
+  // Note: This will release the call combiner.
+  lb_call->StartTransportStreamOpBatch(batch);
+}
+
+}  // namespace
+
+void RetryFilter::CallData::CallAttempt::AddClosureForBatch(
+    grpc_transport_stream_op_batch* batch, const char* reason,
+    CallCombinerClosureList* closures) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: adding batch (%s): %s",
+            calld_->chand_, calld_, this, reason,
+            grpc_transport_stream_op_batch_string(batch).c_str());
+  }
+  batch->handler_private.extra_arg = lb_call_.get();
+  GRPC_CLOSURE_INIT(&batch->handler_private.closure, StartBatchInCallCombiner,
+                    batch, grpc_schedule_on_exec_ctx);
+  closures->Add(&batch->handler_private.closure, GRPC_ERROR_NONE, reason);
+}
+
+void RetryFilter::CallData::CallAttempt::
+    AddBatchForInternalRecvTrailingMetadata(CallCombinerClosureList* closures) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p: call failed but "
+            "recv_trailing_metadata not started; starting it internally",
+            calld_->chand_, calld_, this);
+  }
+  // Create batch_data with 2 refs, since this batch will be unreffed twice:
+  // once for the recv_trailing_metadata_ready callback when the batch
+  // completes, and again when we actually get a recv_trailing_metadata
+  // op from the surface.
+  BatchData* batch_data = CreateBatch(2, false /* set_on_complete */);
+  batch_data->AddRetriableRecvTrailingMetadataOp();
+  recv_trailing_metadata_internal_batch_.reset(batch_data);
+  AddClosureForBatch(batch_data->batch(),
+                     "starting internal recv_trailing_metadata", closures);
+}
+
+void RetryFilter::CallData::CallAttempt::AddBatchForCancelOp(
+    grpc_error_handle error, CallCombinerClosureList* closures) {
+  BatchData* cancel_batch_data = CreateBatch(1, /*set_on_complete=*/true);
+  cancel_batch_data->AddCancelStreamOp(error);
+  AddClosureForBatch(cancel_batch_data->batch(),
+                     "start cancellation batch on call attempt", closures);
+}
+
+void RetryFilter::CallData::CallAttempt::AddBatchesForPendingBatches(
+    CallCombinerClosureList* closures) {
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(calld_->pending_batches_); ++i) {
+    PendingBatch* pending = &calld_->pending_batches_[i];
+    grpc_transport_stream_op_batch* batch = pending->batch;
+    if (batch == nullptr) continue;
+    bool has_send_ops = false;
+    // Skip any batch that either (a) has already been started on this
+    // call attempt or (b) we can't start yet because we're still
+    // replaying send ops that need to be completed first.
+    // TODO(roth): Note that if any one op in the batch can't be sent
+    // yet due to ops that we're replaying, we don't start any of the ops
+    // in the batch.  This is probably okay, but it could conceivably
+    // lead to increased latency in some cases -- e.g., we could delay
+    // starting a recv op due to it being in the same batch with a send
+    // op.  If/when we revamp the callback protocol in
+    // transport_stream_op_batch, we may be able to fix this.
+    if (batch->send_initial_metadata) {
+      if (started_send_initial_metadata_) continue;
+      has_send_ops = true;
+    }
+    if (batch->send_message) {
+      if (completed_send_message_count_ < started_send_message_count_) {
+        continue;
+      }
+      has_send_ops = true;
+    }
+    // Note that we only start send_trailing_metadata if we have no more
+    // send_message ops to start, since we can't send down any more
+    // send_message ops after send_trailing_metadata.
+    if (batch->send_trailing_metadata) {
+      if (started_send_message_count_ + batch->send_message <
+              calld_->send_messages_.size() ||
+          started_send_trailing_metadata_) {
+        continue;
+      }
+      has_send_ops = true;
+    }
+    int num_callbacks = has_send_ops;  // All send ops share one callback.
+    if (batch->recv_initial_metadata) {
+      if (started_recv_initial_metadata_) continue;
+      ++num_callbacks;
+    }
+    if (batch->recv_message) {
+      if (completed_recv_message_count_ < started_recv_message_count_) {
+        continue;
+      }
+      ++num_callbacks;
+    }
+    if (batch->recv_trailing_metadata) {
+      if (started_recv_trailing_metadata_) {
+        seen_recv_trailing_metadata_from_surface_ = true;
+        // If we previously completed a recv_trailing_metadata op
+        // initiated by AddBatchForInternalRecvTrailingMetadata(), use the
+        // result of that instead of trying to re-start this op.
+        if (GPR_UNLIKELY(recv_trailing_metadata_internal_batch_ != nullptr)) {
+          // If the batch completed, then trigger the completion callback
+          // directly, so that we return the previously returned results to
+          // the application.  Otherwise, just unref the internally started
+          // batch, since we'll propagate the completion when it completes.
+          if (completed_recv_trailing_metadata_) {
+            closures->Add(
+                &recv_trailing_metadata_ready_, recv_trailing_metadata_error_,
+                "re-executing recv_trailing_metadata_ready to propagate "
+                "internally triggered result");
+            // Ref will be released by callback.
+            recv_trailing_metadata_internal_batch_.release();
+          } else {
+            recv_trailing_metadata_internal_batch_.reset(
+                DEBUG_LOCATION,
+                "internally started recv_trailing_metadata batch pending and "
+                "recv_trailing_metadata started from surface");
+            GRPC_ERROR_UNREF(recv_trailing_metadata_error_);
+          }
+          recv_trailing_metadata_error_ = GRPC_ERROR_NONE;
+        }
+        // We don't want the fact that we've already started this op internally
+        // to prevent us from adding a batch that may contain other ops.
+        // Instead, we'll just skip adding this op below.
+        if (num_callbacks == 0) continue;
+      } else {
+        ++num_callbacks;
+      }
+    }
+    // If we're already committed and the following conditions are met,
+    // just send the batch down as-is:
+    // - The batch contains no cached send ops.  (If it does, we need
+    //   the logic below to use the cached payloads.)
+    // - The batch does not contain recv_trailing_metadata when we have
+    //   already started an internal recv_trailing_metadata batch.  (If
+    //   we've already started an internal recv_trailing_metadata batch,
+    //   then we need the logic below to send all ops in the batch
+    //   *except* the recv_trailing_metadata op.)
+    if (calld_->retry_committed_ && !pending->send_ops_cached &&
+        (!batch->recv_trailing_metadata || !started_recv_trailing_metadata_)) {
+      AddClosureForBatch(
+          batch,
+          "start non-replayable pending batch on call attempt after commit",
+          closures);
+      calld_->PendingBatchClear(pending);
+      continue;
+    }
+    // Create batch with the right number of callbacks.
+    BatchData* batch_data =
+        CreateBatch(num_callbacks, has_send_ops /* set_on_complete */);
+    // Cache send ops if needed.
+    calld_->MaybeCacheSendOpsForBatch(pending);
+    // send_initial_metadata.
+    if (batch->send_initial_metadata) {
+      batch_data->AddRetriableSendInitialMetadataOp();
+    }
+    // send_message.
+    if (batch->send_message) {
+      batch_data->AddRetriableSendMessageOp();
+    }
+    // send_trailing_metadata.
+    if (batch->send_trailing_metadata) {
+      batch_data->AddRetriableSendTrailingMetadataOp();
+    }
+    // recv_initial_metadata.
+    if (batch->recv_initial_metadata) {
+      // recv_flags is only used on the server side.
+      GPR_ASSERT(batch->payload->recv_initial_metadata.recv_flags == nullptr);
+      batch_data->AddRetriableRecvInitialMetadataOp();
+    }
+    // recv_message.
+    if (batch->recv_message) {
+      batch_data->AddRetriableRecvMessageOp();
+    }
+    // recv_trailing_metadata.
+    if (batch->recv_trailing_metadata && !started_recv_trailing_metadata_) {
+      batch_data->AddRetriableRecvTrailingMetadataOp();
+    }
+    AddClosureForBatch(batch_data->batch(),
+                       "start replayable pending batch on call attempt",
+                       closures);
+  }
+}
+
+void RetryFilter::CallData::CallAttempt::AddRetriableBatches(
+    CallCombinerClosureList* closures) {
+  // Replay previously-returned send_* ops if needed.
+  BatchData* replay_batch_data = MaybeCreateBatchForReplay();
+  if (replay_batch_data != nullptr) {
+    AddClosureForBatch(replay_batch_data->batch(),
+                       "start replay batch on call attempt", closures);
+  }
+  // Now add pending batches.
+  AddBatchesForPendingBatches(closures);
+}
+
+void RetryFilter::CallData::CallAttempt::StartRetriableBatches() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p: constructing retriable batches",
+            calld_->chand_, calld_, this);
+  }
+  // Construct list of closures to execute, one for each pending batch.
+  CallCombinerClosureList closures;
+  AddRetriableBatches(&closures);
+  // Note: This will yield the call combiner.
+  // Start batches on LB call.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p: starting %" PRIuPTR
+            " retriable batches on lb_call=%p",
+            calld_->chand_, calld_, this, closures.size(), lb_call_.get());
+  }
+  closures.RunClosures(calld_->call_combiner_);
+}
+
+void RetryFilter::CallData::CallAttempt::CancelFromSurface(
+    grpc_transport_stream_op_batch* cancel_batch) {
+  MaybeCancelPerAttemptRecvTimer();
+  // Propagate cancellation to LB call.
