diff options
Diffstat (limited to 'src/core/ext/lb_policy/grpclb/grpclb.c')
-rw-r--r-- | src/core/ext/lb_policy/grpclb/grpclb.c | 977 |
1 files changed, 518 insertions, 459 deletions
diff --git a/src/core/ext/lb_policy/grpclb/grpclb.c b/src/core/ext/lb_policy/grpclb/grpclb.c index ae1f2a3b4c..41e177c495 100644 --- a/src/core/ext/lb_policy/grpclb/grpclb.c +++ b/src/core/ext/lb_policy/grpclb/grpclb.c @@ -43,34 +43,27 @@ * policy to select from this list of LB server backends. * * The first time the policy gets a request for a pick, a ping, or to exit the - * idle state, \a query_for_backends() is called. It creates an instance of \a - * lb_client_data, an internal struct meant to contain the data associated with - * the internal communication with the LB server. This instance is created via - * \a lb_client_data_create(). There, the call over lb_channel to pick-first - * from {a1..an} is created, the \a LoadBalancingRequest message is assembled - * and all necessary callbacks for the progress of the internal call configured. + * idle state, \a query_for_backends_locked() is called. This function sets up + * and initiates the internal communication with the LB server. In particular, + * it's responsible for instantiating the internal *streaming* call to the LB + * server (whichever address from {a1..an} pick-first chose). This call is + * serviced by two callbacks, \a lb_on_server_status_received and \a + * lb_on_response_received. The former will be called when the call to the LB + * server completes. This can happen if the LB server closes the connection or + * if this policy itself cancels the call (for example because it's shutting + * down). If the internal call times out, the usual behavior of pick-first + * applies, continuing to pick from the list {a1..an}. * - * Back in \a query_for_backends(), the internal *streaming* call to the LB - * server (whichever address from {a1..an} pick-first chose) is kicked off. - * It'll progress over the callbacks configured in \a lb_client_data_create() - * (see the field docstrings of \a lb_client_data for more details). - * - * If the call fails with UNIMPLEMENTED, the original call will also fail. - * There's a misconfiguration somewhere: at least one of {a1..an} isn't a LB - * server, which contradicts the LB bit being set. If the internal call times - * out, the usual behavior of pick-first applies, continuing to pick from the - * list {a1..an}. - * - * Upon sucesss, a \a LoadBalancingResponse is expected in \a res_recv_cb. An - * invalid one results in the termination of the streaming call. A new streaming - * call should be created if possible, failing the original call otherwise. - * For a valid \a LoadBalancingResponse, the server list of actual backends is - * extracted. A Round Robin policy will be created from this list. There are two - * possible scenarios: + * Upon sucesss, the incoming \a LoadBalancingResponse is processed by \a + * res_recv. An invalid one results in the termination of the streaming call. A + * new streaming call should be created if possible, failing the original call + * otherwise. For a valid \a LoadBalancingResponse, the server list of actual + * backends is extracted. A Round Robin policy will be created from this list. + * There are two possible scenarios: * * 1. This is the first server list received. There was no previous instance of - * the Round Robin policy. \a rr_handover() will instantiate the RR policy - * and perform all the pending operations over it. + * the Round Robin policy. \a rr_handover_locked() will instantiate the RR + * policy and perform all the pending operations over it. * 2. There's already a RR policy instance active. We need to introduce the new * one build from the new serverlist, but taking care not to disrupt the * operations in progress over the old RR instance. This is done by @@ -78,16 +71,16 @@ * references are held on the old RR policy, it'll be destroyed and \a * glb_rr_connectivity_changed notified with a \a GRPC_CHANNEL_SHUTDOWN * state. At this point we can transition to a new RR instance safely, which - * is done once again via \a rr_handover(). + * is done once again via \a rr_handover_locked(). * * * Once a RR policy instance is in place (and getting updated as described), * calls to for a pick, a ping or a cancellation will be serviced right away by * forwarding them to the RR instance. Any time there's no RR policy available - * (ie, right after the creation of the gRPCLB policy, if an empty serverlist - * is received, etc), pick/ping requests are added to a list of pending - * picks/pings to be flushed and serviced as part of \a rr_handover() the moment - * the RR policy instance becomes available. + * (ie, right after the creation of the gRPCLB policy, if an empty serverlist is + * received, etc), pick/ping requests are added to a list of pending picks/pings + * to be flushed and serviced as part of \a rr_handover_locked() the moment the + * RR policy instance becomes available. * * \see https://github.com/grpc/grpc/blob/master/doc/load-balancing.md for the * high level design and details. */ @@ -96,6 +89,12 @@ * - Implement LB service forwarding (point 2c. in the doc's diagram). */ +/* With the addition of a libuv endpoint, sockaddr.h now includes uv.h when + using that endpoint. Because of various transitive includes in uv.h, + including windows.h on Windows, uv.h must be included before other system + headers. Therefore, sockaddr.h must always be included first */ +#include "src/core/lib/iomgr/sockaddr.h" + #include <errno.h> #include <string.h> @@ -107,19 +106,28 @@ #include <grpc/support/string_util.h> #include <grpc/support/time.h> -#include "src/core/ext/client_config/client_channel_factory.h" -#include "src/core/ext/client_config/lb_policy_factory.h" -#include "src/core/ext/client_config/lb_policy_registry.h" -#include "src/core/ext/client_config/parse_address.h" +#include "src/core/ext/client_channel/client_channel_factory.h" +#include "src/core/ext/client_channel/lb_policy_factory.h" +#include "src/core/ext/client_channel/lb_policy_registry.h" +#include "src/core/ext/client_channel/parse_address.h" #include "src/core/ext/lb_policy/grpclb/grpclb.h" #include "src/core/ext/lb_policy/grpclb/load_balancer_api.h" +#include "src/core/lib/channel/channel_args.h" #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 + int grpc_lb_glb_trace = 0; /* add lb_token of selected subchannel (address) to the call's initial @@ -134,6 +142,9 @@ static void initial_metadata_add_lb_token( } typedef struct wrapped_rr_closure_arg { + /* the closure instance using this struct as argument */ + grpc_closure wrapper_closure; + /* the original closure. Usually a on_complete/notify cb for pick() and ping() * calls against the internal RR instance, respectively. */ grpc_closure *wrapped_closure; @@ -155,9 +166,8 @@ typedef struct wrapped_rr_closure_arg { /* The RR instance related to the closure */ grpc_lb_policy *rr_policy; - /* when not NULL, represents a pending_{pick,ping} node to be freed upon - * closure execution */ - void *owning_pending_node; /* to be freed if not NULL */ + /* heap memory to be freed upon closure execution. */ + void *free_when_done; } wrapped_rr_closure_arg; /* The \a on_complete closure passed as part of the pick requires keeping a @@ -166,13 +176,12 @@ typedef struct wrapped_rr_closure_arg { static void wrapped_rr_closure(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { wrapped_rr_closure_arg *wc_arg = arg; - if (wc_arg->rr_policy != NULL) { - if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "Unreffing RR (0x%" PRIxPTR ")", - (intptr_t)wc_arg->rr_policy); - } - GRPC_LB_POLICY_UNREF(exec_ctx, wc_arg->rr_policy, "wrapped_rr_closure"); + GPR_ASSERT(wc_arg->wrapped_closure != NULL); + grpc_exec_ctx_sched(exec_ctx, wc_arg->wrapped_closure, GRPC_ERROR_REF(error), + NULL); + + if (wc_arg->rr_policy != NULL) { /* 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 */ @@ -181,12 +190,14 @@ static void wrapped_rr_closure(grpc_exec_ctx *exec_ctx, void *arg, wc_arg->lb_token_mdelem_storage, GRPC_MDELEM_REF(wc_arg->lb_token)); } + if (grpc_lb_glb_trace) { + gpr_log(GPR_INFO, "Unreffing RR (0x%" PRIxPTR ")", + (intptr_t)wc_arg->rr_policy); + } + GRPC_LB_POLICY_UNREF(exec_ctx, wc_arg->rr_policy, "wrapped_rr_closure"); } - GPR_ASSERT(wc_arg->wrapped_closure != NULL); - - grpc_exec_ctx_sched(exec_ctx, wc_arg->wrapped_closure, GRPC_ERROR_REF(error), - NULL); - gpr_free(wc_arg->owning_pending_node); + GPR_ASSERT(wc_arg->free_when_done != NULL); + gpr_free(wc_arg->free_when_done); } /* Linked list of pending pick requests. It stores all information needed to @@ -207,10 +218,6 @@ typedef struct pending_pick { * upon error. */ grpc_connected_subchannel **target; - /* a closure wrapping the original on_complete one to be invoked once the - * pick() has completed (regardless of success) */ - grpc_closure wrapped_on_complete; - /* args for wrapped_on_complete */ wrapped_rr_closure_arg wrapped_on_complete_arg; } pending_pick; @@ -230,8 +237,9 @@ static void add_pending_pick(pending_pick **root, pp->wrapped_on_complete_arg.initial_metadata = pick_args->initial_metadata; pp->wrapped_on_complete_arg.lb_token_mdelem_storage = pick_args->lb_token_mdelem_storage; - grpc_closure_init(&pp->wrapped_on_complete, wrapped_rr_closure, - &pp->wrapped_on_complete_arg); + pp->wrapped_on_complete_arg.free_when_done = pp; + grpc_closure_init(&pp->wrapped_on_complete_arg.wrapper_closure, + wrapped_rr_closure, &pp->wrapped_on_complete_arg); *root = pp; } @@ -239,10 +247,6 @@ static void add_pending_pick(pending_pick **root, typedef struct pending_ping { struct pending_ping *next; - /* a closure wrapping the original on_complete one to be invoked once the - * ping() has completed (regardless of success) */ - grpc_closure wrapped_notify; - /* args for wrapped_notify */ wrapped_rr_closure_arg wrapped_notify_arg; } pending_ping; @@ -251,10 +255,11 @@ static void add_pending_ping(pending_ping **root, grpc_closure *notify) { pending_ping *pping = gpr_malloc(sizeof(*pping)); memset(pping, 0, sizeof(pending_ping)); memset(&pping->wrapped_notify_arg, 0, sizeof(wrapped_rr_closure_arg)); - pping->next = *root; - grpc_closure_init(&pping->wrapped_notify, wrapped_rr_closure, - &pping->wrapped_notify_arg); pping->wrapped_notify_arg.wrapped_closure = notify; + pping->wrapped_notify_arg.free_when_done = pping; + pping->next = *root; + grpc_closure_init(&pping->wrapped_notify_arg.wrapper_closure, + wrapped_rr_closure, &pping->wrapped_notify_arg); *root = pping; } @@ -262,7 +267,6 @@ static void add_pending_ping(pending_ping **root, grpc_closure *notify) { * glb_lb_policy */ typedef struct rr_connectivity_data rr_connectivity_data; -struct lb_client_data; static const grpc_lb_policy_vtable glb_lb_policy_vtable; typedef struct glb_lb_policy { /** base policy: must be first */ @@ -274,6 +278,7 @@ typedef struct glb_lb_policy { /** who the client is trying to communicate with */ const char *server_name; grpc_client_channel_factory *cc_factory; + grpc_channel_args *args; /** deadline for the LB's call */ gpr_timespec deadline; @@ -293,27 +298,47 @@ typedef struct glb_lb_policy { * response has arrived. */ grpc_grpclb_serverlist *serverlist; - /** addresses from \a serverlist */ - grpc_lb_addresses *addresses; - /** list of picks that are waiting on RR's policy connectivity */ pending_pick *pending_picks; /** list of pings that are waiting on RR's policy connectivity */ pending_ping *pending_pings; - /** client data associated with the LB server communication */ - struct lb_client_data *lb_client; + bool shutting_down; + + /************************************************************/ + /* client data associated with the LB server communication */ + /************************************************************/ + /* Status from the LB server has been received. This signals the end of the LB + * call. */ + grpc_closure lb_on_server_status_received; + + /* A response from the LB server has been received. Process it */ + grpc_closure lb_on_response_received; + + grpc_call *lb_call; /* streaming call to the LB server, */ + + grpc_metadata_array lb_initial_metadata_recv; /* initial MD from LB server */ + grpc_metadata_array + lb_trailing_metadata_recv; /* trailing MD from LB server */ + + /* what's being sent to the LB server. Note that its value may vary if the LB + * server indicates a redirect. */ + grpc_byte_buffer *lb_request_payload; + + /* response the LB server, if any. Processed in lb_on_response_received() */ + grpc_byte_buffer *lb_response_payload; - /** for tracking of the RR connectivity */ - rr_connectivity_data *rr_connectivity; + /* call status code and details, set in lb_on_server_status_received() */ + grpc_status_code lb_call_status; + char *lb_call_status_details; + size_t lb_call_status_details_capacity; - /* a wrapped (see \a wrapped_rr_closure) on-complete closure for readily - * available RR picks */ - grpc_closure wrapped_on_complete; + /** LB call retry backoff state */ + gpr_backoff lb_call_backoff_state; - /* arguments for the wrapped_on_complete closure */ - wrapped_rr_closure_arg wc_arg; + /** LB call retry timer */ + grpc_timer lb_call_retry_timer; } glb_lb_policy; /* Keeps track and reacts to