diff options
Diffstat (limited to 'src/core')
40 files changed, 693 insertions, 1995 deletions
diff --git a/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc b/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc index 1ee1925a25..0061190016 100644 --- a/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc +++ b/src/core/ext/filters/client_channel/lb_policy/grpclb/grpclb.cc @@ -852,10 +852,12 @@ void GrpcLb::BalancerCallState::OnBalancerMessageReceivedLocked( } } else { // No valid initial response or serverlist found. + char* response_slice_str = + grpc_dump_slice(response_slice, GPR_DUMP_ASCII | GPR_DUMP_HEX); gpr_log(GPR_ERROR, "[grpclb %p] Invalid LB response received: '%s'. Ignoring.", - grpclb_policy, - grpc_dump_slice(response_slice, GPR_DUMP_ASCII | GPR_DUMP_HEX)); + grpclb_policy, response_slice_str); + gpr_free(response_slice_str); } grpc_slice_unref_internal(response_slice); if (!grpclb_policy->shutting_down_) { diff --git a/src/core/ext/filters/client_channel/parse_address.cc b/src/core/ext/filters/client_channel/parse_address.cc index 7e726dd562..707beb8876 100644 --- a/src/core/ext/filters/client_channel/parse_address.cc +++ b/src/core/ext/filters/client_channel/parse_address.cc @@ -129,30 +129,37 @@ bool grpc_parse_ipv6_hostport(const char* hostport, grpc_resolved_address* addr, size_t host_without_scope_len = static_cast<size_t>(host_end - host); uint32_t sin6_scope_id = 0; if (host_without_scope_len > GRPC_INET6_ADDRSTRLEN) { - gpr_log(GPR_ERROR, - "invalid ipv6 address length %zu. Length cannot be greater than " - "GRPC_INET6_ADDRSTRLEN i.e %d)", - host_without_scope_len, GRPC_INET6_ADDRSTRLEN); + if (log_errors) { + gpr_log( + GPR_ERROR, + "invalid ipv6 address length %zu. Length cannot be greater than " + "GRPC_INET6_ADDRSTRLEN i.e %d)", + host_without_scope_len, GRPC_INET6_ADDRSTRLEN); + } goto done; } strncpy(host_without_scope, host, host_without_scope_len); host_without_scope[host_without_scope_len] = '\0'; if (grpc_inet_pton(GRPC_AF_INET6, host_without_scope, &in6->sin6_addr) == 0) { - gpr_log(GPR_ERROR, "invalid ipv6 address: '%s'", host_without_scope); + if (log_errors) { + gpr_log(GPR_ERROR, "invalid ipv6 address: '%s'", host_without_scope); + } goto done; } if (gpr_parse_bytes_to_uint32(host_end + 1, strlen(host) - host_without_scope_len - 1, &sin6_scope_id) == 0) { - gpr_log(GPR_ERROR, "invalid ipv6 scope id: '%s'", host_end + 1); + if (log_errors) { + gpr_log(GPR_ERROR, "invalid ipv6 scope id: '%s'", host_end + 1); + } goto done; } // Handle "sin6_scope_id" being type "u_long". See grpc issue #10027. in6->sin6_scope_id = sin6_scope_id; } else { if (grpc_inet_pton(GRPC_AF_INET6, host, &in6->sin6_addr) == 0) { - gpr_log(GPR_ERROR, "invalid ipv6 address: '%s'", host); + if (log_errors) gpr_log(GPR_ERROR, "invalid ipv6 address: '%s'", host); goto done; } } @@ -197,3 +204,12 @@ bool grpc_parse_uri(const grpc_uri* uri, grpc_resolved_address* resolved_addr) { gpr_log(GPR_ERROR, "Can't parse scheme '%s'", uri->scheme); return false; } + +uint16_t grpc_strhtons(const char* port) { + if (strcmp(port, "http") == 0) { + return htons(80); + } else if (strcmp(port, "https") == 0) { + return htons(443); + } + return htons(static_cast<unsigned short>(atoi(port))); +} diff --git a/src/core/ext/filters/client_channel/parse_address.h b/src/core/ext/filters/client_channel/parse_address.h index 9a88b66edc..c2af0e6c49 100644 --- a/src/core/ext/filters/client_channel/parse_address.h +++ b/src/core/ext/filters/client_channel/parse_address.h @@ -47,4 +47,7 @@ bool grpc_parse_ipv4_hostport(const char* hostport, grpc_resolved_address* addr, bool grpc_parse_ipv6_hostport(const char* hostport, grpc_resolved_address* addr, bool log_errors); +/* Converts named or numeric port to a uint16 suitable for use in a sockaddr. */ +uint16_t grpc_strhtons(const char* port); + #endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_PARSE_ADDRESS_H */ diff --git a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.cc b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.cc index 485998f5e4..4c795c34c8 100644 --- a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.cc +++ b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.cc @@ -87,15 +87,6 @@ typedef struct grpc_ares_hostbyname_request { static void do_basic_init(void) { gpr_mu_init(&g_init_mu); } -static uint16_t strhtons(const char* port) { - if (strcmp(port, "http") == 0) { - return htons(80); - } else if (strcmp(port, "https") == 0) { - return htons(443); - } - return htons(static_cast<unsigned short>(atoi(port))); -} - static void log_address_sorting_list(grpc_lb_addresses* lb_addrs, const char* input_output_str) { for (size_t i = 0; i < lb_addrs->num_addresses; i++) { @@ -139,12 +130,6 @@ void grpc_cares_wrapper_address_sorting_sort(grpc_lb_addresses* lb_addrs) { } } -/* Allow tests to access grpc_ares_wrapper_address_sorting_sort */ -void grpc_cares_wrapper_test_only_address_sorting_sort( - grpc_lb_addresses* lb_addrs) { - grpc_cares_wrapper_address_sorting_sort(lb_addrs); -} - static void grpc_ares_request_ref_locked(grpc_ares_request* r) { r->pending_queries++; } @@ -371,7 +356,8 @@ done: grpc_ares_request_unref_locked(r); } -static grpc_ares_request* grpc_dns_lookup_ares_locked_impl( +static grpc_ares_request* +grpc_dns_lookup_ares_continue_after_check_localhost_and_ip_literals_locked( const char* dns_server, const char* name, const char* default_port, grpc_pollset_set* interested_parties, grpc_closure* on_done, grpc_lb_addresses** addrs, bool check_grpclb, char** service_config_json, @@ -454,12 +440,12 @@ static grpc_ares_request* grpc_dns_lookup_ares_locked_impl( } r->pending_queries = 1; if (grpc_ares_query_ipv6()) { - hr = create_hostbyname_request_locked(r, host, strhtons(port), + hr = create_hostbyname_request_locked(r, host, grpc_strhtons(port), false /* is_balancer */); ares_gethostbyname(*channel, hr->host, AF_INET6, on_hostbyname_done_locked, hr); } - hr = create_hostbyname_request_locked(r, host, strhtons(port), + hr = create_hostbyname_request_locked(r, host, grpc_strhtons(port), false /* is_balancer */); ares_gethostbyname(*channel, hr->host, AF_INET, on_hostbyname_done_locked, hr); @@ -494,6 +480,79 @@ error_cleanup: return nullptr; } +static bool inner_resolve_as_ip_literal_locked(const char* name, + const char* default_port, + grpc_lb_addresses** addrs, + char** host, char** port, + char** hostport) { + gpr_split_host_port(name, host, port); + if (*host == nullptr) { + gpr_log(GPR_ERROR, + "Failed to parse %s to host:port while attempting to resolve as ip " + "literal.", + name); + return false; + } + if (*port == nullptr) { + if (default_port == nullptr) { + gpr_log(GPR_ERROR, + "No port or default port for %s while attempting to resolve as " + "ip literal.", + name); + return false; + } + *port = gpr_strdup(default_port); + } + grpc_resolved_address addr; + GPR_ASSERT(gpr_join_host_port(hostport, *host, atoi(*port))); + if (grpc_parse_ipv4_hostport(*hostport, &addr, false /* log errors */) || + grpc_parse_ipv6_hostport(*hostport, &addr, false /* log errors */)) { + GPR_ASSERT(*addrs == nullptr); + *addrs = grpc_lb_addresses_create(1, nullptr); + grpc_lb_addresses_set_address( + *addrs, 0, addr.addr, addr.len, false /* is_balancer */, + nullptr /* balancer_name */, nullptr /* user_data */); + return true; + } + return false; +} + +static bool resolve_as_ip_literal_locked(const char* name, + const char* default_port, + grpc_lb_addresses** addrs) { + char* host = nullptr; + char* port = nullptr; + char* hostport = nullptr; + bool out = inner_resolve_as_ip_literal_locked(name, default_port, addrs, + &host, &port, &hostport); + gpr_free(host); + gpr_free(port); + gpr_free(hostport); + return out; +} + +static grpc_ares_request* grpc_dns_lookup_ares_locked_impl( + const char* dns_server, const char* name, const char* default_port, + grpc_pollset_set* interested_parties, grpc_closure* on_done, + grpc_lb_addresses** addrs, bool check_grpclb, char** service_config_json, + grpc_combiner* combiner) { + // Early out if the target is an ipv4 or ipv6 literal. + if (resolve_as_ip_literal_locked(name, default_port, addrs)) { + GRPC_CLOSURE_SCHED(on_done, GRPC_ERROR_NONE); + return nullptr; + } + // Early out if the target is localhost and we're on Windows. + if (grpc_ares_maybe_resolve_localhost_manually_locked(name, default_port, + addrs)) { + GRPC_CLOSURE_SCHED(on_done, GRPC_ERROR_NONE); + return nullptr; + } + // Look up name using c-ares lib. + return grpc_dns_lookup_ares_continue_after_check_localhost_and_ip_literals_locked( + dns_server, name, default_port, interested_parties, on_done, addrs, + check_grpclb, service_config_json, combiner); +} + grpc_ares_request* (*grpc_dns_lookup_ares_locked)( const char* dns_server, const char* name, const char* default_port, grpc_pollset_set* interested_parties, grpc_closure* on_done, @@ -502,7 +561,9 @@ grpc_ares_request* (*grpc_dns_lookup_ares_locked)( void grpc_cancel_ares_request(grpc_ares_request* r) { if (grpc_dns_lookup_ares_locked == grpc_dns_lookup_ares_locked_impl) { - grpc_ares_ev_driver_shutdown_locked(r->ev_driver); + if (r != nullptr) { + grpc_ares_ev_driver_shutdown_locked(r->ev_driver); + } } } diff --git a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h index ca5779e1d7..1bc457d4cf 100644 --- a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h +++ b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h @@ -81,9 +81,15 @@ void grpc_ares_complete_request_locked(grpc_ares_request* request); /* E.g., return false if ipv6 is known to not be available. */ bool grpc_ares_query_ipv6(); -/* Exposed only for testing */ -void grpc_cares_wrapper_test_only_address_sorting_sort( - grpc_lb_addresses* lb_addrs); +/* Maybe (depending on the current platform) checks if "name" matches + * "localhost" and if so fills in addrs with the correct sockaddr structures. + * Returns a bool indicating whether or not such an action was performed. + * See https://github.com/grpc/grpc/issues/15158. */ +bool grpc_ares_maybe_resolve_localhost_manually_locked( + const char* name, const char* default_port, grpc_lb_addresses** addrs); + +/* Sorts destinations in lb_addrs according to RFC 6724. */ +void grpc_cares_wrapper_address_sorting_sort(grpc_lb_addresses* lb_addrs); #endif /* GRPC_CORE_EXT_FILTERS_CLIENT_CHANNEL_RESOLVER_DNS_C_ARES_GRPC_ARES_WRAPPER_H \ */ diff --git a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_posix.cc b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_posix.cc index 23c0fec74f..639eec2323 100644 --- a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_posix.cc +++ b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_posix.cc @@ -26,4 +26,9 @@ bool grpc_ares_query_ipv6() { return grpc_ipv6_loopback_available(); } +bool grpc_ares_maybe_resolve_localhost_manually_locked( + const char* name, const char* default_port, grpc_lb_addresses** addrs) { + return false; +} + #endif /* GRPC_ARES == 1 && defined(GRPC_POSIX_SOCKET_ARES_EV_DRIVER) */ diff --git a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_windows.cc b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_windows.cc index ee827e284e..7e34784691 100644 --- a/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_windows.cc +++ b/src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper_windows.cc @@ -21,9 +21,79 @@ #include "src/core/lib/iomgr/port.h" #if GRPC_ARES == 1 && defined(GPR_WINDOWS) +#include <grpc/support/string_util.h> + +#include "src/core/ext/filters/client_channel/lb_policy_factory.h" +#include "src/core/ext/filters/client_channel/parse_address.h" #include "src/core/ext/filters/client_channel/resolver/dns/c_ares/grpc_ares_wrapper.h" +#include "src/core/lib/gpr/host_port.h" +#include "src/core/lib/gpr/string.h" #include "src/core/lib/iomgr/socket_windows.h" bool grpc_ares_query_ipv6() { return grpc_ipv6_loopback_available(); } +static bool inner_maybe_resolve_localhost_manually_locked( + const char* name, const char* default_port, grpc_lb_addresses** addrs, + char** host, char** port) { + gpr_split_host_port(name, host, port); + if (*host == nullptr) { + gpr_log(GPR_ERROR, + "Failed to parse %s into host:port during Windows localhost " + "resolution check.", + name); + return false; + } + if (*port == nullptr) { + if (default_port == nullptr) { + gpr_log(GPR_ERROR, + "No port or default port for %s during Windows localhost " + "resolution check.", + name); + return false; + } + *port = gpr_strdup(default_port); + } + if (gpr_stricmp(*host, "localhost") == 0) { + GPR_ASSERT(*addrs == nullptr); + *addrs = grpc_lb_addresses_create(2, nullptr); + uint16_t numeric_port = grpc_strhtons(*port); + // Append the ipv6 loopback address. + struct sockaddr_in6 ipv6_loopback_addr; + memset(&ipv6_loopback_addr, 0, sizeof(ipv6_loopback_addr)); + ((char*)&ipv6_loopback_addr.sin6_addr)[15] = 1; + ipv6_loopback_addr.sin6_family = AF_INET6; + ipv6_loopback_addr.sin6_port = numeric_port; + grpc_lb_addresses_set_address( + *addrs, 0, &ipv6_loopback_addr, sizeof(ipv6_loopback_addr), + false /* is_balancer */, nullptr /* balancer_name */, + nullptr /* user_data */); + // Append the ipv4 loopback address. + struct sockaddr_in ipv4_loopback_addr; + memset(&ipv4_loopback_addr, 0, sizeof(ipv4_loopback_addr)); + ((char*)&ipv4_loopback_addr.sin_addr)[0] = 0x7f; + ((char*)&ipv4_loopback_addr.sin_addr)[3] = 0x01; + ipv4_loopback_addr.sin_family = AF_INET; + ipv4_loopback_addr.sin_port = numeric_port; + grpc_lb_addresses_set_address( + *addrs, 1, &ipv4_loopback_addr, sizeof(ipv4_loopback_addr), + false /* is_balancer */, nullptr /* balancer_name */, + nullptr /* user_data */); + // Let the address sorter figure out which one should be tried first. + grpc_cares_wrapper_address_sorting_sort(*addrs); + return true; + } + return false; +} + +bool grpc_ares_maybe_resolve_localhost_manually_locked( + const char* name, const char* default_port, grpc_lb_addresses** addrs) { + char* host = nullptr; + char* port = nullptr; + bool out = inner_maybe_resolve_localhost_manually_locked(name, default_port, + addrs, &host, &port); + gpr_free(host); + gpr_free(port); + return out; +} + #endif /* GRPC_ARES == 1 && defined(GPR_WINDOWS) */ diff --git a/src/core/ext/filters/client_channel/subchannel_index.cc b/src/core/ext/filters/client_channel/subchannel_index.cc index f2b6c24e8e..1c23a6c4be 100644 --- a/src/core/ext/filters/client_channel/subchannel_index.cc +++ b/src/core/ext/filters/client_channel/subchannel_index.cc @@ -42,7 +42,7 @@ struct grpc_subchannel_key { grpc_subchannel_args args; }; -static gpr_atm g_force_creation = false; +static bool g_force_creation = false; static grpc_subchannel_key* create_key( const grpc_subchannel_args* args, @@ -74,7 +74,7 @@ static grpc_subchannel_key* subchannel_key_copy(grpc_subchannel_key* k) { int grpc_subchannel_key_compare(const grpc_subchannel_key* a, const grpc_subchannel_key* b) { // To pretend the keys are different, return a non-zero value. - if (GPR_UNLIKELY(gpr_atm_no_barrier_load(&g_force_creation))) return 1; + if (GPR_UNLIKELY(g_force_creation)) return 1; int c = GPR_ICMP(a->args.filter_count, b->args.filter_count); if (c != 0) return c; if (a->args.filter_count > 0) { @@ -251,5 +251,5 @@ void grpc_subchannel_index_unregister(grpc_subchannel_key* key, } void grpc_subchannel_index_test_only_set_force_creation(bool force_creation) { - gpr_atm_no_barrier_store(&g_force_creation, force_creation); + g_force_creation = force_creation; } diff --git a/src/core/ext/filters/http/client/http_client_filter.cc b/src/core/ext/filters/http/client/http_client_filter.cc index 91fa163fec..cd459e47cd 100644 --- a/src/core/ext/filters/http/client/http_client_filter.cc +++ b/src/core/ext/filters/http/client/http_client_filter.cc @@ -58,6 +58,8 @@ struct call_data { grpc_metadata_batch* recv_trailing_metadata; grpc_closure* original_recv_trailing_metadata_ready; grpc_closure recv_trailing_metadata_ready; + grpc_error* recv_trailing_metadata_error; + bool seen_recv_trailing_metadata_ready; // State for handling send_message ops. grpc_transport_stream_op_batch* send_message_batch; size_t send_message_bytes_read; @@ -157,12 +159,27 @@ static void recv_initial_metadata_ready(void* user_data, grpc_error* error) { } else { GRPC_ERROR_REF(error); } - GRPC_CLOSURE_RUN(calld->original_recv_initial_metadata_ready, error); + grpc_closure* closure = calld->original_recv_initial_metadata_ready; + calld->original_recv_initial_metadata_ready = nullptr; + if (calld->seen_recv_trailing_metadata_ready) { + GRPC_CALL_COMBINER_START( + calld->call_combiner, &calld->recv_trailing_metadata_ready, + calld->recv_trailing_metadata_error, "continue recv_trailing_metadata"); + } + GRPC_CLOSURE_RUN(closure, error); } static void recv_trailing_metadata_ready(void* user_data, grpc_error* error) { grpc_call_element* elem = static_cast<grpc_call_element*>(user_data); call_data* calld = static_cast<call_data*>(elem->call_data); + if (calld->original_recv_initial_metadata_ready != nullptr) { + calld->recv_trailing_metadata_error = GRPC_ERROR_REF(error); + calld->seen_recv_trailing_metadata_ready = true; + GRPC_CALL_COMBINER_STOP(calld->call_combiner, + "deferring recv_trailing_metadata_ready until " + "after recv_initial_metadata_ready"); + return; + } if (error == GRPC_ERROR_NONE) { error = client_filter_incoming_metadata(elem, calld->recv_trailing_metadata); diff --git a/src/core/ext/filters/http/server/http_server_filter.cc b/src/core/ext/filters/http/server/http_server_filter.cc index 1b3426b120..436ea09d94 100644 --- a/src/core/ext/filters/http/server/http_server_filter.cc +++ b/src/core/ext/filters/http/server/http_server_filter.cc @@ -63,8 +63,11 @@ struct call_data { grpc_core::OrphanablePtr<grpc_core::ByteStream>* recv_message; bool seen_recv_message_ready; + // State for intercepting recv_trailing_metadata grpc_closure recv_trailing_metadata_ready; grpc_closure* original_recv_trailing_metadata_ready; + grpc_error* recv_trailing_metadata_ready_error; + bool seen_recv_trailing_metadata_ready; }; struct channel_data { @@ -301,6 +304,13 @@ static void hs_recv_initial_metadata_ready(void* user_data, grpc_error* err) { } else { GRPC_ERROR_REF(err); } + if (calld->seen_recv_trailing_metadata_ready) { + GRPC_CALL_COMBINER_START(calld->call_combiner, + &calld->recv_trailing_metadata_ready, + calld->recv_trailing_metadata_ready_error, + "resuming hs_recv_trailing_metadata_ready from " + "hs_recv_initial_metadata_ready"); + } GRPC_CLOSURE_RUN(calld->original_recv_initial_metadata_ready, err); } @@ -331,6 +341,14 @@ static void hs_recv_message_ready(void* user_data, grpc_error* err) { static void hs_recv_trailing_metadata_ready(void* user_data, grpc_error* err) { grpc_call_element* elem = static_cast<grpc_call_element*>(user_data); call_data* calld = static_cast<call_data*>(elem->call_data); + if (!calld->seen_recv_initial_metadata_ready) { + calld->recv_trailing_metadata_ready_error = GRPC_ERROR_REF(err); + calld->seen_recv_trailing_metadata_ready = true; + GRPC_CALL_COMBINER_STOP(calld->call_combiner, + "deferring hs_recv_trailing_metadata_ready until " + "ater hs_recv_initial_metadata_ready"); + return; + } err = grpc_error_add_child( GRPC_ERROR_REF(err), GRPC_ERROR_REF(calld->recv_initial_metadata_ready_error)); diff --git a/src/core/ext/filters/message_size/message_size_filter.cc b/src/core/ext/filters/message_size/message_size_filter.cc index c17df86f3d..2d3b16d992 100644 --- a/src/core/ext/filters/message_size/message_size_filter.cc +++ b/src/core/ext/filters/message_size/message_size_filter.cc @@ -108,6 +108,8 @@ struct call_data { grpc_closure* next_recv_message_ready; // Original recv_trailing_metadata callback, invoked after our own. grpc_closure* original_recv_trailing_metadata_ready; + bool seen_recv_trailing_metadata; + grpc_error* recv_trailing_metadata_error; }; struct channel_data { @@ -147,7 +149,21 @@ static void recv_message_ready(void* user_data, grpc_error* error) { GRPC_ERROR_REF(error); } // Invoke the next callback. - GRPC_CLOSURE_RUN(calld->next_recv_message_ready, error); + grpc_closure* closure = calld->next_recv_message_ready; + calld->next_recv_message_ready = nullptr; + if (calld->seen_recv_trailing_metadata) { + /* We might potentially see another RECV_MESSAGE op. In that case, we do not + * want to run the recv_trailing_metadata_ready closure again. The newer + * RECV_MESSAGE op cannot cause any errors since the transport has already + * invoked the recv_trailing_metadata_ready closure and all further + * RECV_MESSAGE ops will get null payloads. */ + calld->seen_recv_trailing_metadata = false; + GRPC_CALL_COMBINER_START(calld->call_combiner, + &calld->recv_trailing_metadata_ready, + calld->recv_trailing_metadata_error, + "continue recv_trailing_metadata_ready"); + } + GRPC_CLOSURE_RUN(closure, error); } // Callback invoked on completion of recv_trailing_metadata @@ -155,6 +171,14 @@ static void recv_message_ready(void* user_data, grpc_error* error) { static void recv_trailing_metadata_ready(void* user_data, grpc_error* error) { grpc_call_element* elem = static_cast<grpc_call_element*>(user_data); call_data* calld = static_cast<call_data*>(elem->call_data); + if (calld->next_recv_message_ready != nullptr) { + calld->seen_recv_trailing_metadata = true; + calld->recv_trailing_metadata_error = GRPC_ERROR_REF(error); + GRPC_CALL_COMBINER_STOP(calld->call_combiner, + "deferring recv_trailing_metadata_ready until " + "after recv_message_ready"); + return; + } error = grpc_error_add_child(GRPC_ERROR_REF(error), GRPC_ERROR_REF(calld->error)); // Invoke the next callback. diff --git a/src/core/ext/transport/chttp2/transport/hpack_encoder.cc b/src/core/ext/transport/chttp2/transport/hpack_encoder.cc index 0eaf63f133..920d52770f 100644 --- a/src/core/ext/transport/chttp2/transport/hpack_encoder.cc +++ b/src/core/ext/transport/chttp2/transport/hpack_encoder.cc @@ -525,7 +525,7 @@ static void hpack_enc(grpc_chttp2_hpack_compressor* c, grpc_mdelem elem, /* should this elem be in the table? */ size_t decoder_space_usage = - grpc_mdelem_get_size_in_hpack_table(elem, st->use_true_binary_metadata); + grpc_chttp2_get_size_in_hpack_table(elem, st->use_true_binary_metadata); bool should_add_elem = elem_interned && decoder_space_usage < MAX_DECODER_SPACE_USAGE && c->filter_elems[HASH_FRAGMENT_1(elem_hash)] >= @@ -688,11 +688,22 @@ void grpc_chttp2_encode_header(grpc_chttp2_hpack_compressor* c, emit_advertise_table_size_change(c, &st); } for (size_t i = 0; i < extra_headers_size; ++i) { - hpack_enc(c, *extra_headers[i], &st); + grpc_mdelem md = *extra_headers[i]; + uint8_t static_index = grpc_chttp2_get_static_hpack_table_index(md); + if (static_index) { + emit_indexed(c, static_index, &st); + } else { + hpack_enc(c, md, &st); + } } grpc_metadata_batch_assert_ok(metadata); for (grpc_linked_mdelem* l = metadata->list.head; l; l = l->next) { - hpack_enc(c, l->md, &st); + uint8_t static_index = grpc_chttp2_get_static_hpack_table_index(l->md); + if (static_index) { + emit_indexed(c, static_index, &st); + } else { + hpack_enc(c, l->md, &st); + } } grpc_millis deadline = metadata->deadline; if (deadline != GRPC_MILLIS_INF_FUTURE) { diff --git a/src/core/ext/transport/chttp2/transport/hpack_mapping.cc b/src/core/ext/transport/chttp2/transport/hpack_mapping.cc new file mode 100644 index 0000000000..fd529f0fd4 --- /dev/null +++ b/src/core/ext/transport/chttp2/transport/hpack_mapping.cc @@ -0,0 +1,39 @@ +/* + * Copyright 2015 gRPC authors. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * WARNING: Auto-generated code. + * + * To make changes to this file, change + * tools/codegen/core/gen_static_metadata.py, and then re-run it. + * + * This file contains the mapping from the index of each metadata element in the + * grpc static metadata table to the index of that element in the hpack static + * metadata table. If the element is not contained in the static hpack table, + * then the returned index is 0. + */ + +#include <grpc/support/port_platform.h> + +#include "src/core/ext/transport/chttp2/transport/hpack_mapping.h" + +const uint8_t grpc_hpack_static_mdelem_indices[GRPC_STATIC_MDELEM_COUNT] = { + 0, 0, 0, 0, 0, 0, 0, 0, 3, 8, 13, 6, 7, 0, 1, 2, 0, 4, + 5, 9, 10, 11, 12, 14, 15, 0, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, + 0, 0, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, + 42, 43, 44, 0, 0, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, + 58, 59, 60, 61, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, +}; diff --git a/src/core/ext/transport/chttp2/transport/hpack_mapping.h b/src/core/ext/transport/chttp2/transport/hpack_mapping.h new file mode 100644 index 0000000000..ebcd65bd9f --- /dev/null +++ b/src/core/ext/transport/chttp2/transport/hpack_mapping.h @@ -0,0 +1,38 @@ +/* + * Copyright 2015 gRPC authors. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +/* + * WARNING: Auto-generated code. + * + * To make changes to this file, change + * tools/codegen/core/gen_static_metadata.py, and then re-run it. + * + * This file contains the mapping from the index of each metadata element in the + * grpc static metadata table to the index of that element in the hpack static + * metadata table. If the element is not contained in the static hpack table, + * then the returned index is 0. + */ + +#ifndef GRPC_CORE_EXT_TRANSPORT_CHTTP2_TRANSPORT_HPACK_MAPPING_H +#define GRPC_CORE_EXT_TRANSPORT_CHTTP2_TRANSPORT_HPACK_MAPPING_H + +#include <grpc/support/port_platform.h> + +#include "src/core/lib/transport/static_metadata.h" + +extern const uint8_t grpc_hpack_static_mdelem_indices[GRPC_STATIC_MDELEM_COUNT]; + +#endif /* GRPC_CORE_EXT_TRANSPORT_CHTTP2_TRANSPORT_HPACK_MAPPING_H */ diff --git a/src/core/ext/transport/chttp2/transport/hpack_table.cc b/src/core/ext/transport/chttp2/transport/hpack_table.cc index 7929258356..117679ab75 100644 --- a/src/core/ext/transport/chttp2/transport/hpack_table.cc +++ b/src/core/ext/transport/chttp2/transport/hpack_table.cc @@ -27,8 +27,10 @@ #include <grpc/support/log.h> #include <grpc/support/string_util.h> +#include "src/core/ext/transport/chttp2/transport/hpack_mapping.h" #include "src/core/lib/debug/trace.h" #include "src/core/lib/gpr/murmur_hash.h" +#include "src/core/lib/transport/static_metadata.h" extern grpc_core::TraceFlag grpc_http_trace; @@ -366,3 +368,30 @@ grpc_chttp2_hptbl_find_result grpc_chttp2_hptbl_find( return r; } + +static size_t get_base64_encoded_size(size_t raw_length) { + static const uint8_t tail_xtra[3] = {0, 2, 3}; + return raw_length / 3 * 4 + tail_xtra[raw_length % 3]; +} + +size_t grpc_chttp2_get_size_in_hpack_table(grpc_mdelem elem, + bool use_true_binary_metadata) { + size_t overhead_and_key = 32 + GRPC_SLICE_LENGTH(GRPC_MDKEY(elem)); + size_t value_len = GRPC_SLICE_LENGTH(GRPC_MDVALUE(elem)); + if (grpc_is_binary_header(GRPC_MDKEY(elem))) { + return overhead_and_key + (use_true_binary_metadata + ? value_len + 1 + : get_base64_encoded_size(value_len)); + } else { + return overhead_and_key + value_len; + } +} + +uint8_t grpc_chttp2_get_static_hpack_table_index(grpc_mdelem md) { + if (GRPC_MDELEM_STORAGE(md) == GRPC_MDELEM_STORAGE_STATIC) { + return grpc_hpack_static_mdelem_indices[GRPC_MDELEM_DATA(md) - + grpc_static_mdelem_table]; + } else { + return 0; + } +} diff --git a/src/core/ext/transport/chttp2/transport/hpack_table.h b/src/core/ext/transport/chttp2/transport/hpack_table.h index 98026a4ba4..a0ffc6fab7 100644 --- a/src/core/ext/transport/chttp2/transport/hpack_table.h +++ b/src/core/ext/transport/chttp2/transport/hpack_table.h @@ -83,6 +83,15 @@ grpc_mdelem grpc_chttp2_hptbl_lookup(const grpc_chttp2_hptbl* tbl, /* add a table entry to the index */ grpc_error* grpc_chttp2_hptbl_add(grpc_chttp2_hptbl* tbl, grpc_mdelem md) GRPC_MUST_USE_RESULT; + +size_t grpc_chttp2_get_size_in_hpack_table(grpc_mdelem elem, + bool use_true_binary_metadata); + +/* Returns the static hpack table index that corresponds to /a elem. Returns 0 + if /a elem is not statically stored or if it is not in the static hpack + table */ +uint8_t grpc_chttp2_get_static_hpack_table_index(grpc_mdelem md); + /* Find a key/value pair in the table... returns the index in the table of the most similar entry, or 0 if the value was not found */ typedef struct { diff --git a/src/core/ext/transport/chttp2/transport/parsing.cc b/src/core/ext/transport/chttp2/transport/parsing.cc index 0694ad743b..f532b084c9 100644 --- a/src/core/ext/transport/chttp2/transport/parsing.cc +++ b/src/core/ext/transport/chttp2/transport/parsing.cc @@ -409,67 +409,81 @@ static void on_initial_header(void* tp, grpc_mdelem md) { gpr_free(value); } - if (grpc_slice_eq(GRPC_MDKEY(md), GRPC_MDSTR_GRPC_STATUS) && - !grpc_mdelem_eq(md, GRPC_MDELEM_GRPC_STATUS_0)) { - /* TODO(ctiller): check for a status like " 0" */ - s->seen_error = true; - } + if (GRPC_MDELEM_STORAGE(md) == GRPC_MDELEM_STORAGE_STATIC) { + // We don't use grpc_mdelem_eq here to avoid executing additional + // instructions. The reasoning is if the payload is not equal, we already + // know that the metadata elements are not equal because the md is + // confirmed to be static. If we had used grpc_mdelem_eq here, then if the + // payloads are not equal, grpc_mdelem_eq executes more instructions to + // determine if they're equal or not. + if (md.payload == GRPC_MDELEM_GRPC_STATUS_1.payload || + md.payload == GRPC_MDELEM_GRPC_STATUS_2.payload) { + s->seen_error = true; + } + } else { + if (grpc_slice_eq(GRPC_MDKEY(md), GRPC_MDSTR_GRPC_STATUS) && + !grpc_mdelem_eq(md, GRPC_MDELEM_GRPC_STATUS_0)) { + /* TODO(ctiller): check for a status like " 0" */ + s->seen_error = true; + } - if (grpc_slice_eq(GRPC_MDKEY(md), GRPC_MDSTR_GRPC_TIMEOUT)) { - grpc_millis* cached_timeout = - static_cast<grpc_millis*>(grpc_mdelem_get_user_data(md, free_timeout)); - grpc_millis timeout; - if (cached_timeout != nullptr) { - timeout = *cached_timeout; - } else { - if (GPR_UNLIKELY( - !grpc_http2_decode_timeout(GRPC_MDVALUE(md), &timeout))) { - char* val = grpc_slice_to_c_string(GRPC_MDVALUE(md)); - gpr_log(GPR_ERROR, "Ignoring bad timeout value '%s'", val); - gpr_free(val); - timeout = GRPC_MILLIS_INF_FUTURE; + if (grpc_slice_eq(GRPC_MDKEY(md), GRPC_MDSTR_GRPC_TIMEOUT)) { + grpc_millis* cached_timeout = static_cast<grpc_millis*>( + grpc_mdelem_get_user_data(md, free_timeout)); + grpc_millis timeout; + if (cached_timeout != nullptr) { + timeout = *cached_timeout; + } else { + if (GPR_UNLIKELY( + !grpc_http2_decode_timeout(GRPC_MDVALUE(md), &timeout))) { + char* val = grpc_slice_to_c_string(GRPC_MDVALUE(md)); + gpr_log(GPR_ERROR, "Ignoring bad timeout value '%s'", val); + gpr_free(val); + timeout = GRPC_MILLIS_INF_FUTURE; + } + if (GRPC_MDELEM_IS_INTERNED(md)) { + /* store the result */ + cached_timeout = + static_cast<grpc_millis*>(gpr_malloc(sizeof(grpc_millis))); + *cached_timeout = timeout; + grpc_mdelem_set_user_data(md, free_timeout, cached_timeout); + } } - if (GRPC_MDELEM_IS_INTERNED(md)) { - /* store the result */ - cached_timeout = - static_cast<grpc_millis*>(gpr_malloc(sizeof(grpc_millis))); - *cached_timeout = timeout; - grpc_mdelem_set_user_data(md, free_timeout, cached_timeout); + if (timeout != GRPC_MILLIS_INF_FUTURE) { + grpc_chttp2_incoming_metadata_buffer_set_deadline( + &s->metadata_buffer[0], grpc_core::ExecCtx::Get()->Now() + timeout); } + GRPC_MDELEM_UNREF(md); + return; } - if (timeout != GRPC_MILLIS_INF_FUTURE) { - grpc_chttp2_incoming_metadata_buffer_set_deadline( - &s->metadata_buffer[0], grpc_core::ExecCtx::Get()->Now() + timeout); - } + } + + const size_t new_size = s->metadata_buffer[0].size + GRPC_MDELEM_LENGTH(md); + const size_t metadata_size_limit = + t->settings[GRPC_ACKED_SETTINGS] + [GRPC_CHTTP2_SETTINGS_MAX_HEADER_LIST_SIZE]; + if (new_size > metadata_size_limit) { + gpr_log(GPR_DEBUG, + "received initial metadata size exceeds limit (%" PRIuPTR + " vs. %" PRIuPTR ")", + new_size, metadata_size_limit); + grpc_chttp2_cancel_stream( + t, s, + grpc_error_set_int(GRPC_ERROR_CREATE_FROM_STATIC_STRING( + "received initial metadata size exceeds limit"), + GRPC_ERROR_INT_GRPC_STATUS, + GRPC_STATUS_RESOURCE_EXHAUSTED)); + grpc_chttp2_parsing_become_skip_parser(t); + s->seen_error = true; GRPC_MDELEM_UNREF(md); } else { - const size_t new_size = s->metadata_buffer[0].size + GRPC_MDELEM_LENGTH(md); - const size_t metadata_size_limit = - t->settings[GRPC_ACKED_SETTINGS] - [GRPC_CHTTP2_SETTINGS_MAX_HEADER_LIST_SIZE]; - if (new_size > metadata_size_limit) { - gpr_log(GPR_DEBUG, - "received initial metadata size exceeds limit (%" PRIuPTR - " vs. %" PRIuPTR ")", - new_size, metadata_size_limit); - grpc_chttp2_cancel_stream( - t, s, - grpc_error_set_int( - GRPC_ERROR_CREATE_FROM_STATIC_STRING( - "received initial metadata size exceeds limit"), - GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_RESOURCE_EXHAUSTED)); + grpc_error* error = + grpc_chttp2_incoming_metadata_buffer_add(&s->metadata_buffer[0], md); + if (error != GRPC_ERROR_NONE) { + grpc_chttp2_cancel_stream(t, s, error); grpc_chttp2_parsing_become_skip_parser(t); s->seen_error = true; GRPC_MDELEM_UNREF(md); - } else { - grpc_error* error = - grpc_chttp2_incoming_metadata_buffer_add(&s->metadata_buffer[0], md); - if (error != GRPC_ERROR_NONE) { - grpc_chttp2_cancel_stream(t, s, error); - grpc_chttp2_parsing_become_skip_parser(t); - s->seen_error = true; - GRPC_MDELEM_UNREF(md); - } } } } @@ -491,8 +505,19 @@ static void on_trailing_header(void* tp, grpc_mdelem md) { gpr_free(value); } - if (grpc_slice_eq(GRPC_MDKEY(md), GRPC_MDSTR_GRPC_STATUS) && - !grpc_mdelem_eq(md, GRPC_MDELEM_GRPC_STATUS_0)) { + if (GRPC_MDELEM_STORAGE(md) == GRPC_MDELEM_STORAGE_STATIC) { + // We don't use grpc_mdelem_eq here to avoid executing additional + // instructions. The reasoning is if the payload is not equal, we already + // know that the metadata elements are not equal because the md is + // confirmed to be static. If we had used grpc_mdelem_eq here, then if the + // payloads are not equal, grpc_mdelem_eq executes more instructions to + // determine if they're equal or not. + if (md.payload == GRPC_MDELEM_GRPC_STATUS_1.payload || + md.payload == GRPC_MDELEM_GRPC_STATUS_2.payload) { + s->seen_error = true; + } + } else if (grpc_slice_eq(GRPC_MDKEY(md), GRPC_MDSTR_GRPC_STATUS) && + !grpc_mdelem_eq(md, GRPC_MDELEM_GRPC_STATUS_0)) { /* TODO(ctiller): check for a status like " 0" */ s->seen_error = true; } diff --git a/src/core/lib/channel/channel_trace.h b/src/core/lib/channel/channel_trace.h index 230faa483e..94fea20b45 100644 --- a/src/core/lib/channel/channel_trace.h +++ b/src/core/lib/channel/channel_trace.h @@ -63,7 +63,7 @@ class ChannelTrace { // stack, determine if it makes more sense to accept a char* instead of a // slice. void AddTraceEventWithReference(Severity severity, grpc_slice data, - RefCountedPtr<BaseNode> referenced_channel); + RefCountedPtr<BaseNode> referenced_entity); // Creates and returns the raw grpc_json object, so a parent channelz // object may incorporate the json before rendering. diff --git a/src/core/lib/iomgr/ev_epollsig_linux.cc b/src/core/lib/iomgr/ev_epollsig_linux.cc deleted file mode 100644 index 5695ac795d..0000000000 --- a/src/core/lib/iomgr/ev_epollsig_linux.cc +++ /dev/null @@ -1,1743 +0,0 @@ -/* - * - * Copyright 2016 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#include <grpc/support/port_platform.h> - -#include "src/core/lib/iomgr/port.h" - -#include <grpc/grpc_posix.h> -#include <grpc/support/log.h> - -/* This polling engine is only relevant on linux kernels supporting epoll() */ -#ifdef GRPC_LINUX_EPOLL_CREATE1 - -#include "src/core/lib/iomgr/ev_epollsig_linux.h" - -#include <assert.h> -#include <errno.h> -#include <limits.h> -#include <poll.h> -#include <pthread.h> -#include <signal.h> -#include <string.h> -#include <sys/epoll.h> -#include <sys/socket.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/string_util.h> - -#include "src/core/lib/debug/stats.h" -#include "src/core/lib/gpr/tls.h" -#include "src/core/lib/gpr/useful.h" -#include "src/core/lib/gprpp/manual_constructor.h" -#include "src/core/lib/iomgr/block_annotate.h" -#include "src/core/lib/iomgr/ev_posix.h" -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/lockfree_event.h" -#include "src/core/lib/iomgr/timer.h" -#include "src/core/lib/iomgr/wakeup_fd_posix.h" -#include "src/core/lib/profiling/timers.h" - -#define GRPC_POLLSET_KICK_BROADCAST ((grpc_pollset_worker*)1) - -#define GRPC_POLLING_TRACE(...) \ - if (grpc_polling_trace.enabled()) { \ - gpr_log(GPR_INFO, __VA_ARGS__); \ - } - -static int grpc_wakeup_signal = -1; -static bool is_grpc_wakeup_signal_initialized = false; - -/* Implements the function defined in grpc_posix.h. This function might be - * called before even calling grpc_init() to set either a different signal to - * use. If signum == -1, then the use of signals is disabled */ -void grpc_use_signal(int signum) { - grpc_wakeup_signal = signum; - is_grpc_wakeup_signal_initialized = true; - - if (grpc_wakeup_signal < 0) { - gpr_log(GPR_INFO, - "Use of signals is disabled. Epoll engine will not be used"); - } else { - gpr_log(GPR_INFO, "epoll engine will be using signal: %d", - grpc_wakeup_signal); - } -} - -struct polling_island; - -typedef enum { - POLL_OBJ_FD, - POLL_OBJ_POLLSET, - POLL_OBJ_POLLSET_SET -} poll_obj_type; - -typedef struct poll_obj { -#ifndef NDEBUG - poll_obj_type obj_type; -#endif - gpr_mu mu; - struct polling_island* pi; -} poll_obj; - -const char* poll_obj_string(poll_obj_type po_type) { - switch (po_type) { - case POLL_OBJ_FD: - return "fd"; - case POLL_OBJ_POLLSET: - return "pollset"; - case POLL_OBJ_POLLSET_SET: - return "pollset_set"; - } - - GPR_UNREACHABLE_CODE(return "UNKNOWN"); -} - - /******************************************************************************* - * Fd Declarations - */ - -#define FD_FROM_PO(po) ((grpc_fd*)(po)) - -struct grpc_fd { - poll_obj po; - - int fd; - /* refst format: - bit 0 : 1=Active / 0=Orphaned - bits 1-n : refcount - Ref/Unref by two to avoid altering the orphaned bit */ - gpr_atm refst; - - /* The fd is either closed or we relinquished control of it. In either - cases, this indicates that the 'fd' on this structure is no longer - valid */ - bool orphaned; - - grpc_core::ManualConstructor<grpc_core::LockfreeEvent> read_closure; - grpc_core::ManualConstructor<grpc_core::LockfreeEvent> write_closure; - grpc_core::ManualConstructor<grpc_core::LockfreeEvent> error_closure; - - struct grpc_fd* freelist_next; - grpc_closure* on_done_closure; - - grpc_iomgr_object iomgr_object; - - /* Do we need to track EPOLLERR events separately? */ - bool track_err; -}; - -/* Reference counting for fds */ -#ifndef NDEBUG -static void fd_ref(grpc_fd* fd, const char* reason, const char* file, int line); -static void fd_unref(grpc_fd* fd, const char* reason, const char* file, - int line); -#define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__) -#define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__) -#else -static void fd_ref(grpc_fd* fd); -static void fd_unref(grpc_fd* fd); -#define GRPC_FD_REF(fd, reason) fd_ref(fd) -#define GRPC_FD_UNREF(fd, reason) fd_unref(fd) -#endif - -static void fd_global_init(void); -static void fd_global_shutdown(void); - -/******************************************************************************* - * Polling island Declarations - */ - -#ifndef NDEBUG - -#define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__) -#define PI_UNREF(p, r) pi_unref_dbg((p), (r), __FILE__, __LINE__) - -#else - -#define PI_ADD_REF(p, r) pi_add_ref((p)) -#define PI_UNREF(p, r) pi_unref((p)) - -#endif - -/* This is also used as grpc_workqueue (by directly casing it) */ -typedef struct polling_island { - gpr_mu mu; - /* Ref count. Use PI_ADD_REF() and PI_UNREF() macros to increment/decrement - the refcount. - Once the ref count becomes zero, this structure is destroyed which means - we should ensure that there is never a scenario where a PI_ADD_REF() is - racing with a PI_UNREF() that just made the ref_count zero. */ - gpr_atm ref_count; - - /* Pointer to the polling_island this merged into. - * merged_to value is only set once in polling_island's lifetime (and that too - * only if the island is merged with another island). Because of this, we can - * use gpr_atm type here so that we can do atomic access on this and reduce - * lock contention on 'mu' mutex. - * - * Note that if this field is not NULL (i.e not 0), all the remaining fields - * (except mu and ref_count) are invalid and must be ignored. */ - gpr_atm merged_to; - - /* Number of threads currently polling on this island */ - gpr_atm poller_count; - - /* The fd of the underlying epoll set */ - int epoll_fd; - - /* The file descriptors in the epoll set */ - size_t fd_cnt; - size_t fd_capacity; - grpc_fd** fds; -} polling_island; - -/******************************************************************************* - * Pollset Declarations - */ -struct grpc_pollset_worker { - /* Thread id of this worker */ - pthread_t pt_id; - - /* Used to prevent a worker from getting kicked multiple times */ - gpr_atm is_kicked; - struct grpc_pollset_worker* next; - struct grpc_pollset_worker* prev; -}; - -struct grpc_pollset { - poll_obj po; - - grpc_pollset_worker root_worker; - bool kicked_without_pollers; - - bool shutting_down; /* Is the pollset shutting down ? */ - bool finish_shutdown_called; /* Is the 'finish_shutdown_locked()' called ? */ - grpc_closure* shutdown_done; /* Called after after shutdown is complete */ -}; - -/******************************************************************************* - * Pollset-set Declarations - */ -struct grpc_pollset_set { - poll_obj po; -}; - -/******************************************************************************* - * Common helpers - */ - -static bool append_error(grpc_error** composite, grpc_error* error, - const char* desc) { - if (error == GRPC_ERROR_NONE) return true; - if (*composite == GRPC_ERROR_NONE) { - *composite = GRPC_ERROR_CREATE_FROM_COPIED_STRING(desc); - } - *composite = grpc_error_add_child(*composite, error); - return false; -} - -/******************************************************************************* - * Polling island Definitions - */ - -/* The wakeup fd that is used to wake up all threads in a Polling island. This - is useful in the polling island merge operation where we need to wakeup all - the threads currently polling the smaller polling island (so that they can - start polling the new/merged polling island) - - NOTE: This fd is initialized to be readable and MUST NOT be consumed i.e the - threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */ -static grpc_wakeup_fd polling_island_wakeup_fd; - -/* The polling island being polled right now. - See comments in workqueue_maybe_wakeup for why this is tracked. */ -static __thread polling_island* g_current_thread_polling_island; - -/* Forward declaration */ -static void polling_island_delete(polling_island* pi); - -#ifdef GRPC_TSAN -/* Currently TSAN may incorrectly flag data races between epoll_ctl and - epoll_wait for any grpc_fd structs that are added to the epoll set via - epoll_ctl and are returned (within a very short window) via epoll_wait(). - - To work-around this race, we establish a happens-before relation between - the code just-before epoll_ctl() and the code after epoll_wait() by using - this atomic */ -gpr_atm g_epoll_sync; -#endif /* defined(GRPC_TSAN) */ - -static void pi_add_ref(polling_island* pi); -static void pi_unref(polling_island* pi); - -#ifndef NDEBUG -static void pi_add_ref_dbg(polling_island* pi, const char* reason, - const char* file, int line) { - if (grpc_polling_trace.enabled()) { - gpr_atm old_cnt = gpr_atm_acq_load(&pi->ref_count); - gpr_log(GPR_INFO, - "Add ref pi: %p, old:%" PRIdPTR " -> new:%" PRIdPTR - " (%s) - (%s, %d)", - pi, old_cnt, old_cnt + 1, reason, file, line); - } - pi_add_ref(pi); -} - -static void pi_unref_dbg(polling_island* pi, const char* reason, - const char* file, int line) { - if (grpc_polling_trace.enabled()) { - gpr_atm old_cnt = gpr_atm_acq_load(&pi->ref_count); - gpr_log(GPR_INFO, - "Unref pi: %p, old:%" PRIdPTR " -> new:%" PRIdPTR - " (%s) - (%s, %d)", - pi, old_cnt, (old_cnt - 1), reason, file, line); - } - pi_unref(pi); -} -#endif - -static void pi_add_ref(polling_island* pi) { - gpr_atm_no_barrier_fetch_add(&pi->ref_count, 1); -} - -static void pi_unref(polling_island* pi) { - /* If ref count went to zero, delete the polling island. - Note that this deletion not be done under a lock. Once the ref count goes - to zero, we are guaranteed that no one else holds a reference to the - polling island (and that there is no racing pi_add_ref() call either). - - Also, if we are deleting the polling island and the merged_to field is - non-empty, we should remove a ref to the merged_to polling island - */ - if (1 == gpr_atm_full_fetch_add(&pi->ref_count, -1)) { - polling_island* next = (polling_island*)gpr_atm_acq_load(&pi->merged_to); - polling_island_delete(pi); - if (next != nullptr) { - PI_UNREF(next, "pi_delete"); /* Recursive call */ - } - } -} - -/* The caller is expected to hold pi->mu lock before calling this function */ -static void polling_island_add_fds_locked(polling_island* pi, grpc_fd** fds, - size_t fd_count, bool add_fd_refs, - grpc_error** error) { - int err; - size_t i; - struct epoll_event ev; - char* err_msg; - const char* err_desc = "polling_island_add_fds"; - -#ifdef GRPC_TSAN - /* See the definition of g_epoll_sync for more context */ - gpr_atm_rel_store(&g_epoll_sync, (gpr_atm)0); -#endif /* defined(GRPC_TSAN) */ - - for (i = 0; i < fd_count; i++) { - ev.events = static_cast<uint32_t>(EPOLLIN | EPOLLOUT | EPOLLET); - /* Use the least significant bit of ev.data.ptr to store track_err to avoid - * synchronization issues when accessing it after receiving an event */ - ev.data.ptr = reinterpret_cast<void*>(reinterpret_cast<intptr_t>(fds[i]) | - (fds[i]->track_err ? 