Merge branch 'master' of github.com:grpc/grpc into grpclb_races

9 files changed, 55 insertions, 63 deletions

diff --git a/src/core/ext/client_config/client_channel.c b/src/core/ext/client_config/client_channel.c
index a6056c3e8d..cbf79afa17 100644
--- a/src/core/ext/client_config/client_channel.c
+++ b/src/core/ext/client_config/client_channel.c
@@ -513,10 +513,14 @@ static void retry_waiting_locked(grpc_exec_ctx *exec_ctx, call_data *calld) {
 
 static void subchannel_ready(grpc_exec_ctx *exec_ctx, void *arg,
                              grpc_error *error) {
-  call_data *calld = arg;
+  grpc_call_element *elem = arg;
+  call_data *calld = elem->call_data;
+  channel_data *chand = elem->channel_data;
   gpr_mu_lock(&calld->mu);
   GPR_ASSERT(calld->creation_phase ==
              GRPC_SUBCHANNEL_CALL_HOLDER_PICKING_SUBCHANNEL);
+  grpc_polling_entity_del_from_pollset_set(exec_ctx, calld->pollent,
+                                           chand->interested_parties);
   calld->creation_phase = GRPC_SUBCHANNEL_CALL_HOLDER_NOT_CREATING;
   if (calld->connected_subchannel == NULL) {
     gpr_atm_no_barrier_store(&calld->subchannel_call, 1);
@@ -564,6 +568,9 @@ typedef struct {
   grpc_closure closure;
 } continue_picking_args;
 
+/** Return true if subchannel is available immediately (in which case on_ready
+    should not be called), or false otherwise (in which case on_ready should be
+    called when the subchannel is available). */
 static bool pick_subchannel(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
                             grpc_metadata_batch *initial_metadata,
                             uint32_t initial_metadata_flags,
@@ -629,8 +636,8 @@ static bool pick_subchannel(grpc_exec_ctx *exec_ctx, grpc_call_element *elem,
     gpr_mu_unlock(&chand->mu);
     // TODO(dgq): make this deadline configurable somehow.
     const grpc_lb_policy_pick_args inputs = {
-        calld->pollent, initial_metadata, initial_metadata_flags,
-        &calld->lb_token_mdelem, gpr_inf_future(GPR_CLOCK_MONOTONIC)};
+        initial_metadata, initial_metadata_flags, &calld->lb_token_mdelem,
+        gpr_inf_future(GPR_CLOCK_MONOTONIC)};
     r = grpc_lb_policy_pick(exec_ctx, lb_policy, &inputs, connected_subchannel,
                             NULL, on_ready);
     GRPC_LB_POLICY_UNREF(exec_ctx, lb_policy, "pick_subchannel");
@@ -672,6 +679,7 @@ static void cc_start_transport_stream_op(grpc_exec_ctx *exec_ctx,
                                          grpc_call_element *elem,
                                          grpc_transport_stream_op *op) {
   call_data *calld = elem->call_data;
+  channel_data *chand = elem->channel_data;
   GRPC_CALL_LOG_OP(GPR_INFO, elem, op);
   grpc_deadline_state_client_start_transport_stream_op(exec_ctx, elem, op);
   /* try to (atomically) get the call */
@@ -739,14 +747,20 @@ retry:
       calld->connected_subchannel == NULL &&
       op->send_initial_metadata != NULL) {
     calld->creation_phase = GRPC_SUBCHANNEL_CALL_HOLDER_PICKING_SUBCHANNEL;
-    grpc_closure_init(&calld->next_step, subchannel_ready, calld);
+    grpc_closure_init(&calld->next_step, subchannel_ready, elem);
     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(exec_ctx, elem, op->send_initial_metadata,
                         op->send_initial_metadata_flags,
                         &calld->connected_subchannel, &calld->next_step,
                         GRPC_ERROR_NONE)) {
       calld->creation_phase = GRPC_SUBCHANNEL_CALL_HOLDER_NOT_CREATING;
       GRPC_CALL_STACK_UNREF(exec_ctx, calld->owning_call, "pick_subchannel");
+    } else {
+      grpc_polling_entity_add_to_pollset_set(exec_ctx, calld->pollent,
+                                             chand->interested_parties);
     }
   }
   /* if we've got a subchannel, then let's ask it to create a call */
diff --git a/src/core/ext/client_config/lb_policy.h b/src/core/ext/client_config/lb_policy.h
index 110d08fcac..de424cd105 100644
--- a/src/core/ext/client_config/lb_policy.h
+++ b/src/core/ext/client_config/lb_policy.h
@@ -35,7 +35,6 @@
 #define GRPC_CORE_EXT_CLIENT_CONFIG_LB_POLICY_H
 
