diff options
author | Yang Gao <yangg@google.com> | 2016-04-04 19:59:26 +0000 |
---|---|---|
committer | Yang Gao <yangg@google.com> | 2016-04-04 19:59:26 +0000 |
commit | b49a1e5849529bdd769587dbd798edfabe516d5f (patch) | |
tree | 500adbeaa6f1172ad2886b0417f216e096eff17b /src/core | |
parent | a1821d64130288db642eb185cb013346f6c62ec6 (diff) | |
parent | 1b7c0a2c5cf27d7a77d9c3476fe6406a98ca3d76 (diff) |
Merge branch 'master' into hpack_table
Diffstat (limited to 'src/core')
66 files changed, 2420 insertions, 2500 deletions
diff --git a/src/core/lib/census/README.md b/src/core/ext/census/README.md index fb615a2194..fb615a2194 100644 --- a/src/core/lib/census/README.md +++ b/src/core/ext/census/README.md diff --git a/src/core/lib/census/aggregation.h b/src/core/ext/census/aggregation.h index 088ad808ef..45f789c772 100644 --- a/src/core/lib/census/aggregation.h +++ b/src/core/ext/census/aggregation.h @@ -33,8 +33,8 @@ #include <stddef.h> -#ifndef GRPC_CORE_LIB_CENSUS_AGGREGATION_H -#define GRPC_CORE_LIB_CENSUS_AGGREGATION_H +#ifndef GRPC_CORE_EXT_CENSUS_AGGREGATION_H +#define GRPC_CORE_EXT_CENSUS_AGGREGATION_H /** Structure used to describe an aggregation type. */ struct census_aggregation_ops { @@ -63,4 +63,4 @@ struct census_aggregation_ops { size_t (*print)(const void *aggregation, char *buffer, size_t n); }; -#endif /* GRPC_CORE_LIB_CENSUS_AGGREGATION_H */ +#endif /* GRPC_CORE_EXT_CENSUS_AGGREGATION_H */ diff --git a/src/core/lib/statistics/census_init.c b/src/core/ext/census/census_init.c index 7a8b8b660a..690b09e789 100644 --- a/src/core/lib/statistics/census_init.c +++ b/src/core/ext/census/census_init.c @@ -31,11 +31,11 @@ * */ -#include "src/core/lib/statistics/census_interface.h" +#include "src/core/ext/census/census_interface.h" #include <grpc/support/log.h> -#include "src/core/lib/statistics/census_rpc_stats.h" -#include "src/core/lib/statistics/census_tracing.h" +#include "src/core/ext/census/census_rpc_stats.h" +#include "src/core/ext/census/census_tracing.h" void census_init(void) { census_tracing_init(); diff --git a/src/core/lib/statistics/census_interface.h b/src/core/ext/census/census_interface.h index 4dce81b0c7..57e75f56ee 100644 --- a/src/core/lib/statistics/census_interface.h +++ b/src/core/ext/census/census_interface.h @@ -31,8 +31,8 @@ * */ -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_INTERFACE_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_INTERFACE_H +#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_INTERFACE_H +#define GRPC_CORE_EXT_CENSUS_CENSUS_INTERFACE_H #include <grpc/support/port_platform.h> @@ -73,4 +73,4 @@ census_op_id census_tracing_start_op(void); /* Ends tracing. Calling this function will invalidate the input op_id. */ void census_tracing_end_op(census_op_id op_id); -#endif /* GRPC_CORE_LIB_STATISTICS_CENSUS_INTERFACE_H */ +#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_INTERFACE_H */ diff --git a/src/core/lib/statistics/census_log.c b/src/core/ext/census/census_log.c index 782b34e3a1..9a7331adc2 100644 --- a/src/core/lib/statistics/census_log.c +++ b/src/core/ext/census/census_log.c @@ -89,7 +89,7 @@ include the name of the structure, which will be passed as the first argument. E.g. cl_block_initialize() will initialize a cl_block. */ -#include "src/core/lib/statistics/census_log.h" +#include "src/core/ext/census/census_log.h" #include <grpc/support/alloc.h> #include <grpc/support/atm.h> #include <grpc/support/cpu.h> diff --git a/src/core/lib/statistics/census_log.h b/src/core/ext/census/census_log.h index 6efcef49a4..534ecc5705 100644 --- a/src/core/lib/statistics/census_log.h +++ b/src/core/ext/census/census_log.h @@ -31,8 +31,8 @@ * */ -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_LOG_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_LOG_H +#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_LOG_H +#define GRPC_CORE_EXT_CENSUS_CENSUS_LOG_H #include <stddef.h> @@ -88,4 +88,4 @@ size_t census_log_remaining_space(void); out-of-space. */ int census_log_out_of_space_count(void); -#endif /* GRPC_CORE_LIB_STATISTICS_CENSUS_LOG_H */ +#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_LOG_H */ diff --git a/src/core/lib/statistics/census_rpc_stats.c b/src/core/ext/census/census_rpc_stats.c index 33522e0f40..09ee12d54b 100644 --- a/src/core/lib/statistics/census_rpc_stats.c +++ b/src/core/ext/census/census_rpc_stats.c @@ -36,11 +36,11 @@ #include <grpc/support/alloc.h> #include <grpc/support/log.h> #include <grpc/support/sync.h> -#include "src/core/lib/statistics/census_interface.h" -#include "src/core/lib/statistics/census_rpc_stats.h" -#include "src/core/lib/statistics/census_tracing.h" -#include "src/core/lib/statistics/hash_table.h" -#include "src/core/lib/statistics/window_stats.h" +#include "src/core/ext/census/census_interface.h" +#include "src/core/ext/census/census_rpc_stats.h" +#include "src/core/ext/census/census_tracing.h" +#include "src/core/ext/census/hash_table.h" +#include "src/core/ext/census/window_stats.h" #include "src/core/lib/support/murmur_hash.h" #include "src/core/lib/support/string.h" diff --git a/src/core/lib/statistics/census_rpc_stats.h b/src/core/ext/census/census_rpc_stats.h index 3853642362..7e4d8d1640 100644 --- a/src/core/lib/statistics/census_rpc_stats.h +++ b/src/core/ext/census/census_rpc_stats.h @@ -31,11 +31,11 @@ * */ -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_RPC_STATS_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_RPC_STATS_H +#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_RPC_STATS_H +#define GRPC_CORE_EXT_CENSUS_CENSUS_RPC_STATS_H #include <grpc/support/port_platform.h> -#include "src/core/lib/statistics/census_interface.h" +#include "src/core/ext/census/census_interface.h" #ifdef __cplusplus extern "C" { @@ -98,4 +98,4 @@ void census_stats_store_shutdown(void); } #endif -#endif /* GRPC_CORE_LIB_STATISTICS_CENSUS_RPC_STATS_H */ +#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_RPC_STATS_H */ diff --git a/src/core/lib/statistics/census_tracing.c b/src/core/ext/census/census_tracing.c index ff71e796ad..f893dc9864 100644 --- a/src/core/lib/statistics/census_tracing.c +++ b/src/core/ext/census/census_tracing.c @@ -31,8 +31,8 @@ * */ -#include "src/core/lib/statistics/census_tracing.h" -#include "src/core/lib/statistics/census_interface.h" +#include "src/core/ext/census/census_tracing.h" +#include "src/core/ext/census/census_interface.h" #include <stdio.h> #include <string.h> @@ -41,7 +41,7 @@ #include <grpc/support/log.h> #include <grpc/support/port_platform.h> #include <grpc/support/sync.h> -#include "src/core/lib/statistics/hash_table.h" +#include "src/core/ext/census/hash_table.h" #include "src/core/lib/support/string.h" void census_trace_obj_destroy(census_trace_obj *obj) { diff --git a/src/core/lib/statistics/census_tracing.h b/src/core/ext/census/census_tracing.h index 7b2db1e008..42a0d7403e 100644 --- a/src/core/lib/statistics/census_tracing.h +++ b/src/core/ext/census/census_tracing.h @@ -31,11 +31,11 @@ * */ -#ifndef GRPC_CORE_LIB_STATISTICS_CENSUS_TRACING_H -#define GRPC_CORE_LIB_STATISTICS_CENSUS_TRACING_H +#ifndef GRPC_CORE_EXT_CENSUS_CENSUS_TRACING_H +#define GRPC_CORE_EXT_CENSUS_CENSUS_TRACING_H #include <grpc/support/time.h> -#include "src/core/lib/statistics/census_rpc_stats.h" +#include "src/core/ext/census/census_rpc_stats.h" /* WARNING: The data structures and APIs provided by this file are for GRPC library's internal use ONLY. They might be changed in backward-incompatible @@ -93,4 +93,4 @@ census_trace_obj **census_get_active_ops(int *num_active_ops); } #endif -#endif /* GRPC_CORE_LIB_STATISTICS_CENSUS_TRACING_H */ +#endif /* GRPC_CORE_EXT_CENSUS_CENSUS_TRACING_H */ diff --git a/src/core/lib/census/context.c b/src/core/ext/census/context.c index 0dfc4ecbf1..0dfc4ecbf1 100644 --- a/src/core/lib/census/context.c +++ b/src/core/ext/census/context.c diff --git a/src/core/lib/census/grpc_context.c b/src/core/ext/census/grpc_context.c index 98285ab2d5..98285ab2d5 100644 --- a/src/core/lib/census/grpc_context.c +++ b/src/core/ext/census/grpc_context.c diff --git a/src/core/lib/census/grpc_filter.c b/src/core/ext/census/grpc_filter.c index 9bca2d434d..abfb3bb5f0 100644 --- a/src/core/lib/census/grpc_filter.c +++ b/src/core/ext/census/grpc_filter.c @@ -31,7 +31,7 @@ * */ -#include "src/core/lib/census/grpc_filter.h" +#include "src/core/ext/census/grpc_filter.h" #include <stdio.h> #include <string.h> @@ -42,9 +42,9 @@ #include <grpc/support/slice.h> #include <grpc/support/time.h> +#include "src/core/ext/census/census_interface.h" +#include "src/core/ext/census/census_rpc_stats.h" #include "src/core/lib/channel/channel_stack.h" -#include "src/core/lib/statistics/census_interface.h" -#include "src/core/lib/statistics/census_rpc_stats.h" #include "src/core/lib/transport/static_metadata.h" typedef struct call_data { diff --git a/src/core/lib/census/grpc_filter.h b/src/core/ext/census/grpc_filter.h index e747fe6590..a39bd82224 100644 --- a/src/core/lib/census/grpc_filter.h +++ b/src/core/ext/census/grpc_filter.h @@ -31,8 +31,8 @@ * */ -#ifndef GRPC_CORE_LIB_CENSUS_GRPC_FILTER_H -#define GRPC_CORE_LIB_CENSUS_GRPC_FILTER_H +#ifndef GRPC_CORE_EXT_CENSUS_GRPC_FILTER_H +#define GRPC_CORE_EXT_CENSUS_GRPC_FILTER_H #include "src/core/lib/channel/channel_stack.h" @@ -41,4 +41,4 @@ extern const grpc_channel_filter grpc_client_census_filter; extern const grpc_channel_filter grpc_server_census_filter; -#endif /* GRPC_CORE_LIB_CENSUS_GRPC_FILTER_H */ +#endif /* GRPC_CORE_EXT_CENSUS_GRPC_FILTER_H */ diff --git a/src/core/lib/census/grpc_plugin.c b/src/core/ext/census/grpc_plugin.c index 7f6df11f00..0f15ecb2c2 100644 --- a/src/core/lib/census/grpc_plugin.c +++ b/src/core/ext/census/grpc_plugin.c @@ -31,13 +31,11 @@ * */ -#include "src/core/lib/census/grpc_plugin.h" - #include <limits.h> #include <grpc/census.h> -#include "src/core/lib/census/grpc_filter.h" +#include "src/core/ext/census/grpc_filter.h" #include "src/core/lib/channel/channel_stack_builder.h" #include "src/core/lib/surface/channel_init.h" @@ -67,4 +65,4 @@ void census_grpc_plugin_init(void) { maybe_add_census_filter, NULL); } -void census_grpc_plugin_destroy(void) { census_shutdown(); } +void census_grpc_plugin_shutdown(void) { census_shutdown(); } diff --git a/src/core/lib/statistics/hash_table.c b/src/core/ext/census/hash_table.c index 79481c3eb1..ee6fdfc6e8 100644 --- a/src/core/lib/statistics/hash_table.c +++ b/src/core/ext/census/hash_table.c @@ -31,7 +31,7 @@ * */ -#include "src/core/lib/statistics/hash_table.h" +#include "src/core/ext/census/hash_table.h" #include <stddef.h> #include <stdio.h> diff --git a/src/core/lib/statistics/hash_table.h b/src/core/ext/census/hash_table.h index 45984a1c41..30ea4264a2 100644 --- a/src/core/lib/statistics/hash_table.h +++ b/src/core/ext/census/hash_table.h @@ -31,8 +31,8 @@ * */ -#ifndef GRPC_CORE_LIB_STATISTICS_HASH_TABLE_H -#define GRPC_CORE_LIB_STATISTICS_HASH_TABLE_H +#ifndef GRPC_CORE_EXT_CENSUS_HASH_TABLE_H +#define GRPC_CORE_EXT_CENSUS_HASH_TABLE_H #include <stddef.h> @@ -128,4 +128,4 @@ typedef void (*census_ht_itr_cb)(census_ht_key key, const void *val_ptr, should not invalidate data entries. */ uint64_t census_ht_for_all(const census_ht *ht, census_ht_itr_cb); -#endif /* GRPC_CORE_LIB_STATISTICS_HASH_TABLE_H */ +#endif /* GRPC_CORE_EXT_CENSUS_HASH_TABLE_H */ diff --git a/src/core/lib/census/initialize.c b/src/core/ext/census/initialize.c index 896276e44a..896276e44a 100644 --- a/src/core/lib/census/initialize.c +++ b/src/core/ext/census/initialize.c diff --git a/src/core/lib/census/mlog.c b/src/core/ext/census/mlog.c index ce060a2d85..698b7096ab 100644 --- a/src/core/lib/census/mlog.c +++ b/src/core/ext/census/mlog.c @@ -88,7 +88,7 @@ // include the name of the structure, which will be passed as the first // argument. E.g. cl_block_initialize() will initialize a cl_block. -#include "src/core/lib/census/mlog.h" +#include "src/core/ext/census/mlog.h" #include <grpc/support/alloc.h> #include <grpc/support/atm.h> #include <grpc/support/cpu.h> diff --git a/src/core/lib/census/mlog.h b/src/core/ext/census/mlog.h index dea43de725..a256426f91 100644 --- a/src/core/lib/census/mlog.h +++ b/src/core/ext/census/mlog.h @@ -33,8 +33,8 @@ /* A very fast in-memory log, optimized for multiple writers. */ -#ifndef GRPC_CORE_LIB_CENSUS_MLOG_H -#define GRPC_CORE_LIB_CENSUS_MLOG_H +#ifndef GRPC_CORE_EXT_CENSUS_MLOG_H +#define GRPC_CORE_EXT_CENSUS_MLOG_H #include <grpc/support/port_platform.h> #include <stddef.h> @@ -92,4 +92,4 @@ size_t census_log_remaining_space(void); out-of-space. */ int64_t census_log_out_of_space_count(void); -#endif /* GRPC_CORE_LIB_CENSUS_MLOG_H */ +#endif /* GRPC_CORE_EXT_CENSUS_MLOG_H */ diff --git a/src/core/lib/census/operation.c b/src/core/ext/census/operation.c index 5c58704372..5c58704372 100644 --- a/src/core/lib/census/operation.c +++ b/src/core/ext/census/operation.c diff --git a/src/core/lib/census/placeholders.c b/src/core/ext/census/placeholders.c index fe23d13971..fe23d13971 100644 --- a/src/core/lib/census/placeholders.c +++ b/src/core/ext/census/placeholders.c diff --git a/src/core/lib/census/rpc_metric_id.h b/src/core/ext/census/rpc_metric_id.h index 0f05b7de8a..888ec500a7 100644 --- a/src/core/lib/census/rpc_metric_id.h +++ b/src/core/ext/census/rpc_metric_id.h @@ -31,8 +31,8 @@ * */ -#ifndef GRPC_CORE_LIB_CENSUS_RPC_METRIC_ID_H -#define GRPC_CORE_LIB_CENSUS_RPC_METRIC_ID_H +#ifndef GRPC_CORE_EXT_CENSUS_RPC_METRIC_ID_H +#define GRPC_CORE_EXT_CENSUS_RPC_METRIC_ID_H /* Metric ID's used for RPC measurements. */ /* Count of client requests sent. */ @@ -48,4 +48,4 @@ /* Server side request latency. */ #define CENSUS_METRIC_RPC_SERVER_LATENCY ((uint32_t)5) -#endif /* GRPC_CORE_LIB_CENSUS_RPC_METRIC_ID_H */ +#endif /* GRPC_CORE_EXT_CENSUS_RPC_METRIC_ID_H */ diff --git a/src/core/lib/census/tracing.c b/src/core/ext/census/tracing.c index 3b5d6dab2b..3b5d6dab2b 100644 --- a/src/core/lib/census/tracing.c +++ b/src/core/ext/census/tracing.c diff --git a/src/core/lib/statistics/window_stats.c b/src/core/ext/census/window_stats.c index 9185e98d60..5f7bd9952e 100644 --- a/src/core/lib/statistics/window_stats.c +++ b/src/core/ext/census/window_stats.c @@ -31,7 +31,7 @@ * */ -#include "src/core/lib/statistics/window_stats.h" +#include "src/core/ext/census/window_stats.h" #include <grpc/support/alloc.h> #include <grpc/support/log.h> #include <grpc/support/time.h> diff --git a/src/core/lib/statistics/window_stats.h b/src/core/ext/census/window_stats.h index 629c08b8d5..25658c9ce0 100644 --- a/src/core/lib/statistics/window_stats.h +++ b/src/core/ext/census/window_stats.h @@ -31,8 +31,8 @@ * */ -#ifndef GRPC_CORE_LIB_STATISTICS_WINDOW_STATS_H -#define GRPC_CORE_LIB_STATISTICS_WINDOW_STATS_H +#ifndef GRPC_CORE_EXT_CENSUS_WINDOW_STATS_H +#define GRPC_CORE_EXT_CENSUS_WINDOW_STATS_H #include <grpc/support/time.h> @@ -170,4 +170,4 @@ void census_window_stats_get_sums(const struct census_window_stats *wstats, assertion failure). This function is thread-compatible. */ void census_window_stats_destroy(struct census_window_stats *wstats); -#endif /* GRPC_CORE_LIB_STATISTICS_WINDOW_STATS_H */ +#endif /* GRPC_CORE_EXT_CENSUS_WINDOW_STATS_H */ diff --git a/src/core/ext/resolver/dns/native/README.md b/src/core/ext/resolver/dns/native/README.md new file mode 100644 index 0000000000..695de47b9f --- /dev/null +++ b/src/core/ext/resolver/dns/native/README.md @@ -0,0 +1,2 @@ +dns: scheme name resolution, using getaddrbyname +(or other OS specific implementation) diff --git a/src/core/lib/client_config/resolvers/dns_resolver.c b/src/core/ext/resolver/dns/native/dns_resolver.c index ad586aa7e4..70d8a3fe2d 100644 --- a/src/core/lib/client_config/resolvers/dns_resolver.c +++ b/src/core/ext/resolver/dns/native/dns_resolver.c @@ -31,8 +31,6 @@ * */ -#include "src/core/lib/client_config/resolvers/dns_resolver.h" - #include <string.h> #include <grpc/support/alloc.h> @@ -40,6 +38,7 @@ #include <grpc/support/string_util.h> #include "src/core/lib/client_config/lb_policy_registry.h" +#include "src/core/lib/client_config/resolver_registry.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/timer.h" #include "src/core/lib/support/backoff.h" @@ -277,8 +276,8 @@ static grpc_resolver *dns_factory_create_resolver( return dns_create(args, "https", "pick_first"); } -char *dns_factory_get_default_host_name(grpc_resolver_factory *factory, - grpc_uri *uri) { +static char *dns_factory_get_default_host_name(grpc_resolver_factory *factory, + grpc_uri *uri) { const char *path = uri->path; if (path[0] == '/') ++path; return gpr_strdup(path); @@ -289,6 +288,12 @@ static const grpc_resolver_factory_vtable dns_factory_vtable = { dns_factory_get_default_host_name, "dns"}; static grpc_resolver_factory dns_resolver_factory = {&dns_factory_vtable}; -grpc_resolver_factory *grpc_dns_resolver_factory_create() { +static grpc_resolver_factory *dns_resolver_factory_create() { return &dns_resolver_factory; } + +void grpc_resolver_dns_native_init(void) { + grpc_register_resolver_type(dns_resolver_factory_create()); +} + +void grpc_resolver_dns_native_shutdown(void) {} diff --git a/src/core/ext/resolver/sockaddr/README.md b/src/core/ext/resolver/sockaddr/README.md new file mode 100644 index 0000000000..e307ba88f5 --- /dev/null +++ b/src/core/ext/resolver/sockaddr/README.md @@ -0,0 +1 @@ +Support for resolving ipv4:, ipv6:, unix: schemes diff --git a/src/core/lib/client_config/resolvers/sockaddr_resolver.c b/src/core/ext/resolver/sockaddr/sockaddr_resolver.c index 0e2656d033..69595ca3db 100644 --- a/src/core/lib/client_config/resolvers/sockaddr_resolver.c +++ b/src/core/ext/resolver/sockaddr/sockaddr_resolver.c @@ -33,8 +33,6 @@ #include <grpc/support/port_platform.h> -#include "src/core/lib/client_config/resolvers/sockaddr_resolver.h" - #include <stdio.h> #include <string.h> @@ -43,6 +41,7 @@ #include <grpc/support/string_util.h> #include "src/core/lib/client_config/lb_policy_registry.h" +#include "src/core/lib/client_config/resolver_registry.