1 : 0)); - err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, fds[i]->fd, &ev); - - if (err < 0) { - if (errno != EEXIST) { - gpr_asprintf( - &err_msg, - "epoll_ctl (epoll_fd: %d) add fd: %d failed with error: %d (%s)", - pi->epoll_fd, fds[i]->fd, errno, strerror(errno)); - append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); - gpr_free(err_msg); - } - - continue; - } - - if (pi->fd_cnt == pi->fd_capacity) { - pi->fd_capacity = GPR_MAX(pi->fd_capacity + 8, pi->fd_cnt * 3 / 2); - pi->fds = static_cast<grpc_fd**>( - gpr_realloc(pi->fds, sizeof(grpc_fd*) * pi->fd_capacity)); - } - - pi->fds[pi->fd_cnt++] = fds[i]; - if (add_fd_refs) { - GRPC_FD_REF(fds[i], "polling_island"); - } - } -} - -/* The caller is expected to hold pi->mu before calling this */ -static void polling_island_add_wakeup_fd_locked(polling_island* pi, - grpc_wakeup_fd* wakeup_fd, - grpc_error** error) { - struct epoll_event ev; - int err; - char* err_msg; - const char* err_desc = "polling_island_add_wakeup_fd"; - - ev.events = static_cast<uint32_t>(EPOLLIN | EPOLLET); - ev.data.ptr = wakeup_fd; - err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, - GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), &ev); - if (err < 0 && errno != EEXIST) { - gpr_asprintf(&err_msg, - "epoll_ctl (epoll_fd: %d) add wakeup fd: %d failed with " - "error: %d (%s)", - pi->epoll_fd, GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), errno, - strerror(errno)); - append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); - gpr_free(err_msg); - } -} - -/* The caller is expected to hold pi->mu lock before calling this function */ -static void polling_island_remove_all_fds_locked(polling_island* pi, - bool remove_fd_refs, - grpc_error** error) { - int err; - size_t i; - char* err_msg; - const char* err_desc = "polling_island_remove_fds"; - - for (i = 0; i < pi->fd_cnt; i++) { - err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, pi->fds[i]->fd, nullptr); - if (err < 0 && errno != ENOENT) { - gpr_asprintf(&err_msg, - "epoll_ctl (epoll_fd: %d) delete fds[%zu]: %d failed with " - "error: %d (%s)", - pi->epoll_fd, i, pi->fds[i]->fd, errno, strerror(errno)); - append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); - gpr_free(err_msg); - } - - if (remove_fd_refs) { - GRPC_FD_UNREF(pi->fds[i], "polling_island"); - } - } - - pi->fd_cnt = 0; -} - -/* The caller is expected to hold pi->mu lock before calling this function */ -static void polling_island_remove_fd_locked(polling_island* pi, grpc_fd* fd, - grpc_error** error) { - int err; - size_t i; - char* err_msg; - const char* err_desc = "polling_island_remove_fd"; - - /* If fd is already closed, then it would have been automatically been removed - from the epoll set */ - err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, fd->fd, nullptr); - if (err < 0 && errno != ENOENT) { - gpr_asprintf( - &err_msg, - "epoll_ctl (epoll_fd: %d) del fd: %d failed with error: %d (%s)", - pi->epoll_fd, fd->fd, errno, strerror(errno)); - append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); - gpr_free(err_msg); - } - - for (i = 0; i < pi->fd_cnt; i++) { - if (pi->fds[i] == fd) { - pi->fds[i] = pi->fds[--pi->fd_cnt]; - GRPC_FD_UNREF(fd, "polling_island"); - break; - } - } -} - -/* Might return NULL in case of an error */ -static polling_island* polling_island_create(grpc_fd* initial_fd, - grpc_error** error) { - polling_island* pi = nullptr; - const char* err_desc = "polling_island_create"; - - *error = GRPC_ERROR_NONE; - - pi = static_cast<polling_island*>(gpr_malloc(sizeof(*pi))); - gpr_mu_init(&pi->mu); - pi->fd_cnt = 0; - pi->fd_capacity = 0; - pi->fds = nullptr; - pi->epoll_fd = -1; - - gpr_atm_rel_store(&pi->ref_count, 0); - gpr_atm_rel_store(&pi->poller_count, 0); - gpr_atm_rel_store(&pi->merged_to, (gpr_atm) nullptr); - - pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC); - - if (pi->epoll_fd < 0) { - append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc); - goto done; - } - - if (initial_fd != nullptr) { - polling_island_add_fds_locked(pi, &initial_fd, 1, true, error); - } - -done: - if (*error != GRPC_ERROR_NONE) { - polling_island_delete(pi); - pi = nullptr; - } - return pi; -} - -static void polling_island_delete(polling_island* pi) { - GPR_ASSERT(pi->fd_cnt == 0); - - if (pi->epoll_fd >= 0) { - close(pi->epoll_fd); - } - gpr_mu_destroy(&pi->mu); - gpr_free(pi->fds); - gpr_free(pi); -} - -/* Attempts to gets the last polling island in the linked list (liked by the - * 'merged_to' field). Since this does not lock the polling island, there are no - * guarantees that the island returned is the last island */ -static polling_island* polling_island_maybe_get_latest(polling_island* pi) { - polling_island* next = (polling_island*)gpr_atm_acq_load(&pi->merged_to); - while (next != nullptr) { - pi = next; - next = (polling_island*)gpr_atm_acq_load(&pi->merged_to); - } - - return pi; -} - -/* Gets the lock on the *latest* polling island i.e the last polling island in - the linked list (linked by the 'merged_to' field). Call gpr_mu_unlock on the - returned polling island's mu. - Usage: To lock/unlock polling island "pi", do the following: - polling_island *pi_latest = polling_island_lock(pi); - ... - ... critical section .. - ... - gpr_mu_unlock(&pi_latest->mu); // NOTE: use pi_latest->mu. NOT pi->mu */ -static polling_island* polling_island_lock(polling_island* pi) { - polling_island* next = nullptr; - - while (true) { - next = (polling_island*)gpr_atm_acq_load(&pi->merged_to); - if (next == nullptr) { - /* Looks like 'pi' is the last node in the linked list but unless we check - this by holding the pi->mu lock, we cannot be sure (i.e without the - pi->mu lock, we don't prevent island merges). - To be absolutely sure, check once more by holding the pi->mu lock */ - gpr_mu_lock(&pi->mu); - next = (polling_island*)gpr_atm_acq_load(&pi->merged_to); - if (next == nullptr) { - /* pi is infact the last node and we have the pi->mu lock. we're done */ - break; - } - - /* pi->merged_to is not NULL i.e pi isn't the last node anymore. pi->mu - * isn't the lock we are interested in. Continue traversing the list */ - gpr_mu_unlock(&pi->mu); - } - - pi = next; - } - - return pi; -} - -/* Gets the lock on the *latest* polling islands in the linked lists pointed by - *p and *q (and also updates *p and *q to point to the latest polling islands) - - This function is needed because calling the following block of code to obtain - locks on polling islands (*p and *q) is prone to deadlocks. - { - polling_island_lock(*p, true); - polling_island_lock(*q, true); - } - - Usage/example: - polling_island *p1; - polling_island *p2; - .. - polling_island_lock_pair(&p1, &p2); - .. - .. Critical section with both p1 and p2 locked - .. - // Release locks: Always call polling_island_unlock_pair() to release locks - polling_island_unlock_pair(p1, p2); -*/ -static void polling_island_lock_pair(polling_island** p, polling_island** q) { - polling_island* pi_1 = *p; - polling_island* pi_2 = *q; - polling_island* next_1 = nullptr; - polling_island* next_2 = nullptr; - - /* The algorithm is simple: - - Go to the last polling islands in the linked lists *pi_1 and *pi_2 (and - keep updating pi_1 and pi_2) - - Then obtain locks on the islands by following a lock order rule of - locking polling_island with lower address first - Special case: Before obtaining the locks, check if pi_1 and pi_2 are - pointing to the same island. If that is the case, we can just call - polling_island_lock() - - After obtaining both the locks, double check that the polling islands - are still the last polling islands in their respective linked lists - (this is because there might have been polling island merges before - we got the lock) - - If the polling islands are the last islands, we are done. If not, - release the locks and continue the process from the first step */ - while (true) { - next_1 = (polling_island*)gpr_atm_acq_load(&pi_1->merged_to); - while (next_1 != nullptr) { - pi_1 = next_1; - next_1 = (polling_island*)gpr_atm_acq_load(&pi_1->merged_to); - } - - next_2 = (polling_island*)gpr_atm_acq_load(&pi_2->merged_to); - while (next_2 != nullptr) { - pi_2 = next_2; - next_2 = (polling_island*)gpr_atm_acq_load(&pi_2->merged_to); - } - - if (pi_1 == pi_2) { - pi_1 = pi_2 = polling_island_lock(pi_1); - break; - } - - if (pi_1 < pi_2) { - gpr_mu_lock(&pi_1->mu); - gpr_mu_lock(&pi_2->mu); - } else { - gpr_mu_lock(&pi_2->mu); - gpr_mu_lock(&pi_1->mu); - } - - next_1 = (polling_island*)gpr_atm_acq_load(&pi_1->merged_to); - next_2 = (polling_island*)gpr_atm_acq_load(&pi_2->merged_to); - if (next_1 == nullptr && next_2 == nullptr) { - break; - } - - gpr_mu_unlock(&pi_1->mu); - gpr_mu_unlock(&pi_2->mu); - } - - *p = pi_1; - *q = pi_2; -} - -static void polling_island_unlock_pair(polling_island* p, polling_island* q) { - if (p == q) { - gpr_mu_unlock(&p->mu); - } else { - gpr_mu_unlock(&p->mu); - gpr_mu_unlock(&q->mu); - } -} - -static polling_island* polling_island_merge(polling_island* p, - polling_island* q, - grpc_error** error) { - /* Get locks on both the polling islands */ - polling_island_lock_pair(&p, &q); - - if (p != q) { - /* Make sure that p points to the polling island with fewer fds than q */ - if (p->fd_cnt > q->fd_cnt) { - GPR_SWAP(polling_island*, p, q); - } - - /* Merge p with q i.e move all the fds from p (The one with fewer fds) to q - Note that the refcounts on the fds being moved will not change here. - This is why the last param in the following two functions is 'false') */ - polling_island_add_fds_locked(q, p->fds, p->fd_cnt, false, error); - polling_island_remove_all_fds_locked(p, false, error); - - /* Wakeup all the pollers (if any) on p so that they pickup this change */ - polling_island_add_wakeup_fd_locked(p, &polling_island_wakeup_fd, error); - - /* Add the 'merged_to' link from p --> q */ - gpr_atm_rel_store(&p->merged_to, (gpr_atm)q); - PI_ADD_REF(q, "pi_merge"); /* To account for the new incoming ref from p */ - } - /* else if p == q, nothing needs to be done */ - - polling_island_unlock_pair(p, q); - - /* Return the merged polling island (Note that no merge would have happened - if p == q which is ok) */ - return q; -} - -static grpc_error* polling_island_global_init() { - grpc_error* error = GRPC_ERROR_NONE; - - error = grpc_wakeup_fd_init(&polling_island_wakeup_fd); - if (error == GRPC_ERROR_NONE) { - error = grpc_wakeup_fd_wakeup(&polling_island_wakeup_fd); - } - - return error; -} - -static void polling_island_global_shutdown() { - grpc_wakeup_fd_destroy(&polling_island_wakeup_fd); -} - -/******************************************************************************* - * Fd Definitions - */ - -/* We need to keep a freelist not because of any concerns of malloc performance - * but instead so that implementations with multiple threads in (for example) - * epoll_wait deal with the race between pollset removal and incoming poll - * notifications. - * - * The problem is that the poller ultimately holds a reference to this - * object, so it is very difficult to know when is safe to free it, at least - * without some expensive synchronization. - * - * If we keep the object freelisted, in the worst case losing this race just - * becomes a spurious read notification on a reused fd. - */ - -/* The alarm system needs to be able to wakeup 'some poller' sometimes - * (specifically when a new alarm needs to be triggered earlier than the next - * alarm 'epoch'). This wakeup_fd gives us something to alert on when such a - * case occurs. */ - -static grpc_fd* fd_freelist = nullptr; -static gpr_mu fd_freelist_mu; - -#ifndef NDEBUG -#define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__) -#define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__) -static void ref_by(grpc_fd* fd, int n, const char* reason, const char* file, - int line) { - if (grpc_trace_fd_refcount.enabled()) { - gpr_log(GPR_DEBUG, - "FD %d %p ref %d %" PRIdPTR " -> %" PRIdPTR " [%s; %s:%d]", - fd->fd, fd, n, gpr_atm_no_barrier_load(&fd->refst), - gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line); - } -#else -#define REF_BY(fd, n, reason) ref_by(fd, n) -#define UNREF_BY(fd, n, reason) unref_by(fd, n) -static void ref_by(grpc_fd* fd, int n) { -#endif - GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0); -} - -#ifndef NDEBUG -static void unref_by(grpc_fd* fd, int n, const char* reason, const char* file, - int line) { - if (grpc_trace_fd_refcount.enabled()) { - gpr_log(GPR_DEBUG, - "FD %d %p unref %d %" PRIdPTR " -> %" PRIdPTR " [%s; %s:%d]", - fd->fd, fd, n, gpr_atm_no_barrier_load(&fd->refst), - gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line); - } -#else -static void unref_by(grpc_fd* fd, int n) { -#endif - gpr_atm old = gpr_atm_full_fetch_add(&fd->refst, -n); - if (old == n) { - /* Add the fd to the freelist */ - gpr_mu_lock(&fd_freelist_mu); - fd->freelist_next = fd_freelist; - fd_freelist = fd; - grpc_iomgr_unregister_object(&fd->iomgr_object); - - fd->read_closure->DestroyEvent(); - fd->write_closure->DestroyEvent(); - fd->error_closure->DestroyEvent(); - - gpr_mu_unlock(&fd_freelist_mu); - } else { - GPR_ASSERT(old > n); - } -} - -/* Increment refcount by two to avoid changing the orphan bit */ -#ifndef NDEBUG -static void fd_ref(grpc_fd* fd, const char* reason, const char* file, - int line) { - ref_by(fd, 2, reason, file, line); -} - -static void fd_unref(grpc_fd* fd, const char* reason, const char* file, - int line) { - unref_by(fd, 2, reason, file, line); -} -#else -static void fd_ref(grpc_fd* fd) { ref_by(fd, 2); } -static void fd_unref(grpc_fd* fd) { unref_by(fd, 2); } -#endif - -static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } - -static void fd_global_shutdown(void) { - gpr_mu_lock(&fd_freelist_mu); - gpr_mu_unlock(&fd_freelist_mu); - while (fd_freelist != nullptr) { - grpc_fd* fd = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - gpr_mu_destroy(&fd->po.mu); - gpr_free(fd); - } - gpr_mu_destroy(&fd_freelist_mu); -} - -static grpc_fd* fd_create(int fd, const char* name, bool track_err) { - grpc_fd* new_fd = nullptr; - - gpr_mu_lock(&fd_freelist_mu); - if (fd_freelist != nullptr) { - new_fd = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - } - gpr_mu_unlock(&fd_freelist_mu); - - if (new_fd == nullptr) { - new_fd = static_cast<grpc_fd*>(gpr_malloc(sizeof(grpc_fd))); - gpr_mu_init(&new_fd->po.mu); - new_fd->read_closure.Init(); - new_fd->write_closure.Init(); - new_fd->error_closure.Init(); - } - - /* Note: It is not really needed to get the new_fd->po.mu lock here. If this - * is a newly created fd (or an fd we got from the freelist), no one else - * would be holding a lock to it anyway. */ - gpr_mu_lock(&new_fd->po.mu); - new_fd->po.pi = nullptr; -#ifndef NDEBUG - new_fd->po.obj_type = POLL_OBJ_FD; -#endif - - gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1); - new_fd->fd = fd; - new_fd->orphaned = false; - new_fd->read_closure->InitEvent(); - new_fd->write_closure->InitEvent(); - new_fd->error_closure->InitEvent(); - new_fd->track_err = track_err; - - new_fd->freelist_next = nullptr; - new_fd->on_done_closure = nullptr; - - gpr_mu_unlock(&new_fd->po.mu); - - char* fd_name; - gpr_asprintf(&fd_name, "%s fd=%d", name, fd); - grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name); - gpr_free(fd_name); - return new_fd; -} - -static int fd_wrapped_fd(grpc_fd* fd) { - int ret_fd = -1; - gpr_mu_lock(&fd->po.mu); - if (!fd->orphaned) { - ret_fd = fd->fd; - } - gpr_mu_unlock(&fd->po.mu); - - return ret_fd; -} - -static void fd_orphan(grpc_fd* fd, grpc_closure* on_done, int* release_fd, - const char* reason) { - grpc_error* error = GRPC_ERROR_NONE; - polling_island* unref_pi = nullptr; - - gpr_mu_lock(&fd->po.mu); - fd->on_done_closure = on_done; - - /* Remove the active