 #include "src/core/ext/client_config/subchannel.h"
-#include "src/core/lib/iomgr/polling_entity.h"
 #include "src/core/lib/transport/connectivity_state.h"
 
 /** A load balancing policy: specified by a vtable and a struct (which
@@ -55,8 +54,6 @@ struct grpc_lb_policy {
 
 /** Extra arguments for an LB pick */
 typedef struct grpc_lb_policy_pick_args {
-  /** Parties interested in the pick's progress */
-  grpc_polling_entity *pollent;
   /** Initial metadata associated with the picking call. */
   grpc_metadata_batch *initial_metadata;
   /** Bitmask used for selective cancelling. See \a
@@ -153,7 +150,8 @@ void grpc_lb_policy_init(grpc_lb_policy *policy,
     once the pick is complete with its error argument set to indicate
     success or failure.
 
-    Any I/O should be done under \a pick_args->pollent. */
+    Any IO should be done under the \a interested_parties \a grpc_pollset_set
+    in the \a grpc_lb_policy struct. */
 int grpc_lb_policy_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *policy,
                         const grpc_lb_policy_pick_args *pick_args,
                         grpc_connected_subchannel **target, void **user_data,
diff --git a/src/core/ext/lb_policy/grpclb/grpclb.c b/src/core/ext/lb_policy/grpclb/grpclb.c
index c18cc0eb9e..140dc9b782 100644
--- a/src/core/ext/lb_policy/grpclb/grpclb.c
+++ b/src/core/ext/lb_policy/grpclb/grpclb.c
@@ -388,14 +388,14 @@ static grpc_lb_addresses *process_serverlist(
           GPR_ARRAY_SIZE(server->load_balance_token) - 1;
       grpc_mdstr *lb_token_mdstr = grpc_mdstr_from_buffer(
           (uint8_t *)server->load_balance_token, lb_token_size);
-      user_data = grpc_mdelem_from_metadata_strings(
-          GRPC_MDSTR_LOAD_REPORTING_INITIAL, lb_token_mdstr);
+      user_data = grpc_mdelem_from_metadata_strings(GRPC_MDSTR_LB_TOKEN,
+                                                    lb_token_mdstr);
     } else {
       gpr_log(GPR_ERROR,
               "Missing LB token for backend address '%s'. The empty token will "
               "be used instead",
               grpc_sockaddr_to_uri((struct sockaddr *)&addr.addr));
-      user_data = GRPC_MDELEM_LOAD_REPORTING_INITIAL_EMPTY;
+      user_data = GRPC_MDELEM_LB_TOKEN_EMPTY;
     }
 
     grpc_lb_addresses_set_address(lb_addresses, addr_idx, &addr.addr, addr.len,
@@ -447,6 +447,9 @@ static void rr_handover_locked(grpc_exec_ctx *exec_ctx,
             (intptr_t)glb_policy->rr_policy);
   }
   GPR_ASSERT(glb_policy->rr_policy != NULL);
+  grpc_pollset_set_add_pollset_set(exec_ctx,
+                                   glb_policy->rr_policy->interested_parties,
+                                   glb_policy->base.interested_parties);
   glb_policy->rr_connectivity->state = grpc_lb_policy_check_connectivity(
       exec_ctx, glb_policy->rr_policy, &error);
   grpc_lb_policy_notify_on_state_change(
@@ -677,8 +680,6 @@ static void glb_cancel_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
   while (pp != NULL) {
     pending_pick *next = pp->next;
     if (pp->target == target) {
-      grpc_polling_entity_del_from_pollset_set(
-          exec_ctx, pp->pick_args.pollent, glb_policy->base.interested_parties);
       *target = NULL;
       grpc_exec_ctx_sched(
           exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure,
@@ -710,8 +711,6 @@ static void glb_cancel_picks(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
     pending_pick *next = pp->next;
     if ((pp->pick_args.initial_metadata_flags & initial_metadata_flags_mask) ==
         initial_metadata_flags_eq) {
-      grpc_polling_entity_del_from_pollset_set(
-          exec_ctx, pp->pick_args.pollent, glb_policy->base.interested_parties);
       grpc_exec_ctx_sched(
           exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure,
           GRPC_ERROR_CREATE_REFERENCING("Pick Cancelled", &error, 1), NULL);
@@ -801,8 +800,6 @@ static int glb_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
      * Eventually, wrapped_on_complete will be called, which will -among other
      * things- add the LB token to the call's initial metadata */
   } else {
-    grpc_polling_entity_add_to_pollset_set(exec_ctx, pick_args->pollent,
-                                           glb_policy->base.interested_parties);
     add_pending_pick(&glb_policy->pending_picks, pick_args, target,
                      on_complete);
 