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/unix_sockets_posix.h" #include "src/core/lib/support/string.h" @@ -343,12 +342,20 @@ static void sockaddr_factory_unref(grpc_resolver_factory *factory) {} name##_factory_create_resolver, prefix##name##_get_default_authority, \ #name}; \ static grpc_resolver_factory name##_resolver_factory = { \ - &name##_factory_vtable}; \ - grpc_resolver_factory *grpc_##name##_resolver_factory_create() { \ - return &name##_resolver_factory; \ - } + &name##_factory_vtable} #ifdef GPR_HAVE_UNIX_SOCKET -DECL_FACTORY(unix, grpc_) +DECL_FACTORY(unix, grpc_); #endif -DECL_FACTORY(ipv4, ) DECL_FACTORY(ipv6, ) +DECL_FACTORY(ipv4, ); +DECL_FACTORY(ipv6, ); + +void grpc_resolver_sockaddr_init(void) { + grpc_register_resolver_type(&ipv4_resolver_factory); + grpc_register_resolver_type(&ipv6_resolver_factory); +#ifdef GPR_HAVE_UNIX_SOCKET + grpc_register_resolver_type(&unix_resolver_factory); +#endif +} + +void grpc_resolver_sockaddr_shutdown(void) {} diff --git a/src/core/ext/resolver/zookeeper/README.md b/src/core/ext/resolver/zookeeper/README.md new file mode 100644 index 0000000000..ce6f39683b --- /dev/null +++ b/src/core/ext/resolver/zookeeper/README.md @@ -0,0 +1 @@ +Zookeeper based name resolver: WIP diff --git a/src/core/lib/client_config/resolvers/zookeeper_resolver.c b/src/core/ext/resolver/zookeeper/zookeeper_resolver.c index a66c9925f0..5acb0940c6 100644 --- a/src/core/lib/client_config/resolvers/zookeeper_resolver.c +++ b/src/core/ext/resolver/zookeeper/zookeeper_resolver.c @@ -31,8 +31,6 @@ * */ -#include "src/core/lib/client_config/resolvers/zookeeper_resolver.h" - #include <string.h> #include <grpc/support/alloc.h> @@ -474,15 +472,6 @@ static grpc_resolver *zookeeper_create(grpc_resolver_args *args, return &r->base; } -static void zookeeper_plugin_init() { - grpc_register_resolver_type(grpc_zookeeper_resolver_factory_create()); -} - -void grpc_zookeeper_register() { - GRPC_API_TRACE("grpc_zookeeper_register(void)", 0, ()); - grpc_register_plugin(zookeeper_plugin_init, NULL); -} - /* * FACTORY */ @@ -509,6 +498,15 @@ static const grpc_resolver_factory_vtable zookeeper_factory_vtable = { static grpc_resolver_factory zookeeper_resolver_factory = { &zookeeper_factory_vtable}; -grpc_resolver_factory *grpc_zookeeper_resolver_factory_create() { +static grpc_resolver_factory *zookeeper_resolver_factory_create() { return &zookeeper_resolver_factory; } + +static void zookeeper_plugin_init() { + grpc_register_resolver_type(zookeeper_resolver_factory_create()); +} + +void grpc_zookeeper_register() { + GRPC_API_TRACE("grpc_zookeeper_register(void)", 0, ()); + grpc_register_plugin(zookeeper_plugin_init, NULL); +} diff --git a/src/core/ext/transport/chttp2/client/insecure/channel_create.c b/src/core/ext/transport/chttp2/client/insecure/channel_create.c index a5f7c6cfbc..606fff5fb4 100644 --- a/src/core/ext/transport/chttp2/client/insecure/channel_create.c +++ b/src/core/ext/transport/chttp2/client/insecure/channel_create.c @@ -40,8 +40,8 @@ #include <grpc/support/slice.h> #include <grpc/support/slice_buffer.h> +#include "src/core/ext/census/grpc_filter.h" #include "src/core/ext/transport/chttp2/transport/chttp2_transport.h" -#include "src/core/lib/census/grpc_filter.h" #include "src/core/lib/channel/channel_args.h" #include "src/core/lib/channel/client_channel.h" #include "src/core/lib/channel/compress_filter.h" diff --git a/src/core/lib/client_config/resolver_registry.c b/src/core/lib/client_config/resolver_registry.c index e8432d165a..5f3db273b5 100644 --- a/src/core/lib/client_config/resolver_registry.c +++ b/src/core/lib/client_config/resolver_registry.c @@ -70,14 +70,11 @@ void grpc_register_resolver_type(grpc_resolver_factory *factory) { g_all_of_the_resolvers[g_number_of_resolvers++] = factory; } -static grpc_resolver_factory *lookup_factory(grpc_uri *uri) { +static grpc_resolver_factory *lookup_factory(const char *name) { int i; - /* handling NULL uri's here simplifies grpc_resolver_create */ - if (!uri) return NULL; - for (i = 0; i < g_number_of_resolvers; i++) { - if (0 == strcmp(uri->scheme, g_all_of_the_resolvers[i]->vtable->scheme)) { + if (0 == strcmp(name, g_all_of_the_resolvers[i]->vtable->scheme)) { return g_all_of_the_resolvers[i]; } } @@ -85,6 +82,17 @@ static grpc_resolver_factory *lookup_factory(grpc_uri *uri) { return NULL; } +grpc_resolver_factory *grpc_resolver_factory_lookup(const char *name) { + grpc_resolver_factory *f = lookup_factory(name); + if (f) grpc_resolver_factory_ref(f); + return f; +} + +static grpc_resolver_factory *lookup_factory_by_uri(grpc_uri *uri) { + if (!uri) return NULL; + return lookup_factory(uri->scheme); +} + static grpc_resolver_factory *resolve_factory(const char *target, grpc_uri **uri) { char *tmp; @@ -92,13 +100,13 @@ static grpc_resolver_factory *resolve_factory(const char *target, GPR_ASSERT(uri != NULL); *uri = grpc_uri_parse(target, 1); - factory = lookup_factory(*uri); + factory = lookup_factory_by_uri(*uri); if (factory == NULL) { if (g_default_resolver_prefix != NULL) { grpc_uri_destroy(*uri); gpr_asprintf(&tmp, "%s%s", g_default_resolver_prefix, target); *uri = grpc_uri_parse(tmp, 1); - factory = lookup_factory(*uri); + factory = lookup_factory_by_uri(*uri); if (factory == NULL) { grpc_uri_destroy(grpc_uri_parse(target, 0)); grpc_uri_destroy(grpc_uri_parse(tmp, 0)); diff --git a/src/core/lib/client_config/resolver_registry.h b/src/core/lib/client_config/resolver_registry.h index 330cb7f544..36c4f2fe03 100644 --- a/src/core/lib/client_config/resolver_registry.h +++ b/src/core/lib/client_config/resolver_registry.h @@ -58,6 +58,10 @@ void grpc_register_resolver_type(grpc_resolver_factory *factory); grpc_resolver *grpc_resolver_create( const char *target, grpc_subchannel_factory *subchannel_factory); +/** Find a resolver factory given a name and return an (owned-by-the-caller) + * reference to it */ +grpc_resolver_factory *grpc_resolver_factory_lookup(const char *name); + /** Given a target, return a (freshly allocated with gpr_malloc) string representing the default authority to pass from a client. */ char *grpc_get_default_authority(const char *target); diff --git a/src/core/lib/client_config/resolvers/dns_resolver.h b/src/core/lib/client_config/resolvers/dns_resolver.h deleted file mode 100644 index de41961ea9..0000000000 --- a/src/core/lib/client_config/resolvers/dns_resolver.h +++ /dev/null @@ -1,42 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_DNS_RESOLVER_H -#define GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_DNS_RESOLVER_H - -#include "src/core/lib/client_config/resolver_factory.h" - -/** Create a dns resolver factory */ -grpc_resolver_factory *grpc_dns_resolver_factory_create(void); - -#endif /* GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_DNS_RESOLVER_H */ diff --git a/src/core/lib/client_config/resolvers/sockaddr_resolver.h b/src/core/lib/client_config/resolvers/sockaddr_resolver.h deleted file mode 100644 index a8b208176c..0000000000 --- a/src/core/lib/client_config/resolvers/sockaddr_resolver.h +++ /dev/null @@ -1,50 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_SOCKADDR_RESOLVER_H -#define GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_SOCKADDR_RESOLVER_H - -#include <grpc/support/port_platform.h> - -#include "src/core/lib/client_config/resolver_factory.h" - -grpc_resolver_factory *grpc_ipv4_resolver_factory_create(void); - -grpc_resolver_factory *grpc_ipv6_resolver_factory_create(void); - -#ifdef GPR_POSIX_SOCKET -/** Create a unix resolver factory */ -grpc_resolver_factory *grpc_unix_resolver_factory_create(void); -#endif - -#endif /* GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_SOCKADDR_RESOLVER_H */ diff --git a/src/core/lib/client_config/resolvers/zookeeper_resolver.h b/src/core/lib/client_config/resolvers/zookeeper_resolver.h deleted file mode 100644 index a067bd0bdb..0000000000 --- a/src/core/lib/client_config/resolvers/zookeeper_resolver.h +++ /dev/null @@ -1,42 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_ZOOKEEPER_RESOLVER_H -#define GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_ZOOKEEPER_RESOLVER_H - -#include "src/core/lib/client_config/resolver_factory.h" - -/** Create a zookeeper resolver factory */ -grpc_resolver_factory *grpc_zookeeper_resolver_factory_create(void); - -#endif /* GRPC_CORE_LIB_CLIENT_CONFIG_RESOLVERS_ZOOKEEPER_RESOLVER_H */ diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c new file mode 100644 index 0000000000..dba335490b --- /dev/null +++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c @@ -0,0 +1,1936 @@ +/* + * + * Copyright 2015-2016, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +/* This file will be removed shortly: it's here to keep refactoring + * steps simple and auditable. + * It's the combination of the old files: + * - fd_posix.{h,c} + * - pollset_posix.{h,c} + * - pullset_multipoller_with_{poll,epoll}.{h,c} + * The new version will be split into: + * - ev_poll_posix.{h,c} + * - ev_epoll_posix.{h,c} + */ + +#include <grpc/support/port_platform.h> + +#ifdef GPR_POSIX_SOCKET + +#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" + +#include <assert.h> +#include <errno.h> +#include <poll.h> +#include <string.h> +#include <sys/socket.h> +#include <unistd.h> + +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> +#include <grpc/support/string_util.h> +#include <grpc/support/tls.h> +#include <grpc/support/useful.h> + +#include "src/core/lib/iomgr/iomgr_internal.h" +#include "src/core/lib/iomgr/wakeup_fd_posix.h" +#include "src/core/lib/profiling/timers.h" +#include "src/core/lib/support/block_annotate.h" + +/******************************************************************************* + * FD declarations + */ + +typedef struct grpc_fd_watcher { + struct grpc_fd_watcher *next; + struct grpc_fd_watcher *prev; + grpc_pollset *pollset; + grpc_pollset_worker *worker; + grpc_fd *fd; +} grpc_fd_watcher; + +struct grpc_fd { + int fd; + /* refst format: + bit0: 1=active/0=orphaned + bit1-n: refcount + meaning that mostly we ref by two to avoid altering the orphaned bit, + and just unref by 1 when we're ready to flag the object as orphaned */ + gpr_atm refst; + + gpr_mu mu; + int shutdown; + int closed; + int released; + + /* The watcher list. + + The following watcher related fields are protected by watcher_mu. + + An fd_watcher is an ephemeral object created when an fd wants to + begin polling, and destroyed after the poll. + + It denotes the fd's interest in whether to read poll or write poll + or both or neither on this fd. + + If a watcher is asked to poll for reads or writes, the read_watcher + or write_watcher fields are set respectively. A watcher may be asked + to poll for both, in which case both fields will be set. + + read_watcher and write_watcher may be NULL if no watcher has been + asked to poll for reads or writes. + + If an fd_watcher is not asked to poll for reads or writes, it's added + to a linked list of inactive watchers, rooted at inactive_watcher_root. + If at a later time there becomes need of a poller to poll, one of + the inactive pollers may be kicked out of their poll loops to take + that responsibility. */ + grpc_fd_watcher inactive_watcher_root; + grpc_fd_watcher *read_watcher; + grpc_fd_watcher *write_watcher; + + grpc_closure *read_closure; + grpc_closure *write_closure; + + struct grpc_fd *freelist_next; + + grpc_closure *on_done_closure; + + grpc_iomgr_object iomgr_object; +}; + +/* Begin polling on an fd. + Registers that the given pollset is interested in this fd - so that if read + or writability interest changes, the pollset can be kicked to pick up that + new interest. + Return value is: + (fd_needs_read? read_mask : 0) | (fd_needs_write? write_mask : 0) + i.e. a combination of read_mask and write_mask determined by the fd's current + interest in said events. + Polling strategies that do not need to alter their behavior depending on the + fd's current interest (such as epoll) do not need to call this function. + MUST NOT be called with a pollset lock taken */ +static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, + grpc_pollset_worker *worker, uint32_t read_mask, + uint32_t write_mask, grpc_fd_watcher *rec); +/* Complete polling previously started with fd_begin_poll + MUST NOT be called with a pollset lock taken + if got_read or got_write are 1, also does the become_{readable,writable} as + appropriate. */ +static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec, + int got_read, int got_write); + +/* Return 1 if this fd is orphaned, 0 otherwise */ +static bool fd_is_orphaned(grpc_fd *fd); + +/* Reference counting for fds */ +/*#define GRPC_FD_REF_COUNT_DEBUG*/ +#ifdef GRPC_FD_REF_COUNT_DEBUG +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); + +#define CLOSURE_NOT_READY ((grpc_closure *)0) +#define CLOSURE_READY ((grpc_closure *)1) + +/******************************************************************************* + * pollset declarations + */ + +typedef struct grpc_pollset_vtable grpc_pollset_vtable; + +typedef struct grpc_cached_wakeup_fd { + grpc_wakeup_fd fd; + struct grpc_cached_wakeup_fd *next; +} grpc_cached_wakeup_fd; + +struct grpc_pollset_worker { + grpc_cached_wakeup_fd *wakeup_fd; + int reevaluate_polling_on_wakeup; + int kicked_specifically; + struct grpc_pollset_worker *next; + struct grpc_pollset_worker *prev; +}; + +struct grpc_pollset { + /* pollsets under posix can mutate representation as fds are added and + removed. + For example, we may choose a poll() based implementation on linux for + few fds, and an epoll() based implementation for many fds */ + const grpc_pollset_vtable *vtable; + gpr_mu mu; + grpc_pollset_worker root_worker; + int in_flight_cbs; + int shutting_down; + int called_shutdown; + int kicked_without_pollers; + grpc_closure *shutdown_done; + grpc_closure_list idle_jobs; + union { + int fd; + void *ptr; + } data; + /* Local cache of eventfds for workers */ + grpc_cached_wakeup_fd *local_wakeup_cache; +}; + +struct grpc_pollset_vtable { + void (*add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd, int and_unlock_pollset); + void (*maybe_work_and_unlock)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker *worker, + gpr_timespec deadline, gpr_timespec now); + void (*finish_shutdown)(grpc_pollset *pollset); + void (*destroy)(grpc_pollset *pollset); +}; + +/* Add an fd to a pollset */ +static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd); + +static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + +/* Convert a timespec to milliseconds: + - very small or negative poll times are clamped to zero to do a + non-blocking poll (which becomes spin polling) + - other small values are rounded up to one millisecond + - longer than a millisecond polls are rounded up to the next nearest + millisecond to avoid spinning + - infinite timeouts are converted to -1 */ +static int poll_deadline_to_millis_timeout(gpr_timespec deadline, + gpr_timespec now); + +/* Allow kick to wakeup the currently polling worker */ +#define GRPC_POLLSET_CAN_KICK_SELF 1 +/* Force the wakee to repoll when awoken */ +#define GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP 2 +/* As per pollset_kick, with an extended set of flags (defined above) + -- mostly for fd_posix's use. */ +static void pollset_kick_ext(grpc_pollset *p, + grpc_pollset_worker *specific_worker, + uint32_t flags); + +/* turn a pollset into a multipoller: platform specific */ +typedef void (*platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + struct grpc_fd **fds, + size_t fd_count); +static platform_become_multipoller_type platform_become_multipoller; + +/* Return 1 if the pollset has active threads in pollset_work (pollset must + * be locked) */ +static int pollset_has_workers(grpc_pollset *pollset); + +static void remove_fd_from_all_epoll_sets(int fd); + +/******************************************************************************* + * pollset_set definitions + */ + +struct grpc_pollset_set { + gpr_mu mu; + + size_t pollset_count; + size_t pollset_capacity; + grpc_pollset **pollsets; + + size_t pollset_set_count; + size_t pollset_set_capacity; + struct grpc_pollset_set **pollset_sets; + + size_t fd_count; + size_t fd_capacity; + grpc_fd **fds; +}; + +/******************************************************************************* + * fd_posix.c + */ + +/* 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. + */ +/* TODO(klempner): We could use some form of polling generation count to know + * when these are safe to free. */ +/* TODO(klempner): Consider disabling freelisting if we don't have multiple + * threads in poll on the same fd */ +/* TODO(klempner): Batch these allocations to reduce fragmentation */ +static grpc_fd *fd_freelist = NULL; +static gpr_mu fd_freelist_mu; + +static void freelist_fd(grpc_fd *fd) { + gpr_mu_lock(&fd_freelist_mu); + fd->freelist_next = fd_freelist; + fd_freelist = fd; + grpc_iomgr_unregister_object(&fd->iomgr_object); + gpr_mu_unlock(&fd_freelist_mu); +} + +static grpc_fd *alloc_fd(int fd) { + grpc_fd *r = NULL; + gpr_mu_lock(&fd_freelist_mu); + if (fd_freelist != NULL) { + r = fd_freelist; + fd_freelist = fd_freelist->freelist_next; + } + gpr_mu_unlock(&fd_freelist_mu); + if (r == NULL) { + r = gpr_malloc(sizeof(grpc_fd)); + gpr_mu_init(&r->mu); + } + + gpr_mu_lock(&r->mu); + gpr_atm_rel_store(&r->refst, 1); + r->shutdown = 0; + r->read_closure = CLOSURE_NOT_READY; + r->write_closure = CLOSURE_NOT_READY; + r->fd = fd; + r->inactive_watcher_root.next = r->inactive_watcher_root.prev = + &r->inactive_watcher_root; + r->freelist_next = NULL; + r->read_watcher = r->write_watcher = NULL; + r->on_done_closure = NULL; + r->closed = 0; + r->released = 0; + gpr_mu_unlock(&r->mu); + return r; +} + +static void destroy(grpc_fd *fd) { + gpr_mu_destroy(&fd->mu); + gpr_free(fd); +} + +#ifdef GRPC_FD_REF_COUNT_DEBUG +#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) { + gpr_log(GPR_DEBUG, "FD %d %p ref %d %d -> %d [%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); +} + +#ifdef GRPC_FD_REF_COUNT_DEBUG +static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file, + int line) { + gpr_atm old; + gpr_log(GPR_DEBUG, "FD %d %p unref %d %d -> %d [%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) { + gpr_atm old; +#endif + old = gpr_atm_full_fetch_add(&fd->refst, -n); + if (old == n) { + freelist_fd(fd); + } else { + GPR_ASSERT(old > n); + } +} + +static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } + +static void fd_global_shutdown(void) { + gpr_mu_lock(&fd_freelist_mu); + gpr_mu_unlock(&fd_freelist_mu); + while (fd_freelist != NULL) { + grpc_fd *fd = fd_freelist; + fd_freelist = fd_freelist->freelist_next; + destroy(fd); + } + gpr_mu_destroy(&fd_freelist_mu); +} + +static grpc_fd *fd_create(int fd, const char *name) { + grpc_fd *r = alloc_fd(fd); + char *name2; + gpr_asprintf(&name2, "%s fd=%d", name, fd); + grpc_iomgr_register_object(&r->iomgr_object, name2); + gpr_free(name2); +#ifdef GRPC_FD_REF_COUNT_DEBUG + gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, r, name); +#endif + return r; +} + +static bool fd_is_orphaned(grpc_fd *fd) { + return (gpr_atm_acq_load(&fd->refst) & 1) == 0; +} + +static void pollset_kick_locked(grpc_fd_watcher *watcher) { + gpr_mu_lock(&watcher->pollset->mu); + GPR_ASSERT(watcher->worker); + pollset_kick_ext(watcher->pollset, watcher->worker, + GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP); + gpr_mu_unlock(&watcher->pollset->mu); +} + +static void maybe_wake_one_watcher_locked(grpc_fd *fd) { + if (fd->inactive_watcher_root.next != &fd->inactive_watcher_root) { + pollset_kick_locked(fd->inactive_watcher_root.next); + } else if (fd->read_watcher) { + pollset_kick_locked(fd->read_watcher); + } else if (fd->write_watcher) { + pollset_kick_locked(fd->write_watcher); + } +} + +static void wake_all_watchers_locked(grpc_fd *fd) { + grpc_fd_watcher *watcher; + for (watcher = fd->inactive_watcher_root.next; + watcher != &fd->inactive_watcher_root; watcher = watcher->next) { + pollset_kick_locked(watcher); + } + if (fd->read_watcher) { + pollset_kick_locked(fd->read_watcher); + } + if (fd->write_watcher && fd->write_watcher != fd->read_watcher) { + pollset_kick_locked(fd->write_watcher); + } +} + +static int has_watchers(grpc_fd *fd) { + return fd->read_watcher != NULL || fd->write_watcher != NULL || + fd->inactive_watcher_root.next != &fd->inactive_watcher_root; +} + +static void close_fd_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + fd->closed = 1; + if (!fd->released) { + close(fd->fd); + } else { + remove_fd_from_all_epoll_sets(fd->fd); + } + grpc_exec_ctx_enqueue(exec_ctx, fd->on_done_closure, true, NULL); +} + +static int fd_wrapped_fd(grpc_fd *fd) { + if (fd->released || fd->closed) { + return -1; + } else { + return fd->fd; + } +} + +static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *on_done, int *release_fd, + const char *reason) { + fd->on_done_closure = on_done; + fd->released = release_fd != NULL; + if (!fd->released) { + shutdown(fd->fd, SHUT_RDWR); + } else { + *release_fd = fd->fd; + } + gpr_mu_lock(&fd->mu); + REF_BY(fd, 1, reason); /* remove active status, but keep referenced */ + if (!has_watchers(fd)) { + close_fd_locked(exec_ctx, fd); + } else { + wake_all_watchers_locked(fd); + } + gpr_mu_unlock(&fd->mu); + UNREF_BY(fd, 2, reason); /* drop the reference */ +} + +/* increment refcount by two to avoid changing the orphan bit */ +#ifdef GRPC_FD_REF_COUNT_DEBUG +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 notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure **st, grpc_closure *closure) { + if (*st == CLOSURE_NOT_READY) { + /* not ready ==> switch to a waiting state by setting the closure */ + *st = closure; + } else if (*st == CLOSURE_READY) { + /* already ready ==> queue the closure to run immediately */ + *st = CLOSURE_NOT_READY; + grpc_exec_ctx_enqueue(exec_ctx, closure, !fd->shutdown, NULL); + maybe_wake_one_watcher_locked(fd); + } else { + /* upcallptr was set to a different closure. This is an error! */ + gpr_log(GPR_ERROR, + "User called a notify_on function with a previous callback still " + "pending"); + abort(); + } +} + +/* returns 1 if state becomes not ready */ +static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure **st) { + if (*st == CLOSURE_READY) { + /* duplicate ready ==> ignore */ + return 0; + } else if (*st == CLOSURE_NOT_READY) { + /* not ready, and not waiting ==> flag ready */ + *st = CLOSURE_READY; + return 0; + } else { + /* waiting ==> queue closure */ + grpc_exec_ctx_enqueue(exec_ctx, *st, !fd->shutdown, NULL); + *st = CLOSURE_NOT_READY; + return 1; + } +} + +static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + gpr_mu_lock(&fd->mu); + GPR_ASSERT(!fd->shutdown); + fd->shutdown = 1; + set_ready_locked(exec_ctx, fd, &fd->read_closure); + set_ready_locked(exec_ctx, fd, &fd->write_closure); + gpr_mu_unlock(&fd->mu); +} + +static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + gpr_mu_lock(&fd->mu); + notify_on_locked(exec_ctx, fd, &fd->read_closure, closure); + gpr_mu_unlock(&fd->mu); +} + +static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + gpr_mu_lock(&fd->mu); + notify_on_locked(exec_ctx, fd, &fd->write_closure, closure); + gpr_mu_unlock(&fd->mu); +} + +static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, + grpc_pollset_worker *worker, uint32_t read_mask, + uint32_t write_mask, grpc_fd_watcher *watcher) { + uint32_t mask = 0; + grpc_closure *cur; + int requested; + /* keep track of pollers that have requested our events, in case they change + */ + GRPC_FD_REF(fd, "poll"); + + gpr_mu_lock(&fd->mu); + + /* if we are shutdown, then don't add to the watcher set */ + if (fd->shutdown) { + watcher->fd = NULL; + watcher->pollset = NULL; + watcher->worker = NULL; + gpr_mu_unlock(&fd->mu); + GRPC_FD_UNREF(fd, "poll"); + return 0; + } + + /* if there