status but keep referenced. We want this grpc_fd struct - to be alive (and not added to freelist) until the end of this function */ - REF_BY(fd, 1, reason); - - /* Remove the fd from the polling island: - - Get a lock on the latest polling island (i.e the last island in the - linked list pointed by fd->po.pi). This is the island that - would actually contain the fd - - Remove the fd from the latest polling island - - Unlock the latest polling island - - Set fd->po.pi to NULL (but remove the ref on the polling island - before doing this.) */ - if (fd->po.pi != nullptr) { - polling_island* pi_latest = polling_island_lock(fd->po.pi); - polling_island_remove_fd_locked(pi_latest, fd, &error); - gpr_mu_unlock(&pi_latest->mu); - - unref_pi = fd->po.pi; - fd->po.pi = nullptr; - } - - /* If release_fd is not NULL, we should be relinquishing control of the file - descriptor fd->fd (but we still own the grpc_fd structure). */ - if (release_fd != nullptr) { - *release_fd = fd->fd; - } else { - close(fd->fd); - } - - fd->orphaned = true; - - GRPC_CLOSURE_SCHED(fd->on_done_closure, GRPC_ERROR_REF(error)); - - gpr_mu_unlock(&fd->po.mu); - UNREF_BY(fd, 2, reason); /* Drop the reference */ - if (unref_pi != nullptr) { - /* Unref stale polling island here, outside the fd lock above. - The polling island owns a workqueue which owns an fd, and unreffing - inside the lock can cause an eventual lock loop that makes TSAN very - unhappy. */ - PI_UNREF(unref_pi, "fd_orphan"); - } - if (error != GRPC_ERROR_NONE) { - const char* msg = grpc_error_string(error); - gpr_log(GPR_DEBUG, "fd_orphan: %s", msg); - } - GRPC_ERROR_UNREF(error); -} - -static bool fd_is_shutdown(grpc_fd* fd) { - return fd->read_closure->IsShutdown(); -} - -/* Might be called multiple times */ -static void fd_shutdown(grpc_fd* fd, grpc_error* why) { - if (fd->read_closure->SetShutdown(GRPC_ERROR_REF(why))) { - shutdown(fd->fd, SHUT_RDWR); - fd->write_closure->SetShutdown(GRPC_ERROR_REF(why)); - fd->error_closure->SetShutdown(GRPC_ERROR_REF(why)); - } - GRPC_ERROR_UNREF(why); -} - -static void fd_notify_on_read(grpc_fd* fd, grpc_closure* closure) { - fd->read_closure->NotifyOn(closure); -} - -static void fd_notify_on_write(grpc_fd* fd, grpc_closure* closure) { - fd->write_closure->NotifyOn(closure); -} - -static void fd_notify_on_error(grpc_fd* fd, grpc_closure* closure) { - fd->error_closure->NotifyOn(closure); -} - -static void fd_become_readable(grpc_fd* fd) { fd->read_closure->SetReady(); } - -static void fd_become_writable(grpc_fd* fd) { fd->write_closure->SetReady(); } - -static void fd_has_errors(grpc_fd* fd) { fd->error_closure->SetReady(); } - -/******************************************************************************* - * Pollset Definitions - */ -GPR_TLS_DECL(g_current_thread_pollset); -GPR_TLS_DECL(g_current_thread_worker); -static __thread bool g_initialized_sigmask; -static __thread sigset_t g_orig_sigmask; - -static void sig_handler(int sig_num) { -#ifdef GRPC_EPOLL_DEBUG - gpr_log(GPR_INFO, "Received signal %d", sig_num); -#endif -} - -static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); } - -/* Global state management */ -static grpc_error* pollset_global_init(void) { - gpr_tls_init(&g_current_thread_pollset); - gpr_tls_init(&g_current_thread_worker); - poller_kick_init(); - return GRPC_ERROR_NONE; -} - -static void pollset_global_shutdown(void) { - gpr_tls_destroy(&g_current_thread_pollset); - gpr_tls_destroy(&g_current_thread_worker); -} - -static grpc_error* pollset_worker_kick(grpc_pollset_worker* worker) { - grpc_error* err = GRPC_ERROR_NONE; - - /* Kick the worker only if it was not already kicked */ - if (gpr_atm_no_barrier_cas(&worker->is_kicked, static_cast<gpr_atm>(0), - static_cast<gpr_atm>(1))) { - GRPC_POLLING_TRACE( - "pollset_worker_kick: Kicking worker: %p (thread id: %ld)", - (void*)worker, (long int)worker->pt_id); - int err_num = pthread_kill(worker->pt_id, grpc_wakeup_signal); - if (err_num != 0) { - err = GRPC_OS_ERROR(err_num, "pthread_kill"); - } - } - return err; -} - -/* Return 1 if the pollset has active threads in pollset_work (pollset must - * be locked) */ -static int pollset_has_workers(grpc_pollset* p) { - return p->root_worker.next != &p->root_worker; -} - -static void remove_worker(grpc_pollset* p, grpc_pollset_worker* worker) { - worker->prev->next = worker->next; - worker->next->prev = worker->prev; -} - -static grpc_pollset_worker* pop_front_worker(grpc_pollset* p) { - if (pollset_has_workers(p)) { - grpc_pollset_worker* w = p->root_worker.next; - remove_worker(p, w); - return w; - } else { - return nullptr; - } -} - -static void push_back_worker(grpc_pollset* p, grpc_pollset_worker* worker) { - worker->next = &p->root_worker; - worker->prev = worker->next->prev; - worker->prev->next = worker->next->prev = worker; -} - -static void push_front_worker(grpc_pollset* p, grpc_pollset_worker* worker) { - worker->prev = &p->root_worker; - worker->next = worker->prev->next; - worker->prev->next = worker->next->prev = worker; -} - -/* p->mu must be held before calling this function */ -static grpc_error* pollset_kick(grpc_pollset* p, - grpc_pollset_worker* specific_worker) { - GPR_TIMER_SCOPE("pollset_kick", 0); - grpc_error* error = GRPC_ERROR_NONE; - GRPC_STATS_INC_POLLSET_KICK(); - const char* err_desc = "Kick Failure"; - grpc_pollset_worker* worker = specific_worker; - if (worker != nullptr) { - if (worker == GRPC_POLLSET_KICK_BROADCAST) { - if (pollset_has_workers(p)) { - GPR_TIMER_SCOPE("pollset_kick.broadcast", 0); - for (worker = p->root_worker.next; worker != &p->root_worker; - worker = worker->next) { - if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) { - append_error(&error, pollset_worker_kick(worker), err_desc); - } - } - } else { - p->kicked_without_pollers = true; - } - } else { - GPR_TIMER_MARK("kicked_specifically", 0); - if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) { - append_error(&error, pollset_worker_kick(worker), err_desc); - } - } - } else if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)p) { - /* Since worker == NULL, it means that we can kick "any" worker on this - pollset 'p'. If 'p' happens to be the same pollset this thread is - currently polling (i.e in pollset_work() function), then there is no need - to kick any other worker since the current thread can just absorb the - kick. This is the reason why we enter this case only when - g_current_thread_pollset is != p */ - - GPR_TIMER_MARK("kick_anonymous", 0); - worker = pop_front_worker(p); - if (worker != nullptr) { - GPR_TIMER_MARK("finally_kick", 0); - push_back_worker(p, worker); - append_error(&error, pollset_worker_kick(worker), err_desc); - } else { - GPR_TIMER_MARK("kicked_no_pollers", 0); - p->kicked_without_pollers = true; - } - } - - GRPC_LOG_IF_ERROR("pollset_kick", GRPC_ERROR_REF(error)); - return error; -} - -static void pollset_init(grpc_pollset* pollset, gpr_mu** mu) { - gpr_mu_init(&pollset->po.mu); - *mu = &pollset->po.mu; - pollset->po.pi = nullptr; -#ifndef NDEBUG - pollset->po.obj_type = POLL_OBJ_POLLSET; -#endif - - pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; - pollset->kicked_without_pollers = false; - - pollset->shutting_down = false; - pollset->finish_shutdown_called = false; - pollset->shutdown_done = nullptr; -} - -static int poll_deadline_to_millis_timeout(grpc_millis millis) { - if (millis == GRPC_MILLIS_INF_FUTURE) return -1; - grpc_millis delta = millis - grpc_core::ExecCtx::Get()->Now(); - if (delta > INT_MAX) - return INT_MAX; - else if (delta < 0) - return 0; - else - return static_cast<int>(delta); -} - -static void pollset_release_polling_island(grpc_pollset* ps, - const char* reason) { - if (ps->po.pi != nullptr) { - PI_UNREF(ps->po.pi, reason); - } - ps->po.pi = nullptr; -} - -static void finish_shutdown_locked(grpc_pollset* pollset) { - /* The pollset cannot have any workers if we are at this stage */ - GPR_ASSERT(!pollset_has_workers(pollset)); - - pollset->finish_shutdown_called = true; - - /* Release the ref and set pollset->po.pi to NULL */ - pollset_release_polling_island(pollset, "ps_shutdown"); - GRPC_CLOSURE_SCHED(pollset->shutdown_done, GRPC_ERROR_NONE); -} - -/* pollset->po.mu lock must be held by the caller before calling this */ -static void pollset_shutdown(grpc_pollset* pollset, grpc_closure* closure) { - GPR_TIMER_SCOPE("pollset_shutdown", 0); - GPR_ASSERT(!pollset->shutting_down); - pollset->shutting_down = true; - pollset->shutdown_done = closure; - pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - - /* If the pollset has any workers, we cannot call finish_shutdown_locked() - because it would release the underlying polling island. In such a case, we - let the last worker call finish_shutdown_locked() from pollset_work() */ - if (!pollset_has_workers(pollset)) { - GPR_ASSERT(!pollset->finish_shutdown_called); - GPR_TIMER_MARK("pollset_shutdown.finish_shutdown_locked", 0); - finish_shutdown_locked(pollset); - } -} - -/* pollset_shutdown is guaranteed to be called before pollset_destroy. So other - * than destroying the mutexes, there is nothing special that needs to be done - * here */ -static void pollset_destroy(grpc_pollset* pollset) { - GPR_ASSERT(!pollset_has_workers(pollset)); - gpr_mu_destroy(&pollset->po.mu); -} - -#define GRPC_EPOLL_MAX_EVENTS 100 -/* Note: sig_mask contains the signal mask to use *during* epoll_wait() */ -static void pollset_work_and_unlock(grpc_pollset* pollset, - grpc_pollset_worker* worker, int timeout_ms, - sigset_t* sig_mask, grpc_error** error) { - GPR_TIMER_SCOPE("pollset_work_and_unlock", 0); - struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; - int epoll_fd = -1; - int ep_rv; - polling_island* pi = nullptr; - char* err_msg; - const char* err_desc = "pollset_work_and_unlock"; - - /* We need to get the epoll_fd to wait on. The epoll_fd is in inside the - latest polling island pointed by pollset->po.pi - - Since epoll_fd is immutable, we can read it without obtaining the polling - island lock. There is however a possibility that the polling island (from - which we got the epoll_fd) got merged with another island while we are - in this function. This is still okay because in such a case, we will wakeup - right-away from epoll_wait() and pick up the latest polling_island the next - this function (i.e pollset_work_and_unlock()) is called */ - - if (pollset->po.pi == nullptr) { - pollset->po.pi = polling_island_create(nullptr, error); - if (pollset->po.pi == nullptr) { - return; /* Fatal error. We cannot continue */ - } - - PI_ADD_REF(pollset->po.pi, "ps"); - GRPC_POLLING_TRACE("pollset_work: pollset: %p created new pi: %p", - (void*)pollset, (void*)pollset->po.pi); - } - - pi = polling_island_maybe_get_latest(pollset->po.pi); - epoll_fd = pi->epoll_fd; - - /* Update the pollset->po.pi since the island being pointed by - pollset->po.pi maybe older than the one pointed by pi) */ - if (pollset->po.pi != pi) { - /* Always do PI_ADD_REF before PI_UNREF because PI_UNREF may cause the - polling island to be deleted */ - PI_ADD_REF(pi, "ps"); - PI_UNREF(pollset->po.pi, "ps"); - pollset->po.pi = pi; - } - - /* Add an extra ref so that the island does not get destroyed (which means - the epoll_fd won't be closed) while we are are doing an epoll_wait() on the - epoll_fd */ - PI_ADD_REF(pi, "ps_work"); - gpr_mu_unlock(&pollset->po.mu); - - gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1); - g_current_thread_polling_island = pi; - - GRPC_SCHEDULING_START_BLOCKING_REGION; - GRPC_STATS_INC_SYSCALL_POLL(); - ep_rv = - epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms, sig_mask); - GRPC_SCHEDULING_END_BLOCKING_REGION; - if (ep_rv < 0) { - if (errno != EINTR) { - gpr_asprintf(&err_msg, - "epoll_wait() epoll fd: %d failed with error: %d (%s)", - epoll_fd, errno, strerror(errno)); - append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc); - } else { - /* We were interrupted. Save an interation by doing a zero timeout - epoll_wait to see if there are any other events of interest */ - GRPC_POLLING_TRACE("pollset_work: pollset: %p, worker: %p received kick", - (void*)pollset, (void*)worker); - ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); - } - } - -#ifdef GRPC_TSAN - /* See the definition of g_poll_sync for more details */ - gpr_atm_acq_load(&g_epoll_sync); -#endif /* defined(GRPC_TSAN) */ - - for (int i = 0; i < ep_rv; ++i) { - void* data_ptr = ep_ev[i].data.ptr; - if (data_ptr == &polling_island_wakeup_fd) { - GRPC_POLLING_TRACE( - "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: " - "%d) got merged", - (void*)pollset, (void*)worker, epoll_fd); - /* This means that our polling island is merged with a different - island. We do not have to do anything here since the subsequent call - to the function pollset_work_and_unlock() will pick up the correct - epoll_fd */ - } else { - grpc_fd* fd = reinterpret_cast<grpc_fd*>( - reinterpret_cast<intptr_t>(data_ptr) & ~static_cast<intptr_t>(1)); - bool track_err = - reinterpret_cast<intptr_t>(data_ptr) & ~static_cast<intptr_t>(1); - bool cancel = (ep_ev[i].events & EPOLLHUP) != 0; - bool error = (ep_ev[i].events & EPOLLERR) != 0; - bool read_ev = (ep_ev[i].events & (EPOLLIN | EPOLLPRI)) != 0; - bool write_ev = (ep_ev[i].events & EPOLLOUT) != 0; - bool err_fallback = error && !track_err; - - if (error && !err_fallback) { - fd_has_errors(fd); - } - if (read_ev || cancel || err_fallback) { - fd_become_readable(fd); - } - if (write_ev || cancel || err_fallback) { - fd_become_writable(fd); - } - } - } - - g_current_thread_polling_island = nullptr; - gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1); - - GPR_ASSERT(pi != nullptr); - - /* Before leaving, release the extra ref we added to the polling island. It - is important to use "pi" here (i.e our old copy of pollset->po.pi - that we got before releasing the polling island lock). This is because - pollset->po.pi pointer might get udpated in other parts of the - code when there is an island merge while we are doing epoll_wait() above */ - PI_UNREF(pi, "ps_work"); -} - -/* pollset->po.mu lock must be held by the caller before calling this. - The function pollset_work() may temporarily release the lock (pollset->po.mu) - during the course of its execution but it will always re-acquire the lock and - ensure that it is held by the time the function returns */ -static grpc_error* pollset_work(grpc_pollset* pollset, - grpc_pollset_worker** worker_hdl, - grpc_millis deadline) { - GPR_TIMER_SCOPE("pollset_work", 0); - grpc_error* error = GRPC_ERROR_NONE; - int timeout_ms = poll_deadline_to_millis_timeout(deadline); - - sigset_t new_mask; - - grpc_pollset_worker worker; - worker.next = worker.prev = nullptr; - worker.pt_id = pthread_self(); - gpr_atm_no_barrier_store(&worker.is_kicked, (gpr_atm)0); - - if (worker_hdl) *worker_hdl = &worker; - - gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset); - gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); - - if (pollset->kicked_without_pollers) { - /* If the pollset was kicked without pollers, pretend that the current - worker got the kick and skip polling. A kick indicates that there is some - work that needs attention like an event on the completion queue or an - alarm */ - GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0); - pollset->kicked_without_pollers = 0; - } else if (!pollset->shutting_down) { - /* We use the posix-signal with number 