@@ -926,7 +923,7 @@ static lb_client_data *lb_client_data_create(glb_lb_policy *glb_policy) {
 
   /* Note the following LB call progresses every time there's activity in \a
    * glb_policy->base.interested_parties, which is comprised of the polling
-   * entities passed to glb_pick(). */
+   * entities from \a client_channel. */
   lb_client->lb_call = grpc_channel_create_pollset_set_call(
       glb_policy->lb_channel, NULL, GRPC_PROPAGATE_DEFAULTS,
       glb_policy->base.interested_parties,
diff --git a/src/core/ext/lb_policy/pick_first/pick_first.c b/src/core/ext/lb_policy/pick_first/pick_first.c
index 961a0c9b19..6533327343 100644
--- a/src/core/ext/lb_policy/pick_first/pick_first.c
+++ b/src/core/ext/lb_policy/pick_first/pick_first.c
@@ -39,7 +39,6 @@
 
 typedef struct pending_pick {
   struct pending_pick *next;
-  grpc_polling_entity *pollent;
   uint32_t initial_metadata_flags;
   grpc_connected_subchannel **target;
   grpc_closure *on_complete;
@@ -119,8 +118,6 @@ static void pf_shutdown(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) {
   while (pp != NULL) {
     pending_pick *next = pp->next;
     *pp->target = NULL;
-    grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                             p->base.interested_parties);
     grpc_exec_ctx_sched(exec_ctx, pp->on_complete, GRPC_ERROR_NONE, NULL);
     gpr_free(pp);
     pp = next;
@@ -138,8 +135,6 @@ static void pf_cancel_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
   while (pp != NULL) {
     pending_pick *next = pp->next;
     if (pp->target == target) {
-      grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                               p->base.interested_parties);
       *target = NULL;
       grpc_exec_ctx_sched(
           exec_ctx, pp->on_complete,
@@ -168,8 +163,6 @@ static void pf_cancel_picks(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
     pending_pick *next = pp->next;
     if ((pp->initial_metadata_flags & initial_metadata_flags_mask) ==
         initial_metadata_flags_eq) {
-      grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                               p->base.interested_parties);
       grpc_exec_ctx_sched(
           exec_ctx, pp->on_complete,
           GRPC_ERROR_CREATE_REFERENCING("Pick Cancelled", &error, 1), NULL);
@@ -229,11 +222,8 @@ static int pf_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
     if (!p->started_picking) {
       start_picking(exec_ctx, p);
     }
-    grpc_polling_entity_add_to_pollset_set(exec_ctx, pick_args->pollent,
-                                           p->base.interested_parties);
     pp = gpr_malloc(sizeof(*pp));
     pp->next = p->pending_picks;
-    pp->pollent = pick_args->pollent;
     pp->target = target;
     pp->initial_metadata_flags = pick_args->initial_metadata_flags;
     pp->on_complete = on_complete;
@@ -319,8 +309,6 @@ static void pf_connectivity_changed(grpc_exec_ctx *exec_ctx, void *arg,
         while ((pp = p->pending_picks)) {
           p->pending_picks = pp->next;
           *pp->target = selected;
-          grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                                   p->base.interested_parties);
           grpc_exec_ctx_sched(exec_ctx, pp->on_complete, GRPC_ERROR_NONE, NULL);
           gpr_free(pp);
         }
diff --git a/src/core/ext/lb_policy/round_robin/round_robin.c b/src/core/ext/lb_policy/round_robin/round_robin.c
index 325202fc6d..9bd3f9da24 100644
--- a/src/core/ext/lb_policy/round_robin/round_robin.c
+++ b/src/core/ext/lb_policy/round_robin/round_robin.c
@@ -78,9 +78,6 @@ int grpc_lb_round_robin_trace = 0;
 typedef struct pending_pick {
   struct pending_pick *next;
 