+  lb_call_->StartTransportStreamOpBatch(cancel_batch);
+}
+
+bool RetryFilter::CallData::CallAttempt::ShouldRetry(
+    absl::optional<grpc_status_code> status, bool is_lb_drop,
+    grpc_mdelem* server_pushback_md, grpc_millis* server_pushback_ms) {
+  // LB drops always inhibit retries.
+  if (is_lb_drop) return false;
+  // TODO(roth): Handle transparent retries here.
+  // If no retry policy, don't retry.
+  if (calld_->retry_policy_ == nullptr) return false;
+  // Check status.
+  if (status.has_value()) {
+    if (GPR_LIKELY(*status == GRPC_STATUS_OK)) {
+      if (calld_->retry_throttle_data_ != nullptr) {
+        calld_->retry_throttle_data_->RecordSuccess();
+      }
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: call succeeded",
+                calld_->chand_, calld_, this);
+      }
+      return false;
+    }
+    // Status is not OK.  Check whether the status is retryable.
+    if (!calld_->retry_policy_->retryable_status_codes().Contains(*status)) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO,
+                "chand=%p calld=%p attempt=%p: status %s not configured as "
+                "retryable",
+                calld_->chand_, calld_, this,
+                grpc_status_code_to_string(*status));
+      }
+      return false;
+    }
+  }
+  // Record the failure and check whether retries are throttled.
+  // Note that it's important for this check to come after the status
+  // code check above, since we should only record failures whose statuses
+  // match the configured retryable status codes, so that we don't count
+  // things like failures due to malformed requests (INVALID_ARGUMENT).
+  // Conversely, it's important for this to come before the remaining
+  // checks, so that we don't fail to record failures due to other factors.
+  if (calld_->retry_throttle_data_ != nullptr &&
+      !calld_->retry_throttle_data_->RecordFailure()) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: retries throttled",
+              calld_->chand_, calld_, this);
+    }
+    return false;
+  }
+  // Check whether the call is committed.
+  if (calld_->retry_committed_) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: retries already committed",
+              calld_->chand_, calld_, this);
+    }
+    return false;
+  }
+  // Check whether we have retries remaining.
+  ++calld_->num_attempts_completed_;
+  if (calld_->num_attempts_completed_ >=
+      calld_->retry_policy_->max_attempts()) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(
+          GPR_INFO, "chand=%p calld=%p attempt=%p: exceeded %d retry attempts",
+          calld_->chand_, calld_, this, calld_->retry_policy_->max_attempts());
+    }
+    return false;
+  }
+  // Check server push-back.
+  if (server_pushback_md != nullptr) {
+    // If the value is "-1" or any other unparseable string, we do not retry.
+    uint32_t ms;
+    if (!grpc_parse_slice_to_uint32(GRPC_MDVALUE(*server_pushback_md), &ms)) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO,
+                "chand=%p calld=%p attempt=%p: not retrying due to server "
+                "push-back",
+                calld_->chand_, calld_, this);
+      }
+      return false;
+    } else {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(
+            GPR_INFO,
+            "chand=%p calld=%p attempt=%p: server push-back: retry in %u ms",
+            calld_->chand_, calld_, this, ms);
+      }
+      *server_pushback_ms = static_cast<grpc_millis>(ms);
+    }
+  }
+  // Check with call dispatch controller.
+  auto* service_config_call_data =
+      static_cast<ClientChannelServiceConfigCallData*>(
+          calld_->call_context_[GRPC_CONTEXT_SERVICE_CONFIG_CALL_DATA].value);
+  if (!service_config_call_data->call_dispatch_controller()->ShouldRetry()) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(
+          GPR_INFO,
+          "chand=%p calld=%p attempt=%p: call dispatch controller denied retry",
+          calld_->chand_, calld_, this);
+    }
+    return false;
+  }
+  // We should retry.
+  return true;
+}
+
+void RetryFilter::CallData::CallAttempt::Abandon() {
+  abandoned_ = true;
+  // Unref batches for deferred completion callbacks that will now never
+  // be invoked.
+  if (started_recv_trailing_metadata_ &&
+      !seen_recv_trailing_metadata_from_surface_) {
+    recv_trailing_metadata_internal_batch_.reset(
+        DEBUG_LOCATION,
+        "internal recv_trailing_metadata completed before that op was "
+        "started from the surface");
+  }
+  GRPC_ERROR_UNREF(recv_trailing_metadata_error_);
+  recv_trailing_metadata_error_ = GRPC_ERROR_NONE;
+  recv_initial_metadata_ready_deferred_batch_.reset(
+      DEBUG_LOCATION,
+      "unref deferred recv_initial_metadata_ready batch due to retry");
+  GRPC_ERROR_UNREF(recv_initial_metadata_error_);
+  recv_initial_metadata_error_ = GRPC_ERROR_NONE;
+  recv_message_ready_deferred_batch_.reset(
+      DEBUG_LOCATION, "unref deferred recv_message_ready batch due to retry");
+  GRPC_ERROR_UNREF(recv_message_error_);
+  recv_message_error_ = GRPC_ERROR_NONE;
+  for (auto& on_complete_deferred_batch : on_complete_deferred_batches_) {
+    on_complete_deferred_batch.batch.reset(
+        DEBUG_LOCATION, "unref deferred on_complete batch due to retry");
+    GRPC_ERROR_UNREF(on_complete_deferred_batch.error);
+  }
+  on_complete_deferred_batches_.clear();
+}
+
+void RetryFilter::CallData::CallAttempt::OnPerAttemptRecvTimer(
+    void* arg, grpc_error_handle error) {
+  auto* call_attempt = static_cast<CallAttempt*>(arg);
+  GRPC_CLOSURE_INIT(&call_attempt->on_per_attempt_recv_timer_,
+                    OnPerAttemptRecvTimerLocked, call_attempt, nullptr);
+  GRPC_CALL_COMBINER_START(call_attempt->calld_->call_combiner_,
+                           &call_attempt->on_per_attempt_recv_timer_,
+                           GRPC_ERROR_REF(error), "per-attempt timer fired");
+}
+
+void RetryFilter::CallData::CallAttempt::OnPerAttemptRecvTimerLocked(
+    void* arg, grpc_error_handle error) {
+  auto* call_attempt = static_cast<CallAttempt*>(arg);
+  auto* calld = call_attempt->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p: perAttemptRecvTimeout timer fired: "
+            "error=%s, per_attempt_recv_timer_pending_=%d",
+            calld->chand_, calld, call_attempt,
+            grpc_error_std_string(error).c_str(),
+            call_attempt->per_attempt_recv_timer_pending_);
+  }
+  CallCombinerClosureList closures;
+  if (error == GRPC_ERROR_NONE &&
+      call_attempt->per_attempt_recv_timer_pending_) {
+    call_attempt->per_attempt_recv_timer_pending_ = false;
+    // Cancel this attempt.
+    // TODO(roth): When implementing hedging, we should not cancel the
+    // current attempt.
+    call_attempt->AddBatchForCancelOp(
+        grpc_error_set_int(GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+                               "retry perAttemptRecvTimeout exceeded"),
+                           GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_CANCELLED),
+        &closures);
+    // Check whether we should retry.
+    if (call_attempt->ShouldRetry(
+            /*status=*/absl::nullopt, /*is_lb_drop=*/false,
+            /*server_pushback_md=*/nullptr, /*server_pushback_ms=*/nullptr)) {
+      // Mark current attempt as abandoned.
+      call_attempt->Abandon();
+      // We are retrying.  Start backoff timer.
+      calld->StartRetryTimer(/*server_pushback_ms=*/-1);
+    } else {
+      // Not retrying, so commit the call.
+      calld->RetryCommit(call_attempt);
+      // If retry state is no longer needed, switch to fast path for
+      // subsequent batches.
+      call_attempt->MaybeSwitchToFastPath();
+    }
+  }
+  closures.RunClosures(calld->call_combiner_);
+  call_attempt->Unref(DEBUG_LOCATION, "OnPerAttemptRecvTimer");
+  GRPC_CALL_STACK_UNREF(calld->owning_call_, "OnPerAttemptRecvTimer");
+}
+
+void RetryFilter::CallData::CallAttempt::MaybeCancelPerAttemptRecvTimer() {
+  if (per_attempt_recv_timer_pending_) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: cancelling "
+              "perAttemptRecvTimeout timer",
+              calld_->chand_, calld_, this);
+    }
+    per_attempt_recv_timer_pending_ = false;
+    grpc_timer_cancel(&per_attempt_recv_timer_);
+  }
+}
+
+//
+// RetryFilter::CallData::CallAttempt::BatchData
+//
+
+RetryFilter::CallData::CallAttempt::BatchData::BatchData(
+    RefCountedPtr<CallAttempt> attempt, int refcount, bool set_on_complete)
+    : RefCounted(
+          GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace) ? "BatchData" : nullptr,
+          refcount),
+      call_attempt_(std::move(attempt)) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: creating batch %p",
+            call_attempt_->calld_->chand_, call_attempt_->calld_,
+            call_attempt_.get(), this);
+  }
+  // We hold a ref to the call stack for every batch sent on a call attempt.