is nobody polling for read, but we need to, then start doing so */ + cur = fd->read_closure; + requested = cur != CLOSURE_READY; + if (read_mask && fd->read_watcher == NULL && requested) { + fd->read_watcher = watcher; + mask |= read_mask; + } + /* if there is nobody polling for write, but we need to, then start doing so + */ + cur = fd->write_closure; + requested = cur != CLOSURE_READY; + if (write_mask && fd->write_watcher == NULL && requested) { + fd->write_watcher = watcher; + mask |= write_mask; + } + /* if not polling, remember this watcher in case we need someone to later */ + if (mask == 0 && worker != NULL) { + watcher->next = &fd->inactive_watcher_root; + watcher->prev = watcher->next->prev; + watcher->next->prev = watcher->prev->next = watcher; + } + watcher->pollset = pollset; + watcher->worker = worker; + watcher->fd = fd; + gpr_mu_unlock(&fd->mu); + + return mask; +} + +static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher, + int got_read, int got_write) { + int was_polling = 0; + int kick = 0; + grpc_fd *fd = watcher->fd; + + if (fd == NULL) { + return; + } + + gpr_mu_lock(&fd->mu); + + if (watcher == fd->read_watcher) { + /* remove read watcher, kick if we still need a read */ + was_polling = 1; + if (!got_read) { + kick = 1; + } + fd->read_watcher = NULL; + } + if (watcher == fd->write_watcher) { + /* remove write watcher, kick if we still need a write */ + was_polling = 1; + if (!got_write) { + kick = 1; + } + fd->write_watcher = NULL; + } + if (!was_polling && watcher->worker != NULL) { + /* remove from inactive list */ + watcher->next->prev = watcher->prev; + watcher->prev->next = watcher->next; + } + if (got_read) { + if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) { + kick = 1; + } + } + if (got_write) { + if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) { + kick = 1; + } + } + if (kick) { + maybe_wake_one_watcher_locked(fd); + } + if (fd_is_orphaned(fd) && !has_watchers(fd) && !fd->closed) { + close_fd_locked(exec_ctx, fd); + } + gpr_mu_unlock(&fd->mu); + + GRPC_FD_UNREF(fd, "poll"); +} + +/******************************************************************************* + * pollset_posix.c + */ + +GPR_TLS_DECL(g_current_thread_poller); +GPR_TLS_DECL(g_current_thread_worker); + +/** 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. */ +grpc_wakeup_fd grpc_global_wakeup_fd; + +static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->prev->next = worker->next; + worker->next->prev = worker->prev; +} + +static int pollset_has_workers(grpc_pollset *p) { + return p->root_worker.next != &p->root_worker; +} + +static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) { + if (pollset_has_workers(p)) { + grpc_pollset_worker *w = p->root_worker.next; + remove_worker(p, w); + return w; + } else { + return NULL; + } +} + +static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->next = &p->root_worker; + worker->prev = worker->next->prev; + worker->prev->next = worker->next->prev = worker; +} + +static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) { + worker->prev = &p->root_worker; + worker->next = worker->prev->next; + worker->prev->next = worker->next->prev = worker; +} + +static void pollset_kick_ext(grpc_pollset *p, + grpc_pollset_worker *specific_worker, + uint32_t flags) { + GPR_TIMER_BEGIN("pollset_kick_ext", 0); + + /* pollset->mu already held */ + if (specific_worker != NULL) { + if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) { + GPR_TIMER_BEGIN("pollset_kick_ext.broadcast", 0); + GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); + for (specific_worker = p->root_worker.next; + specific_worker != &p->root_worker; + specific_worker = specific_worker->next) { + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } + p->kicked_without_pollers = 1; + GPR_TIMER_END("pollset_kick_ext.broadcast", 0); + } else if (gpr_tls_get(&g_current_thread_worker) != + (intptr_t)specific_worker) { + GPR_TIMER_MARK("different_thread_worker", 0); + if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { + specific_worker->reevaluate_polling_on_wakeup = 1; + } + specific_worker->kicked_specifically = 1; + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) { + GPR_TIMER_MARK("kick_yoself", 0); + if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { + specific_worker->reevaluate_polling_on_wakeup = 1; + } + specific_worker->kicked_specifically = 1; + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } + } else if (gpr_tls_get(&g_current_thread_poller) != (intptr_t)p) { + GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); + GPR_TIMER_MARK("kick_anonymous", 0); + specific_worker = pop_front_worker(p); + if (specific_worker != NULL) { + if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) { + GPR_TIMER_MARK("kick_anonymous_not_self", 0); + push_back_worker(p, specific_worker); + specific_worker = pop_front_worker(p); + if ((flags & GRPC_POLLSET_CAN_KICK_SELF) == 0 && + gpr_tls_get(&g_current_thread_worker) == + (intptr_t)specific_worker) { + push_back_worker(p, specific_worker); + specific_worker = NULL; + } + } + if (specific_worker != NULL) { + GPR_TIMER_MARK("finally_kick", 0); + push_back_worker(p, specific_worker); + grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); + } + } else { + GPR_TIMER_MARK("kicked_no_pollers", 0); + p->kicked_without_pollers = 1; + } + } + + GPR_TIMER_END("pollset_kick_ext", 0); +} + +static void pollset_kick(grpc_pollset *p, + grpc_pollset_worker *specific_worker) { + pollset_kick_ext(p, specific_worker, 0); +} + +/* global state management */ + +static void pollset_global_init(void) { + gpr_tls_init(&g_current_thread_poller); + gpr_tls_init(&g_current_thread_worker); + grpc_wakeup_fd_global_init(); + grpc_wakeup_fd_init(&grpc_global_wakeup_fd); +} + +static void pollset_global_shutdown(void) { + grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd); + gpr_tls_destroy(&g_current_thread_poller); + gpr_tls_destroy(&g_current_thread_worker); + grpc_wakeup_fd_global_destroy(); +} + +static void kick_poller(void) { grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); } + +/* main interface */ + +static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null); + +static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) { + gpr_mu_init(&pollset->mu); + *mu = &pollset->mu; + pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; + pollset->in_flight_cbs = 0; + pollset->shutting_down = 0; + pollset->called_shutdown = 0; + pollset->kicked_without_pollers = 0; + pollset->idle_jobs.head = pollset->idle_jobs.tail = NULL; + pollset->local_wakeup_cache = NULL; + pollset->kicked_without_pollers = 0; + become_basic_pollset(pollset, NULL); +} + +static void pollset_destroy(grpc_pollset *pollset) { + GPR_ASSERT(pollset->in_flight_cbs == 0); + GPR_ASSERT(!pollset_has_workers(pollset)); + GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); + pollset->vtable->destroy(pollset); + while (pollset->local_wakeup_cache) { + grpc_cached_wakeup_fd *next = pollset->local_wakeup_cache->next; + grpc_wakeup_fd_destroy(&pollset->local_wakeup_cache->fd); + gpr_free(pollset->local_wakeup_cache); + pollset->local_wakeup_cache = next; + } + gpr_mu_destroy(&pollset->mu); +} + +static void pollset_reset(grpc_pollset *pollset) { + GPR_ASSERT(pollset->shutting_down); + GPR_ASSERT(pollset->in_flight_cbs == 0); + GPR_ASSERT(!pollset_has_workers(pollset)); + GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); + pollset->vtable->destroy(pollset); + pollset->shutting_down = 0; + pollset->called_shutdown = 0; + pollset->kicked_without_pollers = 0; + become_basic_pollset(pollset, NULL); +} + +static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd) { + gpr_mu_lock(&pollset->mu); + pollset->vtable->add_fd(exec_ctx, pollset, fd, 1); +/* the following (enabled only in debug) will reacquire and then release + our lock - meaning that if the unlocking flag passed to add_fd above is + not respected, the code will deadlock (in a way that we have a chance of + debugging) */ +#ifndef NDEBUG + gpr_mu_lock(&pollset->mu); + gpr_mu_unlock(&pollset->mu); +#endif +} + +static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) { + GPR_ASSERT(grpc_closure_list_empty(pollset->idle_jobs)); + pollset->vtable->finish_shutdown(pollset); + grpc_exec_ctx_enqueue(exec_ctx, pollset->shutdown_done, true, NULL); +} + +static void pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker_hdl, gpr_timespec now, + gpr_timespec deadline) { + grpc_pollset_worker worker; + *worker_hdl = &worker; + + /* pollset->mu already held */ + int added_worker = 0; + int locked = 1; + int queued_work = 0; + int keep_polling = 0; + GPR_TIMER_BEGIN("pollset_work", 0); + /* this must happen before we (potentially) drop pollset->mu */ + worker.next = worker.prev = NULL; + worker.reevaluate_polling_on_wakeup = 0; + if (pollset->local_wakeup_cache != NULL) { + worker.wakeup_fd = pollset->local_wakeup_cache; + pollset->local_wakeup_cache = worker.wakeup_fd->next; + } else { + worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd)); + grpc_wakeup_fd_init(&worker.wakeup_fd->fd); + } + worker.kicked_specifically = 0; + /* If there's work waiting for the pollset to be idle, and the + pollset is idle, then do that work */ + if (!pollset_has_workers(pollset) && + !grpc_closure_list_empty(pollset->idle_jobs)) { + GPR_TIMER_MARK("pollset_work.idle_jobs", 0); + grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); + goto done; + } + /* If we're shutting down then we don't execute any extended work */ + if (pollset->shutting_down) { + GPR_TIMER_MARK("pollset_work.shutting_down", 0); + goto done; + } + /* Give do_promote priority so we don't starve it out */ + if (pollset->in_flight_cbs) { + GPR_TIMER_MARK("pollset_work.in_flight_cbs", 0); + gpr_mu_unlock(&pollset->mu); + locked = 0; + goto done; + } + /* Start polling, and keep doing so while we're being asked to + re-evaluate our pollers (this allows poll() based pollers to + ensure they don't miss wakeups) */ + keep_polling = 1; + while (keep_polling) { + keep_polling = 0; + if (!pollset->kicked_without_pollers) { + if (!added_worker) { + push_front_worker(pollset, &worker); + added_worker = 1; + gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); + } + gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset); + GPR_TIMER_BEGIN("maybe_work_and_unlock", 0); + pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, &worker, + deadline, now); + GPR_TIMER_END("maybe_work_and_unlock", 0); + locked = 0; + gpr_tls_set(&g_current_thread_poller, 0); + } else { + GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0); + pollset->kicked_without_pollers = 0; + } + /* Finished execution - start cleaning up. + Note that we may arrive here from outside the enclosing while() loop. + In that case we won't loop though as we haven't added worker to the + worker list, which means nobody could ask us to re-evaluate polling). */ + done: + if (!locked) { + queued_work |= grpc_exec_ctx_flush(exec_ctx); + gpr_mu_lock(&pollset->mu); + locked = 1; + } + /* If we're forced to re-evaluate polling (via pollset_kick with + GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) then we land here and force + a loop */ + if (worker.reevaluate_polling_on_wakeup) { + worker.reevaluate_polling_on_wakeup = 0; + pollset->kicked_without_pollers = 0; + if (queued_work || worker.kicked_specifically) { + /* If there's queued work on the list, then set the deadline to be + immediate so we get back out of the polling loop quickly */ + deadline = gpr_inf_past(GPR_CLOCK_MONOTONIC); + } + keep_polling = 1; + } + } + if (added_worker) { + remove_worker(pollset, &worker); + gpr_tls_set(&g_current_thread_worker, 0); + } + /* release wakeup fd to the local pool */ + worker.wakeup_fd->next = pollset->local_wakeup_cache; + pollset->local_wakeup_cache = worker.wakeup_fd; + /* check shutdown conditions */ + if (pollset->shutting_down) { + if (pollset_has_workers(pollset)) { + pollset_kick(pollset, NULL); + } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) { + pollset->called_shutdown = 1; + gpr_mu_unlock(&pollset->mu); + finish_shutdown(exec_ctx, pollset); + grpc_exec_ctx_flush(exec_ctx); + /* Continuing to access pollset here is safe -- it is the caller's + * responsibility to not destroy when it has outstanding calls to + * pollset_work. + * TODO(dklempner): Can we refactor the shutdown logic to avoid this? */ + gpr_mu_lock(&pollset->mu); + } else if (!grpc_closure_list_empty(pollset->idle_jobs)) { + grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); + gpr_mu_unlock(&pollset->mu); + grpc_exec_ctx_flush(exec_ctx); + gpr_mu_lock(&pollset->mu); + } + } + *worker_hdl = NULL; + GPR_TIMER_END("pollset_work", 0); +} + +static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure) { + GPR_ASSERT(!pollset->shutting_down); + pollset->shutting_down = 1; + pollset->shutdown_done = closure; + pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); + if (!pollset_has_workers(pollset)) { + grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); + } + if (!pollset->called_shutdown && pollset->in_flight_cbs == 0 && + !pollset_has_workers(pollset)) { + pollset->called_shutdown = 1; + finish_shutdown(exec_ctx, pollset); + } +} + +static int poll_deadline_to_millis_timeout(gpr_timespec deadline, + gpr_timespec now) { + gpr_timespec timeout; + static const int64_t max_spin_polling_us = 10; + if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) { + return -1; + } + if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros( + max_spin_polling_us, + GPR_TIMESPAN))) <= 0) { + return 0; + } + timeout = gpr_time_sub(deadline, now); + return gpr_time_to_millis(gpr_time_add( + timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); +} + +/* + * basic_pollset - a vtable that provides polling for zero or one file + * descriptor via poll() + */ + +typedef struct grpc_unary_promote_args { + const grpc_pollset_vtable *original_vtable; + grpc_pollset *pollset; + grpc_fd *fd; + grpc_closure promotion_closure; +} grpc_unary_promote_args; + +static void basic_do_promote(grpc_exec_ctx *exec_ctx, void *args, + bool success) { + grpc_unary_promote_args *up_args = args; + const grpc_pollset_vtable *original_vtable = up_args->original_vtable; + grpc_pollset *pollset = up_args->pollset; + grpc_fd *fd = up_args->fd; + + /* + * This is quite tricky. There are a number of cases to keep in mind here: + * 1. fd may have been orphaned + * 2. The pollset may no longer be a unary poller (and we can't let case #1 + * leak to other pollset types!) + * 3. pollset's fd (which may have changed) may have been orphaned + * 4. The pollset may be shutting down. + */ + + gpr_mu_lock(&pollset->mu); + /* First we need to ensure that nobody is polling concurrently */ + GPR_ASSERT(!pollset_has_workers(pollset)); + + gpr_free(up_args); + /* At this point the pollset may no longer be a unary poller. In that case + * we should just call the right add function and be done. */ + /* TODO(klempner): If we're not careful this could cause infinite recursion. + * That's not a problem for now because empty_pollset has a trivial poller + * and we don't have any mechanism to unbecome multipoller. */ + pollset->in_flight_cbs--; + if (pollset->shutting_down) { + /* We don't care about this pollset anymore. */ + if (pollset->in_flight_cbs == 0 && !pollset->called_shutdown) { + pollset->called_shutdown = 1; + finish_shutdown(exec_ctx, pollset); + } + } else if (fd_is_orphaned(fd)) { + /* Don't try to add it to anything, we'll drop our ref on it below */ + } else if (pollset->vtable != original_vtable) { + pollset->vtable->add_fd(exec_ctx, pollset, fd, 0); + } else if (fd != pollset->data.ptr) { + grpc_fd *fds[2]; + fds[0] = pollset->data.ptr; + fds[1] = fd; + + if (fds[0] && !fd_is_orphaned(fds[0])) { + platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds)); + GRPC_FD_UNREF(fds[0], "basicpoll"); + } else { + /* old fd is orphaned and we haven't cleaned it up until now, so remain a + * unary poller */ + /* Note that it is possible that fds[1] is also orphaned at this point. + * That's okay, we'll correct it at the next add or poll. */ + if (fds[0]) GRPC_FD_UNREF(fds[0], "basicpoll"); + pollset->data.ptr = fd; + GRPC_FD_REF(fd, "basicpoll"); + } + } + + gpr_mu_unlock(&pollset->mu); + + /* Matching ref in basic_pollset_add_fd */ + GRPC_FD_UNREF(fd, "basicpoll_add"); +} + +static void basic_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd, int and_unlock_pollset) { + grpc_unary_promote_args *up_args; + GPR_ASSERT(fd); + if (fd == pollset->data.ptr) goto exit; + + if (!pollset_has_workers(pollset)) { + /* Fast path -- no in flight cbs */ + /* TODO(klempner): Comment this out and fix any test failures or establish + * they are due to timing issues */ + grpc_fd *fds[2]; + fds[0] = pollset->data.ptr; + fds[1] = fd; + + if (fds[0] == NULL) { + pollset->data.ptr = fd; + GRPC_FD_REF(fd, "basicpoll"); + } else if (!fd_is_orphaned(fds[0])) { + platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds)); + GRPC_FD_UNREF(fds[0], "basicpoll"); + } else { + /* old fd is orphaned and we haven't cleaned it up until now, so remain a + * unary poller */ + GRPC_FD_UNREF(fds[0], "basicpoll"); + pollset->data.ptr = fd; + GRPC_FD_REF(fd, "basicpoll"); + } + goto exit; + } + + /* Now we need to promote. This needs to happen when we're not polling. Since + * this may be called from poll, the wait needs to happen asynchronously. */ + GRPC_FD_REF(fd, "basicpoll_add"); + pollset->in_flight_cbs++; + up_args = gpr_malloc(sizeof(*up_args)); + up_args->fd = fd; + up_args->original_vtable = pollset->vtable; + up_args->pollset = pollset; + up_args->promotion_closure.cb = basic_do_promote; + up_args->promotion_closure.cb_arg = up_args; + + grpc_closure_list_add(&pollset->idle_jobs, &up_args->promotion_closure, 1); + pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); + +exit: + if (and_unlock_pollset) { + gpr_mu_unlock(&pollset->mu); + } +} + +static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + grpc_pollset_worker *worker, + gpr_timespec deadline, + gpr_timespec now) { +#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) +#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) + + struct pollfd pfd[3]; + grpc_fd *fd; + grpc_fd_watcher fd_watcher; + int timeout; + int r; + nfds_t nfds; + + fd = pollset->data.ptr; + if (fd && fd_is_orphaned(fd)) { + GRPC_FD_UNREF(fd, "basicpoll"); + fd = pollset->data.ptr = NULL; + } + timeout = poll_deadline_to_millis_timeout(deadline, now); + pfd[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); + pfd[0].events = POLLIN; + pfd[0].revents = 0; + pfd[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); + pfd[1].events = POLLIN; + pfd[1].revents = 0; + nfds = 2; + if (fd) { + pfd[2].fd = fd->fd; + pfd[2].revents = 0; + GRPC_FD_REF(fd, "basicpoll_begin"); + gpr_mu_unlock(&pollset->mu); + pfd[2].events = + (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, &fd_watcher); + if (pfd[2].events != 0) { + nfds++; + } + } else { + gpr_mu_unlock(&pollset->mu); + } + + /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid + even going into the blocking annotation if possible */ + /* poll fd count (argument 2) is shortened by one if we have no events + to poll on - such that it only includes the kicker */ + GPR_TIMER_BEGIN("poll", 0); + GRPC_SCHEDULING_START_BLOCKING_REGION; + r = grpc_poll_function(pfd, nfds, timeout); + GRPC_SCHEDULING_END_BLOCKING_REGION; + GPR_TIMER_END("poll", 0); + + if (r < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); + } + if (fd) { + fd_end_poll(exec_ctx, &fd_watcher, 0, 0); + } + } else if (r == 0) { + if (fd) { + fd_end_poll(exec_ctx, &fd_watcher, 0, 0); + } + } else { + if (pfd[0].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); + } + if (pfd[1].