Merge github.com:grpc/grpc into server_stats

86 files changed, 3739 insertions, 4243 deletions

diff --git a/src/core/ext/census/context.c b/src/core/ext/census/context.c
index 1019b287d7..9b25a32e36 100644
--- a/src/core/ext/census/context.c
+++ b/src/core/ext/census/context.c
@@ -141,7 +141,7 @@ static char *decode_tag(struct raw_tag *tag, char *header, int offset) {
 // Make a copy (in 'to') of an existing tag_set.
 static void tag_set_copy(struct tag_set *to, const struct tag_set *from) {
   memcpy(to, from, sizeof(struct tag_set));
-  to->kvm = gpr_malloc(to->kvm_size);
+  to->kvm = (char *)gpr_malloc(to->kvm_size);
   memcpy(to->kvm, from->kvm, from->kvm_used);
 }
 
@@ -184,7 +184,7 @@ static bool tag_set_add_tag(struct tag_set *tags, const census_tag *tag,
   if (tags->kvm_used + tag_size > tags->kvm_size) {
     // allocate new memory if needed
     tags->kvm_size += 2 * CENSUS_MAX_TAG_KV_LEN + TAG_HEADER_SIZE;
-    char *new_kvm = gpr_malloc(tags->kvm_size);
+    char *new_kvm = (char *)gpr_malloc(tags->kvm_size);
     if (tags->kvm_used > 0) memcpy(new_kvm, tags->kvm, tags->kvm_used);
     gpr_free(tags->kvm);
     tags->kvm = new_kvm;
@@ -274,7 +274,8 @@ static void tag_set_flatten(struct tag_set *tags) {
 census_context *census_context_create(const census_context *base,
                                       const census_tag *tags, int ntags,
                                       census_context_status const **status) {
-  census_context *context = gpr_malloc(sizeof(census_context));
+  census_context *context =
+      (census_context *)gpr_malloc(sizeof(census_context));
   // If we are given a base, copy it into our new tag set. Otherwise set it
   // to zero/NULL everything.
   if (base == NULL) {
@@ -459,7 +460,7 @@ static void tag_set_decode(struct tag_set *tags, const char *buffer,
   }
   tags->kvm_used = size - header_size;
   tags->kvm_size = tags->kvm_used + CENSUS_MAX_TAG_KV_LEN;
-  tags->kvm = gpr_malloc(tags->kvm_size);
+  tags->kvm = (char *)gpr_malloc(tags->kvm_size);
   if (tag_header_size != TAG_HEADER_SIZE) {
     // something new in the tag information. I don't understand it, so
     // don't copy it over.
@@ -481,7 +482,8 @@ static void tag_set_decode(struct tag_set *tags, const char *buffer,
 }
 
 census_context *census_context_decode(const char *buffer, size_t size) {
-  census_context *context = gpr_malloc(sizeof(census_context));
+  census_context *context =
+      (census_context *)gpr_malloc(sizeof(census_context));
   memset(&context->tags[LOCAL_TAGS], 0, sizeof(struct tag_set));
   if (buffer == NULL) {
     memset(&context->tags[PROPAGATED_TAGS], 0, sizeof(struct tag_set));
diff --git a/src/core/ext/census/grpc_filter.c b/src/core/ext/census/grpc_filter.c
index 13fe2e6b1c..b37ab90389 100644
--- a/src/core/ext/census/grpc_filter.c
+++ b/src/core/ext/census/grpc_filter.c
@@ -60,8 +60,8 @@ static void extract_and_annotate_method_tag(grpc_metadata_batch *md,
 
 static void client_mutate_op(grpc_call_element *elem,
                              grpc_transport_stream_op_batch *op) {
-  call_data *calld = elem->call_data;
-  channel_data *chand = elem->channel_data;
+  call_data *calld = (call_data *)elem->call_data;
+  channel_data *chand = (channel_data *)elem->channel_data;
   if (op->send_initial_metadata) {
     extract_and_annotate_method_tag(
         op->payload->send_initial_metadata.send_initial_metadata, calld, chand);
@@ -78,9 +78,9 @@ static void client_start_transport_op(grpc_exec_ctx *exec_ctx,
 static void server_on_done_recv(grpc_exec_ctx *exec_ctx, void *ptr,
                                 grpc_error *error) {
   GPR_TIMER_BEGIN("census-server:server_on_done_recv", 0);
-  grpc_call_element *elem = ptr;
-  call_data *calld = elem->call_data;
-  channel_data *chand = elem->channel_data;
+  grpc_call_element *elem = (grpc_call_element *)ptr;
+  call_data *calld = (call_data *)elem->call_data;
+  channel_data *chand = (channel_data *)elem->channel_data;
   if (error == GRPC_ERROR_NONE) {
     extract_and_annotate_method_tag(calld->recv_initial_metadata, calld, chand);
   }
@@ -90,7 +90,7 @@ static void server_on_done_recv(grpc_exec_ctx *exec_ctx, void *ptr,
 
 static void server_mutate_op(grpc_call_element *elem,
                              grpc_transport_stream_op_batch *op) {
-  call_data *calld = elem->call_data;
+  call_data *calld = (call_data *)elem->call_data;
   if (op->recv_initial_metadata) {
     /* substitute our callback for the op callback */
     calld->recv_initial_metadata =
@@ -117,7 +117,7 @@ static void server_start_transport_op(grpc_exec_ctx *exec_ctx,
 static grpc_error *client_init_call_elem(grpc_exec_ctx *exec_ctx,
                                          grpc_call_element *elem,
                                          const grpc_call_element_args *args) {
-  call_data *d = elem->call_data;
+  call_data *d = (call_data *)elem->call_data;
   GPR_ASSERT(d != NULL);
   memset(d, 0, sizeof(*d));
   d->start_ts = args->start_time;
@@ -128,7 +128,7 @@ static void client_destroy_call_elem(grpc_exec_ctx *exec_ctx,
                                      grpc_call_element *elem,
                                      const grpc_call_final_info *final_info,
                                      grpc_closure *ignored) {
-  call_data *d = elem->call_data;
+  call_data *d = (call_data *)elem->call_data;
   GPR_ASSERT(d != NULL);
   /* TODO(hongyu): record rpc client stats and census_rpc_end_op here */
 }
@@ -136,7 +136,7 @@ static void client_destroy_call_elem(grpc_exec_ctx *exec_ctx,
 static grpc_error *server_init_call_elem(grpc_exec_ctx *exec_ctx,
                                          grpc_call_element *elem,
                                          const grpc_call_element_args *args) {
-  call_data *d = elem->call_data;
+  call_data *d = (call_data *)elem->call_data;
   GPR_ASSERT(d != NULL);
   memset(d, 0, sizeof(*d));
   d->start_ts = args->start_time;
@@ -150,7 +150,7 @@ static void server_destroy_call_elem(grpc_exec_ctx *exec_ctx,
                                      grpc_call_element *elem,
                                      const grpc_call_final_info *final_info,
                                      grpc_closure *ignored) {
-  call_data *d = elem->call_data;
+  call_data *d = (call_data *)elem->call_data;
   GPR_ASSERT(d != NULL);
   /* TODO(hongyu): record rpc server stats and census_tracing_end_op here */
 }
@@ -158,14 +158,14 @@ static void server_destroy_call_elem(grpc_exec_ctx *exec_ctx,
 static grpc_error *init_channel_elem(grpc_exec_ctx *exec_ctx,
                                      grpc_channel_element *elem,
                                      grpc_channel_element_args *args) {
-  channel_data *chand = elem->channel_data;
+  channel_data *chand = (channel_data *)elem->channel_data;
   GPR_ASSERT(chand != NULL);
   return GRPC_ERROR_NONE;
 }
 
 static void destroy_channel_elem(grpc_exec_ctx *exec_ctx,
                                  grpc_channel_element *elem) {
-  channel_data *chand = elem->channel_data;
+  channel_data *chand = (channel_data *)elem->channel_data;
   GPR_ASSERT(chand != NULL);
 }
 
@@ -179,7 +179,6 @@ const grpc_channel_filter grpc_client_census_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "census-client"};
 
@@ -193,6 +192,5 @@ const grpc_channel_filter grpc_server_census_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "census-server"};
diff --git a/src/core/ext/census/mlog.c b/src/core/ext/census/mlog.c
index 937ceb101b..4b8c8466b3 100644
--- a/src/core/ext/census/mlog.c
+++ b/src/core/ext/census/mlog.c
@@ -467,7 +467,8 @@ void census_log_initialize(size_t size_in_mb, int discard_old_records) {
   g_log.blocks = (cl_block*)gpr_malloc_aligned(
       g_log.num_blocks * sizeof(cl_block), GPR_CACHELINE_SIZE_LOG);
   memset(g_log.blocks, 0, g_log.num_blocks * sizeof(cl_block));
-  g_log.buffer = gpr_malloc(g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
+  g_log.buffer =
+      (char*)gpr_malloc(g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
   memset(g_log.buffer, 0, g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
   cl_block_list_initialize(&g_log.free_block_list);
   cl_block_list_initialize(&g_log.dirty_block_list);
diff --git a/src/core/ext/census/resource.c b/src/core/ext/census/resource.c
index 1a676f0e1e..44a887231c 100644
--- a/src/core/ext/census/resource.c
+++ b/src/core/ext/census/resource.c
@@ -87,7 +87,7 @@ static bool validate_string(pb_istream_t *stream, const pb_field_t *field,
         gpr_log(GPR_INFO, "Zero-length Resource name.");
         return false;
       }
-      vresource->name = gpr_malloc(stream->bytes_left + 1);
+      vresource->name = (char *)gpr_malloc(stream->bytes_left + 1);
       vresource->name[stream->bytes_left] = '\0';
       if (!pb_read(stream, (uint8_t *)vresource->name, stream->bytes_left)) {
         return false;
@@ -106,7 +106,7 @@ static bool validate_string(pb_istream_t *stream, const pb_field_t *field,
       if (stream->bytes_left == 0) {
         return true;
       }
-      vresource->description = gpr_malloc(stream->bytes_left + 1);
+      vresource->description = (char *)gpr_malloc(stream->bytes_left + 1);
       vresource->description[stream->bytes_left] = '\0';
       if (!pb_read(stream, (uint8_t *)vresource->description,
                    stream->bytes_left)) {
@@ -134,7 +134,8 @@ static bool validate_units_helper(pb_istream_t *stream, int *count,
     // Have to allocate a new array of values. Normal case is 0 or 1, so
     // this should normally not be an issue.
     google_census_Resource_BasicUnit *new_bup =
-        gpr_malloc((size_t)*count * sizeof(google_census_Resource_BasicUnit));
+        (google_census_Resource_BasicUnit *)gpr_malloc(
+            (size_t)*count * sizeof(google_census_Resource_BasicUnit));
     if (*count != 1) {
       memcpy(new_bup, *bup,
              (size_t)(*count - 1) * sizeof(google_census_Resource_BasicUnit));
@@ -207,7 +208,8 @@ size_t allocate_resource(void) {
   // Expand resources if needed.
   if (n_resources == n_defined_resources) {
     size_t new_n_resources = n_resources ? n_resources * 2 : 2;
-    resource **new_resources = gpr_malloc(new_n_resources * sizeof(resource *));
+    resource **new_resources =
+        (resource **)gpr_malloc(new_n_resources * sizeof(resource *));
     if (n_resources != 0) {
       memcpy(new_resources, resources, n_resources * sizeof(resource *));
     }
@@ -226,7 +228,7 @@ size_t allocate_resource(void) {
     }
   }
   GPR_ASSERT(id < n_resources && resources[id] == NULL);
-  resources[id] = gpr_malloc(sizeof(resource));
+  resources[id] = (resource *)gpr_malloc(sizeof(resource));
   memset(resources[id], 0, sizeof(resource));
   n_defined_resources++;
   next_id = (id + 1) % n_resources;
@@ -276,22 +278,24 @@ int32_t define_resource(const resource *base) {
   gpr_mu_lock(&resource_lock);
   size_t id = allocate_resource();
   size_t len = strlen(base->name) + 1;
-  resources[id]->name = gpr_malloc(len);
+  resources[id]->name = (char *)gpr_malloc(len);
   memcpy(resources[id]->name, base->name, len);
   if (base->description) {
     len = strlen(base->description) + 1;
-    resources[id]->description = gpr_malloc(len);
+    resources[id]->description = (char *)gpr_malloc(len);
     memcpy(resources[id]->description, base->description, len);
   }
   resources[id]->prefix = base->prefix;
   resources[id]->n_numerators = base->n_numerators;
   len = (size_t)base->n_numerators * sizeof(*base->numerators);
-  resources[id]->numerators = gpr_malloc(len);
+  resources[id]->numerators =
+      (google_census_Resource_BasicUnit *)gpr_malloc(len);
   memcpy(resources[id]->numerators, base->numerators, len);
   resources[id]->n_denominators = base->n_denominators;
   if (base->n_denominators != 0) {
     len = (size_t)base->n_denominators * sizeof(*base->denominators);
-    resources[id]->denominators = gpr_malloc(len);
+    resources[id]->denominators =
+        (google_census_Resource_BasicUnit *)gpr_malloc(len);
     memcpy(resources[id]->denominators, base->denominators, len);
   }
   gpr_mu_unlock(&resource_lock);
diff --git a/src/core/ext/filters/client_channel/channel_connectivity.c b/src/core/ext/filters/client_channel/channel_connectivity.c
index b83c95275f..0a9e90d12e 100644
--- a/src/core/ext/filters/client_channel/channel_connectivity.c
+++ b/src/core/ext/filters/client_channel/channel_connectivity.c
@@ -191,6 +191,12 @@ static void watcher_timer_init(grpc_exec_ctx *exec_ctx, void *arg,
   gpr_free(wa);
 }
 
+int grpc_channel_support_connectivity_watcher(grpc_channel *channel) {
+  grpc_channel_element *client_channel_elem =
+      grpc_channel_stack_last_element(grpc_channel_get_channel_stack(channel));
+  return client_channel_elem->filter != &grpc_client_channel_filter ? 0 : 1;
+}
+
 void grpc_channel_watch_connectivity_state(
     grpc_channel *channel, grpc_connectivity_state last_observed_state,
     gpr_timespec deadline, grpc_completion_queue *cq, void *tag) {
diff --git a/src/core/ext/filters/client_channel/client_channel.c b/src/core/ext/filters/client_channel/client_channel.c
index 58e31d7b45..c24e52ec1d 100644
--- a/src/core/ext/filters/client_channel/client_channel.c
+++ b/src/core/ext/filters/client_channel/client_channel.c
@@ -796,7 +796,8 @@ static void cc_destroy_channel_elem(grpc_exec_ctx *exec_ctx,
 //   send_message
 //   recv_trailing_metadata
 //   send_trailing_metadata
-#define MAX_WAITING_BATCHES 6
+// We also add room for a single cancel_stream batch.
+#define MAX_WAITING_BATCHES 7
 
 /** Call data.  Holds a pointer to grpc_subchannel_call and the
     associated machinery to create such a pointer.
@@ -807,24 +808,27 @@ typedef struct client_channel_call_data {
   // State for handling deadlines.
   // The code in deadline_filter.c requires this to be the first field.
   // TODO(roth): This is slightly sub-optimal in that grpc_deadline_state
-  // and this struct both independently store a pointer to the call
-  // stack and each has its own mutex.  If/when we have time, find a way
-  // to avoid this without breaking the grpc_deadline_state abstraction.
+  // and this struct both independently store pointers to the call stack
+  // and call combiner.  If/when we have time, find a way to avoid this
+  // without breaking the grpc_deadline_state abstraction.
   grpc_deadline_state deadline_state;
 
   grpc_slice path;  // Request path.
   gpr_timespec call_start_time;
   gpr_timespec deadline;
+  gpr_arena *arena;
+  grpc_call_stack *owning_call;
+  grpc_call_combiner *call_combiner;
+
   grpc_server_retry_throttle_data *retry_throttle_data;
   method_parameters *method_params;
 
-  /** either 0 for no call, a pointer to a grpc_subchannel_call (if the lowest
-      bit is 0), or a pointer to an error (if the lowest bit is 1) */
-  gpr_atm subchannel_call_or_error;
-  gpr_arena *arena;
+  grpc_subchannel_call *subchannel_call;
+  grpc_error *error;
 
   grpc_lb_policy *lb_policy;  // Holds ref while LB pick is pending.
   grpc_closure lb_pick_closure;
+  grpc_closure lb_pick_cancel_closure;
 
   grpc_connected_subchannel *connected_subchannel;
   grpc_call_context_element subchannel_call_context[GRPC_CONTEXT_COUNT];
@@ -832,10 +836,9 @@ typedef struct client_channel_call_data {
 
   grpc_transport_stream_op_batch *waiting_for_pick_batches[MAX_WAITING_BATCHES];
   size_t waiting_for_pick_batches_count;
+  grpc_closure handle_pending_batch_in_call_combiner[MAX_WAITING_BATCHES];
 
-  grpc_transport_stream_op_batch_payload *initial_metadata_payload;
-
-  grpc_call_stack *owning_call;
+  grpc_transport_stream_op_batch *initial_metadata_batch;
 
   grpc_linked_mdelem lb_token_mdelem;
 
@@ -843,55 +846,42 @@ typedef struct client_channel_call_data {
   grpc_closure *original_on_complete;
 } call_data;
 
-typedef struct {
-  grpc_subchannel_call *subchannel_call;
-  grpc_error *error;
-} call_or_error;
-
-static call_or_error get_call_or_error(call_data *p) {
-  gpr_atm c = gpr_atm_acq_load(&p->subchannel_call_or_error);
-  if (c == 0)
-    return (call_or_error){NULL, NULL};
-  else if (c & 1)
-    return (call_or_error){NULL, (grpc_error *)((c) & ~(gpr_atm)1)};
-  else
-    return (call_or_error){(grpc_subchannel_call *)c, NULL};
+grpc_subchannel_call *grpc_client_channel_get_subchannel_call(
+    grpc_call_element *elem) {
+  call_data *calld = elem->call_data;
+  return calld->subchannel_call;
 }
 
-static bool set_call_or_error(call_data *p, call_or_error coe) {
-  // this should always be under a lock
-  call_or_error existing = get_call_or_error(p);
-  if (existing.error != GRPC_ERROR_NONE) {
-    GRPC_ERROR_UNREF(coe.error);
-    return false;
-  }
-  GPR_ASSERT(existing.subchannel_call == NULL);
-  if (coe.error != GRPC_ERROR_NONE) {
-    GPR_ASSERT(coe.subchannel_call == NULL);
-    gpr_atm_rel_store(&p->subchannel_call_or_error, 1 | (gpr_atm)coe.error);
+// This is called via the call combiner, so access to calld is synchronized.
+static void waiting_for_pick_batches_add(
+    call_data *calld, grpc_transport_stream_op_batch *batch) {
+  if (batch->send_initial_metadata) {
+    GPR_ASSERT(calld->initial_metadata_batch == NULL);
+    calld->initial_metadata_batch = batch;
   } else {
-    GPR_ASSERT(coe.subchannel_call != NULL);
-    gpr_atm_rel_store(&p->subchannel_call_or_error,
-                      (gpr_atm)coe.subchannel_call);
+    GPR_ASSERT(calld->waiting_for_pick_batches_count < MAX_WAITING_BATCHES);
+    calld->waiting_for_pick_batches[calld->waiting_for_pick_batches_count++] =
+        batch;
   }
-  return true;
 }
 
-grpc_subchannel_call *grpc_client_channel_get_subchannel_call(
-    grpc_call_element *call_elem) {
-  return get_call_or_error(call_elem->call_data).subchannel_call;
-}
-
-static void waiting_for_pick_batches_add_locked(
-    call_data *calld, grpc_transport_stream_op_batch *batch) {
-  GPR_ASSERT(calld->waiting_for_pick_batches_count < MAX_WAITING_BATCHES);
-  calld->waiting_for_pick_batches[calld->waiting_for_pick_batches_count++] =
-      batch;
+// This is called via the call combiner, so access to calld is synchronized.
+static void fail_pending_batch_in_call_combiner(grpc_exec_ctx *exec_ctx,
+                                                void *arg, grpc_error *error) {
+  call_data *calld = arg;
+  if (calld->waiting_for_pick_batches_count > 0) {
+    --calld->waiting_for_pick_batches_count;
+    grpc_transport_stream_op_batch_finish_with_failure(
+        exec_ctx,
+        calld->waiting_for_pick_batches[calld->waiting_for_pick_batches_count],
+        GRPC_ERROR_REF(error), calld->call_combiner);
+  }
 }
 
-static void waiting_for_pick_batches_fail_locked(grpc_exec_ctx *exec_ctx,
-                                                 grpc_call_element *elem,
-                                                 grpc_error *error) {
+// This is called via the call combiner, so access to calld is synchronized.
+static void waiting_for_pick_batches_fail(grpc_exec_ctx *exec_ctx,
+                                          grpc_call_element *elem,
+                                          grpc_error *error) {
   call_data *calld = elem->call_data;
   if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
     gpr_log(GPR_DEBUG,
@@ -900,34 +890,60 @@ static void waiting_for_pick_batches_fail_locked(grpc_exec_ctx *exec_ctx,
             grpc_error_string(error));
   }
   for (size_t i = 0; i < calld->waiting_for_pick_batches_count; ++i) {
+    GRPC_CLOSURE_INIT(&calld->handle_pending_batch_in_call_combiner[i],
+                      fail_pending_batch_in_call_combiner, calld,
+                      grpc_schedule_on_exec_ctx);
+    GRPC_CALL_COMBINER_START(exec_ctx, calld->call_combiner,
+                             &calld->handle_pending_batch_in_call_combiner[i],
+                             GRPC_ERROR_REF(error),
+                             "waiting_for_pick_batches_fail");
+  }
+  if (calld->initial_metadata_batch != NULL) {
     grpc_transport_stream_op_batch_finish_with_failure(
-        exec_ctx, calld->waiting_for_pick_batches[i], GRPC_ERROR_REF(error));
+        exec_ctx, calld->initial_metadata_batch, GRPC_ERROR_REF(error),
+        calld->call_combiner);
+  } else {
+    GRPC_CALL_COMBINER_STOP(exec_ctx, calld->call_combiner,
+                            "waiting_for_pick_batches_fail");
   }
-  calld->waiting_for_pick_batches_count = 0;
   GRPC_ERROR_UNREF(error);
 }
 
-static void waiting_for_pick_batches_resume_locked(grpc_exec_ctx *exec_ctx,
-                                                   grpc_call_element *elem) {
-  call_data *calld = elem->call_data;
-  if (calld->waiting_for_pick_batches_count == 0) return;
-  call_or_error coe = get_call_or_error(calld);
-  if (coe.error != GRPC_ERROR_NONE) {
-    waiting_for_pick_batches_fail_locked(exec_ctx, elem,
-                                         GRPC_ERROR_REF(coe.error));
-    return;
+// This is called via the call combiner, so access to calld is synchronized.
+static void run_pending_batch_in_call_combiner(grpc_exec_ctx *exec_ctx,
+                                               void *arg, grpc_error *ignored) {
+  call_data *calld = arg;
+  if (calld->waiting_for_pick_batches_count > 0) {
+    --calld->waiting_for_pick_batches_count;
+    grpc_subchannel_call_process_op(
+        exec_ctx, calld->subchannel_call,
+        calld->waiting_for_pick_batches[calld->waiting_for_pick_batches_count]);
   }
+}
+
+// This is called via the call combiner, so access to calld is synchronized.
+static void waiting_for_pick_batches_resume(grpc_exec_ctx *exec_ctx,
+                                            grpc_call_element *elem) {
+  channel_data *chand = elem->channel_data;
+  call_data *calld = elem->call_data;
   if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
     gpr_log(GPR_DEBUG, "chand=%p calld=%p: sending %" PRIdPTR
                        " pending batches to subchannel_call=%p",
-            elem->channel_data, calld, calld->waiting_for_pick_batches_count,
-            coe.subchannel_call);
+            chand, calld, calld->waiting_for_pick_batches_count,
+            calld->subchannel_call);
   }
   for (size_t i = 0; i < calld->waiting_for_pick_batches_count; ++i) {
-    grpc_subchannel_call_process_op(exec_ctx, coe.subchannel_call,
-                                    calld->waiting_for_pick_batches[i]);
-  }
-  calld->waiting_for_pick_batches_count = 0;
+    GRPC_CLOSURE_INIT(&calld->handle_pending_batch_in_call_combiner[i],
+                      run_pending_batch_in_call_combiner, calld,
+                      grpc_schedule_on_exec_ctx);
+    GRPC_CALL_COMBINER_START(exec_ctx, calld->call_combiner,
+                             &calld->handle_pending_batch_in_call_combiner[i],
+                             GRPC_ERROR_NONE,
+                             "waiting_for_pick_batches_resume");
+  }
+  GPR_ASSERT(calld->initial_metadata_batch != NULL);
+  grpc_subchannel_call_process_op(exec_ctx, calld->subchannel_call,
+                                  calld->initial_metadata_batch);
 }
 
 // Applies service config to the call.  Must be invoked once we know
@@ -968,29 +984,28 @@ static void apply_service_config_to_call_locked(grpc_exec_ctx *exec_ctx,
 static void create_subchannel_call_locked(grpc_exec_ctx *exec_ctx,
                                           grpc_call_element *elem,
                                           grpc_error *error) {
+  channel_data *chand = elem->channel_data;
   call_data *calld = elem->call_data;
-  grpc_subchannel_call *subchannel_call = NULL;
   const grpc_connected_subchannel_call_args call_args = {
       .pollent = calld->pollent,
       .path = calld->path,
       .start_time = calld->call_start_time,
       .deadline = calld->deadline,
       .arena = calld->arena,
-      .context = calld->subchannel_call_context};
+      .context = calld->subchannel_call_context,
+      .call_combiner = calld->call_combiner};
   grpc_error *new_error = grpc_connected_subchannel_create_call(
-      exec_ctx, calld->connected_subchannel, &call_args, &subchannel_call);
+      exec_ctx, calld->connected_subchannel, &call_args,
+      &calld->subchannel_call);
   if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
     gpr_log(GPR_DEBUG, "chand=%p calld=%p: create subchannel_call=%p: error=%s",
-            elem->channel_data, calld, subchannel_call,
-            grpc_error_string(new_error));
+            chand, calld, calld->subchannel_call, grpc_error_string(new_error));
   }
-  GPR_ASSERT(set_call_or_error(
-      calld, (call_or_error){.subchannel_call = subchannel_call}));
   if (new_error != GRPC_ERROR_NONE) {
     new_error = grpc_error_add_child(new_error, error);
-    waiting_for_pick_batches_fail_locked(exec_ctx, elem, new_error);
+    waiting_for_pick_batches_fail(exec_ctx, elem, new_error);
   } else {
-    waiting_for_pick_batches_resume_locked(exec_ctx, elem);
+    waiting_for_pick_batches_resume(exec_ctx, elem);
   }
   GRPC_ERROR_UNREF(error);
 }
@@ -1002,60 +1017,27 @@ static void subchannel_ready_locked(grpc_exec_ctx *exec_ctx,
   channel_data *chand = elem->channel_data;
   grpc_polling_entity_del_from_pollset_set(exec_ctx, calld->pollent,
                                            chand->interested_parties);
-  call_or_error coe = get_call_or_error(calld);
   if (calld->connected_subchannel == NULL) {
     // Failed to create subchannel.
-    grpc_error *failure =
-        error == GRPC_ERROR_NONE
-            ? GRPC_ERROR_CREATE_FROM_STATIC_STRING(
-                  "Call dropped by load balancing policy")
-            : GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
-                  "Failed to create subchannel", &error, 1);
+    GRPC_ERROR_UNREF(calld->error);
+    calld->error = error == GRPC_ERROR_NONE
+                       ? GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+                             "Call dropped by load balancing policy")
+                       : GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
+                             "Failed to create subchannel", &error, 1);
     if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
       gpr_log(GPR_DEBUG,
               "chand=%p calld=%p: failed to create subchannel: error=%s", chand,
-              calld, grpc_error_string(failure));
-    }
-    set_call_or_error(calld, (call_or_error){.error = GRPC_ERROR_REF(failure)});
-    waiting_for_pick_batches_fail_locked(exec_ctx, elem, failure);
-  } else if (coe.error != GRPC_ERROR_NONE) {
-    /* already cancelled before subchannel became ready */
-    grpc_error *child_errors[] = {error, coe.error};
-    grpc_error *cancellation_error =
-        GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
-            "Cancelled before creating subchannel", child_errors,
-            GPR_ARRAY_SIZE(child_errors));
-    /* if due to deadline, attach the deadline exceeded status to the error */
-    if (gpr_time_cmp(calld->deadline, gpr_now(GPR_CLOCK_MONOTONIC)) < 0) {
-      cancellation_error =
-          grpc_error_set_int(cancellation_error, GRPC_ERROR_INT_GRPC_STATUS,
-                             GRPC_STATUS_DEADLINE_EXCEEDED);
-    }
-    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-      gpr_log(GPR_DEBUG,
-              "chand=%p calld=%p: cancelled before subchannel became ready: %s",
-              chand, calld, grpc_error_string(cancellation_error));
+              calld, grpc_error_string(calld->error));
     }
-    waiting_for_pick_batches_fail_locked(exec_ctx, elem, cancellation_error);
+    waiting_for_pick_batches_fail(exec_ctx, elem, GRPC_ERROR_REF(calld->error));
   } else {
     /* Create call on subchannel. */
     create_subchannel_call_locked(exec_ctx, elem, GRPC_ERROR_REF(error));
   }
-  GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call, "pick_subchannel");
   GRPC_ERROR_UNREF(error);
 }
 
-static char *cc_get_peer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem) {
-  call_data *calld = elem->call_data;
-  grpc_subchannel_call *subchannel_call =
-      get_call_or_error(calld).subchannel_call;
-  if (subchannel_call == NULL) {
-    return NULL;
-  } else {
-    return grpc_subchannel_call_get_peer(exec_ctx, subchannel_call);
-  }
-}
-
 /** Return true if subchannel is available immediately (in which case
     subchannel_ready_locked() should not be called), or false otherwise (in
     which case subchannel_ready_locked() should be called when the subchannel
@@ -1065,39 +1047,78 @@ static bool pick_subchannel_locked(grpc_exec_ctx *exec_ctx,
 
 typedef struct {
   grpc_call_element *elem;
-  bool cancelled;
+  bool finished;
   grpc_closure closure;
+  grpc_closure cancel_closure;
 } pick_after_resolver_result_args;
 
+// Note: This runs under the client_channel combiner, but will NOT be
+// holding the call combiner.
+static void pick_after_resolver_result_cancel_locked(grpc_exec_ctx *exec_ctx,
+                                                     void *arg,
+                                                     grpc_error *error) {
+  pick_after_resolver_result_args *args = arg;
+  if (args->finished) {
+    gpr_free(args);
+    return;
+  }
+  args->finished = true;
+  grpc_call_element *elem = args->elem;
+  channel_data *chand = elem->channel_data;
+  call_data *calld = elem->call_data;
+  // If we don't yet have a resolver result, then a closure for
+  // pick_after_resolver_result_done_locked() will have been added to
+  // chand->waiting_for_resolver_result_closures, and it may not be invoked
+  // until after this call has been destroyed.  We mark the operation as
+  // finished, so that when pick_after_resolver_result_done_locked()
+  // is called, it will be a no-op.  We also immediately invoke
+  // subchannel_ready_locked() to propagate the error back to the caller.
+  if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+    gpr_log(GPR_DEBUG,
+            "chand=%p calld=%p: cancelling pick waiting for resolver result",
+            chand, calld);
+  }
+  // Note: Although we are not in the call combiner here, we are
+  // basically stealing the call combiner from the pending pick, so
+  // it's safe to call subchannel_ready_locked() here -- we are
+  // essentially calling it here instead of calling it in
+  // pick_after_resolver_result_done_locked().
+  subchannel_ready_locked(exec_ctx, elem,
+                          GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
+                              "Pick cancelled", &error, 1));
+}
+
 static void pick_after_resolver_result_done_locked(grpc_exec_ctx *exec_ctx,
                                                    void *arg,
                                                    grpc_error *error) {
   pick_after_resolver_result_args *args = arg;
-  if (args->cancelled) {
+  if (args->finished) {
     /* cancelled, do nothing */
     if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
       gpr_log(GPR_DEBUG, "call cancelled before resolver result");
     }
+    gpr_free(args);
+    return;
+  }
+  args->finished = true;
+  grpc_call_element *elem = args->elem;
+  channel_data *chand = elem->channel_data;
+  call_data *calld = elem->call_data;
+  if (error != GRPC_ERROR_NONE) {
+    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+      gpr_log(GPR_DEBUG, "chand=%p calld=%p: resolver failed to return data",
+              chand, calld);
+    }
+    subchannel_ready_locked(exec_ctx, elem, GRPC_ERROR_REF(error));
   } else {
-    channel_data *chand = args->elem->channel_data;
-    call_data *calld = args->elem->call_data;
-    if (error != GRPC_ERROR_NONE) {
-      if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-        gpr_log(GPR_DEBUG, "chand=%p calld=%p: resolver failed to return data",
-                chand, calld);
-      }
-      subchannel_ready_locked(exec_ctx, args->elem, GRPC_ERROR_REF(error));
-    } else {
-      if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-        gpr_log(GPR_DEBUG, "chand=%p calld=%p: resolver returned, doing pick",
-                chand, calld);
-      }
-      if (pick_subchannel_locked(exec_ctx, args->elem)) {
-        subchannel_ready_locked(exec_ctx, args->elem, GRPC_ERROR_NONE);
-      }
+    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+      gpr_log(GPR_DEBUG, "chand=%p calld=%p: resolver returned, doing pick",
+              chand, calld);
+    }
+    if (pick_subchannel_locked(exec_ctx, elem)) {
+      subchannel_ready_locked(exec_ctx, elem, GRPC_ERROR_NONE);
     }
   }
-  gpr_free(args);
 }
 
 static void pick_after_resolver_result_start_locked(grpc_exec_ctx *exec_ctx,
@@ -1116,41 +1137,34 @@ static void pick_after_resolver_result_start_locked(grpc_exec_ctx *exec_ctx,
                     args, grpc_combiner_scheduler(chand->combiner));
   grpc_closure_list_append(&chand->waiting_for_resolver_result_closures,
                            &args->closure, GRPC_ERROR_NONE);
+  grpc_call_combiner_set_notify_on_cancel(
+      exec_ctx, calld->call_combiner,
+      GRPC_CLOSURE_INIT(&args->cancel_closure,
+                        pick_after_resolver_result_cancel_locked, args,
+                        grpc_combiner_scheduler(chand->combiner)));
 }
 
-static void pick_after_resolver_result_cancel_locked(grpc_exec_ctx *exec_ctx,
-                                                     grpc_call_element *elem,
-                                                     grpc_error *error) {
+// Note: This runs under the client_channel combiner, but will NOT be
+// holding the call combiner.
+static void pick_callback_cancel_locked(grpc_exec_ctx *exec_ctx, void *arg,
+                                        grpc_error *error) {
+  grpc_call_element *elem = arg;
   channel_data *chand = elem->channel_data;
   call_data *calld = elem->call_data;
-  // If we don't yet have a resolver result, then a closure for
-  // pick_after_resolver_result_done_locked() will have been added to
-  // chand->waiting_for_resolver_result_closures, and it may not be invoked
-  // until after this call has been destroyed.  We mark the operation as
-  // cancelled, so that when pick_after_resolver_result_done_locked()
-  // is called, it will be a no-op.  We also immediately invoke
-  // subchannel_ready_locked() to propagate the error back to the caller.
-  for (grpc_closure *closure = chand->waiting_for_resolver_result_closures.head;
-       closure != NULL; closure = closure->next_data.next) {
-    pick_after_resolver_result_args *args = closure->cb_arg;
-    if (!args->cancelled && args->elem == elem) {
-      if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-        gpr_log(GPR_DEBUG,
-                "chand=%p calld=%p: "
-                "cancelling pick waiting for resolver result",
-                chand, calld);
-      }
-      args->cancelled = true;
-      subchannel_ready_locked(exec_ctx, elem,
-                              GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
-                                  "Pick cancelled", &error, 1));
+  if (error != GRPC_ERROR_NONE && calld->lb_policy != NULL) {
+    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+      gpr_log(GPR_DEBUG, "chand=%p calld=%p: cancelling pick from LB policy %p",
+              chand, calld, calld->lb_policy);
     }
+    grpc_lb_policy_cancel_pick_locked(exec_ctx, calld->lb_policy,
+                                      &calld->connected_subchannel,
+                                      GRPC_ERROR_REF(error));
   }
-  GRPC_ERROR_UNREF(error);
+  GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call, "pick_callback_cancel");
 }
 
 // Callback invoked by grpc_lb_policy_pick_locked() for async picks.
-// Unrefs the LB policy after invoking subchannel_ready_locked().
+// Unrefs the LB policy and invokes subchannel_ready_locked().
 static void pick_callback_done_locked(grpc_exec_ctx *exec_ctx, void *arg,
                                       grpc_error *error) {
   grpc_call_element *elem = arg;
@@ -1194,24 +1208,17 @@ static bool pick_callback_start_locked(grpc_exec_ctx *exec_ctx,
     }
     GRPC_LB_POLICY_UNREF(exec_ctx, calld->lb_policy, "pick_subchannel");
     calld->lb_policy = NULL;
+  } else {
+    GRPC_CALL_STACK_REF(calld->owning_call, "pick_callback_cancel");
+    grpc_call_combiner_set_notify_on_cancel(
+        exec_ctx, calld->call_combiner,
+        GRPC_CLOSURE_INIT(&calld->lb_pick_cancel_closure,
+                          pick_callback_cancel_locked, elem,
+                          grpc_combiner_scheduler(chand->combiner)));
   }
   return pick_done;
 }
 
-static void pick_callback_cancel_locked(grpc_exec_ctx *exec_ctx,
-                                        grpc_call_element *elem,
-                                        grpc_error *error) {
-  channel_data *chand = elem->channel_data;
-  call_data *calld = elem->call_data;
-  GPR_ASSERT(calld->lb_policy != NULL);
-  if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-    gpr_log(GPR_DEBUG, "chand=%p calld=%p: cancelling pick from LB policy %p",
-            chand, calld, calld->lb_policy);
-  }
-  grpc_lb_policy_cancel_pick_locked(exec_ctx, calld->lb_policy,
-                                    &calld->connected_subchannel, error);
-}
-
 static bool pick_subchannel_locked(grpc_exec_ctx *exec_ctx,
                                    grpc_call_element *elem) {
   GPR_TIMER_BEGIN("pick_subchannel", 0);
@@ -1224,7 +1231,7 @@ static bool pick_subchannel_locked(grpc_exec_ctx *exec_ctx,
     // Otherwise, if the service config specified a value for this
     // method, use that.
     uint32_t initial_metadata_flags =
-        calld->initial_metadata_payload->send_initial_metadata
+        calld->initial_metadata_batch->payload->send_initial_metadata
             .send_initial_metadata_flags;
     const bool wait_for_ready_set_from_api =
         initial_metadata_flags &
@@ -1241,7 +1248,7 @@ static bool pick_subchannel_locked(grpc_exec_ctx *exec_ctx,
       }
     }
     const grpc_lb_policy_pick_args inputs = {
-        calld->initial_metadata_payload->send_initial_metadata
+        calld->initial_metadata_batch->payload->send_initial_metadata
             .send_initial_metadata,
         initial_metadata_flags, &calld->lb_token_mdelem};
     pick_done = pick_callback_start_locked(exec_ctx, elem, &inputs);
@@ -1258,91 +1265,33 @@ static bool pick_subchannel_locked(grpc_exec_ctx *exec_ctx,
   return pick_done;
 }
 
-static void start_transport_stream_op_batch_locked(grpc_exec_ctx *exec_ctx,
-                                                   void *arg,
-                                                   grpc_error *error_ignored) {
-  GPR_TIMER_BEGIN("start_transport_stream_op_batch_locked", 0);
-  grpc_transport_stream_op_batch *batch = arg;
-  grpc_call_element *elem = batch->handler_private.extra_arg;
-  call_data *calld = elem->call_data;
-  channel_data *chand = elem->channel_data;
-  /* need to recheck that another thread hasn't set the call */
-  call_or_error coe = get_call_or_error(calld);
-  if (coe.error != GRPC_ERROR_NONE) {
-    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-      gpr_log(GPR_DEBUG, "chand=%p calld=%p: failing batch with error: %s",
-              chand, calld, grpc_error_string(coe.error));
-    }
-    grpc_transport_stream_op_batch_finish_with_failure(
-        exec_ctx, batch, GRPC_ERROR_REF(coe.error));
-    goto done;
-  }
-  if (coe.subchannel_call != NULL) {
-    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-      gpr_log(GPR_DEBUG,
-              "chand=%p calld=%p: sending batch to subchannel_call=%p", chand,
-              calld, coe.subchannel_call);
-    }
-    grpc_subchannel_call_process_op(exec_ctx, coe.subchannel_call, batch);
-    goto done;
-  }
-  // Add to waiting-for-pick list.  If we succeed in getting a
-  // subchannel call below, we'll handle this batch (along with any
-  // other waiting batches) in waiting_for_pick_batches_resume_locked().
-  waiting_for_pick_batches_add_locked(calld, batch);
-  // If this is a cancellation, cancel the pending pick (if any) and
-  // fail any pending batches.
-  if (batch->cancel_stream) {
-    grpc_error *error = batch->payload->cancel_stream.cancel_error;
-    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-      gpr_log(GPR_DEBUG, "chand=%p calld=%p: recording cancel_error=%s", chand,
-              calld, grpc_error_string(error));
-    }
-    /* Stash a copy of cancel_error in our call data, so that we can use
-       it for subsequent operations.  This ensures that if the call is
-       cancelled before any batches are passed down (e.g., if the deadline
-       is in the past when the call starts), we can return the right
-       error to the caller when the first batch does get passed down. */
-    set_call_or_error(calld, (call_or_error){.error = GRPC_ERROR_REF(error)});
-    if (calld->lb_policy != NULL) {
-      pick_callback_cancel_locked(exec_ctx, elem, GRPC_ERROR_REF(error));
+static void start_pick_locked(grpc_exec_ctx *exec_ctx, void *arg,
+                              grpc_error *error_ignored) {
+  GPR_TIMER_BEGIN("start_pick_locked", 0);
+  grpc_call_element *elem = (grpc_call_element *)arg;
+  call_data *calld = (call_data *)elem->call_data;
+  channel_data *chand = (channel_data *)elem->channel_data;
+  GPR_ASSERT(calld->connected_subchannel == NULL);
+  if (pick_subchannel_locked(exec_ctx, elem)) {
+    // Pick was returned synchronously.
+    if (calld->connected_subchannel == NULL) {
+      GRPC_ERROR_UNREF(calld->error);
+      calld->error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
+          "Call dropped by load balancing policy");
+      waiting_for_pick_batches_fail(exec_ctx, elem,
+                                    GRPC_ERROR_REF(calld->error));
     } else {
-      pick_after_resolver_result_cancel_locked(exec_ctx, elem,
-                                               GRPC_ERROR_REF(error));
-    }
-    waiting_for_pick_batches_fail_locked(exec_ctx, elem, GRPC_ERROR_REF(error));
-    goto done;
-  }
-  /* if we don't have a subchannel, try to get one */
-  if (batch->send_initial_metadata) {
-    GPR_ASSERT(calld->connected_subchannel == NULL);
-    calld->initial_metadata_payload = batch->payload;
-    GRPC_CALL_STACK_REF(calld->owning_call, "pick_subchannel");
-    /* If a subchannel is not available immediately, the polling entity from
-       call_data should be provided to channel_data's interested_parties, so
-       that IO of the lb_policy and resolver could be done under it. */
-    if (pick_subchannel_locked(exec_ctx, elem)) {
-      // Pick was returned synchronously.
-      GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call, "pick_subchannel");
-      if (calld->connected_subchannel == NULL) {
-        grpc_error *error = GRPC_ERROR_CREATE_FROM_STATIC_STRING(
-            "Call dropped by load balancing policy");
-        set_call_or_error(calld,
-                          (call_or_error){.error = GRPC_ERROR_REF(error)});
-        waiting_for_pick_batches_fail_locked(exec_ctx, elem, error);
-      } else {
-        // Create subchannel call.
-        create_subchannel_call_locked(exec_ctx, elem, GRPC_ERROR_NONE);
-      }
-    } else {
-      grpc_polling_entity_add_to_pollset_set(exec_ctx, calld->pollent,
-                                             chand->interested_parties);
+      // Create subchannel call.
+      create_subchannel_call_locked(exec_ctx, elem, GRPC_ERROR_NONE);
     }
+  } else {
+    // Pick will be done asynchronously.  Add the call's polling entity to
+    // the channel's interested_parties, so that I/O for the resolver
+    // and LB policy can be done under it.
+    grpc_polling_entity_add_to_pollset_set(exec_ctx, calld->pollent,
+                                           chand->interested_parties);
   }
-done:
-  GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call,
-                        "start_transport_stream_op_batch");
-  GPR_TIMER_END("start_transport_stream_op_batch_locked", 0);
+  GPR_TIMER_END("start_pick_locked", 0);
 }
 
 static void on_complete(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) {
@@ -1365,27 +1314,49 @@ static void on_complete(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) {
                    GRPC_ERROR_REF(error));
 }
 
-/* The logic here is fairly complicated, due to (a) the fact that we
-   need to handle the case where we receive the send op before the
-   initial metadata op, and (b) the need for efficiency, especially in
-   the streaming case.
-
-   We use double-checked locking to initially see if initialization has been
-   performed. If it has not, we acquire the combiner and perform initialization.
-   If it has, we proceed on the fast path. */
 static void cc_start_transport_stream_op_batch(
     grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     grpc_transport_stream_op_batch *batch) {
   call_data *calld = elem->call_data;
   channel_data *chand = elem->channel_data;
-  if (GRPC_TRACER_ON(grpc_client_channel_trace) ||
-      GRPC_TRACER_ON(grpc_trace_channel)) {
-    grpc_call_log_op(GPR_INFO, elem, batch);
-  }
   if (chand->deadline_checking_enabled) {
     grpc_deadline_state_client_start_transport_stream_op_batch(exec_ctx, elem,
                                                                batch);
   }
+  GPR_TIMER_BEGIN("cc_start_transport_stream_op_batch", 0);
+  // If we've previously been cancelled, immediately fail any new batches.
+  if (calld->error != GRPC_ERROR_NONE) {
+    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+      gpr_log(GPR_DEBUG, "chand=%p calld=%p: failing batch with error: %s",
+              chand, calld, grpc_error_string(calld->error));
+    }
+    grpc_transport_stream_op_batch_finish_with_failure(
+        exec_ctx, batch, GRPC_ERROR_REF(calld->error), calld->call_combiner);
+    goto done;
+  }
+  if (batch->cancel_stream) {
+    // Stash a copy of cancel_error in our call data, so that we can use
+    // it for subsequent operations.  This ensures that if the call is
+    // cancelled before any batches are passed down (e.g., if the deadline
+    // is in the past when the call starts), we can return the right
+    // error to the caller when the first batch does get passed down.
+    GRPC_ERROR_UNREF(calld->error);
+    calld->error = GRPC_ERROR_REF(batch->payload->cancel_stream.cancel_error);
+    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+      gpr_log(GPR_DEBUG, "chand=%p calld=%p: recording cancel_error=%s", chand,
+              calld, grpc_error_string(calld->error));
+    }
+    // If we have a subchannel call, send the cancellation batch down.
+    // Otherwise, fail all pending batches.
+    if (calld->subchannel_call != NULL) {
+      grpc_subchannel_call_process_op(exec_ctx, calld->subchannel_call, batch);
+    } else {
+      waiting_for_pick_batches_add(calld, batch);
+      waiting_for_pick_batches_fail(exec_ctx, elem,
+                                    GRPC_ERROR_REF(calld->error));
+    }
+    goto done;
+  }
   // Intercept on_complete for recv_trailing_metadata so that we can
   // check retry throttle status.
   if (batch->recv_trailing_metadata) {
@@ -1395,38 +1366,43 @@ static void cc_start_transport_stream_op_batch(
                       grpc_schedule_on_exec_ctx);
     batch->on_complete = &calld->on_complete;
   }
-  /* try to (atomically) get the call */
-  call_or_error coe = get_call_or_error(calld);
-  GPR_TIMER_BEGIN("cc_start_transport_stream_op_batch", 0);
-  if (coe.error != GRPC_ERROR_NONE) {
+  // Check if we've already gotten a subchannel call.
+  // Note that once we have completed the pick, we do not need to enter
+  // the channel combiner, which is more efficient (especially for
+  // streaming calls).
+  if (calld->subchannel_call != NULL) {
     if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-      gpr_log(GPR_DEBUG, "chand=%p calld=%p: failing batch with error: %s",
-              chand, calld, grpc_error_string(coe.error));
+      gpr_log(GPR_DEBUG,
+              "chand=%p calld=%p: sending batch to subchannel_call=%p", chand,
+              calld, calld->subchannel_call);
     }
-    grpc_transport_stream_op_batch_finish_with_failure(
-        exec_ctx, batch, GRPC_ERROR_REF(coe.error));
+    grpc_subchannel_call_process_op(exec_ctx, calld->subchannel_call, batch);
     goto done;
   }
-  if (coe.subchannel_call != NULL) {
+  // We do not yet have a subchannel call.
+  // Add the batch to the waiting-for-pick list.
+  waiting_for_pick_batches_add(calld, batch);
+  // For batches containing a send_initial_metadata op, enter the channel
+  // combiner to start a pick.
+  if (batch->send_initial_metadata) {
+    if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
+      gpr_log(GPR_DEBUG, "chand=%p calld=%p: entering combiner", chand, calld);
+    }
+    GRPC_CLOSURE_SCHED(
+        exec_ctx,
+        GRPC_CLOSURE_INIT(&batch->handler_private.closure, start_pick_locked,
+                          elem, grpc_combiner_scheduler(chand->combiner)),
+        GRPC_ERROR_NONE);
+  } else {
+    // For all other batches, release the call combiner.
     if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
       gpr_log(GPR_DEBUG,
-              "chand=%p calld=%p: sending batch to subchannel_call=%p", chand,
-              calld, coe.subchannel_call);
+              "chand=%p calld=%p: saved batch, yeilding call combiner", chand,
+              calld);
     }
-    grpc_subchannel_call_process_op(exec_ctx, coe.subchannel_call, batch);
-    goto done;
-  }
-  /* we failed; lock and figure out what to do */
-  if (GRPC_TRACER_ON(grpc_client_channel_trace)) {
-    gpr_log(GPR_DEBUG, "chand=%p calld=%p: entering combiner", chand, calld);
+    GRPC_CALL_COMBINER_STOP(exec_ctx, calld->call_combiner,
+                            "batch does not include send_initial_metadata");
   }
-  GRPC_CALL_STACK_REF(calld->owning_call, "start_transport_stream_op_batch");
-  batch->handler_private.extra_arg = elem;
-  GRPC_CLOSURE_SCHED(
-      exec_ctx, GRPC_CLOSURE_INIT(&batch->handler_private.closure,
-                                  start_transport_stream_op_batch_locked, batch,
-                                  grpc_combiner_scheduler(chand->combiner)),
-      GRPC_ERROR_NONE);
 done:
   GPR_TIMER_END("cc_start_transport_stream_op_batch", 0);
 }
@@ -1441,10 +1417,12 @@ static grpc_error *cc_init_call_elem(grpc_exec_ctx *exec_ctx,
   calld->path = grpc_slice_ref_internal(args->path);
   calld->call_start_time = args->start_time;
   calld->deadline = gpr_convert_clock_type(args->deadline, GPR_CLOCK_MONOTONIC);
-  calld->owning_call = args->call_stack;
   calld->arena = args->arena;
+  calld->owning_call = args->call_stack;
+  calld->call_combiner = args->call_combiner;
   if (chand->deadline_checking_enabled) {
-    grpc_deadline_state_init(exec_ctx, elem, args->call_stack, calld->deadline);
+    grpc_deadline_state_init(exec_ctx, elem, args->call_stack,
+                             args->call_combiner, calld->deadline);
   }
   return GRPC_ERROR_NONE;
 }
@@ -1463,13 +1441,12 @@ static void cc_destroy_call_elem(grpc_exec_ctx *exec_ctx,
   if (calld->method_params != NULL) {
     method_parameters_unref(calld->method_params);
   }
-  call_or_error coe = get_call_or_error(calld);
-  GRPC_ERROR_UNREF(coe.error);
-  if (coe.subchannel_call != NULL) {
-    grpc_subchannel_call_set_cleanup_closure(coe.subchannel_call,
+  GRPC_ERROR_UNREF(calld->error);
+  if (calld->subchannel_call != NULL) {
+    grpc_subchannel_call_set_cleanup_closure(calld->subchannel_call,
                                              then_schedule_closure);
     then_schedule_closure = NULL;
-    GRPC_SUBCHANNEL_CALL_UNREF(exec_ctx, coe.subchannel_call,
+    GRPC_SUBCHANNEL_CALL_UNREF(exec_ctx, calld->subchannel_call,
                                "client_channel_destroy_call");
   }
   GPR_ASSERT(calld->lb_policy == NULL);
@@ -1508,7 +1485,6 @@ const grpc_channel_filter grpc_client_channel_filter = {
     sizeof(channel_data),
     cc_init_channel_elem,
     cc_destroy_channel_elem,
-    cc_get_peer,
     cc_get_channel_info,
     "client-channel",
 };
diff --git a/src/core/ext/filters/client_channel/lb_policy/grpclb/client_load_reporting_filter.c b/src/core/ext/filters/client_channel/lb_policy/grpclb/client_load_reporting_filter.c
index 568bb2ba8d..299f26b4de 100644
--- a/src/core/ext/filters/client_channel/lb_policy/grpclb/client_load_reporting_filter.c
+++ b/src/core/ext/filters/client_channel/lb_policy/grpclb/client_load_reporting_filter.c
@@ -132,6 +132,5 @@ const grpc_channel_filter grpc_client_load_reporting_filter = {
     0,  // sizeof(channel_data)
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "client_load_reporting"};
diff --git a/src/core/ext/filters/client_channel/resolver/fake/fake_resolver.c b/src/core/ext/filters/client_channel/resolver/fake/fake_resolver.c
index 56ed4371a9..3ff081a514 100644
--- a/src/core/ext/filters/client_channel/resolver/fake/fake_resolver.c
+++ b/src/core/ext/filters/client_channel/resolver/fake/fake_resolver.c
@@ -32,6 +32,7 @@
 #include "src/core/ext/filters/client_channel/parse_address.h"
 #include "src/core/ext/filters/client_channel/resolver_registry.h"
 #include "src/core/lib/channel/channel_args.h"
+#include "src/core/lib/iomgr/closure.h"
 #include "src/core/lib/iomgr/combiner.h"
 #include "src/core/lib/iomgr/resolve_address.h"
 #include "src/core/lib/iomgr/unix_sockets_posix.h"
@@ -125,7 +126,6 @@ static const grpc_resolver_vtable fake_resolver_vtable = {
 
 struct grpc_fake_resolver_response_generator {
   fake_resolver* resolver;  // Set by the fake_resolver constructor to itself.
-  grpc_channel_args* next_response;
   gpr_refcount refcount;
 };
 