+  // This is because some batches on the call attempt may not complete
+  // until after all of the batches are completed at the surface (because
+  // each batch that is pending at the surface holds a ref).  This
+  // can happen for replayed send ops, and it can happen for
+  // recv_initial_metadata and recv_message ops on a call attempt that has
+  // been abandoned.
+  GRPC_CALL_STACK_REF(call_attempt_->calld_->owning_call_, "Retry BatchData");
+  batch_.payload = &call_attempt_->batch_payload_;
+  if (set_on_complete) {
+    GRPC_CLOSURE_INIT(&on_complete_, OnComplete, this, nullptr);
+    batch_.on_complete = &on_complete_;
+  }
+}
+
+RetryFilter::CallData::CallAttempt::BatchData::~BatchData() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: destroying batch %p",
+            call_attempt_->calld_->chand_, call_attempt_->calld_,
+            call_attempt_.get(), this);
+  }
+  GRPC_CALL_STACK_UNREF(call_attempt_->calld_->owning_call_, "Retry BatchData");
+  call_attempt_.reset(DEBUG_LOCATION, "~BatchData");
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    FreeCachedSendOpDataForCompletedBatch() {
+  auto* calld = call_attempt_->calld_;
+  // TODO(roth): When we implement hedging, this logic will need to get
+  // a bit more complex, because there may be other (now abandoned) call
+  // attempts still using this data.  We may need to do some sort of
+  // ref-counting instead.
+  if (batch_.send_initial_metadata) {
+    calld->FreeCachedSendInitialMetadata();
+  }
+  if (batch_.send_message) {
+    calld->FreeCachedSendMessage(call_attempt_->completed_send_message_count_ -
+                                 1);
+  }
+  if (batch_.send_trailing_metadata) {
+    calld->FreeCachedSendTrailingMetadata();
+  }
+}
+
+//
+// recv_initial_metadata callback handling
+//
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    MaybeAddClosureForRecvInitialMetadataCallback(
+        grpc_error_handle error, CallCombinerClosureList* closures) {
+  // Find pending batch.
+  PendingBatch* pending = call_attempt_->calld_->PendingBatchFind(
+      "invoking recv_initial_metadata_ready for",
+      [](grpc_transport_stream_op_batch* batch) {
+        return batch->recv_initial_metadata &&
+               batch->payload->recv_initial_metadata
+                       .recv_initial_metadata_ready != nullptr;
+      });
+  if (pending == nullptr) {
+    GRPC_ERROR_UNREF(error);
+    return;
+  }
+  // Return metadata.
+  *pending->batch->payload->recv_initial_metadata.recv_initial_metadata =
+      std::move(call_attempt_->recv_initial_metadata_);
+  // Propagate trailing_metadata_available.
+  *pending->batch->payload->recv_initial_metadata.trailing_metadata_available =
+      call_attempt_->trailing_metadata_available_;
+  // Update bookkeeping.
+  // Note: Need to do this before invoking the callback, since invoking
+  // the callback will result in yielding the call combiner.
+  grpc_closure* recv_initial_metadata_ready =
+      pending->batch->payload->recv_initial_metadata
+          .recv_initial_metadata_ready;
+  pending->batch->payload->recv_initial_metadata.recv_initial_metadata_ready =
+      nullptr;
+  call_attempt_->calld_->MaybeClearPendingBatch(pending);
+  // Add callback to closures.
+  closures->Add(recv_initial_metadata_ready, error,
+                "recv_initial_metadata_ready for pending batch");
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::RecvInitialMetadataReady(
+    void* arg, grpc_error_handle error) {
+  RefCountedPtr<BatchData> batch_data(static_cast<BatchData*>(arg));
+  CallAttempt* call_attempt = batch_data->call_attempt_.get();
+  CallData* calld = call_attempt->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p batch_data=%p: "
+            "got recv_initial_metadata_ready, error=%s",
+            calld->chand_, calld, call_attempt, batch_data.get(),
+            grpc_error_std_string(error).c_str());
+  }
+  call_attempt->completed_recv_initial_metadata_ = true;
+  // If this attempt has been abandoned, then we're not going to use the
+  // result of this recv_initial_metadata op, so do nothing.
+  if (call_attempt->abandoned_) {
+    GRPC_CALL_COMBINER_STOP(
+        calld->call_combiner_,
+        "recv_initial_metadata_ready for abandoned attempt");
+    return;
+  }
+  // Cancel per-attempt recv timer, if any.
+  call_attempt->MaybeCancelPerAttemptRecvTimer();
+  // If we're not committed, check the response to see if we need to commit.
+  if (!calld->retry_committed_) {
+    // If we got an error or a Trailers-Only response and have not yet gotten
+    // the recv_trailing_metadata_ready callback, then defer propagating this
+    // callback back to the surface.  We can evaluate whether to retry when
+    // recv_trailing_metadata comes back.
+    if (GPR_UNLIKELY((call_attempt->trailing_metadata_available_ ||
+                      error != GRPC_ERROR_NONE) &&
+                     !call_attempt->completed_recv_trailing_metadata_)) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO,
+                "chand=%p calld=%p attempt=%p: deferring "
+                "recv_initial_metadata_ready (Trailers-Only)",
+                calld->chand_, calld, call_attempt);
+      }
+      call_attempt->recv_initial_metadata_ready_deferred_batch_ =
+          std::move(batch_data);
+      call_attempt->recv_initial_metadata_error_ = GRPC_ERROR_REF(error);
+      CallCombinerClosureList closures;
+      if (error != GRPC_ERROR_NONE) {
+        call_attempt->AddBatchForCancelOp(GRPC_ERROR_REF(error), &closures);
+      }
+      if (!call_attempt->started_recv_trailing_metadata_) {
+        // recv_trailing_metadata not yet started by application; start it
+        // ourselves to get status.
+        call_attempt->AddBatchForInternalRecvTrailingMetadata(&closures);
+      }
+      closures.RunClosures(calld->call_combiner_);
+      return;
+    }
+    // Received valid initial metadata, so commit the call.
+    calld->RetryCommit(call_attempt);
+    // If retry state is no longer needed, switch to fast path for
+    // subsequent batches.
+    call_attempt->MaybeSwitchToFastPath();
+  }
+  // Invoke the callback to return the result to the surface.
+  CallCombinerClosureList closures;
+  batch_data->MaybeAddClosureForRecvInitialMetadataCallback(
+      GRPC_ERROR_REF(error), &closures);
+  closures.RunClosures(calld->call_combiner_);
+}
+
+//
+// recv_message callback handling
+//
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    MaybeAddClosureForRecvMessageCallback(grpc_error_handle error,
+                                          CallCombinerClosureList* closures) {
+  // Find pending op.
+  PendingBatch* pending = call_attempt_->calld_->PendingBatchFind(
+      "invoking recv_message_ready for",
+      [](grpc_transport_stream_op_batch* batch) {
+        return batch->recv_message &&
+               batch->payload->recv_message.recv_message_ready != nullptr;
+      });
+  if (pending == nullptr) {
+    GRPC_ERROR_UNREF(error);
+    return;
+  }
+  // Return payload.
+  *pending->batch->payload->recv_message.recv_message =
+      std::move(call_attempt_->recv_message_);
+  // Update bookkeeping.
+  // Note: Need to do this before invoking the callback, since invoking
+  // the callback will result in yielding the call combiner.
+  grpc_closure* recv_message_ready =
+      pending->batch->payload->recv_message.recv_message_ready;
+  pending->batch->payload->recv_message.recv_message_ready = nullptr;
+  call_attempt_->calld_->MaybeClearPendingBatch(pending);
+  // Add callback to closures.
+  closures->Add(recv_message_ready, error,
+                "recv_message_ready for pending batch");
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::RecvMessageReady(
+    void* arg, grpc_error_handle error) {
+  RefCountedPtr<BatchData> batch_data(static_cast<BatchData*>(arg));
+  CallAttempt* call_attempt = batch_data->call_attempt_.get();
+  CallData* calld = call_attempt->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p batch_data=%p: "
+            "got recv_message_ready, error=%s",
+            calld->chand_, calld, call_attempt, batch_data.get(),
+            grpc_error_std_string(error).c_str());
+  }
+  ++call_attempt->completed_recv_message_count_;
+  // If this attempt has been abandoned, then we're not going to use the
+  // result of this recv_message op, so do nothing.
+  if (call_attempt->abandoned_) {
+    GRPC_CALL_COMBINER_STOP(calld->call_combiner_,
+                            "recv_message_ready for abandoned attempt");
+    return;
+  }
+  // Cancel per-attempt recv timer, if any.
+  call_attempt->MaybeCancelPerAttemptRecvTimer();
+  // If we're not committed, check the response to see if we need to commit.
+  if (!calld->retry_committed_) {
+    // If we got an error or the payload was nullptr and we have not yet gotten
+    // the recv_trailing_metadata_ready callback, then defer propagating this
+    // callback back to the surface.  We can evaluate whether to retry when
+    // recv_trailing_metadata comes back.