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); + } + if (nfds > 2) { + fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK, + pfd[2].revents & POLLOUT_CHECK); + } else if (fd) { + fd_end_poll(exec_ctx, &fd_watcher, 0, 0); + } + } + + if (fd) { + GRPC_FD_UNREF(fd, "basicpoll_begin"); + } +} + +static void basic_pollset_destroy(grpc_pollset *pollset) { + if (pollset->data.ptr != NULL) { + GRPC_FD_UNREF(pollset->data.ptr, "basicpoll"); + pollset->data.ptr = NULL; + } +} + +static const grpc_pollset_vtable basic_pollset = { + basic_pollset_add_fd, basic_pollset_maybe_work_and_unlock, + basic_pollset_destroy, basic_pollset_destroy}; + +static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null) { + pollset->vtable = &basic_pollset; + pollset->data.ptr = fd_or_null; + if (fd_or_null != NULL) { + GRPC_FD_REF(fd_or_null, "basicpoll"); + } +} + +/******************************************************************************* + * pollset_multipoller_with_poll_posix.c + */ + +#ifndef GPR_LINUX_MULTIPOLL_WITH_EPOLL + +typedef struct { + /* all polled fds */ + size_t fd_count; + size_t fd_capacity; + grpc_fd **fds; + /* fds that have been removed from the pollset explicitly */ + size_t del_count; + size_t del_capacity; + grpc_fd **dels; +} poll_hdr; + +static void multipoll_with_poll_pollset_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + grpc_fd *fd, + int and_unlock_pollset) { + size_t i; + poll_hdr *h = pollset->data.ptr; + /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */ + for (i = 0; i < h->fd_count; i++) { + if (h->fds[i] == fd) goto exit; + } + if (h->fd_count == h->fd_capacity) { + h->fd_capacity = GPR_MAX(h->fd_capacity + 8, h->fd_count * 3 / 2); + h->fds = gpr_realloc(h->fds, sizeof(grpc_fd *) * h->fd_capacity); + } + h->fds[h->fd_count++] = fd; + GRPC_FD_REF(fd, "multipoller"); +exit: + if (and_unlock_pollset) { + gpr_mu_unlock(&pollset->mu); + } +} + +static void multipoll_with_poll_pollset_maybe_work_and_unlock( + grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, + gpr_timespec deadline, gpr_timespec now) { +#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) +#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) + + int timeout; + int r; + size_t i, j, fd_count; + nfds_t pfd_count; + poll_hdr *h; + /* TODO(ctiller): inline some elements to avoid an allocation */ + grpc_fd_watcher *watchers; + struct pollfd *pfds; + + h = pollset->data.ptr; + timeout = poll_deadline_to_millis_timeout(deadline, now); + /* TODO(ctiller): perform just one malloc here if we exceed the inline case */ + pfds = gpr_malloc(sizeof(*pfds) * (h->fd_count + 2)); + watchers = gpr_malloc(sizeof(*watchers) * (h->fd_count + 2)); + fd_count = 0; + pfd_count = 2; + pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); + pfds[0].events = POLLIN; + pfds[0].revents = 0; + pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); + pfds[1].events = POLLIN; + pfds[1].revents = 0; + for (i = 0; i < h->fd_count; i++) { + int remove = fd_is_orphaned(h->fds[i]); + for (j = 0; !remove && j < h->del_count; j++) { + if (h->fds[i] == h->dels[j]) remove = 1; + } + if (remove) { + GRPC_FD_UNREF(h->fds[i], "multipoller"); + } else { + h->fds[fd_count++] = h->fds[i]; + watchers[pfd_count].fd = h->fds[i]; + GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start"); + pfds[pfd_count].fd = h->fds[i]->fd; + pfds[pfd_count].revents = 0; + pfd_count++; + } + } + for (j = 0; j < h->del_count; j++) { + GRPC_FD_UNREF(h->dels[j], "multipoller_del"); + } + h->del_count = 0; + h->fd_count = fd_count; + gpr_mu_unlock(&pollset->mu); + + for (i = 2; i < pfd_count; i++) { + grpc_fd *fd = watchers[i].fd; + pfds[i].events = (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, + &watchers[i]); + GRPC_FD_UNREF(fd, "multipoller_start"); + } + + /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid + even going into the blocking annotation if possible */ + GRPC_SCHEDULING_START_BLOCKING_REGION; + r = grpc_poll_function(pfds, pfd_count, timeout); + GRPC_SCHEDULING_END_BLOCKING_REGION; + + if (r < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); + } + for (i = 2; i < pfd_count; i++) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0); + } + } else if (r == 0) { + for (i = 2; i < pfd_count; i++) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0); + } + } else { + if (pfds[0].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); + } + if (pfds[1].revents & POLLIN_CHECK) { + grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); + } + for (i = 2; i < pfd_count; i++) { + if (watchers[i].fd == NULL) { + fd_end_poll(exec_ctx, &watchers[i], 0, 0); + continue; + } + fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK, + pfds[i].revents & POLLOUT_CHECK); + } + } + + gpr_free(pfds); + gpr_free(watchers); +} + +static void multipoll_with_poll_pollset_finish_shutdown(grpc_pollset *pollset) { + size_t i; + poll_hdr *h = pollset->data.ptr; + for (i = 0; i < h->fd_count; i++) { + GRPC_FD_UNREF(h->fds[i], "multipoller"); + } + for (i = 0; i < h->del_count; i++) { + GRPC_FD_UNREF(h->dels[i], "multipoller_del"); + } + h->fd_count = 0; + h->del_count = 0; +} + +static void multipoll_with_poll_pollset_destroy(grpc_pollset *pollset) { + poll_hdr *h = pollset->data.ptr; + multipoll_with_poll_pollset_finish_shutdown(pollset); + gpr_free(h->fds); + gpr_free(h->dels); + gpr_free(h); +} + +static const grpc_pollset_vtable multipoll_with_poll_pollset = { + multipoll_with_poll_pollset_add_fd, + multipoll_with_poll_pollset_maybe_work_and_unlock, + multipoll_with_poll_pollset_finish_shutdown, + multipoll_with_poll_pollset_destroy}; + +static void poll_become_multipoller(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, grpc_fd **fds, + size_t nfds) { + size_t i; + poll_hdr *h = gpr_malloc(sizeof(poll_hdr)); + pollset->vtable = &multipoll_with_poll_pollset; + pollset->data.ptr = h; + h->fd_count = nfds; + h->fd_capacity = nfds; + h->fds = gpr_malloc(nfds * sizeof(grpc_fd *)); + h->del_count = 0; + h->del_capacity = 0; + h->dels = NULL; + for (i = 0; i < nfds; i++) { + h->fds[i] = fds[i]; + GRPC_FD_REF(fds[i], "multipoller"); + } +} + +#endif /* !GPR_LINUX_MULTIPOLL_WITH_EPOLL */ + +/******************************************************************************* + * pollset_multipoller_with_epoll_posix.c + */ + +#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL + +#include <errno.h> +#include <poll.h> +#include <string.h> +#include <sys/epoll.h> +#include <unistd.h> + +#include <grpc/support/alloc.h> +#include <grpc/support/log.h> +#include <grpc/support/useful.h> + +#include "src/core/lib/iomgr/ev_posix.h" +#include "src/core/lib/profiling/timers.h" +#include "src/core/lib/support/block_annotate.h" + +static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure **st) { + /* only one set_ready can be active at once (but there may be a racing + notify_on) */ + gpr_mu_lock(&fd->mu); + set_ready_locked(exec_ctx, fd, st); + gpr_mu_unlock(&fd->mu); +} + +static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + set_ready(exec_ctx, fd, &fd->read_closure); +} + +static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + set_ready(exec_ctx, fd, &fd->write_closure); +} + +struct epoll_fd_list { + int *epoll_fds; + size_t count; + size_t capacity; +}; + +static struct epoll_fd_list epoll_fd_global_list; +static gpr_once init_epoll_fd_list_mu = GPR_ONCE_INIT; +static gpr_mu epoll_fd_list_mu; + +static void init_mu(void) { gpr_mu_init(&epoll_fd_list_mu); } + +static void add_epoll_fd_to_global_list(int epoll_fd) { + gpr_once_init(&init_epoll_fd_list_mu, init_mu); + + gpr_mu_lock(&epoll_fd_list_mu); + if (epoll_fd_global_list.count == epoll_fd_global_list.capacity) { + epoll_fd_global_list.capacity = + GPR_MAX((size_t)8, epoll_fd_global_list.capacity * 2); + epoll_fd_global_list.epoll_fds = + gpr_realloc(epoll_fd_global_list.epoll_fds, + epoll_fd_global_list.capacity * sizeof(int)); + } + epoll_fd_global_list.epoll_fds[epoll_fd_global_list.count++] = epoll_fd; + gpr_mu_unlock(&epoll_fd_list_mu); +} + +static void remove_epoll_fd_from_global_list(int epoll_fd) { + gpr_mu_lock(&epoll_fd_list_mu); + GPR_ASSERT(epoll_fd_global_list.count > 0); + for (size_t i = 0; i < epoll_fd_global_list.count; i++) { + if (epoll_fd == epoll_fd_global_list.epoll_fds[i]) { + epoll_fd_global_list.epoll_fds[i] = + epoll_fd_global_list.epoll_fds[--(epoll_fd_global_list.count)]; + break; + } + } + gpr_mu_unlock(&epoll_fd_list_mu); +} + +static void remove_fd_from_all_epoll_sets(int fd) { + int err; + gpr_once_init(&init_epoll_fd_list_mu, init_mu); + gpr_mu_lock(&epoll_fd_list_mu); + if (epoll_fd_global_list.count == 0) { + gpr_mu_unlock(&epoll_fd_list_mu); + return; + } + for (size_t i = 0; i < epoll_fd_global_list.count; i++) { + err = epoll_ctl(epoll_fd_global_list.epoll_fds[i], EPOLL_CTL_DEL, fd, NULL); + if (err < 0 && errno != ENOENT) { + gpr_log(GPR_ERROR, "epoll_ctl del for %d failed: %s", fd, + strerror(errno)); + } + } + gpr_mu_unlock(&epoll_fd_list_mu); +} + +typedef struct { + grpc_pollset *pollset; + grpc_fd *fd; + grpc_closure closure; +} delayed_add; + +typedef struct { int epoll_fd; } epoll_hdr; + +static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_fd *fd) { + epoll_hdr *h = pollset->data.ptr; + struct epoll_event ev; + int err; + grpc_fd_watcher watcher; + + /* We pretend to be polling whilst adding an fd to keep the fd from being + closed during the add. This may result in a spurious wakeup being assigned + to this pollset whilst adding, but that should be benign. */ + GPR_ASSERT(fd_begin_poll(fd, pollset, NULL, 0, 0, &watcher) == 0); + if (watcher.fd != NULL) { + ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); + ev.data.ptr = fd; + err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev); + if (err < 0) { + /* FDs may be added to a pollset multiple times, so EEXIST is normal. */ + if (errno != EEXIST) { + gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", fd->fd, + strerror(errno)); + } + } + } + fd_end_poll(exec_ctx, &watcher, 0, 0); +} + +static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg, + bool iomgr_status) { + delayed_add *da = arg; + + if (!fd_is_orphaned(da->fd)) { + finally_add_fd(exec_ctx, da->pollset, da->fd); + } + + gpr_mu_lock(&da->pollset->mu); + da->pollset->in_flight_cbs--; + if (da->pollset->shutting_down) { + /* We don't care about this pollset anymore. */ + if (da->pollset->in_flight_cbs == 0 && !da->pollset->called_shutdown) { + da->pollset->called_shutdown = 1; + grpc_exec_ctx_enqueue(exec_ctx, da->pollset->shutdown_done, true, NULL); + } + } + gpr_mu_unlock(&da->pollset->mu); + + GRPC_FD_UNREF(da->fd, "delayed_add"); + + gpr_free(da); +} + +static void multipoll_with_epoll_pollset_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, + grpc_fd *fd, + int and_unlock_pollset) { + if (and_unlock_pollset) { + gpr_mu_unlock(&pollset->mu); + finally_add_fd(exec_ctx, pollset, fd); + } else { + delayed_add *da = gpr_malloc(sizeof(*da)); + da->pollset = pollset; + da->fd = fd; + GRPC_FD_REF(fd, "delayed_add"); + grpc_closure_init(&da->closure, perform_delayed_add, da); + pollset->in_flight_cbs++; + grpc_exec_ctx_enqueue(exec_ctx, &da->closure, true, NULL); + } +} + +/* TODO(klempner): We probably want to turn this down a bit */ +#define GRPC_EPOLL_MAX_EVENTS 1000 + +static void multipoll_with_epoll_pollset_maybe_work_and_unlock( + grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, + gpr_timespec deadline, gpr_timespec now) { + struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; + int ep_rv; + int poll_rv; + epoll_hdr *h = pollset->data.ptr; + int timeout_ms; + struct pollfd pfds[2]; + + /* If you want to ignore epoll's ability to sanely handle parallel pollers, + * for a more apples-to-apples performance comparison with poll, add a + * if (pollset->counter != 0) { return 0; } + * here. + */ + + gpr_mu_unlock(&pollset->mu); + + timeout_ms = poll_deadline_to_millis_timeout(deadline, now); + + pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); + pfds[0].events = POLLIN; + pfds[0].revents = 0; + pfds[1].fd = h->epoll_fd; + pfds[1].events = POLLIN; + pfds[1].revents = 0; + + /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid + even going into the blocking annotation if possible */ + GPR_TIMER_BEGIN("poll", 0); + GRPC_SCHEDULING_START_BLOCKING_REGION; + poll_rv = grpc_poll_function(pfds, 2, timeout_ms); + GRPC_SCHEDULING_END_BLOCKING_REGION; + GPR_TIMER_END("poll", 0); + + if (poll_rv < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); + } + } else if (poll_rv == 0) { + /* do nothing */ + } else { + if (pfds[0].revents) { + grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); + } + if (pfds[1].revents) { + do { + /* The following epoll_wait never blocks; it has a timeout of 0 */ + ep_rv = epoll_wait(h->epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); + if (ep_rv < 0) { + if (errno != EINTR) { + gpr_log(GPR_ERROR, "epoll_wait() failed: %s", strerror(errno)); + } + } else { + int i; + for (i = 0; i < ep_rv; ++i) { + grpc_fd *fd = ep_ev[i].data.ptr; + /* TODO(klempner): We might want to consider making err and pri + * separate events */ + int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP); + int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI); + int write_ev = ep_ev[i].events & EPOLLOUT; + if (fd == NULL) { + grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); + } else { + if (read_ev || cancel) { + fd_become_readable(exec_ctx, fd); + } + if (write_ev || cancel) { + fd_become_writable(exec_ctx, fd); + } + } + } + } + } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); + } + } +} + +static void multipoll_with_epoll_pollset_finish_shutdown( + grpc_pollset *pollset) {} + +static void multipoll_with_epoll_pollset_destroy(grpc_pollset *pollset) { + epoll_hdr *h = pollset->data.ptr; + close(h->epoll_fd); + remove_epoll_fd_from_global_list(h->epoll_fd); + gpr_free(h); +} + +static const grpc_pollset_vtable multipoll_with_epoll_pollset = { + multipoll_with_epoll_pollset_add_fd, + multipoll_with_epoll_pollset_maybe_work_and_unlock, + multipoll_with_epoll_pollset_finish_shutdown, + multipoll_with_epoll_pollset_destroy}; + +static void epoll_become_multipoller(grpc_exec_ctx *exec_ctx, + grpc_pollset *pollset, grpc_fd **fds, + size_t nfds) { + size_t i; + epoll_hdr *h = gpr_malloc(sizeof(epoll_hdr)); + struct epoll_event ev; + int err; + + pollset->vtable = &multipoll_with_epoll_pollset; + pollset->data.ptr = h; + h->epoll_fd = epoll_create1(EPOLL_CLOEXEC); + if (h->epoll_fd < 0) { + /* TODO(klempner): Fall back to poll here, especially on ENOSYS */ + gpr_log(GPR_ERROR, "epoll_create1 failed: %s", strerror(errno)); + abort(); + } + add_epoll_fd_to_global_list(h->epoll_fd); + + ev.events = (uint32_t)(EPOLLIN | EPOLLET); + ev.data.ptr = NULL; + err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, + GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), &ev); + if (err < 0) { + gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", + GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), + strerror(errno)); + } + + for (i = 0; i < nfds; i++) { + multipoll_with_epoll_pollset_add_fd(exec_ctx, pollset, fds[i], 0); + } +} + +#else /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ + +static void remove_fd_from_all_epoll_sets(int fd) {} + +#endif /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ + +/******************************************************************************* + * pollset_set_posix.c + */ + +static grpc_pollset_set *pollset_set_create(void) { + grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set)); + memset(pollset_set, 0, sizeof(*pollset_set)); + gpr_mu_init(&pollset_set->mu); + return pollset_set; +} + +static void pollset_set_destroy(grpc_pollset_set *pollset_set) { + size_t i; + gpr_mu_destroy(&pollset_set->mu); + for (i = 0; i < pollset_set->fd_count; i++) { + GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); + } + gpr_free(pollset_set->pollsets); + gpr_free(pollset_set->pollset_sets); + gpr_free(pollset_set->fds); + gpr_free(pollset_set); +} + +static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + size_t i, j; + gpr_mu_lock(&pollset_set->mu); + if (pollset_set->pollset_count == pollset_set->pollset_capacity) { + pollset_set->pollset_capacity = + GPR_MAX(8, 2 * pollset_set->pollset_capacity); + pollset_set->pollsets = + gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity * + sizeof(*pollset_set->pollsets)); + } + pollset_set->pollsets[pollset_set->pollset_count++] = pollset; + for (i = 0, j = 0; i < pollset_set->fd_count; i++) { + if (fd_is_orphaned(pollset_set->fds[i])) { + GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); + } else { + pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]); + pollset_set->fds[j++] = pollset_set->fds[i]; + } + } + pollset_set->fd_count = j; + gpr_mu_unlock(&pollset_set->mu); +} + +static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + size_t i; + gpr_mu_lock(&pollset_set->mu); + for (i = 0; i < pollset_set->pollset_count; i++) { + if (pollset_set->pollsets[i] == pollset) { + pollset_set->pollset_count--; + GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i], + pollset_set->pollsets[pollset_set->pollset_count]); + break; + } + } + gpr_mu_unlock(&pollset_set->mu); +} + +static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + size_t i, j; + gpr_mu_lock(&bag->mu); + if (bag->pollset_set_count == bag->pollset_set_capacity) { + bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity); + bag->pollset_sets = + gpr_realloc(bag->pollset_sets, + bag->pollset_set_capacity * sizeof(*bag->pollset_sets)); + } + bag->pollset_sets[bag->pollset_set_count++] = item; + for (i = 0, j = 0; i < bag->fd_count; i++) { + if (fd_is_orphaned(bag->fds[i])) { + GRPC_FD_UNREF(bag->fds[i], "pollset_set"); + } else { + pollset_set_add_fd(exec_ctx, item, bag->fds[i]); + bag->fds[j++] = bag->fds[i]; + } + } + bag->fd_count = j; + gpr_mu_unlock(&bag->mu); +} + +static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + size_t i; + gpr_mu_lock(&bag->mu); + for (i = 0; i < bag->pollset_set_count; i++) { + if (bag->pollset_sets[i] == item) { + bag->pollset_set_count--; + GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i], + bag->pollset_sets[bag->pollset_set_count]); + break; + } + } + gpr_mu_unlock(&bag->mu); +} + +static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + size_t i; + gpr_mu_lock(&pollset_set->mu); + if (pollset_set->fd_count == pollset_set->fd_capacity) { + pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity); + pollset_set->fds = gpr_realloc( + pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds)); + } + GRPC_FD_REF(fd, "pollset_set"); + pollset_set->fds[pollset_set->fd_count++] = fd; + for (i = 0; i < pollset_set->pollset_count; i++) { + pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd); + } + for (i = 0; i < pollset_set->pollset_set_count; i++) { + pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd); + } + gpr_mu_unlock(&pollset_set->mu); +} + +static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + size_t i; + gpr_mu_lock(&pollset_set->mu); + for (i = 0; i < pollset_set->fd_count; i++) { + if (pollset_set->fds[i] == fd) { + pollset_set->fd_count--; + GPR_SWAP(grpc_fd *, pollset_set->fds[i], + pollset_set->fds[pollset_set->fd_count]); + GRPC_FD_UNREF(fd, "pollset_set"); + break; + } + } + for (i = 0; i < pollset_set->pollset_set_count; i++) { + pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd); + } + gpr_mu_unlock(&pollset_set->mu); +} + +/******************************************************************************* + * event engine binding + */ + +static void shutdown_engine(void) { + fd_global_shutdown(); + pollset_global_shutdown(); +} + +static const grpc_event_engine_vtable vtable = { + .pollset_size = sizeof(grpc_pollset), + + .fd_create = fd_create, + .fd_wrapped_fd = fd_wrapped_fd, + .fd_orphan = fd_orphan, + .fd_shutdown = fd_shutdown, + .fd_notify_on_read = fd_notify_on_read, + .fd_notify_on_write = fd_notify_on_write, + + .pollset_init = pollset_init, + .pollset_shutdown = pollset_shutdown, + .pollset_reset = pollset_reset, + .pollset_destroy = pollset_destroy, + .pollset_work = pollset_work, + .pollset_kick = pollset_kick, + .pollset_add_fd = pollset_add_fd, + + .pollset_set_create = pollset_set_create, + .pollset_set_destroy = pollset_set_destroy, + .pollset_set_add_pollset = pollset_set_add_pollset, + .pollset_set_del_pollset = pollset_set_del_pollset, + .pollset_set_add_pollset_set = pollset_set_add_pollset_set, + .pollset_set_del_pollset_set = pollset_set_del_pollset_set, + .pollset_set_add_fd = pollset_set_add_fd, + .pollset_set_del_fd = pollset_set_del_fd, + + .kick_poller = kick_poller, + + .shutdown_engine = shutdown_engine, +}; + +const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void) { +#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL + platform_become_multipoller = epoll_become_multipoller; +#else + platform_become_multipoller = poll_become_multipoller; +#endif + fd_global_init(); + pollset_global_init(); + return &vtable; +} + +#endif diff --git a/src/core/lib/census/grpc_plugin.h b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h index 4d642a65c7..06d6dbf29d 100644 --- a/src/core/lib/census/grpc_plugin.h +++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h @@ -31,10 +31,11 @@ * */ -#ifndef GRPC_CORE_LIB_CENSUS_GRPC_PLUGIN_H -#define GRPC_CORE_LIB_CENSUS_GRPC_PLUGIN_H +#ifndef GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H +#define GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H -void census_grpc_plugin_init(void); -void census_grpc_plugin_destroy(void); +#include "src/core/lib/iomgr/ev_posix.h" -#endif /* GRPC_CORE_LIB_CENSUS_GRPC_PLUGIN_H */ +const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void); + +#endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H */ diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c new file mode 100644 index 0000000000..235a7df08d --- /dev/null +++ b/src/core/lib/iomgr/ev_posix.c @@ -0,0 +1,165 @@ +/* + * + * Copyright 