@@ -151,19 +151,26 @@ void grpc_fake_resolver_response_generator_unref(
   }
 }
 
-static void set_response_cb(grpc_exec_ctx* exec_ctx, void* arg,
-                            grpc_error* error) {
-  grpc_fake_resolver_response_generator* generator =
-      (grpc_fake_resolver_response_generator*)arg;
+typedef struct set_response_closure_arg {
+  grpc_closure set_response_closure;
+  grpc_fake_resolver_response_generator* generator;
+  grpc_channel_args* next_response;
+} set_response_closure_arg;
+
+static void set_response_closure_fn(grpc_exec_ctx* exec_ctx, void* arg,
+                                    grpc_error* error) {
+  set_response_closure_arg* closure_arg = arg;
+  grpc_fake_resolver_response_generator* generator = closure_arg->generator;
   fake_resolver* r = generator->resolver;
   if (r->next_results != NULL) {
     grpc_channel_args_destroy(exec_ctx, r->next_results);
   }
-  r->next_results = generator->next_response;
+  r->next_results = closure_arg->next_response;
   if (r->results_upon_error != NULL) {
     grpc_channel_args_destroy(exec_ctx, r->results_upon_error);
   }
-  r->results_upon_error = grpc_channel_args_copy(generator->next_response);
+  r->results_upon_error = grpc_channel_args_copy(closure_arg->next_response);
+  gpr_free(closure_arg);
   fake_resolver_maybe_finish_next_locked(exec_ctx, r);
 }
 
@@ -171,12 +178,15 @@ void grpc_fake_resolver_response_generator_set_response(
     grpc_exec_ctx* exec_ctx, grpc_fake_resolver_response_generator* generator,
     grpc_channel_args* next_response) {
   GPR_ASSERT(generator->resolver != NULL);
-  generator->next_response = grpc_channel_args_copy(next_response);
-  GRPC_CLOSURE_SCHED(
-      exec_ctx, GRPC_CLOSURE_CREATE(set_response_cb, generator,
-                                    grpc_combiner_scheduler(
-                                        generator->resolver->base.combiner)),
-      GRPC_ERROR_NONE);
+  set_response_closure_arg* closure_arg = gpr_zalloc(sizeof(*closure_arg));
+  closure_arg->generator = generator;
+  closure_arg->next_response = grpc_channel_args_copy(next_response);
+  GRPC_CLOSURE_SCHED(exec_ctx,
+                     GRPC_CLOSURE_INIT(&closure_arg->set_response_closure,
+                                       set_response_closure_fn, closure_arg,
+                                       grpc_combiner_scheduler(
+                                           generator->resolver->base.combiner)),
+                     GRPC_ERROR_NONE);
 }
 
 static void* response_generator_arg_copy(void* p) {
diff --git a/src/core/ext/filters/client_channel/subchannel.c b/src/core/ext/filters/client_channel/subchannel.c
index 5788819331..6b5b383efd 100644
--- a/src/core/ext/filters/client_channel/subchannel.c
+++ b/src/core/ext/filters/client_channel/subchannel.c
@@ -724,20 +724,14 @@ void grpc_subchannel_call_unref(grpc_exec_ctx *exec_ctx,
   GRPC_CALL_STACK_UNREF(exec_ctx, SUBCHANNEL_CALL_TO_CALL_STACK(c), REF_REASON);
 }
 
-char *grpc_subchannel_call_get_peer(grpc_exec_ctx *exec_ctx,
-                                    grpc_subchannel_call *call) {
-  grpc_call_stack *call_stack = SUBCHANNEL_CALL_TO_CALL_STACK(call);
-  grpc_call_element *top_elem = grpc_call_stack_element(call_stack, 0);
-  return top_elem->filter->get_peer(exec_ctx, top_elem);
-}
-
 void grpc_subchannel_call_process_op(grpc_exec_ctx *exec_ctx,
                                      grpc_subchannel_call *call,
-                                     grpc_transport_stream_op_batch *op) {
+                                     grpc_transport_stream_op_batch *batch) {
   GPR_TIMER_BEGIN("grpc_subchannel_call_process_op", 0);
   grpc_call_stack *call_stack = SUBCHANNEL_CALL_TO_CALL_STACK(call);
   grpc_call_element *top_elem = grpc_call_stack_element(call_stack, 0);
-  top_elem->filter->start_transport_stream_op_batch(exec_ctx, top_elem, op);
+  GRPC_CALL_LOG_OP(GPR_INFO, top_elem, batch);
+  top_elem->filter->start_transport_stream_op_batch(exec_ctx, top_elem, batch);
   GPR_TIMER_END("grpc_subchannel_call_process_op", 0);
 }
 
@@ -760,13 +754,15 @@ grpc_error *grpc_connected_subchannel_create_call(
       args->arena, sizeof(grpc_subchannel_call) + chanstk->call_stack_size);
   grpc_call_stack *callstk = SUBCHANNEL_CALL_TO_CALL_STACK(*call);
   (*call)->connection = GRPC_CONNECTED_SUBCHANNEL_REF(con, "subchannel_call");
-  const grpc_call_element_args call_args = {.call_stack = callstk,
-                                            .server_transport_data = NULL,
-                                            .context = args->context,
-                                            .path = args->path,
-                                            .start_time = args->start_time,
-                                            .deadline = args->deadline,
-                                            .arena = args->arena};
+  const grpc_call_element_args call_args = {
+      .call_stack = callstk,
+      .server_transport_data = NULL,
+      .context = args->context,
+      .path = args->path,
+      .start_time = args->start_time,
+      .deadline = args->deadline,
+      .arena = args->arena,
+      .call_combiner = args->call_combiner};
   grpc_error *error = grpc_call_stack_init(
       exec_ctx, chanstk, 1, subchannel_call_destroy, *call, &call_args);
   if (error != GRPC_ERROR_NONE) {
diff --git a/src/core/ext/filters/client_channel/subchannel.h b/src/core/ext/filters/client_channel/subchannel.h
index 6d2abb04df..51d712f6a7 100644
--- a/src/core/ext/filters/client_channel/subchannel.h
+++ b/src/core/ext/filters/client_channel/subchannel.h
@@ -106,6 +106,7 @@ typedef struct {
   gpr_timespec deadline;
   gpr_arena *arena;
   grpc_call_context_element *context;
+  grpc_call_combiner *call_combiner;
 } grpc_connected_subchannel_call_args;
 
 grpc_error *grpc_connected_subchannel_create_call(
@@ -150,10 +151,6 @@ void grpc_subchannel_call_process_op(grpc_exec_ctx *exec_ctx,
                                      grpc_subchannel_call *subchannel_call,
                                      grpc_transport_stream_op_batch *op);
 
-/** continue querying for peer */
-char *grpc_subchannel_call_get_peer(grpc_exec_ctx *exec_ctx,
-                                    grpc_subchannel_call *subchannel_call);
-
 /** Must be called once per call. Sets the 'then_schedule_closure' argument for
     call stack destruction. */
 void grpc_subchannel_call_set_cleanup_closure(
diff --git a/src/core/ext/filters/deadline/deadline_filter.c b/src/core/ext/filters/deadline/deadline_filter.c
index 6789903c95..565b0679dc 100644
--- a/src/core/ext/filters/deadline/deadline_filter.c
+++ b/src/core/ext/filters/deadline/deadline_filter.c
@@ -34,22 +34,56 @@
 // grpc_deadline_state
 //
 
+// The on_complete callback used when sending a cancel_error batch down the
+// filter stack.  Yields the call combiner when the batch returns.
+static void yield_call_combiner(grpc_exec_ctx* exec_ctx, void* arg,
+                                grpc_error* ignored) {
+  grpc_deadline_state* deadline_state = arg;
+  GRPC_CALL_COMBINER_STOP(exec_ctx, deadline_state->call_combiner,
+                          "got on_complete from cancel_stream batch");
+  GRPC_CALL_STACK_UNREF(exec_ctx, deadline_state->call_stack, "deadline_timer");
+}
+
+// This is called via the call combiner, so access to deadline_state is
+// synchronized.
+static void send_cancel_op_in_call_combiner(grpc_exec_ctx* exec_ctx, void* arg,
+                                            grpc_error* error) {
+  grpc_call_element* elem = arg;
+  grpc_deadline_state* deadline_state = elem->call_data;
+  grpc_transport_stream_op_batch* batch = grpc_make_transport_stream_op(
+      GRPC_CLOSURE_INIT(&deadline_state->timer_callback, yield_call_combiner,
+                        deadline_state, grpc_schedule_on_exec_ctx));
+  batch->cancel_stream = true;
+  batch->payload->cancel_stream.cancel_error = GRPC_ERROR_REF(error);
+  elem->filter->start_transport_stream_op_batch(exec_ctx, elem, batch);
+}
+
 // Timer callback.
 static void timer_callback(grpc_exec_ctx* exec_ctx, void* arg,
                            grpc_error* error) {
   grpc_call_element* elem = (grpc_call_element*)arg;
   grpc_deadline_state* deadline_state = (grpc_deadline_state*)elem->call_data;
   if (error != GRPC_ERROR_CANCELLED) {
-    grpc_call_element_signal_error(
-        exec_ctx, elem,
-        grpc_error_set_int(
-            GRPC_ERROR_CREATE_FROM_STATIC_STRING("Deadline Exceeded"),
-            GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_DEADLINE_EXCEEDED));
+    error = grpc_error_set_int(
+        GRPC_ERROR_CREATE_FROM_STATIC_STRING("Deadline Exceeded"),
+        GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_DEADLINE_EXCEEDED);
+    grpc_call_combiner_cancel(exec_ctx, deadline_state->call_combiner,
+                              GRPC_ERROR_REF(error));
+    GRPC_CLOSURE_INIT(&deadline_state->timer_callback,
+                      send_cancel_op_in_call_combiner, elem,
+                      grpc_schedule_on_exec_ctx);
+    GRPC_CALL_COMBINER_START(exec_ctx, deadline_state->call_combiner,
+                             &deadline_state->timer_callback, error,
+                             "deadline exceeded -- sending cancel_stream op");
+  } else {
+    GRPC_CALL_STACK_UNREF(exec_ctx, deadline_state->call_stack,
+                          "deadline_timer");
   }
-  GRPC_CALL_STACK_UNREF(exec_ctx, deadline_state->call_stack, "deadline_timer");
 }
 
 // Starts the deadline timer.
+// This is called via the call combiner, so access to deadline_state is
+// synchronized.
 static void start_timer_if_needed(grpc_exec_ctx* exec_ctx,
                                   grpc_call_element* elem,
                                   gpr_timespec deadline) {
@@ -58,51 +92,39 @@ static void start_timer_if_needed(grpc_exec_ctx* exec_ctx,
     return;
   }
   grpc_deadline_state* deadline_state = (grpc_deadline_state*)elem->call_data;
-  grpc_deadline_timer_state cur_state;
   grpc_closure* closure = NULL;
-retry:
-  cur_state =
-      (grpc_deadline_timer_state)gpr_atm_acq_load(&deadline_state->timer_state);
-  switch (cur_state) {
+  switch (deadline_state->timer_state) {
     case GRPC_DEADLINE_STATE_PENDING:
       // Note: We do not start the timer if there is already a timer
       return;
     case GRPC_DEADLINE_STATE_FINISHED:
-      if (gpr_atm_rel_cas(&deadline_state->timer_state,
-                          GRPC_DEADLINE_STATE_FINISHED,
-                          GRPC_DEADLINE_STATE_PENDING)) {
-        // If we've already created and destroyed a timer, we always create a
-        // new closure: we have no other guarantee that the inlined closure is
-        // not in use (it may hold a pending call to timer_callback)
-        closure = GRPC_CLOSURE_CREATE(timer_callback, elem,
-                                      grpc_schedule_on_exec_ctx);
-      } else {
-        goto retry;
-      }
+      deadline_state->timer_state = GRPC_DEADLINE_STATE_PENDING;
+      // If we've already created and destroyed a timer, we always create a
+      // new closure: we have no other guarantee that the inlined closure is
+      // not in use (it may hold a pending call to timer_callback)
+      closure =
+          GRPC_CLOSURE_CREATE(timer_callback, elem, grpc_schedule_on_exec_ctx);
       break;
     case GRPC_DEADLINE_STATE_INITIAL:
-      if (gpr_atm_rel_cas(&deadline_state->timer_state,
-                          GRPC_DEADLINE_STATE_INITIAL,
-                          GRPC_DEADLINE_STATE_PENDING)) {
-        closure =
-            GRPC_CLOSURE_INIT(&deadline_state->timer_callback, timer_callback,
-                              elem, grpc_schedule_on_exec_ctx);
-      } else {
-        goto retry;
-      }
+      deadline_state->timer_state = GRPC_DEADLINE_STATE_PENDING;
+      closure =
+          GRPC_CLOSURE_INIT(&deadline_state->timer_callback, timer_callback,
+                            elem, grpc_schedule_on_exec_ctx);
       break;
   }
-  GPR_ASSERT(closure);
+  GPR_ASSERT(closure != NULL);
   GRPC_CALL_STACK_REF(deadline_state->call_stack, "deadline_timer");
   grpc_timer_init(exec_ctx, &deadline_state->timer, deadline, closure,
                   gpr_now(GPR_CLOCK_MONOTONIC));
 }
 
 // Cancels the deadline timer.
+// This is called via the call combiner, so access to deadline_state is
+// synchronized.
 static void cancel_timer_if_needed(grpc_exec_ctx* exec_ctx,
                                    grpc_deadline_state* deadline_state) {
-  if (gpr_atm_rel_cas(&deadline_state->timer_state, GRPC_DEADLINE_STATE_PENDING,
-                      GRPC_DEADLINE_STATE_FINISHED)) {
+  if (deadline_state->timer_state == GRPC_DEADLINE_STATE_PENDING) {
+    deadline_state->timer_state = GRPC_DEADLINE_STATE_FINISHED;
     grpc_timer_cancel(exec_ctx, &deadline_state->timer);
   } else {
     // timer was either in STATE_INITAL (nothing to cancel)
@@ -131,6 +153,7 @@ static void inject_on_complete_cb(grpc_deadline_state* deadline_state,
 // Callback and associated state for starting the timer after call stack
 // initialization has been completed.
 struct start_timer_after_init_state {
+  bool in_call_combiner;
   grpc_call_element* elem;
   gpr_timespec deadline;
   grpc_closure closure;
@@ -138,15 +161,29 @@ struct start_timer_after_init_state {
 static void start_timer_after_init(grpc_exec_ctx* exec_ctx, void* arg,
                                    grpc_error* error) {
   struct start_timer_after_init_state* state = arg;
+  grpc_deadline_state* deadline_state = state->elem->call_data;
+  if (!state->in_call_combiner) {
+    // We are initially called without holding the call combiner, so we
+    // need to bounce ourselves into it.
+    state->in_call_combiner = true;
+    GRPC_CALL_COMBINER_START(exec_ctx, deadline_state->call_combiner,
+                             &state->closure, GRPC_ERROR_REF(error),
+                             "scheduling deadline timer");
+    return;
+  }
   start_timer_if_needed(exec_ctx, state->elem, state->deadline);
   gpr_free(state);
+  GRPC_CALL_COMBINER_STOP(exec_ctx, deadline_state->call_combiner,
+                          "done scheduling deadline timer");
 }
 
 void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
                               grpc_call_stack* call_stack,
+                              grpc_call_combiner* call_combiner,
                               gpr_timespec deadline) {
   grpc_deadline_state* deadline_state = (grpc_deadline_state*)elem->call_data;
   deadline_state->call_stack = call_stack;
+  deadline_state->call_combiner = call_combiner;
   // Deadline will always be infinite on servers, so the timer will only be
   // set on clients with a finite deadline.
   deadline = gpr_convert_clock_type(deadline, GPR_CLOCK_MONOTONIC);
@@ -158,7 +195,7 @@ void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
     // call stack initialization is finished.  To avoid that problem, we
     // create a closure to start the timer, and we schedule that closure
     // to be run after call stack initialization is done.
-    struct start_timer_after_init_state* state = gpr_malloc(sizeof(*state));
+    struct start_timer_after_init_state* state = gpr_zalloc(sizeof(*state));
     state->elem = elem;
     state->deadline = deadline;
     GRPC_CLOSURE_INIT(&state->closure, start_timer_after_init, state,
@@ -232,7 +269,8 @@ typedef struct server_call_data {
 static grpc_error* init_call_elem(grpc_exec_ctx* exec_ctx,
                                   grpc_call_element* elem,
                                   const grpc_call_element_args* args) {
-  grpc_deadline_state_init(exec_ctx, elem, args->call_stack, args->deadline);
+  grpc_deadline_state_init(exec_ctx, elem, args->call_stack,
+                           args->call_combiner, args->deadline);
   return GRPC_ERROR_NONE;
 }
 
@@ -310,7 +348,6 @@ const grpc_channel_filter grpc_client_deadline_filter = {
     0,  // sizeof(channel_data)
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "deadline",
 };
@@ -325,7 +362,6 @@ const grpc_channel_filter grpc_server_deadline_filter = {
     0,  // sizeof(channel_data)
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "deadline",
 };
diff --git a/src/core/ext/filters/deadline/deadline_filter.h b/src/core/ext/filters/deadline/deadline_filter.h
index 420bf7065a..3eb102ad28 100644
--- a/src/core/ext/filters/deadline/deadline_filter.h
+++ b/src/core/ext/filters/deadline/deadline_filter.h
@@ -31,7 +31,8 @@ typedef enum grpc_deadline_timer_state {
 typedef struct grpc_deadline_state {
   // We take a reference to the call stack for the timer callback.
   grpc_call_stack* call_stack;
-  gpr_atm timer_state;
+  grpc_call_combiner* call_combiner;
+  grpc_deadline_timer_state timer_state;
   grpc_timer timer;
   grpc_closure timer_callback;
   // Closure to invoke when the call is complete.
@@ -50,6 +51,7 @@ typedef struct grpc_deadline_state {
 // assumes elem->call_data is zero'd
 void grpc_deadline_state_init(grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
                               grpc_call_stack* call_stack,
+                              grpc_call_combiner* call_combiner,
                               gpr_timespec deadline);
 void grpc_deadline_state_destroy(grpc_exec_ctx* exec_ctx,
                                  grpc_call_element* elem);
@@ -61,6 +63,8 @@ void grpc_deadline_state_destroy(grpc_exec_ctx* exec_ctx,
 // to ensure that the timer callback is not invoked while it is in the
 // process of being reset, which means that attempting to increase the
 // deadline may result in the timer being called twice.
+//
+// Note: Must be called while holding the call combiner.
 void grpc_deadline_state_reset(grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
                                gpr_timespec new_deadline);
 
@@ -70,6 +74,8 @@ void grpc_deadline_state_reset(grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
 //
 // Note: It is the caller's responsibility to chain to the next filter if
 // necessary after this function returns.
+//
+// Note: Must be called while holding the call combiner.
 void grpc_deadline_state_client_start_transport_stream_op_batch(
     grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
     grpc_transport_stream_op_batch* op);
diff --git a/src/core/ext/filters/http/client/http_client_filter.c b/src/core/ext/filters/http/client/http_client_filter.c
index 3ca01a41b5..2d7429c41e 100644
--- a/src/core/ext/filters/http/client/http_client_filter.c
+++ b/src/core/ext/filters/http/client/http_client_filter.c
@@ -36,6 +36,7 @@
 static const size_t kMaxPayloadSizeForGet = 2048;
 
 typedef struct call_data {
+  grpc_call_combiner *call_combiner;
   // State for handling send_initial_metadata ops.
   grpc_linked_mdelem method;
   grpc_linked_mdelem scheme;
@@ -215,13 +216,13 @@ static void on_send_message_next_done(grpc_exec_ctx *exec_ctx, void *arg,
   call_data *calld = (call_data *)elem->call_data;
   if (error != GRPC_ERROR_NONE) {
     grpc_transport_stream_op_batch_finish_with_failure(
-        exec_ctx, calld->send_message_batch, error);
+        exec_ctx, calld->send_message_batch, error, calld->call_combiner);
     return;
   }
   error = pull_slice_from_send_message(exec_ctx, calld);
   if (error != GRPC_ERROR_NONE) {
     grpc_transport_stream_op_batch_finish_with_failure(
-        exec_ctx, calld->send_message_batch, error);
+        exec_ctx, calld->send_message_batch, error, calld->call_combiner);
     return;
   }
   // There may or may not be more to read, but we don't care.  If we got
@@ -302,7 +303,6 @@ static void hc_start_transport_stream_op_batch(
   call_data *calld = elem->call_data;
   channel_data *channeld = elem->channel_data;
   GPR_TIMER_BEGIN("hc_start_transport_stream_op_batch", 0);
-  GRPC_CALL_LOG_OP(GPR_INFO, elem, batch);
 
   if (batch->recv_initial_metadata) {
     /* substitute our callback for the higher callback */
@@ -414,7 +414,7 @@ static void hc_start_transport_stream_op_batch(
 done:
   if (error != GRPC_ERROR_NONE) {
     grpc_transport_stream_op_batch_finish_with_failure(
-        exec_ctx, calld->send_message_batch, error);
+        exec_ctx, calld->send_message_batch, error, calld->call_combiner);
   } else if (!batch_will_be_handled_asynchronously) {
     grpc_call_next_op(exec_ctx, elem, batch);
   }
@@ -426,6 +426,7 @@ static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
                                   grpc_call_element *elem,
                                   const grpc_call_element_args *args) {
   call_data *calld = (call_data *)elem->call_data;
+  calld->call_combiner = args->call_combiner;
   GRPC_CLOSURE_INIT(&calld->recv_initial_metadata_ready,
                     recv_initial_metadata_ready, elem,
                     grpc_schedule_on_exec_ctx);
@@ -565,6 +566,5 @@ const grpc_channel_filter grpc_http_client_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "http-client"};
diff --git a/src/core/ext/filters/http/message_compress/message_compress_filter.c b/src/core/ext/filters/http/message_compress/message_compress_filter.c
index a32819bfe4..98a503cafc 100644
--- a/src/core/ext/filters/http/message_compress/message_compress_filter.c
+++ b/src/core/ext/filters/http/message_compress/message_compress_filter.c
@@ -35,35 +35,29 @@
 #include "src/core/lib/surface/call.h"
 #include "src/core/lib/transport/static_metadata.h"
 
-#define INITIAL_METADATA_UNSEEN 0
-#define HAS_COMPRESSION_ALGORITHM 2
-#define NO_COMPRESSION_ALGORITHM 4
-
-#define CANCELLED_BIT ((gpr_atm)1)
+typedef enum {
+  // Initial metadata not yet seen.
+  INITIAL_METADATA_UNSEEN = 0,
+  // Initial metadata seen; compression algorithm set.
+  HAS_COMPRESSION_ALGORITHM,
+  // Initial metadata seen; no compression algorithm set.
+  NO_COMPRESSION_ALGORITHM,
+} initial_metadata_state;
 
 typedef struct call_data {
-  grpc_slice_buffer slices; /**< Buffers up input slices to be compressed */
+  grpc_call_combiner *call_combiner;
   grpc_linked_mdelem compression_algorithm_storage;
   grpc_linked_mdelem stream_compression_algorithm_storage;
   grpc_linked_mdelem accept_encoding_storage;
   grpc_linked_mdelem accept_stream_encoding_storage;
-  uint32_t remaining_slice_bytes;
   /** Compression algorithm we'll try to use. It may be given by incoming
    * metadata, or by the channel's default compression settings. */
   grpc_compression_algorithm compression_algorithm;
-
-  /* Atomic recording the state of initial metadata; allowed values:
-     INITIAL_METADATA_UNSEEN - initial metadata op not seen
-     HAS_COMPRESSION_ALGORITHM - initial metadata seen; compression algorithm
-                                 set
-     NO_COMPRESSION_ALGORITHM - initial metadata seen; no compression algorithm
-                                set
-     pointer - a stalled op containing a send_message that's waiting on initial
-               metadata
-     pointer | CANCELLED_BIT - request was cancelled with error pointed to */
-  gpr_atm send_initial_metadata_state;
-
+  initial_metadata_state send_initial_metadata_state;
+  grpc_error *cancel_error;
+  grpc_closure start_send_message_batch_in_call_combiner;
   grpc_transport_stream_op_batch *send_message_batch;
+  grpc_slice_buffer slices; /**< Buffers up input slices to be compressed */
   grpc_slice_buffer_stream replacement_stream;
   grpc_closure *original_send_message_on_complete;
   grpc_closure send_message_on_complete;
@@ -92,13 +86,13 @@ static bool skip_compression(grpc_call_element *elem, uint32_t flags,
   channel_data *channeld = elem->channel_data;
 
   if (flags & (GRPC_WRITE_NO_COMPRESS | GRPC_WRITE_INTERNAL_COMPRESS)) {
-    return 1;
+    return true;
   }
   if (has_compression_algorithm) {
     if (calld->compression_algorithm == GRPC_COMPRESS_NONE) {
-      return 1;
+      return true;
     }
-    return 0; /* we have an actual call-specific algorithm */
+    return false; /* we have an actual call-specific algorithm */
   }
   /* no per-call compression override */
   return channeld->default_compression_algorithm == GRPC_COMPRESS_NONE;
@@ -226,6 +220,18 @@ static void send_message_on_complete(grpc_exec_ctx *exec_ctx, void *arg,
                    GRPC_ERROR_REF(error));
 }
 
+static void send_message_batch_continue(grpc_exec_ctx *exec_ctx,
+                                        grpc_call_element *elem) {
+  call_data *calld = (call_data *)elem->call_data;
+  // Note: The call to grpc_call_next_op() results in yielding the
+  // call combiner, so we need to clear calld->send_message_batch
+  // before we do that.
+  grpc_transport_stream_op_batch *send_message_batch =
+      calld->send_message_batch;
+  calld->send_message_batch = NULL;
+  grpc_call_next_op(exec_ctx, elem, send_message_batch);
+}
+
 static void finish_send_message(grpc_exec_ctx *exec_ctx,
                                 grpc_call_element *elem) {
   call_data *calld = (call_data *)elem->call_data;
@@ -234,8 +240,8 @@ static void finish_send_message(grpc_exec_ctx *exec_ctx,
   grpc_slice_buffer_init(&tmp);
   uint32_t send_flags =
       calld->send_message_batch->payload->send_message.send_message->flags;
-  const bool did_compress = grpc_msg_compress(
-      exec_ctx, calld->compression_algorithm, &calld->slices, &tmp);
+  bool did_compress = grpc_msg_compress(exec_ctx, calld->compression_algorithm,
+                                        &calld->slices, &tmp);
   if (did_compress) {
     if (GRPC_TRACER_ON(grpc_compression_trace)) {
       char *algo_name;
@@ -273,7 +279,19 @@ static void finish_send_message(grpc_exec_ctx *exec_ctx,
   calld->original_send_message_on_complete =
       calld->send_message_batch->on_complete;
   calld->send_message_batch->on_complete = &calld->send_message_on_complete;
-  grpc_call_next_op(exec_ctx, elem, calld->send_message_batch);
+  send_message_batch_continue(exec_ctx, elem);
+}
+
+static void fail_send_message_batch_in_call_combiner(grpc_exec_ctx *exec_ctx,
+                                                     void *arg,
+                                                     grpc_error *error) {
+  call_data *calld = arg;
+  if (calld->send_message_batch != NULL) {
+    grpc_transport_stream_op_batch_finish_with_failure(
+        exec_ctx, calld->send_message_batch, GRPC_ERROR_REF(error),
+        calld->call_combiner);
+    calld->send_message_batch = NULL;
+  }
 }
 
 // Pulls a slice from the send_message byte stream and adds it to calld->slices.
@@ -293,21 +311,25 @@ static grpc_error *pull_slice_from_send_message(grpc_exec_ctx *exec_ctx,
 // If all data has been read, invokes finish_send_message().  Otherwise,
 // an async call to grpc_byte_stream_next() has been started, which will
 // eventually result in calling on_send_message_next_done().
-static grpc_error *continue_reading_send_message(grpc_exec_ctx *exec_ctx,
-                                                 grpc_call_element *elem) {
+static void continue_reading_send_message(grpc_exec_ctx *exec_ctx,
+                                          grpc_call_element *elem) {
   call_data *calld = (call_data *)elem->call_data;
   while (grpc_byte_stream_next(
       exec_ctx, calld->send_message_batch->payload->send_message.send_message,
       ~(size_t)0, &calld->on_send_message_next_done)) {
     grpc_error *error = pull_slice_from_send_message(exec_ctx, calld);
-    if (error != GRPC_ERROR_NONE) return error;
+    if (error != GRPC_ERROR_NONE) {
+      // Closure callback; does not take ownership of error.
+      fail_send_message_batch_in_call_combiner(exec_ctx, calld, error);
+      GRPC_ERROR_UNREF(error);
+      return;
+    }
     if (calld->slices.length ==
         calld->send_message_batch->payload->send_message.send_message->length) {
       finish_send_message(exec_ctx, elem);
       break;
     }
   }
-  return GRPC_ERROR_NONE;
 }
 
 // Async callback for grpc_byte_stream_next().
@@ -315,46 +337,37 @@ static void on_send_message_next_done(grpc_exec_ctx *exec_ctx, void *arg,
                                       grpc_error *error) {
   grpc_call_element *elem = (grpc_call_element *)arg;
   call_data *calld = (call_data *)elem->call_data;
-  if (error != GRPC_ERROR_NONE) goto fail;
+  if (error != GRPC_ERROR_NONE) {
+    // Closure callback; does not take ownership of error.
+    fail_send_message_batch_in_call_combiner(exec_ctx, calld, error);
+    return;
+  }
   error = pull_slice_from_send_message(exec_ctx, calld);
-  if (error != GRPC_ERROR_NONE) goto fail;
+  if (error != GRPC_ERROR_NONE) {
+    // Closure callback; does not take ownership of error.
+    fail_send_message_batch_in_call_combiner(exec_ctx, calld, error);
+    GRPC_ERROR_UNREF(error);
+    return;
+  }
   if (calld->slices.length ==
       calld->send_message_batch->payload->send_message.send_message->length) {
     finish_send_message(exec_ctx, elem);
   } else {
-    // This will either finish reading all of the data and invoke
-    // finish_send_message(), or else it will make an async call to
-    // grpc_byte_stream_next(), which will eventually result in calling
-    // this function again.
-    error = continue_reading_send_message(exec_ctx, elem);
-    if (error != GRPC_ERROR_NONE) goto fail;
+    continue_reading_send_message(exec_ctx, elem);
   }
-  return;
-fail:
-  grpc_transport_stream_op_batch_finish_with_failure(
-      exec_ctx, calld->send_message_batch, error);
 }
 
-static void start_send_message_batch(grpc_exec_ctx *exec_ctx,
-                                     grpc_call_element *elem,
-                                     grpc_transport_stream_op_batch *batch,
-                                     bool has_compression_algorithm) {
+static void start_send_message_batch(grpc_exec_ctx *exec_ctx, void *arg,
+                                     grpc_error *unused) {
+  grpc_call_element *elem = (grpc_call_element *)arg;
   call_data *calld = (call_data *)elem->call_data;
-  if (!skip_compression(elem, batch->payload->send_message.send_message->flags,
-                        has_compression_algorithm)) {
-    calld->send_message_batch = batch;
-    // This will either finish reading all of the data and invoke
-    // finish_send_message(), or else it will make an async call to
-    // grpc_byte_stream_next(), which will eventually result in calling
-    // on_send_message_next_done().
-    grpc_error *error = continue_reading_send_message(exec_ctx, elem);
-    if (error != GRPC_ERROR_NONE) {
-      grpc_transport_stream_op_batch_finish_with_failure(
-          exec_ctx, calld->send_message_batch, error);
-    }
+  if (skip_compression(
+          elem,
+          calld->send_message_batch->payload->send_message.send_message->flags,
+          calld->send_initial_metadata_state == HAS_COMPRESSION_ALGORITHM)) {
+    send_message_batch_continue(exec_ctx, elem);
   } else {
-    /* pass control down the stack */
-    grpc_call_next_op(exec_ctx, elem, batch);
+    continue_reading_send_message(exec_ctx, elem);
   }
 }
 
@@ -362,95 +375,80 @@ static void compress_start_transport_stream_op_batch(
     grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     grpc_transport_stream_op_batch *batch) {
   call_data *calld = elem->call_data;
-
   GPR_TIMER_BEGIN("compress_start_transport_stream_op_batch", 0);
-
+  // Handle cancel_stream.
   if (batch->cancel_stream) {
-    // TODO(roth): As part of the upcoming call combiner work, change
-    // this to call grpc_byte_stream_shutdown() on the incoming byte
-    // stream, to cancel any in-flight calls to grpc_byte_stream_next().
-    GRPC_ERROR_REF(batch->payload->cancel_stream.cancel_error);
-    gpr_atm cur = gpr_atm_full_xchg(
-        &calld->send_initial_metadata_state,
-        CANCELLED_BIT | (gpr_atm)batch->payload->cancel_stream.cancel_error);
-    switch (cur) {
-      case HAS_COMPRESSION_ALGORITHM:
-      case NO_COMPRESSION_ALGORITHM:
-      case INITIAL_METADATA_UNSEEN:
-        break;
-      default:
-        if ((cur & CANCELLED_BIT) == 0) {
-          grpc_transport_stream_op_batch_finish_with_failure(
-              exec_ctx, (grpc_transport_stream_op_batch *)cur,
-              GRPC_ERROR_REF(batch->payload->cancel_stream.cancel_error));
-        } else {
-          GRPC_ERROR_UNREF((grpc_error *)(cur & ~CANCELLED_BIT));
-        }
-        break;
+    GRPC_ERROR_UNREF(calld->cancel_error);
+    calld->cancel_error =
+        GRPC_ERROR_REF(batch->payload->cancel_stream.cancel_error);
+    if (calld->send_message_batch != NULL) {
+      if (calld->send_initial_metadata_state == INITIAL_METADATA_UNSEEN) {
+        GRPC_CALL_COMBINER_START(
+            exec_ctx, calld->call_combiner,
+            GRPC_CLOSURE_CREATE(fail_send_message_batch_in_call_combiner, calld,
+                                grpc_schedule_on_exec_ctx),
+            GRPC_ERROR_REF(calld->cancel_error), "failing send_message op");
+      } else {
+        grpc_byte_stream_shutdown(
+            exec_ctx,
+            calld->send_message_batch->payload->send_message.send_message,
+            GRPC_ERROR_REF(calld->cancel_error));
+      }
     }
+  } else if (calld->cancel_error != GRPC_ERROR_NONE) {
+    grpc_transport_stream_op_batch_finish_with_failure(
+        exec_ctx, batch, GRPC_ERROR_REF(calld->cancel_error),
+        calld->call_combiner);
+    goto done;
   }
-
+  // Handle send_initial_metadata.
   if (batch->send_initial_metadata) {
+    GPR_ASSERT(calld->send_initial_metadata_state == INITIAL_METADATA_UNSEEN);
     bool has_compression_algorithm;
     grpc_error *error = process_send_initial_metadata(
         exec_ctx, elem,
         batch->payload->send_initial_metadata.send_initial_metadata,
         &has_compression_algorithm);
     if (error != GRPC_ERROR_NONE) {
-      grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, batch,
-                                                         error);
-      return;
+      grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, batch, error,
+                                                         calld->call_combiner);
+      goto done;
     }
-    gpr_atm cur;
-  retry_send_im:
-    cur = gpr_atm_acq_load(&calld->send_initial_metadata_state);
-    GPR_ASSERT(cur != HAS_COMPRESSION_ALGORITHM &&
-               cur != NO_COMPRESSION_ALGORITHM);
-    if ((cur & CANCELLED_BIT) == 0) {
-      if (!gpr_atm_rel_cas(&calld->send_initial_metadata_state, cur,
-                           has_compression_algorithm
-                               ? HAS_COMPRESSION_ALGORITHM
-                               : NO_COMPRESSION_ALGORITHM)) {
-        goto retry_send_im;
-      }
-      if (cur != INITIAL_METADATA_UNSEEN) {
-        start_send_message_batch(exec_ctx, elem,
-                                 (grpc_transport_stream_op_batch *)cur,
-                                 has_compression_algorithm);
-      }
+    calld->send_initial_metadata_state = has_compression_algorithm
+                                             ? HAS_COMPRESSION_ALGORITHM
+                                             : NO_COMPRESSION_ALGORITHM;
+    // If we had previously received a batch containing a send_message op,
+    // handle it now.  Note that we need to re-enter the call combiner
+    // for this, since we can't send two batches down while holding the
+    // call combiner, since the connected_channel filter (at the bottom of
+    // the call stack) will release the call combiner for each batch it sees.
+    if (calld->send_message_batch != NULL) {
+      GRPC_CALL_COMBINER_START(
+          exec_ctx, calld->call_combiner,
+          &calld->start_send_message_batch_in_call_combiner, GRPC_ERROR_NONE,
+          "starting send_message after send_initial_metadata");
     }
   }
+  // Handle send_message.
   if (batch->send_message) {
-    gpr_atm cur;
-  retry_send:
-    cur = gpr_atm_acq_load(&calld->send_initial_metadata_state);
-    switch (cur) {
-      case INITIAL_METADATA_UNSEEN:
-        if (!gpr_atm_rel_cas(&calld->send_initial_metadata_state, cur,
-                             (gpr_atm)batch)) {
-          goto retry_send;
-        }
-        break;
-      case HAS_COMPRESSION_ALGORITHM:
-      case NO_COMPRESSION_ALGORITHM:
-        start_send_message_batch(exec_ctx, elem, batch,
-                                 cur == HAS_COMPRESSION_ALGORITHM);
-        break;
-      default:
-        if (cur & CANCELLED_BIT) {
-          grpc_transport_stream_op_batch_finish_with_failure(
-              exec_ctx, batch,
-              GRPC_ERROR_REF((grpc_error *)(cur & ~CANCELLED_BIT)));
-        } else {
-          /* >1 send_message concurrently */
-          GPR_UNREACHABLE_CODE(break);
-        }
+    GPR_ASSERT(calld->send_message_batch == NULL);
+    calld->send_message_batch = batch;
+    // If we have not yet seen send_initial_metadata, then we have to
+    // wait.  We save the batch in calld and then drop the call
+    // combiner, which we'll have to pick up again later when we get
+    // send_initial_metadata.
+    if (calld->send_initial_metadata_state == INITIAL_METADATA_UNSEEN) {
+      GRPC_CALL_COMBINER_STOP(
+          exec_ctx, calld->call_combiner,
+          "send_message batch pending send_initial_metadata");
+      goto done;
     }
+    start_send_message_batch(exec_ctx, elem, GRPC_ERROR_NONE);
   } else {
-    /* pass control down the stack */
+    // Pass control down the stack.
     grpc_call_next_op(exec_ctx, elem, batch);
   }
-
+done:
   GPR_TIMER_END("compress_start_transport_stream_op_batch", 0);
 }
 
@@ -458,16 +456,16 @@ static void compress_start_transport_stream_op_batch(
 static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
                                   grpc_call_element *elem,
                                   const grpc_call_element_args *args) {
-  /* grab pointers to our data from the call element */
-  call_data *calld = elem->call_data;
-
-  /* initialize members */
+  call_data *calld = (call_data *)elem->call_data;
+  calld->call_combiner = args->call_combiner;
+  calld->cancel_error = GRPC_ERROR_NONE;
   grpc_slice_buffer_init(&calld->slices);
+  GRPC_CLOSURE_INIT(&calld->start_send_message_batch_in_call_combiner,
+                    start_send_message_batch, elem, grpc_schedule_on_exec_ctx);
   GRPC_CLOSURE_INIT(&calld->on_send_message_next_done,
                     on_send_message_next_done, elem, grpc_schedule_on_exec_ctx);
   GRPC_CLOSURE_INIT(&calld->send_message_on_complete, send_message_on_complete,
                     elem, grpc_schedule_on_exec_ctx);
-
   return GRPC_ERROR_NONE;
 }
 
@@ -475,14 +473,9 @@ static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
 static void destroy_call_elem(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
                               const grpc_call_final_info *final_info,
                               grpc_closure *ignored) {
-  /* grab pointers to our data from the call element */
   call_data *calld = elem->call_data;
   grpc_slice_buffer_destroy_internal(exec_ctx, &calld->slices);
-  gpr_atm imstate =
-      gpr_atm_no_barrier_load(&calld->send_initial_metadata_state);
-  if (imstate & CANCELLED_BIT) {
-    GRPC_ERROR_UNREF((grpc_error *)(imstate & ~CANCELLED_BIT));
-  }
+  GRPC_ERROR_UNREF(calld->cancel_error);
 }
 
 /* Constructor for channel_data */
@@ -550,6 +543,5 @@ const grpc_channel_filter grpc_message_compress_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
-    "compress"};
+    "message_compress"};
diff --git a/src/core/ext/filters/http/server/http_server_filter.c b/src/core/ext/filters/http/server/http_server_filter.c
index b145f12aff..a10e69ba59 100644
--- a/src/core/ext/filters/http/server/http_server_filter.c
+++ b/src/core/ext/filters/http/server/http_server_filter.c
@@ -32,6 +32,8 @@
 #define EXPECTED_CONTENT_TYPE_LENGTH sizeof(EXPECTED_CONTENT_TYPE) - 1
 
 typedef struct call_data {
+  grpc_call_combiner *call_combiner;
+
   grpc_linked_mdelem status;
   grpc_linked_mdelem content_type;
 
@@ -281,7 +283,11 @@ static void hs_on_complete(grpc_exec_ctx *exec_ctx, void *user_data,
     *calld->pp_recv_message = calld->payload_bin_delivered
                                   ? NULL
                                   : (grpc_byte_stream *)&calld->read_stream;
-    GRPC_CLOSURE_RUN(exec_ctx, calld->recv_message_ready, GRPC_ERROR_REF(err));
+    // Re-enter call combiner for recv_message_ready, since the surface
+    // code will release the call combiner for each callback it receives.
+    GRPC_CALL_COMBINER_START(exec_ctx, calld->call_combiner,
+                             calld->recv_message_ready, GRPC_ERROR_REF(err),
+                             "resuming recv_message_ready from on_complete");
     calld->recv_message_ready = NULL;
     calld->payload_bin_delivered = true;
   }
@@ -293,15 +299,20 @@ static void hs_recv_message_ready(grpc_exec_ctx *exec_ctx, void *user_data,
   grpc_call_element *elem = user_data;
   call_data *calld = elem->call_data;
   if (calld->seen_path_with_query) {
-    /* do nothing. This is probably a GET request, and payload will be returned
-    in hs_on_complete callback. */
+    // Do nothing. This is probably a GET request, and payload will be
+    // returned in hs_on_complete callback.
+    // Note that we release the call combiner here, so that other
+    // callbacks can run.
+    GRPC_CALL_COMBINER_STOP(exec_ctx, calld->call_combiner,
+                            "pausing recv_message_ready until on_complete");
   } else {
     GRPC_CLOSURE_RUN(exec_ctx, calld->recv_message_ready, GRPC_ERROR_REF(err));
   }
 }
 
-static void hs_mutate_op(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
-                         grpc_transport_stream_op_batch *op) {
+static grpc_error *hs_mutate_op(grpc_exec_ctx *exec_ctx,
+                                grpc_call_element *elem,
+                                grpc_transport_stream_op_batch *op) {
   /* grab pointers to our data from the call element */
   call_data *calld = elem->call_data;
 
@@ -323,10 +334,7 @@ static void hs_mutate_op(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
               server_filter_outgoing_metadata(
                   exec_ctx, elem,
                   op->payload->send_initial_metadata.send_initial_metadata));
-    if (error != GRPC_ERROR_NONE) {
-      grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, op, error);
-      return;
-    }
+    if (error != GRPC_ERROR_NONE) return error;
   }
 
   if (op->recv_initial_metadata) {
@@ -359,21 +367,25 @@ static void hs_mutate_op(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     grpc_error *error = server_filter_outgoing_metadata(
         exec_ctx, elem,
         op->payload->send_trailing_metadata.send_trailing_metadata);
-    if (error != GRPC_ERROR_NONE) {
-      grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, op, error);
-      return;
-    }
+    if (error != GRPC_ERROR_NONE) return error;
   }
+
+  return GRPC_ERROR_NONE;
 }
 
-static void hs_start_transport_op(grpc_exec_ctx *exec_ctx,
-                                  grpc_call_element *elem,
-                                  grpc_transport_stream_op_batch *op) {
-  GRPC_CALL_LOG_OP(GPR_INFO, elem, op);
-  GPR_TIMER_BEGIN("hs_start_transport_op", 0);
-  hs_mutate_op(exec_ctx, elem, op);
-  grpc_call_next_op(exec_ctx, elem, op);
-  GPR_TIMER_END("hs_start_transport_op", 0);
+static void hs_start_transport_stream_op_batch(
+    grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
+    grpc_transport_stream_op_batch *op) {
+  call_data *calld = elem->call_data;
+  GPR_TIMER_BEGIN("hs_start_transport_stream_op_batch", 0);
+  grpc_error *error = hs_mutate_op(exec_ctx, elem, op);
+  if (error != GRPC_ERROR_NONE) {
+    grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, op, error,
+                                                       calld->call_combiner);
+  } else {
+    grpc_call_next_op(exec_ctx, elem, op);
+  }
+  GPR_TIMER_END("hs_start_transport_stream_op_batch", 0);
 }
 
 /* Constructor for call_data */
@@ -383,6 +395,7 @@ static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
   /* grab pointers to our data from the call element */
   call_data *calld = elem->call_data;
   /* initialize members */
+  calld->call_combiner = args->call_combiner;
   GRPC_CLOSURE_INIT(&calld->hs_on_recv, hs_on_recv, elem,
                     grpc_schedule_on_exec_ctx);
   GRPC_CLOSURE_INIT(&calld->hs_on_complete, hs_on_complete, elem,
@@ -414,7 +427,7 @@ static void destroy_channel_elem(grpc_exec_ctx *exec_ctx,
                                  grpc_channel_element *elem) {}
 
 const grpc_channel_filter grpc_http_server_filter = {
-    hs_start_transport_op,
+    hs_start_transport_stream_op_batch,
     grpc_channel_next_op,
     sizeof(call_data),
     init_call_elem,
@@ -423,6 +436,5 @@ const grpc_channel_filter grpc_http_server_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "http-server"};
diff --git a/src/core/ext/filters/load_reporting/load_reporting.c b/src/core/ext/filters/load_reporting/load_reporting.c
index 9745763c91..b42aa99cdb 100644
--- a/src/core/ext/filters/load_reporting/load_reporting.c
+++ b/src/core/ext/filters/load_reporting/load_reporting.c
@@ -42,9 +42,15 @@ static bool maybe_add_load_reporting_filter(grpc_exec_ctx *exec_ctx,
                                             void *arg) {
   const grpc_channel_args *args =
       grpc_channel_stack_builder_get_channel_arguments(builder);
-  if (is_load_reporting_enabled(args)) {
-    return grpc_channel_stack_builder_prepend_filter(
-        builder, (const grpc_channel_filter *)arg, NULL, NULL);
+  const grpc_channel_filter *filter = arg;
+  grpc_channel_stack_builder_iterator *it =
+      grpc_channel_stack_builder_iterator_find(builder, filter->name);
+  const bool already_has_load_reporting_filter =
+      !grpc_channel_stack_builder_iterator_is_end(it);
+  grpc_channel_stack_builder_iterator_destroy(it);
+  if (is_load_reporting_enabled(args) && !already_has_load_reporting_filter) {
+    return grpc_channel_stack_builder_prepend_filter(builder, filter, NULL,
+                                                     NULL);
   }
   return true;
 }
diff --git a/src/core/ext/filters/load_reporting/load_reporting_filter.c b/src/core/ext/filters/load_reporting/load_reporting_filter.c
index 08474efb2e..17e946937f 100644
--- a/src/core/ext/filters/load_reporting/load_reporting_filter.c
+++ b/src/core/ext/filters/load_reporting/load_reporting_filter.c
@@ -223,6 +223,5 @@ const grpc_channel_filter grpc_load_reporting_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "load_reporting"};
diff --git a/src/core/ext/filters/max_age/max_age_filter.c b/src/core/ext/filters/max_age/max_age_filter.c
index 7d748b9c32..450f67746f 100644
--- a/src/core/ext/filters/max_age/max_age_filter.c
+++ b/src/core/ext/filters/max_age/max_age_filter.c
@@ -273,7 +273,7 @@ static grpc_error* init_call_elem(grpc_exec_ctx* exec_ctx,
 static void destroy_call_elem(grpc_exec_ctx* exec_ctx, grpc_call_element* elem,
                               const grpc_call_final_info* final_info,
                               grpc_closure* ignored) {
-  channel_data* chand = elem->channel_data;
+  channel_data* chand = (channel_data*)elem->channel_data;
   decrease_call_count(exec_ctx, chand);
 }
 
@@ -391,7 +391,6 @@ const grpc_channel_filter grpc_max_age_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "max_age"};
 
diff --git a/src/core/ext/filters/message_size/message_size_filter.c b/src/core/ext/filters/message_size/message_size_filter.c
index 846c7df69a..47763b1deb 100644
--- a/src/core/ext/filters/message_size/message_size_filter.c
+++ b/src/core/ext/filters/message_size/message_size_filter.c
@@ -68,6 +68,7 @@ static void* message_size_limits_create_from_json(const grpc_json* json) {
 }
 
 typedef struct call_data {
+  grpc_call_combiner* call_combiner;
   message_size_limits limits;
   // Receive closures are chained: we inject this closure as the
   // recv_message_ready up-call on transport_stream_op, and remember to
@@ -131,7 +132,8 @@ static void start_transport_stream_op_batch(
         exec_ctx, op,
         grpc_error_set_int(GRPC_ERROR_CREATE_FROM_COPIED_STRING(message_string),
                            GRPC_ERROR_INT_GRPC_STATUS,
-                           GRPC_STATUS_RESOURCE_EXHAUSTED));
+                           GRPC_STATUS_RESOURCE_EXHAUSTED),
+        calld->call_combiner);
     gpr_free(message_string);
     return;
   }
@@ -152,6 +154,7 @@ static grpc_error* init_call_elem(grpc_exec_ctx* exec_ctx,
                                   const grpc_call_element_args* args) {
   channel_data* chand = (channel_data*)elem->channel_data;
   call_data* calld = (call_data*)elem->call_data;
+  calld->call_combiner = args->call_combiner;
   calld->next_recv_message_ready = NULL;
   GRPC_CLOSURE_INIT(&calld->recv_message_ready, recv_message_ready, elem,
                     grpc_schedule_on_exec_ctx);
@@ -259,7 +262,6 @@ const grpc_channel_filter grpc_message_size_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "message_size"};
 
diff --git a/src/core/ext/filters/workarounds/workaround_cronet_compression_filter.c b/src/core/ext/filters/workarounds/workaround_cronet_compression_filter.c
index b4d2cb4b8c..c8b2fe5f99 100644
--- a/src/core/ext/filters/workarounds/workaround_cronet_compression_filter.c
+++ b/src/core/ext/filters/workarounds/workaround_cronet_compression_filter.c
@@ -177,7 +177,6 @@ const grpc_channel_filter grpc_workaround_cronet_compression_filter = {
     0,
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "workaround_cronet_compression"};
 
diff --git a/src/core/ext/transport/chttp2/transport/chttp2_transport.c b/src/core/ext/transport/chttp2/transport/chttp2_transport.c
index 3277cefc2d..7ff77b7cd9 100644
--- a/src/core/ext/transport/chttp2/transport/chttp2_transport.c
+++ b/src/core/ext/transport/chttp2/transport/chttp2_transport.c
@@ -1375,6 +1375,10 @@ static void perform_stream_op_locked(grpc_exec_ctx *exec_ctx, void *stream_op,
             "send_initial_metadata_finished");
       }
     }
+    if (op_payload->send_initial_metadata.peer_string != NULL) {
+      gpr_atm_rel_store(op_payload->send_initial_metadata.peer_string,
+                        (gpr_atm)gpr_strdup(t->peer_string));
+    }
   }
 
   if (op->send_message) {
@@ -1384,11 +1388,18 @@ static void perform_stream_op_locked(grpc_exec_ctx *exec_ctx, void *stream_op,
     on_complete->next_data.scratch |= CLOSURE_BARRIER_MAY_COVER_WRITE;
     s->fetching_send_message_finished = add_closure_barrier(op->on_complete);
     if (s->write_closed) {
+      // Return an error unless the client has already received trailing
+      // metadata from the server, since an application using a
+      // streaming call might send another message before getting a
+      // recv_message failure, breaking out of its loop, and then
+      // starting recv_trailing_metadata.
       grpc_chttp2_complete_closure_step(
           exec_ctx, t, s, &s->fetching_send_message_finished,
-          GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
-              "Attempt to send message after stream was closed",
-              &s->write_closed_error, 1),
+          t->is_client && s->received_trailing_metadata
+              ? GRPC_ERROR_NONE
+              : GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING(
+                    "Attempt to send message after stream was closed",
+                    &s->write_closed_error, 1),
           "fetching_send_message_finished");
     } else {
       GPR_ASSERT(s->fetching_send_message == NULL);
@@ -1476,6 +1487,10 @@ static void perform_stream_op_locked(grpc_exec_ctx *exec_ctx, void *stream_op,
         op_payload->recv_initial_metadata.recv_initial_metadata;
     s->trailing_metadata_available =
         op_payload->recv_initial_metadata.trailing_metadata_available;
+    if (op_payload->recv_initial_metadata.peer_string != NULL) {
+      gpr_atm_rel_store(op_payload->recv_initial_metadata.peer_string,
+                        (gpr_atm)gpr_strdup(t->peer_string));
+    }
     grpc_chttp2_maybe_complete_recv_initial_metadata(exec_ctx, t, s);
   }
 