+    if (GPR_UNLIKELY((call_attempt->recv_message_ == nullptr ||
+                      error != GRPC_ERROR_NONE) &&
+                     !call_attempt->completed_recv_trailing_metadata_)) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO,
+                "chand=%p calld=%p attempt=%p: deferring recv_message_ready "
+                "(nullptr message and recv_trailing_metadata pending)",
+                calld->chand_, calld, call_attempt);
+      }
+      call_attempt->recv_message_ready_deferred_batch_ = std::move(batch_data);
+      call_attempt->recv_message_error_ = GRPC_ERROR_REF(error);
+      CallCombinerClosureList closures;
+      if (error != GRPC_ERROR_NONE) {
+        call_attempt->AddBatchForCancelOp(GRPC_ERROR_REF(error), &closures);
+      }
+      if (!call_attempt->started_recv_trailing_metadata_) {
+        // recv_trailing_metadata not yet started by application; start it
+        // ourselves to get status.
+        call_attempt->AddBatchForInternalRecvTrailingMetadata(&closures);
+      }
+      closures.RunClosures(calld->call_combiner_);
+      return;
+    }
+    // Received a valid message, so commit the call.
+    calld->RetryCommit(call_attempt);
+    // If retry state is no longer needed, switch to fast path for
+    // subsequent batches.
+    call_attempt->MaybeSwitchToFastPath();
+  }
+  // Invoke the callback to return the result to the surface.
+  CallCombinerClosureList closures;
+  batch_data->MaybeAddClosureForRecvMessageCallback(GRPC_ERROR_REF(error),
+                                                    &closures);
+  closures.RunClosures(calld->call_combiner_);
+}
+
+//
+// recv_trailing_metadata handling
+//
+
+namespace {
+
+// Sets *status, *server_pushback_md, and *is_lb_drop based on md_batch
+// and error.
+void GetCallStatus(grpc_millis deadline, grpc_metadata_batch* md_batch,
+                   grpc_error_handle error, grpc_status_code* status,
+                   grpc_mdelem** server_pushback_md, bool* is_lb_drop) {
+  if (error != GRPC_ERROR_NONE) {
+    grpc_error_get_status(error, deadline, status, nullptr, nullptr, nullptr);
+    intptr_t value = 0;
+    if (grpc_error_get_int(error, GRPC_ERROR_INT_LB_POLICY_DROP, &value) &&
+        value != 0) {
+      *is_lb_drop = true;
+    }
+  } else {
+    GPR_ASSERT(md_batch->legacy_index()->named.grpc_status != nullptr);
+    *status = grpc_get_status_code_from_metadata(
+        md_batch->legacy_index()->named.grpc_status->md);
+    if (md_batch->legacy_index()->named.grpc_retry_pushback_ms != nullptr) {
+      *server_pushback_md =
+          &md_batch->legacy_index()->named.grpc_retry_pushback_ms->md;
+    }
+  }
+  GRPC_ERROR_UNREF(error);
+}
+
+}  // namespace
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    MaybeAddClosureForRecvTrailingMetadataReady(
+        grpc_error_handle error, CallCombinerClosureList* closures) {
+  auto* calld = call_attempt_->calld_;
+  // Find pending batch.
+  PendingBatch* pending = calld->PendingBatchFind(
+      "invoking recv_trailing_metadata_ready for",
+      [](grpc_transport_stream_op_batch* batch) {
+        return batch->recv_trailing_metadata &&
+               batch->payload->recv_trailing_metadata
+                       .recv_trailing_metadata_ready != nullptr;
+      });
+  // If we generated the recv_trailing_metadata op internally via
+  // AddBatchForInternalRecvTrailingMetadata(), then there will be no
+  // pending batch.
+  if (pending == nullptr) {
+    call_attempt_->recv_trailing_metadata_error_ = error;
+    return;
+  }
+  // Copy transport stats to be delivered up to the surface.
+  grpc_transport_move_stats(
+      &call_attempt_->collect_stats_,
+      pending->batch->payload->recv_trailing_metadata.collect_stats);
+  // Return metadata.
+  *pending->batch->payload->recv_trailing_metadata.recv_trailing_metadata =
+      std::move(call_attempt_->recv_trailing_metadata_);
+  // Add closure.
+  closures->Add(pending->batch->payload->recv_trailing_metadata
+                    .recv_trailing_metadata_ready,
+                error, "recv_trailing_metadata_ready for pending batch");
+  // Update bookkeeping.
+  pending->batch->payload->recv_trailing_metadata.recv_trailing_metadata_ready =
+      nullptr;
+  calld->MaybeClearPendingBatch(pending);
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddClosuresForDeferredCompletionCallbacks(
+        CallCombinerClosureList* closures) {
+  // Add closure for deferred recv_initial_metadata_ready.
+  if (GPR_UNLIKELY(call_attempt_->recv_initial_metadata_ready_deferred_batch_ !=
+                   nullptr)) {
+    MaybeAddClosureForRecvInitialMetadataCallback(
+        call_attempt_->recv_initial_metadata_error_, closures);
+    call_attempt_->recv_initial_metadata_ready_deferred_batch_.reset(
+        DEBUG_LOCATION, "resuming deferred recv_initial_metadata_ready");
+    call_attempt_->recv_initial_metadata_error_ = GRPC_ERROR_NONE;
+  }
+  // Add closure for deferred recv_message_ready.
+  if (GPR_UNLIKELY(call_attempt_->recv_message_ready_deferred_batch_ !=
+                   nullptr)) {
+    MaybeAddClosureForRecvMessageCallback(call_attempt_->recv_message_error_,
+                                          closures);
+    call_attempt_->recv_message_ready_deferred_batch_.reset(
+        DEBUG_LOCATION, "resuming deferred recv_message_ready");
+    call_attempt_->recv_message_error_ = GRPC_ERROR_NONE;
+  }
+  // Add closures for deferred on_complete callbacks.
+  for (auto& on_complete_deferred_batch :
+       call_attempt_->on_complete_deferred_batches_) {
+    closures->Add(&on_complete_deferred_batch.batch->on_complete_,
+                  on_complete_deferred_batch.error, "resuming on_complete");
+    on_complete_deferred_batch.batch.release();
+  }
+  call_attempt_->on_complete_deferred_batches_.clear();
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddClosuresToFailUnstartedPendingBatches(
+        grpc_error_handle error, CallCombinerClosureList* closures) {
+  auto* calld = call_attempt_->calld_;
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(calld->pending_batches_); ++i) {
+    PendingBatch* pending = &calld->pending_batches_[i];
+    if (pending->batch == nullptr) continue;
+    if (call_attempt_->PendingBatchContainsUnstartedSendOps(pending)) {
+      closures->Add(pending->batch->on_complete, GRPC_ERROR_REF(error),
+                    "failing on_complete for pending batch");
+      pending->batch->on_complete = nullptr;
+      calld->MaybeClearPendingBatch(pending);
+    }
+  }
+  GRPC_ERROR_UNREF(error);
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::RunClosuresForCompletedCall(
+    grpc_error_handle error) {
+  // Construct list of closures to execute.
+  CallCombinerClosureList closures;
+  // First, add closure for recv_trailing_metadata_ready.
+  MaybeAddClosureForRecvTrailingMetadataReady(GRPC_ERROR_REF(error), &closures);
+  // If there are deferred batch completion callbacks, add them to closures.
+  AddClosuresForDeferredCompletionCallbacks(&closures);
+  // Add closures to fail any pending batches that have not yet been started.
+  AddClosuresToFailUnstartedPendingBatches(GRPC_ERROR_REF(error), &closures);
+  // Schedule all of the closures identified above.
+  // Note: This will release the call combiner.
+  closures.RunClosures(call_attempt_->calld_->call_combiner_);
+  GRPC_ERROR_UNREF(error);
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::RecvTrailingMetadataReady(
+    void* arg, grpc_error_handle error) {
+  RefCountedPtr<BatchData> batch_data(static_cast<BatchData*>(arg));
+  CallAttempt* call_attempt = batch_data->call_attempt_.get();
+  CallData* calld = call_attempt->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p batch_data=%p: "
+            "got recv_trailing_metadata_ready, error=%s",
+            calld->chand_, calld, call_attempt, batch_data.get(),
+            grpc_error_std_string(error).c_str());
+  }
+  call_attempt->completed_recv_trailing_metadata_ = true;
+  // If this attempt has been abandoned, then we're not going to use the
+  // result of this recv_trailing_metadata op, so do nothing.
+  if (call_attempt->abandoned_) {
+    GRPC_CALL_COMBINER_STOP(
+        calld->call_combiner_,
+        "recv_trailing_metadata_ready for abandoned attempt");
+    return;
+  }
+  // Cancel per-attempt recv timer, if any.
+  call_attempt->MaybeCancelPerAttemptRecvTimer();
+  // Get the call's status and check for server pushback metadata.
+  grpc_status_code status = GRPC_STATUS_OK;
+  grpc_mdelem* server_pushback_md = nullptr;
+  grpc_metadata_batch* md_batch =
+      batch_data->batch_.payload->recv_trailing_metadata.recv_trailing_metadata;
+  bool is_lb_drop = false;
+  GetCallStatus(calld->deadline_, md_batch, GRPC_ERROR_REF(error), &status,
+                &server_pushback_md, &is_lb_drop);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(
+        GPR_INFO,
+        "chand=%p calld=%p attempt=%p: call finished, status=%s is_lb_drop=%d",
+        calld->chand_, calld, call_attempt, grpc_status_code_to_string(status),
+        is_lb_drop);
+  }
+  // Check if we should retry.
+  grpc_millis server_pushback_ms = -1;
+  if (call_attempt->ShouldRetry(status, is_lb_drop, server_pushback_md,
+                                &server_pushback_ms)) {
+    // Start retry timer.