'grpc_wakeup_signal' for waking up - (i.e 'kicking') a worker in the pollset. A 'kick' is a way to inform the - worker that there is some pending work that needs immediate attention - (like an event on the completion queue, or a polling island merge that - results in a new epoll-fd to wait on) and that the worker should not - spend time waiting in epoll_pwait(). - - A worker can be kicked anytime from the point it is added to the pollset - via push_front_worker() (or push_back_worker()) to the point it is - removed via remove_worker(). - If the worker is kicked before/during it calls epoll_pwait(), it should - immediately exit from epoll_wait(). If the worker is kicked after it - returns from epoll_wait(), then nothing really needs to be done. - - To accomplish this, we mask 'grpc_wakeup_signal' on this thread at all - times *except* when it is in epoll_pwait(). This way, the worker never - misses acting on a kick */ - - if (!g_initialized_sigmask) { - sigemptyset(&new_mask); - sigaddset(&new_mask, grpc_wakeup_signal); - pthread_sigmask(SIG_BLOCK, &new_mask, &g_orig_sigmask); - sigdelset(&g_orig_sigmask, grpc_wakeup_signal); - g_initialized_sigmask = true; - /* new_mask: The new thread mask which blocks 'grpc_wakeup_signal'. - This is the mask used at all times *except during - epoll_wait()*" - g_orig_sigmask: The thread mask which allows 'grpc_wakeup_signal' and - this is the mask to use *during epoll_wait()* - - The new_mask is set on the worker before it is added to the pollset - (i.e before it can be kicked) */ - } - - push_front_worker(pollset, &worker); /* Add worker to pollset */ - - pollset_work_and_unlock(pollset, &worker, timeout_ms, &g_orig_sigmask, - &error); - grpc_core::ExecCtx::Get()->Flush(); - - gpr_mu_lock(&pollset->po.mu); - - /* Note: There is no need to reset worker.is_kicked to 0 since we are no - longer going to use this worker */ - remove_worker(pollset, &worker); - } - - /* If we are the last worker on the pollset (i.e pollset_has_workers() is - false at this point) and the pollset is shutting down, we may have to - finish the shutdown process by calling finish_shutdown_locked(). - See pollset_shutdown() for more details. - - Note: Continuing to access pollset here is safe; it is the caller's - responsibility to not destroy a pollset when it has outstanding calls to - pollset_work() */ - if (pollset->shutting_down && !pollset_has_workers(pollset) && - !pollset->finish_shutdown_called) { - GPR_TIMER_MARK("pollset_work.finish_shutdown_locked", 0); - finish_shutdown_locked(pollset); - - gpr_mu_unlock(&pollset->po.mu); - grpc_core::ExecCtx::Get()->Flush(); - gpr_mu_lock(&pollset->po.mu); - } - - if (worker_hdl) *worker_hdl = nullptr; - - gpr_tls_set(&g_current_thread_pollset, (intptr_t)0); - gpr_tls_set(&g_current_thread_worker, (intptr_t)0); - - GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error)); - return error; -} - -static void add_poll_object(poll_obj* bag, poll_obj_type bag_type, - poll_obj* item, poll_obj_type item_type) { - GPR_TIMER_SCOPE("add_poll_object", 0); - -#ifndef NDEBUG - GPR_ASSERT(item->obj_type == item_type); - GPR_ASSERT(bag->obj_type == bag_type); -#endif - - grpc_error* error = GRPC_ERROR_NONE; - polling_island* pi_new = nullptr; - - gpr_mu_lock(&bag->mu); - gpr_mu_lock(&item->mu); - -retry: - /* - * 1) If item->pi and bag->pi are both non-NULL and equal, do nothing - * 2) If item->pi and bag->pi are both NULL, create a new polling island (with - * a refcount of 2) and point item->pi and bag->pi to the new island - * 3) If exactly one of item->pi or bag->pi is NULL, update it to point to - * the other's non-NULL pi - * 4) Finally if item->pi and bag-pi are non-NULL and not-equal, merge the - * polling islands and update item->pi and bag->pi to point to the new - * island - */ - - /* Early out if we are trying to add an 'fd' to a 'bag' but the fd is already - * orphaned */ - if (item_type == POLL_OBJ_FD && (FD_FROM_PO(item))->orphaned) { - gpr_mu_unlock(&item->mu); - gpr_mu_unlock(&bag->mu); - return; - } - - if (item->pi == bag->pi) { - pi_new = item->pi; - if (pi_new == nullptr) { - /* GPR_ASSERT(item->pi == bag->pi == NULL) */ - - /* If we are adding an fd to a bag (i.e pollset or pollset_set), then - * we need to do some extra work to make TSAN happy */ - if (item_type == POLL_OBJ_FD) { - /* Unlock before creating a new polling island: the polling island will - create a workqueue which creates a file descriptor, and holding an fd - lock here can eventually cause a loop to appear to TSAN (making it - unhappy). We don't think it's a real loop (there's an epoch point - where that loop possibility disappears), but the advantages of - keeping TSAN happy outweigh any performance advantage we might have - by keeping the lock held. */ - gpr_mu_unlock(&item->mu); - pi_new = polling_island_create(FD_FROM_PO(item), &error); - gpr_mu_lock(&item->mu); - - /* Need to reverify any assumptions made between the initial lock and - getting to this branch: if they've changed, we need to throw away our - work and figure things out again. */ - if (item->pi != nullptr) { - GRPC_POLLING_TRACE( - "add_poll_object: Raced creating new polling island. pi_new: %p " - "(fd: %d, %s: %p)", - (void*)pi_new, FD_FROM_PO(item)->fd, poll_obj_string(bag_type), - (void*)bag); - /* No need to lock 'pi_new' here since this is a new polling island - and no one has a reference to it yet */ - polling_island_remove_all_fds_locked(pi_new, true, &error); - - /* Ref and unref so that the polling island gets deleted during unref - */ - PI_ADD_REF(pi_new, "dance_of_destruction"); - PI_UNREF(pi_new, "dance_of_destruction"); - goto retry; - } - } else { - pi_new = polling_island_create(nullptr, &error); - } - - GRPC_POLLING_TRACE( - "add_poll_object: Created new polling island. pi_new: %p (%s: %p, " - "%s: %p)", - (void*)pi_new, poll_obj_string(item_type), (void*)item, - poll_obj_string(bag_type), (void*)bag); - } else { - GRPC_POLLING_TRACE( - "add_poll_object: Same polling island. pi: %p (%s, %s)", - (void*)pi_new, poll_obj_string(item_type), poll_obj_string(bag_type)); - } - } else if (item->pi == nullptr) { - /* GPR_ASSERT(bag->pi != NULL) */ - /* Make pi_new point to latest pi*/ - pi_new = polling_island_lock(bag->pi); - - if (item_type == POLL_OBJ_FD) { - grpc_fd* fd = FD_FROM_PO(item); - polling_island_add_fds_locked(pi_new, &fd, 1, true, &error); - } - - gpr_mu_unlock(&pi_new->mu); - GRPC_POLLING_TRACE( - "add_poll_obj: item->pi was NULL. pi_new: %p (item(%s): %p, " - "bag(%s): %p)", - (void*)pi_new, poll_obj_string(item_type), (void*)item, - poll_obj_string(bag_type), (void*)bag); - } else if (bag->pi == nullptr) { - /* GPR_ASSERT(item->pi != NULL) */ - /* Make pi_new to point to latest pi */ - pi_new = polling_island_lock(item->pi); - gpr_mu_unlock(&pi_new->mu); - GRPC_POLLING_TRACE( - "add_poll_obj: bag->pi was NULL. pi_new: %p (item(%s): %p, " - "bag(%s): %p)", - (void*)pi_new, poll_obj_string(item_type), (void*)item, - poll_obj_string(bag_type), (void*)bag); - } else { - pi_new = polling_island_merge(item->pi, bag->pi, &error); - GRPC_POLLING_TRACE( - "add_poll_obj: polling islands merged. pi_new: %p (item(%s): %p, " - "bag(%s): %p)", - (void*)pi_new, poll_obj_string(item_type), (void*)item, - poll_obj_string(bag_type), (void*)bag); - } - - /* At this point, pi_new is the polling island that both item->pi and bag->pi - MUST be pointing to */ - - if (item->pi != pi_new) { - PI_ADD_REF(pi_new, poll_obj_string(item_type)); - if (item->pi != nullptr) { - PI_UNREF(item->pi, poll_obj_string(item_type)); - } - item->pi = pi_new; - } - - if (bag->pi != pi_new) { - PI_ADD_REF(pi_new, poll_obj_string(bag_type)); - if (bag->pi != nullptr) { - PI_UNREF(bag->pi, poll_obj_string(bag_type)); - } - bag->pi = pi_new; - } - - gpr_mu_unlock(&item->mu); - gpr_mu_unlock(&bag->mu); - - GRPC_LOG_IF_ERROR("add_poll_object", error); -} - -static void pollset_add_fd(grpc_pollset* pollset, grpc_fd* fd) { - add_poll_object(&pollset->po, POLL_OBJ_POLLSET, &fd->po, POLL_OBJ_FD); -} - -/******************************************************************************* - * Pollset-set Definitions - */ - -static grpc_pollset_set* pollset_set_create(void) { - grpc_pollset_set* pss = - static_cast<grpc_pollset_set*>(gpr_malloc(sizeof(*pss))); - gpr_mu_init(&pss->po.mu); - pss->po.pi = nullptr; -#ifndef NDEBUG - pss->po.obj_type = POLL_OBJ_POLLSET_SET; -#endif - return pss; -} - -static void pollset_set_destroy(grpc_pollset_set* pss) { - gpr_mu_destroy(&pss->po.mu); - - if (pss->po.pi != nullptr) { - PI_UNREF(pss->po.pi, "pss_destroy"); - } - - gpr_free(pss); -} - -static void pollset_set_add_fd(grpc_pollset_set* pss, grpc_fd* fd) { - add_poll_object(&pss->po, POLL_OBJ_POLLSET_SET, &fd->po, POLL_OBJ_FD); -} - -static void pollset_set_del_fd(grpc_pollset_set* pss, grpc_fd* fd) { - /* Nothing to do */ -} - -static void pollset_set_add_pollset(grpc_pollset_set* pss, grpc_pollset* ps) { - add_poll_object(&pss->po, POLL_OBJ_POLLSET_SET, &ps->po, POLL_OBJ_POLLSET); -} - -static void pollset_set_del_pollset(grpc_pollset_set* pss, grpc_pollset* ps) { - /* Nothing to do */ -} - -static void pollset_set_add_pollset_set(grpc_pollset_set* bag, - grpc_pollset_set* item) { - add_poll_object(&bag->po, POLL_OBJ_POLLSET_SET, &item->po, - POLL_OBJ_POLLSET_SET); -} - -static void pollset_set_del_pollset_set(grpc_pollset_set* bag, - grpc_pollset_set* item) { - /* Nothing to do */ -} - -/* Test helper functions - * */ -void* grpc_fd_get_polling_island(grpc_fd* fd) { - polling_island* pi; - - gpr_mu_lock(&fd->po.mu); - pi = fd->po.pi; - gpr_mu_unlock(&fd->po.mu); - - return pi; -} - -void* grpc_pollset_get_polling_island(grpc_pollset* ps) { - polling_island* pi; - - gpr_mu_lock(&ps->po.mu); - pi = ps->po.pi; - gpr_mu_unlock(&ps->po.mu); - - return pi; -} - -bool grpc_are_polling_islands_equal(void* p, void* q) { - polling_island* p1 = static_cast<polling_island*>(p); - polling_island* p2 = static_cast<polling_island*>(q); - - /* Note: polling_island_lock_pair() may change p1 and p2 to point to the - latest polling islands in their respective linked lists */ - polling_island_lock_pair(&p1, &p2); - polling_island_unlock_pair(p1, p2); - - return p1 == p2; -} - -/******************************************************************************* - * Event engine binding - */ - -static void shutdown_engine(void) { - fd_global_shutdown(); - pollset_global_shutdown(); - polling_island_global_shutdown(); -} - -static const grpc_event_engine_vtable vtable = { - sizeof(grpc_pollset), - true, - - fd_create, - fd_wrapped_fd, - fd_orphan, - fd_shutdown, - fd_notify_on_read, - fd_notify_on_write, - fd_notify_on_error, - fd_become_readable, - fd_become_writable, - fd_has_errors, - fd_is_shutdown, - - pollset_init, - pollset_shutdown, - pollset_destroy, - pollset_work, - pollset_kick, - pollset_add_fd, - - pollset_set_create, - pollset_set_destroy, - pollset_set_add_pollset, - pollset_set_del_pollset, - pollset_set_add_pollset_set, - pollset_set_del_pollset_set, - pollset_set_add_fd, - pollset_set_del_fd, - - shutdown_engine, -}; - -/* It is possible that GLIBC has epoll but the underlying kernel doesn't. - * Create a dummy epoll_fd to make sure epoll support is available */ -static bool is_epoll_available() { - int fd = epoll_create1(EPOLL_CLOEXEC); - if (fd < 0) { - gpr_log( - GPR_ERROR, - "epoll_create1 failed with error: %d. Not using epoll polling engine", - fd); - return false; - } - close(fd); - return true; -} - -const grpc_event_engine_vtable* grpc_init_epollsig_linux( - bool explicit_request) { - /* If use of signals is disabled, we cannot use epoll engine*/ - if (is_grpc_wakeup_signal_initialized && grpc_wakeup_signal < 0) { - gpr_log(GPR_ERROR, "Skipping epollsig because use of signals is disabled."); - return nullptr; - } - - if (!grpc_has_wakeup_fd()) { - gpr_log(GPR_ERROR, "Skipping epollsig because of no wakeup fd."); - return nullptr; - } - - if (!is_epoll_available()) { - gpr_log(GPR_ERROR, "Skipping epollsig because epoll is unavailable."); - return nullptr; - } - - if (!is_grpc_wakeup_signal_initialized) { - if (explicit_request) { - grpc_use_signal(SIGRTMIN + 6); - } else { - gpr_log(GPR_ERROR, - "Skipping epollsig because uninitialized wakeup signal."); - return nullptr; - } - } - - fd_global_init(); - - if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) { - return nullptr; - } - - if (!GRPC_LOG_IF_ERROR("polling_island_global_init", - polling_island_global_init())) { - return nullptr; - } - - return &vtable; -} - -#else /* defined(GRPC_LINUX_EPOLL_CREATE1) */ -#if defined(GRPC_POSIX_SOCKET_EV_EPOLLSIG) -#include "src/core/lib/iomgr/ev_epollsig_linux.h" -/* If GRPC_LINUX_EPOLL_CREATE1 is not defined, it means - epoll_create1 is not available. Return NULL */ -const grpc_event_engine_vtable* grpc_init_epollsig_linux( - bool explicit_request) { - return nullptr; -} -#endif /* defined(GRPC_POSIX_SOCKET) */ - -void grpc_use_signal(int signum) {} -#endif /* !defined(GRPC_LINUX_EPOLL_CREATE1) */ diff --git a/src/core/lib/iomgr/ev_epollsig_linux.h b/src/core/lib/iomgr/ev_epollsig_linux.h deleted file mode 100644 index 2ba2f0a63b..0000000000 --- a/src/core/lib/iomgr/ev_epollsig_linux.h +++ /dev/null @@ -1,35 +0,0 @@ -/* - * - * Copyright 2015 gRPC authors. - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_EV_EPOLLSIG_LINUX_H -#define GRPC_CORE_LIB_IOMGR_EV_EPOLLSIG_LINUX_H - -#include <grpc/support/port_platform.h> - -#include "src/core/lib/iomgr/ev_posix.h" -#include "src/core/lib/iomgr/port.h" - -const grpc_event_engine_vtable* grpc_init_epollsig_linux(bool explicit_request); - -#ifdef GRPC_LINUX_EPOLL_CREATE1 -void* grpc_fd_get_polling_island(grpc_fd* fd); -void* grpc_pollset_get_polling_island(grpc_pollset* ps); -bool grpc_are_polling_islands_equal(void* p, void* q); -#endif /* defined(GRPC_LINUX_EPOLL_CREATE1) */ - -#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLLSIG_LINUX_H */ diff --git a/src/core/lib/iomgr/ev_posix.cc b/src/core/lib/iomgr/ev_posix.cc index d4377e2d50..8a7dc7b004 100644 --- a/src/core/lib/iomgr/ev_posix.cc +++ b/src/core/lib/iomgr/ev_posix.cc @@ -35,7 +35,6 @@ #include "src/core/lib/gpr/useful.h" #include "src/core/lib/iomgr/ev_epoll1_linux.h" #include "src/core/lib/iomgr/ev_epollex_linux.h" -#include "src/core/lib/iomgr/ev_epollsig_linux.h" #include "src/core/lib/iomgr/ev_poll_posix.h" grpc_core::TraceFlag grpc_polling_trace(false, @@ -123,13 +122,13 @@ const grpc_event_engine_vtable* init_non_polling(bool explicit_request) { // environment variable if that variable is set (which should be a // comma-separated list of one or more event engine names) static event_engine_factory g_factories[] = { - {ENGINE_HEAD_CUSTOM, nullptr}, {ENGINE_HEAD_CUSTOM, nullptr}, - {ENGINE_HEAD_CUSTOM, nullptr}, {ENGINE_HEAD_CUSTOM, nullptr}, - {"epollex", grpc_init_epollex_linux}, {"epoll1", grpc_init_epoll1_linux}, - {"epollsig", grpc_init_epollsig_linux}, {"poll", grpc_init_poll_posix}, - {"poll-cv", grpc_init_poll_cv_posix}, {"none", init_non_polling}, - {ENGINE_TAIL_CUSTOM, nullptr}, {ENGINE_TAIL_CUSTOM, nullptr}, - {ENGINE_TAIL_CUSTOM, nullptr}, {ENGINE_TAIL_CUSTOM, nullptr}, + {ENGINE_HEAD_CUSTOM, nullptr}, {ENGINE_HEAD_CUSTOM, nullptr}, + {ENGINE_HEAD_CUSTOM, nullptr}, {ENGINE_HEAD_CUSTOM, nullptr}, + {"epollex", grpc_init_epollex_linux}, {"epoll1", grpc_init_epoll1_linux}, + {"poll", grpc_init_poll_posix}, {"poll-cv", grpc_init_poll_cv_posix}, + {"none", init_non_polling}, {ENGINE_TAIL_CUSTOM, nullptr}, + {ENGINE_TAIL_CUSTOM, nullptr}, {ENGINE_TAIL_CUSTOM, nullptr}, + {ENGINE_TAIL_CUSTOM, nullptr}, }; static