changes in connectivity of the RR instance */ @@ -329,8 +354,8 @@ static bool is_server_valid(const grpc_grpclb_server *server, size_t idx, if (server->port >> 16 != 0) { if (log) { gpr_log(GPR_ERROR, - "Invalid port '%d' at index %zu of serverlist. Ignoring.", - server->port, idx); + "Invalid port '%d' at index %lu of serverlist. Ignoring.", + server->port, (unsigned long)idx); } return false; } @@ -338,15 +363,52 @@ static bool is_server_valid(const grpc_grpclb_server *server, size_t idx, if (ip->size != 4 && ip->size != 16) { if (log) { gpr_log(GPR_ERROR, - "Expected IP to be 4 or 16 bytes, got %d at index %zu of " + "Expected IP to be 4 or 16 bytes, got %d at index %lu of " "serverlist. Ignoring", - ip->size, idx); + ip->size, (unsigned long)idx); } return false; } return true; } +/* vtable for LB tokens in grpc_lb_addresses. */ +static void *lb_token_copy(void *token) { + return token == NULL ? NULL : GRPC_MDELEM_REF(token); +} +static void lb_token_destroy(void *token) { + if (token != NULL) GRPC_MDELEM_UNREF(token); +} +static int lb_token_cmp(void *token1, void *token2) { + if (token1 > token2) return 1; + if (token1 < token2) return -1; + return 0; +} +static const grpc_lb_user_data_vtable lb_token_vtable = { + lb_token_copy, lb_token_destroy, lb_token_cmp}; + +static void parse_server(const grpc_grpclb_server *server, + grpc_resolved_address *addr) { + const uint16_t netorder_port = htons((uint16_t)server->port); + /* the addresses are given in binary format (a in(6)_addr struct) in + * server->ip_address.bytes. */ + const grpc_grpclb_ip_address *ip = &server->ip_address; + memset(addr, 0, sizeof(*addr)); + if (ip->size == 4) { + addr->len = sizeof(struct sockaddr_in); + struct sockaddr_in *addr4 = (struct sockaddr_in *)&addr->addr; + addr4->sin_family = AF_INET; + memcpy(&addr4->sin_addr, ip->bytes, ip->size); + addr4->sin_port = netorder_port; + } else if (ip->size == 16) { + addr->len = sizeof(struct sockaddr_in6); + struct sockaddr_in6 *addr6 = (struct sockaddr_in6 *)&addr->addr; + addr6->sin6_family = AF_INET6; + memcpy(&addr6->sin6_addr, ip->bytes, ip->size); + addr6->sin6_port = netorder_port; + } +} + /* Returns addresses extracted from \a serverlist. */ static grpc_lb_addresses *process_serverlist( const grpc_grpclb_serverlist *serverlist) { @@ -358,7 +420,8 @@ static grpc_lb_addresses *process_serverlist( } if (num_valid == 0) return NULL; - grpc_lb_addresses *lb_addresses = grpc_lb_addresses_create(num_valid); + grpc_lb_addresses *lb_addresses = + grpc_lb_addresses_create(num_valid, &lb_token_vtable); /* second pass: actually populate the addresses and LB tokens (aka user data * to the outside world) to be read by the RR policy during its creation. @@ -372,41 +435,26 @@ static grpc_lb_addresses *process_serverlist( if (!is_server_valid(serverlist->servers[sl_idx], sl_idx, false)) continue; /* address processing */ - const uint16_t netorder_port = htons((uint16_t)server->port); - /* the addresses are given in binary format (a in(6)_addr struct) in - * server->ip_address.bytes. */ - const grpc_grpclb_ip_address *ip = &server->ip_address; grpc_resolved_address addr; - memset(&addr, 0, sizeof(addr)); - if (ip->size == 4) { - addr.len = sizeof(struct sockaddr_in); - struct sockaddr_in *addr4 = (struct sockaddr_in *)&addr.addr; - addr4->sin_family = AF_INET; - memcpy(&addr4->sin_addr, ip->bytes, ip->size); - addr4->sin_port = netorder_port; - } else if (ip->size == 16) { - addr.len = sizeof(struct sockaddr_in6); - struct sockaddr_in6 *addr6 = (struct sockaddr_in6 *)&addr.addr; - addr6->sin6_family = AF_INET; - memcpy(&addr6->sin6_addr, ip->bytes, ip->size); - addr6->sin6_port = netorder_port; - } + parse_server(server, &addr); /* lb token processing */ void *user_data; if (server->has_load_balance_token) { - const size_t lb_token_size = - GPR_ARRAY_SIZE(server->load_balance_token) - 1; + const size_t lb_token_max_length = + GPR_ARRAY_SIZE(server->load_balance_token); + const size_t lb_token_length = + strnlen(server->load_balance_token, lb_token_max_length); 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); + (uint8_t *)server->load_balance_token, lb_token_length); + 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; + grpc_sockaddr_to_uri(&addr)); + user_data = GRPC_MDELEM_LB_TOKEN_EMPTY; } grpc_lb_addresses_set_address(lb_addresses, addr_idx, &addr.addr, addr.len, @@ -415,57 +463,111 @@ static grpc_lb_addresses *process_serverlist( ++addr_idx; } GPR_ASSERT(addr_idx == num_valid); - return lb_addresses; } -/* A plugin for grpc_lb_addresses_destroy that unrefs the LB token metadata. */ -static void lb_token_destroy(void *token) { - if (token != NULL) GRPC_MDELEM_UNREF(token); +/* 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 */ +static bool pick_from_internal_rr_locked( + grpc_exec_ctx *exec_ctx, grpc_lb_policy *rr_policy, + const grpc_lb_policy_pick_args *pick_args, + grpc_connected_subchannel **target, wrapped_rr_closure_arg *wc_arg) { + GPR_ASSERT(rr_policy != NULL); + const bool pick_done = + grpc_lb_policy_pick(exec_ctx, rr_policy, pick_args, target, + (void **)&wc_arg->lb_token, &wc_arg->wrapper_closure); + if (pick_done) { + /* synchronous grpc_lb_policy_pick call. Unref the RR policy. */ + if (grpc_lb_glb_trace) { + gpr_log(GPR_INFO, "Unreffing RR (0x%" PRIxPTR ")", + (intptr_t)wc_arg->rr_policy); + } + GRPC_LB_POLICY_UNREF(exec_ctx, wc_arg->rr_policy, "glb_pick_sync"); + + /* add the load reporting initial metadata */ + initial_metadata_add_lb_token(pick_args->initial_metadata, + pick_args->lb_token_mdelem_storage, + GRPC_MDELEM_REF(wc_arg->lb_token)); + + gpr_free(wc_arg); + } + /* else, the pending pick will be registered and taken care of by the + * pending pick list inside the RR policy (glb_policy->rr_policy). + * Eventually, wrapped_on_complete will be called, which will -among other + * things- add the LB token to the call's initial metadata */ + return pick_done; } -static grpc_lb_policy *create_rr(grpc_exec_ctx *exec_ctx, - const grpc_grpclb_serverlist *serverlist, - glb_lb_policy *glb_policy) { +static grpc_lb_policy *create_rr_locked( + grpc_exec_ctx *exec_ctx, const grpc_grpclb_serverlist *serverlist, + glb_lb_policy *glb_policy) { GPR_ASSERT(serverlist != NULL && serverlist->num_servers > 0); grpc_lb_policy_args args; memset(&args, 0, sizeof(args)); - args.server_name = glb_policy->server_name; args.client_channel_factory = glb_policy->cc_factory; - args.addresses = process_serverlist(serverlist); - - grpc_lb_policy *rr = grpc_lb_policy_create(exec_ctx, "round_robin", &args); + grpc_lb_addresses *addresses = process_serverlist(serverlist); - if (glb_policy->addresses != NULL) { - /* dispose of the previous version */ - grpc_lb_addresses_destroy(glb_policy->addresses, lb_token_destroy); - } - glb_policy->addresses = args.addresses; + // Replace the LB addresses in the channel args that we pass down to + // the subchannel. + static const char *keys_to_remove[] = {GRPC_ARG_LB_ADDRESSES}; + const grpc_arg arg = grpc_lb_addresses_create_channel_arg(addresses); + args.args = grpc_channel_args_copy_and_add_and_remove( + glb_policy->args, keys_to_remove, GPR_ARRAY_SIZE(keys_to_remove), &arg, + 1); + grpc_lb_policy *rr = grpc_lb_policy_create(exec_ctx, "round_robin", &args); + GPR_ASSERT(rr != NULL); + grpc_lb_addresses_destroy(addresses); + grpc_channel_args_destroy(args.args); return rr; } -static void rr_handover(grpc_exec_ctx *exec_ctx, glb_lb_policy *glb_policy, - grpc_error *error) { +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) { GPR_ASSERT(glb_policy->serverlist != NULL && glb_policy->serverlist->num_servers > 0); - glb_policy->rr_policy = - create_rr(exec_ctx, glb_policy->serverlist, glb_policy); if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "Created RR policy (0x%" PRIxPTR ")", - (intptr_t)glb_policy->rr_policy); + gpr_log(GPR_INFO, "RR handover. Old RR: %p", (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); + } + GPR_ASSERT(glb_policy->rr_policy != NULL); - glb_policy->rr_connectivity->state = grpc_lb_policy_check_connectivity( + grpc_pollset_set_add_pollset_set(exec_ctx, + glb_policy->rr_policy->interested_parties, + glb_policy->base.interested_parties); + + 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); - grpc_lb_policy_notify_on_state_change( - exec_ctx, glb_policy->rr_policy, &glb_policy->rr_connectivity->state, - &glb_policy->rr_connectivity->on_change); + grpc_connectivity_state_set(exec_ctx, &glb_policy->state_tracker, - glb_policy->rr_connectivity->state, - GRPC_ERROR_REF(error), "rr_handover"); + rr_connectivity->state, GRPC_ERROR_REF(error), + "rr_handover"); + /* subscribe */ + 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 */ @@ -478,11 +580,9 @@ static void rr_handover(grpc_exec_ctx *exec_ctx, glb_lb_policy *glb_policy, gpr_log(GPR_INFO, "Pending pick about to PICK from 0x%" PRIxPTR "", (intptr_t)glb_policy->rr_policy); } - grpc_lb_policy_pick(exec_ctx, glb_policy->rr_policy, &pp->pick_args, - pp->target, - (void **)&pp->wrapped_on_complete_arg.lb_token, - &pp->wrapped_on_complete); - pp->wrapped_on_complete_arg.owning_pending_node = pp; + pick_from_internal_rr_locked(exec_ctx, glb_policy->rr_policy, + &pp->pick_args, pp->target, + &pp->wrapped_on_complete_arg); } pending_ping *pping; @@ -495,44 +595,45 @@ static void rr_handover(grpc_exec_ctx *exec_ctx, glb_lb_policy *glb_policy, (intptr_t)glb_policy->rr_policy); } grpc_lb_policy_ping_one(exec_ctx, glb_policy->rr_policy, - &pping->wrapped_notify); - pping->wrapped_notify_arg.owning_pending_node = pping; + &pping->wrapped_notify_arg.wrapper_closure); } } 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) { - if (glb_policy->serverlist != NULL) { - /* a RR policy is shutting down but there's a serverlist available -> - * perform a handover */ - rr_handover(exec_ctx, glb_policy, error); - } else { - /* shutting down and no new serverlist available. Bail out. */ - gpr_free(rr_conn_data); - } + 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. */ + 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 { - if (error == GRPC_ERROR_NONE) { - /* 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), - "glb_rr_connectivity_changed"); - /* resubscribe */ - grpc_lb_policy_notify_on_state_change(exec_ctx, glb_policy->rr_policy, - &rr_conn_data->state, - &rr_conn_data->on_change); - } else { /* error */ - gpr_free(rr_conn_data); - } + GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &glb_policy->base, + "rr_connectivity_cb"); + gpr_free(rr_conn_data); } } static grpc_lb_policy *glb_create(grpc_exec_ctx *exec_ctx, grpc_lb_policy_factory *factory, grpc_lb_policy_args *args) { + /* Get server name. */ + const grpc_arg *arg = + grpc_channel_args_find(args->args, GRPC_ARG_SERVER_NAME); + const char *server_name = + arg != NULL && arg->type == GRPC_ARG_STRING ? arg->value.string : NULL; + /* Count the number of gRPC-LB addresses. There must be at least one. * TODO(roth): For now, we ignore non-balancer addresses, but in the * future, we may change the behavior such that we fall back to using @@ -540,23 +641,27 @@ static grpc_lb_policy *glb_create(grpc_exec_ctx *exec_ctx, * time, this should be changed to allow a list with no balancer addresses, * since the resolver might fail to return a balancer address even when * this is the right LB policy to use. */ + arg = grpc_channel_args_find(args->args, GRPC_ARG_LB_ADDRESSES); + GPR_ASSERT(arg != NULL && arg->type == GRPC_ARG_POINTER); + grpc_lb_addresses *addresses = arg->value.pointer.p; size_t num_grpclb_addrs = 0; - for (size_t i = 0; i < args->addresses->num_addresses; ++i) { - if (args->addresses->addresses[i].is_balancer) ++num_grpclb_addrs; + for (size_t i = 0; i < addresses->num_addresses; ++i) { + if (addresses->addresses[i].is_balancer) ++num_grpclb_addrs; } if (num_grpclb_addrs == 0) return NULL; glb_lb_policy *glb_policy = gpr_malloc(sizeof(*glb_policy)); memset(glb_policy, 0, sizeof(*glb_policy)); - /* All input addresses in args->addresses come from a resolver that claims + /* All input addresses in addresses come from a resolver that claims * they are LB services. It's the resolver's responsibility to make sure * this * policy is only instantiated and used in that case. * * Create a client channel over them to communicate with a LB service */ - glb_policy->server_name = gpr_strdup(args->server_name); + glb_policy->server_name = gpr_strdup(server_name); glb_policy->cc_factory = args->client_channel_factory; + glb_policy->args = grpc_channel_args_copy(args->args); GPR_ASSERT(glb_policy->cc_factory != NULL); /* construct a target from the addresses in args, given in the form @@ -564,22 +669,19 @@ static grpc_lb_policy *glb_create(grpc_exec_ctx *exec_ctx, * TODO(dgq): support mixed ip version */ char **addr_strs = gpr_malloc(sizeof(char *) * num_grpclb_addrs); size_t addr_index = 0; - for (size_t i = 0; i < args->addresses->num_addresses; i++) { - if (args->addresses->addresses[i].user_data != NULL) { + for (size_t i = 0; i < addresses->num_addresses; i++) { + if (addresses->addresses[i].user_data != NULL) { gpr_log(GPR_ERROR, "This LB policy doesn't support user data. It will be ignored"); } - if (args->addresses->addresses[i].is_balancer) { + if (addresses->addresses[i].is_balancer) { if (addr_index == 0) { - addr_strs[addr_index++] = grpc_sockaddr_to_uri( - (const struct sockaddr *)&args->addresses->addresses[i] - .address.addr); + addr_strs[addr_index++] = + grpc_sockaddr_to_uri(&addresses->addresses[i].address); } else { - GPR_ASSERT(grpc_sockaddr_to_string( - &addr_strs[addr_index++], - (const struct sockaddr *)&args->addresses->addresses[i] - .address.addr, - true) > 0); + GPR_ASSERT(grpc_sockaddr_to_string(&addr_strs[addr_index++], + &addresses->addresses[i].address, + true) > 0); } } } @@ -587,10 +689,29 @@ static grpc_lb_policy *glb_create(grpc_exec_ctx *exec_ctx, char *target_uri_str = gpr_strjoin_sep((const char **)addr_strs, num_grpclb_addrs, ",", &uri_path_len); - /* will pick using pick_first */ + /* Create a channel to talk to the LBs. + * + * We strip out the channel arg for the LB policy name, since we want + * to use the default (pick_first) in this case. + * + * We also strip out the channel arg for the resolved addresses, since + * that will be generated by the name resolver used in the LB channel. + * Note that the LB channel will use the sockaddr resolver, so this + * won't actually generate a query to DNS (or some other name service). + * However, the addresses returned by the sockaddr resolver will have + * is_balancer=false, whereas our own addresses have is_balancer=true. + * We need the LB channel to return addresses with is_balancer=false + * so that it does not wind up recursively using the grpclb LB policy, + * as per the special case logic in client_channel.c. + */ + static const char *keys_to_remove[] = {GRPC_ARG_LB_POLICY_NAME, + GRPC_ARG_LB_ADDRESSES}; + grpc_channel_args *new_args = grpc_channel_args_copy_and_remove( + args->args, keys_to_remove, GPR_ARRAY_SIZE(keys_to_remove)); glb_policy->lb_channel = grpc_client_channel_factory_create_channel( exec_ctx, glb_policy->cc_factory, target_uri_str, - GRPC_CLIENT_CHANNEL_TYPE_LOAD_BALANCING, NULL); + GRPC_CLIENT_CHANNEL_TYPE_LOAD_BALANCING, new_args); + grpc_channel_args_destroy(new_args); gpr_free(target_uri_str); for (size_t i = 0; i < num_grpclb_addrs; i++) { @@ -603,18 +724,11 @@ static grpc_lb_policy *glb_create(grpc_exec_ctx *exec_ctx, return NULL; } - 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; - glb_policy->rr_connectivity = rr_connectivity; - grpc_lb_policy_init(&glb_policy->base, &glb_lb_policy_vtable); gpr_mu_init(&glb_policy->mu); grpc_connectivity_state_init(&glb_policy->state_tracker, GRPC_CHANNEL_IDLE, "grpclb"); + return &glb_policy->base; } @@ -623,6 +737,7 @@ static void glb_destroy(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { GPR_ASSERT(glb_policy->pending_picks == NULL); GPR_ASSERT(glb_policy->pending_pings == NULL); gpr_free((void *)glb_policy->server_name); + grpc_channel_args_destroy(glb_policy->args); grpc_channel_destroy(glb_policy->lb_channel); glb_policy->lb_channel = NULL; grpc_connectivity_state_destroy(exec_ctx, &glb_policy->state_tracker); @@ -630,50 +745,46 @@ static void glb_destroy(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { grpc_grpclb_destroy_serverlist(glb_policy->serverlist); } gpr_mu_destroy(&glb_policy->mu); - grpc_lb_addresses_destroy(glb_policy->addresses, lb_token_destroy); gpr_free(glb_policy); } -static void lb_client_data_destroy(struct lb_client_data *lb_client); static void glb_shutdown(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { glb_lb_policy *glb_policy = (glb_lb_policy *)pol; gpr_mu_lock(&glb_policy->mu); + glb_policy->shutting_down = true; pending_pick *pp = glb_policy->pending_picks; 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"); + } + 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 cancel and clean up */ + } + } + grpc_connectivity_state_set( + exec_ctx, &glb_policy->state_tracker, GRPC_CHANNEL_SHUTDOWN, + GRPC_ERROR_CREATE("Channel Shutdown"), "glb_shutdown"); gpr_mu_unlock(&glb_policy->mu); while (pp != NULL) { pending_pick *next = pp->next; *pp->target = NULL; - grpc_exec_ctx_sched(exec_ctx, &pp->wrapped_on_complete, GRPC_ERROR_NONE, - NULL); + grpc_exec_ctx_sched(exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure, + GRPC_ERROR_NONE, NULL); pp = next; } while (pping != NULL) { pending_ping *next = pping->next; - grpc_exec_ctx_sched(exec_ctx, &pping->wrapped_notify, GRPC_ERROR_NONE, - NULL); + grpc_exec_ctx_sched(exec_ctx, &pping->wrapped_notify_arg.wrapper_closure, + GRPC_ERROR_NONE, NULL); pping = next; } - - if (glb_policy->rr_policy) { - /* unsubscribe */ - grpc_lb_policy_notify_on_state_change( - exec_ctx, glb_policy->rr_policy, NULL, - &glb_policy->rr_connectivity->on_change); - GRPC_LB_POLICY_UNREF(exec_ctx, glb_policy->rr_policy, "glb_shutdown"); - } - - lb_client_data_destroy(glb_policy->lb_client); - glb_policy->lb_client = NULL; - - 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, @@ -686,11 +797,9 @@ 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, + exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure, GRPC_ERROR_CREATE_REFERENCING("Pick Cancelled", &error, 1), NULL); } else { pp->next = glb_policy->pending_picks; @@ -702,27 +811,20 @@ static void glb_cancel_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, GRPC_ERROR_UNREF(error); } -static grpc_call *lb_client_data_get_call(struct lb_client_data *lb_client); static void glb_cancel_picks(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, uint32_t initial_metadata_flags_mask, uint32_t initial_metadata_flags_eq, grpc_error *error) { glb_lb_policy *glb_policy = (glb_lb_policy *)pol; gpr_mu_lock(&glb_policy->mu); - if (glb_policy->lb_client != NULL) { - /* cancel the call to the load balancer service, if any */ - grpc_call_cancel(lb_client_data_get_call(glb_policy->lb_client), NULL); - } pending_pick *pp = glb_policy->pending_picks; glb_policy->pending_picks = NULL; while (pp != NULL) { 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, + exec_ctx, &pp->wrapped_on_complete_arg.wrapper_closure, GRPC_ERROR_CREATE_REFERENCING("Pick Cancelled", &error, 1), NULL); } else { pp->next = glb_policy->pending_picks; @@ -734,18 +836,20 @@ static void glb_cancel_picks(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, GRPC_ERROR_UNREF(error); } -static void query_for_backends(grpc_exec_ctx *exec_ctx, - glb_lb_policy *glb_policy); -static void start_picking(grpc_exec_ctx *exec_ctx, glb_lb_policy *glb_policy) { +static void query_for_backends_locked(grpc_exec_ctx *exec_ctx, + glb_lb_policy *glb_policy); +static void start_picking_locked(grpc_exec_ctx *exec_ctx, + glb_lb_policy *glb_policy) { glb_policy->started_picking = true; - query_for_backends(exec_ctx, glb_policy); + gpr_backoff_reset(&glb_policy->lb_call_backoff_state); + query_for_backends_locked(exec_ctx, glb_policy); } static