+    calld->StartRetryTimer(server_pushback_ms);
+    // Cancel call attempt.
+    CallCombinerClosureList closures;
+    call_attempt->AddBatchForCancelOp(
+        error == GRPC_ERROR_NONE
+            ? grpc_error_set_int(
+                  GRPC_ERROR_CREATE_FROM_STATIC_STRING("call attempt failed"),
+                  GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_CANCELLED)
+            : GRPC_ERROR_REF(error),
+        &closures);
+    // Record that this attempt has been abandoned.
+    call_attempt->Abandon();
+    // Yields call combiner.
+    closures.RunClosures(calld->call_combiner_);
+    return;
+  }
+  // Not retrying, so commit the call.
+  calld->RetryCommit(call_attempt);
+  // If retry state is no longer needed, switch to fast path for
+  // subsequent batches.
+  call_attempt->MaybeSwitchToFastPath();
+  // Run any necessary closures.
+  batch_data->RunClosuresForCompletedCall(GRPC_ERROR_REF(error));
+}
+
+//
+// on_complete callback handling
+//
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddClosuresForCompletedPendingBatch(grpc_error_handle error,
+                                        CallCombinerClosureList* closures) {
+  auto* calld = call_attempt_->calld_;
+  PendingBatch* pending = calld->PendingBatchFind(
+      "completed", [this](grpc_transport_stream_op_batch* batch) {
+        // Match the pending batch with the same set of send ops as the
+        // batch we've just completed.
+        return batch->on_complete != nullptr &&
+               batch_.send_initial_metadata == batch->send_initial_metadata &&
+               batch_.send_message == batch->send_message &&
+               batch_.send_trailing_metadata == batch->send_trailing_metadata;
+      });
+  // If batch_data is a replay batch, then there will be no pending
+  // batch to complete.
+  if (pending == nullptr) {
+    GRPC_ERROR_UNREF(error);
+    return;
+  }
+  // Propagate payload.
+  if (batch_.send_message) {
+    pending->batch->payload->send_message.stream_write_closed =
+        batch_.payload->send_message.stream_write_closed;
+  }
+  // Add closure.
+  closures->Add(pending->batch->on_complete, error,
+                "on_complete for pending batch");
+  pending->batch->on_complete = nullptr;
+  calld->MaybeClearPendingBatch(pending);
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddClosuresForReplayOrPendingSendOps(CallCombinerClosureList* closures) {
+  auto* calld = call_attempt_->calld_;
+  bool have_pending_send_ops = call_attempt_->HaveSendOpsToReplay();
+  // We don't check send_initial_metadata here, because that op will always
+  // be started as soon as it is received from the surface, so it will
+  // never need to be started at this point.
+  if (!have_pending_send_ops) {
+    for (size_t i = 0; i < GPR_ARRAY_SIZE(calld->pending_batches_); ++i) {
+      PendingBatch* pending = &calld->pending_batches_[i];
+      grpc_transport_stream_op_batch* batch = pending->batch;
+      if (batch == nullptr || pending->send_ops_cached) continue;
+      if (batch->send_message || batch->send_trailing_metadata) {
+        have_pending_send_ops = true;
+        break;
+      }
+    }
+  }
+  if (have_pending_send_ops) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p attempt=%p: starting next batch for pending "
+              "send op(s)",
+              calld->chand_, calld, call_attempt_.get());
+    }
+    call_attempt_->AddRetriableBatches(closures);
+  }
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::OnComplete(
+    void* arg, grpc_error_handle error) {
+  RefCountedPtr<BatchData> batch_data(static_cast<BatchData*>(arg));
+  CallAttempt* call_attempt = batch_data->call_attempt_.get();
+  CallData* calld = call_attempt->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p batch_data=%p: "
+            "got on_complete, error=%s, batch=%s",
+            calld->chand_, calld, call_attempt, batch_data.get(),
+            grpc_error_std_string(error).c_str(),
+            grpc_transport_stream_op_batch_string(&batch_data->batch_).c_str());
+  }
+  // If this attempt has been abandoned, then we're not going to propagate
+  // the completion of this batch, so do nothing.
+  if (call_attempt->abandoned_) {
+    GRPC_CALL_COMBINER_STOP(calld->call_combiner_,
+                            "on_complete for abandoned attempt");
+    return;
+  }
+  // If we got an error and have not yet gotten the
+  // recv_trailing_metadata_ready callback, then defer propagating this
+  // callback back to the surface.  We can evaluate whether to retry when
+  // recv_trailing_metadata comes back.
+  if (GPR_UNLIKELY(!calld->retry_committed_ && error != GRPC_ERROR_NONE &&
+                   !call_attempt->completed_recv_trailing_metadata_)) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO, "chand=%p calld=%p attempt=%p: deferring on_complete",
+              calld->chand_, calld, call_attempt);
+    }
+    call_attempt->on_complete_deferred_batches_.emplace_back(
+        std::move(batch_data), GRPC_ERROR_REF(error));
+    CallCombinerClosureList closures;
+    call_attempt->AddBatchForCancelOp(GRPC_ERROR_REF(error), &closures);
+    if (!call_attempt->started_recv_trailing_metadata_) {
+      // recv_trailing_metadata not yet started by application; start it
+      // ourselves to get status.
+      call_attempt->AddBatchForInternalRecvTrailingMetadata(&closures);
+    }
+    closures.RunClosures(calld->call_combiner_);
+    return;
+  }
+  // Update bookkeeping in call_attempt.
+  if (batch_data->batch_.send_initial_metadata) {
+    call_attempt->completed_send_initial_metadata_ = true;
+  }
+  if (batch_data->batch_.send_message) {
+    ++call_attempt->completed_send_message_count_;
+  }
+  if (batch_data->batch_.send_trailing_metadata) {
+    call_attempt->completed_send_trailing_metadata_ = true;
+  }
+  // If the call is committed, free cached data for send ops that we've just
+  // completed.
+  if (calld->retry_committed_) {
+    batch_data->FreeCachedSendOpDataForCompletedBatch();
+  }
+  // Construct list of closures to execute.
+  CallCombinerClosureList closures;
+  // Add closure for the completed pending batch, if any.
+  batch_data->AddClosuresForCompletedPendingBatch(GRPC_ERROR_REF(error),
+                                                  &closures);
+  // If needed, add a callback to start any replay or pending send ops on
+  // the LB call.
+  if (!call_attempt->completed_recv_trailing_metadata_) {
+    batch_data->AddClosuresForReplayOrPendingSendOps(&closures);
+  }
+  // If retry state is no longer needed (i.e., we're committed and there
+  // are no more send ops to replay), switch to fast path for subsequent
+  // batches.
+  call_attempt->MaybeSwitchToFastPath();
+  // Schedule all of the closures identified above.
+  // Note: This yields the call combiner.
+  closures.RunClosures(calld->call_combiner_);
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::OnCompleteForCancelOp(
+    void* arg, grpc_error_handle error) {
+  RefCountedPtr<BatchData> batch_data(static_cast<BatchData*>(arg));
+  CallAttempt* call_attempt = batch_data->call_attempt_.get();
+  CallData* calld = call_attempt->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p attempt=%p batch_data=%p: "
+            "got on_complete for cancel_stream batch, error=%s, batch=%s",
+            calld->chand_, calld, call_attempt, batch_data.get(),
+            grpc_error_std_string(error).c_str(),
+            grpc_transport_stream_op_batch_string(&batch_data->batch_).c_str());
+  }
+  GRPC_CALL_COMBINER_STOP(
+      calld->call_combiner_,
+      "on_complete for internally generated cancel_stream op");
+}
+
+//
+// retriable batch construction
+//
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddRetriableSendInitialMetadataOp() {
+  auto* calld = call_attempt_->calld_;
+  // Maps the number of retries to the corresponding metadata value slice.
+  const grpc_slice* retry_count_strings[] = {&GRPC_MDSTR_1, &GRPC_MDSTR_2,
+                                             &GRPC_MDSTR_3, &GRPC_MDSTR_4};
+  // We need to make a copy of the metadata batch for each attempt, since
+  // the filters in the subchannel stack may modify this batch, and we don't
+  // want those modifications to be passed forward to subsequent attempts.
+  //
+  // If we've already completed one or more attempts, add the
+  // grpc-retry-attempts header.