@@ -1836,8 +1851,7 @@ void grpc_chttp2_maybe_complete_recv_trailing_metadata(grpc_exec_ctx *exec_ctx,
         }
       }
     }
-    if (s->read_closed && s->frame_storage.length == 0 &&
-        (!pending_data || s->seen_error) &&
+    if (s->read_closed && s->frame_storage.length == 0 && !pending_data &&
         s->recv_trailing_metadata_finished != NULL) {
       grpc_chttp2_incoming_metadata_buffer_publish(
           exec_ctx, &s->metadata_buffer[1], s->recv_trailing_metadata);
@@ -2944,14 +2958,6 @@ static void destructive_reclaimer_locked(grpc_exec_ctx *exec_ctx, void *arg,
 }
 
 /*******************************************************************************
- * INTEGRATION GLUE
- */
-
-static char *chttp2_get_peer(grpc_exec_ctx *exec_ctx, grpc_transport *t) {
-  return gpr_strdup(((grpc_chttp2_transport *)t)->peer_string);
-}
-
-/*******************************************************************************
  * MONITORING
  */
 static grpc_endpoint *chttp2_get_endpoint(grpc_exec_ctx *exec_ctx,
@@ -2968,7 +2974,6 @@ static const grpc_transport_vtable vtable = {sizeof(grpc_chttp2_stream),
                                              perform_transport_op,
                                              destroy_stream,
                                              destroy_transport,
-                                             chttp2_get_peer,
                                              chttp2_get_endpoint};
 
 grpc_transport *grpc_create_chttp2_transport(
diff --git a/src/core/ext/transport/chttp2/transport/internal.h b/src/core/ext/transport/chttp2/transport/internal.h
index 3c41a8958f..9fff30d54f 100644
--- a/src/core/ext/transport/chttp2/transport/internal.h
+++ b/src/core/ext/transport/chttp2/transport/internal.h
@@ -509,6 +509,8 @@ struct grpc_chttp2_stream {
   /** Are we buffering writes on this stream? If yes, we won't become writable
       until there's enough queued up in the flow_controlled_buffer */
   bool write_buffering;
+  /** Has trailing metadata been received. */
+  bool received_trailing_metadata;
 
   /** the error that resulted in this stream being read-closed */
   grpc_error *read_closed_error;
diff --git a/src/core/ext/transport/chttp2/transport/parsing.c b/src/core/ext/transport/chttp2/transport/parsing.c
index 18d163ee98..19bd86fd0c 100644
--- a/src/core/ext/transport/chttp2/transport/parsing.c
+++ b/src/core/ext/transport/chttp2/transport/parsing.c
@@ -623,6 +623,7 @@ static grpc_error *init_header_frame_parser(grpc_exec_ctx *exec_ctx,
           *s->trailing_metadata_available = true;
         }
         t->hpack_parser.on_header = on_trailing_header;
+        s->received_trailing_metadata = true;
       } else {
         GRPC_CHTTP2_IF_TRACING(gpr_log(GPR_INFO, "parsing initial_metadata"));
         t->hpack_parser.on_header = on_initial_header;
@@ -631,6 +632,7 @@ static grpc_error *init_header_frame_parser(grpc_exec_ctx *exec_ctx,
     case 1:
       GRPC_CHTTP2_IF_TRACING(gpr_log(GPR_INFO, "parsing trailing_metadata"));
       t->hpack_parser.on_header = on_trailing_header;
+      s->received_trailing_metadata = true;
       break;
     case 2:
       gpr_log(GPR_ERROR, "too many header frames received");
diff --git a/src/core/ext/transport/chttp2/transport/writing.c b/src/core/ext/transport/chttp2/transport/writing.c
index c16ffaa5ef..5e9d97d485 100644
--- a/src/core/ext/transport/chttp2/transport/writing.c
+++ b/src/core/ext/transport/chttp2/transport/writing.c
@@ -156,17 +156,8 @@ static uint32_t target_write_size(grpc_chttp2_transport *t) {
 }
 
 // Returns true if initial_metadata contains only default headers.
-//
-// TODO(roth): The fact that we hard-code these particular headers here
-// is fairly ugly.  Need some better way to know which headers are
-// default, maybe via a bit in the static metadata table?
 static bool is_default_initial_metadata(grpc_metadata_batch *initial_metadata) {
-  int num_default_fields =
-      (initial_metadata->idx.named.status != NULL) +
-      (initial_metadata->idx.named.content_type != NULL) +
-      (initial_metadata->idx.named.grpc_encoding != NULL) +
-      (initial_metadata->idx.named.grpc_accept_encoding != NULL);
-  return (size_t)num_default_fields == initial_metadata->list.count;
+  return initial_metadata->list.default_count == initial_metadata->list.count;
 }
 
 grpc_chttp2_begin_write_result grpc_chttp2_begin_write(
diff --git a/src/core/ext/transport/cronet/transport/cronet_transport.c b/src/core/ext/transport/cronet/transport/cronet_transport.c
index abb558982b..587a3b83b5 100644
--- a/src/core/ext/transport/cronet/transport/cronet_transport.c
+++ b/src/core/ext/transport/cronet/transport/cronet_transport.c
@@ -187,9 +187,34 @@ struct stream_obj {
 
   /* Mutex to protect storage */
   gpr_mu mu;
+
+  /* Refcount object of the stream */
+  grpc_stream_refcount *refcount;
 };
 typedef struct stream_obj stream_obj;
 
+#ifndef NDEBUG
+#define GRPC_CRONET_STREAM_REF(stream, reason) \
+  grpc_cronet_stream_ref((stream), (reason))
+#define GRPC_CRONET_STREAM_UNREF(exec_ctx, stream, reason) \
+  grpc_cronet_stream_unref((exec_ctx), (stream), (reason))
+void grpc_cronet_stream_ref(stream_obj *s, const char *reason) {
+  grpc_stream_ref(s->refcount, reason);
+}
+void grpc_cronet_stream_unref(grpc_exec_ctx *exec_ctx, stream_obj *s,
+                              const char *reason) {
+  grpc_stream_unref(exec_ctx, s->refcount, reason);
+}
+#else
+#define GRPC_CRONET_STREAM_REF(stream, reason) grpc_cronet_stream_ref((stream))
+#define GRPC_CRONET_STREAM_UNREF(exec_ctx, stream, reason) \
+  grpc_cronet_stream_unref((exec_ctx), (stream))
+void grpc_cronet_stream_ref(stream_obj *s) { grpc_stream_ref(s->refcount); }
+void grpc_cronet_stream_unref(grpc_exec_ctx *exec_ctx, stream_obj *s) {
+  grpc_stream_unref(exec_ctx, s->refcount);
+}
+#endif
+
 static enum e_op_result execute_stream_op(grpc_exec_ctx *exec_ctx,
                                           struct op_and_state *oas);
 
@@ -346,13 +371,12 @@ static void remove_from_storage(struct stream_obj *s,
   This can get executed from the Cronet network thread via cronet callback
   or on the application supplied thread via the perform_stream_op function.
 */
-static void execute_from_storage(stream_obj *s) {
-  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+static void execute_from_storage(grpc_exec_ctx *exec_ctx, stream_obj *s) {
   gpr_mu_lock(&s->mu);
   for (struct op_and_state *curr = s->storage.head; curr != NULL;) {
     CRONET_LOG(GPR_DEBUG, "calling op at %p. done = %d", curr, curr->done);
     GPR_ASSERT(curr->done == 0);
-    enum e_op_result result = execute_stream_op(&exec_ctx, curr);
+    enum e_op_result result = execute_stream_op(exec_ctx, curr);
     CRONET_LOG(GPR_DEBUG, "execute_stream_op[%p] returns %s", curr,
                op_result_string(result));
     /* if this op is done, then remove it and free memory */
@@ -369,7 +393,6 @@ static void execute_from_storage(stream_obj *s) {
     }
   }
   gpr_mu_unlock(&s->mu);
-  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -377,6 +400,8 @@ static void execute_from_storage(stream_obj *s) {
 */
 static void on_failed(bidirectional_stream *stream, int net_error) {
   CRONET_LOG(GPR_DEBUG, "on_failed(%p, %d)", stream, net_error);
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+
   stream_obj *s = (stream_obj *)stream->annotation;
   gpr_mu_lock(&s->mu);
   bidirectional_stream_destroy(s->cbs);
@@ -392,7 +417,9 @@ static void on_failed(bidirectional_stream *stream, int net_error) {
   }
   null_and_maybe_free_read_buffer(s);
   gpr_mu_unlock(&s->mu);
-  execute_from_storage(s);
+  execute_from_storage(&exec_ctx, s);
+  GRPC_CRONET_STREAM_UNREF(&exec_ctx, s, "cronet transport");
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -400,6 +427,8 @@ static void on_failed(bidirectional_stream *stream, int net_error) {
 */
 static void on_canceled(bidirectional_stream *stream) {
   CRONET_LOG(GPR_DEBUG, "on_canceled(%p)", stream);
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+
   stream_obj *s = (stream_obj *)stream->annotation;
   gpr_mu_lock(&s->mu);
   bidirectional_stream_destroy(s->cbs);
@@ -415,7 +444,9 @@ static void on_canceled(bidirectional_stream *stream) {
   }
   null_and_maybe_free_read_buffer(s);
   gpr_mu_unlock(&s->mu);
-  execute_from_storage(s);
+  execute_from_storage(&exec_ctx, s);
+  GRPC_CRONET_STREAM_UNREF(&exec_ctx, s, "cronet transport");
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -423,6 +454,8 @@ static void on_canceled(bidirectional_stream *stream) {
 */
 static void on_succeeded(bidirectional_stream *stream) {
   CRONET_LOG(GPR_DEBUG, "on_succeeded(%p)", stream);
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+
   stream_obj *s = (stream_obj *)stream->annotation;
   gpr_mu_lock(&s->mu);
   bidirectional_stream_destroy(s->cbs);
@@ -430,7 +463,9 @@ static void on_succeeded(bidirectional_stream *stream) {
   s->cbs = NULL;
   null_and_maybe_free_read_buffer(s);
   gpr_mu_unlock(&s->mu);
-  execute_from_storage(s);
+  execute_from_storage(&exec_ctx, s);
+  GRPC_CRONET_STREAM_UNREF(&exec_ctx, s, "cronet transport");
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -438,6 +473,7 @@ static void on_succeeded(bidirectional_stream *stream) {
 */
 static void on_stream_ready(bidirectional_stream *stream) {
   CRONET_LOG(GPR_DEBUG, "W: on_stream_ready(%p)", stream);
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   stream_obj *s = (stream_obj *)stream->annotation;
   grpc_cronet_transport *t = (grpc_cronet_transport *)s->curr_ct;
   gpr_mu_lock(&s->mu);
@@ -457,7 +493,8 @@ static void on_stream_ready(bidirectional_stream *stream) {
     }
   }
   gpr_mu_unlock(&s->mu);
-  execute_from_storage(s);
+  execute_from_storage(&exec_ctx, s);
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -513,14 +550,15 @@ static void on_response_headers_received(
     s->state.pending_read_from_cronet = true;
   }
   gpr_mu_unlock(&s->mu);
+  execute_from_storage(&exec_ctx, s);
   grpc_exec_ctx_finish(&exec_ctx);
-  execute_from_storage(s);
 }
 
 /*
   Cronet callback
 */
 static void on_write_completed(bidirectional_stream *stream, const char *data) {
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   stream_obj *s = (stream_obj *)stream->annotation;
   CRONET_LOG(GPR_DEBUG, "W: on_write_completed(%p, %s)", stream, data);
   gpr_mu_lock(&s->mu);
@@ -530,7 +568,8 @@ static void on_write_completed(bidirectional_stream *stream, const char *data) {
   }
   s->state.state_callback_received[OP_SEND_MESSAGE] = true;
   gpr_mu_unlock(&s->mu);
-  execute_from_storage(s);
+  execute_from_storage(&exec_ctx, s);
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -538,6 +577,7 @@ static void on_write_completed(bidirectional_stream *stream, const char *data) {
 */
 static void on_read_completed(bidirectional_stream *stream, char *data,
                               int count) {
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   stream_obj *s = (stream_obj *)stream->annotation;
   CRONET_LOG(GPR_DEBUG, "R: on_read_completed(%p, %p, %d)", stream, data,
              count);
@@ -563,14 +603,15 @@ static void on_read_completed(bidirectional_stream *stream, char *data,
       gpr_mu_unlock(&s->mu);
     } else {
       gpr_mu_unlock(&s->mu);
-      execute_from_storage(s);
+      execute_from_storage(&exec_ctx, s);
     }
   } else {
     null_and_maybe_free_read_buffer(s);
     s->state.rs.read_stream_closed = true;
     gpr_mu_unlock(&s->mu);
-    execute_from_storage(s);
+    execute_from_storage(&exec_ctx, s);
   }
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -625,12 +666,11 @@ static void on_response_trailers_received(
     s->state.state_op_done[OP_SEND_TRAILING_METADATA] = true;
 
     gpr_mu_unlock(&s->mu);
-    grpc_exec_ctx_finish(&exec_ctx);
   } else {
     gpr_mu_unlock(&s->mu);
-    grpc_exec_ctx_finish(&exec_ctx);
-    execute_from_storage(s);
+    execute_from_storage(&exec_ctx, s);
   }
+  grpc_exec_ctx_finish(&exec_ctx);
 }
 
 /*
@@ -1313,6 +1353,9 @@ static int init_stream(grpc_exec_ctx *exec_ctx, grpc_transport *gt,
                        grpc_stream *gs, grpc_stream_refcount *refcount,
                        const void *server_data, gpr_arena *arena) {
   stream_obj *s = (stream_obj *)gs;
+
+  s->refcount = refcount;
+  GRPC_CRONET_STREAM_REF(s, "cronet transport");
   memset(&s->storage, 0, sizeof(s->storage));
   s->storage.head = NULL;
   memset(&s->state, 0, sizeof(s->state));
@@ -1370,7 +1413,7 @@ static void perform_stream_op(grpc_exec_ctx *exec_ctx, grpc_transport *gt,
   }
   stream_obj *s = (stream_obj *)gs;
   add_to_storage(s, op);
-  execute_from_storage(s);
+  execute_from_storage(exec_ctx, s);
 }
 
 static void destroy_stream(grpc_exec_ctx *exec_ctx, grpc_transport *gt,
@@ -1386,10 +1429,6 @@ static void destroy_stream(grpc_exec_ctx *exec_ctx, grpc_transport *gt,
 
 static void destroy_transport(grpc_exec_ctx *exec_ctx, grpc_transport *gt) {}
 
-static char *get_peer(grpc_exec_ctx *exec_ctx, grpc_transport *gt) {
-  return NULL;
-}
-
 static grpc_endpoint *get_endpoint(grpc_exec_ctx *exec_ctx,
                                    grpc_transport *gt) {
   return NULL;
@@ -1408,7 +1447,6 @@ static const grpc_transport_vtable grpc_cronet_vtable = {
     perform_op,
     destroy_stream,
     destroy_transport,
-    get_peer,
     get_endpoint};
 
 grpc_transport *grpc_create_cronet_transport(void *engine, const char *target,
diff --git a/src/core/ext/transport/inproc/inproc_transport.c b/src/core/ext/transport/inproc/inproc_transport.c
index 6f4b429ee2..b2d6f2d0c9 100644
--- a/src/core/ext/transport/inproc/inproc_transport.c
+++ b/src/core/ext/transport/inproc/inproc_transport.c
@@ -1251,20 +1251,14 @@ static void set_pollset_set(grpc_exec_ctx *exec_ctx, grpc_transport *gt,
   // Nothing to do here
 }
 
-static char *get_peer(grpc_exec_ctx *exec_ctx, grpc_transport *t) {
-  return gpr_strdup("inproc");
-}
-
 static grpc_endpoint *get_endpoint(grpc_exec_ctx *exec_ctx, grpc_transport *t) {
   return NULL;
 }
 
 static const grpc_transport_vtable inproc_vtable = {
-    sizeof(inproc_stream), "inproc",
-    init_stream,           set_pollset,
-    set_pollset_set,       perform_stream_op,
-    perform_transport_op,  destroy_stream,
-    destroy_transport,     get_peer,
+    sizeof(inproc_stream), "inproc",        init_stream,
+    set_pollset,           set_pollset_set, perform_stream_op,
+    perform_transport_op,  destroy_stream,  destroy_transport,
     get_endpoint};
 
 /*******************************************************************************
diff --git a/src/core/lib/channel/channel_stack.c b/src/core/lib/channel/channel_stack.c
index 0f8e33c4be..775c8bc667 100644
--- a/src/core/lib/channel/channel_stack.c
+++ b/src/core/lib/channel/channel_stack.c
@@ -233,15 +233,10 @@ void grpc_call_stack_destroy(grpc_exec_ctx *exec_ctx, grpc_call_stack *stack,
 void grpc_call_next_op(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
                        grpc_transport_stream_op_batch *op) {
   grpc_call_element *next_elem = elem + 1;
+  GRPC_CALL_LOG_OP(GPR_INFO, next_elem, op);
   next_elem->filter->start_transport_stream_op_batch(exec_ctx, next_elem, op);
 }
 
-char *grpc_call_next_get_peer(grpc_exec_ctx *exec_ctx,
-                              grpc_call_element *elem) {
-  grpc_call_element *next_elem = elem + 1;
-  return next_elem->filter->get_peer(exec_ctx, next_elem);
-}
-
 void grpc_channel_next_get_info(grpc_exec_ctx *exec_ctx,
                                 grpc_channel_element *elem,
                                 const grpc_channel_info *channel_info) {
@@ -265,12 +260,3 @@ grpc_call_stack *grpc_call_stack_from_top_element(grpc_call_element *elem) {
   return (grpc_call_stack *)((char *)(elem)-ROUND_UP_TO_ALIGNMENT_SIZE(
       sizeof(grpc_call_stack)));
 }
-
-void grpc_call_element_signal_error(grpc_exec_ctx *exec_ctx,
-                                    grpc_call_element *elem,
-                                    grpc_error *error) {
-  grpc_transport_stream_op_batch *op = grpc_make_transport_stream_op(NULL);
-  op->cancel_stream = true;
-  op->payload->cancel_stream.cancel_error = error;
-  elem->filter->start_transport_stream_op_batch(exec_ctx, elem, op);
-}
diff --git a/src/core/lib/channel/channel_stack.h b/src/core/lib/channel/channel_stack.h
index a80f8aa826..ae1cac31f7 100644
--- a/src/core/lib/channel/channel_stack.h
+++ b/src/core/lib/channel/channel_stack.h
@@ -40,6 +40,7 @@
 #include <grpc/support/time.h>
 
 #include "src/core/lib/debug/trace.h"
+#include "src/core/lib/iomgr/call_combiner.h"
 #include "src/core/lib/iomgr/polling_entity.h"
 #include "src/core/lib/support/arena.h"
 #include "src/core/lib/transport/transport.h"
@@ -71,6 +72,7 @@ typedef struct {
   gpr_timespec start_time;
   gpr_timespec deadline;
   gpr_arena *arena;
+  grpc_call_combiner *call_combiner;
 } grpc_call_element_args;
 
 typedef struct {
@@ -150,9 +152,6 @@ typedef struct {
   void (*destroy_channel_elem)(grpc_exec_ctx *exec_ctx,
                                grpc_channel_element *elem);
 
-  /* Implement grpc_call_get_peer() */
-  char *(*get_peer)(grpc_exec_ctx *exec_ctx, grpc_call_element *elem);
-
   /* Implement grpc_channel_get_info() */
   void (*get_channel_info)(grpc_exec_ctx *exec_ctx, grpc_channel_element *elem,
                            const grpc_channel_info *channel_info);
@@ -271,8 +270,6 @@ void grpc_call_next_op(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
    stack */
 void grpc_channel_next_op(grpc_exec_ctx *exec_ctx, grpc_channel_element *elem,
                           grpc_transport_op *op);
-/* Pass through a request to get_peer to the next child element */
-char *grpc_call_next_get_peer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem);
 /* Pass through a request to get_channel_info() to the next child element */
 void grpc_channel_next_get_info(grpc_exec_ctx *exec_ctx,
                                 grpc_channel_element *elem,
@@ -288,10 +285,6 @@ void grpc_call_log_op(char *file, int line, gpr_log_severity severity,
                       grpc_call_element *elem,
                       grpc_transport_stream_op_batch *op);
 
-void grpc_call_element_signal_error(grpc_exec_ctx *exec_ctx,
-                                    grpc_call_element *cur_elem,
-                                    grpc_error *error);
-
 extern grpc_tracer_flag grpc_trace_channel;
 
 #define GRPC_CALL_LOG_OP(sev, elem, op) \
diff --git a/src/core/lib/channel/channel_stack_builder.c b/src/core/lib/channel/channel_stack_builder.c
index c369e33073..7f2b8e07ce 100644
--- a/src/core/lib/channel/channel_stack_builder.c
+++ b/src/core/lib/channel/channel_stack_builder.c
@@ -124,6 +124,20 @@ bool grpc_channel_stack_builder_move_prev(
   return true;
 }
 
+grpc_channel_stack_builder_iterator *grpc_channel_stack_builder_iterator_find(
+    grpc_channel_stack_builder *builder, const char *filter_name) {
+  GPR_ASSERT(filter_name != NULL);
+  grpc_channel_stack_builder_iterator *it =
+      grpc_channel_stack_builder_create_iterator_at_first(builder);
+  while (grpc_channel_stack_builder_move_next(it)) {
+    if (grpc_channel_stack_builder_iterator_is_end(it)) break;
+    const char *filter_name_at_it =
+        grpc_channel_stack_builder_iterator_filter_name(it);
+    if (strcmp(filter_name, filter_name_at_it) == 0) break;
+  }
+  return it;
+}
+
 bool grpc_channel_stack_builder_move_prev(
     grpc_channel_stack_builder_iterator *iterator);
 
@@ -169,6 +183,21 @@ bool grpc_channel_stack_builder_append_filter(
   return ok;
 }
 
+bool grpc_channel_stack_builder_remove_filter(
+    grpc_channel_stack_builder *builder, const char *filter_name) {
+  grpc_channel_stack_builder_iterator *it =
+      grpc_channel_stack_builder_iterator_find(builder, filter_name);
+  if (grpc_channel_stack_builder_iterator_is_end(it)) {
+    grpc_channel_stack_builder_iterator_destroy(it);
+    return false;
+  }
+  it->node->prev->next = it->node->next;
+  it->node->next->prev = it->node->prev;
+  gpr_free(it->node);
+  grpc_channel_stack_builder_iterator_destroy(it);
+  return true;
+}
+
 bool grpc_channel_stack_builder_prepend_filter(
     grpc_channel_stack_builder *builder, const grpc_channel_filter *filter,
     grpc_post_filter_create_init_func post_init_func, void *user_data) {
diff --git a/src/core/lib/channel/channel_stack_builder.h b/src/core/lib/channel/channel_stack_builder.h
index d43e427962..fdff2a2b6d 100644
--- a/src/core/lib/channel/channel_stack_builder.h
+++ b/src/core/lib/channel/channel_stack_builder.h
@@ -95,6 +95,11 @@ bool grpc_channel_stack_builder_move_next(
 bool grpc_channel_stack_builder_move_prev(
     grpc_channel_stack_builder_iterator *iterator);
 
+/// Return an iterator at \a filter_name, or at the end of the list if not
+/// found.
+grpc_channel_stack_builder_iterator *grpc_channel_stack_builder_iterator_find(
+    grpc_channel_stack_builder *builder, const char *filter_name);
+
 typedef void (*grpc_post_filter_create_init_func)(
     grpc_channel_stack *channel_stack, grpc_channel_element *elem, void *arg);
 
@@ -132,6 +137,11 @@ bool grpc_channel_stack_builder_append_filter(
     grpc_post_filter_create_init_func post_init_func,
     void *user_data) GRPC_MUST_USE_RESULT;
 
+/// Remove any filter whose name is \a filter_name from \a builder. Returns true
+/// if \a filter_name was not found.
+bool grpc_channel_stack_builder_remove_filter(
+    grpc_channel_stack_builder *builder, const char *filter_name);
+
 /// Terminate iteration and destroy \a iterator
 void grpc_channel_stack_builder_iterator_destroy(
     grpc_channel_stack_builder_iterator *iterator);
diff --git a/src/core/lib/channel/connected_channel.c b/src/core/lib/channel/connected_channel.c
index af06ca802e..8285226fc4 100644
--- a/src/core/lib/channel/connected_channel.c
+++ b/src/core/lib/channel/connected_channel.c
@@ -36,7 +36,57 @@ typedef struct connected_channel_channel_data {
   grpc_transport *transport;
 } channel_data;
 
-typedef struct connected_channel_call_data { void *unused; } call_data;
+typedef struct {
+  grpc_closure closure;
+  grpc_closure *original_closure;
+  grpc_call_combiner *call_combiner;
+  const char *reason;
+} callback_state;
+
+typedef struct connected_channel_call_data {
+  grpc_call_combiner *call_combiner;
+  // Closures used for returning results on the call combiner.
+  callback_state on_complete[6];  // Max number of pending batches.
+  callback_state recv_initial_metadata_ready;
+  callback_state recv_message_ready;
+} call_data;
+
+static void run_in_call_combiner(grpc_exec_ctx *exec_ctx, void *arg,
+                                 grpc_error *error) {
+  callback_state *state = (callback_state *)arg;
+  GRPC_CALL_COMBINER_START(exec_ctx, state->call_combiner,
+                           state->original_closure, GRPC_ERROR_REF(error),
+                           state->reason);
+}
+
+static void run_cancel_in_call_combiner(grpc_exec_ctx *exec_ctx, void *arg,
+                                        grpc_error *error) {
+  run_in_call_combiner(exec_ctx, arg, error);
+  gpr_free(arg);
+}
+
+static void intercept_callback(call_data *calld, callback_state *state,
+                               bool free_when_done, const char *reason,
+                               grpc_closure **original_closure) {
+  state->original_closure = *original_closure;
+  state->call_combiner = calld->call_combiner;
+  state->reason = reason;
+  *original_closure = GRPC_CLOSURE_INIT(
+      &state->closure,
+      free_when_done ? run_cancel_in_call_combiner : run_in_call_combiner,
+      state, grpc_schedule_on_exec_ctx);
+}
+
+static callback_state *get_state_for_batch(
+    call_data *calld, grpc_transport_stream_op_batch *batch) {
+  if (batch->send_initial_metadata) return &calld->on_complete[0];
+  if (batch->send_message) return &calld->on_complete[1];
+  if (batch->send_trailing_metadata) return &calld->on_complete[2];
+  if (batch->recv_initial_metadata) return &calld->on_complete[3];
+  if (batch->recv_message) return &calld->on_complete[4];
+  if (batch->recv_trailing_metadata) return &calld->on_complete[5];
+  GPR_UNREACHABLE_CODE(return NULL);
+}
 
 /* We perform a small hack to locate transport data alongside the connected
    channel data in call allocations, to allow everything to be pulled in minimal
@@ -49,13 +99,38 @@ typedef struct connected_channel_call_data { void *unused; } call_data;
    into transport stream operations */
 static void con_start_transport_stream_op_batch(
     grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
-    grpc_transport_stream_op_batch *op) {
+    grpc_transport_stream_op_batch *batch) {
   call_data *calld = elem->call_data;
   channel_data *chand = elem->channel_data;
-  GRPC_CALL_LOG_OP(GPR_INFO, elem, op);
-
+  if (batch->recv_initial_metadata) {
+    callback_state *state = &calld->recv_initial_metadata_ready;
+    intercept_callback(
+        calld, state, false, "recv_initial_metadata_ready",
+        &batch->payload->recv_initial_metadata.recv_initial_metadata_ready);
+  }
+  if (batch->recv_message) {
+    callback_state *state = &calld->recv_message_ready;
+    intercept_callback(calld, state, false, "recv_message_ready",
+                       &batch->payload->recv_message.recv_message_ready);
+  }
+  if (batch->cancel_stream) {
+    // There can be more than one cancellation batch in flight at any
+    // given time, so we can't just pick out a fixed index into
+    // calld->on_complete like we can for the other ops.  However,
+    // cancellation isn't in the fast path, so we just allocate a new
+    // closure for each one.
+    callback_state *state = (callback_state *)gpr_malloc(sizeof(*state));
+    intercept_callback(calld, state, true, "on_complete (cancel_stream)",
+                       &batch->on_complete);
+  } else {
+    callback_state *state = get_state_for_batch(calld, batch);
+    intercept_callback(calld, state, false, "on_complete", &batch->on_complete);
+  }
   grpc_transport_perform_stream_op(exec_ctx, chand->transport,
-                                   TRANSPORT_STREAM_FROM_CALL_DATA(calld), op);
+                                   TRANSPORT_STREAM_FROM_CALL_DATA(calld),
+                                   batch);
+  GRPC_CALL_COMBINER_STOP(exec_ctx, calld->call_combiner,
+                          "passed batch to transport");
 }
 
 static void con_start_transport_op(grpc_exec_ctx *exec_ctx,
@@ -71,6 +146,7 @@ static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
                                   const grpc_call_element_args *args) {
   call_data *calld = elem->call_data;
   channel_data *chand = elem->channel_data;
+  calld->call_combiner = args->call_combiner;
   int r = grpc_transport_init_stream(
       exec_ctx, chand->transport, TRANSPORT_STREAM_FROM_CALL_DATA(calld),
       &args->call_stack->refcount, args->server_transport_data, args->arena);
@@ -118,11 +194,6 @@ static void destroy_channel_elem(grpc_exec_ctx *exec_ctx,
   }
 }
 