2015-2016, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#include <grpc/support/port_platform.h> + +#ifdef GPR_POSIX_SOCKET + +#include "src/core/lib/iomgr/ev_posix.h" + +#include <grpc/support/log.h> + +#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h" + +static const grpc_event_engine_vtable *g_event_engine; + +grpc_poll_function_type grpc_poll_function = poll; +grpc_wakeup_fd grpc_global_wakeup_fd; + +void grpc_event_engine_init(void) { + if ((g_event_engine = grpc_init_poll_and_epoll_posix())) { + return; + } + gpr_log(GPR_ERROR, "No event engine could be initialized"); + abort(); +} + +void grpc_event_engine_shutdown(void) { g_event_engine->shutdown_engine(); } + +grpc_fd *grpc_fd_create(int fd, const char *name) { + return g_event_engine->fd_create(fd, name); +} + +int grpc_fd_wrapped_fd(grpc_fd *fd) { + return g_event_engine->fd_wrapped_fd(fd); +} + +void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, + int *release_fd, const char *reason) { + g_event_engine->fd_orphan(exec_ctx, fd, on_done, release_fd, reason); +} + +void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { + g_event_engine->fd_shutdown(exec_ctx, fd); +} + +void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + g_event_engine->fd_notify_on_read(exec_ctx, fd, closure); +} + +void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure) { + g_event_engine->fd_notify_on_write(exec_ctx, fd, closure); +} + +size_t grpc_pollset_size(void) { return g_event_engine->pollset_size; } + +void grpc_pollset_init(grpc_pollset *pollset, gpr_mu **mu) { + g_event_engine->pollset_init(pollset, mu); +} + +void grpc_pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure) { + g_event_engine->pollset_shutdown(exec_ctx, pollset, closure); +} + +void grpc_pollset_reset(grpc_pollset *pollset) { + g_event_engine->pollset_reset(pollset); +} + +void grpc_pollset_destroy(grpc_pollset *pollset) { + g_event_engine->pollset_destroy(pollset); +} + +void grpc_pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker, gpr_timespec now, + gpr_timespec deadline) { + g_event_engine->pollset_work(exec_ctx, pollset, worker, now, deadline); +} + +void grpc_pollset_kick(grpc_pollset *pollset, + grpc_pollset_worker *specific_worker) { + g_event_engine->pollset_kick(pollset, specific_worker); +} + +void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd) { + g_event_engine->pollset_add_fd(exec_ctx, pollset, fd); +} + +grpc_pollset_set *grpc_pollset_set_create(void) { + return g_event_engine->pollset_set_create(); +} + +void grpc_pollset_set_destroy(grpc_pollset_set *pollset_set) { + g_event_engine->pollset_set_destroy(pollset_set); +} + +void grpc_pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + g_event_engine->pollset_set_add_pollset(exec_ctx, pollset_set, pollset); +} + +void grpc_pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset) { + g_event_engine->pollset_set_del_pollset(exec_ctx, pollset_set, pollset); +} + +void grpc_pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + g_event_engine->pollset_set_add_pollset_set(exec_ctx, bag, item); +} + +void grpc_pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item) { + g_event_engine->pollset_set_del_pollset_set(exec_ctx, bag, item); +} + +void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + g_event_engine->pollset_set_add_fd(exec_ctx, pollset_set, fd); +} + +void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd) { + g_event_engine->pollset_set_del_fd(exec_ctx, pollset_set, fd); +} + +void grpc_kick_poller(void) { g_event_engine->kick_poller(); } + +#endif // GPR_POSIX_SOCKET diff --git a/src/core/lib/iomgr/ev_posix.h b/src/core/lib/iomgr/ev_posix.h new file mode 100644 index 0000000000..9d27b2bcda --- /dev/null +++ b/src/core/lib/iomgr/ev_posix.h @@ -0,0 +1,158 @@ +/* + * + * Copyright 2015-2016, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + */ + +#ifndef GRPC_CORE_LIB_IOMGR_EV_POSIX_H +#define GRPC_CORE_LIB_IOMGR_EV_POSIX_H + +#include <poll.h> + +#include "src/core/lib/iomgr/exec_ctx.h" +#include "src/core/lib/iomgr/pollset.h" +#include "src/core/lib/iomgr/pollset_set.h" +#include "src/core/lib/iomgr/wakeup_fd_posix.h" + +typedef struct grpc_fd grpc_fd; + +typedef struct grpc_event_engine_vtable { + size_t pollset_size; + + grpc_fd *(*fd_create)(int fd, const char *name); + int (*fd_wrapped_fd)(grpc_fd *fd); + void (*fd_orphan)(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, + int *release_fd, const char *reason); + void (*fd_shutdown)(grpc_exec_ctx *exec_ctx, grpc_fd *fd); + void (*fd_notify_on_read)(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + void (*fd_notify_on_write)(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + + void (*pollset_init)(grpc_pollset *pollset, gpr_mu **mu); + void (*pollset_shutdown)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_closure *closure); + void (*pollset_reset)(grpc_pollset *pollset); + void (*pollset_destroy)(grpc_pollset *pollset); + void (*pollset_work)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + grpc_pollset_worker **worker, gpr_timespec now, + gpr_timespec deadline); + void (*pollset_kick)(grpc_pollset *pollset, + grpc_pollset_worker *specific_worker); + void (*pollset_add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd); + + grpc_pollset_set *(*pollset_set_create)(void); + void (*pollset_set_destroy)(grpc_pollset_set *pollset_set); + void (*pollset_set_add_pollset)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset); + void (*pollset_set_del_pollset)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, + grpc_pollset *pollset); + void (*pollset_set_add_pollset_set)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item); + void (*pollset_set_del_pollset_set)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *bag, + grpc_pollset_set *item); + void (*pollset_set_add_fd)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + void (*pollset_set_del_fd)(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + + void (*kick_poller)(void); + + void (*shutdown_engine)(void); +} grpc_event_engine_vtable; + +void grpc_event_engine_init(void); +void grpc_event_engine_shutdown(void); + +/* Create a wrapped file descriptor. + Requires fd is a non-blocking file descriptor. + This takes ownership of closing fd. */ +grpc_fd *grpc_fd_create(int fd, const char *name); + +/* Return the wrapped fd, or -1 if it has been released or closed. */ +int grpc_fd_wrapped_fd(grpc_fd *fd); + +/* Releases fd to be asynchronously destroyed. + on_done is called when the underlying file descriptor is definitely close()d. + If on_done is NULL, no callback will be made. + If release_fd is not NULL, it's set to fd and fd will not be closed. + Requires: *fd initialized; no outstanding notify_on_read or + notify_on_write. + MUST NOT be called with a pollset lock taken */ +void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, + int *release_fd, const char *reason); + +/* Cause any current callbacks to error out with GRPC_CALLBACK_CANCELLED. */ +void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd); + +/* Register read interest, causing read_cb to be called once when fd becomes + readable, on deadline specified by deadline, or on shutdown triggered by + grpc_fd_shutdown. + read_cb will be called with read_cb_arg when *fd becomes readable. + read_cb is Called with status of GRPC_CALLBACK_SUCCESS if readable, + GRPC_CALLBACK_TIMED_OUT if the call timed out, + and CANCELLED if the call was cancelled. + + Requires:This method must not be called before the read_cb for any previous + call runs. Edge triggered events are used whenever they are supported by the + underlying platform. This means that users must drain fd in read_cb before + calling notify_on_read again. Users are also expected to handle spurious + events, i.e read_cb is called while nothing can be readable from fd */ +void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + +/* Exactly the same semantics as above, except based on writable events. */ +void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, + grpc_closure *closure); + +/* pollset_posix functions */ + +/* Add an fd to a pollset */ +void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, + struct grpc_fd *fd); + +/* pollset_set_posix functions */ + +void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); +void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, + grpc_pollset_set *pollset_set, grpc_fd *fd); + +/* override to allow tests to hook poll() usage */ +typedef int (*grpc_poll_function_type)(struct pollfd *, nfds_t, int); +extern grpc_poll_function_type grpc_poll_function; +extern grpc_wakeup_fd grpc_global_wakeup_fd; + +#endif /* GRPC_CORE_LIB_IOMGR_EV_POSIX_H */ diff --git a/src/core/lib/iomgr/fd_posix.c b/src/core/lib/iomgr/fd_posix.c deleted file mode 100644 index 1397302658..0000000000 --- a/src/core/lib/iomgr/fd_posix.c +++ /dev/null @@ -1,454 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include "src/core/lib/iomgr/fd_posix.h" - -#include <assert.h> -#include <sys/socket.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/string_util.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/pollset_posix.h" - -#define CLOSURE_NOT_READY ((grpc_closure *)0) -#define CLOSURE_READY ((grpc_closure *)1) - -/* 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. - */ -/* TODO(klempner): We could use some form of polling generation count to know - * when these are safe to free. */ -/* TODO(klempner): Consider disabling freelisting if we don't have multiple - * threads in poll on the same fd */ -/* TODO(klempner): Batch these allocations to reduce fragmentation */ -static grpc_fd *fd_freelist = NULL; -static gpr_mu fd_freelist_mu; - -static void freelist_fd(grpc_fd *fd) { - gpr_mu_lock(&fd_freelist_mu); - fd->freelist_next = fd_freelist; - fd_freelist = fd; - grpc_iomgr_unregister_object(&fd->iomgr_object); - gpr_mu_unlock(&fd_freelist_mu); -} - -static grpc_fd *alloc_fd(int fd) { - grpc_fd *r = NULL; - gpr_mu_lock(&fd_freelist_mu); - if (fd_freelist != NULL) { - r = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - } - gpr_mu_unlock(&fd_freelist_mu); - if (r == NULL) { - r = gpr_malloc(sizeof(grpc_fd)); - gpr_mu_init(&r->mu); - } - - gpr_mu_lock(&r->mu); - r->shutdown = 0; - r->read_closure = CLOSURE_NOT_READY; - r->write_closure = CLOSURE_NOT_READY; - r->fd = fd; - r->inactive_watcher_root.next = r->inactive_watcher_root.prev = - &r->inactive_watcher_root; - r->freelist_next = NULL; - r->read_watcher = r->write_watcher = NULL; - r->on_done_closure = NULL; - r->closed = 0; - r->released = 0; - gpr_atm_rel_store(&r->refst, 1); - gpr_mu_unlock(&r->mu); - - return r; -} - -static void destroy(grpc_fd *fd) { - gpr_mu_destroy(&fd->mu); - gpr_free(fd); -} - -#ifdef GRPC_FD_REF_COUNT_DEBUG -#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) { - gpr_log(GPR_DEBUG, "FD %d %p ref %d %d -> %d [%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); -} - -#ifdef GRPC_FD_REF_COUNT_DEBUG -static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file, - int line) { - gpr_atm old; - gpr_log(GPR_DEBUG, "FD %d %p unref %d %d -> %d [%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) { - gpr_atm old; -#endif - old = gpr_atm_full_fetch_add(&fd->refst, -n); - if (old == n) { - freelist_fd(fd); - } else { - GPR_ASSERT(old > n); - } -} - -void grpc_fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); } - -void grpc_fd_global_shutdown(void) { - gpr_mu_lock(&fd_freelist_mu); - gpr_mu_unlock(&fd_freelist_mu); - while (fd_freelist != NULL) { - grpc_fd *fd = fd_freelist; - fd_freelist = fd_freelist->freelist_next; - destroy(fd); - } - gpr_mu_destroy(&fd_freelist_mu); -} - -grpc_fd *grpc_fd_create(int fd, const char *name) { - grpc_fd *r = alloc_fd(fd); - char *name2; - gpr_asprintf(&name2, "%s fd=%d", name, fd); - grpc_iomgr_register_object(&r->iomgr_object, name2); - gpr_free(name2); -#ifdef GRPC_FD_REF_COUNT_DEBUG - gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, r, name); -#endif - return r; -} - -int grpc_fd_is_orphaned(grpc_fd *fd) { - return (gpr_atm_acq_load(&fd->refst) & 1) == 0; -} - -static void pollset_kick_locked(grpc_fd_watcher *watcher) { - gpr_mu_lock(&watcher->pollset->mu); - GPR_ASSERT(watcher->worker); - grpc_pollset_kick_ext(watcher->pollset, watcher->worker, - GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP); - gpr_mu_unlock(&watcher->pollset->mu); -} - -static void maybe_wake_one_watcher_locked(grpc_fd *fd) { - if (fd->inactive_watcher_root.next != &fd->inactive_watcher_root) { - pollset_kick_locked(fd->inactive_watcher_root.next); - } else if (fd->read_watcher) { - pollset_kick_locked(fd->read_watcher); - } else if (fd->write_watcher) { - pollset_kick_locked(fd->write_watcher); - } -} - -static void wake_all_watchers_locked(grpc_fd *fd) { - grpc_fd_watcher *watcher; - for (watcher = fd->inactive_watcher_root.next; - watcher != &fd->inactive_watcher_root; watcher = watcher->next) { - pollset_kick_locked(watcher); - } - if (fd->read_watcher) { - pollset_kick_locked(fd->read_watcher); - } - if (fd->write_watcher && fd->write_watcher != fd->read_watcher) { - pollset_kick_locked(fd->write_watcher); - } -} - -static int has_watchers(grpc_fd *fd) { - return fd->read_watcher != NULL || fd->write_watcher != NULL || - fd->inactive_watcher_root.next != &fd->inactive_watcher_root; -} - -static void close_fd_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - fd->closed = 1; - if (!fd->released) { - close(fd->fd); - } else { - grpc_remove_fd_from_all_epoll_sets(fd->fd); - } - grpc_exec_ctx_enqueue(exec_ctx, fd->on_done_closure, true, NULL); -} - -int grpc_fd_wrapped_fd(grpc_fd *fd) { - if (fd->released || fd->closed) { - return -1; - } else { - return fd->fd; - } -} - -void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, - int *release_fd, const char *reason) { - fd->on_done_closure = on_done; - fd->released = release_fd != NULL; - if (!fd->released) { - shutdown(fd->fd, SHUT_RDWR); - } else { - *release_fd = fd->fd; - } - gpr_mu_lock(&fd->mu); - REF_BY(fd, 1, reason); /* remove active status, but keep referenced */ - if (!has_watchers(fd)) { - close_fd_locked(exec_ctx, fd); - } else { - wake_all_watchers_locked(fd); - } - gpr_mu_unlock(&fd->mu); - UNREF_BY(fd, 2, reason); /* drop the reference */ -} - -/* increment refcount by two to avoid changing the orphan bit */ -#ifdef GRPC_FD_REF_COUNT_DEBUG -void grpc_fd_ref(grpc_fd *fd, const char *reason, const char *file, int line) { - ref_by(fd, 2, reason, file, line); -} - -void grpc_fd_unref(grpc_fd *fd, const char *reason, const char *file, - int line) { - unref_by(fd, 2, reason, file, line); -} -#else -void grpc_fd_ref(grpc_fd *fd) { ref_by(fd, 2); } - -void grpc_fd_unref(grpc_fd *fd) { unref_by(fd, 2); } -#endif - -static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure **st, grpc_closure *closure) { - if (*st == CLOSURE_NOT_READY) { - /* not ready ==> switch to a waiting state by setting the closure */ - *st = closure; - } else if (*st == CLOSURE_READY) { - /* already ready ==> queue the closure to run immediately */ - *st = CLOSURE_NOT_READY; - grpc_exec_ctx_enqueue(exec_ctx, closure, !fd->shutdown, NULL); - maybe_wake_one_watcher_locked(fd); - } else { - /* upcallptr was set to a different closure. This is an error! */ - gpr_log(GPR_ERROR, - "User called a notify_on function with a previous callback still " - "pending"); - abort(); - } -} - -/* returns 1 if state becomes not ready */ -static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure **st) { - if (*st == CLOSURE_READY) { - /* duplicate ready ==> ignore */ - return 0; - } else if (*st == CLOSURE_NOT_READY) { - /* not ready, and not waiting ==> flag ready */ - *st = CLOSURE_READY; - return 0; - } else { - /* waiting ==> queue closure */ - grpc_exec_ctx_enqueue(exec_ctx, *st, !fd->shutdown, NULL); - *st = CLOSURE_NOT_READY; - return 1; - } -} - -static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure **st) { - /* only one set_ready can be active at once (but there may be a racing - notify_on) */ - gpr_mu_lock(&fd->mu); - set_ready_locked(exec_ctx, fd, st); - gpr_mu_unlock(&fd->mu); -} - -void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - gpr_mu_lock(&fd->mu); - GPR_ASSERT(!fd->shutdown); - fd->shutdown = 1; - set_ready_locked(exec_ctx, fd, &fd->read_closure); - set_ready_locked(exec_ctx, fd, &fd->write_closure); - gpr_mu_unlock(&fd->mu); -} - -void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure) { - gpr_mu_lock(&fd->mu); - notify_on_locked(exec_ctx, fd, &fd->read_closure, closure); - gpr_mu_unlock(&fd->mu); -} - -void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure) { - gpr_mu_lock(&fd->mu); - notify_on_locked(exec_ctx, fd, &fd->write_closure, closure); - gpr_mu_unlock(&fd->mu); -} - -uint32_t grpc_fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, - grpc_pollset_worker *worker, uint32_t read_mask, - uint32_t write_mask, grpc_fd_watcher *watcher) { - uint32_t mask = 0; - grpc_closure *cur; - int requested; - /* keep track of pollers that have requested our events, in case they change - */ - GRPC_FD_REF(fd, "poll"); - - gpr_mu_lock(&fd->mu); - - /* if we are shutdown, then don't add to the watcher set */ - if (fd->shutdown) { - watcher->fd = NULL; - watcher->pollset = NULL; - watcher->worker = NULL; - gpr_mu_unlock(&fd->mu); - GRPC_FD_UNREF(fd, "poll"); - return 0; - } - - /* if there is nobody polling for read, but we need to, then start doing so */ - cur = fd->read_closure; - requested = cur != CLOSURE_READY; - if (read_mask && fd->read_watcher == NULL && requested) { - fd->read_watcher = watcher; - mask |= read_mask; - } - /* if there is nobody polling for write, but we need to, then start doing so - */ - cur = fd->write_closure; - requested = cur != CLOSURE_READY; - if (write_mask && fd->write_watcher == NULL && requested) { - fd->write_watcher = watcher; - mask |= write_mask; - } - /* if not polling, remember this watcher in case we need someone to later */ - if (mask == 0 && worker != NULL) { - watcher->next = &fd->inactive_watcher_root; - watcher->prev = watcher->next->prev; - watcher->next->prev = watcher->prev->next = watcher; - } - watcher->pollset = pollset; - watcher->worker = worker; - watcher->fd = fd; - gpr_mu_unlock(&fd->mu); - - return mask; -} - -void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher, - int got_read, int got_write) { - int was_polling = 0; - int kick = 0; - grpc_fd *fd = watcher->fd; - - if (fd == NULL) { - return; - } - - gpr_mu_lock(&fd->mu); - - if (watcher == fd->read_watcher) { - /* remove read watcher, kick if we still need a read */ - was_polling = 1; - if (!got_read) { - kick = 1; - } - fd->read_watcher = NULL; - } - if (watcher == fd->write_watcher) { - /* remove write watcher, kick if we still need a write */ - was_polling = 1; - if (!got_write) { - kick = 1; - } - fd->write_watcher = NULL; - } - if (!was_polling && watcher->worker != NULL) { - /* remove from inactive list */ - watcher->next->prev = watcher->prev; - watcher->prev->next = watcher->next; - } - if (got_read) { - if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) { - kick = 1; - } - } - if (got_write) { - if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) { - kick = 1; - } - } - if (kick) { - maybe_wake_one_watcher_locked(fd); - } - if (grpc_fd_is_orphaned(fd) && !has_watchers(fd) && !fd->closed) { - close_fd_locked(exec_ctx, fd); - } - gpr_mu_unlock(&fd->mu); - - GRPC_FD_UNREF(fd, "poll"); -} - -void grpc_fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - set_ready(exec_ctx, fd, &fd->read_closure); -} - -void grpc_fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) { - set_ready(exec_ctx, fd, &fd->write_closure); -} - -#endif diff --git a/src/core/lib/iomgr/fd_posix.h b/src/core/lib/iomgr/fd_posix.h deleted file mode 100644 index dc389ba743..0000000000 --- a/src/core/lib/iomgr/fd_posix.h +++ /dev/null @@ -1,192 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_FD_POSIX_H -#define GRPC_CORE_LIB_IOMGR_FD_POSIX_H - -#include <grpc/support/atm.h> -#include <grpc/support/sync.h> -#include <grpc/support/time.h> -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/pollset.h" - -typedef struct grpc_fd grpc_fd; - -typedef struct grpc_fd_watcher { - struct grpc_fd_watcher *next; - struct grpc_fd_watcher *prev; - grpc_pollset *pollset; - grpc_pollset_worker *worker; - grpc_fd *fd; -} grpc_fd_watcher; - -struct grpc_fd { - int fd; - /* refst format: - bit0: 1=active/0=orphaned - bit1-n: refcount - meaning that mostly we ref by two to avoid altering the orphaned bit, - and just unref by 1 when we're ready to flag the object as orphaned */ - gpr_atm refst; - - gpr_mu mu; - int shutdown; - int closed; - int released; - - /* The watcher list. - - The following watcher related fields are protected by watcher_mu. - - An fd_watcher is an ephemeral object created when an fd wants to - begin polling, and destroyed after the poll. - - It denotes the fd's interest in whether to