void glb_exit_idle(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol) { glb_lb_policy *glb_policy = (glb_lb_policy *)pol; gpr_mu_lock(&glb_policy->mu); if (!glb_policy->started_picking) { - start_picking(exec_ctx, glb_policy); + start_picking_locked(exec_ctx, glb_policy); } gpr_mu_unlock(&glb_policy->mu); } @@ -771,48 +875,35 @@ static int glb_pick(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, if (glb_policy->rr_policy != NULL) { if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "about to PICK from 0x%" PRIxPTR "", - (intptr_t)glb_policy->rr_policy); + gpr_log(GPR_INFO, "grpclb %p about to PICK from RR %p", + (void *)glb_policy, (void *)glb_policy->rr_policy); } GRPC_LB_POLICY_REF(glb_policy->rr_policy, "glb_pick"); - memset(&glb_policy->wc_arg, 0, sizeof(wrapped_rr_closure_arg)); - glb_policy->wc_arg.rr_policy = glb_policy->rr_policy; - glb_policy->wc_arg.target = target; - glb_policy->wc_arg.wrapped_closure = on_complete; - glb_policy->wc_arg.lb_token_mdelem_storage = - pick_args->lb_token_mdelem_storage; - glb_policy->wc_arg.initial_metadata = pick_args->initial_metadata; - glb_policy->wc_arg.owning_pending_node = NULL; - grpc_closure_init(&glb_policy->wrapped_on_complete, wrapped_rr_closure, - &glb_policy->wc_arg); - - pick_done = - grpc_lb_policy_pick(exec_ctx, glb_policy->rr_policy, pick_args, target, - (void **)&glb_policy->wc_arg.lb_token, - &glb_policy->wrapped_on_complete); - if (pick_done) { - /* synchronous grpc_lb_policy_pick call. Unref the RR policy. */ - if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "Unreffing RR (0x%" PRIxPTR ")", - (intptr_t)glb_policy->wc_arg.rr_policy); - } - GRPC_LB_POLICY_UNREF(exec_ctx, glb_policy->wc_arg.rr_policy, "glb_pick"); - /* add the load reporting initial metadata */ - initial_metadata_add_lb_token( - pick_args->initial_metadata, pick_args->lb_token_mdelem_storage, - GRPC_MDELEM_REF(glb_policy->wc_arg.lb_token)); - } + wrapped_rr_closure_arg *wc_arg = gpr_malloc(sizeof(wrapped_rr_closure_arg)); + memset(wc_arg, 0, sizeof(wrapped_rr_closure_arg)); + + grpc_closure_init(&wc_arg->wrapper_closure, wrapped_rr_closure, wc_arg); + wc_arg->rr_policy = glb_policy->rr_policy; + wc_arg->target = target; + wc_arg->wrapped_closure = on_complete; + wc_arg->lb_token_mdelem_storage = pick_args->lb_token_mdelem_storage; + wc_arg->initial_metadata = pick_args->initial_metadata; + wc_arg->free_when_done = wc_arg; + pick_done = pick_from_internal_rr_locked(exec_ctx, glb_policy->rr_policy, + pick_args, target, wc_arg); } else { - /* else, the pending pick will be registered and taken care of by the - * pending pick list inside the RR policy (glb_policy->rr_policy) */ - grpc_polling_entity_add_to_pollset_set(exec_ctx, pick_args->pollent, - glb_policy->base.interested_parties); + if (grpc_lb_glb_trace) { + gpr_log(GPR_DEBUG, + "No RR policy in grpclb instance %p. Adding to grpclb's pending " + "picks", + (void *)(glb_policy)); + } add_pending_pick(&glb_policy->pending_picks, pick_args, target, on_complete); if (!glb_policy->started_picking) { - start_picking(exec_ctx, glb_policy); + start_picking_locked(exec_ctx, glb_policy); } pick_done = false; } @@ -841,7 +932,7 @@ static void glb_ping_one(grpc_exec_ctx *exec_ctx, grpc_lb_policy *pol, } else { add_pending_ping(&glb_policy->pending_pings, closure); if (!glb_policy->started_picking) { - start_picking(exec_ctx, glb_policy); + start_picking_locked(exec_ctx, glb_policy); } } gpr_mu_unlock(&glb_policy->mu); @@ -859,248 +950,182 @@ static void glb_notify_on_state_change(grpc_exec_ctx *exec_ctx, gpr_mu_unlock(&glb_policy->mu); } -/* - * lb_client_data - * - * Used internally for the client call to the LB */ -typedef struct lb_client_data { - gpr_mu mu; - - /* called once initial metadata's been sent */ - grpc_closure md_sent; - - /* called once the LoadBalanceRequest has been sent to the LB server. See - * src/proto/grpc/.../load_balancer.proto */ - grpc_closure req_sent; - - /* A response from the LB server has been received (or error). Process it */ - grpc_closure res_rcvd; - - /* After the client has sent a close to the LB server */ - grpc_closure close_sent; - - /* ... and the status from the LB server has been received */ - grpc_closure srv_status_rcvd; - - grpc_call *lb_call; /* streaming call to the LB server, */ - gpr_timespec deadline; /* for the streaming call to the LB server */ - - grpc_metadata_array initial_metadata_recv; /* initial MD from LB server */ - grpc_metadata_array trailing_metadata_recv; /* trailing MD from LB server */ - - /* what's being sent to the LB server. Note that its value may vary if the LB - * server indicates a redirect. */ - grpc_byte_buffer *request_payload; - - /* response from the LB server, if any. Processed in res_recv_cb() */ - grpc_byte_buffer *response_payload; - - /* the call's status and status detailset in srv_status_rcvd_cb() */ - grpc_status_code status; - char *status_details; - size_t status_details_capacity; - - /* pointer back to the enclosing policy */ - glb_lb_policy *glb_policy; -} lb_client_data; - -static void md_sent_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); -static void req_sent_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); -static void res_recv_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error); -static void close_sent_cb(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error); -static void srv_status_rcvd_cb(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error); - -static lb_client_data *lb_client_data_create(glb_lb_policy *glb_policy) { +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) { GPR_ASSERT(glb_policy->server_name != NULL); GPR_ASSERT(glb_policy->server_name[0] != '\0'); - lb_client_data *lb_client = gpr_malloc(sizeof(lb_client_data)); - memset(lb_client, 0, sizeof(lb_client_data)); - - gpr_mu_init(&lb_client->mu); - grpc_closure_init(&lb_client->md_sent, md_sent_cb, lb_client); - - grpc_closure_init(&lb_client->req_sent, req_sent_cb, lb_client); - grpc_closure_init(&lb_client->res_rcvd, res_recv_cb, lb_client); - grpc_closure_init(&lb_client->close_sent, close_sent_cb, lb_client); - grpc_closure_init(&lb_client->srv_status_rcvd, srv_status_rcvd_cb, lb_client); - - lb_client->deadline = glb_policy->deadline; - /* 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(). */ - lb_client->lb_call = grpc_channel_create_pollset_set_call( + * entities from \a client_channel. */ + glb_policy->lb_call = grpc_channel_create_pollset_set_call( glb_policy->lb_channel, NULL, GRPC_PROPAGATE_DEFAULTS, glb_policy->base.interested_parties, "/grpc.lb.v1.LoadBalancer/BalanceLoad", glb_policy->server_name, - lb_client->deadline, NULL); + glb_policy->deadline, NULL); - grpc_metadata_array_init(&lb_client->initial_metadata_recv); - grpc_metadata_array_init(&lb_client->trailing_metadata_recv); + grpc_metadata_array_init(&glb_policy->lb_initial_metadata_recv); + grpc_metadata_array_init(&glb_policy->lb_trailing_metadata_recv); grpc_grpclb_request *request = grpc_grpclb_request_create(glb_policy->server_name); - gpr_slice request_payload_slice = grpc_grpclb_request_encode(request); - lb_client->request_payload = + 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); - lb_client->status_details = NULL; - lb_client->status_details_capacity = 0; - lb_client->glb_policy = glb_policy; - return lb_client; + glb_policy->lb_call_status_details = NULL; + glb_policy->lb_call_status_details_capacity = 0; + + grpc_closure_init(&glb_policy->lb_on_server_status_received, + lb_on_server_status_received, 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); } -static void lb_client_data_destroy(lb_client_data *lb_client) { - grpc_call_destroy(lb_client->lb_call); - grpc_metadata_array_destroy(&lb_client->initial_metadata_recv); - grpc_metadata_array_destroy(&lb_client->trailing_metadata_recv); +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; - grpc_byte_buffer_destroy(lb_client->request_payload); + grpc_metadata_array_destroy(&glb_policy->lb_initial_metadata_recv); + grpc_metadata_array_destroy(&glb_policy->lb_trailing_metadata_recv); - gpr_free(lb_client->status_details); - gpr_mu_destroy(&lb_client->mu); - gpr_free(lb_client); -} -static grpc_call *lb_client_data_get_call(lb_client_data *lb_client) { - return lb_client->lb_call; + grpc_byte_buffer_destroy(glb_policy->lb_request_payload); + gpr_free(glb_policy->lb_call_status_details); } /* * Auxiliary functions and LB client callbacks. */ -static void query_for_backends(grpc_exec_ctx *exec_ctx, - 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 (grpc_lb_glb_trace) { + gpr_log(GPR_INFO, "Query for backends (grpclb: %p, lb_call: %p)", + (void *)glb_policy, (void *)glb_policy->lb_call); + } + GPR_ASSERT(glb_policy->lb_call != NULL); - glb_policy->lb_client = lb_client_data_create(glb_policy); grpc_call_error call_error; - grpc_op ops[1]; + grpc_op ops[4]; memset(ops, 0, sizeof(ops)); + grpc_op *op = ops; op->op = GRPC_OP_SEND_INITIAL_METADATA; op->data.send_initial_metadata.count = 0; op->flags = 0; op->reserved = NULL; op++; - call_error = grpc_call_start_batch_and_execute( - exec_ctx, glb_policy->lb_client->lb_call, ops, (size_t)(op - ops), - &glb_policy->lb_client->md_sent); - GPR_ASSERT(GRPC_CALL_OK == call_error); - op = ops; - op->op = GRPC_OP_RECV_STATUS_ON_CLIENT; - op->data.recv_status_on_client.trailing_metadata = - &glb_policy->lb_client->trailing_metadata_recv; - op->data.recv_status_on_client.status = &glb_policy->lb_client->status; - op->data.recv_status_on_client.status_details = - &glb_policy->lb_client->status_details; - op->data.recv_status_on_client.status_details_capacity = - &glb_policy->lb_client->status_details_capacity; + op->op = GRPC_OP_RECV_INITIAL_METADATA; + op->data.recv_initial_metadata = &glb_policy->lb_initial_metadata_recv; op->flags = 0; op->reserved = NULL; op++; - call_error = grpc_call_start_batch_and_execute( - exec_ctx, glb_policy->lb_client->lb_call, ops, (size_t)(op - ops), - &glb_policy->lb_client->srv_status_rcvd); - GPR_ASSERT(GRPC_CALL_OK == call_error); -} - -static void md_sent_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - lb_client_data *lb_client = arg; - GPR_ASSERT(lb_client->lb_call); - grpc_op ops[1]; - memset(ops, 0, sizeof(ops)); - grpc_op *op = ops; + GPR_ASSERT(glb_policy->lb_request_payload != NULL); op->op = GRPC_OP_SEND_MESSAGE; - op->data.send_message = lb_client->request_payload; + op->data.send_message = glb_policy->lb_request_payload; op->flags = 0; op->reserved = NULL; op++; - grpc_call_error call_error = grpc_call_start_batch_and_execute( - exec_ctx, lb_client->lb_call, ops, (size_t)(op - ops), - &lb_client->req_sent); - GPR_ASSERT(GRPC_CALL_OK == call_error); -} - -static void req_sent_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - lb_client_data *lb_client = arg; - GPR_ASSERT(lb_client->lb_call); - grpc_op ops[2]; - memset(ops, 0, sizeof(ops)); - grpc_op *op = ops; - - op->op = GRPC_OP_RECV_INITIAL_METADATA; - op->data.recv_initial_metadata = &lb_client->initial_metadata_recv; + op->op = GRPC_OP_RECV_STATUS_ON_CLIENT; + op->data.recv_status_on_client.trailing_metadata = + &glb_policy->lb_trailing_metadata_recv; + op->data.recv_status_on_client.status = &glb_policy->lb_call_status; + op->data.recv_status_on_client.status_details = + &glb_policy->lb_call_status_details; + op->data.recv_status_on_client.status_details_capacity = + &glb_policy->lb_call_status_details_capacity; op->flags = 0; op->reserved = NULL; op++; + /* take a weak ref (won't prevent calling of \a glb_shutdown if the strong ref + * count goes to zero) to be unref'd in lb_on_server_status_received */ + GRPC_LB_POLICY_WEAK_REF(&glb_policy->base, "lb_on_server_status_received"); + call_error = grpc_call_start_batch_and_execute( + exec_ctx, glb_policy->lb_call, ops, (size_t)(op - ops), + &glb_policy->lb_on_server_status_received); + GPR_ASSERT(GRPC_CALL_OK == call_error); + op = ops; op->op = GRPC_OP_RECV_MESSAGE; - op->data.recv_message = &lb_client->response_payload; + op->data.recv_message = &glb_policy->lb_response_payload; op->flags = 0; op->reserved = NULL; op++; - grpc_call_error call_error = grpc_call_start_batch_and_execute( - exec_ctx, lb_client->lb_call, ops, (size_t)(op - ops), - &lb_client->res_rcvd); + /* take another weak ref to be unref'd in lb_on_response_received */ + GRPC_LB_POLICY_WEAK_REF(&glb_policy->base, "lb_on_response_received"); + call_error = grpc_call_start_batch_and_execute( + exec_ctx, glb_policy->lb_call, ops, (size_t)(op - ops), + &glb_policy->lb_on_response_received); GPR_ASSERT(GRPC_CALL_OK == call_error); } -static void res_recv_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { - lb_client_data *lb_client = arg; +static void lb_on_response_received(grpc_exec_ctx *exec_ctx, void *arg, + grpc_error *error) { + glb_lb_policy *glb_policy = arg; + grpc_op ops[2]; memset(ops, 0, sizeof(ops)); grpc_op *op = ops; - if (lb_client->response_payload != NULL) { + if (glb_policy->lb_response_payload != NULL) { + gpr_backoff_reset(&glb_policy->lb_call_backoff_state); /* Received data from the LB server. Look inside - * lb_client->response_payload, for a serverlist. */ + * glb_policy->lb_response_payload, for