-static char *con_get_peer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem) {
-  channel_data *chand = elem->channel_data;
-  return grpc_transport_get_peer(exec_ctx, chand->transport);
-}
-
 /* No-op. */
 static void con_get_channel_info(grpc_exec_ctx *exec_ctx,
                                  grpc_channel_element *elem,
@@ -138,7 +209,6 @@ const grpc_channel_filter grpc_connected_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    con_get_peer,
     con_get_channel_info,
     "connected",
 };
diff --git a/src/core/lib/debug/stats_data.c b/src/core/lib/debug/stats_data.c
index 73703208da..6d00fc8880 100644
--- a/src/core/lib/debug/stats_data.c
+++ b/src/core/lib/debug/stats_data.c
@@ -56,36 +56,37 @@ const char *grpc_stats_histogram_name[GRPC_STATS_HISTOGRAM_COUNT] = {
     "tcp_read_offer",     "tcp_read_iov_size",  "http2_send_message_size",
     "server_cqs_checked",
 };
-const int grpc_stats_table_0[64] = {
-    0,       1,       2,       3,       4,       6,        8,        11,
-    15,      20,      26,      34,      44,      57,       74,       96,
-    124,     160,     206,     265,     341,     439,      565,      727,
-    935,     1202,    1546,    1988,    2556,    3286,     4225,     5432,
-    6983,    8977,    11540,   14834,   19069,   24513,    31510,    40505,
-    52067,   66929,   86033,   110590,  142157,  182734,   234893,   301940,
-    388125,  498910,  641316,  824370,  1059674, 1362141,  1750943,  2250722,
-    2893155, 3718960, 4780478, 6144988, 7898976, 10153611, 13051794, 16777216};
+const int grpc_stats_table_0[65] = {
+    0,       1,       2,       3,       4,       6,       8,        11,
+    15,      20,      26,      34,      44,      57,      73,       94,
+    121,     155,     199,     255,     327,     419,     537,      688,
+    881,     1128,    1444,    1848,    2365,    3026,    3872,     4954,
+    6338,    8108,    10373,   13270,   16976,   21717,   27782,    35541,
+    45467,   58165,   74409,   95189,   121772,  155778,  199281,   254933,
+    326126,  417200,  533707,  682750,  873414,  1117323, 1429345,  1828502,
+    2339127, 2992348, 3827987, 4896985, 6264509, 8013925, 10251880, 13114801,
+    16777216};
 const uint8_t grpc_stats_table_1[87] = {
-    0,  0,  1,  1,  2,  3,  3,  4,  4,  5,  6,  6,  7,  8,  8,  9,  10, 10,
-    11, 12, 12, 13, 14, 14, 15, 16, 17, 17, 18, 19, 19, 20, 21, 21, 22, 23,
-    24, 24, 25, 25, 26, 27, 28, 28, 29, 30, 30, 31, 32, 33, 33, 34, 35, 35,
-    36, 36, 37, 38, 39, 39, 40, 41, 41, 42, 43, 44, 44, 45, 46, 46, 47, 47,
-    48, 49, 50, 50, 51, 52, 52, 53, 54, 55, 55, 56, 57, 57, 58};
-const int grpc_stats_table_2[64] = {
-    0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  13,
-    15,  17,  19,  21,  23,  25,  28,  31,  34,  37,  41,  45,  49,
-    54,  59,  64,  70,  76,  83,  90,  98,  106, 115, 125, 136, 147,
-    159, 172, 186, 201, 218, 236, 255, 276, 299, 323, 349, 377, 408,
-    441, 477, 515, 556, 601, 649, 701, 757, 817, 881, 950, 1024};
-const uint8_t grpc_stats_table_3[104] = {
-    0,  0,  0,  1,  1,  1,  1,  2,  2,  2,  3,  3,  4,  4,  5,  5,  6,  6,
-    7,  7,  7,  8,  8,  8,  9,  9,  10, 11, 11, 12, 12, 13, 13, 14, 14, 14,
-    15, 15, 16, 16, 16, 17, 18, 18, 19, 20, 20, 21, 21, 22, 22, 23, 23, 24,
-    24, 24, 25, 25, 26, 27, 28, 28, 29, 29, 30, 30, 31, 31, 32, 32, 33, 33,
-    34, 34, 35, 36, 36, 37, 38, 38, 39, 39, 40, 40, 41, 41, 42, 42, 42, 43,
-    44, 45, 45, 46, 47, 47, 48, 48, 49, 49, 50, 50, 51, 51};
-const int grpc_stats_table_4[8] = {0, 1, 2, 4, 8, 16, 32, 64};
-const uint8_t grpc_stats_table_5[4] = {0, 1, 2, 3};
+    0,  0,  1,  1,  2,  3,  3,  4,  4,  5,  6,  6,  7,  8,  8,  9,  10, 11,
+    11, 12, 13, 13, 14, 15, 15, 16, 17, 17, 18, 19, 20, 20, 21, 22, 22, 23,
+    24, 25, 25, 26, 27, 27, 28, 29, 29, 30, 31, 31, 32, 33, 34, 34, 35, 36,
+    36, 37, 38, 39, 39, 40, 41, 41, 42, 43, 44, 44, 45, 45, 46, 47, 48, 48,
+    49, 50, 51, 51, 52, 53, 53, 54, 55, 56, 56, 57, 58, 58, 59};
+const int grpc_stats_table_2[65] = {
+    0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,
+    14,  16,  18,  20,  22,  24,  27,  30,  33,  36,  39,  43,  47,
+    51,  56,  61,  66,  72,  78,  85,  92,  100, 109, 118, 128, 139,
+    151, 164, 178, 193, 209, 226, 244, 264, 285, 308, 333, 359, 387,
+    418, 451, 486, 524, 565, 609, 656, 707, 762, 821, 884, 952, 1024};
+const uint8_t grpc_stats_table_3[102] = {
+    0,  0,  0,  1,  1,  1,  1,  2,  2,  3,  3,  4,  4,  5,  5,  6,  6,
+    6,  7,  7,  7,  8,  8,  9,  9,  10, 11, 11, 12, 12, 13, 13, 14, 14,
+    14, 15, 15, 16, 16, 17, 17, 18, 19, 19, 20, 20, 21, 21, 22, 22, 23,
+    23, 24, 24, 24, 25, 26, 27, 27, 28, 28, 29, 29, 30, 30, 31, 31, 32,
+    32, 33, 33, 34, 35, 35, 36, 37, 37, 38, 38, 39, 39, 40, 40, 41, 41,
+    42, 42, 43, 44, 44, 45, 46, 46, 47, 48, 48, 49, 49, 50, 50, 51, 51};
+const int grpc_stats_table_4[9] = {0, 1, 2, 4, 7, 13, 23, 39, 64};
+const uint8_t grpc_stats_table_5[9] = {0, 0, 1, 2, 2, 3, 4, 4, 5};
 void grpc_stats_inc_tcp_write_size(grpc_exec_ctx *exec_ctx, int value) {
   value = GPR_CLAMP(value, 0, 16777216);
   if (value < 5) {
@@ -98,7 +99,7 @@ void grpc_stats_inc_tcp_write_size(grpc_exec_ctx *exec_ctx, int value) {
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4682617712558473216ull) {
+  if (_val.uint < 4683743612465315840ull) {
     int bucket =
         grpc_stats_table_1[((_val.uint - 4617315517961601024ull) >> 50)] + 5;
     _bkt.dbl = grpc_stats_table_0[bucket];
@@ -113,7 +114,7 @@ void grpc_stats_inc_tcp_write_size(grpc_exec_ctx *exec_ctx, int value) {
 }
 void grpc_stats_inc_tcp_write_iov_size(grpc_exec_ctx *exec_ctx, int value) {
   value = GPR_CLAMP(value, 0, 1024);
-  if (value < 12) {
+  if (value < 13) {
     GRPC_STATS_INC_HISTOGRAM((exec_ctx),
                              GRPC_STATS_HISTOGRAM_TCP_WRITE_IOV_SIZE, value);
     return;
@@ -123,9 +124,9 @@ void grpc_stats_inc_tcp_write_iov_size(grpc_exec_ctx *exec_ctx, int value) {
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4637300241308057600ull) {
+  if (_val.uint < 4637863191261478912ull) {
     int bucket =
-        grpc_stats_table_3[((_val.uint - 4622945017495814144ull) >> 48)] + 12;
+        grpc_stats_table_3[((_val.uint - 4623507967449235456ull) >> 48)] + 13;
     _bkt.dbl = grpc_stats_table_2[bucket];
     bucket -= (_val.uint < _bkt.uint);
     GRPC_STATS_INC_HISTOGRAM((exec_ctx),
@@ -148,7 +149,7 @@ void grpc_stats_inc_tcp_read_size(grpc_exec_ctx *exec_ctx, int value) {
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4682617712558473216ull) {
+  if (_val.uint < 4683743612465315840ull) {
     int bucket =
         grpc_stats_table_1[((_val.uint - 4617315517961601024ull) >> 50)] + 5;
     _bkt.dbl = grpc_stats_table_0[bucket];
@@ -173,7 +174,7 @@ void grpc_stats_inc_tcp_read_offer(grpc_exec_ctx *exec_ctx, int value) {
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4682617712558473216ull) {
+  if (_val.uint < 4683743612465315840ull) {
     int bucket =
         grpc_stats_table_1[((_val.uint - 4617315517961601024ull) >> 50)] + 5;
     _bkt.dbl = grpc_stats_table_0[bucket];
@@ -188,7 +189,7 @@ void grpc_stats_inc_tcp_read_offer(grpc_exec_ctx *exec_ctx, int value) {
 }
 void grpc_stats_inc_tcp_read_iov_size(grpc_exec_ctx *exec_ctx, int value) {
   value = GPR_CLAMP(value, 0, 1024);
-  if (value < 12) {
+  if (value < 13) {
     GRPC_STATS_INC_HISTOGRAM((exec_ctx), GRPC_STATS_HISTOGRAM_TCP_READ_IOV_SIZE,
                              value);
     return;
@@ -198,9 +199,9 @@ void grpc_stats_inc_tcp_read_iov_size(grpc_exec_ctx *exec_ctx, int value) {
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4637300241308057600ull) {
+  if (_val.uint < 4637863191261478912ull) {
     int bucket =
-        grpc_stats_table_3[((_val.uint - 4622945017495814144ull) >> 48)] + 12;
+        grpc_stats_table_3[((_val.uint - 4623507967449235456ull) >> 48)] + 13;
     _bkt.dbl = grpc_stats_table_2[bucket];
     bucket -= (_val.uint < _bkt.uint);
     GRPC_STATS_INC_HISTOGRAM((exec_ctx), GRPC_STATS_HISTOGRAM_TCP_READ_IOV_SIZE,
@@ -224,7 +225,7 @@ void grpc_stats_inc_http2_send_message_size(grpc_exec_ctx *exec_ctx,
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4682617712558473216ull) {
+  if (_val.uint < 4683743612465315840ull) {
     int bucket =
         grpc_stats_table_1[((_val.uint - 4617315517961601024ull) >> 50)] + 5;
     _bkt.dbl = grpc_stats_table_0[bucket];
@@ -250,9 +251,9 @@ void grpc_stats_inc_server_cqs_checked(grpc_exec_ctx *exec_ctx, int value) {
     uint64_t uint;
   } _val, _bkt;
   _val.dbl = value;
-  if (_val.uint < 4627448617123184640ull) {
+  if (_val.uint < 4625196817309499392ull) {
     int bucket =
-        grpc_stats_table_5[((_val.uint - 4613937818241073152ull) >> 52)] + 3;
+        grpc_stats_table_5[((_val.uint - 4613937818241073152ull) >> 51)] + 3;
     _bkt.dbl = grpc_stats_table_4[bucket];
     bucket -= (_val.uint < _bkt.uint);
     GRPC_STATS_INC_HISTOGRAM((exec_ctx),
diff --git a/src/core/lib/iomgr/call_combiner.c b/src/core/lib/iomgr/call_combiner.c
new file mode 100644
index 0000000000..48d8eaec18
--- /dev/null
+++ b/src/core/lib/iomgr/call_combiner.c
@@ -0,0 +1,202 @@
+/*
+ *
+ * Copyright 2017 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 "src/core/lib/iomgr/call_combiner.h"
+
+#include <grpc/support/log.h>
+
+grpc_tracer_flag grpc_call_combiner_trace =
+    GRPC_TRACER_INITIALIZER(false, "call_combiner");
+
+static grpc_error* decode_cancel_state_error(gpr_atm cancel_state) {
+  if (cancel_state & 1) {
+    return (grpc_error*)(cancel_state & ~(gpr_atm)1);
+  }
+  return GRPC_ERROR_NONE;
+}
+
+static gpr_atm encode_cancel_state_error(grpc_error* error) {
+  return (gpr_atm)1 | (gpr_atm)error;
+}
+
+void grpc_call_combiner_init(grpc_call_combiner* call_combiner) {
+  gpr_mpscq_init(&call_combiner->queue);
+}
+
+void grpc_call_combiner_destroy(grpc_call_combiner* call_combiner) {
+  gpr_mpscq_destroy(&call_combiner->queue);
+  GRPC_ERROR_UNREF(decode_cancel_state_error(call_combiner->cancel_state));
+}
+
+#ifndef NDEBUG
+#define DEBUG_ARGS , const char *file, int line
+#define DEBUG_FMT_STR "%s:%d: "
+#define DEBUG_FMT_ARGS , file, line
+#else
+#define DEBUG_ARGS
+#define DEBUG_FMT_STR
+#define DEBUG_FMT_ARGS
+#endif
+
+void grpc_call_combiner_start(grpc_exec_ctx* exec_ctx,
+                              grpc_call_combiner* call_combiner,
+                              grpc_closure* closure,
+                              grpc_error* error DEBUG_ARGS,
+                              const char* reason) {
+  if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+    gpr_log(GPR_DEBUG,
+            "==> grpc_call_combiner_start() [%p] closure=%p [" DEBUG_FMT_STR
+            "%s] error=%s",
+            call_combiner, closure DEBUG_FMT_ARGS, reason,
+            grpc_error_string(error));
+  }
+  size_t prev_size =
+      (size_t)gpr_atm_full_fetch_add(&call_combiner->size, (gpr_atm)1);
+  if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+    gpr_log(GPR_DEBUG, "  size: %" PRIdPTR " -> %" PRIdPTR, prev_size,
+            prev_size + 1);
+  }
+  if (prev_size == 0) {
+    if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+      gpr_log(GPR_DEBUG, "  EXECUTING IMMEDIATELY");
+    }
+    // Queue was empty, so execute this closure immediately.
+    GRPC_CLOSURE_SCHED(exec_ctx, closure, error);
+  } else {
+    if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+      gpr_log(GPR_INFO, "  QUEUING");
+    }
+    // Queue was not empty, so add closure to queue.
+    closure->error_data.error = error;
+    gpr_mpscq_push(&call_combiner->queue, (gpr_mpscq_node*)closure);
+  }
+}
+
+void grpc_call_combiner_stop(grpc_exec_ctx* exec_ctx,
+                             grpc_call_combiner* call_combiner DEBUG_ARGS,
+                             const char* reason) {
+  if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+    gpr_log(GPR_DEBUG,
+            "==> grpc_call_combiner_stop() [%p] [" DEBUG_FMT_STR "%s]",
+            call_combiner DEBUG_FMT_ARGS, reason);
+  }
+  size_t prev_size =
+      (size_t)gpr_atm_full_fetch_add(&call_combiner->size, (gpr_atm)-1);
+  if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+    gpr_log(GPR_DEBUG, "  size: %" PRIdPTR " -> %" PRIdPTR, prev_size,
+            prev_size - 1);
+  }
+  GPR_ASSERT(prev_size >= 1);
+  if (prev_size > 1) {
+    while (true) {
+      if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+        gpr_log(GPR_DEBUG, "  checking queue");
+      }
+      bool empty;
+      grpc_closure* closure = (grpc_closure*)gpr_mpscq_pop_and_check_end(
+          &call_combiner->queue, &empty);
+      if (closure == NULL) {
+        // This can happen either due to a race condition within the mpscq
+        // code or because of a race with grpc_call_combiner_start().
+        if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+          gpr_log(GPR_DEBUG, "  queue returned no result; checking again");
+        }
+        continue;
+      }
+      if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+        gpr_log(GPR_DEBUG, "  EXECUTING FROM QUEUE: closure=%p error=%s",
+                closure, grpc_error_string(closure->error_data.error));
+      }
+      GRPC_CLOSURE_SCHED(exec_ctx, closure, closure->error_data.error);
+      break;
+    }
+  } else if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+    gpr_log(GPR_DEBUG, "  queue empty");
+  }
+}
+
+void grpc_call_combiner_set_notify_on_cancel(grpc_exec_ctx* exec_ctx,
+                                             grpc_call_combiner* call_combiner,
+                                             grpc_closure* closure) {
+  while (true) {
+    // Decode original state.
+    gpr_atm original_state = gpr_atm_acq_load(&call_combiner->cancel_state);
+    grpc_error* original_error = decode_cancel_state_error(original_state);
+    // If error is set, invoke the cancellation closure immediately.
+    // Otherwise, store the new closure.
+    if (original_error != GRPC_ERROR_NONE) {
+      if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+        gpr_log(GPR_DEBUG,
+                "call_combiner=%p: scheduling notify_on_cancel callback=%p "
+                "for pre-existing cancellation",
+                call_combiner, closure);
+      }
+      GRPC_CLOSURE_SCHED(exec_ctx, closure, GRPC_ERROR_REF(original_error));
+      break;
+    } else {
+      if (gpr_atm_full_cas(&call_combiner->cancel_state, original_state,
+                           (gpr_atm)closure)) {
+        if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+          gpr_log(GPR_DEBUG, "call_combiner=%p: setting notify_on_cancel=%p",
+                  call_combiner, closure);
+        }
+        // If we replaced an earlier closure, invoke the original
+        // closure with GRPC_ERROR_NONE.  This allows callers to clean
+        // up any resources they may be holding for the callback.
+        if (original_state != 0) {
+          closure = (grpc_closure*)original_state;
+          if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+            gpr_log(GPR_DEBUG,
+                    "call_combiner=%p: scheduling old cancel callback=%p",
+                    call_combiner, closure);
+          }
+          GRPC_CLOSURE_SCHED(exec_ctx, closure, GRPC_ERROR_NONE);
+        }
+        break;
+      }
+    }
+    // cas failed, try again.
+  }
+}
+
+void grpc_call_combiner_cancel(grpc_exec_ctx* exec_ctx,
+                               grpc_call_combiner* call_combiner,
+                               grpc_error* error) {
+  while (true) {
+    gpr_atm original_state = gpr_atm_acq_load(&call_combiner->cancel_state);
+    grpc_error* original_error = decode_cancel_state_error(original_state);
+    if (original_error != GRPC_ERROR_NONE) {
+      GRPC_ERROR_UNREF(error);
+      break;
+    }
+    if (gpr_atm_full_cas(&call_combiner->cancel_state, original_state,
+                         encode_cancel_state_error(error))) {
+      if (original_state != 0) {
+        grpc_closure* notify_on_cancel = (grpc_closure*)original_state;
+        if (GRPC_TRACER_ON(grpc_call_combiner_trace)) {
+          gpr_log(GPR_DEBUG,
+                  "call_combiner=%p: scheduling notify_on_cancel callback=%p",
+                  call_combiner, notify_on_cancel);
+        }
+        GRPC_CLOSURE_SCHED(exec_ctx, notify_on_cancel, GRPC_ERROR_REF(error));
+      }
+      break;
+    }
+    // cas failed, try again.
+  }
+}
diff --git a/src/core/lib/iomgr/call_combiner.h b/src/core/lib/iomgr/call_combiner.h
new file mode 100644
index 0000000000..5cfb3f0c07
--- /dev/null
+++ b/src/core/lib/iomgr/call_combiner.h
@@ -0,0 +1,121 @@
+/*
+ *
+ * Copyright 2017 gRPC authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ */
+
+#ifndef GRPC_CORE_LIB_IOMGR_CALL_COMBINER_H
+#define GRPC_CORE_LIB_IOMGR_CALL_COMBINER_H
+
+#include <stddef.h>
+
+#include <grpc/support/atm.h>
+
+#include "src/core/lib/iomgr/closure.h"
+#include "src/core/lib/iomgr/exec_ctx.h"
+#include "src/core/lib/support/mpscq.h"
+
+// A simple, lock-free mechanism for serializing activity related to a
+// single call.  This is similar to a combiner but is more lightweight.
+//
+// It requires the callback (or, in the common case where the callback
+// actually kicks off a chain of callbacks, the last callback in that
+// chain) to explicitly indicate (by calling GRPC_CALL_COMBINER_STOP())
+// when it is done with the action that was kicked off by the original
+// callback.
+
+extern grpc_tracer_flag grpc_call_combiner_trace;
+
+typedef struct {
+  gpr_atm size;  // size_t, num closures in queue or currently executing
+  gpr_mpscq queue;
+  // Either 0 (if not cancelled and no cancellation closure set),
+  // a grpc_closure* (if the lowest bit is 0),
+  // or a grpc_error* (if the lowest bit is 1).
+  gpr_atm cancel_state;
+} grpc_call_combiner;
+
+// Assumes memory was initialized to zero.
+void grpc_call_combiner_init(grpc_call_combiner* call_combiner);
+
+void grpc_call_combiner_destroy(grpc_call_combiner* call_combiner);
+
+#ifndef NDEBUG
+#define GRPC_CALL_COMBINER_START(exec_ctx, call_combiner, closure, error,   \
+                                 reason)                                    \
+  grpc_call_combiner_start((exec_ctx), (call_combiner), (closure), (error), \
+                           __FILE__, __LINE__, (reason))
+#define GRPC_CALL_COMBINER_STOP(exec_ctx, call_combiner, reason)           \
+  grpc_call_combiner_stop((exec_ctx), (call_combiner), __FILE__, __LINE__, \
+                          (reason))
+/// Starts processing \a closure on \a call_combiner.
+void grpc_call_combiner_start(grpc_exec_ctx* exec_ctx,
+                              grpc_call_combiner* call_combiner,
+                              grpc_closure* closure, grpc_error* error,
+                              const char* file, int line, const char* reason);
+/// Yields the call combiner to the next closure in the queue, if any.
+void grpc_call_combiner_stop(grpc_exec_ctx* exec_ctx,
+                             grpc_call_combiner* call_combiner,
+                             const char* file, int line, const char* reason);
+#else
+#define GRPC_CALL_COMBINER_START(exec_ctx, call_combiner, closure, error,   \
+                                 reason)                                    \
+  grpc_call_combiner_start((exec_ctx), (call_combiner), (closure), (error), \
+                           (reason))
+#define GRPC_CALL_COMBINER_STOP(exec_ctx, call_combiner, reason) \
+  grpc_call_combiner_stop((exec_ctx), (call_combiner), (reason))
+/// Starts processing \a closure on \a call_combiner.
+void grpc_call_combiner_start(grpc_exec_ctx* exec_ctx,
+                              grpc_call_combiner* call_combiner,
+                              grpc_closure* closure, grpc_error* error,
+                              const char* reason);
+/// Yields the call combiner to the next closure in the queue, if any.
+void grpc_call_combiner_stop(grpc_exec_ctx* exec_ctx,
+                             grpc_call_combiner* call_combiner,
+                             const char* reason);
+#endif
+
+/// Registers \a closure to be invoked by \a call_combiner when
+/// grpc_call_combiner_cancel() is called.
+///
+/// Once a closure is registered, it will always be scheduled exactly
+/// once; this allows the closure to hold references that will be freed
+/// regardless of whether or not the call was cancelled.  If a cancellation
+/// does occur, the closure will be scheduled with the cancellation error;
+/// otherwise, it will be scheduled with GRPC_ERROR_NONE.
+///
+/// The closure will be scheduled in the following cases:
+/// - If grpc_call_combiner_cancel() was called prior to registering the
+///   closure, it will be scheduled immediately with the cancelation error.
+/// - If grpc_call_combiner_cancel() is called after registering the
+///   closure, the closure will be scheduled with the cancellation error.
+/// - If grpc_call_combiner_set_notify_on_cancel() is called again to
+///   register a new cancellation closure, the previous cancellation
+///   closure will be scheduled with GRPC_ERROR_NONE.
+///
+/// If \a closure is NULL, then no closure will be invoked on
+/// cancellation; this effectively unregisters the previously set closure.
+/// However, most filters will not need to explicitly unregister their
+/// callbacks, as this is done automatically when the call is destroyed.
+void grpc_call_combiner_set_notify_on_cancel(grpc_exec_ctx* exec_ctx,
+                                             grpc_call_combiner* call_combiner,
+                                             grpc_closure* closure);
+
+/// Indicates that the call has been cancelled.
+void grpc_call_combiner_cancel(grpc_exec_ctx* exec_ctx,
+                               grpc_call_combiner* call_combiner,
+                               grpc_error* error);
+
+#endif /* GRPC_CORE_LIB_IOMGR_CALL_COMBINER_H */
diff --git a/src/core/lib/iomgr/ev_epoll1_linux.c b/src/core/lib/iomgr/ev_epoll1_linux.c
index 7f053fa728..b76eb9e1c9 100644
--- a/src/core/lib/iomgr/ev_epoll1_linux.c
+++ b/src/core/lib/iomgr/ev_epoll1_linux.c
@@ -698,22 +698,30 @@ static bool begin_worker(grpc_pollset *pollset, grpc_pollset_worker *worker,
         gpr_mu_unlock(&pollset->mu);
         goto retry_lock_neighbourhood;
       }
-      pollset->seen_inactive = false;
-      if (neighbourhood->active_root == NULL) {
-        neighbourhood->active_root = pollset->next = pollset->prev = pollset;
-        /* TODO: sreek. Why would this worker state be other than UNKICKED
-         * here ? (since the worker isn't added to the pollset yet, there is no
-         * way it can be "found" by other threads to get kicked). */
-
-        /* If there is no designated poller, make this the designated poller */
-        if (worker->kick_state == UNKICKED &&
-            gpr_atm_no_barrier_cas(&g_active_poller, 0, (gpr_atm)worker)) {
-          SET_KICK_STATE(worker, DESIGNATED_POLLER);
+
+      /* In the brief time we released the pollset locks above, the worker MAY
+         have been kicked. In this case, the worker should get out of this
+         pollset ASAP and hence this should neither add the pollset to
+         neighbourhood nor mark the pollset as active.
+
+         On a side note, the only way a worker's kick state could have changed
+         at this point is if it were "kicked specifically". Since the worker has
+         not added itself to the pollset yet (by calling worker_insert()), it is
+         not visible in the "kick any" path yet */
+      if (worker->kick_state == UNKICKED) {
+        pollset->seen_inactive = false;
+        if (neighbourhood->active_root == NULL) {
+          neighbourhood->active_root = pollset->next = pollset->prev = pollset;
+          /* Make this the designated poller if there isn't one already */
+          if (worker->kick_state == UNKICKED &&
+              gpr_atm_no_barrier_cas(&g_active_poller, 0, (gpr_atm)worker)) {
+            SET_KICK_STATE(worker, DESIGNATED_POLLER);
+          }
+        } else {
+          pollset->next = neighbourhood->active_root;
+          pollset->prev = pollset->next->prev;
+          pollset->next->prev = pollset->prev->next = pollset;
         }
-      } else {
-        pollset->next = neighbourhood->active_root;
-        pollset->prev = pollset->next->prev;
-        pollset->next->prev = pollset->prev->next = pollset;
       }
     }
     if (is_reassigning) {
@@ -1001,6 +1009,7 @@ static grpc_error *pollset_kick(grpc_pollset *pollset,
     gpr_log(GPR_ERROR, "%s", tmp);
     gpr_free(tmp);
   }
+
   if (specific_worker == NULL) {
     if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)pollset) {
       grpc_pollset_worker *root_worker = pollset->root_worker;
@@ -1076,7 +1085,11 @@ static grpc_error *pollset_kick(grpc_pollset *pollset,
       }
       goto done;
     }
-  } else if (specific_worker->kick_state == KICKED) {
+
+    GPR_UNREACHABLE_CODE(goto done);
+  }
+
+  if (specific_worker->kick_state == KICKED) {
     if (GRPC_TRACER_ON(grpc_polling_trace)) {
       gpr_log(GPR_ERROR, " .. specific worker already kicked");
     }
diff --git a/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c b/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c
deleted file mode 100644
index e2e3cd9003..0000000000
--- a/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c
+++ /dev/null
@@ -1,1961 +0,0 @@
-/*
- *
- * Copyright 2017 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 "src/core/lib/iomgr/port.h"
-
-/* This polling engine is only relevant on linux kernels supporting epoll() */
-#ifdef GRPC_LINUX_EPOLL
-
-#include "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h"
-
-#include <assert.h>
-#include <errno.h>
-#include <limits.h>
-#include <poll.h>
-#include <pthread.h>
-#include <signal.h>
-#include <string.h>
-#include <sys/epoll.h>
-#include <sys/socket.h>
-#include <unistd.h>
-
-#include <grpc/support/alloc.h>
-#include <grpc/support/log.h>
-#include <grpc/support/string_util.h>
-#include <grpc/support/tls.h>
-#include <grpc/support/useful.h>
-
-#include "src/core/lib/debug/stats.h"
-#include "src/core/lib/debug/trace.h"
-#include "src/core/lib/iomgr/ev_posix.h"
-#include "src/core/lib/iomgr/iomgr_internal.h"
-#include "src/core/lib/iomgr/lockfree_event.h"
-#include "src/core/lib/iomgr/timer.h"
-#include "src/core/lib/iomgr/wakeup_fd_posix.h"
-#include "src/core/lib/profiling/timers.h"
-#include "src/core/lib/support/block_annotate.h"
-#include "src/core/lib/support/env.h"
-
-#define GRPC_POLLING_TRACE(fmt, ...)        \
-  if (GRPC_TRACER_ON(grpc_polling_trace)) { \
-    gpr_log(GPR_INFO, (fmt), __VA_ARGS__);  \
-  }
-
-#define GRPC_POLLSET_KICK_BROADCAST ((grpc_pollset_worker *)1)
-
-/* The maximum number of polling threads per polling island. By default no
-   limit */
-static int g_max_pollers_per_pi = INT_MAX;
-
-static int grpc_wakeup_signal = -1;
-static bool is_grpc_wakeup_signal_initialized = false;
-
-/* Implements the function defined in grpc_posix.h. This function might be
- * called before even calling grpc_init() to set either a different signal to
- * use. If signum == -1, then the use of signals is disabled */
-static void grpc_use_signal(int signum) {
-  grpc_wakeup_signal = signum;
-  is_grpc_wakeup_signal_initialized = true;
-
-  if (grpc_wakeup_signal < 0) {
-    gpr_log(GPR_INFO,
-            "Use of signals is disabled. Epoll engine will not be used");
-  } else {
-    gpr_log(GPR_INFO, "epoll engine will be using signal: %d",
-            grpc_wakeup_signal);
-  }
-}
-
-struct polling_island;
-
-typedef enum {
-  POLL_OBJ_FD,
-  POLL_OBJ_POLLSET,
-  POLL_OBJ_POLLSET_SET
-} poll_obj_type;
-
-typedef struct poll_obj {
-#ifndef NDEBUG
-  poll_obj_type obj_type;
-#endif
-  gpr_mu mu;
-  struct polling_island *pi;
-} poll_obj;
-
-static const char *poll_obj_string(poll_obj_type po_type) {
-  switch (po_type) {
-    case POLL_OBJ_FD:
-      return "fd";
-    case POLL_OBJ_POLLSET:
-      return "pollset";
-    case POLL_OBJ_POLLSET_SET:
-      return "pollset_set";
-  }
-
-  GPR_UNREACHABLE_CODE(return "UNKNOWN");
-}
-
-/*******************************************************************************
- * Fd Declarations
- */
-
-#define FD_FROM_PO(po) ((grpc_fd *)(po))
-
-struct grpc_fd {
-  poll_obj po;
-
-  int fd;
-  /* refst format:
-       bit 0    : 1=Active / 0=Orphaned
-       bits 1-n : refcount
-     Ref/Unref by two to avoid altering the orphaned bit */
-  gpr_atm refst;
-
-  /* The fd is either closed or we relinquished control of it. In either
-     cases, this indicates that the 'fd' on this structure is no longer
-     valid */
-  bool orphaned;
-
-  gpr_atm read_closure;
-  gpr_atm write_closure;
-
-  struct grpc_fd *freelist_next;
-  grpc_closure *on_done_closure;
-
-  /* The pollset that last noticed that the fd is readable. The actual type
-   * stored in this is (grpc_pollset *) */
-  gpr_atm read_notifier_pollset;
-
-  grpc_iomgr_object iomgr_object;
-};
-
-/* Reference counting for fds */
-#ifndef NDEBUG
-static void fd_ref(grpc_fd *fd, const char *reason, const char *file, int line);
-static void fd_unref(grpc_fd *fd, const char *reason, const char *file,
-                     int line);
-#define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__)
-#define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__)
-#else
-static void fd_ref(grpc_fd *fd);
-static void fd_unref(grpc_fd *fd);
-#define GRPC_FD_REF(fd, reason) fd_ref(fd)
-#define GRPC_FD_UNREF(fd, reason) fd_unref(fd)
-#endif
-
-static void fd_global_init(void);
-static void fd_global_shutdown(void);
-
-/*******************************************************************************
- * Polling island Declarations
- */
-
-#ifndef NDEBUG
-
-#define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__)
-#define PI_UNREF(exec_ctx, p, r) \
-  pi_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__)
-
-#else
-
-#define PI_ADD_REF(p, r) pi_add_ref((p))
-#define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p))
-
-#endif
-
-typedef struct worker_node {
-  struct worker_node *next;
-  struct worker_node *prev;
-} worker_node;
-
-/* This is also used as grpc_workqueue (by directly casing it) */
-typedef struct polling_island {
-  gpr_mu mu;
-  /* Ref count. Use PI_ADD_REF() and PI_UNREF() macros to increment/decrement
-     the refcount.
-     Once the ref count becomes zero, this structure is destroyed which means
-     we should ensure that there is never a scenario where a PI_ADD_REF() is
-     racing with a PI_UNREF() that just made the ref_count zero. */
-  gpr_atm ref_count;
-
-  /* Pointer to the polling_island this merged into.
-   * merged_to value is only set once in polling_island's lifetime (and that too
-   * only if the island is merged with another island). Because of this, we can
-   * use gpr_atm type here so that we can do atomic access on this and reduce
-   * lock contention on 'mu' mutex.
-   *
-   * Note that if this field is not NULL (i.e not 0), all the remaining fields
-   * (except mu and ref_count) are invalid and must be ignored. */
-  gpr_atm merged_to;
-
-  /* Number of threads currently polling on this island */
-  gpr_atm poller_count;
-
-  /* The list of workers waiting to do polling on this polling island */
-  gpr_mu worker_list_mu;
-  worker_node worker_list_head;
-
-  /* The fd of the underlying epoll set */
-  int epoll_fd;
-
-  /* The file descriptors in the epoll set */
-  size_t fd_cnt;
-  size_t fd_capacity;
-  grpc_fd **fds;
-} polling_island;
-
-/*******************************************************************************
- * Pollset Declarations
- */
-#define WORKER_FROM_WORKER_LIST_NODE(p)          \
-  (struct grpc_pollset_worker *)(((char *)(p)) - \
-                                 offsetof(grpc_pollset_worker, pi_list_link))
-struct grpc_pollset_worker {
-  /* Thread id of this worker */
-  pthread_t pt_id;
-
-  /* Used to prevent a worker from getting kicked multiple times */
-  gpr_atm is_kicked;
-
-  struct grpc_pollset_worker *next;
-  struct grpc_pollset_worker *prev;
-
-  /* Indicates if it is this worker's turn to do epoll */
-  gpr_atm is_polling_turn;
-
-  /* Node in the polling island's worker list. */
-  worker_node pi_list_link;
-};
-
-struct grpc_pollset {
-  poll_obj po;
-
-  grpc_pollset_worker root_worker;
-  bool kicked_without_pollers;
-
-  bool shutting_down;          /* Is the pollset shutting down ? */
-  bool finish_shutdown_called; /* Is the 'finish_shutdown_locked()' called ? */
-  grpc_closure *shutdown_done; /* Called after after shutdown is complete */
-};
-
-/*******************************************************************************
- * Pollset-set Declarations
- */
-struct grpc_pollset_set {
-  poll_obj po;
-};
-
-/*******************************************************************************
- * Common helpers
- */
-
-static bool append_error(grpc_error **composite, grpc_error *error,
-                         const char *desc) {
-  if (error == GRPC_ERROR_NONE) return true;
-  if (*composite == GRPC_ERROR_NONE) {
-    *composite = GRPC_ERROR_CREATE_FROM_COPIED_STRING(desc);
-  }
-  *composite = grpc_error_add_child(*composite, error);
-  return false;
-}
-
-/*******************************************************************************
- * Polling island Definitions
- */
-
-/* The wakeup fd that is used to wake up all threads in a Polling island. This
-   is useful in the polling island merge operation where we need to wakeup all
-   the threads currently polling the smaller polling island (so that they can
-   start polling the new/merged polling island)
-
-   NOTE: This fd is initialized to be readable and MUST NOT be consumed i.e the
-   threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */
-static grpc_wakeup_fd polling_island_wakeup_fd;
-
-/* The polling island being polled right now.
-   See comments in workqueue_maybe_wakeup for why this is tracked. */
-static __thread polling_island *g_current_thread_polling_island;
-
-/* Forward declaration */
-static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi);
-
-#ifdef GRPC_TSAN
-/* Currently TSAN may incorrectly flag data races between epoll_ctl and
-   epoll_wait for any grpc_fd structs that are added to the epoll set via
-   epoll_ctl and are returned (within a very short window) via epoll_wait().
-
-   To work-around this race, we establish a happens-before relation between
-   the code just-before epoll_ctl() and the code after epoll_wait() by using
-   this atomic */
-gpr_atm g_epoll_sync;
-#endif /* defined(GRPC_TSAN) */
-
-static void pi_add_ref(polling_island *pi);
-static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi);
-
-#ifndef NDEBUG
-static void pi_add_ref_dbg(polling_island *pi, const char *reason,
-                           const char *file, int line) {
-  if (GRPC_TRACER_ON(grpc_polling_trace)) {
-    gpr_atm old_cnt = gpr_atm_acq_load(&pi->ref_count);
-    gpr_log(GPR_DEBUG, "Add ref pi: %p, old:%" PRIdPTR " -> new:%" PRIdPTR
-                       " (%s) - (%s, %d)",
-            pi, old_cnt, old_cnt + 1, reason, file, line);
-  }
-  pi_add_ref(pi);
-}
-
-static void pi_unref_dbg(grpc_exec_ctx *exec_ctx, polling_island *pi,
-                         const char *reason, const char *file, int line) {
-  if (GRPC_TRACER_ON(grpc_polling_trace)) {
-    gpr_atm old_cnt = gpr_atm_acq_load(&pi->ref_count);
-    gpr_log(GPR_DEBUG, "Unref pi: %p, old:%" PRIdPTR " -> new:%" PRIdPTR
-                       " (%s) - (%s, %d)",
-            pi, old_cnt, (old_cnt - 1), reason, file, line);
-  }
-  pi_unref(exec_ctx, pi);
-}
-#endif
-
-static void pi_add_ref(polling_island *pi) {
-  gpr_atm_no_barrier_fetch_add(&pi->ref_count, 1);
-}
-
-static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) {
-  /* If ref count went to zero, delete the polling island.
-     Note that this deletion not be done under a lock. Once the ref count goes
-     to zero, we are guaranteed that no one else holds a reference to the
-     polling island (and that there is no racing pi_add_ref() call either).
-
-     Also, if we are deleting the polling island and the merged_to field is
-     non-empty, we should remove a ref to the merged_to polling island
-   */
-  if (1 == gpr_atm_full_fetch_add(&pi->ref_count, -1)) {
-    polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
-    polling_island_delete(exec_ctx, pi);
-    if (next != NULL) {
-      PI_UNREF(exec_ctx, next, "pi_delete"); /* Recursive call */
-    }
-  }
-}
-
-static void worker_node_init(worker_node *node) {
-  node->next = node->prev = node;
-}
-
-/* Not thread safe. Do under a list-level lock */
-static void push_back_worker_node(worker_node *head, worker_node *node) {
-  node->next = head;
-  node->prev = head->prev;
-  head->prev->next = node;
-  head->prev = node;
-}
-
-/* Not thread safe. Do under a list-level lock */
-static void remove_worker_node(worker_node *node) {
-  node->next->prev = node->prev;
-  node->prev->next = node->next;
-  /* If node's next and prev point to itself, the node is considered detached
-   * from the list*/
-  node->next = node->prev = node;
-}
-
-/* Not thread safe. Do under a list-level lock */
-static worker_node *pop_front_worker_node(worker_node *head) {
-  worker_node *node = head->next;
-  if (node != head) {
-    remove_worker_node(node);
-  } else {
-    node = NULL;
-  }
-
-  return node;
-}
-
-/* Returns true if the node's next and prev are pointing to itself (which
-   indicates that the node is not in the list */
-static bool is_worker_node_detached(worker_node *node) {
-  return (node->next == node->prev && node->next == node);
-}
-
-/* The caller is expected to hold pi->mu lock before calling this function
- */
-static void polling_island_add_fds_locked(polling_island *pi, grpc_fd **fds,
-                                          size_t fd_count, bool add_fd_refs,
-                                          grpc_error **error) {
-  int err;
-  size_t i;
-  struct epoll_event ev;
-  char *err_msg;
-  const char *err_desc = "polling_island_add_fds";
-
-#ifdef GRPC_TSAN
-  /* See the definition of g_epoll_sync for more context */
-  gpr_atm_rel_store(&g_epoll_sync, (gpr_atm)0);
-#endif /* defined(GRPC_TSAN) */
-
-  for (i = 0; i < fd_count; i++) {
-    ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET);
-    ev.data.ptr = fds[i];
-    err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, fds[i]->fd, &ev);
-
-    if (err < 0) {
-      if (errno != EEXIST) {
-        gpr_asprintf(
-            &err_msg,
-            "epoll_ctl (epoll_fd: %d) add fd: %d failed with error: %d (%s)",
-            pi->epoll_fd, fds[i]->fd, errno, strerror(errno));
-        append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-        gpr_free(err_msg);
-      }
-
-      continue;
-    }
-
-    if (pi->fd_cnt == pi->fd_capacity) {
-      pi->fd_capacity = GPR_MAX(pi->fd_capacity + 8, pi->fd_cnt * 3 / 2);
-      pi->fds = gpr_realloc(pi->fds, sizeof(grpc_fd *) * pi->fd_capacity);
-    }
-
-    pi->fds[pi->fd_cnt++] = fds[i];
-    if (add_fd_refs) {
-      GRPC_FD_REF(fds[i], "polling_island");
-    }
-  }
-}
-
-/* The caller is expected to hold pi->mu before calling this */
-static void polling_island_add_wakeup_fd_locked(polling_island *pi,
-                                                grpc_wakeup_fd *wakeup_fd,
-                                                grpc_error **error) {
-  struct epoll_event ev;
-  int err;
-  char *err_msg;
-  const char *err_desc = "polling_island_add_wakeup_fd";
-
-  ev.events = (uint32_t)(EPOLLIN | EPOLLET);
-  ev.data.ptr = wakeup_fd;
-  err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD,
-                  GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), &ev);
-  if (err < 0 && errno != EEXIST) {
-    gpr_asprintf(&err_msg,
-                 "epoll_ctl (epoll_fd: %d) add wakeup fd: %d failed with "
-                 "error: %d (%s)",
-                 pi->epoll_fd, GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), errno,
-                 strerror(errno));
-    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-    gpr_free(err_msg);
-  }
-}
-
-/* The caller is expected to hold pi->mu lock before calling this function */
-static void polling_island_remove_all_fds_locked(polling_island *pi,
-                                                 bool remove_fd_refs,
-                                                 grpc_error **error) {
-  int err;
-  size_t i;
-  char *err_msg;
-  const char *err_desc = "polling_island_remove_fds";
-
-  for (i = 0; i < pi->fd_cnt; i++) {
-    err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, pi->fds[i]->fd, NULL);
-    if (err < 0 && errno != ENOENT) {
-      gpr_asprintf(&err_msg,
-                   "epoll_ctl (epoll_fd: %d) delete fds[%zu]: %d failed with "
-                   "error: %d (%s)",
-                   pi->epoll_fd, i, pi->fds[i]->fd, errno, strerror(errno));
-      append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-      gpr_free(err_msg);
-    }
-
-    if (remove_fd_refs) {
-      GRPC_FD_UNREF(pi->fds[i], "polling_island");
-    }
-  }
-
-  pi->fd_cnt = 0;
-}
-
-/* The caller is expected to hold pi->mu lock before calling this function */
-static void polling_island_remove_fd_locked(polling_island *pi, grpc_fd *fd,
-                                            bool is_fd_closed,
-                                            grpc_error **error) {
-  int err;
-  size_t i;
-  char *err_msg;
-  const char *err_desc = "polling_island_remove_fd";
-
-  /* If fd is already closed, then it would have been automatically been removed
-     from the epoll set */
-  if (!is_fd_closed) {
-    err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, fd->fd, NULL);
-    if (err < 0 && errno != ENOENT) {
-      gpr_asprintf(
-          &err_msg,
-          "epoll_ctl (epoll_fd: %d) del fd: %d failed with error: %d (%s)",
-          pi->epoll_fd, fd->fd, errno, strerror(errno));
-      append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-      gpr_free(err_msg);
-    }
-  }
-
-  for (i = 0; i < pi->fd_cnt; i++) {
-    if (pi->fds[i] == fd) {
-      pi->fds[i] = pi->fds[--pi->fd_cnt];
-      GRPC_FD_UNREF(fd, "polling_island");
-      break;
-    }
-  }
-}
-
-/* Might return NULL in case of an error */
-static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx,
-                                             grpc_fd *initial_fd,
-                                             grpc_error **error) {
-  polling_island *pi = NULL;
-  const char *err_desc = "polling_island_create";
-
-  *error = GRPC_ERROR_NONE;
-
-  pi = gpr_malloc(sizeof(*pi));
-  gpr_mu_init(&pi->mu);
-  pi->fd_cnt = 0;
-  pi->fd_capacity = 0;
-  pi->fds = NULL;
-  pi->epoll_fd = -1;
-
-  gpr_atm_rel_store(&pi->ref_count, 0);
-  gpr_atm_rel_store(&pi->poller_count, 0);
-  gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL);
-
-  gpr_mu_init(&pi->worker_list_mu);
-  worker_node_init(&pi->worker_list_head);
-
-  pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
-
-  if (pi->epoll_fd < 0) {
-    append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc);
-    goto done;
-  }
-
-  if (initial_fd != NULL) {
-    polling_island_add_fds_locked(pi, &initial_fd, 1, true, error);
-  }
-
-done:
-  if (*error != GRPC_ERROR_NONE) {
-    polling_island_delete(exec_ctx, pi);
-    pi = NULL;
-  }
-  return pi;
-}
-
-static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) {
-  GPR_ASSERT(pi->fd_cnt == 0);
-
-  if (pi->epoll_fd >= 0) {
-    close(pi->epoll_fd);
-  }
-  gpr_mu_destroy(&pi->mu);
-  gpr_mu_destroy(&pi->worker_list_mu);
-  GPR_ASSERT(is_worker_node_detached(&pi->worker_list_head));
-
-  gpr_free(pi->fds);
-  gpr_free(pi);
-}
-
-/* Attempts to gets the last polling island in the linked list (liked by the
- * 'merged_to' field). Since this does not lock the polling island, there are no
- * guarantees that the island returned is the last island */
-static polling_island *polling_island_maybe_get_latest(polling_island *pi) {
-  polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
-  while (next != NULL) {
-    pi = next;
-    next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
-  }
-
-  return pi;
-}
-
-/* Gets the lock on the *latest* polling island i.e the last polling island in
-   the linked list (linked by the 'merged_to' field). Call gpr_mu_unlock on the
-   returned polling island's mu.
-   Usage: To lock/unlock polling island "pi", do the following:
-      polling_island *pi_latest = polling_island_lock(pi);
-      ...
-      ... critical section ..
-      ...
-      gpr_mu_unlock(&pi_latest->mu); // NOTE: use pi_latest->mu. NOT pi->mu */
-static polling_island *polling_island_lock(polling_island *pi) {
-  polling_island *next = NULL;
-
-  while (true) {
-    next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
-    if (next == NULL) {
-      /* Looks like 'pi' is the last node in the linked list but unless we check
-         this by holding the pi->mu lock, we cannot be sure (i.e without the
-         pi->mu lock, we don't prevent island merges).
-         To be absolutely sure, check once more by holding the pi->mu lock */
-      gpr_mu_lock(&pi->mu);
-      next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
-      if (next == NULL) {
-        /* pi is infact the last node and we have the pi->mu lock. we're done */
-        break;
-      }
-
-      /* pi->merged_to is not NULL i.e pi isn't the last node anymore. pi->mu
-       * isn't the lock we are interested in. Continue traversing the list */
-      gpr_mu_unlock(&pi->mu);
-    }
-
-    pi = next;
-  }
-
-  return pi;
-}
-
-/* Gets the lock on the *latest* polling islands in the linked lists pointed by
-   *p and *q (and also updates *p and *q to point to the latest polling islands)
-
-   This function is needed because calling the following block of code to obtain
-   locks on polling islands (*p and *q) is prone to deadlocks.
-     {
-       polling_island_lock(*p, true);
-       polling_island_lock(*q, true);
-     }
-
-   Usage/example:
-     polling_island *p1;
-     polling_island *p2;
-     ..
-     polling_island_lock_pair(&p1, &p2);
-     ..
-     .. Critical section with both p1 and p2 locked
-     ..
-     // Release locks: Always call polling_island_unlock_pair() to release locks
-     polling_island_unlock_pair(p1, p2);
-*/
-static void polling_island_lock_pair(polling_island **p, polling_island **q) {
-  polling_island *pi_1 = *p;
-  polling_island *pi_2 = *q;
-  polling_island *next_1 = NULL;
-  polling_island *next_2 = NULL;
-
-  /* The algorithm is simple:
-      - Go to the last polling islands in the linked lists *pi_1 and *pi_2 (and
-        keep updating pi_1 and pi_2)
-      - Then obtain locks on the islands by following a lock order rule of
-        locking polling_island with lower address first
-           Special case: Before obtaining the locks, check if pi_1 and pi_2 are
-           pointing to the same island. If that is the case, we can just call
-           polling_island_lock()
-      - After obtaining both the locks, double check that the polling islands
-        are still the last polling islands in their respective linked lists
-        (this is because there might have been polling island merges before
-        we got the lock)
-      - If the polling islands are the last islands, we are done. If not,
-        release the locks and continue the process from the first step */
-  while (true) {
-    next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to);
-    while (next_1 != NULL) {
-      pi_1 = next_1;
-      next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to);
-    }
-
-    next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to);
-    while (next_2 != NULL) {
-      pi_2 = next_2;
-      next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to);
-    }
-
-    if (pi_1 == pi_2) {
-      pi_1 = pi_2 = polling_island_lock(pi_1);
-      break;
-    }
-
-    if (pi_1 < pi_2) {
-      gpr_mu_lock(&pi_1->mu);
-      gpr_mu_lock(&pi_2->mu);
-    } else {
-      gpr_mu_lock(&pi_2->mu);
-      gpr_mu_lock(&pi_1->mu);
-    }
-
-    next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to);
-    next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to);
-    if (next_1 == NULL && next_2 == NULL) {
-      break;
-    }
-
-    gpr_mu_unlock(&pi_1->mu);
-    gpr_mu_unlock(&pi_2->mu);
-  }
-
-  *p = pi_1;
-  *q = pi_2;
-}
-
-static void polling_island_unlock_pair(polling_island *p, polling_island *q) {
-  if (p == q) {
-    gpr_mu_unlock(&p->mu);
-  } else {
-    gpr_mu_unlock(&p->mu);
-    gpr_mu_unlock(&q->mu);
-  }
-}
-
-static polling_island *polling_island_merge(polling_island *p,
-                                            polling_island *q,
-                                            grpc_error **error) {
-  /* Get locks on both the polling islands */
-  polling_island_lock_pair(&p, &q);
-
-  if (p != q) {
-    /* Make sure that p points to the polling island with fewer fds than q */
-    if (p->fd_cnt > q->fd_cnt) {
-      GPR_SWAP(polling_island *, p, q);
-    }
-
-    /* Merge p with q i.e move all the fds from p (The one with fewer fds) to q
-       Note that the refcounts on the fds being moved will not change here.
-       This is why the last param in the following two functions is 'false') */
-    polling_island_add_fds_locked(q, p->fds, p->fd_cnt, false, error);
-    polling_island_remove_all_fds_locked(p, false, error);
-
-    /* Wakeup all the pollers (if any) on p so that they pickup this change */
-    polling_island_add_wakeup_fd_locked(p, &polling_island_wakeup_fd, error);
-
-    /* Add the 'merged_to' link from p --> q */
-    gpr_atm_rel_store(&p->merged_to, (gpr_atm)q);
-    PI_ADD_REF(q, "pi_merge"); /* To account for the new incoming ref from p */
-  }
-  /* else if p == q, nothing needs to be done */
-
-  polling_island_unlock_pair(p, q);
-
-  /* Return the merged polling island (Note that no merge would have happened
-     if p == q which is ok) */
-  return q;
-}
-
-static grpc_error *polling_island_global_init() {
-  grpc_error *error = GRPC_ERROR_NONE;
-
-  error = grpc_wakeup_fd_init(&polling_island_wakeup_fd);
-  if (error == GRPC_ERROR_NONE) {
-    error = grpc_wakeup_fd_wakeup(&polling_island_wakeup_fd);
-  }
-
-  return error;
-}
-
-static void polling_island_global_shutdown() {
-  grpc_wakeup_fd_destroy(&polling_island_wakeup_fd);
-}
-
-/*******************************************************************************
- * Fd Definitions
- */
-
-/* We need to keep a freelist not because of any concerns of malloc performance
- * but instead so that implementations with multiple threads in (for example)
- * epoll_wait deal with the race between pollset removal and incoming poll
- * notifications.
- *
- * The problem is that the poller ultimately holds a reference to this
- * object, so it is very difficult to know when is safe to free it, at least
- * without some expensive synchronization.
- *
- * If we keep the object freelisted, in the worst case losing this race just
- * becomes a spurious read notification on a reused fd.
- */
-
-/* The alarm system needs to be able to wakeup 'some poller' sometimes
- * (specifically when a new alarm needs to be triggered earlier than the next
- * alarm 'epoch'). This wakeup_fd gives us something to alert on when such a
- * case occurs. */
-
-static grpc_fd *fd_freelist = NULL;
-static gpr_mu fd_freelist_mu;
-
-#ifndef NDEBUG
-#define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__)
-#define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__)
-static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file,
-                   int line) {
-  if (GRPC_TRACER_ON(grpc_trace_fd_refcount)) {
-    gpr_log(GPR_DEBUG,
-            "FD %d %p   ref %d %" PRIdPTR " -> %" PRIdPTR " [%s; %s:%d]",
-            fd->fd, fd, n, gpr_atm_no_barrier_load(&fd->refst),
-            gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line);
-  }
-#else
-#define REF_BY(fd, n, reason) ref_by(fd, n)
-#define UNREF_BY(fd, n, reason) unref_by(fd, n)
-static void ref_by(grpc_fd *fd, int n) {
-#endif
-  GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0);
-}
-
-#ifndef NDEBUG
-static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file,
-                     int line) {
-  if (GRPC_TRACER_ON(grpc_trace_fd_refcount)) {
-    gpr_log(GPR_DEBUG,
-            "FD %d %p unref %d %" PRIdPTR " -> %" PRIdPTR " [%s; %s:%d]",
-            fd->fd, fd, n, gpr_atm_no_barrier_load(&fd->refst),
-            gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line);
-  }
-#else
-static void unref_by(grpc_fd *fd, int n) {
-#endif
-  gpr_atm old = gpr_atm_full_fetch_add(&fd->refst, -n);
-  if (old == n) {
-    /* Add the fd to the freelist */
-    gpr_mu_lock(&fd_freelist_mu);
-    fd->freelist_next = fd_freelist;
-    fd_freelist = fd;
-    grpc_iomgr_unregister_object(&fd->iomgr_object);
-
-    grpc_lfev_destroy(&fd->read_closure);
-    grpc_lfev_destroy(&fd->write_closure);
-
-    gpr_mu_unlock(&fd_freelist_mu);
-  } else {
-    GPR_ASSERT(old > n);
-  }
-}
-
-/* Increment refcount by two to avoid changing the orphan bit */
-#ifndef NDEBUG
-static void fd_ref(grpc_fd *fd, const char *reason, const char *file,
-                   int line) {
-  ref_by(fd, 2, reason, file, line);
-}
-
-static void fd_unref(grpc_fd *fd, const char *reason, const char *file,
-                     int line) {
-  unref_by(fd, 2, reason, file, line);
-}
-#else
-static void fd_ref(grpc_fd *fd) { ref_by(fd, 2); }
-static void fd_unref(grpc_fd *fd) { unref_by(fd, 2); }
-#endif
-
-static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); }
-
-static void fd_global_shutdown(void) {
-  gpr_mu_lock(&fd_freelist_mu);
-  gpr_mu_unlock(&fd_freelist_mu);
-  while (fd_freelist != NULL) {
-    grpc_fd *fd = fd_freelist;
-    fd_freelist = fd_freelist->freelist_next;
-    gpr_mu_destroy(&fd->po.mu);
-    gpr_free(fd);
-  }
-  gpr_mu_destroy(&fd_freelist_mu);
-}
-
-static grpc_fd *fd_create(int fd, const char *name) {
-  grpc_fd *new_fd = NULL;
-
-  gpr_mu_lock(&fd_freelist_mu);
-  if (fd_freelist != NULL) {
-    new_fd = fd_freelist;
-    fd_freelist = fd_freelist->freelist_next;
-  }
-  gpr_mu_unlock(&fd_freelist_mu);
-
-  if (new_fd == NULL) {
-    new_fd = gpr_malloc(sizeof(grpc_fd));
-    gpr_mu_init(&new_fd->po.mu);
-  }
-
-  /* Note: It is not really needed to get the new_fd->po.mu lock here. If this
-   * is a newly created fd (or an fd we got from the freelist), no one else
-   * would be holding a lock to it anyway. */
-  gpr_mu_lock(&new_fd->po.mu);
-  new_fd->po.pi = NULL;
-#ifndef NDEBUG
-  new_fd->po.obj_type = POLL_OBJ_FD;
-#endif
-
-  gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1);
-  new_fd->fd = fd;
-  new_fd->orphaned = false;
-  grpc_lfev_init(&new_fd->read_closure);
-  grpc_lfev_init(&new_fd->write_closure);
-  gpr_atm_no_barrier_store(&new_fd->read_notifier_pollset, (gpr_atm)NULL);
-
-  new_fd->freelist_next = NULL;
-  new_fd->on_done_closure = NULL;
-
-  gpr_mu_unlock(&new_fd->po.mu);
-
-  char *fd_name;
-  gpr_asprintf(&fd_name, "%s fd=%d", name, fd);
-  grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name);
-#ifndef NDEBUG
-  if (GRPC_TRACER_ON(grpc_trace_fd_refcount)) {
-    gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, new_fd, fd_name);
-  }
-#endif
-  gpr_free(fd_name);
-  return new_fd;
-}
-
-static int fd_wrapped_fd(grpc_fd *fd) {
-  int ret_fd = -1;
-  gpr_mu_lock(&fd->po.mu);
-  if (!fd->orphaned) {
-    ret_fd = fd->fd;
-  }
-  gpr_mu_unlock(&fd->po.mu);
-
-  return ret_fd;
-}
-
-static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                      grpc_closure *on_done, int *release_fd,
-                      bool already_closed, const char *reason) {
-  grpc_error *error = GRPC_ERROR_NONE;
-  polling_island *unref_pi = NULL;
-
-  gpr_mu_lock(&fd->po.mu);
-  fd->on_done_closure = on_done;
-
-  /* Remove the active status but keep referenced. We want this grpc_fd struct
-     to be alive (and not added to freelist) until the end of this function */
-  REF_BY(fd, 1, reason);
-
-  /* Remove the fd from the polling island:
-     - Get a lock on the latest polling island (i.e the last island in the
-       linked list pointed by fd->po.pi). This is the island that
-       would actually contain the fd
-     - Remove the fd from the latest polling island
-     - Unlock the latest polling island
-     - Set fd->po.pi to NULL (but remove the ref on the polling island
-       before doing this.) */
-  if (fd->po.pi != NULL) {
-    polling_island *pi_latest = polling_island_lock(fd->po.pi);
-    polling_island_remove_fd_locked(pi_latest, fd, already_closed, &error);
-    gpr_mu_unlock(&pi_latest->mu);
-
-    unref_pi = fd->po.pi;
-    fd->po.pi = NULL;
-  }
-
-  /* If release_fd is not NULL, we should be relinquishing control of the file
-     descriptor fd->fd (but we still own the grpc_fd structure). */
-  if (release_fd != NULL) {
-    *release_fd = fd->fd;
-  } else {
-    close(fd->fd);
-  }
-
-  fd->orphaned = true;
-
-  GRPC_CLOSURE_SCHED(exec_ctx, fd->on_done_closure, GRPC_ERROR_REF(error));
-
-  gpr_mu_unlock(&fd->po.mu);
-  UNREF_BY(fd, 2, reason); /* Drop the reference */
-  if (unref_pi != NULL) {
-    /* Unref stale polling island here, outside the fd lock above.
-       The polling island owns a workqueue which owns an fd, and unreffing
-       inside the lock can cause an eventual lock loop that makes TSAN very
-       unhappy. */
-    PI_UNREF(exec_ctx, unref_pi, "fd_orphan");
-  }
-  GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error));
-  GRPC_ERROR_UNREF(error);
-}
-
-static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx,
-                                                  grpc_fd *fd) {
-  gpr_atm notifier = gpr_atm_acq_load(&fd->read_notifier_pollset);
-  return (grpc_pollset *)notifier;
-}
-
-static bool fd_is_shutdown(grpc_fd *fd) {
-  return grpc_lfev_is_shutdown(&fd->read_closure);
-}
-
-/* Might be called multiple times */
-static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_error *why) {
-  if (grpc_lfev_set_shutdown(exec_ctx, &fd->read_closure,
-                             GRPC_ERROR_REF(why))) {
-    shutdown(fd->fd, SHUT_RDWR);
-    grpc_lfev_set_shutdown(exec_ctx, &fd->write_closure, GRPC_ERROR_REF(why));
-  }
-  GRPC_ERROR_UNREF(why);
-}
-
-static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                              grpc_closure *closure) {
-  grpc_lfev_notify_on(exec_ctx, &fd->read_closure, closure, "read");
-}
-
-static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                               grpc_closure *closure) {
-  grpc_lfev_notify_on(exec_ctx, &fd->write_closure, closure, "write");
-}
-
-/*******************************************************************************
- * Pollset Definitions
- */
-GPR_TLS_DECL(g_current_thread_pollset);
-GPR_TLS_DECL(g_current_thread_worker);
-static __thread bool g_initialized_sigmask;
-static __thread sigset_t g_orig_sigmask;
-static __thread sigset_t g_wakeup_sig_set;
-
-static void sig_handler(int sig_num) {
-#ifdef GRPC_EPOLL_DEBUG
-  gpr_log(GPR_INFO, "Received signal %d", sig_num);
-#endif
-}
-
-static void pollset_worker_init(grpc_pollset_worker *worker) {
-  worker->pt_id = pthread_self();
-  worker->next = worker->prev = NULL;
-  gpr_atm_no_barrier_store(&worker->is_kicked, (gpr_atm)0);
-  gpr_atm_no_barrier_store(&worker->is_polling_turn, (gpr_atm)0);
-  worker_node_init(&worker->pi_list_link);
-}
-
-static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); }
-
-/* Global state management */
-static grpc_error *pollset_global_init(void) {
-  gpr_tls_init(&g_current_thread_pollset);
-  gpr_tls_init(&g_current_thread_worker);
-  poller_kick_init();
-  return GRPC_ERROR_NONE;
-}
-
-static void pollset_global_shutdown(void) {
-  gpr_tls_destroy(&g_current_thread_pollset);
-  gpr_tls_destroy(&g_current_thread_worker);
-}
-
-static grpc_error *worker_kick(grpc_pollset_worker *worker,
-                               gpr_atm *is_kicked) {
-  grpc_error *err = GRPC_ERROR_NONE;
-
-  /* Kick the worker only if it was not already kicked */
-  if (gpr_atm_no_barrier_cas(is_kicked, (gpr_atm)0, (gpr_atm)1)) {
-    GRPC_POLLING_TRACE(
-        "pollset_worker_kick: Kicking worker: %p (thread id: %ld)",
-        (void *)worker, (long int)worker->pt_id);
-    int err_num = pthread_kill(worker->pt_id, grpc_wakeup_signal);
-    if (err_num != 0) {
-      err = GRPC_OS_ERROR(err_num, "pthread_kill");
-    }
-  }
-  return err;
-}
-
-static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) {
-  return worker_kick(worker, &worker->is_kicked);
-}
-
-static grpc_error *poller_kick(grpc_pollset_worker *worker) {
-  return worker_kick(worker, &worker->is_polling_turn);
-}
-
-/* Return 1 if the pollset has active threads in pollset_work (pollset must
- * be locked) */
-static int pollset_has_workers(grpc_pollset *p) {
-  return p->root_worker.next != &p->root_worker;
-}
-
-static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
-  worker->prev->next = worker->next;
-  worker->next->prev = worker->prev;
-}
-
-static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) {
-  if (pollset_has_workers(p)) {
-    grpc_pollset_worker *w = p->root_worker.next;
-    remove_worker(p, w);
-    return w;
-  } else {
-    return NULL;
-  }
-}
-
-static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
-  worker->next = &p->root_worker;
-  worker->prev = worker->next->prev;
-  worker->prev->next = worker->next->prev = worker;
-}
-
-static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
-  worker->prev = &p->root_worker;
-  worker->next = worker->prev->next;
-  worker->prev->next = worker->next->prev = worker;
-}
-
-/* p->mu must be held before calling this function */
-static grpc_error *pollset_kick(grpc_pollset *p,
-                                grpc_pollset_worker *specific_worker) {
-  GPR_TIMER_BEGIN("pollset_kick", 0);
-  grpc_error *error = GRPC_ERROR_NONE;
-  const char *err_desc = "Kick Failure";
-  grpc_pollset_worker *worker = specific_worker;
-  if (worker != NULL) {
-    if (worker == GRPC_POLLSET_KICK_BROADCAST) {
-      if (pollset_has_workers(p)) {
-        GPR_TIMER_BEGIN("pollset_kick.broadcast", 0);
-        for (worker = p->root_worker.next; worker != &p->root_worker;
-             worker = worker->next) {
-          if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) {
-            append_error(&error, pollset_worker_kick(worker), err_desc);
-          }
-        }
-        GPR_TIMER_END("pollset_kick.broadcast", 0);
-      } else {
-        p->kicked_without_pollers = true;
-      }
-    } else {
-      GPR_TIMER_MARK("kicked_specifically", 0);
-      if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) {
-        append_error(&error, pollset_worker_kick(worker), err_desc);
-      }
-    }
-  } else if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)p) {
-    /* Since worker == NULL, it means that we can kick "any" worker on this
-       pollset 'p'. If 'p' happens to be the same pollset this thread is
-       currently polling (i.e in pollset_work() function), then there is no need
-       to kick any other worker since the current thread can just absorb the
-       kick. This is the reason why we enter this case only when
-       g_current_thread_pollset is != p */
-
-    GPR_TIMER_MARK("kick_anonymous", 0);
-    worker = pop_front_worker(p);
-    if (worker != NULL) {
-      GPR_TIMER_MARK("finally_kick", 0);
-      push_back_worker(p, worker);
-      append_error(&error, pollset_worker_kick(worker), err_desc);
-    } else {
-      GPR_TIMER_MARK("kicked_no_pollers", 0);
-      p->kicked_without_pollers = true;
-    }
-  }
-
-  GPR_TIMER_END("pollset_kick", 0);
-  GRPC_LOG_IF_ERROR("pollset_kick", GRPC_ERROR_REF(error));
-  return error;
-}
-
-static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) {
-  gpr_mu_init(&pollset->po.mu);
-  *mu = &pollset->po.mu;
-  pollset->po.pi = NULL;
-#ifndef NDEBUG
-  pollset->po.obj_type = POLL_OBJ_POLLSET;
-#endif
-
-  pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker;
-  pollset->kicked_without_pollers = false;
-
-  pollset->shutting_down = false;
-  pollset->finish_shutdown_called = false;
-  pollset->shutdown_done = NULL;
-}
-
-/* Convert millis to timespec (clock-type is assumed to be GPR_TIMESPAN) */
-static struct timespec millis_to_timespec(int millis) {
-  struct timespec linux_ts;
-  gpr_timespec gpr_ts;
-
-  if (millis == -1) {
-    gpr_ts = gpr_inf_future(GPR_TIMESPAN);
-  } else {
-    gpr_ts = gpr_time_from_millis(millis, GPR_TIMESPAN);
-  }
-
-  linux_ts.tv_sec = (time_t)gpr_ts.tv_sec;
-  linux_ts.tv_nsec = gpr_ts.tv_nsec;
-  return linux_ts;
-}
-
-/* Convert a timespec to milliseconds:
-   - Very small or negative poll times are clamped to zero to do a non-blocking
-     poll (which becomes spin polling)
-   - Other small values are rounded up to one millisecond
-   - Longer than a millisecond polls are rounded up to the next nearest
-     millisecond to avoid spinning
-   - Infinite timeouts are converted to -1 */
-static int poll_deadline_to_millis_timeout(gpr_timespec deadline,
-                                           gpr_timespec now) {
-  gpr_timespec timeout;
-  static const int64_t max_spin_polling_us = 10;
-  if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) {
-    return -1;
-  }
-
-  if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros(
-                                                   max_spin_polling_us,
-                                                   GPR_TIMESPAN))) <= 0) {
-    return 0;
-  }
-  timeout = gpr_time_sub(deadline, now);
-  int millis = gpr_time_to_millis(gpr_time_add(
-      timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN)));
-  return millis >= 1 ? millis : 1;
-}
-
-static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                               grpc_pollset *notifier) {
-  grpc_lfev_set_ready(exec_ctx, &fd->read_closure, "read");
-
-  /* Note, it is possible that fd_become_readable might be called twice with
-     different 'notifier's when an fd becomes readable and it is in two epoll
-     sets (This can happen briefly during polling island merges). In such cases
-     it does not really matter which notifer is set as the read_notifier_pollset
-     (They would both point to the same polling island anyway) */
-  /* Use release store to match with acquire load in fd_get_read_notifier */
-  gpr_atm_rel_store(&fd->read_notifier_pollset, (gpr_atm)notifier);
-}
-
-static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  grpc_lfev_set_ready(exec_ctx, &fd->write_closure, "write");
-}
-
-static void pollset_release_polling_island(grpc_exec_ctx *exec_ctx,
-                                           grpc_pollset *ps, char *reason) {
-  if (ps->po.pi != NULL) {
-    PI_UNREF(exec_ctx, ps->po.pi, reason);
-  }
-  ps->po.pi = NULL;
-}
-
-static void finish_shutdown_locked(grpc_exec_ctx *exec_ctx,
-                                   grpc_pollset *pollset) {
-  /* The pollset cannot have any workers if we are at this stage */
-  GPR_ASSERT(!pollset_has_workers(pollset));
-
-  pollset->finish_shutdown_called = true;
-
-  /* Release the ref and set pollset->po.pi to NULL */
-  pollset_release_polling_island(exec_ctx, pollset, "ps_shutdown");
-  GRPC_CLOSURE_SCHED(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE);
-}
-
-/* pollset->po.mu lock must be held by the caller before calling this */
-static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                             grpc_closure *closure) {
-  GPR_TIMER_BEGIN("pollset_shutdown", 0);
-  GPR_ASSERT(!pollset->shutting_down);
-  pollset->shutting_down = true;
-  pollset->shutdown_done = closure;
-  pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST);
-
-  /* If the pollset has any workers, we cannot call finish_shutdown_locked()
-     because it would release the underlying polling island. In such a case, we
-     let the last worker call finish_shutdown_locked() from pollset_work() */
-  if (!pollset_has_workers(pollset)) {
-    GPR_ASSERT(!pollset->finish_shutdown_called);
-    GPR_TIMER_MARK("pollset_shutdown.finish_shutdown_locked", 0);
-    finish_shutdown_locked(exec_ctx, pollset);
-  }
-  GPR_TIMER_END("pollset_shutdown", 0);
-}
-
-/* pollset_shutdown is guaranteed to be called before pollset_destroy. So other
- * than destroying the mutexes, there is nothing special that needs to be done
- * here */
-static void pollset_destroy(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) {
-  GPR_ASSERT(!pollset_has_workers(pollset));
-  gpr_mu_destroy(&pollset->po.mu);
-}
-
-/* NOTE: This function may modify 'now' */
-static bool acquire_polling_lease(grpc_exec_ctx *exec_ctx,
-                                  grpc_pollset_worker *worker,
-                                  polling_island *pi, gpr_timespec deadline,
-                                  gpr_timespec *now) {
-  bool is_lease_acquired = false;
-
-  gpr_mu_lock(&pi->worker_list_mu);  //  LOCK
-  long num_pollers = gpr_atm_no_barrier_load(&pi->poller_count);
-
-  if (num_pollers >= g_max_pollers_per_pi) {
-    push_back_worker_node(&pi->worker_list_head, &worker->pi_list_link);
-    gpr_mu_unlock(&pi->worker_list_mu);  // UNLOCK
-
-    bool is_timeout = false;
-    int ret;
-    int timeout_ms = poll_deadline_to_millis_timeout(deadline, *now);
-    if (timeout_ms == -1) {
-      ret = sigwaitinfo(&g_wakeup_sig_set, NULL);
-    } else {
-      struct timespec sigwait_timeout = millis_to_timespec(timeout_ms);
-      GRPC_SCHEDULING_START_BLOCKING_REGION;
-      GRPC_STATS_INC_SYSCALL_WAIT(exec_ctx);
-      ret = sigtimedwait(&g_wakeup_sig_set, NULL, &sigwait_timeout);
-      GRPC_SCHEDULING_END_BLOCKING_REGION;
-    }
-
-    if (ret == -1) {
-      if (errno == EAGAIN) {
-        is_timeout = true;
-      } else {
-        /* NOTE: This should not happen. If we see these log messages, it means
-           we are most likely doing something incorrect in the setup * needed
-           for sigwaitinfo/sigtimedwait */
-        gpr_log(GPR_ERROR,
-                "sigtimedwait failed with retcode: %d (timeout_ms: %d)", errno,
-                timeout_ms);
-      }
-    }
-
-    /* Did the worker come out of sigtimedwait due to a thread that just
-       exited epoll and kicking it (in release_polling_lease function). */
-    bool is_polling_turn = gpr_atm_acq_load(&worker->is_polling_turn);
-
-    /* Did the worker come out of sigtimedwait due to a thread alerting it that
-       some completion event was (likely) available in the completion queue */
-    bool is_kicked = gpr_atm_no_barrier_load(&worker->is_kicked);
-
-    if (is_kicked || is_timeout) {
-      *now = deadline; /* Essentially make the epoll timeout = 0 */
-    } else if (is_polling_turn) {
-      *now = gpr_now(GPR_CLOCK_MONOTONIC); /* Reduce the epoll timeout */
-    }
-
-    gpr_mu_lock(&pi->worker_list_mu);  // LOCK
-    /* The node might have already been removed from the list by the poller
-       that kicked this. However it is safe to call 'remove_worker_node' on
-       an already detached node */
-    remove_worker_node(&worker->pi_list_link);
-    /* It is important to read the num_pollers again under the lock so that we
-     * have the latest num_pollers value that doesn't change while we are doing
-     * the "(num_pollers < g_max_pollers_per_pi)" a a few lines below */
-    num_pollers = gpr_atm_no_barrier_load(&pi->poller_count);
-  }
-
-  if (num_pollers < g_max_pollers_per_pi) {
-    gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1);
-    is_lease_acquired = true;
-  }
-
-  gpr_mu_unlock(&pi->worker_list_mu);  // UNLOCK
-  return is_lease_acquired;
-}
-
-static void release_polling_lease(polling_island *pi, grpc_error **error) {
-  gpr_mu_lock(&pi->worker_list_mu);
-
-  gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1);
-  worker_node *node = pop_front_worker_node(&pi->worker_list_head);
-  if (node != NULL) {
-    grpc_pollset_worker *next_worker = WORKER_FROM_WORKER_LIST_NODE(node);
-    append_error(error, poller_kick(next_worker), "poller kick error");
-  }
-
-  gpr_mu_unlock(&pi->worker_list_mu);
-}
-
-#define GRPC_EPOLL_MAX_EVENTS 100
-static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd,
-                                   grpc_pollset *pollset, polling_island *pi,
-                                   grpc_pollset_worker *worker,
-                                   gpr_timespec now, gpr_timespec deadline,
-                                   sigset_t *sig_mask, grpc_error **error) {
-  /* Only g_max_pollers_per_pi threads can be doing polling in parallel.
-     If we cannot get a lease, we cannot continue to do epoll_pwait() */
-  if (!acquire_polling_lease(exec_ctx, worker, pi, deadline, &now)) {
-    return;
-  }
-
-  struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS];
-  int ep_rv;
-  char *err_msg;
-  const char *err_desc = "pollset_work_and_unlock";
-
-  /* timeout_ms is the time between 'now' and 'deadline' */
-  int timeout_ms = poll_deadline_to_millis_timeout(deadline, now);
-
-  GRPC_SCHEDULING_START_BLOCKING_REGION;
-  GRPC_STATS_INC_SYSCALL_POLL(exec_ctx);
-  ep_rv =
-      epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask);
-  GRPC_SCHEDULING_END_BLOCKING_REGION;
-
-  /* Give back the lease right away so that some other thread can enter */
-  release_polling_lease(pi, error);
-
-  if (ep_rv < 0) {
-    if (errno != EINTR) {
-      gpr_asprintf(&err_msg,
-                   "epoll_wait() epoll fd: %d failed with error: %d (%s)",
-                   epoll_fd, errno, strerror(errno));
-      append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-    } else {
-      /* We were interrupted. Save an interation by doing a zero timeout
-         epoll_wait to see if there are any other events of interest */
-      GRPC_POLLING_TRACE("pollset_work: pollset: %p, worker: %p received kick",
-                         (void *)pollset, (void *)worker);
-      ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0);
-    }
-  }
-
-#ifdef GRPC_TSAN
-  /* See the definition of g_poll_sync for more details */
-  gpr_atm_acq_load(&g_epoll_sync);
-#endif /* defined(GRPC_TSAN) */
-
-  for (int i = 0; i < ep_rv; ++i) {
-    void *data_ptr = ep_ev[i].data.ptr;
-    if (data_ptr == &polling_island_wakeup_fd) {
-      GRPC_POLLING_TRACE(
-          "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: "
-          "%d) got merged",
-          (void *)pollset, (void *)worker, epoll_fd);
-      /* This means that our polling island is merged with a different
-         island. We do not have to do anything here since the subsequent call
-         to the function pollset_work_and_unlock() will pick up the correct
-         epoll_fd */
-    } else {
-      grpc_fd *fd = data_ptr;
-      int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP);
-      int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI);
-      int write_ev = ep_ev[i].events & EPOLLOUT;
-      if (read_ev || cancel) {
-        fd_become_readable(exec_ctx, fd, pollset);
-      }
-      if (write_ev || cancel) {
-        fd_become_writable(exec_ctx, fd);
-      }
-    }
-  }
-}
-
-/* Note: sig_mask contains the signal mask to use *during* epoll_wait() */
-static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx,
-                                    grpc_pollset *pollset,
-                                    grpc_pollset_worker *worker,
-                                    gpr_timespec now, gpr_timespec deadline,
-                                    sigset_t *sig_mask, grpc_error **error) {
-  int epoll_fd = -1;
-  polling_island *pi = NULL;
-  GPR_TIMER_BEGIN("pollset_work_and_unlock", 0);
-
-  /* We need to get the epoll_fd to wait on. The epoll_fd is in inside the
-     latest polling island pointed by pollset->po.pi
-
-     Since epoll_fd is immutable, it is safe to read it without a lock on the
-     polling island. There is however a possibility that the polling island from
-     which we got the epoll_fd, got merged with another island in the meantime.
-     This is okay because in such a case, we will wakeup right-away from
-     epoll_pwait() (because any merge will poison the old polling island's epoll
-     set 'polling_island_wakeup_fd') and then pick up the latest polling_island
-     the next time this function - pollset_work_and_unlock()) is called */
-
-  if (pollset->po.pi == NULL) {
-    pollset->po.pi = polling_island_create(exec_ctx, NULL, error);
-    if (pollset->po.pi == NULL) {
-      GPR_TIMER_END("pollset_work_and_unlock", 0);
-      return; /* Fatal error. Cannot continue */
-    }
-
-    PI_ADD_REF(pollset->po.pi, "ps");
-    GRPC_POLLING_TRACE("pollset_work: pollset: %p created new pi: %p",
-                       (void *)pollset, (void *)pollset->po.pi);
-  }
-
-  pi = polling_island_maybe_get_latest(pollset->po.pi);
-  epoll_fd = pi->epoll_fd;
-
-  /* Update the pollset->po.pi since the island being pointed by
-     pollset->po.pi maybe older than the one pointed by pi) */
-  if (pollset->po.pi != pi) {
-    /* Always do PI_ADD_REF before PI_UNREF because PI_UNREF may cause the
-       polling island to be deleted */
-    PI_ADD_REF(pi, "ps");
-    PI_UNREF(exec_ctx, pollset->po.pi, "ps");
-    pollset->po.pi = pi;
-  }
-
-  /* Add an extra ref so that the island does not get destroyed (which means
-     the epoll_fd won't be closed) while we are are doing an epoll_wait() on the
-     epoll_fd */
-  PI_ADD_REF(pi, "ps_work");
-  gpr_mu_unlock(&pollset->po.mu);
-
-  g_current_thread_polling_island = pi;
-  pollset_do_epoll_pwait(exec_ctx, epoll_fd, pollset, pi, worker, now, deadline,
-                         sig_mask, error);
-  g_current_thread_polling_island = NULL;
-
-  GPR_ASSERT(pi != NULL);
-
-  /* Before leaving, release the extra ref we added to the polling island. It
-     is important to use "pi" here (i.e our old copy of pollset->po.pi
-     that we got before releasing the polling island lock). This is because
-     pollset->po.pi pointer might get udpated in other parts of the
-     code when there is an island merge while we are doing epoll_wait() above */
-  PI_UNREF(exec_ctx, pi, "ps_work");
-
-  GPR_TIMER_END("pollset_work_and_unlock", 0);
-}
-
-/* pollset->po.mu lock must be held by the caller before calling this.
-   The function pollset_work() may temporarily release the lock (pollset->po.mu)
-   during the course of its execution but it will always re-acquire the lock and
-   ensure that it is held by the time the function returns */
-static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                                grpc_pollset_worker **worker_hdl,
-                                gpr_timespec now, gpr_timespec deadline) {
-  GPR_TIMER_BEGIN("pollset_work", 0);
-  grpc_error *error = GRPC_ERROR_NONE;
-
-  grpc_pollset_worker worker;
-  pollset_worker_init(&worker);
-
-  if (worker_hdl) *worker_hdl = &worker;
-
-  gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset);
-  gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker);
-
-  if (pollset->kicked_without_pollers) {
-    /* If the pollset was kicked without pollers, pretend that the current
-       worker got the kick and skip polling. A kick indicates that there is some
-       work that needs attention like an event on the completion queue or an
-       alarm */
-    GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0);
-    pollset->kicked_without_pollers = 0;
-  } else if (!pollset->shutting_down) {
-    /* We use the posix-signal with number 'grpc_wakeup_signal' for waking up
-       (i.e 'kicking') a worker in the pollset. A 'kick' is a way to inform the
-       worker that there is some pending work that needs immediate attention
-       (like an event on the completion queue, or a polling island merge that
-       results in a new epoll-fd to wait on) and that the worker should not
-       spend time waiting in epoll_pwait().
-
-       A worker can be kicked anytime from the point it is added to the pollset
-       via push_front_worker() (or push_back_worker()) to the point it is
-       removed via remove_worker().
-       If the worker is kicked before/during it calls epoll_pwait(), it should
-       immediately exit from epoll_wait(). If the worker is kicked after it
-       returns from epoll_wait(), then nothing really needs to be done.
-
-       To accomplish this, we mask 'grpc_wakeup_signal' on this thread at all
-       times *except* when it is in epoll_pwait(). This way, the worker never
-       misses acting on a kick */
-
-    if (!g_initialized_sigmask) {
-      sigemptyset(&g_wakeup_sig_set);
-      sigaddset(&g_wakeup_sig_set, grpc_wakeup_signal);
-      pthread_sigmask(SIG_BLOCK, &g_wakeup_sig_set, &g_orig_sigmask);
-      sigdelset(&g_orig_sigmask, grpc_wakeup_signal);
-      g_initialized_sigmask = true;
-      /* new_mask:       The new thread mask which blocks 'grpc_wakeup_signal'.
-                         This is the mask used at all times *except during
-                         epoll_wait()*"
-         g_orig_sigmask: The thread mask which allows 'grpc_wakeup_signal' and
-                         this is the mask to use *during epoll_wait()*
-
-         The new_mask is set on the worker before it is added to the pollset
-         (i.e before it can be kicked) */
-    }
-
-    push_front_worker(pollset, &worker); /* Add worker to pollset */
-
-    pollset_work_and_unlock(exec_ctx, pollset, &worker, now, deadline,
-                            &g_orig_sigmask, &error);
-    grpc_exec_ctx_flush(exec_ctx);
-
-    gpr_mu_lock(&pollset->po.mu);
-
-    /* Note: There is no need to reset worker.is_kicked to 0 since we are no
-       longer going to use this worker */
-    remove_worker(pollset, &worker);
-  }
-
-  /* If we are the last worker on the pollset (i.e pollset_has_workers() is
-     false at this point) and the pollset is shutting down, we may have to
-     finish the shutdown process by calling finish_shutdown_locked().
-     See pollset_shutdown() for more details.
-
-     Note: Continuing to access pollset here is safe; it is the caller's
-     responsibility to not destroy a pollset when it has outstanding calls to
-     pollset_work() */
-  if (pollset->shutting_down && !pollset_has_workers(pollset) &&
-      !pollset->finish_shutdown_called) {
-    GPR_TIMER_MARK("pollset_work.finish_shutdown_locked", 0);
-    finish_shutdown_locked(exec_ctx, pollset);
-
-    gpr_mu_unlock(&pollset->po.mu);
-    grpc_exec_ctx_flush(exec_ctx);
-    gpr_mu_lock(&pollset->po.mu);
-  }
-
-  if (worker_hdl) *worker_hdl = NULL;
-
-  gpr_tls_set(&g_current_thread_pollset, (intptr_t)0);
-  gpr_tls_set(&g_current_thread_worker, (intptr_t)0);
-
-  GPR_TIMER_END("pollset_work", 0);
-
-  GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error));
-  return error;
-}
-
-static void add_poll_object(grpc_exec_ctx *exec_ctx, poll_obj *bag,
-                            poll_obj_type bag_type, poll_obj *item,
-                            poll_obj_type item_type) {
-  GPR_TIMER_BEGIN("add_poll_object", 0);
-
-#ifndef NDEBUG
-  GPR_ASSERT(item->obj_type == item_type);
-  GPR_ASSERT(bag->obj_type == bag_type);
-#endif
-
-  grpc_error *error = GRPC_ERROR_NONE;
-  polling_island *pi_new = NULL;
-
-  gpr_mu_lock(&bag->mu);
-  gpr_mu_lock(&item->mu);
-
-retry:
-  /*
-   * 1) If item->pi and bag->pi are both non-NULL and equal, do nothing
-   * 2) If item->pi and bag->pi are both NULL, create a new polling island (with
-   *    a refcount of 2) and point item->pi and bag->pi to the new island
-   * 3) If exactly one of item->pi or bag->pi is NULL, update it to point to
-   *    the other's non-NULL pi
-   * 4) Finally if item->pi and bag-pi are non-NULL and not-equal, merge the
-   *    polling islands and update item->pi and bag->pi to point to the new
-   *    island
-   */
-
-  /* Early out if we are trying to add an 'fd' to a 'bag' but the fd is already
-   * orphaned */
-  if (item_type == POLL_OBJ_FD && (FD_FROM_PO(item))->orphaned) {
-    gpr_mu_unlock(&item->mu);
-    gpr_mu_unlock(&bag->mu);
-    return;
-  }
-
-  if (item->pi == bag->pi) {
-    pi_new = item->pi;
-    if (pi_new == NULL) {
-      /* GPR_ASSERT(item->pi == bag->pi == NULL) */
-
-      /* If we are adding an fd to a bag (i.e pollset or pollset_set), then
-       * we need to do some extra work to make TSAN happy */
-      if (item_type == POLL_OBJ_FD) {
-        /* Unlock before creating a new polling island: the polling island will
-           create a workqueue which creates a file descriptor, and holding an fd
-           lock here can eventually cause a loop to appear to TSAN (making it
-           unhappy). We don't think it's a real loop (there's an epoch point
-           where that loop possibility disappears), but the advantages of
-           keeping TSAN happy outweigh any performance advantage we might have
-           by keeping the lock held. */
-        gpr_mu_unlock(&item->mu);
-        pi_new = polling_island_create(exec_ctx, FD_FROM_PO(item), &error);
-        gpr_mu_lock(&item->mu);
-
-        /* Need to reverify any assumptions made between the initial lock and
-           getting to this branch: if they've changed, we need to throw away our
-           work and figure things out again. */
-        if (item->pi != NULL) {
-          GRPC_POLLING_TRACE(
-              "add_poll_object: Raced creating new polling island. pi_new: %p "
-              "(fd: %d, %s: %p)",
-              (void *)pi_new, FD_FROM_PO(item)->fd, poll_obj_string(bag_type),
-              (void *)bag);
-          /* No need to lock 'pi_new' here since this is a new polling island
-             and no one has a reference to it yet */
-          polling_island_remove_all_fds_locked(pi_new, true, &error);
-
-          /* Ref and unref so that the polling island gets deleted during unref
-           */
-          PI_ADD_REF(pi_new, "dance_of_destruction");
-          PI_UNREF(exec_ctx, pi_new, "dance_of_destruction");
-          goto retry;
-        }
-      } else {
-        pi_new = polling_island_create(exec_ctx, NULL, &error);
-      }
-
-      GRPC_POLLING_TRACE(
-          "add_poll_object: Created new polling island. pi_new: %p (%s: %p, "
-          "%s: %p)",
-          (void *)pi_new, poll_obj_string(item_type), (void *)item,
-          poll_obj_string(bag_type), (void *)bag);
-    } else {
-      GRPC_POLLING_TRACE(
-          "add_poll_object: Same polling island. pi: %p (%s, %s)",
-          (void *)pi_new, poll_obj_string(item_type),
-          poll_obj_string(bag_type));
-    }
-  } else if (item->pi == NULL) {
-    /* GPR_ASSERT(bag->pi != NULL) */
-    /* Make pi_new point to latest pi*/
-    pi_new = polling_island_lock(bag->pi);
-
-    if (item_type == POLL_OBJ_FD) {
-      grpc_fd *fd = FD_FROM_PO(item);
-      polling_island_add_fds_locked(pi_new, &fd, 1, true, &error);
-    }
-
-    gpr_mu_unlock(&pi_new->mu);
-    GRPC_POLLING_TRACE(
-        "add_poll_obj: item->pi was NULL. pi_new: %p (item(%s): %p, "
-        "bag(%s): %p)",
-        (void *)pi_new, poll_obj_string(item_type), (void *)item,
-        poll_obj_string(bag_type), (void *)bag);
-  } else if (bag->pi == NULL) {
-    /* GPR_ASSERT(item->pi != NULL) */
-    /* Make pi_new to point to latest pi */
-    pi_new = polling_island_lock(item->pi);
-    gpr_mu_unlock(&pi_new->mu);
-    GRPC_POLLING_TRACE(
-        "add_poll_obj: bag->pi was NULL. pi_new: %p (item(%s): %p, "
-        "bag(%s): %p)",
-        (void *)pi_new, poll_obj_string(item_type), (void *)item,
-        poll_obj_string(bag_type), (void *)bag);
-  } else {
-    pi_new = polling_island_merge(item->pi, bag->pi, &error);
-    GRPC_POLLING_TRACE(
-        "add_poll_obj: polling islands merged. pi_new: %p (item(%s): %p, "
-        "bag(%s): %p)",
-        (void *)pi_new, poll_obj_string(item_type), (void *)item,
-        poll_obj_string(bag_type), (void *)bag);
-  }
-
-  /* At this point, pi_new is the polling island that both item->pi and bag->pi
-     MUST be pointing to */
-
-  if (item->pi != pi_new) {
-    PI_ADD_REF(pi_new, poll_obj_string(item_type));
-    if (item->pi != NULL) {
-      PI_UNREF(exec_ctx, item->pi, poll_obj_string(item_type));
-    }
-    item->pi = pi_new;
-  }
-
-  if (bag->pi != pi_new) {
-    PI_ADD_REF(pi_new, poll_obj_string(bag_type));
-    if (bag->pi != NULL) {
-      PI_UNREF(exec_ctx, bag->pi, poll_obj_string(bag_type));
-    }
-    bag->pi = pi_new;
-  }
-
-  gpr_mu_unlock(&item->mu);
-  gpr_mu_unlock(&bag->mu);
-
-  GRPC_LOG_IF_ERROR("add_poll_object", error);
-  GPR_TIMER_END("add_poll_object", 0);
-}
-
-static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                           grpc_fd *fd) {
-  add_poll_object(exec_ctx, &pollset->po, POLL_OBJ_POLLSET, &fd->po,
-                  POLL_OBJ_FD);
-}
-
-/*******************************************************************************
- * Pollset-set Definitions
- */
-
-static grpc_pollset_set *pollset_set_create(void) {
-  grpc_pollset_set *pss = gpr_malloc(sizeof(*pss));
-  gpr_mu_init(&pss->po.mu);
-  pss->po.pi = NULL;
-#ifndef NDEBUG
-  pss->po.obj_type = POLL_OBJ_POLLSET_SET;
-#endif
-  return pss;
-}
-
-static void pollset_set_destroy(grpc_exec_ctx *exec_ctx,
-                                grpc_pollset_set *pss) {
-  gpr_mu_destroy(&pss->po.mu);
-
-  if (pss->po.pi != NULL) {
-    PI_UNREF(exec_ctx, pss->po.pi, "pss_destroy");
-  }
-
-  gpr_free(pss);
-}
-
-static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
-                               grpc_fd *fd) {
-  add_poll_object(exec_ctx, &pss->po, POLL_OBJ_POLLSET_SET, &fd->po,
-                  POLL_OBJ_FD);
-}
-
-static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
-                               grpc_fd *fd) {
-  /* Nothing to do */
-}
-
-static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx,
-                                    grpc_pollset_set *pss, grpc_pollset *ps) {
-  add_poll_object(exec_ctx, &pss->po, POLL_OBJ_POLLSET_SET, &ps->po,
-                  POLL_OBJ_POLLSET);
-}
-
-static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx,
-                                    grpc_pollset_set *pss, grpc_pollset *ps) {
-  /* Nothing to do */
-}
-
-static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx,
-                                        grpc_pollset_set *bag,
-                                        grpc_pollset_set *item) {
-  add_poll_object(exec_ctx, &bag->po, POLL_OBJ_POLLSET_SET, &item->po,
-                  POLL_OBJ_POLLSET_SET);
-}
-
-static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx,
-                                        grpc_pollset_set *bag,
-                                        grpc_pollset_set *item) {
-  /* Nothing to do */
-}
-
-/*******************************************************************************
- * Event engine binding
- */
-
-static void shutdown_engine(void) {
-  fd_global_shutdown();
-  pollset_global_shutdown();
-  polling_island_global_shutdown();
-}
-
-static const grpc_event_engine_vtable vtable = {
-    .pollset_size = sizeof(grpc_pollset),
-
-    .fd_create = fd_create,
-    .fd_wrapped_fd = fd_wrapped_fd,
-    .fd_orphan = fd_orphan,
-    .fd_shutdown = fd_shutdown,
-    .fd_is_shutdown = fd_is_shutdown,
-    .fd_notify_on_read = fd_notify_on_read,
-    .fd_notify_on_write = fd_notify_on_write,
-    .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset,
-
-    .pollset_init = pollset_init,
-    .pollset_shutdown = pollset_shutdown,
-    .pollset_destroy = pollset_destroy,
-    .pollset_work = pollset_work,
-    .pollset_kick = pollset_kick,
-    .pollset_add_fd = pollset_add_fd,
-
-    .pollset_set_create = pollset_set_create,
-    .pollset_set_destroy = pollset_set_destroy,
-    .pollset_set_add_pollset = pollset_set_add_pollset,
-    .pollset_set_del_pollset = pollset_set_del_pollset,
-    .pollset_set_add_pollset_set = pollset_set_add_pollset_set,
-    .pollset_set_del_pollset_set = pollset_set_del_pollset_set,
-    .pollset_set_add_fd = pollset_set_add_fd,
-    .pollset_set_del_fd = pollset_set_del_fd,
-
-    .shutdown_engine = shutdown_engine,
-};
-
-/* It is possible that GLIBC has epoll but the underlying kernel doesn't.
- * Create a dummy epoll_fd to make sure epoll support is available */
-static bool is_epoll_available() {
-  int fd = epoll_create1(EPOLL_CLOEXEC);
-  if (fd < 0) {
-    gpr_log(
-        GPR_ERROR,
-        "epoll_create1 failed with error: %d. Not using epoll polling engine",
-        fd);
-    return false;
-  }
-  close(fd);
-  return true;
-}
-
-/* This is mainly for testing purposes. Checks to see if environment variable
- * GRPC_MAX_POLLERS_PER_PI is set and if so, assigns that value to
- * g_max_pollers_per_pi (any negative value is considered INT_MAX) */
-static void set_max_pollers_per_island() {
-  char *s = gpr_getenv("GRPC_MAX_POLLERS_PER_PI");
-  if (s) {
-    g_max_pollers_per_pi = (int)strtol(s, NULL, 10);
-    if (g_max_pollers_per_pi < 0) {
-      g_max_pollers_per_pi = INT_MAX;
-    }
-  } else {
-    g_max_pollers_per_pi = INT_MAX;
-  }
-
-  gpr_log(GPR_INFO, "Max number of pollers per polling island: %d",
-          g_max_pollers_per_pi);
-}
-
-const grpc_event_engine_vtable *grpc_init_epoll_limited_pollers_linux(
-    bool explicitly_requested) {
-  if (!explicitly_requested) {
-    return NULL;
-  }
-
-  /* If use of signals is disabled, we cannot use epoll engine*/
-  if (is_grpc_wakeup_signal_initialized && grpc_wakeup_signal < 0) {
-    return NULL;
-  }
-
-  if (!grpc_has_wakeup_fd()) {
-    return NULL;
-  }
-
-  if (!is_epoll_available()) {
-    return NULL;
-  }
-
-  if (!is_grpc_wakeup_signal_initialized) {
-    grpc_use_signal(SIGRTMIN + 6);
-  }
-
-  set_max_pollers_per_island();
-
-  fd_global_init();
-
-  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
-    return NULL;
-  }
-
-  if (!GRPC_LOG_IF_ERROR("polling_island_global_init",
-                         polling_island_global_init())) {
-    return NULL;
-  }
-
-  return &vtable;
-}
-
-#else /* defined(GRPC_LINUX_EPOLL) */
-#if defined(GRPC_POSIX_SOCKET)
-#include "src/core/lib/iomgr/ev_posix.h"
-/* If GRPC_LINUX_EPOLL is not defined, it means epoll is not available. Return
- * NULL */
-const grpc_event_engine_vtable *grpc_init_epoll_limited_pollers_linux(
-    bool explicitly_requested) {
-  return NULL;
-}
-#endif /* defined(GRPC_POSIX_SOCKET) */
-#endif /* !defined(GRPC_LINUX_EPOLL) */
diff --git a/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h b/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h
deleted file mode 100644
index 1d6af5f52c..0000000000
--- a/src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h
+++ /dev/null
@@ -1,28 +0,0 @@
-/*
- *
- * 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.
- *
- */
-
-#ifndef GRPC_CORE_LIB_IOMGR_EV_EPOLL_LIMITED_POLLERS_LINUX_H
-#define GRPC_CORE_LIB_IOMGR_EV_EPOLL_LIMITED_POLLERS_LINUX_H
-
-#include "src/core/lib/iomgr/ev_posix.h"
-#include "src/core/lib/iomgr/port.h"
-
-const grpc_event_engine_vtable *grpc_init_epoll_limited_pollers_linux(
-    bool explicitly_requested);
-
-#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLL_LIMITED_POLLERS_LINUX_H */
diff --git a/src/core/lib/iomgr/ev_epoll_thread_pool_linux.c b/src/core/lib/iomgr/ev_epoll_thread_pool_linux.c
deleted file mode 100644
index ddb8c74d2d..0000000000
--- a/src/core/lib/iomgr/ev_epoll_thread_pool_linux.c
+++ /dev/null
@@ -1,1184 +0,0 @@
-/*
- *
- * Copyright 2017 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 "src/core/lib/iomgr/port.h"
-
-/* This polling engine is only relevant on linux kernels supporting epoll() */
-#ifdef GRPC_LINUX_EPOLL
-
-#include "src/core/lib/iomgr/ev_epoll_thread_pool_linux.h"
-
-#include <assert.h>
-#include <errno.h>
-#include <limits.h>
-#include <poll.h>
-#include <pthread.h>
-#include <string.h>
-#include <sys/epoll.h>
-#include <sys/socket.h>
-#include <unistd.h>
-
-#include <grpc/support/alloc.h>
-#include <grpc/support/cpu.h>
-#include <grpc/support/log.h>
-#include <grpc/support/string_util.h>
-#include <grpc/support/thd.h>
-#include <grpc/support/tls.h>
-#include <grpc/support/useful.h>
-
-#include "src/core/lib/debug/stats.h"
-#include "src/core/lib/iomgr/ev_posix.h"
-#include "src/core/lib/iomgr/iomgr_internal.h"
-#include "src/core/lib/iomgr/lockfree_event.h"
-#include "src/core/lib/iomgr/timer.h"
-#include "src/core/lib/iomgr/wakeup_fd_posix.h"
-#include "src/core/lib/profiling/timers.h"
-#include "src/core/lib/support/block_annotate.h"
-
-/* TODO: sreek - Move this to init.c and initialize this like other tracers. */
-#define GRPC_POLLING_TRACE(fmt, ...)        \
-  if (GRPC_TRACER_ON(grpc_polling_trace)) { \
-    gpr_log(GPR_INFO, (fmt), __VA_ARGS__);  \
-  }
-
-/* The alarm system needs to be able to wakeup 'some poller' sometimes
- * (specifically when a new alarm needs to be triggered earlier than the next
- * alarm 'epoch'). This wakeup_fd gives us something to alert on when such a
- * case occurs. */
-
-struct epoll_set;
-
-#define GRPC_POLLSET_KICK_BROADCAST ((grpc_pollset_worker *)1)
-
-/*******************************************************************************
- * Fd Declarations
- */
-struct grpc_fd {
-  gpr_mu mu;
-  struct epoll_set *eps;
-
-  int fd;
-
-  /* The fd is either closed or we relinquished control of it. In either cases,
-     this indicates that the 'fd' on this structure is no longer valid */
-  bool orphaned;
-
-  gpr_atm read_closure;
-  gpr_atm write_closure;
-
-  struct grpc_fd *freelist_next;
-  grpc_closure *on_done_closure;
-
-  grpc_iomgr_object iomgr_object;
-};
-
-static void fd_global_init(void);
-static void fd_global_shutdown(void);
-
-/*******************************************************************************
- * epoll set Declarations
- */
-
-#ifndef NDEBUG
-
-#define EPS_ADD_REF(p, r) eps_add_ref_dbg((p), (r), __FILE__, __LINE__)
-#define EPS_UNREF(exec_ctx, p, r) \
-  eps_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__)
-
-#else
-
-#define EPS_ADD_REF(p, r) eps_add_ref((p))
-#define EPS_UNREF(exec_ctx, p, r) eps_unref((exec_ctx), (p))
-
-#endif
-
-typedef struct epoll_set {
-  /* Mutex poller should acquire to poll this. This enforces that only one
-   * poller can be polling on epoll_set at any time */
-  gpr_mu mu;
-
-  /* Ref count. Use EPS_ADD_REF() and EPS_UNREF() macros to increment/decrement
-     the refcount. Once the ref count becomes zero, this structure is destroyed
-     which means we should ensure that there is never a scenario where a
-     EPS_ADD_REF() is racing with a EPS_UNREF() that just made the ref_count
-     zero. */
-  gpr_atm ref_count;
-
-  /* Number of threads currently polling on this epoll set*/
-  gpr_atm poller_count;
-
-  /* Is the epoll set shutdown */
-  gpr_atm is_shutdown;
-
-  /* The fd of the underlying epoll set */
-  int epoll_fd;
-} epoll_set;
-
-/*******************************************************************************
- * Pollset Declarations
- */
-struct grpc_pollset_worker {
-  gpr_cv kick_cv;
-
-  struct grpc_pollset_worker *next;
-  struct grpc_pollset_worker *prev;
-};
-
-struct grpc_pollset {
-  gpr_mu mu;
-  struct epoll_set *eps;
-
-  grpc_pollset_worker root_worker;
-  bool kicked_without_pollers;
-
-  bool shutting_down;          /* Is the pollset shutting down ? */
-  bool finish_shutdown_called; /* Is the 'finish_shutdown_locked()' called ? */
-  grpc_closure *shutdown_done; /* Called after after shutdown is complete */
-};
-
-/*******************************************************************************
- * Pollset-set Declarations
- */
-struct grpc_pollset_set {
-  char unused;
-};
-
-/*****************************************************************************
- * Dedicated polling threads and pollsets - Declarations
- */
-
-size_t g_num_eps = 1;
-struct epoll_set **g_epoll_sets = NULL;
-gpr_atm g_next_eps;
-size_t g_num_threads_per_eps = 1;
-gpr_thd_id *g_poller_threads = NULL;
-
-/* Used as read-notifier pollsets for fds. We won't be using read notifier
- * pollsets with this polling engine. So it does not matter what pollset we
- * return */
-grpc_pollset g_read_notifier;
-
-static void add_fd_to_eps(grpc_fd *fd);
-static bool init_epoll_sets();
-static void shutdown_epoll_sets();
-static void poller_thread_loop(void *arg);
-static void start_poller_threads();
-static void shutdown_poller_threads();
-
-/*******************************************************************************
- * Common helpers
- */
-
-static bool append_error(grpc_error **composite, grpc_error *error,
-                         const char *desc) {
-  if (error == GRPC_ERROR_NONE) return true;
-  if (*composite == GRPC_ERROR_NONE) {
-    *composite = GRPC_ERROR_CREATE_FROM_COPIED_STRING(desc);
-  }
-  *composite = grpc_error_add_child(*composite, error);
-  return false;
-}
-
-/*******************************************************************************
- * epoll set Definitions
- */
-
-/* The wakeup fd that is used to wake up all threads in an epoll_set informing
-   that the epoll set is shutdown.  This wakeup fd initialized to be readable
-   and MUST NOT be consumed i.e the threads that woke up MUST NOT call
-   grpc_wakeup_fd_consume_wakeup() */
-static grpc_wakeup_fd epoll_set_wakeup_fd;
-
-/* The epoll set being polled right now.
-   See comments in workqueue_maybe_wakeup for why this is tracked. */
-static __thread epoll_set *g_current_thread_epoll_set;
-
-/* Forward declaration */
-static void epoll_set_delete(epoll_set *eps);
-
-#ifdef GRPC_TSAN
-/* Currently TSAN may incorrectly flag data races between epoll_ctl and
-   epoll_wait for any grpc_fd structs that are added to the epoll set via
-   epoll_ctl and are returned (within a very short window) via epoll_wait().
-
-   To work-around this race, we establish a happens-before relation between
-   the code just-before epoll_ctl() and the code after epoll_wait() by using
-   this atomic */
-gpr_atm g_epoll_sync;
-#endif /* defined(GRPC_TSAN) */
-
-static void eps_add_ref(epoll_set *eps);
-static void eps_unref(grpc_exec_ctx *exec_ctx, epoll_set *eps);
-
-#ifndef NDEBUG
-static void eps_add_ref_dbg(epoll_set *eps, const char *reason,
-                            const char *file, int line) {
-  if (GRPC_TRACER_ON(grpc_polling_trace)) {
-    gpr_atm old_cnt = gpr_atm_acq_load(&eps->ref_count);
-    gpr_log(GPR_DEBUG, "Add ref eps: %p, old:%" PRIdPTR " -> new:%" PRIdPTR
-                       " (%s) - (%s, %d)",
-            eps, old_cnt, old_cnt + 1, reason, file, line);
-  }
-  eps_add_ref(eps);
-}
-
-static void eps_unref_dbg(grpc_exec_ctx *exec_ctx, epoll_set *eps,
-                          const char *reason, const char *file, int line) {
-  if (GRPC_TRACER_ON(grpc_polling_trace)) {
-    gpr_atm old_cnt = gpr_atm_acq_load(&eps->ref_count);
-    gpr_log(GPR_DEBUG, "Unref eps: %p, old:%" PRIdPTR " -> new:%" PRIdPTR
-                       " (%s) - (%s, %d)",
-            eps, old_cnt, (old_cnt - 1), reason, file, line);
-  }
-  eps_unref(exec_ctx, eps);
-}
-#endif
-
-static void eps_add_ref(epoll_set *eps) {
-  gpr_atm_no_barrier_fetch_add(&eps->ref_count, 1);
-}
-
-static void eps_unref(grpc_exec_ctx *exec_ctx, epoll_set *eps) {
-  /* If ref count went to zero, delete the epoll set. This deletion is
-     not done under a lock since once the ref count goes to zero, we are
-     guaranteed that no one else holds a reference to the epoll set (and
-     that there is no racing eps_add_ref() call either).*/
-  if (1 == gpr_atm_full_fetch_add(&eps->ref_count, -1)) {
-    epoll_set_delete(eps);
-  }
-}
-
-static void epoll_set_add_fd_locked(epoll_set *eps, grpc_fd *fd,
-                                    grpc_error **error) {
-  int err;
-  struct epoll_event ev;
-  char *err_msg;
-  const char *err_desc = "epoll_set_add_fd_locked";
-
-#ifdef GRPC_TSAN
-  /* See the definition of g_epoll_sync for more context */
-  gpr_atm_rel_store(&g_epoll_sync, (gpr_atm)0);
-#endif /* defined(GRPC_TSAN) */
-
-  ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET);
-  ev.data.ptr = fd;
-  err = epoll_ctl(eps->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev);
-  if (err < 0 && errno != EEXIST) {
-    gpr_asprintf(
-        &err_msg,
-        "epoll_ctl (epoll_fd: %d) add fd: %d failed with error: %d (%s)",
-        eps->epoll_fd, fd->fd, errno, strerror(errno));
-    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-    gpr_free(err_msg);
-  }
-}
-
-static void epoll_set_add_wakeup_fd_locked(epoll_set *eps,
-                                           grpc_wakeup_fd *wakeup_fd,
-                                           grpc_error **error) {
-  struct epoll_event ev;
-  int err;
-  char *err_msg;
-  const char *err_desc = "epoll_set_add_wakeup_fd";
-
-  ev.events = (uint32_t)(EPOLLIN | EPOLLET);
-  ev.data.ptr = wakeup_fd;
-  err = epoll_ctl(eps->epoll_fd, EPOLL_CTL_ADD,
-                  GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), &ev);
-  if (err < 0 && errno != EEXIST) {
-    gpr_asprintf(&err_msg,
-                 "epoll_ctl (epoll_fd: %d) add wakeup fd: %d failed with "
-                 "error: %d (%s)",
-                 eps->epoll_fd, GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), errno,
-                 strerror(errno));
-    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-    gpr_free(err_msg);
-  }
-}
-
-static void epoll_set_remove_fd(epoll_set *eps, grpc_fd *fd, bool is_fd_closed,
-                                grpc_error **error) {
-  int err;
-  char *err_msg;
-  const char *err_desc = "epoll_set_remove_fd";
-
-  /* If fd is already closed, then it would have been automatically been removed
-     from the epoll set */
-  if (!is_fd_closed) {
-    err = epoll_ctl(eps->epoll_fd, EPOLL_CTL_DEL, fd->fd, NULL);
-    if (err < 0 && errno != ENOENT) {
-      gpr_asprintf(
-          &err_msg,
-          "epoll_ctl (epoll_fd: %d) del fd: %d failed with error: %d (%s)",
-          eps->epoll_fd, fd->fd, errno, strerror(errno));
-      append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-      gpr_free(err_msg);
-    }
-  }
-}
-
-/* Might return NULL in case of an error */
-static epoll_set *epoll_set_create(grpc_error **error) {
-  epoll_set *eps = NULL;
-  const char *err_desc = "epoll_set_create";
-
-  *error = GRPC_ERROR_NONE;
-
-  eps = gpr_malloc(sizeof(*eps));
-  eps->epoll_fd = -1;
-
-  gpr_mu_init(&eps->mu);
-
-  gpr_atm_rel_store(&eps->ref_count, 0);
-  gpr_atm_rel_store(&eps->poller_count, 0);
-
-  gpr_atm_rel_store(&eps->is_shutdown, false);
-
-  eps->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
-
-  if (eps->epoll_fd < 0) {
-    append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc);
-    goto done;
-  }
-
-done:
-  if (*error != GRPC_ERROR_NONE) {
-    epoll_set_delete(eps);
-    eps = NULL;
-  }
-  return eps;
-}
-
-static void epoll_set_delete(epoll_set *eps) {
-  if (eps->epoll_fd >= 0) {
-    close(eps->epoll_fd);
-  }
-
-  gpr_mu_destroy(&eps->mu);
-
-  gpr_free(eps);
-}
-
-static grpc_error *epoll_set_global_init() {
-  grpc_error *error = GRPC_ERROR_NONE;
-
-  error = grpc_wakeup_fd_init(&epoll_set_wakeup_fd);
-  if (error == GRPC_ERROR_NONE) {
-    error = grpc_wakeup_fd_wakeup(&epoll_set_wakeup_fd);
-  }
-
-  return error;
-}
-
-static void epoll_set_global_shutdown() {
-  grpc_wakeup_fd_destroy(&epoll_set_wakeup_fd);
-}
-
-/*******************************************************************************
- * Fd Definitions
- */
-
-/* We need to keep a freelist not because of any concerns of malloc performance
- * but instead so that implementations with multiple threads in (for example)
- * epoll_wait deal with the race between pollset removal and incoming poll
- * notifications.
- *
- * The problem is that the poller ultimately holds a reference to this
- * object, so it is very difficult to know when is safe to free it, at least
- * without some expensive synchronization.
- *
- * If we keep the object freelisted, in the worst case losing this race just
- * becomes a spurious read notification on a reused fd.
- */
-
-static grpc_fd *fd_freelist = NULL;
-static gpr_mu fd_freelist_mu;
-
-static grpc_fd *get_fd_from_freelist() {
-  grpc_fd *new_fd = NULL;
-
-  gpr_mu_lock(&fd_freelist_mu);
-  if (fd_freelist != NULL) {
-    new_fd = fd_freelist;
-    fd_freelist = fd_freelist->freelist_next;
-  }
-  gpr_mu_unlock(&fd_freelist_mu);
-  return new_fd;
-}
-
-static void add_fd_to_freelist(grpc_fd *fd) {
-  gpr_mu_lock(&fd_freelist_mu);
-  fd->freelist_next = fd_freelist;
-  fd_freelist = fd;
-  grpc_iomgr_unregister_object(&fd->iomgr_object);
-
-  grpc_lfev_destroy(&fd->read_closure);
-  grpc_lfev_destroy(&fd->write_closure);
-
-  gpr_mu_unlock(&fd_freelist_mu);
-}
-
-static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); }
-
-static void fd_global_shutdown(void) {
-  gpr_mu_lock(&fd_freelist_mu);
-  gpr_mu_unlock(&fd_freelist_mu);
-  while (fd_freelist != NULL) {
-    grpc_fd *fd = fd_freelist;
-    fd_freelist = fd_freelist->freelist_next;
-    gpr_mu_destroy(&fd->mu);
-    gpr_free(fd);
-  }
-  gpr_mu_destroy(&fd_freelist_mu);
-}
-
-static grpc_fd *fd_create(int fd, const char *name) {
-  grpc_fd *new_fd = get_fd_from_freelist();
-  if (new_fd == NULL) {
-    new_fd = gpr_malloc(sizeof(grpc_fd));
-    gpr_mu_init(&new_fd->mu);
-  }
-
-  /* Note: It is not really needed to get the new_fd->mu lock here. If this
-   * is a newly created fd (or an fd we got from the freelist), no one else
-   * would be holding a lock to it anyway. */
-  gpr_mu_lock(&new_fd->mu);
-  new_fd->eps = NULL;
-
-  new_fd->fd = fd;
-  new_fd->orphaned = false;
-  grpc_lfev_init(&new_fd->read_closure);
-  grpc_lfev_init(&new_fd->write_closure);
-
-  new_fd->freelist_next = NULL;
-  new_fd->on_done_closure = NULL;
-
-  gpr_mu_unlock(&new_fd->mu);
-
-  char *fd_name;
-  gpr_asprintf(&fd_name, "%s fd=%d", name, fd);
-  grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name);
-  gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, (void *)new_fd, fd_name);
-  gpr_free(fd_name);
-
-  /* Associate the fd with one of the eps */
-  add_fd_to_eps(new_fd);
-  return new_fd;
-}
-
-static int fd_wrapped_fd(grpc_fd *fd) {
-  int ret_fd = -1;
-  gpr_mu_lock(&fd->mu);
-  if (!fd->orphaned) {
-    ret_fd = fd->fd;
-  }
-  gpr_mu_unlock(&fd->mu);
-
-  return ret_fd;
-}
-
-static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                      grpc_closure *on_done, int *release_fd,
-                      bool already_closed, const char *reason) {
-  bool is_fd_closed = already_closed;
-  grpc_error *error = GRPC_ERROR_NONE;
-  epoll_set *unref_eps = NULL;
-
-  gpr_mu_lock(&fd->mu);
-  fd->on_done_closure = on_done;
-
-  /* If release_fd is not NULL, we should be relinquishing control of the file
-     descriptor fd->fd (but we still own the grpc_fd structure). */
-  if (release_fd != NULL) {
-    *release_fd = fd->fd;
-  } else if (!is_fd_closed) {
-    close(fd->fd);
-    is_fd_closed = true;
-  }
-
-  fd->orphaned = true;
-
-  /* Remove the fd from the epoll set */
-  if (fd->eps != NULL) {
-    epoll_set_remove_fd(fd->eps, fd, is_fd_closed, &error);
-    unref_eps = fd->eps;
-    fd->eps = NULL;
-  }
-
-  GRPC_CLOSURE_SCHED(exec_ctx, fd->on_done_closure, GRPC_ERROR_REF(error));
-
-  gpr_mu_unlock(&fd->mu);
-
-  /* We are done with this fd. Release it (i.e add back to freelist) */
-  add_fd_to_freelist(fd);
-
-  if (unref_eps != NULL) {
-    /* Unref stale epoll set here, outside the fd lock above.
-       The epoll set owns a workqueue which owns an fd, and unreffing
-       inside the lock can cause an eventual lock loop that makes TSAN very
-       unhappy. */
-    EPS_UNREF(exec_ctx, unref_eps, "fd_orphan");
-  }
-  GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error));
-  GRPC_ERROR_UNREF(error);
-}
-
-/* This polling engine doesn't really need the read notifier functionality. So
- * it just returns a dummy read notifier pollset */
-static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx,
-                                                  grpc_fd *fd) {
-  return &g_read_notifier;
-}
-
-static bool fd_is_shutdown(grpc_fd *fd) {
-  return grpc_lfev_is_shutdown(&fd->read_closure);
-}
-
-/* Might be called multiple times */
-static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_error *why) {
-  if (grpc_lfev_set_shutdown(exec_ctx, &fd->read_closure,
-                             GRPC_ERROR_REF(why))) {
-    shutdown(fd->fd, SHUT_RDWR);
-    grpc_lfev_set_shutdown(exec_ctx, &fd->write_closure, GRPC_ERROR_REF(why));
-  }
-  GRPC_ERROR_UNREF(why);
-}
-
-static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                              grpc_closure *closure) {
-  grpc_lfev_notify_on(exec_ctx, &fd->read_closure, closure, "read");
-}
-
-static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                               grpc_closure *closure) {
-  grpc_lfev_notify_on(exec_ctx, &fd->write_closure, closure, "write");
-}
-
-/*******************************************************************************
- * Pollset Definitions
- */
-/* TODO: sreek - Not needed anymore */
-GPR_TLS_DECL(g_current_thread_pollset);
-GPR_TLS_DECL(g_current_thread_worker);
-
-static void pollset_worker_init(grpc_pollset_worker *worker) {
-  worker->next = worker->prev = NULL;
-  gpr_cv_init(&worker->kick_cv);
-}
-
-/* Global state management */
-static grpc_error *pollset_global_init(void) {
-  gpr_tls_init(&g_current_thread_pollset);
-  gpr_tls_init(&g_current_thread_worker);
-  return GRPC_ERROR_NONE;
-}
-
-static void pollset_global_shutdown(void) {
-  gpr_tls_destroy(&g_current_thread_pollset);
-  gpr_tls_destroy(&g_current_thread_worker);
-}
-
-static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) {
-  gpr_cv_signal(&worker->kick_cv);
-  return GRPC_ERROR_NONE;
-}
-
-/* Return 1 if the pollset has active threads in pollset_work (pollset must
- * be locked) */
-static int pollset_has_workers(grpc_pollset *p) {
-  return p->root_worker.next != &p->root_worker;
-}
-
-static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
-  worker->prev->next = worker->next;
-  worker->next->prev = worker->prev;
-}
-
-static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) {
-  if (pollset_has_workers(p)) {
-    grpc_pollset_worker *w = p->root_worker.next;
-    remove_worker(p, w);
-    return w;
-  } else {
-    return NULL;
-  }
-}
-
-static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
-  worker->next = &p->root_worker;
-  worker->prev = worker->next->prev;
-  worker->prev->next = worker->next->prev = worker;
-}
-
-static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
-  worker->prev = &p->root_worker;
-  worker->next = worker->prev->next;
-  worker->prev->next = worker->next->prev = worker;
-}
-
-/* p->mu must be held before calling this function */
-static grpc_error *pollset_kick(grpc_pollset *p,
-                                grpc_pollset_worker *specific_worker) {
-  GPR_TIMER_BEGIN("pollset_kick", 0);
-  grpc_error *error = GRPC_ERROR_NONE;
-  const char *err_desc = "Kick Failure";
-  grpc_pollset_worker *worker = specific_worker;
-  if (worker != NULL) {
-    if (worker == GRPC_POLLSET_KICK_BROADCAST) {
-      if (pollset_has_workers(p)) {
-        GPR_TIMER_BEGIN("pollset_kick.broadcast", 0);
-        for (worker = p->root_worker.next; worker != &p->root_worker;
-             worker = worker->next) {
-          if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) {
-            append_error(&error, pollset_worker_kick(worker), err_desc);
-          }
-        }
-        GPR_TIMER_END("pollset_kick.broadcast", 0);
-      } else {
-        p->kicked_without_pollers = true;
-      }
-    } else {
-      GPR_TIMER_MARK("kicked_specifically", 0);
-      if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) {
-        append_error(&error, pollset_worker_kick(worker), err_desc);
-      }
-    }
-  } else if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)p) {
-    /* Since worker == NULL, it means that we can kick "any" worker on this
-       pollset 'p'. If 'p' happens to be the same pollset this thread is
-       currently polling (i.e in pollset_work() function), then there is no need
-       to kick any other worker since the current thread can just absorb the
-       kick. This is the reason why we enter this case only when
-       g_current_thread_pollset is != p */
-
-    GPR_TIMER_MARK("kick_anonymous", 0);
-    worker = pop_front_worker(p);
-    if (worker != NULL) {
-      GPR_TIMER_MARK("finally_kick", 0);
-      push_back_worker(p, worker);
-      append_error(&error, pollset_worker_kick(worker), err_desc);
-    } else {
-      GPR_TIMER_MARK("kicked_no_pollers", 0);
-      p->kicked_without_pollers = true;
-    }
-  }
-
-  GPR_TIMER_END("pollset_kick", 0);
-  GRPC_LOG_IF_ERROR("pollset_kick", GRPC_ERROR_REF(error));
-  return error;
-}
-
-static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) {
-  gpr_mu_init(&pollset->mu);
-  *mu = &pollset->mu;
-  pollset->eps = NULL;
-
-  pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker;
-  pollset->kicked_without_pollers = false;
-
-  pollset->shutting_down = false;
-  pollset->finish_shutdown_called = false;
-  pollset->shutdown_done = NULL;
-}
-
-static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  grpc_lfev_set_ready(exec_ctx, &fd->read_closure, "read");
-}
-
-static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  grpc_lfev_set_ready(exec_ctx, &fd->write_closure, "write");
-}
-
-static void pollset_release_epoll_set(grpc_exec_ctx *exec_ctx, grpc_pollset *ps,
-                                      char *reason) {
-  if (ps->eps != NULL) {
-    EPS_UNREF(exec_ctx, ps->eps, reason);
-  }
-  ps->eps = NULL;
-}
-
-static void finish_shutdown_locked(grpc_exec_ctx *exec_ctx,
-                                   grpc_pollset *pollset) {
-  /* The pollset cannot have any workers if we are at this stage */
-  GPR_ASSERT(!pollset_has_workers(pollset));
-
-  pollset->finish_shutdown_called = true;
-
-  /* Release the ref and set pollset->eps to NULL */
-  pollset_release_epoll_set(exec_ctx, pollset, "ps_shutdown");
-  GRPC_CLOSURE_SCHED(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE);
-}
-
-/* pollset->mu lock must be held by the caller before calling this */
-static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                             grpc_closure *closure) {
-  GPR_TIMER_BEGIN("pollset_shutdown", 0);
-  GPR_ASSERT(!pollset->shutting_down);
-  pollset->shutting_down = true;
-  pollset->shutdown_done = closure;
-  pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST);
-
-  /* If the pollset has any workers, we cannot call finish_shutdown_locked()
-     because it would release the underlying epoll set. In such a case, we
-     let the last worker call finish_shutdown_locked() from pollset_work() */
-  if (!pollset_has_workers(pollset)) {
-    GPR_ASSERT(!pollset->finish_shutdown_called);
-    GPR_TIMER_MARK("pollset_shutdown.finish_shutdown_locked", 0);
-    finish_shutdown_locked(exec_ctx, pollset);
-  }
-  GPR_TIMER_END("pollset_shutdown", 0);
-}
-
-/* pollset_shutdown is guaranteed to be called before pollset_destroy. So other
- * than destroying the mutexes, there is nothing special that needs to be done
- * here */
-static void pollset_destroy(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) {
-  GPR_ASSERT(!pollset_has_workers(pollset));
-  gpr_mu_destroy(&pollset->mu);
-}
-
-/* Blocking call */
-static void acquire_epoll_lease(epoll_set *eps) {
-  if (g_num_threads_per_eps > 1) {
-    gpr_mu_lock(&eps->mu);
-  }
-}
-
-static void release_epoll_lease(epoll_set *eps) {
-  if (g_num_threads_per_eps > 1) {
-    gpr_mu_unlock(&eps->mu);
-  }
-}
-
-#define GRPC_EPOLL_MAX_EVENTS 100
-static void do_epoll_wait(grpc_exec_ctx *exec_ctx, int epoll_fd, epoll_set *eps,
-                          grpc_error **error) {
-  struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS];
-  int ep_rv;
-  char *err_msg;
-  const char *err_desc = "do_epoll_wait";
-
-  int timeout_ms = -1;
-
-  GRPC_SCHEDULING_START_BLOCKING_REGION;
-  acquire_epoll_lease(eps);
-  GRPC_STATS_INC_SYSCALL_POLL(exec_ctx);
-  ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms);
-  release_epoll_lease(eps);
-  GRPC_SCHEDULING_END_BLOCKING_REGION;
-
-  if (ep_rv < 0) {
-    gpr_asprintf(&err_msg,
-                 "epoll_wait() epoll fd: %d failed with error: %d (%s)",
-                 epoll_fd, errno, strerror(errno));
-    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-  }
-
-#ifdef GRPC_TSAN
-  /* See the definition of g_poll_sync for more details */
-  gpr_atm_acq_load(&g_epoll_sync);
-#endif /* defined(GRPC_TSAN) */
-
-  for (int i = 0; i < ep_rv; ++i) {
-    void *data_ptr = ep_ev[i].data.ptr;
-    if (data_ptr == &epoll_set_wakeup_fd) {
-      gpr_atm_rel_store(&eps->is_shutdown, 1);
-      gpr_log(GPR_INFO, "pollset poller: shutdown set");
-    } else {
-      grpc_fd *fd = data_ptr;
-      int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP);
-      int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI);
-      int write_ev = ep_ev[i].events & EPOLLOUT;
-      if (read_ev || cancel) {
-        fd_become_readable(exec_ctx, fd);
-      }
-      if (write_ev || cancel) {
-        fd_become_writable(exec_ctx, fd);
-      }
-    }
-  }
-}
-
-static void epoll_set_work(grpc_exec_ctx *exec_ctx, epoll_set *eps,
-                           grpc_error **error) {
-  int epoll_fd = -1;
-  GPR_TIMER_BEGIN("epoll_set_work", 0);
-
-  /* Since epoll_fd is immutable, it is safe to read it without a lock on the
-     epoll set. */
-  epoll_fd = eps->epoll_fd;
-
-  gpr_atm_no_barrier_fetch_add(&eps->poller_count, 1);
-  g_current_thread_epoll_set = eps;
-
-  do_epoll_wait(exec_ctx, epoll_fd, eps, error);
-
-  g_current_thread_epoll_set = NULL;
-  gpr_atm_no_barrier_fetch_add(&eps->poller_count, -1);
-
-  GPR_TIMER_END("epoll_set_work", 0);
-}
-
-/* pollset->mu lock must be held by the caller before calling this.
-   The function pollset_work() may temporarily release the lock (pollset->mu)
-   during the course of its execution but it will always re-acquire the lock and
-   ensure that it is held by the time the function returns */
-static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                                grpc_pollset_worker **worker_hdl,
-                                gpr_timespec now, gpr_timespec deadline) {
-  GPR_TIMER_BEGIN("pollset_work", 0);
-  grpc_error *error = GRPC_ERROR_NONE;
-
-  grpc_pollset_worker worker;
-  pollset_worker_init(&worker);
-
-  if (worker_hdl) *worker_hdl = &worker;
-
-  gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset);
-  gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker);
-
-  if (pollset->kicked_without_pollers) {
-    /* If the pollset was kicked without pollers, pretend that the current
-       worker got the kick and skip polling. A kick indicates that there is some
-       work that needs attention like an event on the completion queue or an
-       alarm */
-    GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0);
-    pollset->kicked_without_pollers = 0;
-  } else if (!pollset->shutting_down) {
-    push_front_worker(pollset, &worker);
-
-    gpr_cv_wait(&worker.kick_cv, &pollset->mu,
-                gpr_convert_clock_type(deadline, GPR_CLOCK_REALTIME));
-    /* pollset->mu locked here */
-
-    remove_worker(pollset, &worker);
-  }
-
-  /* If we are the last worker on the pollset (i.e pollset_has_workers() is
-     false at this point) and the pollset is shutting down, we may have to
-     finish the shutdown process by calling finish_shutdown_locked().
-     See pollset_shutdown() for more details.
-
-     Note: Continuing to access pollset here is safe; it is the caller's
-     responsibility to not destroy a pollset when it has outstanding calls to
-     pollset_work() */
-  if (pollset->shutting_down && !pollset_has_workers(pollset) &&
-      !pollset->finish_shutdown_called) {
-    GPR_TIMER_MARK("pollset_work.finish_shutdown_locked", 0);
-    finish_shutdown_locked(exec_ctx, pollset);
-
-    gpr_mu_unlock(&pollset->mu);
-    grpc_exec_ctx_flush(exec_ctx);
-    gpr_mu_lock(&pollset->mu);
-  }
-
-  if (worker_hdl) *worker_hdl = NULL;
-
-  gpr_tls_set(&g_current_thread_pollset, (intptr_t)0);
-  gpr_tls_set(&g_current_thread_worker, (intptr_t)0);
-
-  GPR_TIMER_END("pollset_work", 0);
-
-  GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error));
-  return error;
-}
-
-static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                           grpc_fd *fd) {
-  /* Nothing to do */
-}
-
-/*******************************************************************************
- * Pollset-set Definitions
- */
-grpc_pollset_set g_dummy_pollset_set;
-static grpc_pollset_set *pollset_set_create(void) {
-  return &g_dummy_pollset_set;
-}
-
-static void pollset_set_destroy(grpc_exec_ctx *exec_ctx,
-                                grpc_pollset_set *pss) {
-  /* Nothing to do */
-}
-
-static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
-                               grpc_fd *fd) {
-  /* Nothing to do */
-}
-
-static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
-                               grpc_fd *fd) {
-  /* Nothing to do */
-}
-
-static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx,
-                                    grpc_pollset_set *pss, grpc_pollset *ps) {
-  /* Nothing to do */
-}
-
-static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx,
-                                    grpc_pollset_set *pss, grpc_pollset *ps) {
-  /* Nothing to do */
-}
-
-static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx,
-                                        grpc_pollset_set *bag,
-                                        grpc_pollset_set *item) {
-  /* Nothing to do */
-}
-
-static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx,
-                                        grpc_pollset_set *bag,
-                                        grpc_pollset_set *item) {
-  /* Nothing to do */
-}
-
-/*******************************************************************************
- * Event engine binding
- */
-
-static void shutdown_engine(void) {
-  shutdown_poller_threads();
-  shutdown_epoll_sets();
-  fd_global_shutdown();
-  pollset_global_shutdown();
-  epoll_set_global_shutdown();
-  gpr_log(GPR_INFO, "ev-epoll-threadpool engine shutdown complete");
-}
-
-static const grpc_event_engine_vtable vtable = {
-    .pollset_size = sizeof(grpc_pollset),
-
-    .fd_create = fd_create,
-    .fd_wrapped_fd = fd_wrapped_fd,
-    .fd_orphan = fd_orphan,
-    .fd_shutdown = fd_shutdown,
-    .fd_is_shutdown = fd_is_shutdown,
-    .fd_notify_on_read = fd_notify_on_read,
-    .fd_notify_on_write = fd_notify_on_write,
-    .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset,
-
-    .pollset_init = pollset_init,
-    .pollset_shutdown = pollset_shutdown,
-    .pollset_destroy = pollset_destroy,
-    .pollset_work = pollset_work,
-    .pollset_kick = pollset_kick,
-    .pollset_add_fd = pollset_add_fd,
-
-    .pollset_set_create = pollset_set_create,
-    .pollset_set_destroy = pollset_set_destroy,
-    .pollset_set_add_pollset = pollset_set_add_pollset,
-    .pollset_set_del_pollset = pollset_set_del_pollset,
-    .pollset_set_add_pollset_set = pollset_set_add_pollset_set,
-    .pollset_set_del_pollset_set = pollset_set_del_pollset_set,
-    .pollset_set_add_fd = pollset_set_add_fd,
-    .pollset_set_del_fd = pollset_set_del_fd,
-
-    .shutdown_engine = shutdown_engine,
-};
-
-/*****************************************************************************
- * Dedicated polling threads and pollsets - Definitions
- */
-static void add_fd_to_eps(grpc_fd *fd) {
-  GPR_ASSERT(fd->eps == NULL);
-  GPR_TIMER_BEGIN("add_fd_to_eps", 0);
-
-  grpc_error *error = GRPC_ERROR_NONE;
-  size_t idx = (size_t)gpr_atm_no_barrier_fetch_add(&g_next_eps, 1) % g_num_eps;
-  epoll_set *eps = g_epoll_sets[idx];
-
-  gpr_mu_lock(&fd->mu);
-
-  if (fd->orphaned) {
-    gpr_mu_unlock(&fd->mu);
-    return; /* Early out */
-  }
-
-  epoll_set_add_fd_locked(eps, fd, &error);
-  EPS_ADD_REF(eps, "fd");
-  fd->eps = eps;
-
-  GRPC_POLLING_TRACE("add_fd_to_eps (fd: %d, eps idx = %" PRIdPTR ")", fd->fd,
-                     idx);
-  gpr_mu_unlock(&fd->mu);
-
-  GRPC_LOG_IF_ERROR("add_fd_to_eps", error);
-  GPR_TIMER_END("add_fd_to_eps", 0);
-}
-
-static bool init_epoll_sets() {
-  grpc_error *error = GRPC_ERROR_NONE;
-  bool is_success = true;
-
-  g_epoll_sets = (epoll_set **)malloc(g_num_eps * sizeof(epoll_set *));
-
-  for (size_t i = 0; i < g_num_eps; i++) {
-    g_epoll_sets[i] = epoll_set_create(&error);
-    if (g_epoll_sets[i] == NULL) {
-      gpr_log(GPR_ERROR, "Error in creating a epoll set");
-      g_num_eps = i; /* Helps cleanup */
-      shutdown_epoll_sets();
-      is_success = false;
-      goto done;
-    }
-
-    EPS_ADD_REF(g_epoll_sets[i], "init_epoll_sets");
-  }
-
-  gpr_atm_no_barrier_store(&g_next_eps, 0);
-  gpr_mu *mu;
-  pollset_init(&g_read_notifier, &mu);
-
-done:
-  GRPC_LOG_IF_ERROR("init_epoll_sets", error);
-  return is_success;
-}
-
-static void shutdown_epoll_sets() {
-  if (!g_epoll_sets) {
-    return;
-  }
-
-  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
-  for (size_t i = 0; i < g_num_eps; i++) {
-    EPS_UNREF(&exec_ctx, g_epoll_sets[i], "shutdown_epoll_sets");
-  }
-  grpc_exec_ctx_flush(&exec_ctx);
-
-  gpr_free(g_epoll_sets);
-  g_epoll_sets = NULL;
-  pollset_destroy(&exec_ctx, &g_read_notifier);
-  grpc_exec_ctx_finish(&exec_ctx);
-}
-
-static void poller_thread_loop(void *arg) {
-  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
-  grpc_error *error = GRPC_ERROR_NONE;
-  epoll_set *eps = (epoll_set *)arg;
-
-  while (!gpr_atm_acq_load(&eps->is_shutdown)) {
-    epoll_set_work(&exec_ctx, eps, &error);
-    grpc_exec_ctx_flush(&exec_ctx);
-  }
-
-  grpc_exec_ctx_finish(&exec_ctx);
-  GRPC_LOG_IF_ERROR("poller_thread_loop", error);
-}
-
-/* g_epoll_sets MUST be initialized before calling this */
-static void start_poller_threads() {
-  GPR_ASSERT(g_epoll_sets);
-
-  gpr_log(GPR_INFO, "Starting poller threads");
-
-  size_t num_threads = g_num_eps * g_num_threads_per_eps;
-  g_poller_threads = (gpr_thd_id *)malloc(num_threads * sizeof(gpr_thd_id));
-  gpr_thd_options options = gpr_thd_options_default();
-  gpr_thd_options_set_joinable(&options);
-
-  for (size_t i = 0; i < num_threads; i++) {
-    gpr_thd_new(&g_poller_threads[i], poller_thread_loop,
-                (void *)g_epoll_sets[i % g_num_eps], &options);
-  }
-}
-
-static void shutdown_poller_threads() {
-  GPR_ASSERT(g_poller_threads);
-  GPR_ASSERT(g_epoll_sets);
-  grpc_error *error = GRPC_ERROR_NONE;
-
-  gpr_log(GPR_INFO, "Shutting down pollers");
-
-  epoll_set *eps = NULL;
-  size_t num_threads = g_num_eps * g_num_threads_per_eps;
-  for (size_t i = 0; i < num_threads; i++) {
-    eps = g_epoll_sets[i];
-    epoll_set_add_wakeup_fd_locked(eps, &epoll_set_wakeup_fd, &error);
-  }
-
-  for (size_t i = 0; i < g_num_eps; i++) {
-    gpr_thd_join(g_poller_threads[i]);
-  }
-
-  GRPC_LOG_IF_ERROR("shutdown_poller_threads", error);
-  gpr_free(g_poller_threads);
-  g_poller_threads = NULL;
-}
-
-/****************************************************************************/
-
-/* It is possible that GLIBC has epoll but the underlying kernel doesn't.
- * Create a dummy epoll_fd to make sure epoll support is available */
-static bool is_epoll_available() {
-  int fd = epoll_create1(EPOLL_CLOEXEC);
-  if (fd < 0) {
-    gpr_log(
-        GPR_ERROR,
-        "epoll_create1 failed with error: %d. Not using epoll polling engine",
-        fd);
-    return false;
-  }
-  close(fd);
-  return true;
-}
-
-const grpc_event_engine_vtable *grpc_init_epoll_thread_pool_linux(
-    bool requested_explicitly) {
-  if (!requested_explicitly) return NULL;
-
-  if (!grpc_has_wakeup_fd()) {
-    return NULL;
-  }
-
-  if (!is_epoll_available()) {
-    return NULL;
-  }
-
-  fd_global_init();
-
-  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
-    return NULL;
-  }
-
-  if (!GRPC_LOG_IF_ERROR("epoll_set_global_init", epoll_set_global_init())) {
-    return NULL;
-  }
-
-  if (!init_epoll_sets()) {
-    return NULL;
-  }
-
-  /* TODO (sreek): Maynot be a good idea to start threads here (especially if
-   * this engine doesn't get picked. Consider introducing an engine_init
-   * function in the vtable */
-  start_poller_threads();
-  return &vtable;
-}
-
-#else /* defined(GRPC_LINUX_EPOLL) */
-#if defined(GRPC_POSIX_SOCKET)
-#include "src/core/lib/iomgr/ev_posix.h"
-/* If GRPC_LINUX_EPOLL is not defined, it means epoll is not available. Return
- * NULL */
-const grpc_event_engine_vtable *grpc_init_epoll_thread_pool_linux(
-    bool requested_explicitly) {
-  return NULL;
-}
-#endif /* defined(GRPC_POSIX_SOCKET) */
-#endif /* !defined(GRPC_LINUX_EPOLL) */
diff --git a/src/core/lib/iomgr/ev_epoll_thread_pool_linux.h b/src/core/lib/iomgr/ev_epoll_thread_pool_linux.h
deleted file mode 100644
index 2ca68cc9d1..0000000000
--- a/src/core/lib/iomgr/ev_epoll_thread_pool_linux.h
+++ /dev/null
@@ -1,28 +0,0 @@
-/*
- *
- * Copyright 2017 gRPC authors.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- *
- */
-
-#ifndef GRPC_CORE_LIB_IOMGR_EV_EPOLL_THREAD_POOL_LINUX_H
-#define GRPC_CORE_LIB_IOMGR_EV_EPOLL_THREAD_POOL_LINUX_H
-
-#include "src/core/lib/iomgr/ev_posix.h"
-#include "src/core/lib/iomgr/port.h"
-
-const grpc_event_engine_vtable *grpc_init_epoll_thread_pool_linux(
-    bool requested_explicitly);
-
-#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLL_THREAD_POOL_LINUX_H */
diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c
index 424b40e425..bb43061ff5 100644
--- a/src/core/lib/iomgr/ev_posix.c
+++ b/src/core/lib/iomgr/ev_posix.c
@@ -31,8 +31,6 @@
 
