diff options
Diffstat (limited to 'src/core/ext/lb_policy')
| -rw-r--r-- | src/core/ext/lb_policy/grpclb/grpclb.c | 291 | ||||
| -rw-r--r-- | src/core/ext/lb_policy/grpclb/load_balancer_api.c | 20 | ||||
| -rw-r--r-- | src/core/ext/lb_policy/grpclb/load_balancer_api.h | 8 | ||||
| -rw-r--r-- | src/core/ext/lb_policy/pick_first/pick_first.c | 2 | ||||
| -rw-r--r-- | src/core/ext/lb_policy/round_robin/round_robin.c | 312 |
5 files changed, 420 insertions, 213 deletions
diff --git a/src/core/ext/lb_policy/grpclb/grpclb.c b/src/core/ext/lb_policy/grpclb/grpclb.c index 30e412e358..df0db61c22 100644 --- a/src/core/ext/lb_policy/grpclb/grpclb.c +++ b/src/core/ext/lb_policy/grpclb/grpclb.c @@ -116,16 +116,18 @@ #include "src/core/lib/iomgr/sockaddr.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/timer.h" +#include "src/core/lib/slice/slice_string_helpers.h" #include "src/core/lib/support/backoff.h" #include "src/core/lib/support/string.h" #include "src/core/lib/surface/call.h" #include "src/core/lib/surface/channel.h" #include "src/core/lib/transport/static_metadata.h" -#define BACKOFF_MULTIPLIER 1.6 -#define BACKOFF_JITTER 0.2 -#define BACKOFF_MIN_SECONDS 10 -#define BACKOFF_MAX_SECONDS 60 +#define GRPC_GRPCLB_MIN_CONNECT_TIMEOUT_SECONDS 20 +#define GRPC_GRPCLB_INITIAL_CONNECT_BACKOFF_SECONDS 1 +#define GRPC_GRPCLB_RECONNECT_BACKOFF_MULTIPLIER 1.6 +#define GRPC_GRPCLB_RECONNECT_MAX_BACKOFF_SECONDS 120 +#define GRPC_GRPCLB_RECONNECT_JITTER 0.2 int grpc_lb_glb_trace = 0; @@ -181,17 +183,24 @@ static void wrapped_rr_closure(grpc_exec_ctx *exec_ctx, void *arg, NULL); if (wc_arg->rr_policy != NULL) { - /* if target is NULL, no pick has been made by the RR policy (eg, all + /* if *target is NULL, no pick has been made by the RR policy (eg, all * addresses failed to connect). There won't be any user_data/token * available */ - if (wc_arg->target != NULL) { - initial_metadata_add_lb_token(wc_arg->initial_metadata, - wc_arg->lb_token_mdelem_storage, - GRPC_MDELEM_REF(wc_arg->lb_token)); + if (*wc_arg->target != NULL) { + if (wc_arg->lb_token != NULL) { + initial_metadata_add_lb_token(wc_arg->initial_metadata, + wc_arg->lb_token_mdelem_storage, + GRPC_MDELEM_REF(wc_arg->lb_token)); + } else { + gpr_log(GPR_ERROR, + "No LB token for connected subchannel pick %p (from RR " + "instance %p).", + (void *)*wc_arg->target, (void *)wc_arg->rr_policy); + abort(); + } } if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "Unreffing RR (0x%" PRIxPTR ")", - (intptr_t)wc_arg->rr_policy); + gpr_log(GPR_INFO, "Unreffing RR %p", (void *)wc_arg->rr_policy); } GRPC_LB_POLICY_UNREF(exec_ctx, wc_arg->rr_policy, "wrapped_rr_closure"); } @@ -409,7 +418,7 @@ static void parse_server(const grpc_grpclb_server *server, } /* Returns addresses extracted from \a serverlist. */ -static grpc_lb_addresses *process_serverlist( +static grpc_lb_addresses *process_serverlist_locked( const grpc_grpclb_serverlist *serverlist) { size_t num_valid = 0; /* first pass: count how many are valid in order to allocate the necessary @@ -449,10 +458,12 @@ static grpc_lb_addresses *process_serverlist( user_data = grpc_mdelem_from_metadata_strings(GRPC_MDSTR_LB_TOKEN, lb_token_mdstr); } else { - gpr_log(GPR_ERROR, + char *uri = grpc_sockaddr_to_uri(&addr); + gpr_log(GPR_INFO, "Missing LB token for backend address '%s'. The empty token will " "be used instead", - grpc_sockaddr_to_uri(&addr)); + uri); + gpr_free(uri); user_data = GRPC_MDELEM_LB_TOKEN_EMPTY; } @@ -465,6 +476,68 @@ static grpc_lb_addresses *process_serverlist( return lb_addresses; } +/* returns true if the new RR policy should replace the current one, if any */ +static bool update_lb_connectivity_status_locked( + grpc_exec_ctx *exec_ctx, glb_lb_policy *glb_policy, + grpc_connectivity_state new_rr_state, grpc_error *new_rr_state_error) { + grpc_error *curr_state_error; + const grpc_connectivity_state curr_glb_state = grpc_connectivity_state_check( + &glb_policy->state_tracker, &curr_state_error); + + /* The new connectivity status is a function of the previous one and the new + * input coming from the status of the RR policy. + * + * current state (grpclb's) + * | + * v || I | C | R | TF | SD | <- new state (RR's) + * ===++====+=====+=====+======+======+ + * I || I | C | R | [I] | [I] | + * ---++----+-----+-----+------+------+ + * C || I | C | R | [C] | [C] | + * ---++----+-----+-----+------+------+ + * R || I | C | R | [R] | [R] | + * ---++----+-----+-----+------+------+ + * TF || I | C | R | [TF] | [TF] | + * ---++----+-----+-----+------+------+ + * SD || NA | NA | NA | NA | NA | (*) + * ---++----+-----+-----+------+------+ + * + * A [STATE] indicates that the old RR policy is kept. In those cases, STATE + * is the current state of grpclb, which is left untouched. + * + * In summary, if the new state is TRANSIENT_FAILURE or SHUTDOWN, stick to + * the previous RR instance. + * + * Note that the status is never updated to SHUTDOWN as a result of calling + * this function. Only glb_shutdown() has the power to set that state. + * + * (*) This function mustn't be called during shutting down. */ + GPR_ASSERT(curr_glb_state != GRPC_CHANNEL_SHUTDOWN); + + switch (new_rr_state) { + case GRPC_CHANNEL_TRANSIENT_FAILURE: + case GRPC_CHANNEL_SHUTDOWN: + GPR_ASSERT(new_rr_state_error != GRPC_ERROR_NONE); + return false; /* don't replace the RR policy */ + case GRPC_CHANNEL_INIT: + case GRPC_CHANNEL_IDLE: + case GRPC_CHANNEL_CONNECTING: + case GRPC_CHANNEL_READY: + GPR_ASSERT(new_rr_state_error == GRPC_ERROR_NONE); + } + + if (grpc_lb_glb_trace) { + gpr_log(GPR_INFO, + "Setting grpclb's state to %s from new RR policy %p state.", + grpc_connectivity_state_name(new_rr_state), + (void *)glb_policy->rr_policy); + } + grpc_connectivity_state_set(exec_ctx, &glb_policy->state_tracker, + new_rr_state, GRPC_ERROR_REF(new_rr_state_error), + "update_lb_connectivity_status_locked"); + return true; +} + /* perform a pick over \a rr_policy. Given that a pick can return immediately * (ignoring its completion callback) we need to perform the cleanups this * callback would be otherwise resposible for */ @@ -506,7 +579,7 @@ static grpc_lb_policy *create_rr_locked( grpc_lb_policy_args args; memset(&args, 0, sizeof(args)); args.client_channel_factory = glb_policy->cc_factory; - grpc_lb_addresses *addresses = process_serverlist(serverlist); + grpc_lb_addresses *addresses = process_serverlist_locked(serverlist); // Replace the LB addresses in the channel args that we pass down to // the subchannel. @@ -527,49 +600,84 @@ static void glb_rr_connectivity_changed(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); /* glb_policy->rr_policy may be NULL (initial handover) */ static void rr_handover_locked(grpc_exec_ctx *exec_ctx, - glb_lb_policy *glb_policy, grpc_error *error) { + glb_lb_policy *glb_policy) { GPR_ASSERT(glb_policy->serverlist != NULL && glb_policy->serverlist->num_servers > 0); + if (glb_policy->shutting_down) return; + + grpc_lb_policy *new_rr_policy = + create_rr_locked(exec_ctx, glb_policy->serverlist, glb_policy); + if (new_rr_policy == NULL) { + gpr_log(GPR_ERROR, + "Failure creating a RoundRobin policy for serverlist update with " + "%lu entries. The previous RR instance (%p), if any, will continue " + "to be used. Future updates from the LB will attempt to create new " + "instances.", + (unsigned long)glb_policy->serverlist->num_servers, + (void *)glb_policy->rr_policy); + return; + } + + grpc_error *new_rr_state_error = NULL; + const grpc_connectivity_state new_rr_state = + grpc_lb_policy_check_connectivity(exec_ctx, new_rr_policy, + &new_rr_state_error); + /* Connectivity state is a function of the new RR policy just created */ + const bool replace_old_rr = update_lb_connectivity_status_locked( + exec_ctx, glb_policy, new_rr_state, new_rr_state_error); + + if (!replace_old_rr) { + /* dispose of the new RR policy that won't be used after all */ + GRPC_LB_POLICY_UNREF(exec_ctx, new_rr_policy, "rr_handover_no_replace"); + if (grpc_lb_glb_trace) { + gpr_log(GPR_INFO, + "Keeping old RR policy (%p) despite new serverlist: new RR " + "policy was in %s connectivity state.", + (void *)glb_policy->rr_policy, + grpc_connectivity_state_name(new_rr_state)); + } + return; + } + if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "RR handover. Old RR: %p", (void *)glb_policy->rr_policy); + gpr_log(GPR_INFO, "Created RR policy (%p) to replace old RR (%p)", + (void *)new_rr_policy, (void *)glb_policy->rr_policy); } + if (glb_policy->rr_policy != NULL) { /* if we are phasing out an existing RR instance, unref it. */ GRPC_LB_POLICY_UNREF(exec_ctx, glb_policy->rr_policy, "rr_handover"); } - glb_policy->rr_policy = - create_rr_locked(exec_ctx, glb_policy->serverlist, glb_policy); - if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "Created RR policy (%p)", (void *)glb_policy->rr_policy); - } + /* Finally update the RR policy to the newly created one */ + glb_policy->rr_policy = new_rr_policy; - GPR_ASSERT(glb_policy->rr_policy != NULL); + /* Add the gRPC LB's interested_parties pollset_set to that of the newly + * created RR policy. This will make the RR policy progress upon activity on + * gRPC LB, which in turn is tied to the application's call */ grpc_pollset_set_add_pollset_set(exec_ctx, glb_policy->rr_policy->interested_parties, glb_policy->base.interested_parties); + /* Allocate the data for the tracking of the new RR policy's connectivity. + * It'll be deallocated in glb_rr_connectivity_changed() */ rr_connectivity_data *rr_connectivity = gpr_malloc(sizeof(rr_connectivity_data)); memset(rr_connectivity, 0, sizeof(rr_connectivity_data)); grpc_closure_init(&rr_connectivity->on_change, glb_rr_connectivity_changed, rr_connectivity); rr_connectivity->glb_policy = glb_policy; - rr_connectivity->state = grpc_lb_policy_check_connectivity( - exec_ctx, glb_policy->rr_policy, &error); + rr_connectivity->state = new_rr_state; - grpc_connectivity_state_set(exec_ctx, &glb_policy->state_tracker, - rr_connectivity->state, GRPC_ERROR_REF(error), - "rr_handover"); - /* subscribe */ + /* Subscribe to changes to the connectivity of the new RR */ GRPC_LB_POLICY_WEAK_REF(&glb_policy->base, "rr_connectivity_cb"); grpc_lb_policy_notify_on_state_change(exec_ctx, glb_policy->rr_policy, &rr_connectivity->state, &rr_connectivity->on_change); grpc_lb_policy_exit_idle(exec_ctx, glb_policy->rr_policy); - /* flush pending ops */ + /* Update picks and pings in wait */ pending_pick *pp; while ((pp = glb_policy->pending_picks)) { glb_policy->pending_picks = pp->next; @@ -600,28 +708,36 @@ static void rr_handover_locked(grpc_exec_ctx *exec_ctx, static void glb_rr_connectivity_changed(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - /* If shutdown or error free the arg. Rely on the rest of the code to set the - * right grpclb status. */ - rr_connectivity_data *rr_conn_data = arg; - glb_lb_policy *glb_policy = rr_conn_data->glb_policy; - - if (rr_conn_data->state != GRPC_CHANNEL_SHUTDOWN && - !glb_policy->shutting_down) { - gpr_mu_lock(&glb_policy->mu); - /* RR not shutting down. Mimic the RR's policy state */ - grpc_connectivity_state_set(exec_ctx, &glb_policy->state_tracker, - rr_conn_data->state, GRPC_ERROR_REF(error), - "rr_connectivity_cb"); - /* resubscribe. Reuse the "rr_connectivity_cb" weak ref. */ + rr_connectivity_data *rr_connectivity = arg; + glb_lb_policy *glb_policy = rr_connectivity->glb_policy; + + gpr_mu_lock(&glb_policy->mu); + const bool shutting_down = glb_policy->shutting_down; + bool unref_needed = false; + GRPC_ERROR_REF(error); + + if (rr_connectivity->state == GRPC_CHANNEL_SHUTDOWN || shutting_down) { + /* RR policy shutting down. Don't renew subscription and free the arg of + * this callback. In addition we need to stash away the current policy to + * be UNREF'd after releasing the lock. Otherwise, if the UNREF is the last + * one, the policy would be destroyed, alongside the lock, which would + * result in a use-after-free */ + unref_needed = true; + gpr_free(rr_connectivity); + } else { /* rr state != SHUTDOWN && !shutting down: biz as usual */ + update_lb_connectivity_status_locked(exec_ctx, glb_policy, + rr_connectivity->state, error); + /* Resubscribe. Reuse the "rr_connectivity_cb" weak ref. */ grpc_lb_policy_notify_on_state_change(exec_ctx, glb_policy->rr_policy, - &rr_conn_data->state, - &rr_conn_data->on_change); - gpr_mu_unlock(&glb_policy->mu); - } else { + &rr_connectivity->state, + &rr_connectivity->on_change); + } + gpr_mu_unlock(&glb_policy->mu); + if (unref_needed) { GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &glb_policy->base, "rr_connectivity_cb"); - gpr_free(rr_conn_data); } + GRPC_ERROR_UNREF(error); } static grpc_lb_policy *glb_create(grpc_exec_ctx *exec_ctx, @@ -756,8 +872,26 @@ static void glb_shutdown(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { glb_policy->pending_picks = NULL; pending_ping *pping = glb_policy->pending_pings; glb_policy->pending_pings = NULL; + if (glb_policy->rr_policy) { + GRPC_LB_POLICY_UNREF(exec_ctx, glb_policy->rr_policy, "glb_shutdown"); + } + grpc_connectivity_state_set( + exec_ctx, &glb_policy->state_tracker, GRPC_CHANNEL_SHUTDOWN, + GRPC_ERROR_CREATE("Channel Shutdown"), "glb_shutdown"); + /* We need a copy of the lb_call pointer because we can't cancell the call + * while holding glb_policy->mu: lb_on_server_status_received, invoked due to + * the cancel, needs to acquire that same lock */ + grpc_call *lb_call = glb_policy->lb_call; gpr_mu_unlock(&glb_policy->mu); + /* glb_policy->lb_call and this local lb_call must be consistent at this point + * because glb_policy->lb_call is only assigned in lb_call_init_locked as part + * of query_for_backends_locked, which can only be invoked while + * glb_policy->shutting_down is false. */ + if (lb_call != NULL) { + grpc_call_cancel(lb_call, NULL); + /* lb_on_server_status_received will pick up the cancel and clean up */ + } while (pp != NULL) { pending_pick *next = pp->next; *pp->target = NULL; @@ -772,22 +906,6 @@ static void glb_shutdown(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { GRPC_ERROR_NONE, NULL); pping = next; } - - if (glb_policy->rr_policy) { - GRPC_LB_POLICY_UNREF(exec_ctx, glb_policy->rr_policy, "glb_shutdown"); - } - - if (glb_policy->started_picking) { - if (glb_policy->lb_call != NULL) { - grpc_call_cancel(glb_policy->lb_call, NULL); - /* lb_on_server_status_received will pick up the cancellation and clean up - */ - } - } - - grpc_connectivity_state_set( - exec_ctx, &glb_policy->state_tracker, GRPC_CHANNEL_SHUTDOWN, - GRPC_ERROR_CREATE("Channel Shutdown"), "glb_shutdown"); } static void glb_cancel_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, @@ -957,9 +1075,10 @@ static void lb_on_server_status_received(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); static void lb_on_response_received(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); -static void lb_call_init(glb_lb_policy *glb_policy) { +static void lb_call_init_locked(glb_lb_policy *glb_policy) { GPR_ASSERT(glb_policy->server_name != NULL); GPR_ASSERT(glb_policy->server_name[0] != '\0'); + GPR_ASSERT(!glb_policy->shutting_down); /* Note the following LB call progresses every time there's activity in \a * glb_policy->base.interested_parties, which is comprised of the polling @@ -975,10 +1094,10 @@ static void lb_call_init(glb_lb_policy *glb_policy) { grpc_grpclb_request *request = grpc_grpclb_request_create(glb_policy->server_name); - gpr_slice request_payload_slice = grpc_grpclb_request_encode(request); + grpc_slice request_payload_slice = grpc_grpclb_request_encode(request); glb_policy->lb_request_payload = grpc_raw_byte_buffer_create(&request_payload_slice, 1); - gpr_slice_unref(request_payload_slice); + grpc_slice_unref(request_payload_slice); grpc_grpclb_request_destroy(request); glb_policy->lb_call_status_details = NULL; @@ -989,12 +1108,15 @@ static void lb_call_init(glb_lb_policy *glb_policy) { grpc_closure_init(&glb_policy->lb_on_response_received, lb_on_response_received, glb_policy); - gpr_backoff_init(&glb_policy->lb_call_backoff_state, BACKOFF_MULTIPLIER, - BACKOFF_JITTER, BACKOFF_MIN_SECONDS * 1000, - BACKOFF_MAX_SECONDS * 1000); + gpr_backoff_init(&glb_policy->lb_call_backoff_state, + GRPC_GRPCLB_INITIAL_CONNECT_BACKOFF_SECONDS, + GRPC_GRPCLB_RECONNECT_BACKOFF_MULTIPLIER, + GRPC_GRPCLB_RECONNECT_JITTER, + GRPC_GRPCLB_MIN_CONNECT_TIMEOUT_SECONDS * 1000, + GRPC_GRPCLB_RECONNECT_MAX_BACKOFF_SECONDS * 1000); } -static void lb_call_destroy(glb_lb_policy *glb_policy) { +static void lb_call_destroy_locked(glb_lb_policy *glb_policy) { GPR_ASSERT(glb_policy->lb_call != NULL); grpc_call_destroy(glb_policy->lb_call); glb_policy->lb_call = NULL; @@ -1012,7 +1134,9 @@ static void lb_call_destroy(glb_lb_policy *glb_policy) { static void query_for_backends_locked(grpc_exec_ctx *exec_ctx, glb_lb_policy *glb_policy) { GPR_ASSERT(glb_policy->lb_channel != NULL); - lb_call_init(glb_policy); + if (glb_policy->shutting_down) return; + + lb_call_init_locked(glb_policy); if (grpc_lb_glb_trace) { gpr_log(GPR_INFO, "Query for backends (grpclb: %p, lb_call: %p)", @@ -1084,19 +1208,20 @@ static void lb_on_response_received(grpc_exec_ctx *exec_ctx, void *arg, grpc_op ops[2]; memset(ops, 0, sizeof(ops)); grpc_op *op = ops; + gpr_mu_lock(&glb_policy->mu); if (glb_policy->lb_response_payload != NULL) { gpr_backoff_reset(&glb_policy->lb_call_backoff_state); /* Received data from the LB server. Look inside * glb_policy->lb_response_payload, for a serverlist. */ grpc_byte_buffer_reader bbr; grpc_byte_buffer_reader_init(&bbr, glb_policy->lb_response_payload); - gpr_slice response_slice = grpc_byte_buffer_reader_readall(&bbr); + grpc_slice response_slice = grpc_byte_buffer_reader_readall(&bbr); grpc_byte_buffer_destroy(glb_policy->lb_response_payload); grpc_grpclb_serverlist *serverlist = grpc_grpclb_response_parse_serverlist(response_slice); if (serverlist != NULL) { GPR_ASSERT(glb_policy->lb_call != NULL); - gpr_slice_unref(response_slice); + grpc_slice_unref(response_slice); if (grpc_lb_glb_trace) { gpr_log(GPR_INFO, "Serverlist with %lu servers received", (unsigned long)serverlist->num_servers); @@ -1112,23 +1237,24 @@ static void lb_on_response_received(grpc_exec_ctx *exec_ctx, void *arg, /* update serverlist */ if (serverlist->num_servers > 0) { - gpr_mu_lock(&glb_policy->mu); if (grpc_grpclb_serverlist_equals(glb_policy->serverlist, serverlist)) { if (grpc_lb_glb_trace) { gpr_log(GPR_INFO, "Incoming server list identical to current, ignoring."); } + grpc_grpclb_destroy_serverlist(serverlist); } else { /* new serverlist */ if (glb_policy->serverlist != NULL) { /* dispose of the old serverlist */ grpc_grpclb_destroy_serverlist(glb_policy->serverlist); } - /* and update the copy in the glb_lb_policy instance */ + /* and update the copy in the glb_lb_policy instance. This serverlist + * instance will be destroyed either upon the next update or in + * glb_destroy() */ glb_policy->serverlist = serverlist; - rr_handover_locked(exec_ctx, glb_policy, error); + rr_handover_locked(exec_ctx, glb_policy); } - gpr_mu_unlock(&glb_policy->mu); } else { if (grpc_lb_glb_trace) { gpr_log(GPR_INFO, @@ -1138,8 +1264,8 @@ static void lb_on_response_received(grpc_exec_ctx *exec_ctx, void *arg, } } else { /* serverlist == NULL */ gpr_log(GPR_ERROR, "Invalid LB response received: '%s'. Ignoring.", - gpr_dump_slice(response_slice, GPR_DUMP_ASCII | GPR_DUMP_HEX)); - gpr_slice_unref(response_slice); + grpc_dump_slice(response_slice, GPR_DUMP_ASCII | GPR_DUMP_HEX)); + grpc_slice_unref(response_slice); } if (!glb_policy->shutting_down) { @@ -1156,9 +1282,11 @@ static void lb_on_response_received(grpc_exec_ctx *exec_ctx, void *arg, &glb_policy->lb_on_response_received); /* loop */ GPR_ASSERT(GRPC_CALL_OK == call_error); } + gpr_mu_unlock(&glb_policy->mu); } else { /* empty payload: call cancelled. */ /* dispose of the "lb_on_response_received" weak ref taken in * query_for_backends_locked() and reused in every reception loop */ + gpr_mu_unlock(&glb_policy->mu); GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &glb_policy->base, "lb_on_response_received_empty_payload"); } @@ -1178,7 +1306,6 @@ static void lb_call_on_retry_timer(grpc_exec_ctx *exec_ctx, void *arg, query_for_backends_locked(exec_ctx, glb_policy); } gpr_mu_unlock(&glb_policy->mu); - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &glb_policy->base, "grpclb_on_retry_timer"); } @@ -1199,7 +1326,7 @@ static void lb_on_server_status_received(grpc_exec_ctx *exec_ctx, void *arg, } /* We need to performe cleanups no matter what. */ - lb_call_destroy(glb_policy); + lb_call_destroy_locked(glb_policy); if (!glb_policy->shutting_down) { /* if we aren't shutting down, restart the LB client call after some time */ diff --git a/src/core/ext/lb_policy/grpclb/load_balancer_api.c b/src/core/ext/lb_policy/grpclb/load_balancer_api.c index a8881004a0..837e9c1113 100644 --- a/src/core/ext/lb_policy/grpclb/load_balancer_api.c +++ b/src/core/ext/lb_policy/grpclb/load_balancer_api.c @@ -90,18 +90,18 @@ grpc_grpclb_request *grpc_grpclb_request_create(const char *lb_service_name) { return req; } -gpr_slice grpc_grpclb_request_encode(const grpc_grpclb_request *request) { +grpc_slice grpc_grpclb_request_encode(const grpc_grpclb_request *request) { size_t encoded_length; pb_ostream_t sizestream; pb_ostream_t outputstream; - gpr_slice slice; + grpc_slice slice; memset(&sizestream, 0, sizeof(pb_ostream_t)); pb_encode(&sizestream, grpc_lb_v1_LoadBalanceRequest_fields, request); encoded_length = sizestream.bytes_written; - slice = gpr_slice_malloc(encoded_length); + slice = grpc_slice_malloc(encoded_length); outputstream = - pb_ostream_from_buffer(GPR_SLICE_START_PTR(slice), encoded_length); + pb_ostream_from_buffer(GRPC_SLICE_START_PTR(slice), encoded_length); GPR_ASSERT(pb_encode(&outputstream, grpc_lb_v1_LoadBalanceRequest_fields, request) != 0); return slice; @@ -113,10 +113,10 @@ void grpc_grpclb_request_destroy(grpc_grpclb_request *request) { typedef grpc_lb_v1_LoadBalanceResponse grpc_grpclb_response; grpc_grpclb_initial_response *grpc_grpclb_initial_response_parse( - gpr_slice encoded_grpc_grpclb_response) { + grpc_slice encoded_grpc_grpclb_response) { pb_istream_t stream = - pb_istream_from_buffer(GPR_SLICE_START_PTR(encoded_grpc_grpclb_response), - GPR_SLICE_LENGTH(encoded_grpc_grpclb_response)); + pb_istream_from_buffer(GRPC_SLICE_START_PTR(encoded_grpc_grpclb_response), + GRPC_SLICE_LENGTH(encoded_grpc_grpclb_response)); grpc_grpclb_response res; memset(&res, 0, sizeof(grpc_grpclb_response)); if (!pb_decode(&stream, grpc_lb_v1_LoadBalanceResponse_fields, &res)) { @@ -132,12 +132,12 @@ grpc_grpclb_initial_response *grpc_grpclb_initial_response_parse( } grpc_grpclb_serverlist *grpc_grpclb_response_parse_serverlist( - gpr_slice encoded_grpc_grpclb_response) { + grpc_slice encoded_grpc_grpclb_response) { bool status; decode_serverlist_arg arg; pb_istream_t stream = - pb_istream_from_buffer(GPR_SLICE_START_PTR(encoded_grpc_grpclb_response), - GPR_SLICE_LENGTH(encoded_grpc_grpclb_response)); + pb_istream_from_buffer(GRPC_SLICE_START_PTR(encoded_grpc_grpclb_response), + GRPC_SLICE_LENGTH(encoded_grpc_grpclb_response)); pb_istream_t stream_at_start = stream; grpc_grpclb_response res; memset(&res, 0, sizeof(grpc_grpclb_response)); diff --git a/src/core/ext/lb_policy/grpclb/load_balancer_api.h b/src/core/ext/lb_policy/grpclb/load_balancer_api.h index 079a64a3f3..b4c967e426 100644 --- a/src/core/ext/lb_policy/grpclb/load_balancer_api.h +++ b/src/core/ext/lb_policy/grpclb/load_balancer_api.h @@ -34,7 +34,7 @@ #ifndef GRPC_CORE_EXT_LB_POLICY_GRPCLB_LOAD_BALANCER_API_H #define GRPC_CORE_EXT_LB_POLICY_GRPCLB_LOAD_BALANCER_API_H -#include <grpc/support/slice_buffer.h> +#include <grpc/slice_buffer.h> #include "src/core/ext/client_channel/lb_policy_factory.h" #include "src/core/ext/lb_policy/grpclb/proto/grpc/lb/v1/load_balancer.pb.h" @@ -60,7 +60,7 @@ typedef struct grpc_grpclb_serverlist { grpc_grpclb_request *grpc_grpclb_request_create(const char *lb_service_name); /** Protocol Buffers v3-encode \a request */ -gpr_slice grpc_grpclb_request_encode(const grpc_grpclb_request *request); +grpc_slice grpc_grpclb_request_encode(const grpc_grpclb_request *request); /** Destroy \a request */ void grpc_grpclb_request_destroy(grpc_grpclb_request *request); @@ -68,11 +68,11 @@ void grpc_grpclb_request_destroy(grpc_grpclb_request *request); /** Parse (ie, decode) the bytes in \a encoded_grpc_grpclb_response as a \a * grpc_grpclb_initial_response */ grpc_grpclb_initial_response *grpc_grpclb_initial_response_parse( - gpr_slice encoded_grpc_grpclb_response); + grpc_slice encoded_grpc_grpclb_response); /** Parse the list of servers from an encoded \a grpc_grpclb_response */ grpc_grpclb_serverlist *grpc_grpclb_response_parse_serverlist( - gpr_slice encoded_grpc_grpclb_response); + grpc_slice encoded_grpc_grpclb_response); /** Return a copy of \a sl. The caller is responsible for calling \a * grpc_grpclb_destroy_serverlist on the returned copy. */ 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 ac3c6a305a..c69f773e78 100644 --- a/src/core/ext/lb_policy/pick_first/pick_first.c +++ b/src/core/ext/lb_policy/pick_first/pick_first.c @@ -292,6 +292,8 @@ static void pf_connectivity_changed(grpc_exec_ctx *exec_ctx, void *arg, } else { loop: switch (p->checking_connectivity) { + case GRPC_CHANNEL_INIT: + GPR_UNREACHABLE_CODE(return ); case GRPC_CHANNEL_READY: grpc_connectivity_state_set(exec_ctx, &p->state_tracker, GRPC_CHANNEL_READY, GRPC_ERROR_NONE, 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 427999aa6b..59f84054c4 100644 --- a/src/core/ext/lb_policy/round_robin/round_robin.c +++ b/src/core/ext/lb_policy/round_robin/round_robin.c @@ -116,8 +116,13 @@ typedef struct { grpc_closure connectivity_changed_closure; /** this subchannels current position in subchannel->ready_list */ ready_list *ready_list_node; - /** last observed connectivity */ - grpc_connectivity_state connectivity_state; + /** last observed connectivity. Not updated by + * \a grpc_subchannel_notify_on_state_change. Used to determine the previous + * state while processing the new state in \a rr_connectivity_changed */ + grpc_connectivity_state prev_connectivity_state; + /** current connectivity state. Updated by \a + * grpc_subchannel_notify_on_state_change */ + grpc_connectivity_state curr_connectivity_state; /** the subchannel's target user data */ void *user_data; /** vtable to operate over \a user_data */ @@ -127,6 +132,7 @@ typedef struct { struct round_robin_lb_policy { /** base policy: must be first */ grpc_lb_policy base; + gpr_mu mu; /** total number of addresses received at creation time */ size_t num_addresses; @@ -135,8 +141,11 @@ struct round_robin_lb_policy { size_t num_subchannels; subchannel_data **subchannels; - /** mutex protecting remaining members */ - gpr_mu mu; + /** how many subchannels are in TRANSIENT_FAILURE */ + size_t num_transient_failures; + /** how many subchannels are IDLE */ + size_t num_idle; + /** have we started picking? */ int started_picking; /** are we shutting down? */ @@ -258,6 +267,10 @@ static void remove_disconnected_sc_locked(round_robin_lb_policy *p, gpr_free(node); } +static bool is_ready_list_empty(round_robin_lb_policy *p) { + return p->ready_list.prev == NULL; +} + static void rr_destroy(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { round_robin_lb_policy *p = (round_robin_lb_policy *)pol; ready_list *elem; @@ -268,7 +281,7 @@ static void rr_destroy(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { for (size_t i = 0; i < p->num_subchannels; i++) { subchannel_data *sd = p->subchannels[i]; - GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, "round_robin_destroy"); + GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, "rr_destroy"); if (sd->user_data != NULL) { GPR_ASSERT(sd->user_data_vtable != NULL); sd->user_data_vtable->destroy(sd->user_data); @@ -381,18 +394,18 @@ static void start_picking(grpc_exec_ctx *exec_ctx, round_robin_lb_policy *p) { size_t i; p->started_picking = 1; - if (grpc_lb_round_robin_trace) { - gpr_log(GPR_DEBUG, "LB_POLICY: p=%p num_subchannels=%" PRIuPTR, (void *)p, - p->num_subchannels); - } - for (i = 0; i < p->num_subchannels; i++) { subchannel_data *sd = p->subchannels[i]; - sd->connectivity_state = GRPC_CHANNEL_IDLE; + /* use some sentinel value outside of the range of grpc_connectivity_state + * to signal an undefined previous state. We won't be referring to this + * value again and it'll be overwritten after the first call to + * rr_connectivity_changed */ + sd->prev_connectivity_state = GRPC_CHANNEL_INIT; + sd->curr_connectivity_state = GRPC_CHANNEL_IDLE; + GRPC_LB_POLICY_WEAK_REF(&p->base, "rr_connectivity"); grpc_subchannel_notify_on_state_change( exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->connectivity_state, &sd->connectivity_changed_closure); - GRPC_LB_POLICY_WEAK_REF(&p->base, "round_robin_connectivity"); + &sd->curr_connectivity_state, &sd->connectivity_changed_closure); } } @@ -422,7 +435,7 @@ static int rr_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, /* readily available, report right away */ *target = GRPC_CONNECTED_SUBCHANNEL_REF( grpc_subchannel_get_connected_subchannel(selected->subchannel), - "picked"); + "rr_picked"); if (user_data != NULL) { *user_data = selected->user_data; @@ -453,125 +466,184 @@ static int rr_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, } } +static void update_state_counters(subchannel_data *sd) { + round_robin_lb_policy *p = sd->policy; + + /* update p->num_transient_failures (resp. p->num_idle): if the previous + * state was TRANSIENT_FAILURE (resp. IDLE), decrement + * p->num_transient_failures (resp. p->num_idle). */ + if (sd->prev_connectivity_state == GRPC_CHANNEL_TRANSIENT_FAILURE) { + GPR_ASSERT(p->num_transient_failures > 0); + --p->num_transient_failures; + } else if (sd->prev_connectivity_state == GRPC_CHANNEL_IDLE) { + GPR_ASSERT(p->num_idle > 0); + --p->num_idle; + } +} + +/* sd is the subchannel_data associted with the updated subchannel. + * shutdown_error will only be used upon policy transition to TRANSIENT_FAILURE + * or SHUTDOWN */ +static grpc_connectivity_state update_lb_connectivity_status( + grpc_exec_ctx *exec_ctx, subchannel_data *sd, grpc_error *error) { + /* In priority order. The first rule to match terminates the search (ie, if we + * are on rule n, all previous rules were unfulfilled). + * + * 1) RULE: ANY subchannel is READY => policy is READY. + * CHECK: At least one subchannel is ready iff p->ready_list is