void add(const char* beg, const char* end, char*** ss, size_t* ns) { diff --git a/src/core/lib/iomgr/port.h b/src/core/lib/iomgr/port.h index abf96662f5..c8046b21dc 100644 --- a/src/core/lib/iomgr/port.h +++ b/src/core/lib/iomgr/port.h @@ -82,6 +82,11 @@ #define GRPC_LINUX_SOCKETUTILS 1 #endif #endif +#ifdef LINUX_VERSION_CODE +#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 37) +#define GRPC_HAVE_TCP_USER_TIMEOUT +#endif /* LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 37) */ +#endif /* LINUX_VERSION_CODE */ #ifndef __GLIBC__ #define GRPC_LINUX_EPOLL 1 #define GRPC_LINUX_EPOLL_CREATE1 1 @@ -110,7 +115,6 @@ #define GRPC_POSIX_SOCKET_EV 1 #define GRPC_POSIX_SOCKET_EV_EPOLL1 1 #define GRPC_POSIX_SOCKET_EV_EPOLLEX 1 -#define GRPC_POSIX_SOCKET_EV_EPOLLSIG 1 #define GRPC_POSIX_SOCKET_EV_POLL 1 #define GRPC_POSIX_SOCKET_RESOLVE_ADDRESS 1 #define GRPC_POSIX_SOCKET_SOCKADDR 1 @@ -178,7 +182,6 @@ #define GRPC_POSIX_SOCKET_ARES_EV_DRIVER 1 #define GRPC_POSIX_SOCKET_EV 1 #define GRPC_POSIX_SOCKET_EV_EPOLLEX 1 -#define GRPC_POSIX_SOCKET_EV_EPOLLSIG 1 #define GRPC_POSIX_SOCKET_EV_POLL 1 #define GRPC_POSIX_SOCKET_EV_EPOLL1 1 #define GRPC_POSIX_SOCKET_IOMGR 1 diff --git a/src/core/lib/iomgr/socket_utils_common_posix.cc b/src/core/lib/iomgr/socket_utils_common_posix.cc index c4b991c94d..50674b0845 100644 --- a/src/core/lib/iomgr/socket_utils_common_posix.cc +++ b/src/core/lib/iomgr/socket_utils_common_posix.cc @@ -41,6 +41,7 @@ #include <grpc/support/log.h> #include <grpc/support/sync.h> +#include "src/core/lib/channel/channel_args.h" #include "src/core/lib/gpr/host_port.h" #include "src/core/lib/gpr/string.h" #include "src/core/lib/iomgr/sockaddr.h" @@ -222,6 +223,95 @@ grpc_error* grpc_set_socket_low_latency(int fd, int low_latency) { return GRPC_ERROR_NONE; } +/* The default values for TCP_USER_TIMEOUT are currently configured to be in + * line with the default values of KEEPALIVE_TIMEOUT as proposed in + * https://github.com/grpc/proposal/blob/master/A18-tcp-user-timeout.md */ +#define DEFAULT_CLIENT_TCP_USER_TIMEOUT_MS 20000 /* 20 seconds */ +#define DEFAULT_SERVER_TCP_USER_TIMEOUT_MS 20000 /* 20 seconds */ + +static int g_default_client_tcp_user_timeout_ms = + DEFAULT_CLIENT_TCP_USER_TIMEOUT_MS; +static int g_default_server_tcp_user_timeout_ms = + DEFAULT_SERVER_TCP_USER_TIMEOUT_MS; +static bool g_default_client_tcp_user_timeout_enabled = false; +static bool g_default_server_tcp_user_timeout_enabled = true; + +void config_default_tcp_user_timeout(bool enable, int timeout, bool is_client) { + if (is_client) { + g_default_client_tcp_user_timeout_enabled = enable; + if (timeout > 0) { + g_default_client_tcp_user_timeout_ms = timeout; + } + } else { + g_default_server_tcp_user_timeout_enabled = enable; + if (timeout > 0) { + g_default_server_tcp_user_timeout_ms = timeout; + } + } +} + +/* Set TCP_USER_TIMEOUT */ +grpc_error* grpc_set_socket_tcp_user_timeout( + int fd, const grpc_channel_args* channel_args, bool is_client) { +#ifdef GRPC_HAVE_TCP_USER_TIMEOUT + bool enable; + int timeout; + if (is_client) { + enable = g_default_client_tcp_user_timeout_enabled; + timeout = g_default_client_tcp_user_timeout_ms; + } else { + enable = g_default_server_tcp_user_timeout_enabled; + timeout = g_default_server_tcp_user_timeout_ms; + } + if (channel_args) { + for (unsigned int i = 0; i < channel_args->num_args; i++) { + if (0 == strcmp(channel_args->args[i].key, GRPC_ARG_KEEPALIVE_TIME_MS)) { + const int value = grpc_channel_arg_get_integer( + &channel_args->args[i], grpc_integer_options{0, 1, INT_MAX}); + /* Continue using default if value is 0 */ + if (value == 0) { + continue; + } + /* Disable if value is INT_MAX */ + enable = value != INT_MAX; + } else if (0 == strcmp(channel_args->args[i].key, + GRPC_ARG_KEEPALIVE_TIMEOUT_MS)) { + const int value = grpc_channel_arg_get_integer( + &channel_args->args[i], grpc_integer_options{0, 1, INT_MAX}); + /* Continue using default if value is 0 */ + if (value == 0) { + continue; + } + timeout = value; + } + } + } + if (enable) { + extern grpc_core::TraceFlag grpc_tcp_trace; + if (grpc_tcp_trace.enabled()) { + gpr_log(GPR_INFO, "Enabling TCP_USER_TIMEOUT with a timeout of %d ms", + timeout); + } + int newval; + socklen_t len = sizeof(newval); + if (0 != setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &timeout, + sizeof(timeout))) { + return GRPC_OS_ERROR(errno, "setsockopt(TCP_USER_TIMEOUT)"); + } + if (0 != getsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &newval, &len)) { + return GRPC_OS_ERROR(errno, "getsockopt(TCP_USER_TIMEOUT)"); + } + if (newval != timeout) { + return GRPC_ERROR_CREATE_FROM_STATIC_STRING( + "Failed to set TCP_USER_TIMEOUT"); + } + } +#else + gpr_log(GPR_INFO, "TCP_USER_TIMEOUT not supported for this platform"); +#endif /* GRPC_HAVE_TCP_USER_TIMEOUT */ + return GRPC_ERROR_NONE; +} + /* set a socket using a grpc_socket_mutator */ grpc_error* grpc_set_socket_with_mutator(int fd, grpc_socket_mutator* mutator) { GPR_ASSERT(mutator); diff --git a/src/core/lib/iomgr/socket_utils_posix.h b/src/core/lib/iomgr/socket_utils_posix.h index b3fd58a530..71a304dc4e 100644 --- a/src/core/lib/iomgr/socket_utils_posix.h +++ b/src/core/lib/iomgr/socket_utils_posix.h @@ -53,6 +53,13 @@ grpc_error* grpc_set_socket_low_latency(int fd, int low_latency); /* set SO_REUSEPORT */ grpc_error* grpc_set_socket_reuse_port(int fd, int reuse); +/* Configure the default values for TCP_USER_TIMEOUT */ +void config_default_tcp_user_timeout(bool enable, int timeout, bool is_client); + +/* Set TCP_USER_TIMEOUT */ +grpc_error* grpc_set_socket_tcp_user_timeout( + int fd, const grpc_channel_args* channel_args, bool is_client); + /* Returns true if this system can create AF_INET6 sockets bound to ::1. The value is probed once, and cached for the life of the process. diff --git a/src/core/lib/iomgr/tcp_client_posix.cc b/src/core/lib/iomgr/tcp_client_posix.cc index 9c989b7dfe..8553ed0db4 100644 --- a/src/core/lib/iomgr/tcp_client_posix.cc +++ b/src/core/lib/iomgr/tcp_client_posix.cc @@ -76,6 +76,9 @@ static grpc_error* prepare_socket(const grpc_resolved_address* addr, int fd, if (!grpc_is_unix_socket(addr)) { err = grpc_set_socket_low_latency(fd, 1); if (err != GRPC_ERROR_NONE) goto error; + err = grpc_set_socket_tcp_user_timeout(fd, channel_args, + true /* is_client */); + if (err != GRPC_ERROR_NONE) goto error; } err = grpc_set_socket_no_sigpipe_if_possible(fd); if (err != GRPC_ERROR_NONE) goto error; diff --git a/src/core/lib/iomgr/tcp_server_utils_posix_common.cc b/src/core/lib/iomgr/tcp_server_utils_posix_common.cc index 9595c028ce..8d8d3f4273 100644 --- a/src/core/lib/iomgr/tcp_server_utils_posix_common.cc +++ b/src/core/lib/iomgr/tcp_server_utils_posix_common.cc @@ -166,6 +166,9 @@ grpc_error* grpc_tcp_server_prepare_socket(grpc_tcp_server* s, int fd, if (err != GRPC_ERROR_NONE) goto error; err = grpc_set_socket_reuse_addr(fd, 1); if (err != GRPC_ERROR_NONE) goto error; + err = grpc_set_socket_tcp_user_timeout(fd, s->channel_args, + false /* is_client */); + if (err != GRPC_ERROR_NONE) goto error; } err = grpc_set_socket_no_sigpipe_if_possible(fd); if (err != GRPC_ERROR_NONE) goto error; diff --git a/src/core/lib/security/credentials/alts/grpc_alts_credentials_client_options.cc b/src/core/lib/security/credentials/alts/grpc_alts_credentials_client_options.cc index 0a39c6c485..118d18d119 100644 --- a/src/core/lib/security/credentials/alts/grpc_alts_credentials_client_options.cc +++ b/src/core/lib/security/credentials/alts/grpc_alts_credentials_client_options.cc @@ -72,7 +72,8 @@ static void target_service_account_destroy( static const grpc_alts_credentials_options_vtable vtable = { alts_client_options_copy, alts_client_options_destroy}; -grpc_alts_credentials_options* grpc_alts_credentials_client_options_create() { +grpc_alts_credentials_options* grpc_alts_credentials_client_options_create( + void) { auto client_options = static_cast<grpc_alts_credentials_client_options*>( gpr_zalloc(sizeof(grpc_alts_credentials_client_options))); client_options->base.vtable = &vtable; diff --git a/src/core/lib/security/credentials/alts/grpc_alts_credentials_server_options.cc b/src/core/lib/security/credentials/alts/grpc_alts_credentials_server_options.cc index 62aa7a620a..1a59c45675 100644 --- a/src/core/lib/security/credentials/alts/grpc_alts_credentials_server_options.cc +++ b/src/core/lib/security/credentials/alts/grpc_alts_credentials_server_options.cc @@ -36,7 +36,8 @@ static void alts_server_options_destroy( static const grpc_alts_credentials_options_vtable vtable = { alts_server_options_copy, alts_server_options_destroy}; -grpc_alts_credentials_options* grpc_alts_credentials_server_options_create() { +grpc_alts_credentials_options* grpc_alts_credentials_server_options_create( + void) { grpc_alts_credentials_server_options* server_options = static_cast<grpc_alts_credentials_server_options*>( gpr_zalloc(sizeof(*server_options))); diff --git a/src/core/lib/security/transport/server_auth_filter.cc b/src/core/lib/security/transport/server_auth_filter.cc index 552e70130a..b99fc5e178 100644 --- a/src/core/lib/security/transport/server_auth_filter.cc +++ b/src/core/lib/security/transport/server_auth_filter.cc @@ -41,9 +41,11 @@ struct call_data { grpc_transport_stream_op_batch* recv_initial_metadata_batch; grpc_closure* original_recv_initial_metadata_ready; grpc_closure recv_initial_metadata_ready; - grpc_error* error; + grpc_error* recv_initial_metadata_error; grpc_closure recv_trailing_metadata_ready; grpc_closure* original_recv_trailing_metadata_ready; + grpc_error* recv_trailing_metadata_error; + bool seen_recv_trailing_metadata_ready; grpc_metadata_array md; const grpc_metadata* consumed_md; size_t num_consumed_md; @@ -114,8 +116,16 @@ static void on_md_processing_done_inner(grpc_call_element* elem, batch->payload->recv_initial_metadata.recv_initial_metadata, remove_consumed_md, elem, "Response metadata filtering error"); } - calld->error = GRPC_ERROR_REF(error); - GRPC_CLOSURE_SCHED(calld->original_recv_initial_metadata_ready, error); + calld->recv_initial_metadata_error = GRPC_ERROR_REF(error); + grpc_closure* closure = calld->original_recv_initial_metadata_ready; + calld->original_recv_initial_metadata_ready = nullptr; + if (calld->seen_recv_trailing_metadata_ready) { + GRPC_CALL_COMBINER_START(calld->call_combiner, + &calld->recv_trailing_metadata_ready, + calld->recv_trailing_metadata_error, + "continue recv_trailing_metadata_ready"); + } + GRPC_CLOSURE_SCHED(closure, error); } // Called from application code. @@ -184,14 +194,30 @@ static void recv_initial_metadata_ready(void* arg, grpc_error* error) { return; } } - GRPC_CLOSURE_RUN(calld->original_recv_initial_metadata_ready, - GRPC_ERROR_REF(error)); + grpc_closure* closure = calld->original_recv_initial_metadata_ready; + calld->original_recv_initial_metadata_ready = nullptr; + if (calld->seen_recv_trailing_metadata_ready) { + GRPC_CALL_COMBINER_START(calld->call_combiner, + &calld->recv_trailing_metadata_ready, + calld->recv_trailing_metadata_error, + "continue recv_trailing_metadata_ready"); + } + GRPC_CLOSURE_RUN(closure, GRPC_ERROR_REF(error)); } static void recv_trailing_metadata_ready(void* user_data, grpc_error* err) { grpc_call_element* elem = static_cast<grpc_call_element*>(user_data); call_data* calld = static_cast<call_data*>(elem->call_data); - err = grpc_error_add_child(GRPC_ERROR_REF(err), GRPC_ERROR_REF(calld->error)); + if (calld->original_recv_initial_metadata_ready != nullptr) { + calld->recv_trailing_metadata_error = GRPC_ERROR_REF(err); + calld->seen_recv_trailing_metadata_ready = true; + GRPC_CALL_COMBINER_STOP(calld->call_combiner, + "deferring recv_trailing_metadata_ready until " + "after recv_initial_metadata_ready"); + return; + } + err = grpc_error_add_child( + GRPC_ERROR_REF(err), GRPC_ERROR_REF(calld->recv_initial_metadata_error)); GRPC_CLOSURE_RUN(calld->original_recv_trailing_metadata_ready, err); } @@ -249,7 +275,7 @@ static void destroy_call_elem(grpc_call_element* elem, const grpc_call_final_info* final_info, grpc_closure* ignored) { call_data* calld = static_cast<call_data*>(elem->call_data); - GRPC_ERROR_UNREF(calld->error); + GRPC_ERROR_UNREF(calld->recv_initial_metadata_error); } /* Constructor for channel_data */ diff --git a/src/core/lib/surface/call.cc b/src/core/lib/surface/call.cc index 11b438f5dc..a9349afa68 100644 --- a/src/core/lib/surface/call.cc +++ b/src/core/lib/surface/call.cc @@ -207,7 +207,7 @@ struct grpc_call { grpc_server* server; } server; } final_op; - grpc_error* status_error; + gpr_atm status_error; /* recv_state can contain one of the following values: RECV_NONE : : no initial metadata and messages received @@ -519,10 +519,12 @@ static void destroy_call(void* call, grpc_error* error) { GRPC_CQ_INTERNAL_UNREF(c->cq, "bind"); } - grpc_error_get_status(c->status_error, c->send_deadline, + grpc_error* status_error = + reinterpret_cast<grpc_error*>(gpr_atm_acq_load(&c->status_error)); + grpc_error_get_status(status_error, c->send_deadline, &c->final_info.final_status, nullptr, nullptr, &(c->final_info.error_string)); - GRPC_ERROR_UNREF(c->status_error); + GRPC_ERROR_UNREF(status_error); c->final_info.stats.latency = gpr_time_sub(gpr_now(GPR_CLOCK_MONOTONIC), c->start_time); @@ -705,7 +707,7 @@ static void set_final_status(grpc_call* call, grpc_error* error) { call->final_op.client.error_string); // explicitly take a ref grpc_slice_ref_internal(*call->final_op.client.status_details); - call->status_error = error; + gpr_atm_rel_store(&call->status_error, reinterpret_cast<gpr_atm>(error)); grpc_core::channelz::ChannelNode* channelz_channel = grpc_channel_get_channelz_node(call->channel); if (channelz_channel != nullptr) { @@ -717,7 +719,9 @@ static void set_final_status(grpc_call* call, grpc_error* error) { } } else { *call->final_op.server.cancelled = - error != GRPC_ERROR_NONE || call->status_error != GRPC_ERROR_NONE; + error != GRPC_ERROR_NONE || + reinterpret_cast<grpc_error*>(gpr_atm_acq_load(&call->status_error)) != + GRPC_ERROR_NONE; grpc_core::channelz::ServerNode* channelz_server = grpc_server_get_channelz_node(call->final_op.server.server); if (channelz_server != nullptr) { @@ -1686,7 +1690,8 @@ static grpc_call_error call_start_batch(grpc_call* call, const grpc_op* ops, } } - call->status_error = status_error; + gpr_atm_rel_store(&call->status_error, + reinterpret_cast<gpr_atm>(status_error)); if (!prepare_application_metadata( call, static_cast<int>( diff --git a/src/core/lib/surface/completion_queue.cc b/src/core/lib/surface/completion_queue.cc index 0769d9e4f6..5dc9991f70 100644 --- a/src/core/lib/surface/completion_queue.cc +++ b/src/core/lib/surface/completion_queue.cc @@ -184,7 +184,8 @@ static const cq_poller_vtable g_poller_vtable_by_poller_type[] = { typedef struct cq_vtable { grpc_cq_completion_type cq_completion_type; size_t data_size; - void (*init)(void* data, grpc_core::CQCallbackInterface* shutdown_callback); + void (*init)(void* data, + grpc_experimental_completion_queue_functor* shutdown_callback); void (*shutdown)(grpc_completion_queue* cq); void (*destroy)(void* data); bool (*begin_op)(grpc_completion_queue* cq, void* tag); @@ -267,7 +268,7 @@ typedef struct cq_callback_data { bool shutdown_called; /** A callback