 #include "src/core/lib/debug/trace.h"
 #include "src/core/lib/iomgr/ev_epoll1_linux.h"
-#include "src/core/lib/iomgr/ev_epoll_limited_pollers_linux.h"
-#include "src/core/lib/iomgr/ev_epoll_thread_pool_linux.h"
 #include "src/core/lib/iomgr/ev_epollex_linux.h"
 #include "src/core/lib/iomgr/ev_epollsig_linux.h"
 #include "src/core/lib/iomgr/ev_poll_posix.h"
@@ -66,8 +64,6 @@ typedef struct {
 static const event_engine_factory g_factories[] = {
     {"epoll1", grpc_init_epoll1_linux},
     {"epollsig", grpc_init_epollsig_linux},
-    {"epoll-threadpool", grpc_init_epoll_thread_pool_linux},
-    {"epoll-limited", grpc_init_epoll_limited_pollers_linux},
     {"poll", grpc_init_poll_posix},
     {"poll-cv", grpc_init_poll_cv_posix},
     {"epollex", grpc_init_epollex_linux},
diff --git a/src/core/lib/iomgr/iomgr.c b/src/core/lib/iomgr/iomgr.c
index 3d19953eeb..1feea6d628 100644
--- a/src/core/lib/iomgr/iomgr.c
+++ b/src/core/lib/iomgr/iomgr.c
@@ -164,13 +164,7 @@ void grpc_iomgr_unregister_object(grpc_iomgr_object *obj) {
 