-  /* polling entity for the pick()'s async notification */
-  grpc_polling_entity *pollent;
-
   /* output argument where to store the pick()ed user_data. It'll be NULL if no
    * such data is present or there's an error (the definite test for errors is
    * \a target being NULL). */
@@ -318,8 +315,6 @@ static void rr_cancel_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
   while (pp != NULL) {
     pending_pick *next = pp->next;
     if (pp->target == target) {
-      grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                               p->base.interested_parties);
       *target = NULL;
       grpc_exec_ctx_sched(
           exec_ctx, pp->on_complete,
@@ -348,8 +343,6 @@ static void rr_cancel_picks(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
     pending_pick *next = pp->next;
     if ((pp->initial_metadata_flags & initial_metadata_flags_mask) ==
         initial_metadata_flags_eq) {
-      grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                               p->base.interested_parties);
       *pp->target = NULL;
       grpc_exec_ctx_sched(
           exec_ctx, pp->on_complete,
@@ -422,11 +415,8 @@ static int rr_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol,
     if (!p->started_picking) {
       start_picking(exec_ctx, p);
     }
-    grpc_polling_entity_add_to_pollset_set(exec_ctx, pick_args->pollent,
-                                           p->base.interested_parties);
     pp = gpr_malloc(sizeof(*pp));
     pp->next = p->pending_picks;
-    pp->pollent = pick_args->pollent;
     pp->target = target;
     pp->on_complete = on_complete;
     pp->initial_metadata_flags = pick_args->initial_metadata_flags;
@@ -482,8 +472,6 @@ static void rr_connectivity_changed(grpc_exec_ctx *exec_ctx, void *arg,
                     "[RR CONN CHANGED] TARGET <-- SUBCHANNEL %p (NODE %p)",
                     (void *)selected->subchannel, (void *)selected);
           }
-          grpc_polling_entity_del_from_pollset_set(exec_ctx, pp->pollent,
-                                                   p->base.interested_parties);
           grpc_exec_ctx_sched(exec_ctx, pp->on_complete, GRPC_ERROR_NONE, NULL);
           gpr_free(pp);
         }
diff --git a/src/core/ext/load_reporting/load_reporting.h b/src/core/ext/load_reporting/load_reporting.h
index e37817d8c2..e13097654d 100644
--- a/src/core/ext/load_reporting/load_reporting.h
+++ b/src/core/ext/load_reporting/load_reporting.h
@@ -37,13 +37,21 @@
 #include <grpc/impl/codegen/grpc_types.h>
 #include "src/core/lib/channel/channel_stack.h"
 
-/** Metadata key for initial metadata coming from clients */
-/* TODO(dgq): change to the final value TBD */
-#define GRPC_LOAD_REPORTING_INITIAL_MD_KEY "load-reporting-initial"
+/** Metadata key for the gRPC LB load balancer token.
+ *
+ * The value corresponding to this key is an opaque token that is given to the
+ * frontend as part of each pick; the frontend sends this token to the backend
+ * in each request it sends when using that pick. The token is used by the
+ * backend to verify the request and to allow the backend to report load to the
+ * gRPC LB system. */
+#define GRPC_LB_TOKEN_MD_KEY "lb-token"
 
-/** Metadata key for trailing metadata from servers */
-/* TODO(dgq): change to the final value TBD */
-#define GRPC_LOAD_REPORTING_TRAILING_MD_KEY "load-reporting-trailing"
+/** Metadata key for gRPC LB cost reporting.
+ *
+ * The value corresponding to this key is an opaque binary blob reported by the
+ * backend as part of its trailing metadata containing cost information for the
+ * call. */
+#define GRPC_LB_COST_MD_KEY "lb-cost"
 
 /** Identifiers for the invocation point of the users LR callback */
 typedef enum grpc_load_reporting_source {
diff --git a/src/core/ext/load_reporting/load_reporting_filter.c b/src/core/ext/load_reporting/load_reporting_filter.c
index 394f0cb832..22bf36367f 100644
--- a/src/core/ext/load_reporting/load_reporting_filter.c
+++ b/src/core/ext/load_reporting/load_reporting_filter.c
@@ -75,7 +75,7 @@ static grpc_mdelem *recv_md_filter(void *user_data, grpc_mdelem *md) {
 
   if (md->key == GRPC_MDSTR_PATH) {
     calld->service_method = grpc_mdstr_as_c_string(md->value);
-  } else if (md->key == GRPC_MDSTR_LOAD_REPORTING_INITIAL) {
+  } else if (md->key == GRPC_MDSTR_LB_TOKEN) {
     calld->initial_md_string = gpr_strdup(grpc_mdstr_as_c_string(md->value));
     return NULL;
   }
@@ -193,7 +193,7 @@ static grpc_mdelem *lr_trailing_md_filter(void *user_data, grpc_mdelem *md) {
   grpc_call_element *elem = user_data;
   call_data *calld = elem->call_data;
 