that gets invoked when the CQ completes shutdown */ - grpc_core::CQCallbackInterface* shutdown_callback; + grpc_experimental_completion_queue_functor* shutdown_callback; } cq_callback_data; /* Completion queue structure */ @@ -333,12 +334,12 @@ static grpc_event cq_pluck(grpc_completion_queue* cq, void* tag, gpr_timespec deadline, void* reserved); // Note that cq_init_next and cq_init_pluck do not use the shutdown_callback -static void cq_init_next(void* data, - grpc_core::CQCallbackInterface* shutdown_callback); -static void cq_init_pluck(void* data, - grpc_core::CQCallbackInterface* shutdown_callback); -static void cq_init_callback(void* data, - grpc_core::CQCallbackInterface* shutdown_callback); +static void cq_init_next( + void* data, grpc_experimental_completion_queue_functor* shutdown_callback); +static void cq_init_pluck( + void* data, grpc_experimental_completion_queue_functor* shutdown_callback); +static void cq_init_callback( + void* data, grpc_experimental_completion_queue_functor* shutdown_callback); static void cq_destroy_next(void* data); static void cq_destroy_pluck(void* data); static void cq_destroy_callback(void* data); @@ -462,7 +463,7 @@ static long cq_event_queue_num_items(grpc_cq_event_queue* q) { grpc_completion_queue* grpc_completion_queue_create_internal( grpc_cq_completion_type completion_type, grpc_cq_polling_type polling_type, - grpc_core::CQCallbackInterface* shutdown_callback) { + grpc_experimental_completion_queue_functor* shutdown_callback) { GPR_TIMER_SCOPE("grpc_completion_queue_create_internal", 0); grpc_completion_queue* cq; @@ -497,8 +498,8 @@ grpc_completion_queue* grpc_completion_queue_create_internal( return cq; } -static void cq_init_next(void* data, - grpc_core::CQCallbackInterface* shutdown_callback) { +static void cq_init_next( + void* data, grpc_experimental_completion_queue_functor* shutdown_callback) { cq_next_data* cqd = static_cast<cq_next_data*>(data); /* Initial count is dropped by grpc_completion_queue_shutdown */ gpr_atm_no_barrier_store(&cqd->pending_events, 1); @@ -513,8 +514,8 @@ static void cq_destroy_next(void* data) { cq_event_queue_destroy(&cqd->queue); } -static void cq_init_pluck(void* data, - grpc_core::CQCallbackInterface* shutdown_callback) { +static void cq_init_pluck( + void* data, grpc_experimental_completion_queue_functor* shutdown_callback) { cq_pluck_data* cqd = static_cast<cq_pluck_data*>(data); /* Initial count is dropped by grpc_completion_queue_shutdown */ gpr_atm_no_barrier_store(&cqd->pending_events, 1); @@ -532,7 +533,7 @@ static void cq_destroy_pluck(void* data) { } static void cq_init_callback( - void* data, grpc_core::CQCallbackInterface* shutdown_callback) { + void* data, grpc_experimental_completion_queue_functor* shutdown_callback) { cq_callback_data* cqd = static_cast<cq_callback_data*>(data); /* Initial count is dropped by grpc_completion_queue_shutdown */ gpr_atm_no_barrier_store(&cqd->pending_events, 1); @@ -859,7 +860,8 @@ static void cq_end_op_for_callback( GRPC_ERROR_UNREF(error); - (static_cast<grpc_core::CQCallbackInterface*>(tag))->Run(is_success); + auto* functor = static_cast<grpc_experimental_completion_queue_functor*>(tag); + (*functor->functor_run)(functor, is_success); } void grpc_cq_end_op(grpc_completion_queue* cq, void* tag, grpc_error* error, @@ -1343,7 +1345,7 @@ static void cq_finish_shutdown_callback(grpc_completion_queue* cq) { GPR_ASSERT(cqd->shutdown_called); cq->poller_vtable->shutdown(POLLSET_FROM_CQ(cq), &cq->pollset_shutdown_done); - callback->Run(true); + (*callback->functor_run)(callback, true); } static void cq_shutdown_callback(grpc_completion_queue* cq) { @@ -1364,9 +1366,11 @@ static void cq_shutdown_callback(grpc_completion_queue* cq) { } cqd->shutdown_called = true; if (gpr_atm_full_fetch_add(&cqd->pending_events, -1) == 1) { + gpr_mu_unlock(cq->mu); cq_finish_shutdown_callback(cq); + } else { + gpr_mu_unlock(cq->mu); } - gpr_mu_unlock(cq->mu); GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down (callback cq)"); } diff --git a/src/core/lib/surface/completion_queue.h b/src/core/lib/surface/completion_queue.h index a7c524d8e8..d60fe6d6ef 100644 --- a/src/core/lib/surface/completion_queue.h +++ b/src/core/lib/surface/completion_queue.h @@ -48,23 +48,6 @@ typedef struct grpc_cq_completion { uintptr_t next; } grpc_cq_completion; -/// For callback CQs, the tag that is passed in for an operation must -/// actually be a pointer to an implementation of the following class. -/// When the operation completes, the tag will be typecasted from void* -/// to grpc_core::CQCallbackInterface* and then the Run method will be -/// invoked on it. In practice, the language binding (e.g., C++ API -/// implementation) is responsible for providing and using an implementation -/// of this abstract base class. -namespace grpc_core { -class CQCallbackInterface { - public: - virtual ~CQCallbackInterface() {} - virtual void Run(bool) GRPC_ABSTRACT; - - GRPC_ABSTRACT_BASE_CLASS -}; -} // namespace grpc_core - #ifndef NDEBUG void grpc_cq_internal_ref(grpc_completion_queue* cc, const char* reason, const char* file, int line); @@ -106,6 +89,6 @@ int grpc_get_cq_poll_num(grpc_completion_queue* cc); grpc_completion_queue* grpc_completion_queue_create_internal( grpc_cq_completion_type completion_type, grpc_cq_polling_type polling_type, - grpc_core::CQCallbackInterface* shutdown_callback); + grpc_experimental_completion_queue_functor* shutdown_callback); #endif /* GRPC_CORE_LIB_SURFACE_COMPLETION_QUEUE_H */ diff --git a/src/core/lib/surface/completion_queue_factory.cc b/src/core/lib/surface/completion_queue_factory.cc index ed92dd7eba..2616c156e4 100644 --- a/src/core/lib/surface/completion_queue_factory.cc +++ b/src/core/lib/surface/completion_queue_factory.cc @@ -31,8 +31,7 @@ static grpc_completion_queue* default_create( const grpc_completion_queue_factory* factory, const grpc_completion_queue_attributes* attr) { return grpc_completion_queue_create_internal( - attr->cq_completion_type, attr->cq_polling_type, - static_cast<grpc_core::CQCallbackInterface*>(attr->cq_shutdown_cb)); + attr->cq_completion_type, attr->cq_polling_type, attr->cq_shutdown_cb); } static grpc_completion_queue_factory_vtable default_vtable = {default_create}; @@ -73,7 +72,8 @@ grpc_completion_queue* grpc_completion_queue_create_for_pluck(void* reserved) { } grpc_completion_queue* grpc_completion_queue_create_for_callback( - void* shutdown_callback, void* reserved) { + grpc_experimental_completion_queue_functor* shutdown_callback, + void* reserved) { GPR_ASSERT(!reserved); grpc_completion_queue_attributes attr = { 2, GRPC_CQ_CALLBACK, GRPC_CQ_DEFAULT_POLLING, shutdown_callback}; diff --git a/src/core/lib/surface/server.cc b/src/core/lib/surface/server.cc index 5fa58ffdec..72ddc2648d 100644 --- a/src/core/lib/surface/server.cc +++ b/src/core/lib/surface/server.cc @@ -150,12 +150,15 @@ struct call_data { grpc_closure kill_zombie_closure; grpc_closure* on_done_recv_initial_metadata; grpc_closure recv_trailing_metadata_ready; - grpc_error* error; + grpc_error* recv_initial_metadata_error; grpc_closure* original_recv_trailing_metadata_ready; + grpc_error* recv_trailing_metadata_error; + bool seen_recv_trailing_metadata_ready; grpc_closure publish; call_data* pending_next; + grpc_call_combiner* call_combiner; }; struct request_matcher { @@ -727,21 +730,43 @@ static void server_on_recv_initial_metadata(void* ptr, grpc_error* error) { if (calld->host_set && calld->path_set) { /* do nothing */ } else { + /* Pass the error reference to calld->recv_initial_metadata_error */ grpc_error* src_error = error; error = GRPC_ERROR_CREATE_REFERENCING_FROM_STATIC_STRING( - "Missing :authority or :path", &error, 1); + "Missing :authority or :path", &src_error, 1); GRPC_ERROR_UNREF(src_error); + calld->recv_initial_metadata_error = GRPC_ERROR_REF(error); } - - GRPC_CLOSURE_RUN(calld->on_done_recv_initial_metadata, error); + grpc_closure* closure = calld->on_done_recv_initial_metadata; + calld->on_done_recv_initial_metadata = nullptr; + if (calld->seen_recv_trailing_metadata_ready) { + GRPC_CALL_COMBINER_START(calld->call_combiner, + &calld->recv_trailing_metadata_ready, + calld->recv_trailing_metadata_error, + "continue server_recv_trailing_metadata_ready"); + } + GRPC_CLOSURE_RUN(closure, error); } static void server_recv_trailing_metadata_ready(void* user_data, - grpc_error* err) { + grpc_error* error) { grpc_call_element* elem = static_cast<grpc_call_element*>(user_data); call_data* calld = static_cast<call_data*>(elem->call_data); - err = grpc_error_add_child(GRPC_ERROR_REF(err), GRPC_ERROR_REF(calld->error)); - GRPC_CLOSURE_RUN(calld->original_recv_trailing_metadata_ready, err); + if (calld->on_done_recv_initial_metadata != nullptr) { + calld->recv_trailing_metadata_error = GRPC_ERROR_REF(error); + calld->seen_recv_trailing_metadata_ready = true; + GRPC_CLOSURE_INIT(&calld->recv_trailing_metadata_ready, + server_recv_trailing_metadata_ready, elem, + grpc_schedule_on_exec_ctx); + GRPC_CALL_COMBINER_STOP(calld->call_combiner, + "deferring server_recv_trailing_metadata_ready " + "until after server_on_recv_initial_metadata"); + return; + } + error = + grpc_error_add_child(GRPC_ERROR_REF(error), + GRPC_ERROR_REF(calld->recv_initial_metadata_error)); + GRPC_CLOSURE_RUN(calld->original_recv_trailing_metadata_ready, error); } static void server_mutate_op(grpc_call_element* elem, @@ -845,6 +870,7 @@ static grpc_error* init_call_elem(grpc_call_element* elem, memset(calld, 0, sizeof(call_data)); calld->deadline = GRPC_MILLIS_INF_FUTURE; calld->call = grpc_call_from_top_element(elem); + calld->call_combiner = args->call_combiner; GRPC_CLOSURE_INIT(&calld->server_on_recv_initial_metadata, server_on_recv_initial_metadata, elem, @@ -863,7 +889,7 @@ static void destroy_call_elem(grpc_call_element* elem, call_data* calld = static_cast<call_data*>(elem->call_data); GPR_ASSERT(calld->state != PENDING); - GRPC_ERROR_UNREF(calld->error); + GRPC_ERROR_UNREF(calld->recv_initial_metadata_error); if (calld->host_set) { grpc_slice_unref_internal(calld->host); } diff --git a/src/core/lib/transport/metadata.cc b/src/core/lib/transport/metadata.cc index d10194a2fe..d164502280 100644 --- a/src/core/lib/transport/metadata.cc +++ b/src/core/lib/transport/metadata.cc @@ -342,24 +342,6 @@ grpc_mdelem grpc_mdelem_from_grpc_metadata(grpc_metadata* metadata) { changed ? nullptr : reinterpret_cast<grpc_mdelem_data*>(metadata)); } -static size_t get_base64_encoded_size(size_t raw_length) { - static const uint8_t tail_xtra[3] = {0, 2, 3}; - return raw_length / 3 * 4 + tail_xtra[raw_length % 3]; -} - -size_t grpc_mdelem_get_size_in_hpack_table(grpc_mdelem elem, - bool use_true_binary_metadata) { - size_t overhead_and_key = 32 + GRPC_SLICE_LENGTH(GRPC_MDKEY(elem)); - size_t value_len = GRPC_SLICE_LENGTH(GRPC_MDVALUE(elem)); - if (grpc_is_binary_header(GRPC_MDKEY(elem))) { - return overhead_and_key + (use_true_binary_metadata - ? value_len + 1 - : get_base64_encoded_size(value_len)); - } else { - return overhead_and_key + value_len; - } -} - grpc_mdelem grpc_mdelem_ref(grpc_mdelem gmd DEBUG_ARGS) { switch (GRPC_MDELEM_STORAGE(gmd)) { case GRPC_MDELEM_STORAGE_EXTERNAL: diff --git a/src/core/lib/transport/metadata.h b/src/core/lib/transport/metadata.h index 78df4bc3a3..338082276c 100644 --- a/src/core/lib/transport/metadata.h +++ b/src/core/lib/transport/metadata.h @@ -125,9 +125,6 @@ grpc_mdelem grpc_mdelem_create( bool grpc_mdelem_eq(grpc_mdelem a, grpc_mdelem b); -size_t grpc_mdelem_get_size_in_hpack_table(grpc_mdelem elem, - bool use_true_binary_metadata); - /* Mutator and accessor for grpc_mdelem user data. The destructor function is used as a type tag and is checked during user_data fetch. */ void* grpc_mdelem_get_user_data(grpc_mdelem md, void (*if_destroy_func)(void*)); diff --git a/src/core/lib/transport/metadata_batch.cc b/src/core/lib/transport/metadata_batch.cc index 49740fcd1e..761aad19c4 100644 --- a/src/core/lib/transport/metadata_batch.cc +++ b/src/core/lib/transport/metadata_batch.cc @@ -105,7 +105,7 @@ static grpc_error* maybe_link_callout(grpc_metadata_batch* batch, return GRPC_ERROR_NONE; } if (batch->idx.array[idx] == nullptr) { - if (grpc_static_callout_is_default[idx]) ++batch->list.default_count; + ++batch->list.default_count; batch->idx.array[idx] = storage; return GRPC_ERROR_NONE; } @@ -121,7 +121,7 @@ static void maybe_unlink_callout(grpc_metadata_batch* batch, if (idx == GRPC_BATCH_CALLOUTS_COUNT) { return; } - if (grpc_static_callout_is_default[idx]) --batch->list.default_count; + --batch->list.default_count; GPR_ASSERT(batch->idx.array[idx] != nullptr); batch->idx.array[idx] = nullptr; } diff --git a/src/core/lib/transport/metadata_batch.h b/src/core/lib/transport/metadata_batch.h index 7068750b6f..0bcbb32d1f 100644 --- a/src/core/lib/transport/metadata_batch.h +++ b/src/core/lib/transport/metadata_batch.h @@ -82,6 +82,7 @@ void grpc_metadata_batch_set_value(grpc_linked_mdelem* storage, grpc_error* grpc_metadata_batch_link_head(grpc_metadata_batch* batch, grpc_linked_mdelem* storage) GRPC_MUST_USE_RESULT; + /** Add \a storage to the end of \a batch. storage->md is assumed to be valid. \a storage is owned by the caller and must survive for the @@ -100,6 +101,7 @@ grpc_error* grpc_metadata_batch_link_tail(grpc_metadata_batch* batch, grpc_error* grpc_metadata_batch_add_head( grpc_metadata_batch* batch, grpc_linked_mdelem* storage, grpc_mdelem elem_to_add) GRPC_MUST_USE_RESULT; + /** Add \a elem_to_add as the last element in \a batch, using \a storage as backing storage for the linked list element. \a storage is owned by the caller and must survive for the diff --git a/src/core/lib/transport/static_metadata.cc b/src/core/lib/transport/static_metadata.cc index 6a5144f21a..ed88aa3f28 100644 --- a/src/core/lib/transport/static_metadata.cc +++ b/src/core/lib/transport/static_metadata.cc @@ -568,33 +568,6 @@ grpc_mdelem_data grpc_static_mdelem_table[GRPC_STATIC_MDELEM_COUNT] = { {{&grpc_static_metadata_refcounts[16], {{g_bytes + 186, 15}}}, {&grpc_static_metadata_refcounts[102], {{g_bytes + 1101, 13}}}}, }; -bool grpc_static_callout_is_default[GRPC_BATCH_CALLOUTS_COUNT] = { - true, // :path - true, // :method - true, // :status - true, // :authority - true, // :scheme - true, // te - true, // grpc-message - true, // grpc-status - true, // grpc-payload-bin - true, // grpc-encoding - true, // grpc-accept-encoding - true, // grpc-server-stats-bin - true, // grpc-tags-bin - true, // grpc-trace-bin - true, // content-type - true, // content-encoding - true, // accept-encoding - true, // grpc-internal-encoding-request - true, // grpc-internal-stream-encoding-request - true, // user-agent - true, // host - true, // lb-token - true, // grpc-previous-rpc-attempts - true, // grpc-retry-pushback-ms -}; - const uint8_t grpc_static_accept_encoding_metadata[8] = {0, 76, 77, 78, 79, 80, 81, 82}; diff --git a/src/core/lib/transport/static_metadata.h b/src/core/lib/transport/static_metadata.h index b3a10f5873..1e95f4e5bb 100644 --- a/src/core/lib/transport/static_metadata.h +++ b/src/core/lib/transport/static_metadata.h @@ -587,8 +587,6 @@ typedef union { GRPC_BATCH_CALLOUTS_COUNT) \ : GRPC_BATCH_CALLOUTS_COUNT) -extern bool grpc_static_callout_is_default[GRPC_BATCH_CALLOUTS_COUNT]; - extern const uint8_t grpc_static_accept_encoding_metadata[8]; #define GRPC_MDELEM_ACCEPT_ENCODING_FOR_ALGORITHMS(algs) \ (GRPC_MAKE_MDELEM( \ |