+  grpc_metadata_batch_copy(&calld->send_initial_metadata_,
+                           &call_attempt_->send_initial_metadata_);
+  if (GPR_UNLIKELY(call_attempt_->send_initial_metadata_.legacy_index()
+                       ->named.grpc_previous_rpc_attempts != nullptr)) {
+    call_attempt_->send_initial_metadata_.Remove(
+        GRPC_BATCH_GRPC_PREVIOUS_RPC_ATTEMPTS);
+  }
+  if (GPR_UNLIKELY(calld->num_attempts_completed_ > 0)) {
+    grpc_mdelem retry_md = grpc_mdelem_create(
+        GRPC_MDSTR_GRPC_PREVIOUS_RPC_ATTEMPTS,
+        *retry_count_strings[calld->num_attempts_completed_ - 1], nullptr);
+    grpc_error_handle error = grpc_metadata_batch_add_tail(
+        &call_attempt_->send_initial_metadata_,
+        &call_attempt_->retry_attempts_metadata_, retry_md,
+        GRPC_BATCH_GRPC_PREVIOUS_RPC_ATTEMPTS);
+    if (GPR_UNLIKELY(error != GRPC_ERROR_NONE)) {
+      gpr_log(GPR_ERROR, "error adding retry metadata: %s",
+              grpc_error_std_string(error).c_str());
+      GPR_ASSERT(false);
+    }
+  }
+  call_attempt_->started_send_initial_metadata_ = true;
+  batch_.send_initial_metadata = true;
+  batch_.payload->send_initial_metadata.send_initial_metadata =
+      &call_attempt_->send_initial_metadata_;
+  batch_.payload->send_initial_metadata.send_initial_metadata_flags =
+      calld->send_initial_metadata_flags_;
+  batch_.payload->send_initial_metadata.peer_string = calld->peer_string_;
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddRetriableSendMessageOp() {
+  auto* calld = call_attempt_->calld_;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(
+        GPR_INFO,
+        "chand=%p calld=%p attempt=%p: starting calld->send_messages[%" PRIuPTR
+        "]",
+        calld->chand_, calld, call_attempt_.get(),
+        call_attempt_->started_send_message_count_);
+  }
+  ByteStreamCache* cache =
+      calld->send_messages_[call_attempt_->started_send_message_count_];
+  ++call_attempt_->started_send_message_count_;
+  call_attempt_->send_message_.Init(cache);
+  batch_.send_message = true;
+  batch_.payload->send_message.send_message.reset(
+      call_attempt_->send_message_.get());
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddRetriableSendTrailingMetadataOp() {
+  auto* calld = call_attempt_->calld_;
+  // We need to make a copy of the metadata batch for each attempt, since
+  // the filters in the subchannel stack may modify this batch, and we don't
+  // want those modifications to be passed forward to subsequent attempts.
+  grpc_metadata_batch_copy(&calld->send_trailing_metadata_,
+                           &call_attempt_->send_trailing_metadata_);
+  call_attempt_->started_send_trailing_metadata_ = true;
+  batch_.send_trailing_metadata = true;
+  batch_.payload->send_trailing_metadata.send_trailing_metadata =
+      &call_attempt_->send_trailing_metadata_;
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddRetriableRecvInitialMetadataOp() {
+  call_attempt_->started_recv_initial_metadata_ = true;
+  batch_.recv_initial_metadata = true;
+  call_attempt_->recv_initial_metadata_.Clear();
+  batch_.payload->recv_initial_metadata.recv_initial_metadata =
+      &call_attempt_->recv_initial_metadata_;
+  batch_.payload->recv_initial_metadata.trailing_metadata_available =
+      &call_attempt_->trailing_metadata_available_;
+  GRPC_CLOSURE_INIT(&call_attempt_->recv_initial_metadata_ready_,
+                    RecvInitialMetadataReady, this, grpc_schedule_on_exec_ctx);
+  batch_.payload->recv_initial_metadata.recv_initial_metadata_ready =
+      &call_attempt_->recv_initial_metadata_ready_;
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddRetriableRecvMessageOp() {
+  ++call_attempt_->started_recv_message_count_;
+  batch_.recv_message = true;
+  batch_.payload->recv_message.recv_message = &call_attempt_->recv_message_;
+  batch_.payload->recv_message.call_failed_before_recv_message = nullptr;
+  GRPC_CLOSURE_INIT(&call_attempt_->recv_message_ready_, RecvMessageReady, this,
+                    grpc_schedule_on_exec_ctx);
+  batch_.payload->recv_message.recv_message_ready =
+      &call_attempt_->recv_message_ready_;
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::
+    AddRetriableRecvTrailingMetadataOp() {
+  call_attempt_->started_recv_trailing_metadata_ = true;
+  batch_.recv_trailing_metadata = true;
+  call_attempt_->recv_trailing_metadata_.Clear();
+  batch_.payload->recv_trailing_metadata.recv_trailing_metadata =
+      &call_attempt_->recv_trailing_metadata_;
+  batch_.payload->recv_trailing_metadata.collect_stats =
+      &call_attempt_->collect_stats_;
+  GRPC_CLOSURE_INIT(&call_attempt_->recv_trailing_metadata_ready_,
+                    RecvTrailingMetadataReady, this, grpc_schedule_on_exec_ctx);
+  batch_.payload->recv_trailing_metadata.recv_trailing_metadata_ready =
+      &call_attempt_->recv_trailing_metadata_ready_;
+}
+
+void RetryFilter::CallData::CallAttempt::BatchData::AddCancelStreamOp(
+    grpc_error_handle error) {
+  batch_.cancel_stream = true;
+  batch_.payload->cancel_stream.cancel_error = error;
+  // Override on_complete callback.
+  GRPC_CLOSURE_INIT(&on_complete_, OnCompleteForCancelOp, this, nullptr);
+}
+
+//
+// CallData vtable functions
+//
+
+grpc_error_handle RetryFilter::CallData::Init(
+    grpc_call_element* elem, const grpc_call_element_args* args) {
+  auto* chand = static_cast<RetryFilter*>(elem->channel_data);
+  new (elem->call_data) CallData(chand, *args);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p: created call", chand,
+            elem->call_data);
+  }
+  return GRPC_ERROR_NONE;
+}
+
+void RetryFilter::CallData::Destroy(grpc_call_element* elem,
+                                    const grpc_call_final_info* /*final_info*/,
+                                    grpc_closure* then_schedule_closure) {
+  auto* calld = static_cast<CallData*>(elem->call_data);
+  // Save our ref to the CallStackDestructionBarrier until after our
+  // dtor is invoked.
+  RefCountedPtr<CallStackDestructionBarrier> call_stack_destruction_barrier =
+      std::move(calld->call_stack_destruction_barrier_);
+  calld->~CallData();
+  // Now set the callback in the CallStackDestructionBarrier object,
+  // right before we release our ref to it (implicitly upon returning).
+  // The callback will be invoked when the CallStackDestructionBarrier
+  // is destroyed.
+  call_stack_destruction_barrier->set_on_call_stack_destruction(
+      then_schedule_closure);
+}
+
+void RetryFilter::CallData::StartTransportStreamOpBatch(
+    grpc_call_element* elem, grpc_transport_stream_op_batch* batch) {
+  auto* calld = static_cast<CallData*>(elem->call_data);
+  calld->StartTransportStreamOpBatch(batch);
+}
+
+void RetryFilter::CallData::SetPollent(grpc_call_element* elem,
+                                       grpc_polling_entity* pollent) {
+  auto* calld = static_cast<CallData*>(elem->call_data);
+  calld->pollent_ = pollent;
+}
+
+//
+// CallData implementation
+//
+
+const RetryMethodConfig* GetRetryPolicy(
+    const grpc_call_context_element* context) {
+  if (context == nullptr) return nullptr;
+  auto* svc_cfg_call_data = static_cast<ServiceConfigCallData*>(
+      context[GRPC_CONTEXT_SERVICE_CONFIG_CALL_DATA].value);
+  if (svc_cfg_call_data == nullptr) return nullptr;
+  return static_cast<const RetryMethodConfig*>(
+      svc_cfg_call_data->GetMethodParsedConfig(
+          RetryServiceConfigParser::ParserIndex()));
+}
+
+RetryFilter::CallData::CallData(RetryFilter* chand,
+                                const grpc_call_element_args& args)
+    : chand_(chand),
+      retry_throttle_data_(chand->retry_throttle_data_),
+      retry_policy_(GetRetryPolicy(args.context)),
+      retry_backoff_(
+          BackOff::Options()
+              .set_initial_backoff(retry_policy_ == nullptr
+                                       ? 0
+                                       : retry_policy_->initial_backoff())
+              .set_multiplier(retry_policy_ == nullptr
+                                  ? 0
+                                  : retry_policy_->backoff_multiplier())
+              .set_jitter(RETRY_BACKOFF_JITTER)
+              .set_max_backoff(
+                  retry_policy_ == nullptr ? 0 : retry_policy_->max_backoff())),
+      path_(grpc_slice_ref_internal(args.path)),
+      deadline_(args.deadline),
+      arena_(args.arena),
+      owning_call_(args.call_stack),
+      call_combiner_(args.call_combiner),
+      call_context_(args.context),
+      call_stack_destruction_barrier_(
+          arena_->New<CallStackDestructionBarrier>()),
+      pending_send_initial_metadata_(false),
+      pending_send_message_(false),
+      pending_send_trailing_metadata_(false),
+      retry_committed_(false),
+      retry_timer_pending_(false) {}
+
+RetryFilter::CallData::~CallData() {
+  grpc_slice_unref_internal(path_);
+  // Make sure there are no remaining pending batches.
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(pending_batches_); ++i) {
+    GPR_ASSERT(pending_batches_[i].batch == nullptr);
+  }
+  GRPC_ERROR_UNREF(cancelled_from_surface_);
+}
+
+void RetryFilter::CallData::StartTransportStreamOpBatch(
+    grpc_transport_stream_op_batch* batch) {
+  // If we have an LB call, delegate to the LB call.
+  if (committed_call_ != nullptr) {
+    // Note: This will release the call combiner.
+    committed_call_->StartTransportStreamOpBatch(batch);
+    return;
+  }
+  // Handle cancellation.