-  if (md->key == GRPC_MDSTR_LOAD_REPORTING_TRAILING) {
+  if (md->key == GRPC_MDSTR_LB_COST) {
     calld->trailing_md_string = gpr_strdup(grpc_mdstr_as_c_string(md->value));
     return NULL;
   }
diff --git a/src/core/lib/transport/static_metadata.c b/src/core/lib/transport/static_metadata.c
index 5e0352a467..f019ef156a 100644
--- a/src/core/lib/transport/static_metadata.c
+++ b/src/core/lib/transport/static_metadata.c
@@ -126,9 +126,9 @@ const char *const grpc_static_metadata_strings[GRPC_STATIC_MDSTR_COUNT] = {
     "if-range",
     "if-unmodified-since",
     "last-modified",
+    "lb-cost",
+    "lb-token",
     "link",
-    "load-reporting-initial",
-    "load-reporting-trailing",
     "location",
     "max-forwards",
     ":method",
diff --git a/src/core/lib/transport/static_metadata.h b/src/core/lib/transport/static_metadata.h
index 5b9ee1a60a..e0a8196419 100644
--- a/src/core/lib/transport/static_metadata.h
+++ b/src/core/lib/transport/static_metadata.h
@@ -175,12 +175,12 @@ extern grpc_mdstr grpc_static_mdstr_table[GRPC_STATIC_MDSTR_COUNT];
 #define GRPC_MDSTR_IF_UNMODIFIED_SINCE (&grpc_static_mdstr_table[62])
 /* "last-modified" */
 #define GRPC_MDSTR_LAST_MODIFIED (&grpc_static_mdstr_table[63])
+/* "lb-cost" */
+#define GRPC_MDSTR_LB_COST (&grpc_static_mdstr_table[64])
+/* "lb-token" */
+#define GRPC_MDSTR_LB_TOKEN (&grpc_static_mdstr_table[65])
 /* "link" */
-#define GRPC_MDSTR_LINK (&grpc_static_mdstr_table[64])
-/* "load-reporting-initial" */
-#define GRPC_MDSTR_LOAD_REPORTING_INITIAL (&grpc_static_mdstr_table[65])
-/* "load-reporting-trailing" */
-#define GRPC_MDSTR_LOAD_REPORTING_TRAILING (&grpc_static_mdstr_table[66])
+#define GRPC_MDSTR_LINK (&grpc_static_mdstr_table[66])
 /* "location" */
 #define GRPC_MDSTR_LOCATION (&grpc_static_mdstr_table[67])
 /* "max-forwards" */
@@ -337,13 +337,12 @@ extern uintptr_t grpc_static_mdelem_user_data[GRPC_STATIC_MDELEM_COUNT];
 #define GRPC_MDELEM_IF_UNMODIFIED_SINCE_EMPTY (&grpc_static_mdelem_table[44])
 /* "last-modified": "" */
 #define GRPC_MDELEM_LAST_MODIFIED_EMPTY (&grpc_static_mdelem_table[45])
+/* "lb-cost": "" */
+#define GRPC_MDELEM_LB_COST_EMPTY (&grpc_static_mdelem_table[46])
+/* "lb-token": "" */
+#define GRPC_MDELEM_LB_TOKEN_EMPTY (&grpc_static_mdelem_table[47])
 /* "link": "" */
-#define GRPC_MDELEM_LINK_EMPTY (&grpc_static_mdelem_table[46])
-/* "load-reporting-initial": "" */
-#define GRPC_MDELEM_LOAD_REPORTING_INITIAL_EMPTY (&grpc_static_mdelem_table[47])
-/* "load-reporting-trailing": "" */
-#define GRPC_MDELEM_LOAD_REPORTING_TRAILING_EMPTY \
-  (&grpc_static_mdelem_table[48])
+#define GRPC_MDELEM_LINK_EMPTY (&grpc_static_mdelem_table[48])
 /* "location": "" */
 #define GRPC_MDELEM_LOCATION_EMPTY (&grpc_static_mdelem_table[49])
 /* "max-forwards": "" */