 bool grpc_iomgr_abort_on_leaks(void) {
   char *env = gpr_getenv("GRPC_ABORT_ON_LEAKS");
-  if (env == NULL) return false;
-  static const char *truthy[] = {"yes",  "Yes",  "YES", "true",
-                                 "True", "TRUE", "1"};
-  bool should_we = false;
-  for (size_t i = 0; i < GPR_ARRAY_SIZE(truthy); i++) {
-    if (0 == strcmp(env, truthy[i])) should_we = true;
-  }
+  bool should_we = gpr_is_true(env);
   gpr_free(env);
   return should_we;
 }
diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c
index 793147f30a..98a2afd78b 100644
--- a/src/core/lib/iomgr/tcp_posix.c
+++ b/src/core/lib/iomgr/tcp_posix.c
@@ -250,7 +250,7 @@ static void tcp_do_read(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp) {
   msg.msg_name = NULL;
   msg.msg_namelen = 0;
   msg.msg_iov = iov;
-  msg.msg_iovlen = tcp->incoming_buffer->count;
+  msg.msg_iovlen = (msg_iovlen_type)tcp->incoming_buffer->count;
   msg.msg_control = NULL;
   msg.msg_controllen = 0;
   msg.msg_flags = 0;
diff --git a/src/core/lib/security/transport/client_auth_filter.c b/src/core/lib/security/transport/client_auth_filter.c
index 531a88434f..dd7dd44e79 100644
--- a/src/core/lib/security/transport/client_auth_filter.c
+++ b/src/core/lib/security/transport/client_auth_filter.c
@@ -39,6 +39,8 @@
 
 /* We can have a per-call credentials. */
 typedef struct {
+  grpc_call_stack *owning_call;
+  grpc_call_combiner *call_combiner;
   grpc_call_credentials *creds;
   bool have_host;
   bool have_method;
@@ -49,17 +51,12 @@ typedef struct {
      pollset_set so that work can progress when this call wants work to progress
   */
   grpc_polling_entity *pollent;
-  gpr_atm security_context_set;
-  gpr_mu security_context_mu;
   grpc_credentials_mdelem_array md_array;
   grpc_linked_mdelem md_links[MAX_CREDENTIALS_METADATA_COUNT];
   grpc_auth_metadata_context auth_md_context;
-  grpc_closure closure;
-  // Either 0 (no cancellation and no async operation in flight),
-  // a grpc_closure* (if the lowest bit is 0),
-  // or a grpc_error* (if the lowest bit is 1).
-  gpr_atm cancellation_state;
-  grpc_closure cancel_closure;
+  grpc_closure async_result_closure;
+  grpc_closure check_call_host_cancel_closure;
+  grpc_closure get_request_metadata_cancel_closure;
 } call_data;
 
 /* We can have a per-channel credentials. */
@@ -68,43 +65,6 @@ typedef struct {
   grpc_auth_context *auth_context;
 } channel_data;
 
-static void decode_cancel_state(gpr_atm cancel_state, grpc_closure **func,
-                                grpc_error **error) {
-  // If the lowest bit is 1, the value is a grpc_error*.
-  // Otherwise, if non-zdero, the value is a grpc_closure*.
-  if (cancel_state & 1) {
-    *error = (grpc_error *)(cancel_state & ~(gpr_atm)1);
-  } else if (cancel_state != 0) {
-    *func = (grpc_closure *)cancel_state;
-  }
-}
-
-static gpr_atm encode_cancel_state_error(grpc_error *error) {
-  // Set the lowest bit to 1 to indicate that it's an error.
-  return (gpr_atm)1 | (gpr_atm)error;
-}
-
-// Returns an error if the call has been cancelled.  Otherwise, sets the
-// cancellation function to be called upon cancellation.
-static grpc_error *set_cancel_func(grpc_call_element *elem,
-                                   grpc_iomgr_cb_func func) {
-  call_data *calld = (call_data *)elem->call_data;
-  // Decode original state.
-  gpr_atm original_state = gpr_atm_acq_load(&calld->cancellation_state);
-  grpc_error *original_error = GRPC_ERROR_NONE;
-  grpc_closure *original_func = NULL;
-  decode_cancel_state(original_state, &original_func, &original_error);
-  // If error is set, return it.
-  if (original_error != GRPC_ERROR_NONE) return GRPC_ERROR_REF(original_error);
-  // Otherwise, store func.
-  GRPC_CLOSURE_INIT(&calld->cancel_closure, func, elem,
-                    grpc_schedule_on_exec_ctx);
-  GPR_ASSERT(((gpr_atm)&calld->cancel_closure & (gpr_atm)1) == 0);
-  gpr_atm_rel_store(&calld->cancellation_state,
-                    (gpr_atm)&calld->cancel_closure);
-  return GRPC_ERROR_NONE;
-}
-
 static void reset_auth_metadata_context(
     grpc_auth_metadata_context *auth_md_context) {
   if (auth_md_context->service_url != NULL) {
@@ -153,7 +113,8 @@ static void on_credentials_metadata(grpc_exec_ctx *exec_ctx, void *arg,
   } else {
     error = grpc_error_set_int(error, GRPC_ERROR_INT_GRPC_STATUS,
                                GRPC_STATUS_UNAUTHENTICATED);
-    grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, batch, error);
+    grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, batch, error,
+                                                       calld->call_combiner);
   }
 }
 
@@ -191,8 +152,12 @@ static void cancel_get_request_metadata(grpc_exec_ctx *exec_ctx, void *arg,
                                         grpc_error *error) {
   grpc_call_element *elem = (grpc_call_element *)arg;
   call_data *calld = (call_data *)elem->call_data;
-  grpc_call_credentials_cancel_get_request_metadata(
-      exec_ctx, calld->creds, &calld->md_array, GRPC_ERROR_REF(error));
+  if (error != GRPC_ERROR_NONE) {
+    grpc_call_credentials_cancel_get_request_metadata(
+        exec_ctx, calld->creds, &calld->md_array, GRPC_ERROR_REF(error));
+  }
+  GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call,
+                        "cancel_get_request_metadata");
 }
 
 static void send_security_metadata(grpc_exec_ctx *exec_ctx,
@@ -223,7 +188,8 @@ static void send_security_metadata(grpc_exec_ctx *exec_ctx,
           grpc_error_set_int(
               GRPC_ERROR_CREATE_FROM_STATIC_STRING(
                   "Incompatible credentials set on channel and call."),
-              GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_UNAUTHENTICATED));
+              GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_UNAUTHENTICATED),
+          calld->call_combiner);
       return;
     }
   } else {
@@ -234,22 +200,25 @@ static void send_security_metadata(grpc_exec_ctx *exec_ctx,
   build_auth_metadata_context(&chand->security_connector->base,
                               chand->auth_context, calld);
 
-  grpc_error *cancel_error = set_cancel_func(elem, cancel_get_request_metadata);
-  if (cancel_error != GRPC_ERROR_NONE) {
-    grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, batch,
-                                                       cancel_error);
-    return;
-  }
   GPR_ASSERT(calld->pollent != NULL);
-  GRPC_CLOSURE_INIT(&calld->closure, on_credentials_metadata, batch,
-                    grpc_schedule_on_exec_ctx);
+
+  GRPC_CLOSURE_INIT(&calld->async_result_closure, on_credentials_metadata,
+                    batch, grpc_schedule_on_exec_ctx);
   grpc_error *error = GRPC_ERROR_NONE;
   if (grpc_call_credentials_get_request_metadata(
           exec_ctx, calld->creds, calld->pollent, calld->auth_md_context,
-          &calld->md_array, &calld->closure, &error)) {
+          &calld->md_array, &calld->async_result_closure, &error)) {
     // Synchronous return; invoke on_credentials_metadata() directly.
     on_credentials_metadata(exec_ctx, batch, error);
     GRPC_ERROR_UNREF(error);
+  } else {
+    // Async return; register cancellation closure with call combiner.
+    GRPC_CALL_STACK_REF(calld->owning_call, "cancel_get_request_metadata");
+    grpc_call_combiner_set_notify_on_cancel(
+        exec_ctx, calld->call_combiner,
+        GRPC_CLOSURE_INIT(&calld->get_request_metadata_cancel_closure,
+                          cancel_get_request_metadata, elem,
+                          grpc_schedule_on_exec_ctx));
   }
 }
 
@@ -258,7 +227,6 @@ static void on_host_checked(grpc_exec_ctx *exec_ctx, void *arg,
   grpc_transport_stream_op_batch *batch = (grpc_transport_stream_op_batch *)arg;
   grpc_call_element *elem = batch->handler_private.extra_arg;
   call_data *calld = elem->call_data;
-
   if (error == GRPC_ERROR_NONE) {
     send_security_metadata(exec_ctx, elem, batch);
   } else {
@@ -271,7 +239,8 @@ static void on_host_checked(grpc_exec_ctx *exec_ctx, void *arg,
         exec_ctx, batch,
         grpc_error_set_int(GRPC_ERROR_CREATE_FROM_COPIED_STRING(error_msg),
                            GRPC_ERROR_INT_GRPC_STATUS,
-                           GRPC_STATUS_UNAUTHENTICATED));
+                           GRPC_STATUS_UNAUTHENTICATED),
+        calld->call_combiner);
     gpr_free(error_msg);
   }
 }
@@ -281,9 +250,12 @@ static void cancel_check_call_host(grpc_exec_ctx *exec_ctx, void *arg,
   grpc_call_element *elem = (grpc_call_element *)arg;
   call_data *calld = (call_data *)elem->call_data;
   channel_data *chand = (channel_data *)elem->channel_data;
-  grpc_channel_security_connector_cancel_check_call_host(
-      exec_ctx, chand->security_connector, &calld->closure,
-      GRPC_ERROR_REF(error));
+  if (error != GRPC_ERROR_NONE) {
+    grpc_channel_security_connector_cancel_check_call_host(
+        exec_ctx, chand->security_connector, &calld->async_result_closure,
+        GRPC_ERROR_REF(error));
+  }
+  GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call, "cancel_check_call_host");
 }
 
 static void auth_start_transport_stream_op_batch(
@@ -295,52 +267,19 @@ static void auth_start_transport_stream_op_batch(
   call_data *calld = elem->call_data;
   channel_data *chand = elem->channel_data;
 
-  if (batch->cancel_stream) {
-    while (true) {
-      // Decode the original cancellation state.
-      gpr_atm original_state = gpr_atm_acq_load(&calld->cancellation_state);
-      grpc_error *cancel_error = GRPC_ERROR_NONE;
-      grpc_closure *func = NULL;
-      decode_cancel_state(original_state, &func, &cancel_error);
-      // If we had already set a cancellation error, there's nothing
-      // more to do.
-      if (cancel_error != GRPC_ERROR_NONE) break;
-      // If there's a cancel func, call it.
-      // Note that even if the cancel func has been changed by some
-      // other thread between when we decoded it and now, it will just
-      // be a no-op.
-      cancel_error = GRPC_ERROR_REF(batch->payload->cancel_stream.cancel_error);
-      if (func != NULL) {
-        GRPC_CLOSURE_SCHED(exec_ctx, func, GRPC_ERROR_REF(cancel_error));
-      }
-      // Encode the new error into cancellation state.
-      if (gpr_atm_full_cas(&calld->cancellation_state, original_state,
-                           encode_cancel_state_error(cancel_error))) {
-        break;  // Success.
-      }
-      // The cas failed, so try again.
-    }
-  } else {
-    /* double checked lock over security context to ensure it's set once */
-    if (gpr_atm_acq_load(&calld->security_context_set) == 0) {
-      gpr_mu_lock(&calld->security_context_mu);
-      if (gpr_atm_acq_load(&calld->security_context_set) == 0) {
-        GPR_ASSERT(batch->payload->context != NULL);
-        if (batch->payload->context[GRPC_CONTEXT_SECURITY].value == NULL) {
-          batch->payload->context[GRPC_CONTEXT_SECURITY].value =
-              grpc_client_security_context_create();
-          batch->payload->context[GRPC_CONTEXT_SECURITY].destroy =
-              grpc_client_security_context_destroy;
-        }
-        grpc_client_security_context *sec_ctx =
-            batch->payload->context[GRPC_CONTEXT_SECURITY].value;
-        GRPC_AUTH_CONTEXT_UNREF(sec_ctx->auth_context, "client auth filter");
-        sec_ctx->auth_context =
-            GRPC_AUTH_CONTEXT_REF(chand->auth_context, "client_auth_filter");
-        gpr_atm_rel_store(&calld->security_context_set, 1);
-      }
-      gpr_mu_unlock(&calld->security_context_mu);
+  if (!batch->cancel_stream) {
+    GPR_ASSERT(batch->payload->context != NULL);
+    if (batch->payload->context[GRPC_CONTEXT_SECURITY].value == NULL) {
+      batch->payload->context[GRPC_CONTEXT_SECURITY].value =
+          grpc_client_security_context_create();
+      batch->payload->context[GRPC_CONTEXT_SECURITY].destroy =
+          grpc_client_security_context_destroy;
     }
+    grpc_client_security_context *sec_ctx =
+        batch->payload->context[GRPC_CONTEXT_SECURITY].value;
+    GRPC_AUTH_CONTEXT_UNREF(sec_ctx->auth_context, "client auth filter");
+    sec_ctx->auth_context =
+        GRPC_AUTH_CONTEXT_REF(chand->auth_context, "client_auth_filter");
   }
 
   if (batch->send_initial_metadata) {
@@ -365,26 +304,27 @@ static void auth_start_transport_stream_op_batch(
       }
     }
     if (calld->have_host) {
-      grpc_error *cancel_error = set_cancel_func(elem, cancel_check_call_host);
-      if (cancel_error != GRPC_ERROR_NONE) {
-        grpc_transport_stream_op_batch_finish_with_failure(exec_ctx, batch,
-                                                           cancel_error);
+      batch->handler_private.extra_arg = elem;
+      GRPC_CLOSURE_INIT(&calld->async_result_closure, on_host_checked, batch,
+                        grpc_schedule_on_exec_ctx);
+      char *call_host = grpc_slice_to_c_string(calld->host);
+      grpc_error *error = GRPC_ERROR_NONE;
+      if (grpc_channel_security_connector_check_call_host(
+              exec_ctx, chand->security_connector, call_host,
+              chand->auth_context, &calld->async_result_closure, &error)) {
+        // Synchronous return; invoke on_host_checked() directly.
+        on_host_checked(exec_ctx, batch, error);
+        GRPC_ERROR_UNREF(error);
       } else {
-        char *call_host = grpc_slice_to_c_string(calld->host);
-        batch->handler_private.extra_arg = elem;
-        grpc_error *error = GRPC_ERROR_NONE;
-        if (grpc_channel_security_connector_check_call_host(
-                exec_ctx, chand->security_connector, call_host,
-                chand->auth_context,
-                GRPC_CLOSURE_INIT(&calld->closure, on_host_checked, batch,
-                                  grpc_schedule_on_exec_ctx),
-                &error)) {
-          // Synchronous return; invoke on_host_checked() directly.
-          on_host_checked(exec_ctx, batch, error);
-          GRPC_ERROR_UNREF(error);
-        }
-        gpr_free(call_host);
+        // Async return; register cancellation closure with call combiner.
+        GRPC_CALL_STACK_REF(calld->owning_call, "cancel_check_call_host");
+        grpc_call_combiner_set_notify_on_cancel(
+            exec_ctx, calld->call_combiner,
+            GRPC_CLOSURE_INIT(&calld->check_call_host_cancel_closure,
+                              cancel_check_call_host, elem,
+                              grpc_schedule_on_exec_ctx));
       }
+      gpr_free(call_host);
       GPR_TIMER_END("auth_start_transport_stream_op_batch", 0);
       return; /* early exit */
     }
@@ -400,8 +340,8 @@ static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
                                   grpc_call_element *elem,
                                   const grpc_call_element_args *args) {
   call_data *calld = elem->call_data;
-  memset(calld, 0, sizeof(*calld));
-  gpr_mu_init(&calld->security_context_mu);
+  calld->owning_call = args->call_stack;
+  calld->call_combiner = args->call_combiner;
   return GRPC_ERROR_NONE;
 }
 
@@ -426,12 +366,6 @@ static void destroy_call_elem(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     grpc_slice_unref_internal(exec_ctx, calld->method);
   }
   reset_auth_metadata_context(&calld->auth_md_context);
-  gpr_mu_destroy(&calld->security_context_mu);
-  gpr_atm cancel_state = gpr_atm_acq_load(&calld->cancellation_state);
-  grpc_error *cancel_error = GRPC_ERROR_NONE;
-  grpc_closure *cancel_func = NULL;
-  decode_cancel_state(cancel_state, &cancel_func, &cancel_error);
-  GRPC_ERROR_UNREF(cancel_error);
 }
 
 /* Constructor for channel_data */
@@ -490,6 +424,5 @@ const grpc_channel_filter grpc_client_auth_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "client-auth"};
diff --git a/src/core/lib/security/transport/server_auth_filter.c b/src/core/lib/security/transport/server_auth_filter.c
index 9bf3f0ca0f..7f523c0883 100644
--- a/src/core/lib/security/transport/server_auth_filter.c
+++ b/src/core/lib/security/transport/server_auth_filter.c
@@ -26,7 +26,15 @@
 #include "src/core/lib/security/transport/auth_filters.h"
 #include "src/core/lib/slice/slice_internal.h"
 
+typedef enum {
+  STATE_INIT = 0,
+  STATE_DONE,
+  STATE_CANCELLED,
+} async_state;
+
 typedef struct call_data {
+  grpc_call_combiner *call_combiner;
+  grpc_call_stack *owning_call;
   grpc_transport_stream_op_batch *recv_initial_metadata_batch;
   grpc_closure *original_recv_initial_metadata_ready;
   grpc_closure recv_initial_metadata_ready;
@@ -34,6 +42,8 @@ typedef struct call_data {
   const grpc_metadata *consumed_md;
   size_t num_consumed_md;
   grpc_auth_context *auth_context;
+  grpc_closure cancel_closure;
+  gpr_atm state;  // async_state
 } call_data;
 
 typedef struct channel_data {
@@ -78,54 +88,94 @@ static grpc_filtered_mdelem remove_consumed_md(grpc_exec_ctx *exec_ctx,
   return GRPC_FILTERED_MDELEM(md);
 }
 
-/* called from application code */
-static void on_md_processing_done(
-    void *user_data, const grpc_metadata *consumed_md, size_t num_consumed_md,
-    const grpc_metadata *response_md, size_t num_response_md,
-    grpc_status_code status, const char *error_details) {
-  grpc_call_element *elem = user_data;
+static void on_md_processing_done_inner(grpc_exec_ctx *exec_ctx,
+                                        grpc_call_element *elem,
+                                        const grpc_metadata *consumed_md,
+                                        size_t num_consumed_md,
+                                        const grpc_metadata *response_md,
+                                        size_t num_response_md,
+                                        grpc_error *error) {
   call_data *calld = elem->call_data;
   grpc_transport_stream_op_batch *batch = calld->recv_initial_metadata_batch;
-  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
   /* TODO(jboeuf): Implement support for response_md. */
   if (response_md != NULL && num_response_md > 0) {
     gpr_log(GPR_INFO,
             "response_md in auth metadata processing not supported for now. "
             "Ignoring...");
   }
-  grpc_error *error = GRPC_ERROR_NONE;
-  if (status == GRPC_STATUS_OK) {
+  if (error == GRPC_ERROR_NONE) {
     calld->consumed_md = consumed_md;
     calld->num_consumed_md = num_consumed_md;
     error = grpc_metadata_batch_filter(
-        &exec_ctx, batch->payload->recv_initial_metadata.recv_initial_metadata,
+        exec_ctx, batch->payload->recv_initial_metadata.recv_initial_metadata,
         remove_consumed_md, elem, "Response metadata filtering error");
-  } else {
-    if (error_details == NULL) {
-      error_details = "Authentication metadata processing failed.";
+  }
+  GRPC_CLOSURE_SCHED(exec_ctx, calld->original_recv_initial_metadata_ready,
+                     error);
+}
+
+// Called from application code.
+static void on_md_processing_done(
+    void *user_data, const grpc_metadata *consumed_md, size_t num_consumed_md,
+    const grpc_metadata *response_md, size_t num_response_md,
+    grpc_status_code status, const char *error_details) {
+  grpc_call_element *elem = user_data;
+  call_data *calld = elem->call_data;
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+  // If the call was not cancelled while we were in flight, process the result.
+  if (gpr_atm_full_cas(&calld->state, (gpr_atm)STATE_INIT,
+                       (gpr_atm)STATE_DONE)) {
+    grpc_error *error = GRPC_ERROR_NONE;
+    if (status != GRPC_STATUS_OK) {
+      if (error_details == NULL) {
+        error_details = "Authentication metadata processing failed.";
+      }
+      error = grpc_error_set_int(
+          GRPC_ERROR_CREATE_FROM_COPIED_STRING(error_details),
+          GRPC_ERROR_INT_GRPC_STATUS, status);
     }
-    error =
-        grpc_error_set_int(GRPC_ERROR_CREATE_FROM_COPIED_STRING(error_details),
-                           GRPC_ERROR_INT_GRPC_STATUS, status);
+    on_md_processing_done_inner(&exec_ctx, elem, consumed_md, num_consumed_md,
+                                response_md, num_response_md, error);
   }
+  // Clean up.
   for (size_t i = 0; i < calld->md.count; i++) {
     grpc_slice_unref_internal(&exec_ctx, calld->md.metadata[i].key);
     grpc_slice_unref_internal(&exec_ctx, calld->md.metadata[i].value);
   }
   grpc_metadata_array_destroy(&calld->md);
-  GRPC_CLOSURE_SCHED(&exec_ctx, calld->original_recv_initial_metadata_ready,
-                     error);
+  GRPC_CALL_STACK_UNREF(&exec_ctx, calld->owning_call, "server_auth_metadata");
   grpc_exec_ctx_finish(&exec_ctx);
 }
 
+static void cancel_call(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) {
+  grpc_call_element *elem = (grpc_call_element *)arg;
+  call_data *calld = elem->call_data;
+  // If the result was not already processed, invoke the callback now.
+  if (error != GRPC_ERROR_NONE &&
+      gpr_atm_full_cas(&calld->state, (gpr_atm)STATE_INIT,
+                       (gpr_atm)STATE_CANCELLED)) {
+    on_md_processing_done_inner(exec_ctx, elem, NULL, 0, NULL, 0,
+                                GRPC_ERROR_REF(error));
+  }
+  GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call, "cancel_call");
+}
+
 static void recv_initial_metadata_ready(grpc_exec_ctx *exec_ctx, void *arg,
                                         grpc_error *error) {
-  grpc_call_element *elem = arg;
+  grpc_call_element *elem = (grpc_call_element *)arg;
   channel_data *chand = elem->channel_data;
   call_data *calld = elem->call_data;
   grpc_transport_stream_op_batch *batch = calld->recv_initial_metadata_batch;
   if (error == GRPC_ERROR_NONE) {
     if (chand->creds != NULL && chand->creds->processor.process != NULL) {
+      // We're calling out to the application, so we need to make sure
+      // to drop the call combiner early if we get cancelled.
+      GRPC_CALL_STACK_REF(calld->owning_call, "cancel_call");
+      GRPC_CLOSURE_INIT(&calld->cancel_closure, cancel_call, elem,
+                        grpc_schedule_on_exec_ctx);
+      grpc_call_combiner_set_notify_on_cancel(exec_ctx, calld->call_combiner,
+                                              &calld->cancel_closure);
+      GRPC_CALL_STACK_REF(calld->owning_call, "server_auth_metadata");
       calld->md = metadata_batch_to_md_array(
           batch->payload->recv_initial_metadata.recv_initial_metadata);
       chand->creds->processor.process(
@@ -159,6 +209,8 @@ static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
                                   const grpc_call_element_args *args) {
   call_data *calld = elem->call_data;
   channel_data *chand = elem->channel_data;
+  calld->call_combiner = args->call_combiner;
+  calld->owning_call = args->call_stack;
   GRPC_CLOSURE_INIT(&calld->recv_initial_metadata_ready,
                     recv_initial_metadata_ready, elem,
                     grpc_schedule_on_exec_ctx);
@@ -218,6 +270,5 @@ const grpc_channel_filter grpc_server_auth_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "server-auth"};
diff --git a/src/core/lib/support/string.c b/src/core/lib/support/string.c
index b65009754a..ec93303024 100644
--- a/src/core/lib/support/string.c
+++ b/src/core/lib/support/string.c
@@ -298,3 +298,16 @@ void *gpr_memrchr(const void *s, int c, size_t n) {
   }
   return NULL;
 }
+
+bool gpr_is_true(const char *s) {
+  if (s == NULL) {
+    return false;
+  }
+  static const char *truthy[] = {"yes", "true", "1"};
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(truthy); i++) {
+    if (0 == gpr_stricmp(s, truthy[i])) {
+      return true;
+    }
+  }
+  return false;
+}
diff --git a/src/core/lib/support/string.h b/src/core/lib/support/string.h
index e11df8439d..5a56fa3a0a 100644
--- a/src/core/lib/support/string.h
+++ b/src/core/lib/support/string.h
@@ -19,6 +19,7 @@
 #ifndef GRPC_CORE_LIB_SUPPORT_STRING_H
 #define GRPC_CORE_LIB_SUPPORT_STRING_H
 
+#include <stdbool.h>
 #include <stddef.h>
 
 #include <grpc/support/port_platform.h>
@@ -106,6 +107,8 @@ int gpr_stricmp(const char *a, const char *b);
 
 void *gpr_memrchr(const void *s, int c, size_t n);
 
+/** Return true if lower(s) equals "true", "yes" or "1", otherwise false. */
+bool gpr_is_true(const char *s);
 #ifdef __cplusplus
 }
 #endif
diff --git a/src/core/lib/surface/call.c b/src/core/lib/surface/call.c
index e68c201134..3aa20ffcd7 100644
--- a/src/core/lib/surface/call.c
+++ b/src/core/lib/surface/call.c
@@ -122,6 +122,7 @@ typedef struct batch_control {
       bool is_closure;
     } notify_tag;
   } completion_data;
+  grpc_closure start_batch;
   grpc_closure finish_batch;
   gpr_refcount steps_to_complete;
 
@@ -151,6 +152,7 @@ typedef struct {
 struct grpc_call {
   gpr_refcount ext_ref;
   gpr_arena *arena;
+  grpc_call_combiner call_combiner;
   grpc_completion_queue *cq;
   grpc_polling_entity pollent;
   grpc_channel *channel;
@@ -184,6 +186,11 @@ struct grpc_call {
      Element 0 is initial metadata, element 1 is trailing metadata. */
   grpc_metadata_array *buffered_metadata[2];
 
+  grpc_metadata compression_md;
+
+  // A char* indicating the peer name.
+  gpr_atm peer_string;
+
   /* Packed received call statuses from various sources */
   gpr_atm status[STATUS_SOURCE_COUNT];
 
@@ -262,8 +269,9 @@ grpc_tracer_flag grpc_compression_trace =
 #define CALL_FROM_TOP_ELEM(top_elem) \
   CALL_FROM_CALL_STACK(grpc_call_stack_from_top_element(top_elem))
 
-static void execute_op(grpc_exec_ctx *exec_ctx, grpc_call *call,
-                       grpc_transport_stream_op_batch *op);
+static void execute_batch(grpc_exec_ctx *exec_ctx, grpc_call *call,
+                          grpc_transport_stream_op_batch *op,
+                          grpc_closure *start_batch_closure);
 static void cancel_with_status(grpc_exec_ctx *exec_ctx, grpc_call *c,
                                status_source source, grpc_status_code status,
                                const char *description);
@@ -328,6 +336,7 @@ grpc_error *grpc_call_create(grpc_exec_ctx *exec_ctx,
                          sizeof(grpc_call) + channel_stack->call_stack_size);
   gpr_ref_init(&call->ext_ref, 1);
   call->arena = arena;
+  grpc_call_combiner_init(&call->call_combiner);
   *out_call = call;
   call->channel = args->channel;
   call->cq = args->cq;
@@ -436,7 +445,8 @@ grpc_error *grpc_call_create(grpc_exec_ctx *exec_ctx,
       .path = path,
       .start_time = call->start_time,
       .deadline = send_deadline,
-      .arena = call->arena};
+      .arena = call->arena,
+      .call_combiner = &call->call_combiner};
   add_init_error(&error, grpc_call_stack_init(exec_ctx, channel_stack, 1,
                                               destroy_call, call, &call_args));
   if (error != GRPC_ERROR_NONE) {
@@ -503,6 +513,8 @@ static void release_call(grpc_exec_ctx *exec_ctx, void *call,
                          grpc_error *error) {
   grpc_call *c = call;
   grpc_channel *channel = c->channel;
+  grpc_call_combiner_destroy(&c->call_combiner);
+  gpr_free((char *)c->peer_string);
   grpc_channel_update_call_size_estimate(channel, gpr_arena_destroy(c->arena));
   GRPC_CHANNEL_INTERNAL_UNREF(exec_ctx, channel, "call");
 }
@@ -586,6 +598,12 @@ void grpc_call_unref(grpc_call *c) {
   if (cancel) {
     cancel_with_error(&exec_ctx, c, STATUS_FROM_API_OVERRIDE,
                       GRPC_ERROR_CANCELLED);
+  } else {
+    // Unset the call combiner cancellation closure.  This has the
+    // effect of scheduling the previously set cancellation closure, if
+    // any, so that it can release any internal references it may be
+    // holding to the call stack.
+    grpc_call_combiner_set_notify_on_cancel(&exec_ctx, &c->call_combiner, NULL);
   }
   GRPC_CALL_INTERNAL_UNREF(&exec_ctx, c, "destroy");
   grpc_exec_ctx_finish(&exec_ctx);
@@ -602,30 +620,37 @@ grpc_call_error grpc_call_cancel(grpc_call *call, void *reserved) {
   return GRPC_CALL_OK;
 }
 
-static void execute_op(grpc_exec_ctx *exec_ctx, grpc_call *call,
-                       grpc_transport_stream_op_batch *op) {
-  grpc_call_element *elem;
-
-  GPR_TIMER_BEGIN("execute_op", 0);
-  elem = CALL_ELEM_FROM_CALL(call, 0);
-  elem->filter->start_transport_stream_op_batch(exec_ctx, elem, op);
-  GPR_TIMER_END("execute_op", 0);
+// This is called via the call combiner to start sending a batch down
+// the filter stack.
+static void execute_batch_in_call_combiner(grpc_exec_ctx *exec_ctx, void *arg,
+                                           grpc_error *ignored) {
+  grpc_transport_stream_op_batch *batch = arg;
+  grpc_call *call = batch->handler_private.extra_arg;
+  GPR_TIMER_BEGIN("execute_batch", 0);
+  grpc_call_element *elem = CALL_ELEM_FROM_CALL(call, 0);
+  GRPC_CALL_LOG_OP(GPR_INFO, elem, batch);
+  elem->filter->start_transport_stream_op_batch(exec_ctx, elem, batch);
+  GPR_TIMER_END("execute_batch", 0);
+}
+
+// start_batch_closure points to a caller-allocated closure to be used
+// for entering the call combiner.
+static void execute_batch(grpc_exec_ctx *exec_ctx, grpc_call *call,
+                          grpc_transport_stream_op_batch *batch,
+                          grpc_closure *start_batch_closure) {
+  batch->handler_private.extra_arg = call;
+  GRPC_CLOSURE_INIT(start_batch_closure, execute_batch_in_call_combiner, batch,
+                    grpc_schedule_on_exec_ctx);
+  GRPC_CALL_COMBINER_START(exec_ctx, &call->call_combiner, start_batch_closure,
+                           GRPC_ERROR_NONE, "executing batch");
 }
 
 char *grpc_call_get_peer(grpc_call *call) {
-  grpc_call_element *elem = CALL_ELEM_FROM_CALL(call, 0);
-  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
-  char *result;
-  GRPC_API_TRACE("grpc_call_get_peer(%p)", 1, (call));
-  result = elem->filter->get_peer(&exec_ctx, elem);
-  if (result == NULL) {
-    result = grpc_channel_get_target(call->channel);
-  }
-  if (result == NULL) {
-    result = gpr_strdup("unknown");
-  }
-  grpc_exec_ctx_finish(&exec_ctx);
-  return result;
+  char *peer_string = (char *)gpr_atm_acq_load(&call->peer_string);
+  if (peer_string != NULL) return gpr_strdup(peer_string);
+  peer_string = grpc_channel_get_target(call->channel);
+  if (peer_string != NULL) return peer_string;
+  return gpr_strdup("unknown");
 }
 
 grpc_call *grpc_call_from_top_element(grpc_call_element *elem) {
@@ -652,20 +677,41 @@ grpc_call_error grpc_call_cancel_with_status(grpc_call *c,
   return GRPC_CALL_OK;
 }
 
-static void done_termination(grpc_exec_ctx *exec_ctx, void *call,
+typedef struct {
+  grpc_call *call;
+  grpc_closure start_batch;
+  grpc_closure finish_batch;
+} cancel_state;
+
+// The on_complete callback used when sending a cancel_stream batch down
+// the filter stack.  Yields the call combiner when the batch is done.
+static void done_termination(grpc_exec_ctx *exec_ctx, void *arg,
                              grpc_error *error) {
-  GRPC_CALL_INTERNAL_UNREF(exec_ctx, call, "termination");
+  cancel_state *state = (cancel_state *)arg;
+  GRPC_CALL_COMBINER_STOP(exec_ctx, &state->call->call_combiner,
+                          "on_complete for cancel_stream op");
+  GRPC_CALL_INTERNAL_UNREF(exec_ctx, state->call, "termination");
+  gpr_free(state);
 }
 
 static void cancel_with_error(grpc_exec_ctx *exec_ctx, grpc_call *c,
                               status_source source, grpc_error *error) {
   GRPC_CALL_INTERNAL_REF(c, "termination");
+  // Inform the call combiner of the cancellation, so that it can cancel
+  // any in-flight asynchronous actions that may be holding the call
+  // combiner.  This ensures that the cancel_stream batch can be sent
+  // down the filter stack in a timely manner.
+  grpc_call_combiner_cancel(exec_ctx, &c->call_combiner, GRPC_ERROR_REF(error));
   set_status_from_error(exec_ctx, c, source, GRPC_ERROR_REF(error));
-  grpc_transport_stream_op_batch *op = grpc_make_transport_stream_op(
-      GRPC_CLOSURE_CREATE(done_termination, c, grpc_schedule_on_exec_ctx));
+  cancel_state *state = (cancel_state *)gpr_malloc(sizeof(*state));
+  state->call = c;
+  GRPC_CLOSURE_INIT(&state->finish_batch, done_termination, state,
+                    grpc_schedule_on_exec_ctx);
+  grpc_transport_stream_op_batch *op =
+      grpc_make_transport_stream_op(&state->finish_batch);
   op->cancel_stream = true;
   op->payload->cancel_stream.cancel_error = error;
-  execute_op(exec_ctx, c, op);
+  execute_batch(exec_ctx, c, op, &state->start_batch);
 }
 
 static grpc_error *error_from_status(grpc_status_code status,
@@ -1431,6 +1477,18 @@ static void receiving_stream_ready(grpc_exec_ctx *exec_ctx, void *bctlp,
   }
 }
 
+// The recv_message_ready callback used when sending a batch containing
+// a recv_message op down the filter stack.  Yields the call combiner
+// before processing the received message.
+static void receiving_stream_ready_in_call_combiner(grpc_exec_ctx *exec_ctx,
+                                                    void *bctlp,
+                                                    grpc_error *error) {
+  batch_control *bctl = bctlp;
+  grpc_call *call = bctl->call;
+  GRPC_CALL_COMBINER_STOP(exec_ctx, &call->call_combiner, "recv_message_ready");
+  receiving_stream_ready(exec_ctx, bctlp, error);
+}
+
 static void validate_filtered_metadata(grpc_exec_ctx *exec_ctx,
                                        batch_control *bctl) {
   grpc_call *call = bctl->call;
@@ -1537,6 +1595,9 @@ static void receiving_initial_metadata_ready(grpc_exec_ctx *exec_ctx,
   batch_control *bctl = bctlp;
   grpc_call *call = bctl->call;
 