+  if (GPR_UNLIKELY(batch->cancel_stream)) {
+    grpc_error_handle cancel_error = batch->payload->cancel_stream.cancel_error;
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO, "chand=%p calld=%p: cancelled from surface: %s", chand_,
+              this, grpc_error_std_string(cancel_error).c_str());
+    }
+    // If we have a current call attempt, commit the call, then send
+    // the cancellation down to that attempt.  When the call fails, it
+    // will not be retried, because we have committed it here.
+    if (call_attempt_ != nullptr) {
+      RetryCommit(call_attempt_.get());
+      // TODO(roth): When implementing hedging, this will get more
+      // complex, because instead of just passing the batch down to a
+      // single call attempt, we'll need to cancel multiple call
+      // attempts and wait for the cancellation on_complete from each call
+      // attempt before we propagate the on_complete from this batch
+      // back to the surface.
+      // Note: This will release the call combiner.
+      call_attempt_->CancelFromSurface(batch);
+      return;
+    }
+    // Save cancel_error in case subsequent batches are started.
+    GRPC_ERROR_UNREF(cancelled_from_surface_);
+    cancelled_from_surface_ = GRPC_ERROR_REF(cancel_error);
+    // Cancel retry timer.
+    if (retry_timer_pending_) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO, "chand=%p calld=%p: cancelling retry timer", chand_,
+                this);
+      }
+      retry_timer_pending_ = false;  // Lame timer callback.
+      grpc_timer_cancel(&retry_timer_);
+      FreeAllCachedSendOpData();
+    }
+    // Fail pending batches.
+    PendingBatchesFail(GRPC_ERROR_REF(cancel_error));
+    // Note: This will release the call combiner.
+    grpc_transport_stream_op_batch_finish_with_failure(
+        batch, GRPC_ERROR_REF(cancel_error), call_combiner_);
+    return;
+  }
+  // Add the batch to the pending list.
+  PendingBatch* pending = PendingBatchesAdd(batch);
+  // If the timer is pending, yield the call combiner and wait for it to
+  // run, since we don't want to start another call attempt until it does.
+  if (retry_timer_pending_) {
+    GRPC_CALL_COMBINER_STOP(call_combiner_,
+                            "added pending batch while retry timer pending");
+    return;
+  }
+  // If we do not yet have a call attempt, create one.
+  if (call_attempt_ == nullptr) {
+    // If we were previously cancelled from the surface, cancel this
+    // batch instead of creating a call attempt.
+    if (cancelled_from_surface_ != GRPC_ERROR_NONE) {
+      PendingBatchClear(pending);
+      // Note: This will release the call combiner.
+      grpc_transport_stream_op_batch_finish_with_failure(
+          batch, GRPC_ERROR_REF(cancelled_from_surface_), call_combiner_);
+      return;
+    }
+    // If there is no retry policy, then commit retries immediately.
+    // This ensures that the code below will always jump to the fast path.
+    // TODO(roth): Remove this special case when we implement
+    // transparent retries.
+    if (retry_policy_ == nullptr) retry_committed_ = true;
+    // If this is the first batch and retries are already committed
+    // (e.g., if this batch put the call above the buffer size limit), then
+    // immediately create an LB call and delegate the batch to it.  This
+    // avoids the overhead of unnecessarily allocating a CallAttempt
+    // object or caching any of the send op data.
+    // Note that we would ideally like to do this also on subsequent
+    // attempts (e.g., if a batch puts the call above the buffer size
+    // limit since the last attempt was complete), but in practice that's
+    // not really worthwhile, because we will almost always have cached and
+    // completed at least the send_initial_metadata op on the previous
+    // attempt, which means that we'd need special logic to replay the
+    // batch anyway, which is exactly what the CallAttempt object provides.
+    // We also skip this optimization if perAttemptRecvTimeout is set in the
+    // retry policy, because we need the code in CallAttempt to handle
+    // the associated timer.
+    if (num_attempts_completed_ == 0 && retry_committed_ &&
+        (retry_policy_ == nullptr ||
+         !retry_policy_->per_attempt_recv_timeout().has_value())) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO,
+                "chand=%p calld=%p: retry committed before first attempt; "
+                "creating LB call",
+                chand_, this);
+      }
+      PendingBatchClear(pending);
+      auto* service_config_call_data =
+          static_cast<ClientChannelServiceConfigCallData*>(
+              call_context_[GRPC_CONTEXT_SERVICE_CONFIG_CALL_DATA].value);
+      committed_call_ = CreateLoadBalancedCall(
+          service_config_call_data->call_dispatch_controller());
+      committed_call_->StartTransportStreamOpBatch(batch);
+      return;
+    }
+    // Otherwise, create a call attempt.
+    // The attempt will automatically start any necessary replays or
+    // pending batches.
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO, "chand=%p calld=%p: creating call attempt", chand_,
+              this);
+    }
+    CreateCallAttempt();
+    return;
+  }
+  // Send batches to call attempt.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p: starting batch on attempt=%p", chand_,
+            this, call_attempt_.get());
+  }
+  call_attempt_->StartRetriableBatches();
+}
+
+OrphanablePtr<ClientChannel::LoadBalancedCall>
+RetryFilter::CallData::CreateLoadBalancedCall(
+    ConfigSelector::CallDispatchController* call_dispatch_controller) {
+  grpc_call_element_args args = {owning_call_, nullptr,          call_context_,
+                                 path_,        /*start_time=*/0, deadline_,
+                                 arena_,       call_combiner_};
+  return chand_->client_channel_->CreateLoadBalancedCall(
+      args, pollent_,
+      // This callback holds a ref to the CallStackDestructionBarrier
+      // object until the LB call is destroyed.
+      call_stack_destruction_barrier_->MakeLbCallDestructionClosure(this),
+      call_dispatch_controller,
+      // TODO(roth): Change this when we support transparent retries.
+      /*is_transparent_retry=*/false);
+}
+
+void RetryFilter::CallData::CreateCallAttempt() {
+  call_attempt_ = MakeRefCounted<CallAttempt>(this);
+  call_attempt_->StartRetriableBatches();
+}
+
+//
+// send op data caching
+//
+
+void RetryFilter::CallData::MaybeCacheSendOpsForBatch(PendingBatch* pending) {
+  if (pending->send_ops_cached) return;
+  pending->send_ops_cached = true;
+  grpc_transport_stream_op_batch* batch = pending->batch;
+  // Save a copy of metadata for send_initial_metadata ops.
+  if (batch->send_initial_metadata) {
+    seen_send_initial_metadata_ = true;
+    grpc_metadata_batch* send_initial_metadata =
+        batch->payload->send_initial_metadata.send_initial_metadata;
+    grpc_metadata_batch_copy(send_initial_metadata, &send_initial_metadata_);
+    send_initial_metadata_flags_ =
+        batch->payload->send_initial_metadata.send_initial_metadata_flags;
+    peer_string_ = batch->payload->send_initial_metadata.peer_string;
+  }
+  // Set up cache for send_message ops.
+  if (batch->send_message) {
+    ByteStreamCache* cache = arena_->New<ByteStreamCache>(
+        std::move(batch->payload->send_message.send_message));
+    send_messages_.push_back(cache);
+  }
+  // Save metadata batch for send_trailing_metadata ops.
+  if (batch->send_trailing_metadata) {
+    seen_send_trailing_metadata_ = true;
+    grpc_metadata_batch* send_trailing_metadata =
+        batch->payload->send_trailing_metadata.send_trailing_metadata;
+    grpc_metadata_batch_copy(send_trailing_metadata, &send_trailing_metadata_);
+  }
+}
+
+void RetryFilter::CallData::FreeCachedSendInitialMetadata() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p: destroying send_initial_metadata",
+            chand_, this);
+  }
+  send_initial_metadata_.Clear();
+}
+
+void RetryFilter::CallData::FreeCachedSendMessage(size_t idx) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p: destroying send_messages[%" PRIuPTR "]", chand_,
+            this, idx);
+  }
+  send_messages_[idx]->Destroy();
+}
+
+void RetryFilter::CallData::FreeCachedSendTrailingMetadata() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p: destroying send_trailing_metadata",
+            chand_, this);
+  }
+  send_trailing_metadata_.Clear();
+}
+
+void RetryFilter::CallData::FreeAllCachedSendOpData() {
+  if (seen_send_initial_metadata_) {
+    FreeCachedSendInitialMetadata();
+  }
+  for (size_t i = 0; i < send_messages_.size(); ++i) {
+    FreeCachedSendMessage(i);
+  }
+  if (seen_send_trailing_metadata_) {
+    FreeCachedSendTrailingMetadata();
+  }
+}
+
+//
+// pending_batches management
+//
+
+size_t RetryFilter::CallData::GetBatchIndex(
+    grpc_transport_stream_op_batch* batch) {
+  if (batch->send_initial_metadata) return 0;
+  if (batch->send_message) return 1;
+  if (batch->send_trailing_metadata) return 2;
+  if (batch->recv_initial_metadata) return 3;
+  if (batch->recv_message) return 4;
+  if (batch->recv_trailing_metadata) return 5;
+  GPR_UNREACHABLE_CODE(return (size_t)-1);
+}
+
+// This is called via the call combiner, so access to calld is synchronized.