NOT empty. + * + * 2) RULE: ANY subchannel is CONNECTING => policy is CONNECTING. + * CHECK: sd->curr_connectivity_state == CONNECTING. + * + * 3) RULE: ALL subchannels are SHUTDOWN => policy is SHUTDOWN. + * CHECK: p->num_subchannels = 0. + * + * 4) RULE: ALL subchannels are TRANSIENT_FAILURE => policy is + * TRANSIENT_FAILURE. + * CHECK: p->num_transient_failures == p->num_subchannels. + * + * 5) RULE: ALL subchannels are IDLE => policy is IDLE. + * CHECK: p->num_idle == p->num_subchannels. + */ + round_robin_lb_policy *p = sd->policy; + if (!is_ready_list_empty(p)) { /* 1) READY */ + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, GRPC_CHANNEL_READY, + GRPC_ERROR_NONE, "rr_ready"); + return GRPC_CHANNEL_READY; + } else if (sd->curr_connectivity_state == + GRPC_CHANNEL_CONNECTING) { /* 2) CONNECTING */ + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, + GRPC_CHANNEL_CONNECTING, GRPC_ERROR_NONE, + "rr_connecting"); + return GRPC_CHANNEL_CONNECTING; + } else if (p->num_subchannels == 0) { /* 3) SHUTDOWN */ + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, + GRPC_CHANNEL_SHUTDOWN, GRPC_ERROR_REF(error), + "rr_shutdown"); + return GRPC_CHANNEL_SHUTDOWN; + } else if (p->num_transient_failures == + p->num_subchannels) { /* 4) TRANSIENT_FAILURE */ + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, + GRPC_CHANNEL_TRANSIENT_FAILURE, + GRPC_ERROR_REF(error), "rr_transient_failure"); + return GRPC_CHANNEL_TRANSIENT_FAILURE; + } else if (p->num_idle == p->num_subchannels) { /* 5) IDLE */ + grpc_connectivity_state_set(exec_ctx, &p->state_tracker, GRPC_CHANNEL_IDLE, + GRPC_ERROR_NONE, "rr_idle"); + return GRPC_CHANNEL_IDLE; + } + /* no change */ + return sd->curr_connectivity_state; +} + static void rr_connectivity_changed(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { subchannel_data *sd = arg; round_robin_lb_policy *p = sd->policy; pending_pick *pp; - int unref = 0; - GRPC_ERROR_REF(error); gpr_mu_lock(&p->mu); if (p->shutdown) { - unref = 1; - } else { - switch (sd->connectivity_state) { - case GRPC_CHANNEL_READY: - grpc_connectivity_state_set(exec_ctx, &p->state_tracker, - GRPC_CHANNEL_READY, GRPC_ERROR_REF(error), - "connecting_ready"); - /* add the newly connected subchannel to the list of connected ones. - * Note that it goes to the "end of the line". */ - sd->ready_list_node = add_connected_sc_locked(p, sd); - /* at this point we know there's at least one suitable subchannel. Go - * ahead and pick one and notify the pending suitors in - * p->pending_picks. This preemtively replicates rr_pick()'s actions. */ - ready_list *selected = peek_next_connected_locked(p); - GPR_ASSERT(selected != NULL); - if (p->pending_picks != NULL) { - /* if the selected subchannel is going to be used for the pending - * picks, update the last picked pointer */ - advance_last_picked_locked(p); + gpr_mu_unlock(&p->mu); + GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "rr_connectivity"); + GRPC_ERROR_UNREF(error); + return; + } + switch (sd->curr_connectivity_state) { + case GRPC_CHANNEL_INIT: + GPR_UNREACHABLE_CODE(return ); + case GRPC_CHANNEL_READY: + /* add the newly connected subchannel to the list of connected ones. + * Note that it goes to the "end of the line". */ + sd->ready_list_node = add_connected_sc_locked(p, sd); + /* at this point we know there's at least one suitable subchannel. Go + * ahead and pick one and notify the pending suitors in + * p->pending_picks. This preemtively replicates rr_pick()'s actions. */ + ready_list *selected = peek_next_connected_locked(p); + GPR_ASSERT(selected != NULL); + if (p->pending_picks != NULL) { + /* if the selected subchannel is going to be used for the pending + * picks, update the last picked pointer */ + advance_last_picked_locked(p); + } + while ((pp = p->pending_picks)) { + p->pending_picks = pp->next; + *pp->target = GRPC_CONNECTED_SUBCHANNEL_REF( + grpc_subchannel_get_connected_subchannel(selected->subchannel), + "rr_picked"); + if (pp->user_data != NULL) { + *pp->user_data = selected->user_data; } - + if (grpc_lb_round_robin_trace) { + gpr_log(GPR_DEBUG, + "[RR CONN CHANGED] TARGET <-- SUBCHANNEL %p (NODE %p)", + (void *)selected->subchannel, (void *)selected); + } + grpc_exec_ctx_sched(exec_ctx, pp->on_complete, GRPC_ERROR_NONE, NULL); + gpr_free(pp); + } + update_lb_connectivity_status(exec_ctx, sd, error); + sd->prev_connectivity_state = sd->curr_connectivity_state; + /* renew notification: reuses the "rr_connectivity" weak ref */ + grpc_subchannel_notify_on_state_change( + exec_ctx, sd->subchannel, p->base.interested_parties, + &sd->curr_connectivity_state, &sd->connectivity_changed_closure); + break; + case GRPC_CHANNEL_IDLE: + ++p->num_idle; + /* fallthrough */ + case GRPC_CHANNEL_CONNECTING: + update_state_counters(sd); + update_lb_connectivity_status(exec_ctx, sd, error); + sd->prev_connectivity_state = sd->curr_connectivity_state; + /* renew notification: reuses the "rr_connectivity" weak ref */ + grpc_subchannel_notify_on_state_change( + exec_ctx, sd->subchannel, p->base.interested_parties, + &sd->curr_connectivity_state, &sd->connectivity_changed_closure); + break; + case