read poll or write poll - or both or neither on this fd. - - If a watcher is asked to poll for reads or writes, the read_watcher - or write_watcher fields are set respectively. A watcher may be asked - to poll for both, in which case both fields will be set. - - read_watcher and write_watcher may be NULL if no watcher has been - asked to poll for reads or writes. - - If an fd_watcher is not asked to poll for reads or writes, it's added - to a linked list of inactive watchers, rooted at inactive_watcher_root. - If at a later time there becomes need of a poller to poll, one of - the inactive pollers may be kicked out of their poll loops to take - that responsibility. */ - grpc_fd_watcher inactive_watcher_root; - grpc_fd_watcher *read_watcher; - grpc_fd_watcher *write_watcher; - - grpc_closure *read_closure; - grpc_closure *write_closure; - - struct grpc_fd *freelist_next; - - grpc_closure *on_done_closure; - - grpc_iomgr_object iomgr_object; -}; - -/* Create a wrapped file descriptor. - Requires fd is a non-blocking file descriptor. - This takes ownership of closing fd. */ -grpc_fd *grpc_fd_create(int fd, const char *name); - -/* Return the wrapped fd, or -1 if it has been released or closed. */ -int grpc_fd_wrapped_fd(grpc_fd *fd); - -/* Releases fd to be asynchronously destroyed. - on_done is called when the underlying file descriptor is definitely close()d. - If on_done is NULL, no callback will be made. - If release_fd is not NULL, it's set to fd and fd will not be closed. - Requires: *fd initialized; no outstanding notify_on_read or - notify_on_write. - MUST NOT be called with a pollset lock taken */ -void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done, - int *release_fd, const char *reason); - -/* Begin polling on an fd. - Registers that the given pollset is interested in this fd - so that if read - or writability interest changes, the pollset can be kicked to pick up that - new interest. - Return value is: - (fd_needs_read? read_mask : 0) | (fd_needs_write? write_mask : 0) - i.e. a combination of read_mask and write_mask determined by the fd's current - interest in said events. - Polling strategies that do not need to alter their behavior depending on the - fd's current interest (such as epoll) do not need to call this function. - MUST NOT be called with a pollset lock taken */ -uint32_t grpc_fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset, - grpc_pollset_worker *worker, uint32_t read_mask, - uint32_t write_mask, grpc_fd_watcher *rec); -/* Complete polling previously started with grpc_fd_begin_poll - MUST NOT be called with a pollset lock taken - if got_read or got_write are 1, also does the become_{readable,writable} as - appropriate. */ -void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec, - int got_read, int got_write); - -/* Return 1 if this fd is orphaned, 0 otherwise */ -int grpc_fd_is_orphaned(grpc_fd *fd); - -/* Cause any current callbacks to error out with GRPC_CALLBACK_CANCELLED. */ -void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd); - -/* Register read interest, causing read_cb to be called once when fd becomes - readable, on deadline specified by deadline, or on shutdown triggered by - grpc_fd_shutdown. - read_cb will be called with read_cb_arg when *fd becomes readable. - read_cb is Called with status of GRPC_CALLBACK_SUCCESS if readable, - GRPC_CALLBACK_TIMED_OUT if the call timed out, - and CANCELLED if the call was cancelled. - - Requires:This method must not be called before the read_cb for any previous - call runs. Edge triggered events are used whenever they are supported by the - underlying platform. This means that users must drain fd in read_cb before - calling notify_on_read again. Users are also expected to handle spurious - events, i.e read_cb is called while nothing can be readable from fd */ -void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure); - -/* Exactly the same semantics as above, except based on writable events. */ -void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd, - grpc_closure *closure); - -/* Notification from the poller to an fd that it has become readable or - writable. - If allow_synchronous_callback is 1, allow running the fd callback inline - in this callstack, otherwise register an asynchronous callback and return */ -void grpc_fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd); -void grpc_fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd); - -/* Reference counting for fds */ -/*#define GRPC_FD_REF_COUNT_DEBUG*/ -#ifdef GRPC_FD_REF_COUNT_DEBUG -void grpc_fd_ref(grpc_fd *fd, const char *reason, const char *file, int line); -void grpc_fd_unref(grpc_fd *fd, const char *reason, const char *file, int line); -#define GRPC_FD_REF(fd, reason) grpc_fd_ref(fd, reason, __FILE__, __LINE__) -#define GRPC_FD_UNREF(fd, reason) grpc_fd_unref(fd, reason, __FILE__, __LINE__) -#else -void grpc_fd_ref(grpc_fd *fd); -void grpc_fd_unref(grpc_fd *fd); -#define GRPC_FD_REF(fd, reason) grpc_fd_ref(fd) -#define GRPC_FD_UNREF(fd, reason) grpc_fd_unref(fd) -#endif - -void grpc_fd_global_init(void); -void grpc_fd_global_shutdown(void); - -#endif /* GRPC_CORE_LIB_IOMGR_FD_POSIX_H */ diff --git a/src/core/lib/iomgr/iomgr.c b/src/core/lib/iomgr/iomgr.c index 5ac47f42f2..146663984d 100644 --- a/src/core/lib/iomgr/iomgr.c +++ b/src/core/lib/iomgr/iomgr.c @@ -63,7 +63,6 @@ void grpc_iomgr_init(void) { g_root_object.next = g_root_object.prev = &g_root_object; g_root_object.name = "root"; grpc_iomgr_platform_init(); - grpc_pollset_global_init(); } static size_t count_objects(void) { @@ -138,7 +137,6 @@ void grpc_iomgr_shutdown(void) { gpr_mu_lock(&g_mu); gpr_mu_unlock(&g_mu); - grpc_pollset_global_shutdown(); grpc_iomgr_platform_shutdown(); grpc_exec_ctx_global_shutdown(); gpr_mu_destroy(&g_mu); diff --git a/src/core/lib/iomgr/iomgr_internal.h b/src/core/lib/iomgr/iomgr_internal.h index 87d5a3d672..805be1f1e4 100644 --- a/src/core/lib/iomgr/iomgr_internal.h +++ b/src/core/lib/iomgr/iomgr_internal.h @@ -36,7 +36,6 @@ #include <stdbool.h> -#include <grpc/support/sync.h> #include "src/core/lib/iomgr/iomgr.h" typedef struct grpc_iomgr_object { @@ -45,9 +44,6 @@ typedef struct grpc_iomgr_object { struct grpc_iomgr_object *prev; } grpc_iomgr_object; -void grpc_pollset_global_init(void); -void grpc_pollset_global_shutdown(void); - void grpc_iomgr_register_object(grpc_iomgr_object *obj, const char *name); void grpc_iomgr_unregister_object(grpc_iomgr_object *obj); diff --git a/src/core/lib/iomgr/iomgr_posix.c b/src/core/lib/iomgr/iomgr_posix.c index 5d909bd220..016c501f75 100644 --- a/src/core/lib/iomgr/iomgr_posix.c +++ b/src/core/lib/iomgr/iomgr_posix.c @@ -36,17 +36,17 @@ #ifdef GPR_POSIX_SOCKET #include "src/core/lib/debug/trace.h" -#include "src/core/lib/iomgr/fd_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/iomgr_posix.h" #include "src/core/lib/iomgr/tcp_posix.h" void grpc_iomgr_platform_init(void) { - grpc_fd_global_init(); + grpc_event_engine_init(); grpc_register_tracer("tcp", &grpc_tcp_trace); } void grpc_iomgr_platform_flush(void) {} -void grpc_iomgr_platform_shutdown(void) { grpc_fd_global_shutdown(); } +void grpc_iomgr_platform_shutdown(void) { grpc_event_engine_shutdown(); } #endif /* GRPC_POSIX_SOCKET */ diff --git a/src/core/lib/iomgr/iomgr_windows.c b/src/core/lib/iomgr/iomgr_windows.c index 25c88425c4..398517fc75 100644 --- a/src/core/lib/iomgr/iomgr_windows.c +++ b/src/core/lib/iomgr/iomgr_windows.c @@ -41,6 +41,7 @@ #include "src/core/lib/iomgr/iocp_windows.h" #include "src/core/lib/iomgr/iomgr.h" +#include "src/core/lib/iomgr/pollset_windows.h" #include "src/core/lib/iomgr/socket_windows.h" /* Windows' io manager is going to be fully designed using IO completion @@ -61,11 +62,13 @@ static void winsock_shutdown(void) { void grpc_iomgr_platform_init(void) { winsock_init(); grpc_iocp_init(); + grpc_pollset_global_init(); } void grpc_iomgr_platform_flush(void) { grpc_iocp_flush(); } void grpc_iomgr_platform_shutdown(void) { + grpc_pollset_global_shutdown(); grpc_iocp_shutdown(); winsock_shutdown(); } diff --git a/src/core/lib/iomgr/pollset_multipoller_with_epoll.c b/src/core/lib/iomgr/pollset_multipoller_with_epoll.c deleted file mode 100644 index 04d912e94d..0000000000 --- a/src/core/lib/iomgr/pollset_multipoller_with_epoll.c +++ /dev/null @@ -1,324 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL - -#include <errno.h> -#include <poll.h> -#include <string.h> -#include <sys/epoll.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/useful.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/support/block_annotate.h" - -struct epoll_fd_list { - int *epoll_fds; - size_t count; - size_t capacity; -}; - -static struct epoll_fd_list epoll_fd_global_list; -static gpr_once init_epoll_fd_list_mu = GPR_ONCE_INIT; -static gpr_mu epoll_fd_list_mu; - -static void init_mu(void) { gpr_mu_init(&epoll_fd_list_mu); } - -static void add_epoll_fd_to_global_list(int epoll_fd) { - gpr_once_init(&init_epoll_fd_list_mu, init_mu); - - gpr_mu_lock(&epoll_fd_list_mu); - if (epoll_fd_global_list.count == epoll_fd_global_list.capacity) { - epoll_fd_global_list.capacity = - GPR_MAX((size_t)8, epoll_fd_global_list.capacity * 2); - epoll_fd_global_list.epoll_fds = - gpr_realloc(epoll_fd_global_list.epoll_fds, - epoll_fd_global_list.capacity * sizeof(int)); - } - epoll_fd_global_list.epoll_fds[epoll_fd_global_list.count++] = epoll_fd; - gpr_mu_unlock(&epoll_fd_list_mu); -} - -static void remove_epoll_fd_from_global_list(int epoll_fd) { - gpr_mu_lock(&epoll_fd_list_mu); - GPR_ASSERT(epoll_fd_global_list.count > 0); - for (size_t i = 0; i < epoll_fd_global_list.count; i++) { - if (epoll_fd == epoll_fd_global_list.epoll_fds[i]) { - epoll_fd_global_list.epoll_fds[i] = - epoll_fd_global_list.epoll_fds[--(epoll_fd_global_list.count)]; - break; - } - } - gpr_mu_unlock(&epoll_fd_list_mu); -} - -void grpc_remove_fd_from_all_epoll_sets(int fd) { - int err; - gpr_once_init(&init_epoll_fd_list_mu, init_mu); - gpr_mu_lock(&epoll_fd_list_mu); - if (epoll_fd_global_list.count == 0) { - gpr_mu_unlock(&epoll_fd_list_mu); - return; - } - for (size_t i = 0; i < epoll_fd_global_list.count; i++) { - err = epoll_ctl(epoll_fd_global_list.epoll_fds[i], EPOLL_CTL_DEL, fd, NULL); - if (err < 0 && errno != ENOENT) { - gpr_log(GPR_ERROR, "epoll_ctl del for %d failed: %s", fd, - strerror(errno)); - } - } - gpr_mu_unlock(&epoll_fd_list_mu); -} - -typedef struct { - grpc_pollset *pollset; - grpc_fd *fd; - grpc_closure closure; -} delayed_add; - -typedef struct { int epoll_fd; } pollset_hdr; - -static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd) { - pollset_hdr *h = pollset->data.ptr; - struct epoll_event ev; - int err; - grpc_fd_watcher watcher; - - /* We pretend to be polling whilst adding an fd to keep the fd from being - closed during the add. This may result in a spurious wakeup being assigned - to this pollset whilst adding, but that should be benign. */ - GPR_ASSERT(grpc_fd_begin_poll(fd, pollset, NULL, 0, 0, &watcher) == 0); - if (watcher.fd != NULL) { - ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET); - ev.data.ptr = fd; - err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev); - if (err < 0) { - /* FDs may be added to a pollset multiple times, so EEXIST is normal. */ - if (errno != EEXIST) { - gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", fd->fd, - strerror(errno)); - } - } - } - grpc_fd_end_poll(exec_ctx, &watcher, 0, 0); -} - -static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg, - bool iomgr_status) { - delayed_add *da = arg; - - if (!grpc_fd_is_orphaned(da->fd)) { - finally_add_fd(exec_ctx, da->pollset, da->fd); - } - - gpr_mu_lock(&da->pollset->mu); - da->pollset->in_flight_cbs--; - if (da->pollset->shutting_down) { - /* We don't care about this pollset anymore. */ - if (da->pollset->in_flight_cbs == 0 && !da->pollset->called_shutdown) { - da->pollset->called_shutdown = 1; - grpc_exec_ctx_enqueue(exec_ctx, da->pollset->shutdown_done, true, NULL); - } - } - gpr_mu_unlock(&da->pollset->mu); - - GRPC_FD_UNREF(da->fd, "delayed_add"); - - gpr_free(da); -} - -static void multipoll_with_epoll_pollset_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_fd *fd, - int and_unlock_pollset) { - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - finally_add_fd(exec_ctx, pollset, fd); - } else { - delayed_add *da = gpr_malloc(sizeof(*da)); - da->pollset = pollset; - da->fd = fd; - GRPC_FD_REF(fd, "delayed_add"); - grpc_closure_init(&da->closure, perform_delayed_add, da); - pollset->in_flight_cbs++; - grpc_exec_ctx_enqueue(exec_ctx, &da->closure, true, NULL); - } -} - -/* TODO(klempner): We probably want to turn this down a bit */ -#define GRPC_EPOLL_MAX_EVENTS 1000 - -static void multipoll_with_epoll_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { - struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS]; - int ep_rv; - int poll_rv; - pollset_hdr *h = pollset->data.ptr; - int timeout_ms; - struct pollfd pfds[2]; - - /* If you want to ignore epoll's ability to sanely handle parallel pollers, - * for a more apples-to-apples performance comparison with poll, add a - * if (pollset->counter != 0) { return 0; } - * here. - */ - - gpr_mu_unlock(&pollset->mu); - - timeout_ms = grpc_poll_deadline_to_millis_timeout(deadline, now); - - pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfds[0].events = POLLIN; - pfds[0].revents = 0; - pfds[1].fd = h->epoll_fd; - pfds[1].events = POLLIN; - pfds[1].revents = 0; - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - GPR_TIMER_BEGIN("poll", 0); - GRPC_SCHEDULING_START_BLOCKING_REGION; - poll_rv = grpc_poll_function(pfds, 2, timeout_ms); - GRPC_SCHEDULING_END_BLOCKING_REGION; - GPR_TIMER_END("poll", 0); - - if (poll_rv < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); - } - } else if (poll_rv == 0) { - /* do nothing */ - } else { - if (pfds[0].revents) { - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); - } - if (pfds[1].revents) { - do { - /* The following epoll_wait never blocks; it has a timeout of 0 */ - ep_rv = epoll_wait(h->epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0); - if (ep_rv < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "epoll_wait() failed: %s", strerror(errno)); - } - } else { - int i; - for (i = 0; i < ep_rv; ++i) { - grpc_fd *fd = ep_ev[i].data.ptr; - /* TODO(klempner): We might want to consider making err and pri - * separate events */ - int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP); - int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI); - int write_ev = ep_ev[i].events & EPOLLOUT; - if (fd == NULL) { - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); - } else { - if (read_ev || cancel) { - grpc_fd_become_readable(exec_ctx, fd); - } - if (write_ev || cancel) { - grpc_fd_become_writable(exec_ctx, fd); - } - } - } - } - } while (ep_rv == GRPC_EPOLL_MAX_EVENTS); - } - } -} - -static void multipoll_with_epoll_pollset_finish_shutdown( - grpc_pollset *pollset) {} - -static void multipoll_with_epoll_pollset_destroy(grpc_pollset *pollset) { - pollset_hdr *h = pollset->data.ptr; - close(h->epoll_fd); - remove_epoll_fd_from_global_list(h->epoll_fd); - gpr_free(h); -} - -static const grpc_pollset_vtable multipoll_with_epoll_pollset = { - multipoll_with_epoll_pollset_add_fd, - multipoll_with_epoll_pollset_maybe_work_and_unlock, - multipoll_with_epoll_pollset_finish_shutdown, - multipoll_with_epoll_pollset_destroy}; - -static void epoll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, grpc_fd **fds, - size_t nfds) { - size_t i; - pollset_hdr *h = gpr_malloc(sizeof(pollset_hdr)); - struct epoll_event ev; - int err; - - pollset->vtable = &multipoll_with_epoll_pollset; - pollset->data.ptr = h; - h->epoll_fd = epoll_create1(EPOLL_CLOEXEC); - if (h->epoll_fd < 0) { - /* TODO(klempner): Fall back to poll here, especially on ENOSYS */ - gpr_log(GPR_ERROR, "epoll_create1 failed: %s", strerror(errno)); - abort(); - } - add_epoll_fd_to_global_list(h->epoll_fd); - - ev.events = (uint32_t)(EPOLLIN | EPOLLET); - ev.data.ptr = NULL; - err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, - GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), &ev); - if (err < 0) { - gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", - GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), - strerror(errno)); - } - - for (i = 0; i < nfds; i++) { - multipoll_with_epoll_pollset_add_fd(exec_ctx, pollset, fds[i], 0); - } -} - -grpc_platform_become_multipoller_type grpc_platform_become_multipoller = - epoll_become_multipoller; - -#else /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ - -void grpc_remove_fd_from_all_epoll_sets(int fd) {} - -#endif /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */ diff --git a/src/core/lib/iomgr/pollset_multipoller_with_poll_posix.c b/src/core/lib/iomgr/pollset_multipoller_with_poll_posix.c deleted file mode 100644 index 7668bf63f5..0000000000 --- a/src/core/lib/iomgr/pollset_multipoller_with_poll_posix.c +++ /dev/null @@ -1,234 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include "src/core/lib/iomgr/pollset_posix.h" - -#include <errno.h> -#include <poll.h> -#include <stdlib.h> -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/support/block_annotate.h" - -typedef struct { - /* all polled fds */ - size_t fd_count; - size_t fd_capacity; - grpc_fd **fds; - /* fds that have been removed from the pollset explicitly */ - size_t del_count; - size_t del_capacity; - grpc_fd **dels; -} pollset_hdr; - -static void multipoll_with_poll_pollset_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_fd *fd, - int and_unlock_pollset) { - size_t i; - pollset_hdr *h = pollset->data.ptr; - /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */ - for (i = 0; i < h->fd_count; i++) { - if (h->fds[i] == fd) goto exit; - } - if (h->fd_count == h->fd_capacity) { - h->fd_capacity = GPR_MAX(h->fd_capacity + 8, h->fd_count * 3 / 2); - h->fds = gpr_realloc(h->fds, sizeof(grpc_fd *) * h->fd_capacity); - } - h->fds[h->fd_count++] = fd; - GRPC_FD_REF(fd, "multipoller"); -exit: - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - } -} - -static void multipoll_with_poll_pollset_maybe_work_and_unlock( - grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now) { -#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) -#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) - - int timeout; - int r; - size_t i, j, fd_count; - nfds_t pfd_count; - pollset_hdr *h; - /* TODO(ctiller): inline some elements to avoid an allocation */ - grpc_fd_watcher *watchers; - struct pollfd *pfds; - - h = pollset->data.ptr; - timeout = grpc_poll_deadline_to_millis_timeout(deadline, now); - /* TODO(ctiller): perform just one malloc here if we exceed the inline case */ - pfds = gpr_malloc(sizeof(*pfds) * (h->fd_count + 2)); - watchers = gpr_malloc(sizeof(*watchers) * (h->fd_count + 2)); - fd_count = 0; - pfd_count = 2; - pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); - pfds[0].events = POLLIN; - pfds[0].revents = 0; - pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfds[1].events = POLLIN; - pfds[1].revents = 0; - for (i = 0; i < h->fd_count; i++) { - int remove = grpc_fd_is_orphaned(h->fds[i]); - for (j = 0; !remove && j < h->del_count; j++) { - if (h->fds[i] == h->dels[j]) remove = 1; - } - if (remove) { - GRPC_FD_UNREF(h->fds[i], "multipoller"); - } else { - h->fds[fd_count++] = h->fds[i]; - watchers[pfd_count].fd = h->fds[i]; - GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start"); - pfds[pfd_count].fd = h->fds[i]->fd; - pfds[pfd_count].revents = 0; - pfd_count++; - } - } - for (j = 0; j < h->del_count; j++) { - GRPC_FD_UNREF(h->dels[j], "multipoller_del"); - } - h->del_count = 0; - h->fd_count = fd_count; - gpr_mu_unlock(&pollset->mu); - - for (i = 2; i < pfd_count; i++) { - grpc_fd *fd = watchers[i].fd; - pfds[i].events = (short)grpc_fd_begin_poll(fd, pollset, worker, POLLIN, - POLLOUT, &watchers[i]); - GRPC_FD_UNREF(fd, "multipoller_start"); - } - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - GRPC_SCHEDULING_START_BLOCKING_REGION; - r = grpc_poll_function(pfds, pfd_count, timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - - if (r < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); - } - for (i = 2; i < pfd_count; i++) { - grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0); - } - } else if (r == 0) { - for (i = 2; i < pfd_count; i++) { - grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0); - } - } else { - if (pfds[0].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); - } - if (pfds[1].