+  GRPC_CALL_COMBINER_STOP(exec_ctx, &call->call_combiner,
+                          "recv_initial_metadata_ready");
+
   add_batch_error(exec_ctx, bctl, GRPC_ERROR_REF(error), false);
   if (error == GRPC_ERROR_NONE) {
     grpc_metadata_batch *md =
@@ -1590,7 +1651,8 @@ static void receiving_initial_metadata_ready(grpc_exec_ctx *exec_ctx,
 static void finish_batch(grpc_exec_ctx *exec_ctx, void *bctlp,
                          grpc_error *error) {
   batch_control *bctl = bctlp;
-
+  grpc_call *call = bctl->call;
+  GRPC_CALL_COMBINER_STOP(exec_ctx, &call->call_combiner, "on_complete");
   add_batch_error(exec_ctx, bctl, GRPC_ERROR_REF(error), false);
   finish_batch_step(exec_ctx, bctl);
 }
@@ -1610,9 +1672,6 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
   int num_completion_callbacks_needed = 1;
   grpc_call_error error = GRPC_CALL_OK;
 
-  // sent_initial_metadata guards against variable reuse.
-  grpc_metadata compression_md;
-
   GPR_TIMER_BEGIN("grpc_call_start_batch", 0);
   GRPC_CALL_LOG_BATCH(GPR_INFO, call, ops, nops, notify_tag);
 
@@ -1660,7 +1719,7 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
           goto done_with_error;
         }
         /* process compression level */
-        memset(&compression_md, 0, sizeof(compression_md));
+        memset(&call->compression_md, 0, sizeof(call->compression_md));
         size_t additional_metadata_count = 0;
         grpc_compression_level effective_compression_level =
             GRPC_COMPRESS_LEVEL_NONE;
@@ -1698,9 +1757,9 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
             const grpc_stream_compression_algorithm calgo =
                 stream_compression_algorithm_for_level_locked(
                     call, effective_stream_compression_level);
-            compression_md.key =
+            call->compression_md.key =
                 GRPC_MDSTR_GRPC_INTERNAL_STREAM_ENCODING_REQUEST;
-            compression_md.value =
+            call->compression_md.value =
                 grpc_stream_compression_algorithm_slice(calgo);
           } else {
             const grpc_compression_algorithm calgo =
@@ -1708,8 +1767,10 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
                     call, effective_compression_level);
             /* the following will be picked up by the compress filter and used
              * as the call's compression algorithm. */
-            compression_md.key = GRPC_MDSTR_GRPC_INTERNAL_ENCODING_REQUEST;
-            compression_md.value = grpc_compression_algorithm_slice(calgo);
+            call->compression_md.key =
+                GRPC_MDSTR_GRPC_INTERNAL_ENCODING_REQUEST;
+            call->compression_md.value =
+                grpc_compression_algorithm_slice(calgo);
             additional_metadata_count++;
           }
         }
@@ -1724,7 +1785,7 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
         if (!prepare_application_metadata(
                 exec_ctx, call, (int)op->data.send_initial_metadata.count,
                 op->data.send_initial_metadata.metadata, 0, call->is_client,
-                &compression_md, (int)additional_metadata_count)) {
+                &call->compression_md, (int)additional_metadata_count)) {
           error = GRPC_CALL_ERROR_INVALID_METADATA;
           goto done_with_error;
         }
@@ -1736,6 +1797,10 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
             &call->metadata_batch[0 /* is_receiving */][0 /* is_trailing */];
         stream_op_payload->send_initial_metadata.send_initial_metadata_flags =
             op->flags;
+        if (call->is_client) {
+          stream_op_payload->send_initial_metadata.peer_string =
+              &call->peer_string;
+        }
         break;
       case GRPC_OP_SEND_MESSAGE:
         if (!are_write_flags_valid(op->flags)) {
@@ -1868,6 +1933,10 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
             &call->metadata_batch[1 /* is_receiving */][0 /* is_trailing */];
         stream_op_payload->recv_initial_metadata.recv_initial_metadata_ready =
             &call->receiving_initial_metadata_ready;
+        if (!call->is_client) {
+          stream_op_payload->recv_initial_metadata.peer_string =
+              &call->peer_string;
+        }
         num_completion_callbacks_needed++;
         break;
       case GRPC_OP_RECV_MESSAGE:
@@ -1884,8 +1953,9 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
         stream_op->recv_message = true;
         call->receiving_buffer = op->data.recv_message.recv_message;
         stream_op_payload->recv_message.recv_message = &call->receiving_stream;
-        GRPC_CLOSURE_INIT(&call->receiving_stream_ready, receiving_stream_ready,
-                          bctl, grpc_schedule_on_exec_ctx);
+        GRPC_CLOSURE_INIT(&call->receiving_stream_ready,
+                          receiving_stream_ready_in_call_combiner, bctl,
+                          grpc_schedule_on_exec_ctx);
         stream_op_payload->recv_message.recv_message_ready =
             &call->receiving_stream_ready;
         num_completion_callbacks_needed++;
@@ -1955,7 +2025,7 @@ static grpc_call_error call_start_batch(grpc_exec_ctx *exec_ctx,
   stream_op->on_complete = &bctl->finish_batch;
   gpr_atm_rel_store(&call->any_ops_sent_atm, 1);
 
-  execute_op(exec_ctx, call, stream_op);
+  execute_batch(exec_ctx, call, stream_op, &bctl->start_batch);
 
 done:
   GPR_TIMER_END("grpc_call_start_batch", 0);
diff --git a/src/core/lib/surface/call.h b/src/core/lib/surface/call.h
index 185bfccb77..d537637cbb 100644
--- a/src/core/lib/surface/call.h
+++ b/src/core/lib/surface/call.h
@@ -19,6 +19,10 @@
 #ifndef GRPC_CORE_LIB_SURFACE_CALL_H
 #define GRPC_CORE_LIB_SURFACE_CALL_H
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 #include "src/core/lib/channel/channel_stack.h"
 #include "src/core/lib/channel/context.h"
 #include "src/core/lib/surface/api_trace.h"
@@ -26,10 +30,6 @@
 #include <grpc/grpc.h>
 #include <grpc/impl/codegen/compression_types.h>
 
-#ifdef __cplusplus
-extern "C" {
-#endif
-
 typedef void (*grpc_ioreq_completion_func)(grpc_exec_ctx *exec_ctx,
                                            grpc_call *call, int success,
                                            void *user_data);
@@ -89,7 +89,7 @@ grpc_call_error grpc_call_start_batch_and_execute(grpc_exec_ctx *exec_ctx,
 /* Given the top call_element, get the call object. */
 grpc_call *grpc_call_from_top_element(grpc_call_element *surface_element);
 
-void grpc_call_log_batch(char *file, int line, gpr_log_severity severity,
+void grpc_call_log_batch(const char *file, int line, gpr_log_severity severity,
                          grpc_call *call, const grpc_op *ops, size_t nops,
                          void *tag);
 
diff --git a/src/core/lib/surface/call_log_batch.c b/src/core/lib/surface/call_log_batch.c
index 4443aba58a..4a1c265817 100644
--- a/src/core/lib/surface/call_log_batch.c
+++ b/src/core/lib/surface/call_log_batch.c
@@ -103,7 +103,7 @@ char *grpc_op_string(const grpc_op *op) {
   return out;
 }
 
-void grpc_call_log_batch(char *file, int line, gpr_log_severity severity,
+void grpc_call_log_batch(const char *file, int line, gpr_log_severity severity,
                          grpc_call *call, const grpc_op *ops, size_t nops,
                          void *tag) {
   char *tmp;
diff --git a/src/core/lib/surface/init.c b/src/core/lib/surface/init.c
index 898476daee..280315036f 100644
--- a/src/core/lib/surface/init.c
+++ b/src/core/lib/surface/init.c
@@ -31,6 +31,7 @@
 #include "src/core/lib/debug/stats.h"
 #include "src/core/lib/debug/trace.h"
 #include "src/core/lib/http/parser.h"
+#include "src/core/lib/iomgr/call_combiner.h"
 #include "src/core/lib/iomgr/combiner.h"
 #include "src/core/lib/iomgr/executor.h"
 #include "src/core/lib/iomgr/iomgr.h"
@@ -129,6 +130,7 @@ void grpc_init(void) {
     grpc_register_tracer(&grpc_trace_channel_stack_builder);
     grpc_register_tracer(&grpc_http1_trace);
     grpc_register_tracer(&grpc_cq_pluck_trace);  // default on
+    grpc_register_tracer(&grpc_call_combiner_trace);
     grpc_register_tracer(&grpc_combiner_trace);
     grpc_register_tracer(&grpc_server_channel_trace);
     grpc_register_tracer(&grpc_bdp_estimator_trace);
diff --git a/src/core/lib/surface/lame_client.cc b/src/core/lib/surface/lame_client.cc
index a0791080a9..6286f9159d 100644
--- a/src/core/lib/surface/lame_client.cc
+++ b/src/core/lib/surface/lame_client.cc
@@ -40,6 +40,7 @@ namespace grpc_core {
 namespace {
 
 struct CallData {
+  grpc_call_combiner *call_combiner;
   grpc_linked_mdelem status;
   grpc_linked_mdelem details;
   grpc_core::atomic<bool> filled_metadata;
@@ -52,14 +53,14 @@ struct ChannelData {
 
 static void fill_metadata(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
                           grpc_metadata_batch *mdb) {
-  CallData *calld = static_cast<CallData *>(elem->call_data);
+  CallData *calld = reinterpret_cast<CallData *>(elem->call_data);
   bool expected = false;
   if (!calld->filled_metadata.compare_exchange_strong(
           expected, true, grpc_core::memory_order_relaxed,
           grpc_core::memory_order_relaxed)) {
     return;
   }
-  ChannelData *chand = static_cast<ChannelData *>(elem->channel_data);
+  ChannelData *chand = reinterpret_cast<ChannelData *>(elem->channel_data);
   char tmp[GPR_LTOA_MIN_BUFSIZE];
   gpr_ltoa(chand->error_code, tmp);
   calld->status.md = grpc_mdelem_from_slices(
@@ -79,6 +80,7 @@ static void fill_metadata(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
 static void lame_start_transport_stream_op_batch(
     grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     grpc_transport_stream_op_batch *op) {
+  CallData *calld = reinterpret_cast<CallData *>(elem->call_data);
   if (op->recv_initial_metadata) {
     fill_metadata(exec_ctx, elem,
                   op->payload->recv_initial_metadata.recv_initial_metadata);
@@ -87,12 +89,8 @@ static void lame_start_transport_stream_op_batch(
                   op->payload->recv_trailing_metadata.recv_trailing_metadata);
   }
   grpc_transport_stream_op_batch_finish_with_failure(
-      exec_ctx, op,
-      GRPC_ERROR_CREATE_FROM_STATIC_STRING("lame client channel"));
-}
-
-static char *lame_get_peer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem) {
-  return NULL;
+      exec_ctx, op, GRPC_ERROR_CREATE_FROM_STATIC_STRING("lame client channel"),
+      calld->call_combiner);
 }
 
 static void lame_get_channel_info(grpc_exec_ctx *exec_ctx,
@@ -122,6 +120,8 @@ static void lame_start_transport_op(grpc_exec_ctx *exec_ctx,
 static grpc_error *init_call_elem(grpc_exec_ctx *exec_ctx,
                                   grpc_call_element *elem,
                                   const grpc_call_element_args *args) {
+  CallData *calld = reinterpret_cast<CallData *>(elem->call_data);
+  calld->call_combiner = args->call_combiner;
   return GRPC_ERROR_NONE;
 }
 
@@ -156,7 +156,6 @@ extern "C" const grpc_channel_filter grpc_lame_filter = {
     sizeof(grpc_core::ChannelData),
     grpc_core::init_channel_elem,
     grpc_core::destroy_channel_elem,
-    grpc_core::lame_get_peer,
     grpc_core::lame_get_channel_info,
     "lame-client",
 };
@@ -176,7 +175,7 @@ grpc_channel *grpc_lame_client_channel_create(const char *target,
       "error_message=%s)",
       3, (target, (int)error_code, error_message));
   GPR_ASSERT(elem->filter == &grpc_lame_filter);
-  auto chand = static_cast<grpc_core::ChannelData *>(elem->channel_data);
+  auto chand = reinterpret_cast<grpc_core::ChannelData *>(elem->channel_data);
   chand->error_code = error_code;
   chand->error_message = error_message;
   grpc_exec_ctx_finish(&exec_ctx);
diff --git a/src/core/lib/surface/server.c b/src/core/lib/surface/server.c
index 786d754b49..df21467545 100644
--- a/src/core/lib/surface/server.c
+++ b/src/core/lib/surface/server.c
@@ -792,7 +792,6 @@ static void server_mutate_op(grpc_call_element *elem,
 static void server_start_transport_stream_op_batch(
     grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     grpc_transport_stream_op_batch *op) {
-  GRPC_CALL_LOG_OP(GPR_INFO, elem, op);
   server_mutate_op(elem, op);
   grpc_call_next_op(exec_ctx, elem, op);
 }
@@ -965,7 +964,6 @@ const grpc_channel_filter grpc_server_top_filter = {
     sizeof(channel_data),
     init_channel_elem,
     destroy_channel_elem,
-    grpc_call_next_get_peer,
     grpc_channel_next_get_info,
     "server",
 };
diff --git a/src/core/lib/transport/metadata_batch.c b/src/core/lib/transport/metadata_batch.c
index 8f24b8527c..a077052561 100644
--- a/src/core/lib/transport/metadata_batch.c
+++ b/src/core/lib/transport/metadata_batch.c
@@ -105,6 +105,7 @@ static grpc_error *maybe_link_callout(grpc_metadata_batch *batch,
     return GRPC_ERROR_NONE;
   }
   if (batch->idx.array[idx] == NULL) {
+    if (grpc_static_callout_is_default[idx]) ++batch->list.default_count;
     batch->idx.array[idx] = storage;
     return GRPC_ERROR_NONE;
   }
@@ -120,6 +121,7 @@ static void maybe_unlink_callout(grpc_metadata_batch *batch,
   if (idx == GRPC_BATCH_CALLOUTS_COUNT) {
     return;
   }
+  if (grpc_static_callout_is_default[idx]) --batch->list.default_count;
   GPR_ASSERT(batch->idx.array[idx] != NULL);
   batch->idx.array[idx] = NULL;
 }
diff --git a/src/core/lib/transport/metadata_batch.h b/src/core/lib/transport/metadata_batch.h
index 1b11a3e252..57d298c75c 100644
--- a/src/core/lib/transport/metadata_batch.h
+++ b/src/core/lib/transport/metadata_batch.h
@@ -41,6 +41,7 @@ typedef struct grpc_linked_mdelem {
 
 typedef struct grpc_mdelem_list {
   size_t count;
+  size_t default_count;  // Number of default keys.
   grpc_linked_mdelem *head;
   grpc_linked_mdelem *tail;
 } grpc_mdelem_list;
diff --git a/src/core/lib/transport/static_metadata.c b/src/core/lib/transport/static_metadata.c
index 28f05d5c44..b20d94aeac 100644
--- a/src/core/lib/transport/static_metadata.c
+++ b/src/core/lib/transport/static_metadata.c
@@ -823,6 +823,31 @@ grpc_mdelem_data grpc_static_mdelem_table[GRPC_STATIC_MDELEM_COUNT] = {
      {.refcount = &grpc_static_metadata_refcounts[97],
       .data.refcounted = {g_bytes + 1040, 13}}},
 };
+bool grpc_static_callout_is_default[GRPC_BATCH_CALLOUTS_COUNT] = {
+    true,  // :path
+    true,  // :method
+    true,  // :status
+    true,  // :authority
+    true,  // :scheme
+    true,  // te
+    true,  // grpc-message
+    true,  // grpc-status
+    true,  // grpc-payload-bin
+    true,  // grpc-encoding
+    true,  // grpc-accept-encoding
+    true,  // grpc-server-stats-bin
+    true,  // grpc-tags-bin
+    true,  // grpc-trace-bin
+    true,  // content-type
+    true,  // content-encoding
+    true,  // accept-encoding
+    true,  // grpc-internal-encoding-request
+    true,  // grpc-internal-stream-encoding-request
+    true,  // user-agent
+    true,  // host
+    true,  // lb-token
+};
+
 const uint8_t grpc_static_accept_encoding_metadata[8] = {0,  76, 77, 78,
                                                          79, 80, 81, 82};
 
diff --git a/src/core/lib/transport/static_metadata.h b/src/core/lib/transport/static_metadata.h
index 93ab90dff8..f03a9d23b1 100644
--- a/src/core/lib/transport/static_metadata.h
+++ b/src/core/lib/transport/static_metadata.h
@@ -571,6 +571,8 @@ typedef union {
              GRPC_BATCH_CALLOUTS_COUNT)                 \
        : GRPC_BATCH_CALLOUTS_COUNT)
 
+extern bool grpc_static_callout_is_default[GRPC_BATCH_CALLOUTS_COUNT];
+
 extern const uint8_t grpc_static_accept_encoding_metadata[8];
 #define GRPC_MDELEM_ACCEPT_ENCODING_FOR_ALGORITHMS(algs)                       \
   (GRPC_MAKE_MDELEM(                                                           \
diff --git a/src/core/lib/transport/transport.c b/src/core/lib/transport/transport.c
index 6c61f4b8d9..650b0559aa 100644
--- a/src/core/lib/transport/transport.c
+++ b/src/core/lib/transport/transport.c
@@ -197,11 +197,6 @@ void grpc_transport_destroy_stream(grpc_exec_ctx *exec_ctx,
                                     then_schedule_closure);
 }
 
-char *grpc_transport_get_peer(grpc_exec_ctx *exec_ctx,
-                              grpc_transport *transport) {
-  return transport->vtable->get_peer(exec_ctx, transport);
-}
-
 grpc_endpoint *grpc_transport_get_endpoint(grpc_exec_ctx *exec_ctx,
                                            grpc_transport *transport) {
   return transport->vtable->get_endpoint(exec_ctx, transport);
@@ -214,24 +209,24 @@ grpc_endpoint *grpc_transport_get_endpoint(grpc_exec_ctx *exec_ctx,
 // is a function that must always unref cancel_error
 // though it lives in lib, it handles transport stream ops sure
 // it's grpc_transport_stream_op_batch_finish_with_failure
-
 void grpc_transport_stream_op_batch_finish_with_failure(
     grpc_exec_ctx *exec_ctx, grpc_transport_stream_op_batch *batch,
-    grpc_error *error) {
+    grpc_error *error, grpc_call_combiner *call_combiner) {
   if (batch->send_message) {
     grpc_byte_stream_destroy(exec_ctx,
                              batch->payload->send_message.send_message);
   }
   if (batch->recv_message) {
-    GRPC_CLOSURE_SCHED(exec_ctx,
-                       batch->payload->recv_message.recv_message_ready,
-                       GRPC_ERROR_REF(error));
+    GRPC_CALL_COMBINER_START(exec_ctx, call_combiner,
+                             batch->payload->recv_message.recv_message_ready,
+                             GRPC_ERROR_REF(error),
+                             "failing recv_message_ready");
   }
   if (batch->recv_initial_metadata) {
-    GRPC_CLOSURE_SCHED(
-        exec_ctx,
+    GRPC_CALL_COMBINER_START(
+        exec_ctx, call_combiner,
         batch->payload->recv_initial_metadata.recv_initial_metadata_ready,
-        GRPC_ERROR_REF(error));
+        GRPC_ERROR_REF(error), "failing recv_initial_metadata_ready");
   }
   GRPC_CLOSURE_SCHED(exec_ctx, batch->on_complete, error);
   if (batch->cancel_stream) {
diff --git a/src/core/lib/transport/transport.h b/src/core/lib/transport/transport.h
index 099138ea14..fbf5dcb8b5 100644
--- a/src/core/lib/transport/transport.h
+++ b/src/core/lib/transport/transport.h
@@ -22,6 +22,7 @@
 #include <stddef.h>
 
 #include "src/core/lib/channel/context.h"
+#include "src/core/lib/iomgr/call_combiner.h"
 #include "src/core/lib/iomgr/endpoint.h"
 #include "src/core/lib/iomgr/polling_entity.h"
 #include "src/core/lib/iomgr/pollset.h"
@@ -152,6 +153,9 @@ struct grpc_transport_stream_op_batch_payload {
     /** Iff send_initial_metadata != NULL, flags associated with
         send_initial_metadata: a bitfield of GRPC_INITIAL_METADATA_xxx */
     uint32_t send_initial_metadata_flags;
+    // If non-NULL, will be set by the transport to the peer string
+    // (a char*, which the caller takes ownership of).
+    gpr_atm *peer_string;
   } send_initial_metadata;
 
   struct {
@@ -176,6 +180,9 @@ struct grpc_transport_stream_op_batch_payload {
     // immediately available.  This may be a signal that we received a
     // Trailers-Only response.
     bool *trailing_metadata_available;
+    // If non-NULL, will be set by the transport to the peer string
+    // (a char*, which the caller takes ownership of).
+    gpr_atm *peer_string;
   } recv_initial_metadata;
 
   struct {
@@ -293,7 +300,7 @@ void grpc_transport_destroy_stream(grpc_exec_ctx *exec_ctx,
 
 void grpc_transport_stream_op_batch_finish_with_failure(
     grpc_exec_ctx *exec_ctx, grpc_transport_stream_op_batch *op,
-    grpc_error *error);
+    grpc_error *error, grpc_call_combiner *call_combiner);
 
 char *grpc_transport_stream_op_batch_string(grpc_transport_stream_op_batch *op);
 char *grpc_transport_op_string(grpc_transport_op *op);
@@ -332,10 +339,6 @@ void grpc_transport_close(grpc_transport *transport);
 /* Destroy the transport */
 void grpc_transport_destroy(grpc_exec_ctx *exec_ctx, grpc_transport *transport);
 
-/* Get the transports peer */
-char *grpc_transport_get_peer(grpc_exec_ctx *exec_ctx,
-                              grpc_transport *transport);
-
 /* Get the endpoint used by \a transport */
 grpc_endpoint *grpc_transport_get_endpoint(grpc_exec_ctx *exec_ctx,
                                            grpc_transport *transport);
diff --git a/src/core/lib/transport/transport_impl.h b/src/core/lib/transport/transport_impl.h
index fc772c6dd1..bbae69c223 100644
--- a/src/core/lib/transport/transport_impl.h
+++ b/src/core/lib/transport/transport_impl.h
@@ -59,9 +59,6 @@ typedef struct grpc_transport_vtable {
   /* implementation of grpc_transport_destroy */
   void (*destroy)(grpc_exec_ctx *exec_ctx, grpc_transport *self);
 
-  /* implementation of grpc_transport_get_peer */
-  char *(*get_peer)(grpc_exec_ctx *exec_ctx, grpc_transport *self);
-
   /* implementation of grpc_transport_get_endpoint */
   grpc_endpoint *(*get_endpoint)(grpc_exec_ctx *exec_ctx, grpc_transport *self);
 } grpc_transport_vtable;
diff --git a/src/core/lib/transport/transport_op_string.c b/src/core/lib/transport/transport_op_string.c
index 7b18229ba6..409a6c4103 100644
--- a/src/core/lib/transport/transport_op_string.c
+++ b/src/core/lib/transport/transport_op_string.c
@@ -112,6 +112,13 @@ char *grpc_transport_stream_op_batch_string(
     gpr_strvec_add(&b, tmp);
   }
 
+  if (op->collect_stats) {
+    gpr_strvec_add(&b, gpr_strdup(" "));
+    gpr_asprintf(&tmp, "COLLECT_STATS:%p",
+                 op->payload->collect_stats.collect_stats);
+    gpr_strvec_add(&b, tmp);
+  }
+
   out = gpr_strvec_flatten(&b, NULL);
   gpr_strvec_destroy(&b);
 
diff --git a/src/core/tsi/test_creds/BUILD b/src/core/tsi/test_creds/BUILD
index 4b0786d7b8..732f6d91b2 100644
--- a/src/core/tsi/test_creds/BUILD
+++ b/src/core/tsi/test_creds/BUILD
@@ -15,7 +15,15 @@
 licenses(["notice"])  # Apache v2
 
 exports_files([
-    "ca.pem",
-    "server1.key",
-    "server1.pem",
+        "ca.pem",
+        "server1.key",
+        "server1.pem",
+        "server0.key",
+        "server0.pem",
+        "client.key",
+        "client.pem",
+        "badserver.key",
+        "badserver.pem",
+        "badclient.key",
+        "badclient.pem",
 ])
diff --git a/src/cpp/client/channel_cc.cc b/src/cpp/client/channel_cc.cc
index f2d9bb07c9..19a25c838f 100644
--- a/src/cpp/client/channel_cc.cc
+++ b/src/cpp/client/channel_cc.cc
@@ -18,7 +18,10 @@
 
 #include <grpc++/channel.h>
 
+#include <chrono>
+#include <condition_variable>
 #include <memory>
+#include <mutex>
 
 #include <grpc++/client_context.h>
 #include <grpc++/completion_queue.h>
@@ -35,17 +38,197 @@
 #include <grpc/slice.h>
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
+#include <grpc/support/sync.h>
+#include <grpc/support/thd.h>
+#include <grpc/support/time.h>
+#include <grpc/support/useful.h>
 #include "src/core/lib/profiling/timers.h"
+#include "src/core/lib/support/env.h"
+#include "src/core/lib/support/string.h"
 
 namespace grpc {
 
+namespace {
+int kConnectivityCheckIntervalMsec = 500;
+void WatchStateChange(void* arg);
+
+class TagSaver final : public CompletionQueueTag {
+ public:
+  explicit TagSaver(void* tag) : tag_(tag) {}
+  ~TagSaver() override {}
+  bool FinalizeResult(void** tag, bool* status) override {
+    *tag = tag_;
+    delete this;
+    return true;
+  }
+
+ private:
+  void* tag_;
+};
+
+// Constantly watches channel connectivity status to reconnect a transiently
+// disconnected channel. This is a temporary work-around before we have retry
+// support.
+class ChannelConnectivityWatcher : private GrpcLibraryCodegen {
+ public:
+  static void StartWatching(grpc_channel* channel) {
+    if (!IsDisabled()) {
+      std::unique_lock<std::mutex> lock(g_watcher_mu_);
+      if (g_watcher_ == nullptr) {
+        g_watcher_ = new ChannelConnectivityWatcher();
+      }
+      g_watcher_->StartWatchingLocked(channel);
+    }
+  }
+
+  static void StopWatching() {
+    if (!IsDisabled()) {
+      std::unique_lock<std::mutex> lock(g_watcher_mu_);
+      if (g_watcher_->StopWatchingLocked()) {
+        delete g_watcher_;
+        g_watcher_ = nullptr;
+      }
+    }
+  }
+
+ private:
+  ChannelConnectivityWatcher() : channel_count_(0), shutdown_(false) {
+    gpr_ref_init(&ref_, 0);
+    gpr_thd_options options = gpr_thd_options_default();
+    gpr_thd_options_set_joinable(&options);
+    gpr_thd_new(&thd_id_, &WatchStateChange, this, &options);
+  }
+
+  static bool IsDisabled() {
+    char* env = gpr_getenv("GRPC_DISABLE_CHANNEL_CONNECTIVITY_WATCHER");
+    bool disabled = gpr_is_true(env);
+    gpr_free(env);
+    return disabled;
+  }
+
+  void WatchStateChangeImpl() {
+    bool ok = false;
+    void* tag = NULL;
+    CompletionQueue::NextStatus status = CompletionQueue::GOT_EVENT;
+    while (true) {
+      {
+        std::unique_lock<std::mutex> lock(shutdown_mu_);
+        if (shutdown_) {
+          // Drain cq_ if the watcher is shutting down
+          status = cq_.AsyncNext(&tag, &ok, gpr_inf_future(GPR_CLOCK_REALTIME));
+        } else {
+          status = cq_.AsyncNext(&tag, &ok, gpr_inf_past(GPR_CLOCK_REALTIME));
+          // Make sure we've seen 2 TIMEOUTs before going to sleep
+          if (status == CompletionQueue::TIMEOUT) {
+            status = cq_.AsyncNext(&tag, &ok, gpr_inf_past(GPR_CLOCK_REALTIME));
+            if (status == CompletionQueue::TIMEOUT) {
+              shutdown_cv_.wait_for(lock, std::chrono::milliseconds(
+                                              kConnectivityCheckIntervalMsec));
+              continue;
+            }
+          }
+        }
+      }
+      ChannelState* channel_state = static_cast<ChannelState*>(tag);
+      channel_state->state =
+          grpc_channel_check_connectivity_state(channel_state->channel, false);
+      if (channel_state->state == GRPC_CHANNEL_SHUTDOWN) {
+        void* shutdown_tag = NULL;
+        channel_state->shutdown_cq.Next(&shutdown_tag, &ok);
+        delete channel_state;
+        if (gpr_unref(&ref_)) {
+          break;
+        }
+      } else {
+        TagSaver* tag_saver = new TagSaver(channel_state);
+        grpc_channel_watch_connectivity_state(
+            channel_state->channel, channel_state->state,
+            gpr_inf_future(GPR_CLOCK_REALTIME), cq_.cq(), tag_saver);
+      }
+    }
+  }
+
+  void StartWatchingLocked(grpc_channel* channel) {
+    if (thd_id_ != 0) {
+      gpr_ref(&ref_);
+      ++channel_count_;
+      ChannelState* channel_state = new ChannelState(channel);
+      // The first grpc_channel_watch_connectivity_state() is not used to
+      // monitor the channel state change, but to hold a reference of the
+      // c channel. So that WatchStateChangeImpl() can observe state ==
+      // GRPC_CHANNEL_SHUTDOWN before the channel gets destroyed.
+      grpc_channel_watch_connectivity_state(
+          channel_state->channel, channel_state->state,
+          gpr_inf_future(GPR_CLOCK_REALTIME), channel_state->shutdown_cq.cq(),
+          new TagSaver(nullptr));
+      grpc_channel_watch_connectivity_state(
+          channel_state->channel, channel_state->state,
+          gpr_inf_future(GPR_CLOCK_REALTIME), cq_.cq(),
+          new TagSaver(channel_state));
+    }
+  }
+
+  bool StopWatchingLocked() {
+    if (--channel_count_ == 0) {
+      {
+        std::unique_lock<std::mutex> lock(shutdown_mu_);
+        shutdown_ = true;
+        shutdown_cv_.notify_one();
+      }
+      gpr_thd_join(thd_id_);
+      return true;
+    }
+    return false;
+  }
+
+  friend void WatchStateChange(void* arg);
+  struct ChannelState {
+    explicit ChannelState(grpc_channel* channel)
+        : channel(channel), state(GRPC_CHANNEL_IDLE){};
+    grpc_channel* channel;
+    grpc_connectivity_state state;
+    CompletionQueue shutdown_cq;
+  };
+  gpr_thd_id thd_id_;
+  CompletionQueue cq_;
+  gpr_refcount ref_;
+  int channel_count_;
+
+  std::mutex shutdown_mu_;
+  std::condition_variable shutdown_cv_;  // protected by shutdown_mu_
+  bool shutdown_;                        // protected by shutdown_mu_
+
+  static std::mutex g_watcher_mu_;
+  static ChannelConnectivityWatcher* g_watcher_;  // protected by g_watcher_mu_
+};
+
+std::mutex ChannelConnectivityWatcher::g_watcher_mu_;
+ChannelConnectivityWatcher* ChannelConnectivityWatcher::g_watcher_ = nullptr;
+
+void WatchStateChange(void* arg) {
+  ChannelConnectivityWatcher* watcher =
+      static_cast<ChannelConnectivityWatcher*>(arg);
+  watcher->WatchStateChangeImpl();
+}
+}  // namespace
+
 static internal::GrpcLibraryInitializer g_gli_initializer;
 Channel::Channel(const grpc::string& host, grpc_channel* channel)
     : host_(host), c_channel_(channel) {
   g_gli_initializer.summon();
+  if (grpc_channel_support_connectivity_watcher(channel)) {
+    ChannelConnectivityWatcher::StartWatching(channel);
+  }
 }
 
-Channel::~Channel() { grpc_channel_destroy(c_channel_); }
+Channel::~Channel() {
+  const bool stop_watching =
+      grpc_channel_support_connectivity_watcher(c_channel_);
+  grpc_channel_destroy(c_channel_);
+  if (stop_watching) {
+    ChannelConnectivityWatcher::StopWatching();
+  }
+}
 
 namespace {
 
@@ -130,23 +313,6 @@ grpc_connectivity_state Channel::GetState(bool try_to_connect) {
   return grpc_channel_check_connectivity_state(c_channel_, try_to_connect);
 }
 
-namespace {
-class TagSaver final : public CompletionQueueTag {
- public:
-  explicit TagSaver(void* tag) : tag_(tag) {}
-  ~TagSaver() override {}
-  bool FinalizeResult(void** tag, bool* status) override {
-    *tag = tag_;
-    delete this;
-    return true;
-  }
-
- private:
-  void* tag_;
-};
-
-}  // namespace
-
 void Channel::NotifyOnStateChangeImpl(grpc_connectivity_state last_observed,
                                       gpr_timespec deadline,
                                       CompletionQueue* cq, void* tag) {
diff --git a/src/cpp/common/channel_filter.cc b/src/cpp/common/channel_filter.cc
index 448d9fbcf2..ea44cff832 100644
--- a/src/cpp/common/channel_filter.cc
+++ b/src/cpp/common/channel_filter.cc
@@ -68,10 +68,6 @@ void CallData::SetPollsetOrPollsetSet(grpc_exec_ctx *exec_ctx,
   grpc_call_stack_ignore_set_pollset_or_pollset_set(exec_ctx, elem, pollent);
 }
 
-char *CallData::GetPeer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem) {
-  return grpc_call_next_get_peer(exec_ctx, elem);
-}
-
 // internal code used by RegisterChannelFilter()
 
 namespace internal {
diff --git a/src/cpp/common/channel_filter.h b/src/cpp/common/channel_filter.h
index c3d187d7e1..c1aeb3f724 100644
--- a/src/cpp/common/channel_filter.h
+++ b/src/cpp/common/channel_filter.h
@@ -268,9 +268,6 @@ class CallData {
   virtual void SetPollsetOrPollsetSet(grpc_exec_ctx *exec_ctx,
                                       grpc_call_element *elem,
                                       grpc_polling_entity *pollent);
-
-  /// Gets the peer name.
-  virtual char *GetPeer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem);
 };
 
 namespace internal {
@@ -349,11 +346,6 @@ class ChannelFilter final {
     CallDataType *call_data = reinterpret_cast<CallDataType *>(elem->call_data);
     call_data->SetPollsetOrPollsetSet(exec_ctx, elem, pollent);
   }
-
-  static char *GetPeer(grpc_exec_ctx *exec_ctx, grpc_call_element *elem) {
-    CallDataType *call_data = reinterpret_cast<CallDataType *>(elem->call_data);
-    return call_data->GetPeer(exec_ctx, elem);
-  }
 };
 
 struct FilterRecord {
@@ -396,8 +388,7 @@ void RegisterChannelFilter(
        FilterType::call_data_size, FilterType::InitCallElement,
        FilterType::SetPollsetOrPollsetSet, FilterType::DestroyCallElement,
        FilterType::channel_data_size, FilterType::InitChannelElement,
-       FilterType::DestroyChannelElement, FilterType::GetPeer,
-       FilterType::GetChannelInfo, name}};
+       FilterType::DestroyChannelElement, FilterType::GetChannelInfo, name}};
   internal::channel_filters->push_back(filter_record);
 }
 
diff --git a/src/cpp/server/server_cc.cc b/src/cpp/server/server_cc.cc
index 2483300cb1..6bd3ecda32 100644
--- a/src/cpp/server/server_cc.cc
+++ b/src/cpp/server/server_cc.cc
@@ -17,6 +17,7 @@
 
 #include <grpc++/server.h>
 
+#include <cstdlib>
 #include <sstream>
 #include <utility>
 
@@ -38,6 +39,7 @@
 
 #include "src/core/ext/transport/inproc/inproc_transport.h"
 #include "src/core/lib/profiling/timers.h"
+#include "src/core/lib/surface/call.h"
 #include "src/cpp/client/create_channel_internal.h"
 #include "src/cpp/server/health/default_health_check_service.h"
 #include "src/cpp/thread_manager/thread_manager.h"
@@ -607,7 +609,12 @@ void Server::PerformOpsOnCall(CallOpSetInterface* ops, Call* call) {
   grpc_op cops[MAX_OPS];
   ops->FillOps(call->call(), cops, &nops);
   auto result = grpc_call_start_batch(call->call(), cops, nops, ops, nullptr);
-  GPR_ASSERT(GRPC_CALL_OK == result);
+  if (result != GRPC_CALL_OK) {
+    gpr_log(GPR_ERROR, "Fatal: grpc_call_start_batch returned %d", result);
+    grpc_call_log_batch(__FILE__, __LINE__, GPR_LOG_SEVERITY_ERROR,
+                        call->call(), cops, nops, ops);
+    abort();
+  }
 }
 
 ServerInterface::BaseAsyncRequest::BaseAsyncRequest(
diff --git a/src/objective-c/!ProtoCompiler-gRPCPlugin.podspec b/src/objective-c/!ProtoCompiler-gRPCPlugin.podspec
index 3a282b0526..7d073c9a84 100644
--- a/src/objective-c/!ProtoCompiler-gRPCPlugin.podspec
+++ b/src/objective-c/!ProtoCompiler-gRPCPlugin.podspec
@@ -101,7 +101,7 @@ Pod::Spec.new do |s|
   s.preserve_paths = plugin
 
   # Restrict the protoc version to the one supported by this plugin.
-  s.dependency '!ProtoCompiler', '3.3.0'
+  s.dependency '!ProtoCompiler', '3.4.0'
   # For the Protobuf dependency not to complain:
   s.ios.deployment_target = '7.0'
   s.osx.deployment_target = '10.9'
diff --git a/src/objective-c/!ProtoCompiler.podspec b/src/objective-c/!ProtoCompiler.podspec
index c3f95f9f42..25c437911f 100644
--- a/src/objective-c/!ProtoCompiler.podspec
+++ b/src/objective-c/!ProtoCompiler.podspec
@@ -36,7 +36,7 @@ Pod::Spec.new do |s|
   # exclamation mark ensures that other "regular" pods will be able to find it as it'll be installed
   # before them.
   s.name     = '!ProtoCompiler'
-  v = '3.3.0'
+  v = '3.4.0'
   s.version  = v
   s.summary  = 'The Protobuf Compiler (protoc) generates Objective-C files from .proto files'
   s.description = <<-DESC
diff --git a/src/proto/grpc/core/BUILD b/src/proto/grpc/core/BUILD
new file mode 100644
index 0000000000..46de9fae18
--- /dev/null
+++ b/src/proto/grpc/core/BUILD
@@ -0,0 +1,24 @@
+# Copyright 2017 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.
+
+licenses(["notice"])  # Apache v2
+
+load("//bazel:grpc_build_system.bzl", "grpc_proto_library", "grpc_package")
+
+grpc_package(name = "core", visibility = "public")
+
+grpc_proto_library(
+    name = "stats_proto",
+    srcs = ["stats.proto"],
+)
diff --git a/src/proto/grpc/testing/BUILD b/src/proto/grpc/testing/BUILD
index 07e08117f0..36d3782262 100644
--- a/src/proto/grpc/testing/BUILD
+++ b/src/proto/grpc/testing/BUILD
@@ -84,6 +84,9 @@ grpc_proto_library(
     name = "stats_proto",
     srcs = ["stats.proto"],
     has_services = False,
+    deps = [
+        "//src/proto/grpc/core:stats_proto",
+    ]
 )
 
 grpc_proto_library(
diff --git a/src/python/grpcio/grpc_core_dependencies.py b/src/python/grpcio/grpc_core_dependencies.py
index 1cbf345ab6..2071827b64 100644
--- a/src/python/grpcio/grpc_core_dependencies.py
+++ b/src/python/grpcio/grpc_core_dependencies.py
@@ -77,6 +77,7 @@ CORE_SOURCE_FILES = [
   'src/core/lib/http/format_request.c',
   'src/core/lib/http/httpcli.c',
   'src/core/lib/http/parser.c',
+  'src/core/lib/iomgr/call_combiner.c',
   'src/core/lib/iomgr/closure.c',
   'src/core/lib/iomgr/combiner.c',
   'src/core/lib/iomgr/endpoint.c',
@@ -85,8 +86,6 @@ CORE_SOURCE_FILES = [
   'src/core/lib/iomgr/endpoint_pair_windows.c',
   'src/core/lib/iomgr/error.c',
   'src/core/lib/iomgr/ev_epoll1_linux.c',
-  'src/core/lib/iomgr/ev_epoll_limited_pollers_linux.c',
-  'src/core/lib/iomgr/ev_epoll_thread_pool_linux.c',
   'src/core/lib/iomgr/ev_epollex_linux.c',
   'src/core/lib/iomgr/ev_epollsig_linux.c',
   'src/core/lib/iomgr/ev_poll_posix.c',
diff --git a/src/python/grpcio_testing/grpc_testing/__init__.py b/src/python/grpcio_testing/grpc_testing/__init__.py
index 14e25f09e2..917e11808e 100644
--- a/src/python/grpcio_testing/grpc_testing/__init__.py
+++ b/src/python/grpcio_testing/grpc_testing/__init__.py
@@ -293,6 +293,278 @@ class Channel(six.with_metaclass(abc.ABCMeta), grpc.Channel):
         raise NotImplementedError()
 
 
+class UnaryUnaryServerRpc(six.with_metaclass(abc.ABCMeta)):
+    """Fixture for a unary-unary RPC serviced by a system under test.
+
+    Enables users to "play client" for the RPC.
+    """
+
+    @abc.abstractmethod
+    def initial_metadata(self):
+        """Accesses the initial metadata emitted by the system under test.
+
+        This method blocks until the system under test has added initial
+        metadata to the RPC (or has provided one or more response messages or
+        has terminated the RPC, either of which will cause gRPC Python to
+        synthesize initial metadata for the RPC).
+
+        Returns:
+          The initial metadata for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def cancel(self):
+        """Cancels the RPC."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def termination(self):
+        """Blocks until the system under test has terminated the RPC.
+
+        Returns:
+          A (response, trailing_metadata, code, details) sequence with the RPC's
+            response, trailing metadata, code, and details.
+        """
+        raise NotImplementedError()
+
+
+class UnaryStreamServerRpc(six.with_metaclass(abc.ABCMeta)):
+    """Fixture for a unary-stream RPC serviced by a system under test.
+
+    Enables users to "play client" for the RPC.
+    """
+
+    @abc.abstractmethod
+    def initial_metadata(self):
+        """Accesses the initial metadata emitted by the system under test.
+
+        This method blocks until the system under test has added initial
+        metadata to the RPC (or has provided one or more response messages or
+        has terminated the RPC, either of which will cause gRPC Python to
+        synthesize initial metadata for the RPC).
+
+        Returns:
+          The initial metadata for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def take_response(self):
+        """Draws one of the responses added to the RPC by the system under test.
+
+        Successive calls to this method return responses in the same order in
+        which the system under test added them to the RPC.
+
+        Returns:
+          A response message added to the RPC by the system under test.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def cancel(self):
+        """Cancels the RPC."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def termination(self):
+        """Blocks until the system under test has terminated the RPC.
+
+        Returns:
+          A (trailing_metadata, code, details) sequence with the RPC's trailing
+            metadata, code, and details.
+        """
+        raise NotImplementedError()
+
+
+class StreamUnaryServerRpc(six.with_metaclass(abc.ABCMeta)):
+    """Fixture for a stream-unary RPC serviced by a system under test.
+
+    Enables users to "play client" for the RPC.
+    """
+
+    @abc.abstractmethod
+    def initial_metadata(self):
+        """Accesses the initial metadata emitted by the system under test.
+
+        This method blocks until the system under test has added initial
+        metadata to the RPC (or has provided one or more response messages or
+        has terminated the RPC, either of which will cause gRPC Python to
+        synthesize initial metadata for the RPC).
+
+        Returns:
+          The initial metadata for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def send_request(self, request):
+        """Sends a request to the system under test.
+
+        Args:
+          request: A request message for the RPC to be "sent" to the system
+            under test.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def requests_closed(self):
+        """Indicates the end of the RPC's request stream."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def cancel(self):
+        """Cancels the RPC."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def termination(self):
+        """Blocks until the system under test has terminated the RPC.
+
+        Returns:
+          A (response, trailing_metadata, code, details) sequence with the RPC's
+            response, trailing metadata, code, and details.
+        """
+        raise NotImplementedError()
+
+
+class StreamStreamServerRpc(six.with_metaclass(abc.ABCMeta)):
+    """Fixture for a stream-stream RPC serviced by a system under test.
+
+    Enables users to "play client" for the RPC.
+    """
+
+    @abc.abstractmethod
+    def initial_metadata(self):
+        """Accesses the initial metadata emitted by the system under test.
+
+        This method blocks until the system under test has added initial
+        metadata to the RPC (or has provided one or more response messages or
+        has terminated the RPC, either of which will cause gRPC Python to
+        synthesize initial metadata for the RPC).
+
+        Returns:
+          The initial metadata for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def send_request(self, request):
+        """Sends a request to the system under test.
+
+        Args:
+          request: A request message for the RPC to be "sent" to the system
+            under test.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def requests_closed(self):
+        """Indicates the end of the RPC's request stream."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def take_response(self):
+        """Draws one of the responses added to the RPC by the system under test.
+
+        Successive calls to this method return responses in the same order in
+        which the system under test added them to the RPC.
+
+        Returns:
+          A response message added to the RPC by the system under test.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def cancel(self):
+        """Cancels the RPC."""
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def termination(self):
+        """Blocks until the system under test has terminated the RPC.
+
+        Returns:
+          A (trailing_metadata, code, details) sequence with the RPC's trailing
+            metadata, code, and details.
+        """
+        raise NotImplementedError()
+
+
+class Server(six.with_metaclass(abc.ABCMeta)):
+    """A server with which to test a system that services RPCs."""
+
+    @abc.abstractmethod
+    def invoke_unary_unary(
+            self, method_descriptor, invocation_metadata, request, timeout):
+        """Invokes an RPC to be serviced by the system under test.
+
+        Args:
+          method_descriptor: A descriptor.MethodDescriptor describing a unary-unary
+            RPC method.
+          invocation_metadata: The RPC's invocation metadata.
+          request: The RPC's request.
+          timeout: A duration of time in seconds for the RPC or None to
+            indicate that the RPC has no time limit.
+
+        Returns:
+          A UnaryUnaryServerRpc with which to "play client" for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def invoke_unary_stream(
+            self, method_descriptor, invocation_metadata, request, timeout):
+        """Invokes an RPC to be serviced by the system under test.
+
+        Args:
+          method_descriptor: A descriptor.MethodDescriptor describing a unary-stream
+            RPC method.
+          invocation_metadata: The RPC's invocation metadata.
+          request: The RPC's request.
+          timeout: A duration of time in seconds for the RPC or None to
+            indicate that the RPC has no time limit.
+
+        Returns:
+          A UnaryStreamServerRpc with which to "play client" for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def invoke_stream_unary(
+            self, method_descriptor, invocation_metadata, timeout):
+        """Invokes an RPC to be serviced by the system under test.
+
+        Args:
+          method_descriptor: A descriptor.MethodDescriptor describing a stream-unary
+            RPC method.
+          invocation_metadata: The RPC's invocation metadata.
+          timeout: A duration of time in seconds for the RPC or None to
+            indicate that the RPC has no time limit.
+
+        Returns:
+          A StreamUnaryServerRpc with which to "play client" for the RPC.
+        """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def invoke_stream_stream(
+            self, method_descriptor, invocation_metadata, timeout):
+        """Invokes an RPC to be serviced by the system under test.
+
+        Args:
+          method_descriptor: A descriptor.MethodDescriptor describing a stream-stream
+            RPC method.
+          invocation_metadata: The RPC's invocation metadata.
+          timeout: A duration of time in seconds for the RPC or None to
+            indicate that the RPC has no time limit.
+
+        Returns:
+          A StreamStreamServerRpc with which to "play client" for the RPC.
+        """
+        raise NotImplementedError()
+
+
 class Time(six.with_metaclass(abc.ABCMeta)):
     """A simulation of time.
 