GRPC_CHANNEL_TRANSIENT_FAILURE: + ++p->num_transient_failures; + /* remove from ready list if still present */ + if (sd->ready_list_node != NULL) { + remove_disconnected_sc_locked(p, sd->ready_list_node); + sd->ready_list_node = NULL; + } + update_lb_connectivity_status(exec_ctx, sd, error); + sd->prev_connectivity_state = sd->curr_connectivity_state; + /* renew notification: reuses the "rr_connectivity" weak ref */ + grpc_subchannel_notify_on_state_change( + exec_ctx, sd->subchannel, p->base.interested_parties, + &sd->curr_connectivity_state, &sd->connectivity_changed_closure); + break; + case GRPC_CHANNEL_SHUTDOWN: + update_state_counters(sd); + if (sd->ready_list_node != NULL) { + remove_disconnected_sc_locked(p, sd->ready_list_node); + sd->ready_list_node = NULL; + } + --p->num_subchannels; + GPR_SWAP(subchannel_data *, p->subchannels[sd->index], + p->subchannels[p->num_subchannels]); + GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, "rr_subchannel_shutdown"); + p->subchannels[sd->index]->index = sd->index; + if (update_lb_connectivity_status(exec_ctx, sd, error) == + GRPC_CHANNEL_SHUTDOWN) { + /* the policy is shutting down. Flush all the pending picks... */ while ((pp = p->pending_picks)) { p->pending_picks = pp->next; - - *pp->target = GRPC_CONNECTED_SUBCHANNEL_REF( - grpc_subchannel_get_connected_subchannel(selected->subchannel), - "picked"); - if (pp->user_data != NULL) { - *pp->user_data = selected->user_data; - } - if (grpc_lb_round_robin_trace) { - gpr_log(GPR_DEBUG, - "[RR CONN CHANGED] TARGET <-- SUBCHANNEL %p (NODE %p)", - (void *)selected->subchannel, (void *)selected); - } + *pp->target = NULL; grpc_exec_ctx_sched(exec_ctx, pp->on_complete, GRPC_ERROR_NONE, NULL); gpr_free(pp); } - grpc_subchannel_notify_on_state_change( - exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->connectivity_state, &sd->connectivity_changed_closure); - break; - case GRPC_CHANNEL_CONNECTING: - case GRPC_CHANNEL_IDLE: - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, sd->connectivity_state, - GRPC_ERROR_REF(error), "connecting_changed"); - grpc_subchannel_notify_on_state_change( - exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->connectivity_state, &sd->connectivity_changed_closure); - break; - case GRPC_CHANNEL_TRANSIENT_FAILURE: - /* renew state notification */ - grpc_subchannel_notify_on_state_change( - exec_ctx, sd->subchannel, p->base.interested_parties, - &sd->connectivity_state, &sd->connectivity_changed_closure); - - /* remove from ready list if still present */ - if (sd->ready_list_node != NULL) { - remove_disconnected_sc_locked(p, sd->ready_list_node); - sd->ready_list_node = NULL; - } - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, - GRPC_ERROR_REF(error), "connecting_transient_failure"); - break; - case GRPC_CHANNEL_SHUTDOWN: - if (sd->ready_list_node != NULL) { - remove_disconnected_sc_locked(p, sd->ready_list_node); - sd->ready_list_node = NULL; - } - - p->num_subchannels--; - GPR_SWAP(subchannel_data *, p->subchannels[sd->index], - p->subchannels[p->num_subchannels]); - GRPC_SUBCHANNEL_UNREF(exec_ctx, sd->subchannel, "round_robin"); - p->subchannels[sd->index]->index = sd->index; - gpr_free(sd); - - unref = 1; - if (p->num_subchannels == 0) { - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_SHUTDOWN, - GRPC_ERROR_CREATE_REFERENCING("Round Robin Channels Exhausted", - &error, 1), - "no_more_channels"); - while ((pp = p->pending_picks)) { - p->pending_picks = pp->next; - *pp->target = NULL; - grpc_exec_ctx_sched(exec_ctx, pp->on_complete, GRPC_ERROR_NONE, - NULL); - gpr_free(pp); - } - } else { - grpc_connectivity_state_set( - exec_ctx, &p->state_tracker, GRPC_CHANNEL_TRANSIENT_FAILURE, - GRPC_ERROR_REF(error), "subchannel_failed"); - } - } /* switch */ - } /* !unref */ - - gpr_mu_unlock(&p->mu); - - if (unref) { - GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "round_robin_connectivity"); + } + gpr_free(sd); + /* unref the "rr_connectivity" weak ref from start_picking */ + GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &p->base, "rr_connectivity"); + break; } - + gpr_mu_unlock(&p->mu); GRPC_ERROR_UNREF(error); } @@ -607,8 +679,9 @@ static void rr_ping_one(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, gpr_mu_unlock(&p->mu); target = GRPC_CONNECTED_SUBCHANNEL_REF( grpc_subchannel_get_connected_subchannel(selected->subchannel), - "picked"); + "rr_picked"); grpc_connected_subchannel_ping(exec_ctx, target, closure); + GRPC_CONNECTED_SUBCHANNEL_UNREF(exec_ctx, target, "rr_picked"); } else { gpr_mu_unlock(&p->mu); grpc_exec_ctx_sched(exec_ctx, closure, @@ -704,6 +777,11 @@ static grpc_lb_policy *round_robin_create(grpc_exec_ctx *exec_ctx, grpc_lb_policy_init(&p->base, &round_robin_lb_policy_vtable); grpc_connectivity_state_init(&p->state_tracker, GRPC_CHANNEL_IDLE, "round_robin"); + + if (grpc_lb_round_robin_trace) { + gpr_log(GPR_DEBUG, "Created RR policy at %p with %lu subchannels", + (void *)p, (unsigned long)p->num_subchannels); + } gpr_mu_init(&p->mu); return &p->base; } |