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); - } - for (i = 2; i < pfd_count; i++) { - if (watchers[i].fd == NULL) { - grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0); - continue; - } - grpc_fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK, - pfds[i].revents & POLLOUT_CHECK); - } - } - - gpr_free(pfds); - gpr_free(watchers); -} - -static void multipoll_with_poll_pollset_finish_shutdown(grpc_pollset *pollset) { - size_t i; - pollset_hdr *h = pollset->data.ptr; - for (i = 0; i < h->fd_count; i++) { - GRPC_FD_UNREF(h->fds[i], "multipoller"); - } - for (i = 0; i < h->del_count; i++) { - GRPC_FD_UNREF(h->dels[i], "multipoller_del"); - } - h->fd_count = 0; - h->del_count = 0; -} - -static void multipoll_with_poll_pollset_destroy(grpc_pollset *pollset) { - pollset_hdr *h = pollset->data.ptr; - multipoll_with_poll_pollset_finish_shutdown(pollset); - gpr_free(h->fds); - gpr_free(h->dels); - gpr_free(h); -} - -static const grpc_pollset_vtable multipoll_with_poll_pollset = { - multipoll_with_poll_pollset_add_fd, - multipoll_with_poll_pollset_maybe_work_and_unlock, - multipoll_with_poll_pollset_finish_shutdown, - multipoll_with_poll_pollset_destroy}; - -void grpc_poll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, grpc_fd **fds, - size_t nfds) { - size_t i; - pollset_hdr *h = gpr_malloc(sizeof(pollset_hdr)); - pollset->vtable = &multipoll_with_poll_pollset; - pollset->data.ptr = h; - h->fd_count = nfds; - h->fd_capacity = nfds; - h->fds = gpr_malloc(nfds * sizeof(grpc_fd *)); - h->del_count = 0; - h->del_capacity = 0; - h->dels = NULL; - for (i = 0; i < nfds; i++) { - h->fds[i] = fds[i]; - GRPC_FD_REF(fds[i], "multipoller"); - } -} - -#endif /* GPR_POSIX_SOCKET */ - -#ifdef GPR_POSIX_MULTIPOLL_WITH_POLL -grpc_platform_become_multipoller_type grpc_platform_become_multipoller = - grpc_poll_become_multipoller; -#endif diff --git a/src/core/lib/iomgr/pollset_posix.c b/src/core/lib/iomgr/pollset_posix.c deleted file mode 100644 index 84a9153034..0000000000 --- a/src/core/lib/iomgr/pollset_posix.c +++ /dev/null @@ -1,633 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include "src/core/lib/iomgr/pollset_posix.h" - -#include <errno.h> -#include <stdlib.h> -#include <string.h> -#include <unistd.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/log.h> -#include <grpc/support/thd.h> -#include <grpc/support/tls.h> -#include <grpc/support/useful.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/iomgr_internal.h" -#include "src/core/lib/iomgr/socket_utils_posix.h" -#include "src/core/lib/profiling/timers.h" -#include "src/core/lib/support/block_annotate.h" - -GPR_TLS_DECL(g_current_thread_poller); -GPR_TLS_DECL(g_current_thread_worker); - -/** Default poll() function - a pointer so that it can be overridden by some - * tests */ -grpc_poll_function_type grpc_poll_function = poll; - -/** 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. */ -grpc_wakeup_fd grpc_global_wakeup_fd; - -static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->prev->next = worker->next; - worker->next->prev = worker->prev; -} - -int grpc_pollset_has_workers(grpc_pollset *p) { - return p->root_worker.next != &p->root_worker; -} - -static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) { - if (grpc_pollset_has_workers(p)) { - grpc_pollset_worker *w = p->root_worker.next; - remove_worker(p, w); - return w; - } else { - return NULL; - } -} - -static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->next = &p->root_worker; - worker->prev = worker->next->prev; - worker->prev->next = worker->next->prev = worker; -} - -static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) { - worker->prev = &p->root_worker; - worker->next = worker->prev->next; - worker->prev->next = worker->next->prev = worker; -} - -size_t grpc_pollset_size(void) { return sizeof(grpc_pollset); } - -void grpc_pollset_kick_ext(grpc_pollset *p, - grpc_pollset_worker *specific_worker, - uint32_t flags) { - GPR_TIMER_BEGIN("grpc_pollset_kick_ext", 0); - - /* pollset->mu already held */ - if (specific_worker != NULL) { - if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) { - GPR_TIMER_BEGIN("grpc_pollset_kick_ext.broadcast", 0); - GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); - for (specific_worker = p->root_worker.next; - specific_worker != &p->root_worker; - specific_worker = specific_worker->next) { - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } - p->kicked_without_pollers = 1; - GPR_TIMER_END("grpc_pollset_kick_ext.broadcast", 0); - } else if (gpr_tls_get(&g_current_thread_worker) != - (intptr_t)specific_worker) { - GPR_TIMER_MARK("different_thread_worker", 0); - if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { - specific_worker->reevaluate_polling_on_wakeup = 1; - } - specific_worker->kicked_specifically = 1; - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) { - GPR_TIMER_MARK("kick_yoself", 0); - if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) { - specific_worker->reevaluate_polling_on_wakeup = 1; - } - specific_worker->kicked_specifically = 1; - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } - } else if (gpr_tls_get(&g_current_thread_poller) != (intptr_t)p) { - GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0); - GPR_TIMER_MARK("kick_anonymous", 0); - specific_worker = pop_front_worker(p); - if (specific_worker != NULL) { - if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) { - GPR_TIMER_MARK("kick_anonymous_not_self", 0); - push_back_worker(p, specific_worker); - specific_worker = pop_front_worker(p); - if ((flags & GRPC_POLLSET_CAN_KICK_SELF) == 0 && - gpr_tls_get(&g_current_thread_worker) == - (intptr_t)specific_worker) { - push_back_worker(p, specific_worker); - specific_worker = NULL; - } - } - if (specific_worker != NULL) { - GPR_TIMER_MARK("finally_kick", 0); - push_back_worker(p, specific_worker); - grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd); - } - } else { - GPR_TIMER_MARK("kicked_no_pollers", 0); - p->kicked_without_pollers = 1; - } - } - - GPR_TIMER_END("grpc_pollset_kick_ext", 0); -} - -void grpc_pollset_kick(grpc_pollset *p, grpc_pollset_worker *specific_worker) { - grpc_pollset_kick_ext(p, specific_worker, 0); -} - -/* global state management */ - -void grpc_pollset_global_init(void) { - gpr_tls_init(&g_current_thread_poller); - gpr_tls_init(&g_current_thread_worker); - grpc_wakeup_fd_global_init(); - grpc_wakeup_fd_init(&grpc_global_wakeup_fd); -} - -void grpc_pollset_global_shutdown(void) { - grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd); - gpr_tls_destroy(&g_current_thread_poller); - gpr_tls_destroy(&g_current_thread_worker); - grpc_wakeup_fd_global_destroy(); -} - -void grpc_kick_poller(void) { grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); } - -/* main interface */ - -static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null); - -void grpc_pollset_init(grpc_pollset *pollset, gpr_mu **mu) { - gpr_mu_init(&pollset->mu); - *mu = &pollset->mu; - pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker; - pollset->in_flight_cbs = 0; - pollset->shutting_down = 0; - pollset->called_shutdown = 0; - pollset->kicked_without_pollers = 0; - pollset->idle_jobs.head = pollset->idle_jobs.tail = NULL; - pollset->local_wakeup_cache = NULL; - pollset->kicked_without_pollers = 0; - become_basic_pollset(pollset, NULL); -} - -void grpc_pollset_destroy(grpc_pollset *pollset) { - GPR_ASSERT(pollset->in_flight_cbs == 0); - GPR_ASSERT(!grpc_pollset_has_workers(pollset)); - GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); - pollset->vtable->destroy(pollset); - while (pollset->local_wakeup_cache) { - grpc_cached_wakeup_fd *next = pollset->local_wakeup_cache->next; - grpc_wakeup_fd_destroy(&pollset->local_wakeup_cache->fd); - gpr_free(pollset->local_wakeup_cache); - pollset->local_wakeup_cache = next; - } -} - -void grpc_pollset_reset(grpc_pollset *pollset) { - GPR_ASSERT(pollset->shutting_down); - GPR_ASSERT(pollset->in_flight_cbs == 0); - GPR_ASSERT(!grpc_pollset_has_workers(pollset)); - GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail); - pollset->vtable->destroy(pollset); - pollset->shutting_down = 0; - pollset->called_shutdown = 0; - pollset->kicked_without_pollers = 0; - become_basic_pollset(pollset, NULL); -} - -void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd) { - gpr_mu_lock(&pollset->mu); - pollset->vtable->add_fd(exec_ctx, pollset, fd, 1); -/* the following (enabled only in debug) will reacquire and then release - our lock - meaning that if the unlocking flag passed to add_fd above is - not respected, the code will deadlock (in a way that we have a chance of - debugging) */ -#ifndef NDEBUG - gpr_mu_lock(&pollset->mu); - gpr_mu_unlock(&pollset->mu); -#endif -} - -static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) { - GPR_ASSERT(grpc_closure_list_empty(pollset->idle_jobs)); - pollset->vtable->finish_shutdown(pollset); - grpc_exec_ctx_enqueue(exec_ctx, pollset->shutdown_done, true, NULL); -} - -void grpc_pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_pollset_worker **worker_hdl, gpr_timespec now, - gpr_timespec deadline) { - grpc_pollset_worker worker; - *worker_hdl = &worker; - - /* pollset->mu already held */ - int added_worker = 0; - int locked = 1; - int queued_work = 0; - int keep_polling = 0; - GPR_TIMER_BEGIN("grpc_pollset_work", 0); - /* this must happen before we (potentially) drop pollset->mu */ - worker.next = worker.prev = NULL; - worker.reevaluate_polling_on_wakeup = 0; - if (pollset->local_wakeup_cache != NULL) { - worker.wakeup_fd = pollset->local_wakeup_cache; - pollset->local_wakeup_cache = worker.wakeup_fd->next; - } else { - worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd)); - grpc_wakeup_fd_init(&worker.wakeup_fd->fd); - } - worker.kicked_specifically = 0; - /* If there's work waiting for the pollset to be idle, and the - pollset is idle, then do that work */ - if (!grpc_pollset_has_workers(pollset) && - !grpc_closure_list_empty(pollset->idle_jobs)) { - GPR_TIMER_MARK("grpc_pollset_work.idle_jobs", 0); - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - goto done; - } - /* If we're shutting down then we don't execute any extended work */ - if (pollset->shutting_down) { - GPR_TIMER_MARK("grpc_pollset_work.shutting_down", 0); - goto done; - } - /* Give do_promote priority so we don't starve it out */ - if (pollset->in_flight_cbs) { - GPR_TIMER_MARK("grpc_pollset_work.in_flight_cbs", 0); - gpr_mu_unlock(&pollset->mu); - locked = 0; - goto done; - } - /* Start polling, and keep doing so while we're being asked to - re-evaluate our pollers (this allows poll() based pollers to - ensure they don't miss wakeups) */ - keep_polling = 1; - while (keep_polling) { - keep_polling = 0; - if (!pollset->kicked_without_pollers) { - if (!added_worker) { - push_front_worker(pollset, &worker); - added_worker = 1; - gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker); - } - gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset); - GPR_TIMER_BEGIN("maybe_work_and_unlock", 0); - pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, &worker, - deadline, now); - GPR_TIMER_END("maybe_work_and_unlock", 0); - locked = 0; - gpr_tls_set(&g_current_thread_poller, 0); - } else { - GPR_TIMER_MARK("grpc_pollset_work.kicked_without_pollers", 0); - pollset->kicked_without_pollers = 0; - } - /* Finished execution - start cleaning up. - Note that we may arrive here from outside the enclosing while() loop. - In that case we won't loop though as we haven't added worker to the - worker list, which means nobody could ask us to re-evaluate polling). */ - done: - if (!locked) { - queued_work |= grpc_exec_ctx_flush(exec_ctx); - gpr_mu_lock(&pollset->mu); - locked = 1; - } - /* If we're forced to re-evaluate polling (via grpc_pollset_kick with - GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) then we land here and force - a loop */ - if (worker.reevaluate_polling_on_wakeup) { - worker.reevaluate_polling_on_wakeup = 0; - pollset->kicked_without_pollers = 0; - if (queued_work || worker.kicked_specifically) { - /* If there's queued work on the list, then set the deadline to be - immediate so we get back out of the polling loop quickly */ - deadline = gpr_inf_past(GPR_CLOCK_MONOTONIC); - } - keep_polling = 1; - } - } - if (added_worker) { - remove_worker(pollset, &worker); - gpr_tls_set(&g_current_thread_worker, 0); - } - /* release wakeup fd to the local pool */ - worker.wakeup_fd->next = pollset->local_wakeup_cache; - pollset->local_wakeup_cache = worker.wakeup_fd; - /* check shutdown conditions */ - if (pollset->shutting_down) { - if (grpc_pollset_has_workers(pollset)) { - grpc_pollset_kick(pollset, NULL); - } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) { - pollset->called_shutdown = 1; - gpr_mu_unlock(&pollset->mu); - finish_shutdown(exec_ctx, pollset); - grpc_exec_ctx_flush(exec_ctx); - /* Continuing to access pollset here is safe -- it is the caller's - * responsibility to not destroy when it has outstanding calls to - * grpc_pollset_work. - * TODO(dklempner): Can we refactor the shutdown logic to avoid this? */ - gpr_mu_lock(&pollset->mu); - } else if (!grpc_closure_list_empty(pollset->idle_jobs)) { - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - gpr_mu_unlock(&pollset->mu); - grpc_exec_ctx_flush(exec_ctx); - gpr_mu_lock(&pollset->mu); - } - } - *worker_hdl = NULL; - GPR_TIMER_END("grpc_pollset_work", 0); -} - -void grpc_pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_closure *closure) { - GPR_ASSERT(!pollset->shutting_down); - pollset->shutting_down = 1; - pollset->shutdown_done = closure; - grpc_pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - if (!grpc_pollset_has_workers(pollset)) { - grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL); - } - if (!pollset->called_shutdown && pollset->in_flight_cbs == 0 && - !grpc_pollset_has_workers(pollset)) { - pollset->called_shutdown = 1; - finish_shutdown(exec_ctx, pollset); - } -} - -int grpc_poll_deadline_to_millis_timeout(gpr_timespec deadline, - gpr_timespec now) { - gpr_timespec timeout; - static const int64_t max_spin_polling_us = 10; - if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) { - return -1; - } - if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros( - max_spin_polling_us, - GPR_TIMESPAN))) <= 0) { - return 0; - } - timeout = gpr_time_sub(deadline, now); - return gpr_time_to_millis(gpr_time_add( - timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN))); -} - -/* - * basic_pollset - a vtable that provides polling for zero or one file - * descriptor via poll() - */ - -typedef struct grpc_unary_promote_args { - const grpc_pollset_vtable *original_vtable; - grpc_pollset *pollset; - grpc_fd *fd; - grpc_closure promotion_closure; -} grpc_unary_promote_args; - -static void basic_do_promote(grpc_exec_ctx *exec_ctx, void *args, - bool success) { - grpc_unary_promote_args *up_args = args; - const grpc_pollset_vtable *original_vtable = up_args->original_vtable; - grpc_pollset *pollset = up_args->pollset; - grpc_fd *fd = up_args->fd; - - /* - * This is quite tricky. There are a number of cases to keep in mind here: - * 1. fd may have been orphaned - * 2. The pollset may no longer be a unary poller (and we can't let case #1 - * leak to other pollset types!) - * 3. pollset's fd (which may have changed) may have been orphaned - * 4. The pollset may be shutting down. - */ - - gpr_mu_lock(&pollset->mu); - /* First we need to ensure that nobody is polling concurrently */ - GPR_ASSERT(!grpc_pollset_has_workers(pollset)); - - gpr_free(up_args); - /* At this point the pollset may no longer be a unary poller. In that case - * we should just call the right add function and be done. */ - /* TODO(klempner): If we're not careful this could cause infinite recursion. - * That's not a problem for now because empty_pollset has a trivial poller - * and we don't have any mechanism to unbecome multipoller. */ - pollset->in_flight_cbs--; - if (pollset->shutting_down) { - /* We don't care about this pollset anymore. */ - if (pollset->in_flight_cbs == 0 && !pollset->called_shutdown) { - pollset->called_shutdown = 1; - finish_shutdown(exec_ctx, pollset); - } - } else if (grpc_fd_is_orphaned(fd)) { - /* Don't try to add it to anything, we'll drop our ref on it below */ - } else if (pollset->vtable != original_vtable) { - pollset->vtable->add_fd(exec_ctx, pollset, fd, 0); - } else if (fd != pollset->data.ptr) { - grpc_fd *fds[2]; - fds[0] = pollset->data.ptr; - fds[1] = fd; - - if (fds[0] && !grpc_fd_is_orphaned(fds[0])) { - grpc_platform_become_multipoller(exec_ctx, pollset, fds, - GPR_ARRAY_SIZE(fds)); - GRPC_FD_UNREF(fds[0], "basicpoll"); - } else { - /* old fd is orphaned and we haven't cleaned it up until now, so remain a - * unary poller */ - /* Note that it is possible that fds[1] is also orphaned at this point. - * That's okay, we'll correct it at the next add or poll. */ - if (fds[0]) GRPC_FD_UNREF(fds[0], "basicpoll"); - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } - } - - gpr_mu_unlock(&pollset->mu); - - /* Matching ref in basic_pollset_add_fd */ - GRPC_FD_UNREF(fd, "basicpoll_add"); -} - -static void basic_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_fd *fd, int and_unlock_pollset) { - grpc_unary_promote_args *up_args; - GPR_ASSERT(fd); - if (fd == pollset->data.ptr) goto exit; - - if (!grpc_pollset_has_workers(pollset)) { - /* Fast path -- no in flight cbs */ - /* TODO(klempner): Comment this out and fix any test failures or establish - * they are due to timing issues */ - grpc_fd *fds[2]; - fds[0] = pollset->data.ptr; - fds[1] = fd; - - if (fds[0] == NULL) { - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } else if (!grpc_fd_is_orphaned(fds[0])) { - grpc_platform_become_multipoller(exec_ctx, pollset, fds, - GPR_ARRAY_SIZE(fds)); - GRPC_FD_UNREF(fds[0], "basicpoll"); - } else { - /* old fd is orphaned and we haven't cleaned it up until now, so remain a - * unary poller */ - GRPC_FD_UNREF(fds[0], "basicpoll"); - pollset->data.ptr = fd; - GRPC_FD_REF(fd, "basicpoll"); - } - goto exit; - } - - /* Now we need to promote. This needs to happen when we're not polling. Since - * this may be called from poll, the wait needs to happen asynchronously. */ - GRPC_FD_REF(fd, "basicpoll_add"); - pollset->in_flight_cbs++; - up_args = gpr_malloc(sizeof(*up_args)); - up_args->fd = fd; - up_args->original_vtable = pollset->vtable; - up_args->pollset = pollset; - up_args->promotion_closure.cb = basic_do_promote; - up_args->promotion_closure.cb_arg = up_args; - - grpc_closure_list_add(&pollset->idle_jobs, &up_args->promotion_closure, 1); - grpc_pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST); - -exit: - if (and_unlock_pollset) { - gpr_mu_unlock(&pollset->mu); - } -} - -static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - grpc_pollset_worker *worker, - gpr_timespec deadline, - gpr_timespec now) { -#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR) -#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR) - - struct pollfd pfd[3]; - grpc_fd *fd; - grpc_fd_watcher fd_watcher; - int timeout; - int r; - nfds_t nfds; - - fd = pollset->data.ptr; - if (fd && grpc_fd_is_orphaned(fd)) { - GRPC_FD_UNREF(fd, "basicpoll"); - fd = pollset->data.ptr = NULL; - } - timeout = grpc_poll_deadline_to_millis_timeout(deadline, now); - pfd[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd); - pfd[0].events = POLLIN; - pfd[0].revents = 0; - pfd[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd); - pfd[1].events = POLLIN; - pfd[1].revents = 0; - nfds = 2; - if (fd) { - pfd[2].fd = fd->fd; - pfd[2].revents = 0; - GRPC_FD_REF(fd, "basicpoll_begin"); - gpr_mu_unlock(&pollset->mu); - pfd[2].events = (short)grpc_fd_begin_poll(fd, pollset, worker, POLLIN, - POLLOUT, &fd_watcher); - if (pfd[2].events != 0) { - nfds++; - } - } else { - gpr_mu_unlock(&pollset->mu); - } - - /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid - even going into the blocking annotation if possible */ - /* poll fd count (argument 2) is shortened by one if we have no events - to poll on - such that it only includes the kicker */ - GPR_TIMER_BEGIN("poll", 0); - GRPC_SCHEDULING_START_BLOCKING_REGION; - r = grpc_poll_function(pfd, nfds, timeout); - GRPC_SCHEDULING_END_BLOCKING_REGION; - GPR_TIMER_END("poll", 0); - - if (r < 0) { - if (errno != EINTR) { - gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno)); - } - if (fd) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0); - } - } else if (r == 0) { - if (fd) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0); - } - } else { - if (pfd[0].