@@ -406,3 +678,20 @@ def channel(service_descriptors, time):
     """
     from grpc_testing import _channel
     return _channel.testing_channel(service_descriptors, time)
+
+
+def server_from_dictionary(descriptors_to_servicers, time):
+    """Creates a Server for use in tests of a gRPC Python-using system.
+
+    Args:
+      descriptors_to_servicers: A dictionary from descriptor.ServiceDescriptors
+        defining RPC services to servicer objects (usually instances of classes
+        that implement "Servicer" interfaces defined in generated "_pb2_grpc"
+        modules) implementing those services.
+      time: A Time to be used for tests.
+
+    Returns:
+      A Server for use in tests.
+    """
+    from grpc_testing import _server
+    return _server.server_from_dictionary(descriptors_to_servicers, time)
diff --git a/src/python/grpcio_testing/grpc_testing/_common.py b/src/python/grpcio_testing/grpc_testing/_common.py
index cb4a7f5fa2..1517434ca7 100644
--- a/src/python/grpcio_testing/grpc_testing/_common.py
+++ b/src/python/grpcio_testing/grpc_testing/_common.py
@@ -37,6 +37,16 @@ def fuss_with_metadata(metadata):
         return _fuss(tuple(metadata))
 
 
+def rpc_names(service_descriptors):
+    rpc_names_to_descriptors = {}
+    for service_descriptor in service_descriptors:
+        for method_descriptor in service_descriptor.methods_by_name.values():
+            rpc_name = '/{}/{}'.format(
+                service_descriptor.full_name, method_descriptor.name)
+            rpc_names_to_descriptors[rpc_name] = method_descriptor
+    return rpc_names_to_descriptors
+
+
 class ChannelRpcRead(
         collections.namedtuple(
             'ChannelRpcRead',
@@ -90,3 +100,61 @@ class ChannelHandler(six.with_metaclass(abc.ABCMeta)):
             self, method_full_rpc_name, invocation_metadata, requests,
             requests_closed, timeout):
         raise NotImplementedError()
+
+
+class ServerRpcRead(
+        collections.namedtuple('ServerRpcRead',
+                               ('request', 'requests_closed', 'terminated',))):
+    pass
+
+
+REQUESTS_CLOSED = ServerRpcRead(None, True, False)
+TERMINATED = ServerRpcRead(None, False, True)
+
+
+class ServerRpcHandler(six.with_metaclass(abc.ABCMeta)):
+
+    @abc.abstractmethod
+    def send_initial_metadata(self, initial_metadata):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def take_request(self):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def add_response(self, response):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def send_termination(self, trailing_metadata, code, details):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def add_termination_callback(self, callback):
+        raise NotImplementedError()
+
+
+class Serverish(six.with_metaclass(abc.ABCMeta)):
+
+    @abc.abstractmethod
+    def invoke_unary_unary(
+            self, method_descriptor, handler, invocation_metadata, request,
+            deadline):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def invoke_unary_stream(
+            self, method_descriptor, handler, invocation_metadata, request,
+            deadline):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def invoke_stream_unary(
+            self, method_descriptor, handler, invocation_metadata, deadline):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def invoke_stream_stream(
+            self, method_descriptor, handler, invocation_metadata, deadline):
+        raise NotImplementedError()
diff --git a/src/python/grpcio_testing/grpc_testing/_server/__init__.py b/src/python/grpcio_testing/grpc_testing/_server/__init__.py
new file mode 100644
index 0000000000..759512949a
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/__init__.py
@@ -0,0 +1,20 @@
+# Copyright 2017 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.
+
+from grpc_testing._server import _server
+
+
+def server_from_dictionary(descriptors_to_servicers, time):
+    return _server.server_from_descriptor_to_servicers(
+        descriptors_to_servicers, time)
diff --git a/src/python/grpcio_testing/grpc_testing/_server/_handler.py b/src/python/grpcio_testing/grpc_testing/_server/_handler.py
new file mode 100644
index 0000000000..b47e04c718
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/_handler.py
@@ -0,0 +1,215 @@
+# Copyright 2017 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.
+
+import abc
+import threading
+
+import grpc
+from grpc_testing import _common
+
+_CLIENT_INACTIVE = object()
+
+
+class Handler(_common.ServerRpcHandler):
+
+    @abc.abstractmethod
+    def initial_metadata(self):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def add_request(self, request):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def take_response(self):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def requests_closed(self):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def cancel(self):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def unary_response_termination(self):
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def stream_response_termination(self):
+        raise NotImplementedError()
+
+
+class _Handler(Handler):
+
+    def __init__(self, requests_closed):
+        self._condition = threading.Condition()
+        self._requests = []
+        self._requests_closed = requests_closed
+        self._initial_metadata = None
+        self._responses = []
+        self._trailing_metadata = None
+        self._code = None
+        self._details = None
+        self._unary_response = None
+        self._expiration_future = None
+        self._termination_callbacks = []
+
+    def send_initial_metadata(self, initial_metadata):
+        with self._condition:
+            self._initial_metadata = initial_metadata
+            self._condition.notify_all()
+
+    def take_request(self):
+        with self._condition:
+            while True:
+                if self._code is None:
+                    if self._requests:
+                        request = self._requests.pop(0)
+                        self._condition.notify_all()
+                        return _common.ServerRpcRead(request, False, False)
+                    elif self._requests_closed:
+                        return _common.REQUESTS_CLOSED
+                    else:
+                        self._condition.wait()
+                else:
+                    return _common.TERMINATED
+
+    def is_active(self):
+        with self._condition:
+            return self._code is None
+
+    def add_response(self, response):
+        with self._condition:
+            self._responses.append(response)
+            self._condition.notify_all()
+
+    def send_termination(self, trailing_metadata, code, details):
+        with self._condition:
+            self._trailing_metadata = trailing_metadata
+            self._code = code
+            self._details = details
+            if self._expiration_future is not None:
+                self._expiration_future.cancel()
+            self._condition.notify_all()
+
+    def add_termination_callback(self, termination_callback):
+        with self._condition:
+            if self._code is None:
+                self._termination_callbacks.append(termination_callback)
+                return True
+            else:
+                return False
+
+    def initial_metadata(self):
+        with self._condition:
+            while True:
+                if self._initial_metadata is None:
+                    if self._code is None:
+                        self._condition.wait()
+                    else:
+                        raise ValueError(
+                            'No initial metadata despite status code!')
+                else:
+                    return self._initial_metadata
+
+    def add_request(self, request):
+        with self._condition:
+            self._requests.append(request)
+            self._condition.notify_all()
+
+    def take_response(self):
+        with self._condition:
+            while True:
+                if self._responses:
+                    response = self._responses.pop(0)
+                    self._condition.notify_all()
+                    return response
+                elif self._code is None:
+                    self._condition.wait()
+                else:
+                    raise ValueError('No more responses!')
+
+    def requests_closed(self):
+        with self._condition:
+            self._requests_closed = True
+            self._condition.notify_all()
+
+    def cancel(self):
+        with self._condition:
+            if self._code is None:
+                self._code = _CLIENT_INACTIVE
+                termination_callbacks = self._termination_callbacks
+                self._termination_callbacks = None
+                if self._expiration_future is not None:
+                    self._expiration_future.cancel()
+                self._condition.notify_all()
+        for termination_callback in termination_callbacks:
+            termination_callback()
+
+    def unary_response_termination(self):
+        with self._condition:
+            while True:
+                if self._code is _CLIENT_INACTIVE:
+                    raise ValueError('Huh? Cancelled but wanting status?')
+                elif self._code is None:
+                    self._condition.wait()
+                else:
+                    if self._unary_response is None:
+                        if self._responses:
+                            self._unary_response = self._responses.pop(0)
+                    return (
+                        self._unary_response, self._trailing_metadata,
+                        self._code, self._details,)
+
+
+    def stream_response_termination(self):
+        with self._condition:
+            while True:
+                if self._code is _CLIENT_INACTIVE:
+                    raise ValueError('Huh? Cancelled but wanting status?')
+                elif self._code is None:
+                    self._condition.wait()
+                else:
+                    return self._trailing_metadata, self._code, self._details,
+
+    def expire(self):
+        with self._condition:
+            if self._code is None:
+                if self._initial_metadata is None:
+                    self._initial_metadata = _common.FUSSED_EMPTY_METADATA
+                self._trailing_metadata = _common.FUSSED_EMPTY_METADATA
+                self._code = grpc.StatusCode.DEADLINE_EXCEEDED
+                self._details = 'Took too much time!'
+                termination_callbacks = self._termination_callbacks
+                self._termination_callbacks = None
+                self._condition.notify_all()
+        for termination_callback in termination_callbacks:
+            termination_callback()
+
+    def set_expiration_future(self, expiration_future):
+        with self._condition:
+            self._expiration_future = expiration_future
+
+
+def handler_without_deadline(requests_closed):
+    return _Handler(requests_closed)
+
+
+def handler_with_deadline(requests_closed, time, deadline):
+    handler = _Handler(requests_closed)
+    expiration_future = time.call_at(handler.expire, deadline)
+    handler.set_expiration_future(expiration_future)
+    return handler
diff --git a/src/python/grpcio_testing/grpc_testing/_server/_rpc.py b/src/python/grpcio_testing/grpc_testing/_server/_rpc.py
new file mode 100644
index 0000000000..f81876f4b2
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/_rpc.py
@@ -0,0 +1,153 @@
+# Copyright 2017 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.
+
+import logging
+import threading
+
+import grpc
+from grpc_testing import _common
+
+
+class Rpc(object):
+
+    def __init__(self, handler, invocation_metadata):
+        self._condition = threading.Condition()
+        self._handler = handler
+        self._invocation_metadata = invocation_metadata
+        self._initial_metadata_sent = False
+        self._pending_trailing_metadata = None
+        self._pending_code = None
+        self._pending_details = None
+        self._callbacks = []
+        self._active = True
+        self._rpc_errors = []
+
+    def _ensure_initial_metadata_sent(self):
+        if not self._initial_metadata_sent:
+            self._handler.send_initial_metadata(_common.FUSSED_EMPTY_METADATA)
+            self._initial_metadata_sent = True
+
+    def _call_back(self):
+        callbacks = tuple(self._callbacks)
+        self._callbacks = None
+
+        def call_back():
+            for callback in callbacks:
+                try:
+                    callback()
+                except Exception:  # pylint: disable=broad-except
+                    logging.exception('Exception calling server-side callback!')
+
+        callback_calling_thread = threading.Thread(target=call_back)
+        callback_calling_thread.start()
+
+    def _terminate(self, trailing_metadata, code, details):
+        if self._active:
+            self._active = False
+            self._handler.send_termination(trailing_metadata, code, details)
+            self._call_back()
+            self._condition.notify_all()
+
+    def _complete(self):
+        if self._pending_trailing_metadata is None:
+            trailing_metadata = _common.FUSSED_EMPTY_METADATA
+        else:
+            trailing_metadata = self._pending_trailing_metadata
+        if self._pending_code is None:
+            code = grpc.StatusCode.OK
+        else:
+            code = self._pending_code
+        details = '' if self._pending_details is None else self._pending_details
+        self._terminate(trailing_metadata, code, details)
+
+    def _abort(self, code, details):
+        self._terminate(_common.FUSSED_EMPTY_METADATA, code, details)
+
+    def add_rpc_error(self, rpc_error):
+        with self._condition:
+            self._rpc_errors.append(rpc_error)
+
+    def application_cancel(self):
+        with self._condition:
+            self._abort(
+                grpc.StatusCode.CANCELLED,
+                'Cancelled by server-side application!')
+
+    def application_exception_abort(self, exception):
+        with self._condition:
+            if exception not in self._rpc_errors:
+                logging.exception('Exception calling application!')
+                self._abort(
+                    grpc.StatusCode.UNKNOWN,
+                    'Exception calling application: {}'.format(exception))
+
+    def extrinsic_abort(self):
+        with self._condition:
+            if self._active:
+                self._active = False
+                self._call_back()
+                self._condition.notify_all()
+
+    def unary_response_complete(self, response):
+        with self._condition:
+            self._ensure_initial_metadata_sent()
+            self._handler.add_response(response)
+            self._complete()
+
+    def stream_response(self, response):
+        with self._condition:
+            self._ensure_initial_metadata_sent()
+            self._handler.add_response(response)
+
+    def stream_response_complete(self):
+        with self._condition:
+            self._ensure_initial_metadata_sent()
+            self._complete()
+
+    def send_initial_metadata(self, initial_metadata):
+        with self._condition:
+            if self._initial_metadata_sent:
+                return False
+            else:
+                self._handler.send_initial_metadata(initial_metadata)
+                self._initial_metadata_sent = True
+                return True
+
+    def is_active(self):
+        with self._condition:
+            return self._active
+
+    def add_callback(self, callback):
+        with self._condition:
+            if self._callbacks is None:
+                return False
+            else:
+                self._callbacks.append(callback)
+                return True
+
+    def invocation_metadata(self):
+        with self._condition:
+            return self._invocation_metadata
+
+    def set_trailing_metadata(self, trailing_metadata):
+        with self._condition:
+            self._pending_trailing_metadata = trailing_metadata
+
+    def set_code(self, code):
+        with self._condition:
+            self._pending_code = code
+
+    def set_details(self, details):
+        with self._condition:
+            self._pending_details = details
diff --git a/src/python/grpcio_testing/grpc_testing/_server/_server.py b/src/python/grpcio_testing/grpc_testing/_server/_server.py
new file mode 100644
index 0000000000..66bcfc13c0
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/_server.py
@@ -0,0 +1,149 @@
+# Copyright 2017 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.
+
+import threading
+
+import grpc_testing
+from grpc_testing import _common
+from grpc_testing._server import _handler
+from grpc_testing._server import _rpc
+from grpc_testing._server import _server_rpc
+from grpc_testing._server import _service
+from grpc_testing._server import _servicer_context
+
+
+def _implementation(descriptors_to_servicers, method_descriptor):
+    servicer = descriptors_to_servicers[method_descriptor.containing_service]
+    return getattr(servicer, method_descriptor.name)
+
+
+def _unary_unary_service(request):
+    def service(implementation, rpc, servicer_context):
+        _service.unary_unary(
+            implementation, rpc, request, servicer_context)
+    return service
+
+
+def _unary_stream_service(request):
+    def service(implementation, rpc, servicer_context):
+        _service.unary_stream(
+            implementation, rpc, request, servicer_context)
+    return service
+
+
+def _stream_unary_service(handler):
+    def service(implementation, rpc, servicer_context):
+        _service.stream_unary(implementation, rpc, handler, servicer_context)
+    return service
+
+
+def _stream_stream_service(handler):
+    def service(implementation, rpc, servicer_context):
+        _service.stream_stream(implementation, rpc, handler, servicer_context)
+    return service
+
+
+class _Serverish(_common.Serverish):
+
+    def __init__(self, descriptors_to_servicers, time):
+        self._descriptors_to_servicers = descriptors_to_servicers
+        self._time = time
+
+    def _invoke(
+            self, service_behavior, method_descriptor, handler,
+            invocation_metadata, deadline):
+        implementation = _implementation(
+            self._descriptors_to_servicers, method_descriptor)
+        rpc = _rpc.Rpc(handler, invocation_metadata)
+        if handler.add_termination_callback(rpc.extrinsic_abort):
+            servicer_context = _servicer_context.ServicerContext(
+                rpc, self._time, deadline)
+            service_thread = threading.Thread(
+                target=service_behavior,
+                args=(implementation, rpc, servicer_context,))
+            service_thread.start()
+
+    def invoke_unary_unary(
+            self, method_descriptor, handler, invocation_metadata, request,
+            deadline):
+        self._invoke(
+            _unary_unary_service(request), method_descriptor, handler,
+            invocation_metadata, deadline)
+
+    def invoke_unary_stream(
+            self, method_descriptor, handler, invocation_metadata, request,
+            deadline):
+        self._invoke(
+            _unary_stream_service(request), method_descriptor, handler,
+            invocation_metadata, deadline)
+
+    def invoke_stream_unary(
+            self, method_descriptor, handler, invocation_metadata, deadline):
+        self._invoke(
+            _stream_unary_service(handler), method_descriptor, handler,
+            invocation_metadata, deadline)
+
+    def invoke_stream_stream(
+            self, method_descriptor, handler, invocation_metadata, deadline):
+        self._invoke(
+            _stream_stream_service(handler), method_descriptor, handler,
+            invocation_metadata, deadline)
+
+
+def _deadline_and_handler(requests_closed, time, timeout):
+    if timeout is None:
+        return None, _handler.handler_without_deadline(requests_closed)
+    else:
+        deadline = time.time() + timeout
+        handler = _handler.handler_with_deadline(requests_closed, time, deadline)
+        return deadline, handler
+
+
+class _Server(grpc_testing.Server):
+
+    def __init__(self, serverish, time):
+        self._serverish = serverish
+        self._time = time
+
+    def invoke_unary_unary(
+            self, method_descriptor, invocation_metadata, request, timeout):
+        deadline, handler = _deadline_and_handler(True, self._time, timeout)
+        self._serverish.invoke_unary_unary(
+            method_descriptor, handler, invocation_metadata, request, deadline)
+        return _server_rpc.UnaryUnaryServerRpc(handler)
+
+    def invoke_unary_stream(
+            self, method_descriptor, invocation_metadata, request, timeout):
+        deadline, handler = _deadline_and_handler(True, self._time, timeout)
+        self._serverish.invoke_unary_stream(
+            method_descriptor, handler, invocation_metadata, request, deadline)
+        return _server_rpc.UnaryStreamServerRpc(handler)
+
+    def invoke_stream_unary(
+            self, method_descriptor, invocation_metadata, timeout):
+        deadline, handler = _deadline_and_handler(False, self._time, timeout)
+        self._serverish.invoke_stream_unary(
+            method_descriptor, handler, invocation_metadata, deadline)
+        return _server_rpc.StreamUnaryServerRpc(handler)
+
+    def invoke_stream_stream(
+            self, method_descriptor, invocation_metadata, timeout):
+        deadline, handler = _deadline_and_handler(False, self._time, timeout)
+        self._serverish.invoke_stream_stream(
+            method_descriptor, handler, invocation_metadata, deadline)
+        return _server_rpc.StreamStreamServerRpc(handler)
+
+
+def server_from_descriptor_to_servicers(descriptors_to_servicers, time):
+    return _Server(_Serverish(descriptors_to_servicers, time), time)
diff --git a/src/python/grpcio_testing/grpc_testing/_server/_server_rpc.py b/src/python/grpcio_testing/grpc_testing/_server/_server_rpc.py
new file mode 100644
index 0000000000..30de8ff0e2
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/_server_rpc.py
@@ -0,0 +1,93 @@
+# Copyright 2017 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.
+
+import grpc_testing
+
+
+class UnaryUnaryServerRpc(grpc_testing.UnaryUnaryServerRpc):
+
+    def __init__(self, handler):
+        self._handler = handler
+
+    def initial_metadata(self):
+        return self._handler.initial_metadata()
+
+    def cancel(self):
+        self._handler.cancel()
+
+    def termination(self):
+        return self._handler.unary_response_termination()
+
+
+class UnaryStreamServerRpc(grpc_testing.UnaryStreamServerRpc):
+
+    def __init__(self, handler):
+        self._handler = handler
+
+    def initial_metadata(self):
+        return self._handler.initial_metadata()
+
+    def take_response(self):
+        return self._handler.take_response()
+
+    def cancel(self):
+        self._handler.cancel()
+
+    def termination(self):
+        return self._handler.stream_response_termination()
+
+
+class StreamUnaryServerRpc(grpc_testing.StreamUnaryServerRpc):
+
+    def __init__(self, handler):
+        self._handler = handler
+
+    def initial_metadata(self):
+        return self._handler.initial_metadata()
+
+    def send_request(self, request):
+        self._handler.add_request(request)
+
+    def requests_closed(self):
+        self._handler.requests_closed()
+
+    def cancel(self):
+        self._handler.cancel()
+
+    def termination(self):
+        return self._handler.unary_response_termination()
+
+
+class StreamStreamServerRpc(grpc_testing.StreamStreamServerRpc):
+
+    def __init__(self, handler):
+        self._handler = handler
+
+    def initial_metadata(self):
+        return self._handler.initial_metadata()
+
+    def send_request(self, request):
+        self._handler.add_request(request)
+
+    def requests_closed(self):
+        self._handler.requests_closed()
+
+    def take_response(self):
+        return self._handler.take_response()
+
+    def cancel(self):
+        self._handler.cancel()
+
+    def termination(self):
+        return self._handler.stream_response_termination()
diff --git a/src/python/grpcio_testing/grpc_testing/_server/_service.py b/src/python/grpcio_testing/grpc_testing/_server/_service.py
new file mode 100644
index 0000000000..36b0a2f7ff
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/_service.py
@@ -0,0 +1,88 @@
+# Copyright 2017 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.
+
+import grpc
+
+
+class _RequestIterator(object):
+
+    def __init__(self, rpc, handler):
+        self._rpc = rpc
+        self._handler = handler
+
+    def _next(self):
+        read = self._handler.take_request()
+        if read.requests_closed:
+            raise StopIteration()
+        elif read.terminated:
+            rpc_error = grpc.RpcError()
+            self._rpc.add_rpc_error(rpc_error)
+            raise rpc_error
+        else:
+            return read.request
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        return self._next()
+
+    def next(self):
+        return self._next()
+
+
+def _unary_response(argument, implementation, rpc, servicer_context):
+    try:
+        response = implementation(argument, servicer_context)
+    except Exception as exception:  # pylint: disable=broad-except
+        rpc.application_exception_abort(exception)
+    else:
+        rpc.unary_response_complete(response)
+
+
+def _stream_response(argument, implementation, rpc, servicer_context):
+    try:
+        response_iterator = implementation(argument, servicer_context)
+    except Exception as exception:  # pylint: disable=broad-except
+        rpc.application_exception_abort(exception)
+    else:
+        while True:
+            try:
+                response = next(response_iterator)
+            except StopIteration:
+                rpc.stream_response_complete()
+                break
+            except Exception as exception:  # pylint: disable=broad-except
+                rpc.application_exception_abort(exception)
+                break
+            else:
+                rpc.stream_response(response)
+
+
+def unary_unary(implementation, rpc, request, servicer_context):
+    _unary_response(request, implementation, rpc, servicer_context)
+
+
+def unary_stream(implementation, rpc, request, servicer_context):
+    _stream_response(request, implementation, rpc, servicer_context)
+
+
+def stream_unary(implementation, rpc, handler, servicer_context):
+    _unary_response(
+        _RequestIterator(rpc, handler), implementation, rpc, servicer_context)
+
+
+def stream_stream(implementation, rpc, handler, servicer_context):
+    _stream_response(
+        _RequestIterator(rpc, handler), implementation, rpc, servicer_context)
diff --git a/src/python/grpcio_testing/grpc_testing/_server/_servicer_context.py b/src/python/grpcio_testing/grpc_testing/_server/_servicer_context.py
new file mode 100644
index 0000000000..496689ded0
--- /dev/null
+++ b/src/python/grpcio_testing/grpc_testing/_server/_servicer_context.py
@@ -0,0 +1,74 @@
+# Copyright 2017 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.
+
+import grpc
+from grpc_testing import _common
+
+
+class ServicerContext(grpc.ServicerContext):
+
+    def __init__(self, rpc, time, deadline):
+        self._rpc = rpc
+        self._time = time
+        self._deadline = deadline
+
+    def is_active(self):
+        return self._rpc.is_active()
+
+    def time_remaining(self):
+        if self._rpc.is_active():
+            if self._deadline is None:
+                return None
+            else:
+                return max(0.0, self._deadline - self._time.time())
+        else:
+            return 0.0
+
+    def cancel(self):
+        self._rpc.application_cancel()
+
+    def add_callback(self, callback):
+        return self._rpc.add_callback(callback)
+
+    def invocation_metadata(self):
+        return self._rpc.invocation_metadata()
+
+    def peer(self):
+        raise NotImplementedError()
+
+    def peer_identities(self):
+        raise NotImplementedError()
+
+    def peer_identity_key(self):
+        raise NotImplementedError()
+
+    def auth_context(self):
+        raise NotImplementedError()
+
+    def send_initial_metadata(self, initial_metadata):
+        initial_metadata_sent = self._rpc.send_initial_metadata(
+            _common.fuss_with_metadata(initial_metadata))
+        if not initial_metadata_sent:
+            raise ValueError(
+                'ServicerContext.send_initial_metadata called too late!')
+
+    def set_trailing_metadata(self, trailing_metadata):
+        self._rpc.set_trailing_metadata(
+            _common.fuss_with_metadata(trailing_metadata))
+
+    def set_code(self, code):
+        self._rpc.set_code(code)
+
+    def set_details(self, details):
+        self._rpc.set_details(details)
diff --git a/src/python/grpcio_tests/tests/testing/_server_application.py b/src/python/grpcio_tests/tests/testing/_server_application.py
new file mode 100644
index 0000000000..06f09c8cb4
--- /dev/null
+++ b/src/python/grpcio_tests/tests/testing/_server_application.py
@@ -0,0 +1,66 @@
+# Copyright 2017 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.
+"""An example gRPC Python-using server-side application."""
+
+import grpc
+
+# requests_pb2 is a semantic dependency of this module.
+from tests.testing import _application_common
+from tests.testing.proto import requests_pb2  # pylint: disable=unused-import
+from tests.testing.proto import services_pb2
+from tests.testing.proto import services_pb2_grpc
+
+
+class FirstServiceServicer(services_pb2_grpc.FirstServiceServicer):
+    """Services RPCs."""
+
+    def UnUn(self, request, context):
+        if _application_common.UNARY_UNARY_REQUEST == request:
+            return _application_common.UNARY_UNARY_RESPONSE
+        else:
+            context.set_code(grpc.StatusCode.INVALID_ARGUMENT)
+            context.set_details('Something is wrong with your request!')
+            return services_pb2.Down()
+
+    def UnStre(self, request, context):
+        if _application_common.UNARY_STREAM_REQUEST != request:
+            context.set_code(grpc.StatusCode.INVALID_ARGUMENT)
+            context.set_details('Something is wrong with your request!')
+        return
+        yield services_pb2.Strange()
+
+    def StreUn(self, request_iterator, context):
+        context.send_initial_metadata((
+            ('server_application_metadata_key', 'Hi there!',),))
+        for request in request_iterator:
+            if request != _application_common.STREAM_UNARY_REQUEST:
+                context.set_code(grpc.StatusCode.INVALID_ARGUMENT)
+                context.set_details('Something is wrong with your request!')
+                return services_pb2.Strange()
+            elif not context.is_active():
+                return services_pb2.Strange()
+        else:
+            return _application_common.STREAM_UNARY_RESPONSE
+
+    def StreStre(self, request_iterator, context):
+        for request in request_iterator:
+            if request != _application_common.STREAM_STREAM_REQUEST:
+                context.set_code(grpc.StatusCode.INVALID_ARGUMENT)
+                context.set_details('Something is wrong with your request!')
+                return
+            elif not context.is_active():
+                return
+            else:
+                yield _application_common.STREAM_STREAM_RESPONSE
+                yield _application_common.STREAM_STREAM_RESPONSE
diff --git a/src/python/grpcio_tests/tests/testing/_server_test.py b/src/python/grpcio_tests/tests/testing/_server_test.py
new file mode 100644
index 0000000000..7897bcce01
--- /dev/null
+++ b/src/python/grpcio_tests/tests/testing/_server_test.py
@@ -0,0 +1,169 @@
+# Copyright 2017 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.
+
+import time
+import unittest
+
+import grpc
+import grpc_testing
+
+from tests.testing import _application_common
+from tests.testing import _application_testing_common
+from tests.testing import _server_application
+from tests.testing.proto import services_pb2
+
+
+# TODO(https://github.com/google/protobuf/issues/3452): Drop this skip.
+@unittest.skipIf(
+    services_pb2.DESCRIPTOR.services_by_name.get('FirstService') is None,
+    'Fix protobuf issue 3452!')
+class FirstServiceServicerTest(unittest.TestCase):
+
+    def setUp(self):
+        self._real_time = grpc_testing.strict_real_time()
+        self._fake_time = grpc_testing.strict_fake_time(time.time())
+        servicer = _server_application.FirstServiceServicer()
+        descriptors_to_servicers = {
+            _application_testing_common.FIRST_SERVICE: servicer
+        }
+        self._real_time_server = grpc_testing.server_from_dictionary(
+            descriptors_to_servicers, self._real_time)
+        self._fake_time_server = grpc_testing.server_from_dictionary(
+            descriptors_to_servicers, self._fake_time)
+
+    def test_successful_unary_unary(self):
+        rpc = self._real_time_server.invoke_unary_unary(
+            _application_testing_common.FIRST_SERVICE_UNUN, (),
+            _application_common.UNARY_UNARY_REQUEST, None)
+        initial_metadata = rpc.initial_metadata()
+        response, trailing_metadata, code, details = rpc.termination()
+
+        self.assertEqual(_application_common.UNARY_UNARY_RESPONSE, response)
+        self.assertIs(code, grpc.StatusCode.OK)
+
+    def test_successful_unary_stream(self):
+        rpc = self._real_time_server.invoke_unary_stream(
+            _application_testing_common.FIRST_SERVICE_UNSTRE, (),
+            _application_common.UNARY_STREAM_REQUEST, None)
+        initial_metadata = rpc.initial_metadata()
+        trailing_metadata, code, details = rpc.termination()
+
+        self.assertIs(code, grpc.StatusCode.OK)
+
+    def test_successful_stream_unary(self):
+        rpc = self._real_time_server.invoke_stream_unary(
+            _application_testing_common.FIRST_SERVICE_STREUN, (), None)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.requests_closed()
+        initial_metadata = rpc.initial_metadata()
+        response, trailing_metadata, code, details = rpc.termination()
+
+        self.assertEqual(_application_common.STREAM_UNARY_RESPONSE, response)
+        self.assertIs(code, grpc.StatusCode.OK)
+
+    def test_successful_stream_stream(self):
+        rpc = self._real_time_server.invoke_stream_stream(
+            _application_testing_common.FIRST_SERVICE_STRESTRE, (), None)
+        rpc.send_request(_application_common.STREAM_STREAM_REQUEST)
+        initial_metadata = rpc.initial_metadata()
+        responses = [
+            rpc.take_response(),
+            rpc.take_response(),
+        ]
+        rpc.send_request(_application_common.STREAM_STREAM_REQUEST)
+        rpc.send_request(_application_common.STREAM_STREAM_REQUEST)
+        responses.extend([
+            rpc.take_response(),
+            rpc.take_response(),
+            rpc.take_response(),
+            rpc.take_response(),
+        ])
+        rpc.requests_closed()
+        trailing_metadata, code, details = rpc.termination()
+
+        for response in responses:
+            self.assertEqual(_application_common.STREAM_STREAM_RESPONSE,
+                             response)
+        self.assertIs(code, grpc.StatusCode.OK)
+
+    def test_server_rpc_idempotence(self):
+        rpc = self._real_time_server.invoke_unary_unary(
+            _application_testing_common.FIRST_SERVICE_UNUN, (),
+            _application_common.UNARY_UNARY_REQUEST, None)
+        first_initial_metadata = rpc.initial_metadata()
+        second_initial_metadata = rpc.initial_metadata()
+        third_initial_metadata = rpc.initial_metadata()
+        first_termination = rpc.termination()
+        second_termination = rpc.termination()
+        third_termination = rpc.termination()
+
+        for later_initial_metadata in (second_initial_metadata,
+                                       third_initial_metadata,):
+            self.assertEqual(first_initial_metadata, later_initial_metadata)
+        response = first_termination[0]
+        terminal_metadata = first_termination[1]
+        code = first_termination[2]
+        details = first_termination[3]
+        for later_termination in (second_termination, third_termination,):
+            self.assertEqual(response, later_termination[0])
+            self.assertEqual(terminal_metadata, later_termination[1])
+            self.assertIs(code, later_termination[2])
+            self.assertEqual(details, later_termination[3])
+        self.assertEqual(_application_common.UNARY_UNARY_RESPONSE, response)
+        self.assertIs(code, grpc.StatusCode.OK)
+
+    def test_misbehaving_client_unary_unary(self):
+        rpc = self._real_time_server.invoke_unary_unary(
+            _application_testing_common.FIRST_SERVICE_UNUN, (),
+            _application_common.ERRONEOUS_UNARY_UNARY_REQUEST, None)
+        initial_metadata = rpc.initial_metadata()
+        response, trailing_metadata, code, details = rpc.termination()
+
+        self.assertIsNot(code, grpc.StatusCode.OK)
+
+    def test_infinite_request_stream_real_time(self):
+        rpc = self._real_time_server.invoke_stream_unary(
+            _application_testing_common.FIRST_SERVICE_STREUN, (),
+            _application_common.INFINITE_REQUEST_STREAM_TIMEOUT)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        initial_metadata = rpc.initial_metadata()
+        self._real_time.sleep_for(
+            _application_common.INFINITE_REQUEST_STREAM_TIMEOUT * 2)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        response, trailing_metadata, code, details = rpc.termination()
+
+        self.assertIs(code, grpc.StatusCode.DEADLINE_EXCEEDED)
+
+    def test_infinite_request_stream_fake_time(self):
+        rpc = self._fake_time_server.invoke_stream_unary(
+            _application_testing_common.FIRST_SERVICE_STREUN, (),
+            _application_common.INFINITE_REQUEST_STREAM_TIMEOUT)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        initial_metadata = rpc.initial_metadata()
+        self._fake_time.sleep_for(
+            _application_common.INFINITE_REQUEST_STREAM_TIMEOUT * 2)
+        rpc.send_request(_application_common.STREAM_UNARY_REQUEST)
+        response, trailing_metadata, code, details = rpc.termination()
+
+        self.assertIs(code, grpc.StatusCode.DEADLINE_EXCEEDED)
+
+
+if __name__ == '__main__':
+    unittest.main(verbosity=2)
diff --git a/src/python/grpcio_tests/tests/tests.json b/src/python/grpcio_tests/tests/tests.json
index c10719b86f..d61297b918 100644
--- a/src/python/grpcio_tests/tests/tests.json
+++ b/src/python/grpcio_tests/tests/tests.json
@@ -10,6 +10,7 @@
   "protoc_plugin.beta_python_plugin_test.PythonPluginTest",
   "reflection._reflection_servicer_test.ReflectionServicerTest",
   "testing._client_test.ClientTest",
+  "testing._server_test.FirstServiceServicerTest",
   "testing._time_test.StrictFakeTimeTest",
   "testing._time_test.StrictRealTimeTest",
   "unit._api_test.AllTest",
diff --git a/src/ruby/.rubocop_todo.yml b/src/ruby/.rubocop_todo.yml
index 05db404582..32b84b8de0 100644
--- a/src/ruby/.rubocop_todo.yml
+++ b/src/ruby/.rubocop_todo.yml
@@ -1,44 +1,569 @@
-# This configuration was generated by `rubocop --auto-gen-config`
-# on 2015-05-22 13:23:34 -0700 using RuboCop version 0.30.1.
+# This configuration was generated by
+# `rubocop --auto-gen-config`
+# on 2017-09-04 17:00:36 +0200 using RuboCop version 0.49.1.
 # The point is for the user to remove these configuration records
 # one by one as the offenses are removed from the code base.
 # Note that changes in the inspected code, or installation of new
 # versions of RuboCop, may require this file to be generated again.
 
-# Offense count: 30
-Metrics/AbcSize:
-  Max: 38
+# Offense count: 3
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, IndentOneStep, IndentationWidth.
+# SupportedStyles: case, end
+Layout/CaseIndentation:
+  Exclude:
+    - 'tools/platform_check.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Layout/CommentIndentation:
+  Exclude:
+    - 'qps/client.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Layout/EmptyLineAfterMagicComment:
+  Exclude:
+    - 'tools/grpc-tools.gemspec'
+
+# Offense count: 33
+# Cop supports --auto-correct.
+# Configuration parameters: AllowAdjacentOneLineDefs, NumberOfEmptyLines.
+Layout/EmptyLineBetweenDefs:
+  Exclude:
+    - 'qps/client.rb'
+    - 'qps/histogram.rb'
+    - 'qps/proxy-worker.rb'
+    - 'qps/server.rb'
+    - 'qps/worker.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Layout/EmptyLines:
+  Exclude:
+    - 'qps/qps-common.rb'
+
+# Offense count: 8
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: empty_lines, empty_lines_except_namespace, empty_lines_special, no_empty_lines
+Layout/EmptyLinesAroundClassBody:
+  Exclude:
+    - 'pb/grpc/testing/duplicate/echo_duplicate_services_pb.rb'
+    - 'pb/grpc/testing/metrics_services_pb.rb'
+    - 'pb/src/proto/grpc/testing/test_services_pb.rb'
+    - 'qps/src/proto/grpc/testing/proxy-service_services_pb.rb'
+    - 'qps/src/proto/grpc/testing/services_services_pb.rb'
+
+# Offense count: 28
+# Cop supports --auto-correct.
+# Configuration parameters: AllowForAlignment, ForceEqualSignAlignment.
+Layout/ExtraSpacing:
+  Enabled: false
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: normal, rails
+Layout/IndentationConsistency:
+  Exclude:
+    - 'pb/grpc/health/checker.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: Width, IgnoredPatterns.
+Layout/IndentationWidth:
+  Exclude:
+    - 'pb/grpc/health/checker.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: symmetrical, new_line, same_line
+Layout/MultilineHashBraceLayout:
+  Exclude:
+    - 'spec/generic/active_call_spec.rb'
+
+# Offense count: 70
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: symmetrical, new_line, same_line
+Layout/MultilineMethodCallBraceLayout:
+  Enabled: false
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, IndentationWidth.
+# SupportedStyles: aligned, indented, indented_relative_to_receiver
+Layout/MultilineMethodCallIndentation:
+  Exclude:
+    - 'spec/generic/rpc_desc_spec.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: symmetrical, new_line, same_line
+Layout/MultilineMethodDefinitionBraceLayout:
+  Exclude:
+    - 'spec/generic/client_stub_spec.rb'
+
+# Offense count: 5
+# Cop supports --auto-correct.
+Layout/SpaceAfterColon:
+  Exclude:
+    - 'lib/grpc/generic/rpc_server.rb'
+
+# Offense count: 7
+# Cop supports --auto-correct.
+Layout/SpaceAfterComma:
+  Exclude:
+    - 'qps/client.rb'
+
+# Offense count: 27
+# Cop supports --auto-correct.
+# Configuration parameters: AllowForAlignment.
+Layout/SpaceAroundOperators:
+  Exclude:
+    - 'qps/client.rb'
+    - 'qps/histogram.rb'
+    - 'qps/proxy-worker.rb'
+    - 'qps/server.rb'
+    - 'spec/generic/active_call_spec.rb'
+    - 'spec/generic/rpc_server_spec.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, EnforcedStyleForEmptyBraces, SupportedStylesForEmptyBraces, SpaceBeforeBlockParameters.
+# SupportedStyles: space, no_space
+# SupportedStylesForEmptyBraces: space, no_space
+Layout/SpaceInsideBlockBraces:
+  Exclude:
+    - 'stress/stress_client.rb'
+
+# Offense count: 4
+# Cop supports --auto-correct.
+Layout/SpaceInsideBrackets:
+  Exclude:
+    - 'tools/bin/grpc_tools_ruby_protoc'
+    - 'tools/bin/grpc_tools_ruby_protoc_plugin'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, EnforcedStyleForEmptyBraces, SupportedStylesForEmptyBraces.
+# SupportedStyles: space, no_space, compact
+# SupportedStylesForEmptyBraces: space, no_space
+Layout/SpaceInsideHashLiteralBraces:
+  Exclude:
+    - 'qps/server.rb'
+
+# Offense count: 6
+# Cop supports --auto-correct.
+Layout/SpaceInsidePercentLiteralDelimiters:
+  Exclude:
+    - 'spec/generic/client_stub_spec.rb'
+    - 'tools/grpc-tools.gemspec'
 
 # Offense count: 3
-# Configuration parameters: CountComments.
-Metrics/ClassLength:
-  Max: 200
+# Cop supports --auto-correct.
+Layout/Tab:
+  Exclude:
+    - 'pb/grpc/health/checker.rb'
+    - 'qps/client.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Layout/TrailingWhitespace:
+  Exclude:
+    - 'qps/worker.rb'
+
+# Offense count: 1
+Lint/IneffectiveAccessModifier:
+  Exclude:
+    - 'lib/grpc/generic/active_call.rb'
+
+# Offense count: 4
+# Cop supports --auto-correct.
+Lint/PercentStringArray:
+  Exclude:
+    - 'spec/client_server_spec.rb'
+    - 'spec/generic/active_call_spec.rb'
+    - 'spec/generic/client_stub_spec.rb'
+
+# Offense count: 4
+Lint/ScriptPermission:
+  Exclude:
+    - 'qps/client.rb'
+    - 'qps/histogram.rb'
+    - 'qps/qps-common.rb'
+    - 'qps/server.rb'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: IgnoreEmptyBlocks, AllowUnusedKeywordArguments.
+Lint/UnusedBlockArgument:
+  Exclude:
+    - 'qps/client.rb'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: AllowUnusedKeywordArguments, IgnoreEmptyMethods.
+Lint/UnusedMethodArgument:
+  Exclude:
+    - 'qps/client.rb'
+
+# Offense count: 1
+# Configuration parameters: ContextCreatingMethods, MethodCreatingMethods.
+Lint/UselessAccessModifier:
+  Exclude:
+    - 'lib/grpc/logconfig.rb'
+
+# Offense count: 1
+Lint/UselessAssignment:
+  Exclude:
+    - 'qps/client.rb'
 
-# Offense count: 35
+# Offense count: 4
+Lint/Void:
+  Exclude:
+    - 'stress/metrics_server.rb'
+    - 'stress/stress_client.rb'
+
+# Offense count: 53
+Metrics/AbcSize:
+  Max: 57
+
+# Offense count: 81
+# Configuration parameters: CountComments, ExcludedMethods.
+Metrics/BlockLength:
+  Max: 715
+
+# Offense count: 82
+# Configuration parameters: AllowHeredoc, AllowURI, URISchemes, IgnoreCopDirectives, IgnoredPatterns.
+# URISchemes: http, https
+Metrics/LineLength:
+  Max: 141
+
+# Offense count: 82
 # Configuration parameters: CountComments.
 Metrics/MethodLength:
-  Max: 36
+  Max: 54
 
-# Offense count: 7
+# Offense count: 5
 # Configuration parameters: CountKeywordArgs.
 Metrics/ParameterLists:
-  Max: 8
+  Max: 7
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Performance/RedundantBlockCall:
+  Exclude:
+    - 'spec/generic/client_stub_spec.rb'
+
+# Offense count: 5
+# Cop supports --auto-correct.
+# Configuration parameters: MaxKeyValuePairs.
+Performance/RedundantMerge:
+  Exclude:
+    - 'spec/generic/active_call_spec.rb'
+    - 'spec/generic/client_stub_spec.rb'
+
+# Offense count: 8
+# Cop supports --auto-correct.
+Performance/TimesMap:
+  Exclude:
+    - 'spec/channel_spec.rb'
+    - 'spec/client_server_spec.rb'
+    - 'spec/server_spec.rb'
+
+# Offense count: 7
+Style/AccessorMethodName:
+  Exclude:
+    - 'qps/server.rb'
+    - 'stress/metrics_server.rb'
+    - 'stress/stress_client.rb'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: prefer_alias, prefer_alias_method
+Style/Alias:
+  Exclude:
+    - 'lib/grpc/generic/rpc_server.rb'
+    - 'lib/grpc/notifier.rb'
+
+# Offense count: 7
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, ProceduralMethods, FunctionalMethods, IgnoredMethods.
+# SupportedStyles: line_count_based, semantic, braces_for_chaining
+# ProceduralMethods: benchmark, bm, bmbm, create, each_with_object, measure, new, realtime, tap, with_object
+# FunctionalMethods: let, let!, subject, watch
+# IgnoredMethods: lambda, proc, it
+Style/BlockDelimiters:
+  Exclude:
+    - 'qps/client.rb'
+    - 'qps/proxy-worker.rb'
+    - 'qps/server.rb'
+    - 'qps/worker.rb'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+Style/ClassMethods:
+  Exclude:
+    - 'tools/platform_check.rb'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, SingleLineConditionsOnly, IncludeTernaryExpressions.
+# SupportedStyles: assign_to_condition, assign_inside_condition
+Style/ConditionalAssignment:
+  Exclude:
+    - 'lib/grpc/generic/rpc_server.rb'
+    - 'lib/grpc/generic/service.rb'
+
+# Offense count: 19
+Style/Documentation:
+  Exclude:
+    - 'spec/**/*'
+    - 'test/**/*'
+    - 'pb/grpc/testing/duplicate/echo_duplicate_services_pb.rb'
+    - 'pb/grpc/testing/metrics_services_pb.rb'
+    - 'pb/src/proto/grpc/testing/test_pb.rb'
+    - 'qps/client.rb'
+    - 'qps/histogram.rb'
+    - 'qps/proxy-worker.rb'
+    - 'qps/server.rb'
+    - 'qps/src/proto/grpc/testing/proxy-service_services_pb.rb'
+    - 'qps/src/proto/grpc/testing/services_pb.rb'
+    - 'qps/src/proto/grpc/testing/services_services_pb.rb'
+    - 'qps/worker.rb'
+    - 'stress/metrics_server.rb'
+    - 'stress/stress_client.rb'
+    - 'tools/platform_check.rb'
 
-# Offense count: 9
+# Offense count: 8
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: compact, expanded
+Style/EmptyMethod:
+  Exclude:
+    - 'bin/noproto_server.rb'
+    - 'lib/grpc/logconfig.rb'
+    - 'spec/generic/rpc_desc_spec.rb'
+
+# Offense count: 2
+# Configuration parameters: ExpectMatchingDefinition, Regex, IgnoreExecutableScripts, AllowedAcronyms.
+# AllowedAcronyms: CLI, DSL, ACL, API, ASCII, CPU, CSS, DNS, EOF, GUID, HTML, HTTP, HTTPS, ID, IP, JSON, LHS, QPS, RAM, RHS, RPC, SLA, SMTP, SQL, SSH, TCP, TLS, TTL, UDP, UI, UID, UUID, URI, URL, UTF8, VM, XML, XMPP, XSRF, XSS
+Style/FileName:
+  Exclude:
+    - 'qps/src/proto/grpc/testing/proxy-service_pb.rb'
+    - 'qps/src/proto/grpc/testing/proxy-service_services_pb.rb'
+
+# Offense count: 12
 # Configuration parameters: AllowedVariables.
 Style/GlobalVars:
-  Enabled: false
+  Exclude:
+    - 'ext/grpc/extconf.rb'
+
+# Offense count: 3
+# Configuration parameters: MinBodyLength.
+Style/GuardClause:
+  Exclude:
+    - 'lib/grpc/generic/bidi_call.rb'
+    - 'lib/grpc/generic/rpc_server.rb'
+    - 'qps/client.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, UseHashRocketsWithSymbolValues, PreferHashRocketsForNonAlnumEndingSymbols.
+# SupportedStyles: ruby19, hash_rockets, no_mixed_keys, ruby19_no_mixed_keys
+Style/HashSyntax:
+  Exclude:
+    - 'stress/metrics_server.rb'
+
+# Offense count: 1
+Style/IfInsideElse:
+  Exclude:
+    - 'lib/grpc/generic/rpc_desc.rb'
+
+# Offense count: 4
+# Cop supports --auto-correct.
+# Configuration parameters: MaxLineLength.
+Style/IfUnlessModifier:
+  Exclude:
+    - 'ext/grpc/extconf.rb'
+    - 'qps/histogram.rb'
+    - 'stress/stress_client.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Style/MethodCallWithoutArgsParentheses:
+  Exclude:
+    - 'qps/client.rb'
+
+# Offense count: 3
+# Cop supports --auto-correct.
+Style/MultilineIfModifier:
+  Exclude:
+    - 'lib/grpc/generic/bidi_call.rb'
+    - 'lib/grpc/generic/client_stub.rb'
+    - 'spec/spec_helper.rb'
+
+# Offense count: 7
+# Cop supports --auto-correct.
+Style/MutableConstant:
+  Exclude:
+    - 'ext/grpc/extconf.rb'
+    - 'lib/grpc/version.rb'
+    - 'spec/compression_options_spec.rb'
+    - 'spec/generic/active_call_spec.rb'
+    - 'tools/version.rb'
 
 # Offense count: 1
-# Configuration parameters: EnforcedStyle, MinBodyLength, SupportedStyles.
-Style/Next:
+# Cop supports --auto-correct.
+Style/NegatedWhile:
+  Exclude:
+    - 'qps/client.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+# Configuration parameters: AutoCorrect, EnforcedStyle, SupportedStyles.
+# SupportedStyles: predicate, comparison
+Style/NumericPredicate:
+  Exclude:
+    - 'spec/**/*'
+    - 'ext/grpc/extconf.rb'
+
+# Offense count: 7
+# Cop supports --auto-correct.
+Style/ParallelAssignment:
+  Exclude:
+    - 'bin/math_server.rb'
+    - 'lib/grpc/generic/rpc_server.rb'
+    - 'spec/generic/client_stub_spec.rb'
+    - 'spec/generic/rpc_desc_spec.rb'
+    - 'spec/generic/rpc_server_pool_spec.rb'
+    - 'spec/generic/rpc_server_spec.rb'
+
+# Offense count: 8
+# Cop supports --auto-correct.
+# Configuration parameters: PreferredDelimiters.
+Style/PercentLiteralDelimiters:
+  Exclude:
+    - 'end2end/grpc_class_init_driver.rb'
+    - 'spec/client_server_spec.rb'
+    - 'spec/generic/active_call_spec.rb'
+    - 'spec/generic/client_stub_spec.rb'
+    - 'tools/grpc-tools.gemspec'
+
+# Offense count: 3
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: compact, exploded
+Style/RaiseArgs:
+  Exclude:
+    - 'stress/metrics_server.rb'
+
+# Offense count: 4
+# Cop supports --auto-correct.
+Style/RedundantParentheses:
+  Exclude:
+    - 'lib/grpc/generic/rpc_server.rb'
+    - 'qps/client.rb'
+    - 'qps/proxy-worker.rb'
+    - 'spec/generic/rpc_desc_spec.rb'
+
+# Offense count: 5
+# Cop supports --auto-correct.
+# Configuration parameters: AllowMultipleReturnValues.
+Style/RedundantReturn:
+  Exclude:
+    - 'end2end/grpc_class_init_client.rb'
+
+# Offense count: 77
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: only_raise, only_fail, semantic
+Style/SignalException:
   Enabled: false
 
 # Offense count: 2
-# Configuration parameters: Methods.
-Style/SingleLineBlockParams:
-  Enabled: false
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles.
+# SupportedStyles: use_perl_names, use_english_names
+Style/SpecialGlobalVars:
+  Exclude:
+    - 'ext/grpc/extconf.rb'
+    - 'stress/stress_client.rb'
+
+# Offense count: 189
+# Cop supports --auto-correct.
+# Configuration parameters: EnforcedStyle, SupportedStyles, ConsistentQuotesInMultiline.
+# SupportedStyles: single_quotes, double_quotes
+Style/StringLiterals:
+  Exclude:
+    - 'pb/grpc/testing/metrics_pb.rb'
+    - 'pb/src/proto/grpc/testing/empty_pb.rb'
+    - 'pb/src/proto/grpc/testing/messages_pb.rb'
+    - 'qps/proxy-worker.rb'
+    - 'qps/server.rb'
+    - 'qps/src/proto/grpc/testing/control_pb.rb'
+    - 'qps/src/proto/grpc/testing/messages_pb.rb'
+    - 'qps/src/proto/grpc/testing/payloads_pb.rb'
+    - 'qps/src/proto/grpc/testing/proxy-service_pb.rb'
+    - 'qps/src/proto/grpc/testing/stats_pb.rb'
+    - 'qps/worker.rb'
 
 # Offense count: 1
 Style/StructInheritance:
-  Enabled: false
+  Exclude:
+    - 'lib/grpc/generic/rpc_desc.rb'
+
+# Offense count: 10
+# Cop supports --auto-correct.
+# Configuration parameters: MinSize, SupportedStyles.
+# SupportedStyles: percent, brackets
+Style/SymbolArray:
+  EnforcedStyle: brackets
+
+# Offense count: 2
+# Cop supports --auto-correct.
+# Configuration parameters: IgnoredMethods.
+# IgnoredMethods: respond_to, define_method
+Style/SymbolProc:
+  Exclude:
+    - 'qps/client.rb'
+    - 'stress/stress_client.rb'
+
+# Offense count: 6
+# Cop supports --auto-correct.
+# Configuration parameters: AllowNamedUnderscoreVariables.
+Style/TrailingUnderscoreVariable:
+  Exclude:
+    - 'spec/channel_credentials_spec.rb'
+    - 'spec/server_credentials_spec.rb'
+
+# Offense count: 3
+# Cop supports --auto-correct.
+# Configuration parameters: ExactNameMatch, AllowPredicates, AllowDSLWriters, IgnoreClassMethods, Whitelist.
+# Whitelist: to_ary, to_a, to_c, to_enum, to_h, to_hash, to_i, to_int, to_io, to_open, to_path, to_proc, to_r, to_regexp, to_str, to_s, to_sym
+Style/TrivialAccessors:
+  Exclude:
+    - 'qps/histogram.rb'
+
+# Offense count: 3
+# Cop supports --auto-correct.
+Style/UnneededInterpolation:
+  Exclude:
+    - 'pb/grpc/health/checker.rb'
+
+# Offense count: 1
+# Cop supports --auto-correct.
+Style/YodaCondition:
+  Exclude:
+    - 'stress/stress_client.rb'
+
+# Offense count: 2
+# Cop supports --auto-correct.
+Style/ZeroLengthPredicate:
+  Exclude:
+    - 'lib/grpc/generic/rpc_server.rb'
diff --git a/src/ruby/end2end/grpc_class_init_client.rb b/src/ruby/end2end/grpc_class_init_client.rb
index b9e3d6054f..c35719a71f 100755
--- a/src/ruby/end2end/grpc_class_init_client.rb
+++ b/src/ruby/end2end/grpc_class_init_client.rb
@@ -41,7 +41,7 @@ def run_gc_stress_test(test_proc)
   GC.enable
   construct_many(test_proc)
 
-  GC.start(full_mark: true, immediate_sweep: true)
+  GC.start
   construct_many(test_proc)
 end
 
diff --git a/src/ruby/ext/grpc/rb_grpc_imports.generated.c b/src/ruby/ext/grpc/rb_grpc_imports.generated.c
index 9671d794c5..0402ce34fb 100644
--- a/src/ruby/ext/grpc/rb_grpc_imports.generated.c
+++ b/src/ruby/ext/grpc/rb_grpc_imports.generated.c
@@ -93,6 +93,7 @@ grpc_alarm_destroy_type grpc_alarm_destroy_import;
 grpc_channel_check_connectivity_state_type grpc_channel_check_connectivity_state_import;
 grpc_channel_num_external_connectivity_watchers_type grpc_channel_num_external_connectivity_watchers_import;
 grpc_channel_watch_connectivity_state_type grpc_channel_watch_connectivity_state_import;
+grpc_channel_support_connectivity_watcher_type grpc_channel_support_connectivity_watcher_import;
 grpc_channel_create_call_type grpc_channel_create_call_import;
 grpc_channel_ping_type grpc_channel_ping_import;
 grpc_channel_register_call_type grpc_channel_register_call_import;
@@ -399,6 +400,7 @@ void grpc_rb_load_imports(HMODULE library) {
   grpc_channel_check_connectivity_state_import = (grpc_channel_check_connectivity_state_type) GetProcAddress(library, "grpc_channel_check_connectivity_state");
   grpc_channel_num_external_connectivity_watchers_import = (grpc_channel_num_external_connectivity_watchers_type) GetProcAddress(library, "grpc_channel_num_external_connectivity_watchers");
   grpc_channel_watch_connectivity_state_import = (grpc_channel_watch_connectivity_state_type) GetProcAddress(library, "grpc_channel_watch_connectivity_state");
+  grpc_channel_support_connectivity_watcher_import = (grpc_channel_support_connectivity_watcher_type) GetProcAddress(library, "grpc_channel_support_connectivity_watcher");
   grpc_channel_create_call_import = (grpc_channel_create_call_type) GetProcAddress(library, "grpc_channel_create_call");
   grpc_channel_ping_import = (grpc_channel_ping_type) GetProcAddress(library, "grpc_channel_ping");
   grpc_channel_register_call_import = (grpc_channel_register_call_type) GetProcAddress(library, "grpc_channel_register_call");
diff --git a/src/ruby/ext/grpc/rb_grpc_imports.generated.h b/src/ruby/ext/grpc/rb_grpc_imports.generated.h
index b64199be8e..e3704e592b 100644
--- a/src/ruby/ext/grpc/rb_grpc_imports.generated.h
+++ b/src/ruby/ext/grpc/rb_grpc_imports.generated.h
@@ -260,6 +260,9 @@ extern grpc_channel_num_external_connectivity_watchers_type grpc_channel_num_ext
 typedef void(*grpc_channel_watch_connectivity_state_type)(grpc_channel *channel, grpc_connectivity_state last_observed_state, gpr_timespec deadline, grpc_completion_queue *cq, void *tag);
 extern grpc_channel_watch_connectivity_state_type grpc_channel_watch_connectivity_state_import;
 #define grpc_channel_watch_connectivity_state grpc_channel_watch_connectivity_state_import
+typedef int(*grpc_channel_support_connectivity_watcher_type)(grpc_channel *channel);
+extern grpc_channel_support_connectivity_watcher_type grpc_channel_support_connectivity_watcher_import;
+#define grpc_channel_support_connectivity_watcher grpc_channel_support_connectivity_watcher_import
 typedef grpc_call *(*grpc_channel_create_call_type)(grpc_channel *channel, grpc_call *parent_call, uint32_t propagation_mask, grpc_completion_queue *completion_queue, grpc_slice method, const grpc_slice *host, gpr_timespec deadline, void *reserved);
 extern grpc_channel_create_call_type grpc_channel_create_call_import;
 #define grpc_channel_create_call grpc_channel_create_call_import