+RetryFilter::CallData::PendingBatch* RetryFilter::CallData::PendingBatchesAdd(
+    grpc_transport_stream_op_batch* batch) {
+  const size_t idx = GetBatchIndex(batch);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p: adding pending batch at index %" PRIuPTR,
+            chand_, this, idx);
+  }
+  PendingBatch* pending = &pending_batches_[idx];
+  GPR_ASSERT(pending->batch == nullptr);
+  pending->batch = batch;
+  pending->send_ops_cached = false;
+  // Update state in calld about pending batches.
+  // Also check if the batch takes us over the retry buffer limit.
+  // Note: We don't check the size of trailing metadata here, because
+  // gRPC clients do not send trailing metadata.
+  if (batch->send_initial_metadata) {
+    pending_send_initial_metadata_ = true;
+    bytes_buffered_for_retry_ += batch->payload->send_initial_metadata
+                                     .send_initial_metadata->TransportSize();
+  }
+  if (batch->send_message) {
+    pending_send_message_ = true;
+    bytes_buffered_for_retry_ +=
+        batch->payload->send_message.send_message->length();
+  }
+  if (batch->send_trailing_metadata) {
+    pending_send_trailing_metadata_ = true;
+  }
+  // TODO(roth): When we implement hedging, if there are currently attempts
+  // in flight, we will need to pick the one on which the max number of send
+  // ops have already been sent, and we commit to that attempt.
+  if (GPR_UNLIKELY(bytes_buffered_for_retry_ >
+                   chand_->per_rpc_retry_buffer_size_)) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO,
+              "chand=%p calld=%p: exceeded retry buffer size, committing",
+              chand_, this);
+    }
+    RetryCommit(call_attempt_.get());
+  }
+  return pending;
+}
+
+void RetryFilter::CallData::PendingBatchClear(PendingBatch* pending) {
+  if (pending->batch->send_initial_metadata) {
+    pending_send_initial_metadata_ = false;
+  }
+  if (pending->batch->send_message) {
+    pending_send_message_ = false;
+  }
+  if (pending->batch->send_trailing_metadata) {
+    pending_send_trailing_metadata_ = false;
+  }
+  pending->batch = nullptr;
+}
+
+void RetryFilter::CallData::MaybeClearPendingBatch(PendingBatch* pending) {
+  grpc_transport_stream_op_batch* batch = pending->batch;
+  // We clear the pending batch if all of its callbacks have been
+  // scheduled and reset to nullptr.
+  if (batch->on_complete == nullptr &&
+      (!batch->recv_initial_metadata ||
+       batch->payload->recv_initial_metadata.recv_initial_metadata_ready ==
+           nullptr) &&
+      (!batch->recv_message ||
+       batch->payload->recv_message.recv_message_ready == nullptr) &&
+      (!batch->recv_trailing_metadata ||
+       batch->payload->recv_trailing_metadata.recv_trailing_metadata_ready ==
+           nullptr)) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+      gpr_log(GPR_INFO, "chand=%p calld=%p: clearing pending batch", chand_,
+              this);
+    }
+    PendingBatchClear(pending);
+  }
+}
+
+// This is called via the call combiner, so access to calld is synchronized.
+void RetryFilter::CallData::FailPendingBatchInCallCombiner(
+    void* arg, grpc_error_handle error) {
+  grpc_transport_stream_op_batch* batch =
+      static_cast<grpc_transport_stream_op_batch*>(arg);
+  CallData* call = static_cast<CallData*>(batch->handler_private.extra_arg);
+  // Note: This will release the call combiner.
+  grpc_transport_stream_op_batch_finish_with_failure(
+      batch, GRPC_ERROR_REF(error), call->call_combiner_);
+}
+
+// This is called via the call combiner, so access to calld is synchronized.
+void RetryFilter::CallData::PendingBatchesFail(grpc_error_handle error) {
+  GPR_ASSERT(error != GRPC_ERROR_NONE);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    size_t num_batches = 0;
+    for (size_t i = 0; i < GPR_ARRAY_SIZE(pending_batches_); ++i) {
+      if (pending_batches_[i].batch != nullptr) ++num_batches;
+    }
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p: failing %" PRIuPTR " pending batches: %s",
+            chand_, this, num_batches, grpc_error_std_string(error).c_str());
+  }
+  CallCombinerClosureList closures;
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(pending_batches_); ++i) {
+    PendingBatch* pending = &pending_batches_[i];
+    grpc_transport_stream_op_batch* batch = pending->batch;
+    if (batch != nullptr) {
+      batch->handler_private.extra_arg = this;
+      GRPC_CLOSURE_INIT(&batch->handler_private.closure,
+                        FailPendingBatchInCallCombiner, batch,
+                        grpc_schedule_on_exec_ctx);
+      closures.Add(&batch->handler_private.closure, GRPC_ERROR_REF(error),
+                   "PendingBatchesFail");
+      PendingBatchClear(pending);
+    }
+  }
+  closures.RunClosuresWithoutYielding(call_combiner_);
+  GRPC_ERROR_UNREF(error);
+}
+
+template <typename Predicate>
+RetryFilter::CallData::PendingBatch* RetryFilter::CallData::PendingBatchFind(
+    const char* log_message, Predicate predicate) {
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(pending_batches_); ++i) {
+    PendingBatch* pending = &pending_batches_[i];
+    grpc_transport_stream_op_batch* batch = pending->batch;
+    if (batch != nullptr && predicate(batch)) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+        gpr_log(GPR_INFO,
+                "chand=%p calld=%p: %s pending batch at index %" PRIuPTR,
+                chand_, this, log_message, i);
+      }
+      return pending;
+    }
+  }
+  return nullptr;
+}
+
+//
+// retry code
+//
+
+void RetryFilter::CallData::RetryCommit(CallAttempt* call_attempt) {
+  if (retry_committed_) return;
+  retry_committed_ = true;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO, "chand=%p calld=%p: committing retries", chand_, this);
+  }
+  if (call_attempt != nullptr) {
+    // If the call attempt's LB call has been committed, inform the call
+    // dispatch controller that the call has been committed.
+    // Note: If call_attempt is null, this is happening before the first
+    // retry attempt is started, in which case we'll just pass the real
+    // call dispatch controller down into the LB call, and it won't be
+    // our problem anymore.
+    if (call_attempt->lb_call_committed()) {
+      auto* service_config_call_data =
+          static_cast<ClientChannelServiceConfigCallData*>(
+              call_context_[GRPC_CONTEXT_SERVICE_CONFIG_CALL_DATA].value);
+      service_config_call_data->call_dispatch_controller()->Commit();
+    }
+    // Free cached send ops.
+    call_attempt->FreeCachedSendOpDataAfterCommit();
+  }
+}
+
+void RetryFilter::CallData::StartRetryTimer(grpc_millis server_pushback_ms) {
+  // Reset call attempt.
+  call_attempt_.reset(DEBUG_LOCATION, "StartRetryTimer");
+  // Compute backoff delay.
+  grpc_millis next_attempt_time;
+  if (server_pushback_ms >= 0) {
+    next_attempt_time = ExecCtx::Get()->Now() + server_pushback_ms;
+    retry_backoff_.Reset();
+  } else {
+    next_attempt_time = retry_backoff_.NextAttemptTime();
+  }
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_retry_trace)) {
+    gpr_log(GPR_INFO,
+            "chand=%p calld=%p: retrying failed call in %" PRId64 " ms", chand_,
+            this, next_attempt_time - ExecCtx::Get()->Now());
+  }
+  // Schedule retry after computed delay.
+  GRPC_CLOSURE_INIT(&retry_closure_, OnRetryTimer, this, nullptr);
+  GRPC_CALL_STACK_REF(owning_call_, "OnRetryTimer");
+  retry_timer_pending_ = true;
+  grpc_timer_init(&retry_timer_, next_attempt_time, &retry_closure_);
+}
+
+void RetryFilter::CallData::OnRetryTimer(void* arg, grpc_error_handle error) {
+  auto* calld = static_cast<CallData*>(arg);
+  GRPC_CLOSURE_INIT(&calld->retry_closure_, OnRetryTimerLocked, calld, nullptr);
+  GRPC_CALL_COMBINER_START(calld->call_combiner_, &calld->retry_closure_,
+                           GRPC_ERROR_REF(error), "retry timer fired");
+}
+
+void RetryFilter::CallData::OnRetryTimerLocked(void* arg,
+                                               grpc_error_handle error) {
+  auto* calld = static_cast<CallData*>(arg);
+  if (error == GRPC_ERROR_NONE && calld->retry_timer_pending_) {
+    calld->retry_timer_pending_ = false;
+    calld->CreateCallAttempt();
+  } else {
+    GRPC_CALL_COMBINER_STOP(calld->call_combiner_, "retry timer cancelled");
+  }
+  GRPC_CALL_STACK_UNREF(calld->owning_call_, "OnRetryTimer");
+}
+
+}  // namespace
+
+const grpc_channel_filter kRetryFilterVtable = {
+    RetryFilter::CallData::StartTransportStreamOpBatch,
+    RetryFilter::StartTransportOp,
+    sizeof(RetryFilter::CallData),
+    RetryFilter::CallData::Init,
+    RetryFilter::CallData::SetPollent,
+    RetryFilter::CallData::Destroy,
+    sizeof(RetryFilter),
+    RetryFilter::Init,
+    RetryFilter::Destroy,
+    RetryFilter::GetChannelInfo,
+    "retry_filter",
+};
+
+}  // namespace grpc_core