a serverlist. */ grpc_byte_buffer_reader bbr; - grpc_byte_buffer_reader_init(&bbr, lb_client->response_payload); - gpr_slice response_slice = grpc_byte_buffer_reader_readall(&bbr); - grpc_byte_buffer_destroy(lb_client->response_payload); + grpc_byte_buffer_reader_init(&bbr, glb_policy->lb_response_payload); + 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_slice_unref(response_slice); + GPR_ASSERT(glb_policy->lb_call != NULL); + grpc_slice_unref(response_slice); if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, "Serverlist with %zu servers received", - serverlist->num_servers); + gpr_log(GPR_INFO, "Serverlist with %lu servers received", + (unsigned long)serverlist->num_servers); + for (size_t i = 0; i < serverlist->num_servers; ++i) { + grpc_resolved_address addr; + parse_server(serverlist->servers[i], &addr); + char *ipport; + grpc_sockaddr_to_string(&ipport, &addr, false); + gpr_log(GPR_INFO, "Serverlist[%lu]: %s", (unsigned long)i, ipport); + gpr_free(ipport); + } } /* update serverlist */ if (serverlist->num_servers > 0) { - if (grpc_grpclb_serverlist_equals(lb_client->glb_policy->serverlist, - serverlist)) { + 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."); } } else { /* new serverlist */ - if (lb_client->glb_policy->serverlist != NULL) { + if (glb_policy->serverlist != NULL) { /* dispose of the old serverlist */ - grpc_grpclb_destroy_serverlist(lb_client->glb_policy->serverlist); + grpc_grpclb_destroy_serverlist(glb_policy->serverlist); } /* and update the copy in the glb_lb_policy instance */ - lb_client->glb_policy->serverlist = serverlist; - } - if (lb_client->glb_policy->rr_policy == NULL) { - /* initial "handover", in this case from a null RR policy, meaning - * it'll just create the first RR policy instance */ - rr_handover(exec_ctx, lb_client->glb_policy, error); - } else { - /* unref the RR policy, eventually leading to its substitution with a - * new one constructed from the received serverlist (see - * glb_rr_connectivity_changed) */ - GRPC_LB_POLICY_UNREF(exec_ctx, lb_client->glb_policy->rr_policy, - "serverlist_received"); + glb_policy->serverlist = serverlist; + + rr_handover_locked(exec_ctx, glb_policy, error); } + gpr_mu_unlock(&glb_policy->mu); } else { if (grpc_lb_glb_trace) { gpr_log(GPR_INFO, @@ -1108,60 +1133,94 @@ static void res_recv_cb(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error) { "response with > 0 servers is received"); } } + } else { /* serverlist == NULL */ + gpr_log(GPR_ERROR, "Invalid LB response received: '%s'. Ignoring.", + grpc_dump_slice(response_slice, GPR_DUMP_ASCII | GPR_DUMP_HEX)); + grpc_slice_unref(response_slice); + } + if (!glb_policy->shutting_down) { /* keep listening for serverlist updates */ op->op = GRPC_OP_RECV_MESSAGE; - op->data.recv_message = &lb_client->response_payload; + op->data.recv_message = &glb_policy->lb_response_payload; op->flags = 0; op->reserved = NULL; op++; + /* reuse the "lb_on_response_received" weak ref taken in + * query_for_backends_locked() */ const grpc_call_error call_error = grpc_call_start_batch_and_execute( - exec_ctx, lb_client->lb_call, ops, (size_t)(op - ops), - &lb_client->res_rcvd); /* loop */ + exec_ctx, glb_policy->lb_call, ops, (size_t)(op - ops), + &glb_policy->lb_on_response_received); /* loop */ GPR_ASSERT(GRPC_CALL_OK == call_error); - return; } - - GPR_ASSERT(serverlist == NULL); - gpr_log(GPR_ERROR, "Invalid LB response received: '%s'", - gpr_dump_slice(response_slice, GPR_DUMP_ASCII)); - gpr_slice_unref(response_slice); - - /* Disconnect from server returning invalid response. */ - op->op = GRPC_OP_SEND_CLOSE_FROM_CLIENT; - op->flags = 0; - op->reserved = NULL; - op++; - grpc_call_error call_error = grpc_call_start_batch_and_execute( - exec_ctx, lb_client->lb_call, ops, (size_t)(op - ops), - &lb_client->close_sent); - GPR_ASSERT(GRPC_CALL_OK == call_error); + } 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 */ + GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &glb_policy->base, + "lb_on_response_received_empty_payload"); } - /* empty payload: call cancelled by server. Cleanups happening in - * srv_status_rcvd_cb */ } -static void close_sent_cb(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, - "Close from LB client sent. Waiting from server status now"); +static void lb_call_on_retry_timer(grpc_exec_ctx *exec_ctx, void *arg, + grpc_error *error) { + glb_lb_policy *glb_policy = arg; + gpr_mu_lock(&glb_policy->mu); + + if (!glb_policy->shutting_down) { + if (grpc_lb_glb_trace) { + gpr_log(GPR_INFO, "Restaring call to LB server (grpclb %p)", + (void *)glb_policy); + } + GPR_ASSERT(glb_policy->lb_call == NULL); + 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"); } -static void srv_status_rcvd_cb(grpc_exec_ctx *exec_ctx, void *arg, - grpc_error *error) { - lb_client_data *lb_client = arg; +static void lb_on_server_status_received(grpc_exec_ctx *exec_ctx, void *arg, + grpc_error *error) { + glb_lb_policy *glb_policy = arg; + gpr_mu_lock(&glb_policy->mu); + + GPR_ASSERT(glb_policy->lb_call != NULL); + if (grpc_lb_glb_trace) { - gpr_log(GPR_INFO, - "status from lb server received. Status = %d, Details = '%s', " - "Capaticy " - "= %zu", - lb_client->status, lb_client->status_details, - lb_client->status_details_capacity); + gpr_log(GPR_DEBUG, + "Status from LB server received. Status = %d, Details = '%s', " + "(call: %p)", + glb_policy->lb_call_status, glb_policy->lb_call_status_details, + (void *)glb_policy->lb_call); + } + + /* We need to performe cleanups no matter what. */ + 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 */ + gpr_timespec now = gpr_now(GPR_CLOCK_MONOTONIC); + gpr_timespec next_try = + gpr_backoff_step(&glb_policy->lb_call_backoff_state, now); + if (grpc_lb_glb_trace) { + gpr_log(GPR_DEBUG, "Connection to LB server lost (grpclb: %p)...", + (void *)glb_policy); + gpr_timespec timeout = gpr_time_sub(next_try, now); + if (gpr_time_cmp(timeout, gpr_time_0(timeout.clock_type)) > 0) { + gpr_log(GPR_DEBUG, "... retrying in %" PRId64 ".%09d seconds.", + timeout.tv_sec, timeout.tv_nsec); + } else { + gpr_log(GPR_DEBUG, "... retrying immediately."); + } + } + GRPC_LB_POLICY_WEAK_REF(&glb_policy->base, "grpclb_retry_timer"); + grpc_timer_init(exec_ctx, &glb_policy->lb_call_retry_timer, next_try, + lb_call_on_retry_timer, glb_policy, now); } - /* TODO(dgq): deal with stream termination properly (fire up another one? - * fail the original call?) */ + gpr_mu_unlock(&glb_policy->mu); + GRPC_LB_POLICY_WEAK_UNREF(exec_ctx, &glb_policy->base, + "lb_on_server_status_received"); } /* Code wiring the policy with the rest of the core */ |