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd); - } - if (pfd[1].revents & POLLIN_CHECK) { - grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd); - } - if (nfds > 2) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK, - pfd[2].revents & POLLOUT_CHECK); - } else if (fd) { - grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0); - } - } - - if (fd) { - GRPC_FD_UNREF(fd, "basicpoll_begin"); - } -} - -static void basic_pollset_destroy(grpc_pollset *pollset) { - if (pollset->data.ptr != NULL) { - GRPC_FD_UNREF(pollset->data.ptr, "basicpoll"); - pollset->data.ptr = NULL; - } -} - -static const grpc_pollset_vtable basic_pollset = { - basic_pollset_add_fd, basic_pollset_maybe_work_and_unlock, - basic_pollset_destroy, basic_pollset_destroy}; - -static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null) { - pollset->vtable = &basic_pollset; - pollset->data.ptr = fd_or_null; - if (fd_or_null != NULL) { - GRPC_FD_REF(fd_or_null, "basicpoll"); - } -} - -#endif /* GPR_POSIX_POLLSET */ diff --git a/src/core/lib/iomgr/pollset_posix.h b/src/core/lib/iomgr/pollset_posix.h deleted file mode 100644 index f53ef8855c..0000000000 --- a/src/core/lib/iomgr/pollset_posix.h +++ /dev/null @@ -1,153 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_POLLSET_POSIX_H -#define GRPC_CORE_LIB_IOMGR_POLLSET_POSIX_H - -#include <poll.h> - -#include <grpc/support/sync.h> - -#include "src/core/lib/iomgr/exec_ctx.h" -#include "src/core/lib/iomgr/iomgr.h" -#include "src/core/lib/iomgr/pollset.h" -#include "src/core/lib/iomgr/wakeup_fd_posix.h" - -typedef struct grpc_pollset_vtable grpc_pollset_vtable; - -/* forward declare only in this file to avoid leaking impl details via - pollset.h; real users of grpc_fd should always include 'fd_posix.h' and not - use the struct tag */ -struct grpc_fd; - -typedef struct grpc_cached_wakeup_fd { - grpc_wakeup_fd fd; - struct grpc_cached_wakeup_fd *next; -} grpc_cached_wakeup_fd; - -struct grpc_pollset_worker { - grpc_cached_wakeup_fd *wakeup_fd; - int reevaluate_polling_on_wakeup; - int kicked_specifically; - struct grpc_pollset_worker *next; - struct grpc_pollset_worker *prev; -}; - -struct grpc_pollset { - /* pollsets under posix can mutate representation as fds are added and - removed. - For example, we may choose a poll() based implementation on linux for - few fds, and an epoll() based implementation for many fds */ - const grpc_pollset_vtable *vtable; - gpr_mu mu; - grpc_pollset_worker root_worker; - int in_flight_cbs; - int shutting_down; - int called_shutdown; - int kicked_without_pollers; - grpc_closure *shutdown_done; - grpc_closure_list idle_jobs; - union { - int fd; - void *ptr; - } data; - /* Local cache of eventfds for workers */ - grpc_cached_wakeup_fd *local_wakeup_cache; -}; - -struct grpc_pollset_vtable { - void (*add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - struct grpc_fd *fd, int and_unlock_pollset); - void (*maybe_work_and_unlock)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - grpc_pollset_worker *worker, - gpr_timespec deadline, gpr_timespec now); - void (*finish_shutdown)(grpc_pollset *pollset); - void (*destroy)(grpc_pollset *pollset); -}; - -/* Add an fd to a pollset */ -void grpc_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, - struct grpc_fd *fd); - -/* Returns the fd to listen on for kicks */ -int grpc_kick_read_fd(grpc_pollset *p); -/* Call after polling has been kicked to leave the kicked state */ -void grpc_kick_drain(grpc_pollset *p); - -/* Convert a timespec to milliseconds: - - very small or negative poll times are clamped to zero to do a - non-blocking poll (which becomes spin polling) - - other small values are rounded up to one millisecond - - longer than a millisecond polls are rounded up to the next nearest - millisecond to avoid spinning - - infinite timeouts are converted to -1 */ -int grpc_poll_deadline_to_millis_timeout(gpr_timespec deadline, - gpr_timespec now); - -/* Allow kick to wakeup the currently polling worker */ -#define GRPC_POLLSET_CAN_KICK_SELF 1 -/* Force the wakee to repoll when awoken */ -#define GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP 2 -/* As per grpc_pollset_kick, with an extended set of flags (defined above) - -- mostly for fd_posix's use. */ -void grpc_pollset_kick_ext(grpc_pollset *p, - grpc_pollset_worker *specific_worker, - uint32_t flags); - -/* turn a pollset into a multipoller: platform specific */ -typedef void (*grpc_platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, - struct grpc_fd **fds, - size_t fd_count); -extern grpc_platform_become_multipoller_type grpc_platform_become_multipoller; - -void grpc_poll_become_multipoller(grpc_exec_ctx *exec_ctx, - grpc_pollset *pollset, struct grpc_fd **fds, - size_t fd_count); - -/* Return 1 if the pollset has active threads in grpc_pollset_work (pollset must - * be locked) */ -int grpc_pollset_has_workers(grpc_pollset *pollset); - -void grpc_remove_fd_from_all_epoll_sets(int fd); - -/* override to allow tests to hook poll() usage */ -/* NOTE: Any changes to grpc_poll_function must take place when the gRPC - is certainly not doing any polling anywhere. - Otherwise, there might be a race between changing the variable and actually - doing a polling operation */ -typedef int (*grpc_poll_function_type)(struct pollfd *, nfds_t, int); -extern grpc_poll_function_type grpc_poll_function; -extern grpc_wakeup_fd grpc_global_wakeup_fd; - -#endif /* GRPC_CORE_LIB_IOMGR_POLLSET_POSIX_H */ diff --git a/src/core/lib/iomgr/pollset_set_posix.c b/src/core/lib/iomgr/pollset_set_posix.c deleted file mode 100644 index ef85075f38..0000000000 --- a/src/core/lib/iomgr/pollset_set_posix.c +++ /dev/null @@ -1,202 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#include <grpc/support/port_platform.h> - -#ifdef GPR_POSIX_SOCKET - -#include <stdlib.h> -#include <string.h> - -#include <grpc/support/alloc.h> -#include <grpc/support/useful.h> - -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/iomgr/pollset_set_posix.h" - -struct grpc_pollset_set { - gpr_mu mu; - - size_t pollset_count; - size_t pollset_capacity; - grpc_pollset **pollsets; - - size_t pollset_set_count; - size_t pollset_set_capacity; - struct grpc_pollset_set **pollset_sets; - - size_t fd_count; - size_t fd_capacity; - grpc_fd **fds; -}; - -grpc_pollset_set *grpc_pollset_set_create(void) { - grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set)); - memset(pollset_set, 0, sizeof(*pollset_set)); - gpr_mu_init(&pollset_set->mu); - return pollset_set; -} - -void grpc_pollset_set_destroy(grpc_pollset_set *pollset_set) { - size_t i; - gpr_mu_destroy(&pollset_set->mu); - for (i = 0; i < pollset_set->fd_count; i++) { - GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); - } - gpr_free(pollset_set->pollsets); - gpr_free(pollset_set->pollset_sets); - gpr_free(pollset_set->fds); - gpr_free(pollset_set); -} - -void grpc_pollset_set_add_pollset(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, - grpc_pollset *pollset) { - size_t i, j; - gpr_mu_lock(&pollset_set->mu); - if (pollset_set->pollset_count == pollset_set->pollset_capacity) { - pollset_set->pollset_capacity = - GPR_MAX(8, 2 * pollset_set->pollset_capacity); - pollset_set->pollsets = - gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity * - sizeof(*pollset_set->pollsets)); - } - pollset_set->pollsets[pollset_set->pollset_count++] = pollset; - for (i = 0, j = 0; i < pollset_set->fd_count; i++) { - if (grpc_fd_is_orphaned(pollset_set->fds[i])) { - GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set"); - } else { - grpc_pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]); - pollset_set->fds[j++] = pollset_set->fds[i]; - } - } - pollset_set->fd_count = j; - gpr_mu_unlock(&pollset_set->mu); -} - -void grpc_pollset_set_del_pollset(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, - grpc_pollset *pollset) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - for (i = 0; i < pollset_set->pollset_count; i++) { - if (pollset_set->pollsets[i] == pollset) { - pollset_set->pollset_count--; - GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i], - pollset_set->pollsets[pollset_set->pollset_count]); - break; - } - } - gpr_mu_unlock(&pollset_set->mu); -} - -void grpc_pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *bag, - grpc_pollset_set *item) { - size_t i, j; - gpr_mu_lock(&bag->mu); - if (bag->pollset_set_count == bag->pollset_set_capacity) { - bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity); - bag->pollset_sets = - gpr_realloc(bag->pollset_sets, - bag->pollset_set_capacity * sizeof(*bag->pollset_sets)); - } - bag->pollset_sets[bag->pollset_set_count++] = item; - for (i = 0, j = 0; i < bag->fd_count; i++) { - if (grpc_fd_is_orphaned(bag->fds[i])) { - GRPC_FD_UNREF(bag->fds[i], "pollset_set"); - } else { - grpc_pollset_set_add_fd(exec_ctx, item, bag->fds[i]); - bag->fds[j++] = bag->fds[i]; - } - } - bag->fd_count = j; - gpr_mu_unlock(&bag->mu); -} - -void grpc_pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *bag, - grpc_pollset_set *item) { - size_t i; - gpr_mu_lock(&bag->mu); - for (i = 0; i < bag->pollset_set_count; i++) { - if (bag->pollset_sets[i] == item) { - bag->pollset_set_count--; - GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i], - bag->pollset_sets[bag->pollset_set_count]); - break; - } - } - gpr_mu_unlock(&bag->mu); -} - -void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - if (pollset_set->fd_count == pollset_set->fd_capacity) { - pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity); - pollset_set->fds = gpr_realloc( - pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds)); - } - GRPC_FD_REF(fd, "pollset_set"); - pollset_set->fds[pollset_set->fd_count++] = fd; - for (i = 0; i < pollset_set->pollset_count; i++) { - grpc_pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd); - } - for (i = 0; i < pollset_set->pollset_set_count; i++) { - grpc_pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd); - } - gpr_mu_unlock(&pollset_set->mu); -} - -void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd) { - size_t i; - gpr_mu_lock(&pollset_set->mu); - for (i = 0; i < pollset_set->fd_count; i++) { - if (pollset_set->fds[i] == fd) { - pollset_set->fd_count--; - GPR_SWAP(grpc_fd *, pollset_set->fds[i], - pollset_set->fds[pollset_set->fd_count]); - GRPC_FD_UNREF(fd, "pollset_set"); - break; - } - } - for (i = 0; i < pollset_set->pollset_set_count; i++) { - grpc_pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd); - } - gpr_mu_unlock(&pollset_set->mu); -} - -#endif /* GPR_POSIX_SOCKET */ diff --git a/src/core/lib/iomgr/pollset_set_posix.h b/src/core/lib/iomgr/pollset_set_posix.h deleted file mode 100644 index e23e974921..0000000000 --- a/src/core/lib/iomgr/pollset_set_posix.h +++ /dev/null @@ -1,45 +0,0 @@ -/* - * - * Copyright 2015, Google Inc. - * All rights reserved. - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * * Neither the name of Google Inc. nor the names of its - * contributors may be used to endorse or promote products derived from - * this software without specific prior written permission. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - */ - -#ifndef GRPC_CORE_LIB_IOMGR_POLLSET_SET_POSIX_H -#define GRPC_CORE_LIB_IOMGR_POLLSET_SET_POSIX_H - -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_set.h" - -void grpc_pollset_set_add_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd); -void grpc_pollset_set_del_fd(grpc_exec_ctx *exec_ctx, - grpc_pollset_set *pollset_set, grpc_fd *fd); - -#endif /* GRPC_CORE_LIB_IOMGR_POLLSET_SET_POSIX_H */ diff --git a/src/core/lib/iomgr/pollset_windows.c b/src/core/lib/iomgr/pollset_windows.c index c2e7f647b5..bff5c586f8 100644 --- a/src/core/lib/iomgr/pollset_windows.c +++ b/src/core/lib/iomgr/pollset_windows.c @@ -47,7 +47,7 @@ gpr_mu grpc_polling_mu; static grpc_pollset_worker *g_active_poller; static grpc_pollset_worker g_global_root_worker; -void grpc_pollset_global_init() { +void grpc_pollset_global_init(void) { gpr_mu_init(&grpc_polling_mu); g_active_poller = NULL; g_global_root_worker.links[GRPC_POLLSET_WORKER_LINK_GLOBAL].next = @@ -55,7 +55,7 @@ void grpc_pollset_global_init() { &g_global_root_worker; } -void grpc_pollset_global_shutdown() { gpr_mu_destroy(&grpc_polling_mu); } +void grpc_pollset_global_shutdown(void) { gpr_mu_destroy(&grpc_polling_mu); } static void remove_worker(grpc_pollset_worker *worker, grpc_pollset_worker_link_type type) { diff --git a/src/core/lib/iomgr/pollset_windows.h b/src/core/lib/iomgr/pollset_windows.h index fef0e87b71..2642013afc 100644 --- a/src/core/lib/iomgr/pollset_windows.h +++ b/src/core/lib/iomgr/pollset_windows.h @@ -72,4 +72,7 @@ struct grpc_pollset { grpc_closure *on_shutdown; }; +void grpc_pollset_global_init(void); +void grpc_pollset_global_shutdown(void); + #endif /* GRPC_CORE_LIB_IOMGR_POLLSET_WINDOWS_H */ diff --git a/src/core/lib/iomgr/tcp_client_posix.c b/src/core/lib/iomgr/tcp_client_posix.c index 77718a5b5c..6430cb629f 100644 --- a/src/core/lib/iomgr/tcp_client_posix.c +++ b/src/core/lib/iomgr/tcp_client_posix.c @@ -47,9 +47,8 @@ #include <grpc/support/string_util.h> #include <grpc/support/time.h> +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/iomgr_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/iomgr/pollset_set_posix.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/socket_utils_posix.h" #include "src/core/lib/iomgr/tcp_posix.h" @@ -140,7 +139,8 @@ static void on_writable(grpc_exec_ctx *exec_ctx, void *acp, bool success) { if (success) { do { so_error_size = sizeof(so_error); - err = getsockopt(fd->fd, SOL_SOCKET, SO_ERROR, &so_error, &so_error_size); + err = getsockopt(grpc_fd_wrapped_fd(fd), SOL_SOCKET, SO_ERROR, &so_error, + &so_error_size); } while (err < 0 && errno == EINTR); if (err < 0) { gpr_log(GPR_ERROR, "failed to connect to '%s': getsockopt(ERROR): %s", diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c index 33ec29a411..7210aef5d5 100644 --- a/src/core/lib/iomgr/tcp_posix.c +++ b/src/core/lib/iomgr/tcp_posix.c @@ -52,8 +52,7 @@ #include <grpc/support/time.h> #include "src/core/lib/debug/trace.h" -#include "src/core/lib/iomgr/pollset_posix.h" -#include "src/core/lib/iomgr/pollset_set_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/profiling/timers.h" #include "src/core/lib/support/string.h" @@ -454,7 +453,7 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, size_t slice_size, grpc_tcp *tcp = (grpc_tcp *)gpr_malloc(sizeof(grpc_tcp)); tcp->base.vtable = &vtable; tcp->peer_string = gpr_strdup(peer_string); - tcp->fd = em_fd->fd; + tcp->fd = grpc_fd_wrapped_fd(em_fd); tcp->read_cb = NULL; tcp->write_cb = NULL; tcp->release_fd_cb = NULL; diff --git a/src/core/lib/iomgr/tcp_posix.h b/src/core/lib/iomgr/tcp_posix.h index acce56467a..99125836d6 100644 --- a/src/core/lib/iomgr/tcp_posix.h +++ b/src/core/lib/iomgr/tcp_posix.h @@ -45,7 +45,7 @@ */ #include "src/core/lib/iomgr/endpoint.h" -#include "src/core/lib/iomgr/fd_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #define GRPC_TCP_DEFAULT_READ_SLICE_SIZE 8192 diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c index 76aeb1c2f0..cfb5251684 100644 --- a/src/core/lib/iomgr/tcp_server_posix.c +++ b/src/core/lib/iomgr/tcp_server_posix.c @@ -59,7 +59,6 @@ #include <grpc/support/string_util.h> #include <grpc/support/sync.h> #include <grpc/support/time.h> -#include "src/core/lib/iomgr/pollset_posix.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/socket_utils_posix.h" diff --git a/src/core/lib/iomgr/udp_server.c b/src/core/lib/iomgr/udp_server.c index 2d8220a9b1..a0b9709be5 100644 --- a/src/core/lib/iomgr/udp_server.c +++ b/src/core/lib/iomgr/udp_server.c @@ -60,12 +60,10 @@ #include <grpc/support/string_util.h> #include <grpc/support/sync.h> #include <grpc/support/time.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" #include "src/core/lib/iomgr/resolve_address.h" #include "src/core/lib/iomgr/sockaddr_utils.h" #include "src/core/lib/iomgr/socket_utils_posix.h" -#include "src/core/lib/iomgr/unix_sockets_posix.h" #include "src/core/lib/support/string.h" #define INIT_PORT_CAP 2 diff --git a/src/core/lib/iomgr/udp_server.h b/src/core/lib/iomgr/udp_server.h index b0ca15101f..d8cf957a22 100644 --- a/src/core/lib/iomgr/udp_server.h +++ b/src/core/lib/iomgr/udp_server.h @@ -35,7 +35,7 @@ #define GRPC_CORE_LIB_IOMGR_UDP_SERVER_H #include "src/core/lib/iomgr/endpoint.h" -#include "src/core/lib/iomgr/fd_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" /* Forward decl of struct grpc_server */ /* This is not typedef'ed to avoid a typedef-redefinition error */ diff --git a/src/core/lib/iomgr/workqueue_posix.c b/src/core/lib/iomgr/workqueue_posix.c index 9d68d9e16a..80e7a0b206 100644 --- a/src/core/lib/iomgr/workqueue_posix.c +++ b/src/core/lib/iomgr/workqueue_posix.c @@ -43,8 +43,7 @@ #include <grpc/support/log.h> #include <grpc/support/useful.h> -#include "src/core/lib/iomgr/fd_posix.h" -#include "src/core/lib/iomgr/pollset_posix.h" +#include "src/core/lib/iomgr/ev_posix.h" static void on_readable(grpc_exec_ctx *exec_ctx, void *arg, bool success); diff --git a/src/core/lib/surface/init.c b/src/core/lib/surface/init.c index 566b030912..aca4ce9d07 100644 --- a/src/core/lib/surface/init.c +++ b/src/core/lib/surface/init.c @@ -39,9 +39,7 @@ #include <grpc/grpc.h> #include <grpc/support/alloc.h> #include <grpc/support/time.h> -/* TODO(ctiller): find another way? - better not to include census here */ #include "src/core/ext/transport/chttp2/transport/chttp2_transport.h" -#include "src/core/lib/census/grpc_plugin.h" #include "src/core/lib/channel/channel_stack.h" #include "src/core/lib/channel/client_channel.h" #include "src/core/lib/channel/compress_filter.h" @@ -50,8 +48,6 @@ #include "src/core/lib/channel/http_server_filter.h" #include "src/core/lib/client_config/lb_policy_registry.h" #include "src/core/lib/client_config/resolver_registry.h" -#include "src/core/lib/client_config/resolvers/dns_resolver.h" -#include "src/core/lib/client_config/resolvers/sockaddr_resolver.h" #include "src/core/lib/client_config/subchannel.h" #include "src/core/lib/client_config/subchannel_index.h" #include "src/core/lib/debug/trace.h" @@ -85,8 +81,6 @@ static int g_initializations; static void do_basic_init(void) { gpr_mu_init(&g_init_mu); grpc_register_built_in_plugins(); - /* TODO(ctiller): ideally remove this strict linkage */ - grpc_register_plugin(census_grpc_plugin_init, census_grpc_plugin_destroy); g_initializations = 0; } @@ -169,12 +163,6 @@ void grpc_init(void) { grpc_channel_init_init(); grpc_lb_policy_registry_init(); grpc_resolver_registry_init(GRPC_DEFAULT_NAME_PREFIX); - grpc_register_resolver_type(grpc_dns_resolver_factory_create()); - grpc_register_resolver_type(grpc_ipv4_resolver_factory_create()); - grpc_register_resolver_type(grpc_ipv6_resolver_factory_create()); -#ifdef GPR_HAVE_UNIX_SOCKET - grpc_register_resolver_type(grpc_unix_resolver_factory_create()); -#endif grpc_register_tracer("api", &grpc_api_trace); grpc_register_tracer("channel", &grpc_trace_channel); grpc_register_tracer("http", &grpc_http_trace); diff --git a/src/core/plugin_registry/grpc_plugin_registry.c b/src/core/plugin_registry/grpc_plugin_registry.c index 3e3c214c22..79df85516e 100644 --- a/src/core/plugin_registry/grpc_plugin_registry.c +++ b/src/core/plugin_registry/grpc_plugin_registry.c @@ -37,10 +37,22 @@ extern void grpc_lb_policy_pick_first_init(void); extern void grpc_lb_policy_pick_first_shutdown(void); extern void grpc_lb_policy_round_robin_init(void); extern void grpc_lb_policy_round_robin_shutdown(void); +extern void grpc_resolver_dns_native_init(void); +extern void grpc_resolver_dns_native_shutdown(void); +extern void grpc_resolver_sockaddr_init(void); +extern void grpc_resolver_sockaddr_shutdown(void); +extern void census_grpc_plugin_init(void); +extern void census_grpc_plugin_shutdown(void); void grpc_register_built_in_plugins(void) { grpc_register_plugin(grpc_lb_policy_pick_first_init, grpc_lb_policy_pick_first_shutdown); grpc_register_plugin(grpc_lb_policy_round_robin_init, grpc_lb_policy_round_robin_shutdown); + grpc_register_plugin(grpc_resolver_dns_native_init, + grpc_resolver_dns_native_shutdown); + grpc_register_plugin(grpc_resolver_sockaddr_init, + grpc_resolver_sockaddr_shutdown); + grpc_register_plugin(census_grpc_plugin_init, + census_grpc_plugin_shutdown); } diff --git a/src/core/plugin_registry/grpc_unsecure_plugin_registry.c b/src/core/plugin_registry/grpc_unsecure_plugin_registry.c index 3e3c214c22..b3786c927d 100644 --- a/src/core/plugin_registry/grpc_unsecure_plugin_registry.c +++ b/src/core/plugin_registry/grpc_unsecure_plugin_registry.c @@ -33,14 +33,26 @@ #include <grpc/grpc.h> +extern void grpc_resolver_dns_native_init(void); +extern void grpc_resolver_dns_native_shutdown(void); +extern void grpc_resolver_sockaddr_init(void); +extern void grpc_resolver_sockaddr_shutdown(void); extern void grpc_lb_policy_pick_first_init(void); extern void grpc_lb_policy_pick_first_shutdown(void); extern void grpc_lb_policy_round_robin_init(void); extern void grpc_lb_policy_round_robin_shutdown(void); +extern void census_grpc_plugin_init(void); +extern void census_grpc_plugin_shutdown(void); void grpc_register_built_in_plugins(void) { + grpc_register_plugin(grpc_resolver_dns_native_init, + grpc_resolver_dns_native_shutdown); + grpc_register_plugin(grpc_resolver_sockaddr_init, + grpc_resolver_sockaddr_shutdown); grpc_register_plugin(grpc_lb_policy_pick_first_init, grpc_lb_policy_pick_first_shutdown); grpc_register_plugin(grpc_lb_policy_round_robin_init, grpc_lb_policy_round_robin_shutdown); + grpc_register_plugin(census_grpc_plugin_init, + census_grpc_plugin_shutdown); } |